1 /*
2 * Copyright (c) 2001, 2011, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/symbolTable.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "code/codeCache.hpp"
29 #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp"
30 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp"
31 #include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.hpp"
32 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp"
33 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp"
34 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp"
35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp"
36 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp"
37 #include "gc_implementation/parNew/parNewGeneration.hpp"
38 #include "gc_implementation/shared/collectorCounters.hpp"
39 #include "gc_implementation/shared/isGCActiveMark.hpp"
40 #include "gc_interface/collectedHeap.inline.hpp"
41 #include "memory/cardTableRS.hpp"
42 #include "memory/collectorPolicy.hpp"
43 #include "memory/gcLocker.inline.hpp"
44 #include "memory/genCollectedHeap.hpp"
45 #include "memory/genMarkSweep.hpp"
46 #include "memory/genOopClosures.inline.hpp"
47 #include "memory/iterator.hpp"
48 #include "memory/referencePolicy.hpp"
49 #include "memory/resourceArea.hpp"
50 #include "oops/oop.inline.hpp"
51 #include "prims/jvmtiExport.hpp"
52 #include "runtime/globals_extension.hpp"
53 #include "runtime/handles.inline.hpp"
54 #include "runtime/java.hpp"
55 #include "runtime/vmThread.hpp"
56 #include "services/memoryService.hpp"
57 #include "services/runtimeService.hpp"
58
59 // statics
60 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
61 bool CMSCollector::_full_gc_requested = false;
62
63 //////////////////////////////////////////////////////////////////
64 // In support of CMS/VM thread synchronization
65 //////////////////////////////////////////////////////////////////
66 // We split use of the CGC_lock into 2 "levels".
67 // The low-level locking is of the usual CGC_lock monitor. We introduce
68 // a higher level "token" (hereafter "CMS token") built on top of the
69 // low level monitor (hereafter "CGC lock").
70 // The token-passing protocol gives priority to the VM thread. The
71 // CMS-lock doesn't provide any fairness guarantees, but clients
72 // should ensure that it is only held for very short, bounded
73 // durations.
74 //
75 // When either of the CMS thread or the VM thread is involved in
76 // collection operations during which it does not want the other
77 // thread to interfere, it obtains the CMS token.
78 //
79 // If either thread tries to get the token while the other has
80 // it, that thread waits. However, if the VM thread and CMS thread
81 // both want the token, then the VM thread gets priority while the
82 // CMS thread waits. This ensures, for instance, that the "concurrent"
83 // phases of the CMS thread's work do not block out the VM thread
84 // for long periods of time as the CMS thread continues to hog
85 // the token. (See bug 4616232).
86 //
87 // The baton-passing functions are, however, controlled by the
88 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
89 // and here the low-level CMS lock, not the high level token,
90 // ensures mutual exclusion.
91 //
92 // Two important conditions that we have to satisfy:
93 // 1. if a thread does a low-level wait on the CMS lock, then it
94 // relinquishes the CMS token if it were holding that token
95 // when it acquired the low-level CMS lock.
96 // 2. any low-level notifications on the low-level lock
97 // should only be sent when a thread has relinquished the token.
98 //
99 // In the absence of either property, we'd have potential deadlock.
100 //
101 // We protect each of the CMS (concurrent and sequential) phases
102 // with the CMS _token_, not the CMS _lock_.
103 //
104 // The only code protected by CMS lock is the token acquisition code
105 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
106 // baton-passing code.
107 //
108 // Unfortunately, i couldn't come up with a good abstraction to factor and
109 // hide the naked CGC_lock manipulation in the baton-passing code
110 // further below. That's something we should try to do. Also, the proof
111 // of correctness of this 2-level locking scheme is far from obvious,
112 // and potentially quite slippery. We have an uneasy supsicion, for instance,
113 // that there may be a theoretical possibility of delay/starvation in the
114 // low-level lock/wait/notify scheme used for the baton-passing because of
115 // potential intereference with the priority scheme embodied in the
116 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
117 // invocation further below and marked with "XXX 20011219YSR".
118 // Indeed, as we note elsewhere, this may become yet more slippery
119 // in the presence of multiple CMS and/or multiple VM threads. XXX
120
121 class CMSTokenSync: public StackObj {
122 private:
123 bool _is_cms_thread;
124 public:
125 CMSTokenSync(bool is_cms_thread):
126 _is_cms_thread(is_cms_thread) {
127 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
128 "Incorrect argument to constructor");
129 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
130 }
131
132 ~CMSTokenSync() {
133 assert(_is_cms_thread ?
134 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
135 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
136 "Incorrect state");
137 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
138 }
139 };
140
141 // Convenience class that does a CMSTokenSync, and then acquires
142 // upto three locks.
143 class CMSTokenSyncWithLocks: public CMSTokenSync {
144 private:
145 // Note: locks are acquired in textual declaration order
146 // and released in the opposite order
147 MutexLockerEx _locker1, _locker2, _locker3;
148 public:
149 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
150 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
151 CMSTokenSync(is_cms_thread),
152 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
153 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
154 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
155 { }
156 };
157
158
159 // Wrapper class to temporarily disable icms during a foreground cms collection.
160 class ICMSDisabler: public StackObj {
161 public:
162 // The ctor disables icms and wakes up the thread so it notices the change;
163 // the dtor re-enables icms. Note that the CMSCollector methods will check
164 // CMSIncrementalMode.
165 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); }
166 ~ICMSDisabler() { CMSCollector::enable_icms(); }
167 };
168
169 //////////////////////////////////////////////////////////////////
170 // Concurrent Mark-Sweep Generation /////////////////////////////
171 //////////////////////////////////////////////////////////////////
172
173 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
174
175 // This struct contains per-thread things necessary to support parallel
176 // young-gen collection.
177 class CMSParGCThreadState: public CHeapObj {
178 public:
179 CFLS_LAB lab;
180 PromotionInfo promo;
181
182 // Constructor.
183 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
184 promo.setSpace(cfls);
185 }
186 };
187
188 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
189 ReservedSpace rs, size_t initial_byte_size, int level,
190 CardTableRS* ct, bool use_adaptive_freelists,
191 FreeBlockDictionary::DictionaryChoice dictionaryChoice) :
192 CardGeneration(rs, initial_byte_size, level, ct),
193 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
194 _debug_collection_type(Concurrent_collection_type)
195 {
196 HeapWord* bottom = (HeapWord*) _virtual_space.low();
197 HeapWord* end = (HeapWord*) _virtual_space.high();
198
199 _direct_allocated_words = 0;
200 NOT_PRODUCT(
201 _numObjectsPromoted = 0;
202 _numWordsPromoted = 0;
203 _numObjectsAllocated = 0;
204 _numWordsAllocated = 0;
205 )
206
207 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
208 use_adaptive_freelists,
209 dictionaryChoice);
210 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
211 if (_cmsSpace == NULL) {
212 vm_exit_during_initialization(
213 "CompactibleFreeListSpace allocation failure");
214 }
215 _cmsSpace->_gen = this;
216
217 _gc_stats = new CMSGCStats();
218
219 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
220 // offsets match. The ability to tell free chunks from objects
221 // depends on this property.
222 debug_only(
223 FreeChunk* junk = NULL;
224 assert(UseCompressedOops ||
225 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
226 "Offset of FreeChunk::_prev within FreeChunk must match"
227 " that of OopDesc::_klass within OopDesc");
228 )
229 if (CollectedHeap::use_parallel_gc_threads()) {
230 typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
231 _par_gc_thread_states =
232 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads);
233 if (_par_gc_thread_states == NULL) {
234 vm_exit_during_initialization("Could not allocate par gc structs");
235 }
236 for (uint i = 0; i < ParallelGCThreads; i++) {
237 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
238 if (_par_gc_thread_states[i] == NULL) {
239 vm_exit_during_initialization("Could not allocate par gc structs");
240 }
241 }
242 } else {
243 _par_gc_thread_states = NULL;
244 }
245 _incremental_collection_failed = false;
246 // The "dilatation_factor" is the expansion that can occur on
247 // account of the fact that the minimum object size in the CMS
248 // generation may be larger than that in, say, a contiguous young
249 // generation.
250 // Ideally, in the calculation below, we'd compute the dilatation
251 // factor as: MinChunkSize/(promoting_gen's min object size)
252 // Since we do not have such a general query interface for the
253 // promoting generation, we'll instead just use the mimimum
254 // object size (which today is a header's worth of space);
255 // note that all arithmetic is in units of HeapWords.
256 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
257 assert(_dilatation_factor >= 1.0, "from previous assert");
258 }
259
260
261 // The field "_initiating_occupancy" represents the occupancy percentage
262 // at which we trigger a new collection cycle. Unless explicitly specified
263 // via CMSInitiating[Perm]OccupancyFraction (argument "io" below), it
264 // is calculated by:
265 //
266 // Let "f" be MinHeapFreeRatio in
267 //
268 // _intiating_occupancy = 100-f +
269 // f * (CMSTrigger[Perm]Ratio/100)
270 // where CMSTrigger[Perm]Ratio is the argument "tr" below.
271 //
272 // That is, if we assume the heap is at its desired maximum occupancy at the
273 // end of a collection, we let CMSTrigger[Perm]Ratio of the (purported) free
274 // space be allocated before initiating a new collection cycle.
275 //
276 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) {
277 assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments");
278 if (io >= 0) {
279 _initiating_occupancy = (double)io / 100.0;
280 } else {
281 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
282 (double)(tr * MinHeapFreeRatio) / 100.0)
283 / 100.0;
284 }
285 }
286
287 void ConcurrentMarkSweepGeneration::ref_processor_init() {
288 assert(collector() != NULL, "no collector");
289 collector()->ref_processor_init();
290 }
291
292 void CMSCollector::ref_processor_init() {
293 if (_ref_processor == NULL) {
294 // Allocate and initialize a reference processor
295 _ref_processor =
296 new ReferenceProcessor(_span, // span
297 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
298 (int) ParallelGCThreads, // mt processing degree
299 _cmsGen->refs_discovery_is_mt(), // mt discovery
300 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
301 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
302 &_is_alive_closure, // closure for liveness info
303 false); // next field updates do not need write barrier
304 // Initialize the _ref_processor field of CMSGen
305 _cmsGen->set_ref_processor(_ref_processor);
306
307 // Allocate a dummy ref processor for perm gen.
308 ReferenceProcessor* rp2 = new ReferenceProcessor();
309 if (rp2 == NULL) {
310 vm_exit_during_initialization("Could not allocate ReferenceProcessor object");
311 }
312 _permGen->set_ref_processor(rp2);
313 }
314 }
315
316 CMSAdaptiveSizePolicy* CMSCollector::size_policy() {
317 GenCollectedHeap* gch = GenCollectedHeap::heap();
318 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
319 "Wrong type of heap");
320 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
321 gch->gen_policy()->size_policy();
322 assert(sp->is_gc_cms_adaptive_size_policy(),
323 "Wrong type of size policy");
324 return sp;
325 }
326
327 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() {
328 CMSGCAdaptivePolicyCounters* results =
329 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters();
330 assert(
331 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
332 "Wrong gc policy counter kind");
333 return results;
334 }
335
336
337 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
338
339 const char* gen_name = "old";
340
341 // Generation Counters - generation 1, 1 subspace
342 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space);
343
344 _space_counters = new GSpaceCounters(gen_name, 0,
345 _virtual_space.reserved_size(),
346 this, _gen_counters);
347 }
348
349 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
350 _cms_gen(cms_gen)
351 {
352 assert(alpha <= 100, "bad value");
353 _saved_alpha = alpha;
354
355 // Initialize the alphas to the bootstrap value of 100.
356 _gc0_alpha = _cms_alpha = 100;
357
358 _cms_begin_time.update();
359 _cms_end_time.update();
360
361 _gc0_duration = 0.0;
362 _gc0_period = 0.0;
363 _gc0_promoted = 0;
364
365 _cms_duration = 0.0;
366 _cms_period = 0.0;
367 _cms_allocated = 0;
368
369 _cms_used_at_gc0_begin = 0;
370 _cms_used_at_gc0_end = 0;
371 _allow_duty_cycle_reduction = false;
372 _valid_bits = 0;
373 _icms_duty_cycle = CMSIncrementalDutyCycle;
374 }
375
376 double CMSStats::cms_free_adjustment_factor(size_t free) const {
377 // TBD: CR 6909490
378 return 1.0;
379 }
380
381 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
382 }
383
384 // If promotion failure handling is on use
385 // the padded average size of the promotion for each
386 // young generation collection.
387 double CMSStats::time_until_cms_gen_full() const {
388 size_t cms_free = _cms_gen->cmsSpace()->free();
389 GenCollectedHeap* gch = GenCollectedHeap::heap();
390 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(),
391 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
392 if (cms_free > expected_promotion) {
393 // Start a cms collection if there isn't enough space to promote
394 // for the next minor collection. Use the padded average as
395 // a safety factor.
396 cms_free -= expected_promotion;
397
398 // Adjust by the safety factor.
399 double cms_free_dbl = (double)cms_free;
400 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
401 // Apply a further correction factor which tries to adjust
402 // for recent occurance of concurrent mode failures.
403 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
404 cms_free_dbl = cms_free_dbl * cms_adjustment;
405
406 if (PrintGCDetails && Verbose) {
407 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
408 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
409 cms_free, expected_promotion);
410 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
411 cms_free_dbl, cms_consumption_rate() + 1.0);
412 }
413 // Add 1 in case the consumption rate goes to zero.
414 return cms_free_dbl / (cms_consumption_rate() + 1.0);
415 }
416 return 0.0;
417 }
418
419 // Compare the duration of the cms collection to the
420 // time remaining before the cms generation is empty.
421 // Note that the time from the start of the cms collection
422 // to the start of the cms sweep (less than the total
423 // duration of the cms collection) can be used. This
424 // has been tried and some applications experienced
425 // promotion failures early in execution. This was
426 // possibly because the averages were not accurate
427 // enough at the beginning.
428 double CMSStats::time_until_cms_start() const {
429 // We add "gc0_period" to the "work" calculation
430 // below because this query is done (mostly) at the
431 // end of a scavenge, so we need to conservatively
432 // account for that much possible delay
433 // in the query so as to avoid concurrent mode failures
434 // due to starting the collection just a wee bit too
435 // late.
436 double work = cms_duration() + gc0_period();
437 double deadline = time_until_cms_gen_full();
438 // If a concurrent mode failure occurred recently, we want to be
439 // more conservative and halve our expected time_until_cms_gen_full()
440 if (work > deadline) {
441 if (Verbose && PrintGCDetails) {
442 gclog_or_tty->print(
443 " CMSCollector: collect because of anticipated promotion "
444 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
445 gc0_period(), time_until_cms_gen_full());
446 }
447 return 0.0;
448 }
449 return work - deadline;
450 }
451
452 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the
453 // amount of change to prevent wild oscillation.
454 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle,
455 unsigned int new_duty_cycle) {
456 assert(old_duty_cycle <= 100, "bad input value");
457 assert(new_duty_cycle <= 100, "bad input value");
458
459 // Note: use subtraction with caution since it may underflow (values are
460 // unsigned). Addition is safe since we're in the range 0-100.
461 unsigned int damped_duty_cycle = new_duty_cycle;
462 if (new_duty_cycle < old_duty_cycle) {
463 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U);
464 if (new_duty_cycle + largest_delta < old_duty_cycle) {
465 damped_duty_cycle = old_duty_cycle - largest_delta;
466 }
467 } else if (new_duty_cycle > old_duty_cycle) {
468 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U);
469 if (new_duty_cycle > old_duty_cycle + largest_delta) {
470 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U);
471 }
472 }
473 assert(damped_duty_cycle <= 100, "invalid duty cycle computed");
474
475 if (CMSTraceIncrementalPacing) {
476 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ",
477 old_duty_cycle, new_duty_cycle, damped_duty_cycle);
478 }
479 return damped_duty_cycle;
480 }
481
482 unsigned int CMSStats::icms_update_duty_cycle_impl() {
483 assert(CMSIncrementalPacing && valid(),
484 "should be handled in icms_update_duty_cycle()");
485
486 double cms_time_so_far = cms_timer().seconds();
487 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M;
488 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far);
489
490 // Avoid division by 0.
491 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01);
492 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full;
493
494 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U);
495 if (new_duty_cycle > _icms_duty_cycle) {
496 // Avoid very small duty cycles (1 or 2); 0 is allowed.
497 if (new_duty_cycle > 2) {
498 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle,
499 new_duty_cycle);
500 }
501 } else if (_allow_duty_cycle_reduction) {
502 // The duty cycle is reduced only once per cms cycle (see record_cms_end()).
503 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle);
504 // Respect the minimum duty cycle.
505 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin;
506 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle);
507 }
508
509 if (PrintGCDetails || CMSTraceIncrementalPacing) {
510 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle);
511 }
512
513 _allow_duty_cycle_reduction = false;
514 return _icms_duty_cycle;
515 }
516
517 #ifndef PRODUCT
518 void CMSStats::print_on(outputStream *st) const {
519 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
520 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
521 gc0_duration(), gc0_period(), gc0_promoted());
522 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
523 cms_duration(), cms_duration_per_mb(),
524 cms_period(), cms_allocated());
525 st->print(",cms_since_beg=%g,cms_since_end=%g",
526 cms_time_since_begin(), cms_time_since_end());
527 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
528 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
529 if (CMSIncrementalMode) {
530 st->print(",dc=%d", icms_duty_cycle());
531 }
532
533 if (valid()) {
534 st->print(",promo_rate=%g,cms_alloc_rate=%g",
535 promotion_rate(), cms_allocation_rate());
536 st->print(",cms_consumption_rate=%g,time_until_full=%g",
537 cms_consumption_rate(), time_until_cms_gen_full());
538 }
539 st->print(" ");
540 }
541 #endif // #ifndef PRODUCT
542
543 CMSCollector::CollectorState CMSCollector::_collectorState =
544 CMSCollector::Idling;
545 bool CMSCollector::_foregroundGCIsActive = false;
546 bool CMSCollector::_foregroundGCShouldWait = false;
547
548 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
549 ConcurrentMarkSweepGeneration* permGen,
550 CardTableRS* ct,
551 ConcurrentMarkSweepPolicy* cp):
552 _cmsGen(cmsGen),
553 _permGen(permGen),
554 _ct(ct),
555 _ref_processor(NULL), // will be set later
556 _conc_workers(NULL), // may be set later
557 _abort_preclean(false),
558 _start_sampling(false),
559 _between_prologue_and_epilogue(false),
560 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
561 _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_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 and perm gen
567 _span(cmsGen->reserved()._union(permGen->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_array(NULL), // may be set in ctor body
574 _eden_chunk_capacity(0), // -- ditto --
575 _eden_chunk_index(0), // -- ditto --
576 _survivor_plab_array(NULL), // -- ditto --
577 _survivor_chunk_array(NULL), // -- ditto --
578 _survivor_chunk_capacity(0), // -- ditto --
579 _survivor_chunk_index(0), // -- ditto --
580 _ser_pmc_preclean_ovflw(0),
581 _ser_kac_preclean_ovflw(0),
582 _ser_pmc_remark_ovflw(0),
583 _par_pmc_remark_ovflw(0),
584 _ser_kac_ovflw(0),
585 _par_kac_ovflw(0),
586 #ifndef PRODUCT
587 _num_par_pushes(0),
588 #endif
589 _collection_count_start(0),
590 _verifying(false),
591 _icms_start_limit(NULL),
592 _icms_stop_limit(NULL),
593 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
594 _completed_initialization(false),
595 _collector_policy(cp),
596 _should_unload_classes(false),
597 _concurrent_cycles_since_last_unload(0),
598 _roots_scanning_options(0),
599 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
600 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding)
601 {
602 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
603 ExplicitGCInvokesConcurrent = true;
604 }
605 // Now expand the span and allocate the collection support structures
606 // (MUT, marking bit map etc.) to cover both generations subject to
607 // collection.
608
609 // First check that _permGen is adjacent to _cmsGen and above it.
610 assert( _cmsGen->reserved().word_size() > 0
611 && _permGen->reserved().word_size() > 0,
612 "generations should not be of zero size");
613 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(),
614 "_cmsGen and _permGen should not overlap");
615 assert(_cmsGen->reserved().end() == _permGen->reserved().start(),
616 "_cmsGen->end() different from _permGen->start()");
617
618 // For use by dirty card to oop closures.
619 _cmsGen->cmsSpace()->set_collector(this);
620 _permGen->cmsSpace()->set_collector(this);
621
622 // Allocate MUT and marking bit map
623 {
624 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
625 if (!_markBitMap.allocate(_span)) {
626 warning("Failed to allocate CMS Bit Map");
627 return;
628 }
629 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
630 }
631 {
632 _modUnionTable.allocate(_span);
633 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
634 }
635
636 if (!_markStack.allocate(MarkStackSize)) {
637 warning("Failed to allocate CMS Marking Stack");
638 return;
639 }
640 if (!_revisitStack.allocate(CMSRevisitStackSize)) {
641 warning("Failed to allocate CMS Revisit Stack");
642 return;
643 }
644
645 // Support for multi-threaded concurrent phases
646 if (CMSConcurrentMTEnabled) {
647 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
648 // just for now
649 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
650 }
651 if (ConcGCThreads > 1) {
652 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
653 ConcGCThreads, true);
654 if (_conc_workers == NULL) {
655 warning("GC/CMS: _conc_workers allocation failure: "
656 "forcing -CMSConcurrentMTEnabled");
657 CMSConcurrentMTEnabled = false;
658 } else {
659 _conc_workers->initialize_workers();
660 }
661 } else {
662 CMSConcurrentMTEnabled = false;
663 }
664 }
665 if (!CMSConcurrentMTEnabled) {
666 ConcGCThreads = 0;
667 } else {
668 // Turn off CMSCleanOnEnter optimization temporarily for
669 // the MT case where it's not fixed yet; see 6178663.
670 CMSCleanOnEnter = false;
671 }
672 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
673 "Inconsistency");
674
675 // Parallel task queues; these are shared for the
676 // concurrent and stop-world phases of CMS, but
677 // are not shared with parallel scavenge (ParNew).
678 {
679 uint i;
680 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
681
682 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
683 || ParallelRefProcEnabled)
684 && num_queues > 0) {
685 _task_queues = new OopTaskQueueSet(num_queues);
686 if (_task_queues == NULL) {
687 warning("task_queues allocation failure.");
688 return;
689 }
690 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues);
691 if (_hash_seed == NULL) {
692 warning("_hash_seed array allocation failure");
693 return;
694 }
695
696 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
697 for (i = 0; i < num_queues; i++) {
698 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
699 if (q == NULL) {
700 warning("work_queue allocation failure.");
701 return;
702 }
703 _task_queues->register_queue(i, q);
704 }
705 for (i = 0; i < num_queues; i++) {
706 _task_queues->queue(i)->initialize();
707 _hash_seed[i] = 17; // copied from ParNew
708 }
709 }
710 }
711
712 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
713 _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio);
714
715 // Clip CMSBootstrapOccupancy between 0 and 100.
716 _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy)))
717 /(double)100;
718
719 _full_gcs_since_conc_gc = 0;
720
721 // Now tell CMS generations the identity of their collector
722 ConcurrentMarkSweepGeneration::set_collector(this);
723
724 // Create & start a CMS thread for this CMS collector
725 _cmsThread = ConcurrentMarkSweepThread::start(this);
726 assert(cmsThread() != NULL, "CMS Thread should have been created");
727 assert(cmsThread()->collector() == this,
728 "CMS Thread should refer to this gen");
729 assert(CGC_lock != NULL, "Where's the CGC_lock?");
730
731 // Support for parallelizing young gen rescan
732 GenCollectedHeap* gch = GenCollectedHeap::heap();
733 _young_gen = gch->prev_gen(_cmsGen);
734 if (gch->supports_inline_contig_alloc()) {
735 _top_addr = gch->top_addr();
736 _end_addr = gch->end_addr();
737 assert(_young_gen != NULL, "no _young_gen");
738 _eden_chunk_index = 0;
739 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
740 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity);
741 if (_eden_chunk_array == NULL) {
742 _eden_chunk_capacity = 0;
743 warning("GC/CMS: _eden_chunk_array allocation failure");
744 }
745 }
746 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
747
748 // Support for parallelizing survivor space rescan
749 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) {
750 const size_t max_plab_samples =
751 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize;
752
753 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads);
754 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples);
755 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads);
756 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
757 || _cursor == NULL) {
758 warning("Failed to allocate survivor plab/chunk array");
759 if (_survivor_plab_array != NULL) {
760 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
761 _survivor_plab_array = NULL;
762 }
763 if (_survivor_chunk_array != NULL) {
764 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
765 _survivor_chunk_array = NULL;
766 }
767 if (_cursor != NULL) {
768 FREE_C_HEAP_ARRAY(size_t, _cursor);
769 _cursor = NULL;
770 }
771 } else {
772 _survivor_chunk_capacity = 2*max_plab_samples;
773 for (uint i = 0; i < ParallelGCThreads; i++) {
774 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples);
775 if (vec == NULL) {
776 warning("Failed to allocate survivor plab array");
777 for (int j = i; j > 0; j--) {
778 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array());
779 }
780 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
781 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
782 _survivor_plab_array = NULL;
783 _survivor_chunk_array = NULL;
784 _survivor_chunk_capacity = 0;
785 break;
786 } else {
787 ChunkArray* cur =
788 ::new (&_survivor_plab_array[i]) ChunkArray(vec,
789 max_plab_samples);
790 assert(cur->end() == 0, "Should be 0");
791 assert(cur->array() == vec, "Should be vec");
792 assert(cur->capacity() == max_plab_samples, "Error");
793 }
794 }
795 }
796 }
797 assert( ( _survivor_plab_array != NULL
798 && _survivor_chunk_array != NULL)
799 || ( _survivor_chunk_capacity == 0
800 && _survivor_chunk_index == 0),
801 "Error");
802
803 // Choose what strong roots should be scanned depending on verification options
804 // and perm gen collection mode.
805 if (!CMSClassUnloadingEnabled) {
806 // If class unloading is disabled we want to include all classes into the root set.
807 add_root_scanning_option(SharedHeap::SO_AllClasses);
808 } else {
809 add_root_scanning_option(SharedHeap::SO_SystemClasses);
810 }
811
812 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
813 _gc_counters = new CollectorCounters("CMS", 1);
814 _completed_initialization = true;
815 _inter_sweep_timer.start(); // start of time
816 #ifdef SPARC
817 // Issue a stern warning, but allow use for experimentation and debugging.
818 if (VM_Version::is_sun4v() && UseMemSetInBOT) {
819 assert(!FLAG_IS_DEFAULT(UseMemSetInBOT), "Error");
820 warning("Experimental flag -XX:+UseMemSetInBOT is known to cause instability"
821 " on sun4v; please understand that you are using at your own risk!");
822 }
823 #endif
824 }
825
826 const char* ConcurrentMarkSweepGeneration::name() const {
827 return "concurrent mark-sweep generation";
828 }
829 void ConcurrentMarkSweepGeneration::update_counters() {
830 if (UsePerfData) {
831 _space_counters->update_all();
832 _gen_counters->update_all();
833 }
834 }
835
836 // this is an optimized version of update_counters(). it takes the
837 // used value as a parameter rather than computing it.
838 //
839 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
840 if (UsePerfData) {
841 _space_counters->update_used(used);
842 _space_counters->update_capacity();
843 _gen_counters->update_all();
844 }
845 }
846
847 void ConcurrentMarkSweepGeneration::print() const {
848 Generation::print();
849 cmsSpace()->print();
850 }
851
852 #ifndef PRODUCT
853 void ConcurrentMarkSweepGeneration::print_statistics() {
854 cmsSpace()->printFLCensus(0);
855 }
856 #endif
857
858 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
859 GenCollectedHeap* gch = GenCollectedHeap::heap();
860 if (PrintGCDetails) {
861 if (Verbose) {
862 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
863 level(), short_name(), s, used(), capacity());
864 } else {
865 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
866 level(), short_name(), s, used() / K, capacity() / K);
867 }
868 }
869 if (Verbose) {
870 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
871 gch->used(), gch->capacity());
872 } else {
873 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
874 gch->used() / K, gch->capacity() / K);
875 }
876 }
877
878 size_t
879 ConcurrentMarkSweepGeneration::contiguous_available() const {
880 // dld proposes an improvement in precision here. If the committed
881 // part of the space ends in a free block we should add that to
882 // uncommitted size in the calculation below. Will make this
883 // change later, staying with the approximation below for the
884 // time being. -- ysr.
885 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
886 }
887
888 size_t
889 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
890 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
891 }
892
893 size_t ConcurrentMarkSweepGeneration::max_available() const {
894 return free() + _virtual_space.uncommitted_size();
895 }
896
897 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
898 size_t available = max_available();
899 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
900 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
901 if (Verbose && PrintGCDetails) {
902 gclog_or_tty->print_cr(
903 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
904 "max_promo("SIZE_FORMAT")",
905 res? "":" not", available, res? ">=":"<",
906 av_promo, max_promotion_in_bytes);
907 }
908 return res;
909 }
910
911 // At a promotion failure dump information on block layout in heap
912 // (cms old generation).
913 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
914 if (CMSDumpAtPromotionFailure) {
915 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
916 }
917 }
918
919 CompactibleSpace*
920 ConcurrentMarkSweepGeneration::first_compaction_space() const {
921 return _cmsSpace;
922 }
923
924 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
925 // Clear the promotion information. These pointers can be adjusted
926 // along with all the other pointers into the heap but
927 // compaction is expected to be a rare event with
928 // a heap using cms so don't do it without seeing the need.
929 if (CollectedHeap::use_parallel_gc_threads()) {
930 for (uint i = 0; i < ParallelGCThreads; i++) {
931 _par_gc_thread_states[i]->promo.reset();
932 }
933 }
934 }
935
936 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
937 blk->do_space(_cmsSpace);
938 }
939
940 void ConcurrentMarkSweepGeneration::compute_new_size() {
941 assert_locked_or_safepoint(Heap_lock);
942
943 // If incremental collection failed, we just want to expand
944 // to the limit.
945 if (incremental_collection_failed()) {
946 clear_incremental_collection_failed();
947 grow_to_reserved();
948 return;
949 }
950
951 size_t expand_bytes = 0;
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 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
961 }
962 if (expand_bytes > 0) {
963 if (PrintGCDetails && Verbose) {
964 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
965 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
966 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
967 gclog_or_tty->print_cr(" Desired free fraction %f",
968 desired_free_percentage);
969 gclog_or_tty->print_cr(" Maximum free fraction %f",
970 maximum_free_percentage);
971 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000);
972 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
973 desired_capacity/1000);
974 int prev_level = level() - 1;
975 if (prev_level >= 0) {
976 size_t prev_size = 0;
977 GenCollectedHeap* gch = GenCollectedHeap::heap();
978 Generation* prev_gen = gch->_gens[prev_level];
979 prev_size = prev_gen->capacity();
980 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
981 prev_size/1000);
982 }
983 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
984 unsafe_max_alloc_nogc()/1000);
985 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
986 contiguous_available()/1000);
987 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
988 expand_bytes);
989 }
990 // safe if expansion fails
991 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
992 if (PrintGCDetails && Verbose) {
993 gclog_or_tty->print_cr(" Expanded free fraction %f",
994 ((double) free()) / capacity());
995 }
996 }
997 }
998
999 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
1000 return cmsSpace()->freelistLock();
1001 }
1002
1003 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
1004 bool tlab) {
1005 CMSSynchronousYieldRequest yr;
1006 MutexLockerEx x(freelistLock(),
1007 Mutex::_no_safepoint_check_flag);
1008 return have_lock_and_allocate(size, tlab);
1009 }
1010
1011 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
1012 bool tlab /* ignored */) {
1013 assert_lock_strong(freelistLock());
1014 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
1015 HeapWord* res = cmsSpace()->allocate(adjustedSize);
1016 // Allocate the object live (grey) if the background collector has
1017 // started marking. This is necessary because the marker may
1018 // have passed this address and consequently this object will
1019 // not otherwise be greyed and would be incorrectly swept up.
1020 // Note that if this object contains references, the writing
1021 // of those references will dirty the card containing this object
1022 // allowing the object to be blackened (and its references scanned)
1023 // either during a preclean phase or at the final checkpoint.
1024 if (res != NULL) {
1025 // We may block here with an uninitialized object with
1026 // its mark-bit or P-bits not yet set. Such objects need
1027 // to be safely navigable by block_start().
1028 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
1029 assert(!((FreeChunk*)res)->isFree(), "Error, block will look free but show wrong size");
1030 collector()->direct_allocated(res, adjustedSize);
1031 _direct_allocated_words += adjustedSize;
1032 // allocation counters
1033 NOT_PRODUCT(
1034 _numObjectsAllocated++;
1035 _numWordsAllocated += (int)adjustedSize;
1036 )
1037 }
1038 return res;
1039 }
1040
1041 // In the case of direct allocation by mutators in a generation that
1042 // is being concurrently collected, the object must be allocated
1043 // live (grey) if the background collector has started marking.
1044 // This is necessary because the marker may
1045 // have passed this address and consequently this object will
1046 // not otherwise be greyed and would be incorrectly swept up.
1047 // Note that if this object contains references, the writing
1048 // of those references will dirty the card containing this object
1049 // allowing the object to be blackened (and its references scanned)
1050 // either during a preclean phase or at the final checkpoint.
1051 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
1052 assert(_markBitMap.covers(start, size), "Out of bounds");
1053 if (_collectorState >= Marking) {
1054 MutexLockerEx y(_markBitMap.lock(),
1055 Mutex::_no_safepoint_check_flag);
1056 // [see comments preceding SweepClosure::do_blk() below for details]
1057 // 1. need to mark the object as live so it isn't collected
1058 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
1059 // 3. need to mark the end of the object so marking, precleaning or sweeping
1060 // can skip over uninitialized or unparsable objects. An allocated
1061 // object is considered uninitialized for our purposes as long as
1062 // its klass word is NULL. (Unparsable objects are those which are
1063 // initialized in the sense just described, but whose sizes can still
1064 // not be correctly determined. Note that the class of unparsable objects
1065 // can only occur in the perm gen. All old gen objects are parsable
1066 // as soon as they are initialized.)
1067 _markBitMap.mark(start); // object is live
1068 _markBitMap.mark(start + 1); // object is potentially uninitialized?
1069 _markBitMap.mark(start + size - 1);
1070 // mark end of object
1071 }
1072 // check that oop looks uninitialized
1073 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
1074 }
1075
1076 void CMSCollector::promoted(bool par, HeapWord* start,
1077 bool is_obj_array, size_t obj_size) {
1078 assert(_markBitMap.covers(start), "Out of bounds");
1079 // See comment in direct_allocated() about when objects should
1080 // be allocated live.
1081 if (_collectorState >= Marking) {
1082 // we already hold the marking bit map lock, taken in
1083 // the prologue
1084 if (par) {
1085 _markBitMap.par_mark(start);
1086 } else {
1087 _markBitMap.mark(start);
1088 }
1089 // We don't need to mark the object as uninitialized (as
1090 // in direct_allocated above) because this is being done with the
1091 // world stopped and the object will be initialized by the
1092 // time the marking, precleaning or sweeping get to look at it.
1093 // But see the code for copying objects into the CMS generation,
1094 // where we need to ensure that concurrent readers of the
1095 // block offset table are able to safely navigate a block that
1096 // is in flux from being free to being allocated (and in
1097 // transition while being copied into) and subsequently
1098 // becoming a bona-fide object when the copy/promotion is complete.
1099 assert(SafepointSynchronize::is_at_safepoint(),
1100 "expect promotion only at safepoints");
1101
1102 if (_collectorState < Sweeping) {
1103 // Mark the appropriate cards in the modUnionTable, so that
1104 // this object gets scanned before the sweep. If this is
1105 // not done, CMS generation references in the object might
1106 // not get marked.
1107 // For the case of arrays, which are otherwise precisely
1108 // marked, we need to dirty the entire array, not just its head.
1109 if (is_obj_array) {
1110 // The [par_]mark_range() method expects mr.end() below to
1111 // be aligned to the granularity of a bit's representation
1112 // in the heap. In the case of the MUT below, that's a
1113 // card size.
1114 MemRegion mr(start,
1115 (HeapWord*)round_to((intptr_t)(start + obj_size),
1116 CardTableModRefBS::card_size /* bytes */));
1117 if (par) {
1118 _modUnionTable.par_mark_range(mr);
1119 } else {
1120 _modUnionTable.mark_range(mr);
1121 }
1122 } else { // not an obj array; we can just mark the head
1123 if (par) {
1124 _modUnionTable.par_mark(start);
1125 } else {
1126 _modUnionTable.mark(start);
1127 }
1128 }
1129 }
1130 }
1131 }
1132
1133 static inline size_t percent_of_space(Space* space, HeapWord* addr)
1134 {
1135 size_t delta = pointer_delta(addr, space->bottom());
1136 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
1137 }
1138
1139 void CMSCollector::icms_update_allocation_limits()
1140 {
1141 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
1142 EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
1143
1144 const unsigned int duty_cycle = stats().icms_update_duty_cycle();
1145 if (CMSTraceIncrementalPacing) {
1146 stats().print();
1147 }
1148
1149 assert(duty_cycle <= 100, "invalid duty cycle");
1150 if (duty_cycle != 0) {
1151 // The duty_cycle is a percentage between 0 and 100; convert to words and
1152 // then compute the offset from the endpoints of the space.
1153 size_t free_words = eden->free() / HeapWordSize;
1154 double free_words_dbl = (double)free_words;
1155 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
1156 size_t offset_words = (free_words - duty_cycle_words) / 2;
1157
1158 _icms_start_limit = eden->top() + offset_words;
1159 _icms_stop_limit = eden->end() - offset_words;
1160
1161 // The limits may be adjusted (shifted to the right) by
1162 // CMSIncrementalOffset, to allow the application more mutator time after a
1163 // young gen gc (when all mutators were stopped) and before CMS starts and
1164 // takes away one or more cpus.
1165 if (CMSIncrementalOffset != 0) {
1166 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
1167 size_t adjustment = (size_t)adjustment_dbl;
1168 HeapWord* tmp_stop = _icms_stop_limit + adjustment;
1169 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
1170 _icms_start_limit += adjustment;
1171 _icms_stop_limit = tmp_stop;
1172 }
1173 }
1174 }
1175 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
1176 _icms_start_limit = _icms_stop_limit = eden->end();
1177 }
1178
1179 // Install the new start limit.
1180 eden->set_soft_end(_icms_start_limit);
1181
1182 if (CMSTraceIncrementalMode) {
1183 gclog_or_tty->print(" icms alloc limits: "
1184 PTR_FORMAT "," PTR_FORMAT
1185 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
1186 _icms_start_limit, _icms_stop_limit,
1187 percent_of_space(eden, _icms_start_limit),
1188 percent_of_space(eden, _icms_stop_limit));
1189 if (Verbose) {
1190 gclog_or_tty->print("eden: ");
1191 eden->print_on(gclog_or_tty);
1192 }
1193 }
1194 }
1195
1196 // Any changes here should try to maintain the invariant
1197 // that if this method is called with _icms_start_limit
1198 // and _icms_stop_limit both NULL, then it should return NULL
1199 // and not notify the icms thread.
1200 HeapWord*
1201 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
1202 size_t word_size)
1203 {
1204 // A start_limit equal to end() means the duty cycle is 0, so treat that as a
1205 // nop.
1206 if (CMSIncrementalMode && _icms_start_limit != space->end()) {
1207 if (top <= _icms_start_limit) {
1208 if (CMSTraceIncrementalMode) {
1209 space->print_on(gclog_or_tty);
1210 gclog_or_tty->stamp();
1211 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
1212 ", new limit=" PTR_FORMAT
1213 " (" SIZE_FORMAT "%%)",
1214 top, _icms_stop_limit,
1215 percent_of_space(space, _icms_stop_limit));
1216 }
1217 ConcurrentMarkSweepThread::start_icms();
1218 assert(top < _icms_stop_limit, "Tautology");
1219 if (word_size < pointer_delta(_icms_stop_limit, top)) {
1220 return _icms_stop_limit;
1221 }
1222
1223 // The allocation will cross both the _start and _stop limits, so do the
1224 // stop notification also and return end().
1225 if (CMSTraceIncrementalMode) {
1226 space->print_on(gclog_or_tty);
1227 gclog_or_tty->stamp();
1228 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
1229 ", new limit=" PTR_FORMAT
1230 " (" SIZE_FORMAT "%%)",
1231 top, space->end(),
1232 percent_of_space(space, space->end()));
1233 }
1234 ConcurrentMarkSweepThread::stop_icms();
1235 return space->end();
1236 }
1237
1238 if (top <= _icms_stop_limit) {
1239 if (CMSTraceIncrementalMode) {
1240 space->print_on(gclog_or_tty);
1241 gclog_or_tty->stamp();
1242 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
1243 ", new limit=" PTR_FORMAT
1244 " (" SIZE_FORMAT "%%)",
1245 top, space->end(),
1246 percent_of_space(space, space->end()));
1247 }
1248 ConcurrentMarkSweepThread::stop_icms();
1249 return space->end();
1250 }
1251
1252 if (CMSTraceIncrementalMode) {
1253 space->print_on(gclog_or_tty);
1254 gclog_or_tty->stamp();
1255 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
1256 ", new limit=" PTR_FORMAT,
1257 top, NULL);
1258 }
1259 }
1260
1261 return NULL;
1262 }
1263
1264 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
1265 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1266 // allocate, copy and if necessary update promoinfo --
1267 // delegate to underlying space.
1268 assert_lock_strong(freelistLock());
1269
1270 #ifndef PRODUCT
1271 if (Universe::heap()->promotion_should_fail()) {
1272 return NULL;
1273 }
1274 #endif // #ifndef PRODUCT
1275
1276 oop res = _cmsSpace->promote(obj, obj_size);
1277 if (res == NULL) {
1278 // expand and retry
1279 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
1280 expand(s*HeapWordSize, MinHeapDeltaBytes,
1281 CMSExpansionCause::_satisfy_promotion);
1282 // Since there's currently no next generation, we don't try to promote
1283 // into a more senior generation.
1284 assert(next_gen() == NULL, "assumption, based upon which no attempt "
1285 "is made to pass on a possibly failing "
1286 "promotion to next generation");
1287 res = _cmsSpace->promote(obj, obj_size);
1288 }
1289 if (res != NULL) {
1290 // See comment in allocate() about when objects should
1291 // be allocated live.
1292 assert(obj->is_oop(), "Will dereference klass pointer below");
1293 collector()->promoted(false, // Not parallel
1294 (HeapWord*)res, obj->is_objArray(), obj_size);
1295 // promotion counters
1296 NOT_PRODUCT(
1297 _numObjectsPromoted++;
1298 _numWordsPromoted +=
1299 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1300 )
1301 }
1302 return res;
1303 }
1304
1305
1306 HeapWord*
1307 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
1308 HeapWord* top,
1309 size_t word_sz)
1310 {
1311 return collector()->allocation_limit_reached(space, top, word_sz);
1312 }
1313
1314 // IMPORTANT: Notes on object size recognition in CMS.
1315 // ---------------------------------------------------
1316 // A block of storage in the CMS generation is always in
1317 // one of three states. A free block (FREE), an allocated
1318 // object (OBJECT) whose size() method reports the correct size,
1319 // and an intermediate state (TRANSIENT) in which its size cannot
1320 // be accurately determined.
1321 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1322 // -----------------------------------------------------
1323 // FREE: klass_word & 1 == 1; mark_word holds block size
1324 //
1325 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1326 // obj->size() computes correct size
1327 // [Perm Gen objects needs to be "parsable" before they can be navigated]
1328 //
1329 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1330 //
1331 // STATE IDENTIFICATION: (64 bit+COOPS)
1332 // ------------------------------------
1333 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1334 //
1335 // OBJECT: klass_word installed; klass_word != 0;
1336 // obj->size() computes correct size
1337 // [Perm Gen comment above continues to hold]
1338 //
1339 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1340 //
1341 //
1342 // STATE TRANSITION DIAGRAM
1343 //
1344 // mut / parnew mut / parnew
1345 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1346 // ^ |
1347 // |------------------------ DEAD <------------------------------------|
1348 // sweep mut
1349 //
1350 // While a block is in TRANSIENT state its size cannot be determined
1351 // so readers will either need to come back later or stall until
1352 // the size can be determined. Note that for the case of direct
1353 // allocation, P-bits, when available, may be used to determine the
1354 // size of an object that may not yet have been initialized.
1355
1356 // Things to support parallel young-gen collection.
1357 oop
1358 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1359 oop old, markOop m,
1360 size_t word_sz) {
1361 #ifndef PRODUCT
1362 if (Universe::heap()->promotion_should_fail()) {
1363 return NULL;
1364 }
1365 #endif // #ifndef PRODUCT
1366
1367 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1368 PromotionInfo* promoInfo = &ps->promo;
1369 // if we are tracking promotions, then first ensure space for
1370 // promotion (including spooling space for saving header if necessary).
1371 // then allocate and copy, then track promoted info if needed.
1372 // When tracking (see PromotionInfo::track()), the mark word may
1373 // be displaced and in this case restoration of the mark word
1374 // occurs in the (oop_since_save_marks_)iterate phase.
1375 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1376 // Out of space for allocating spooling buffers;
1377 // try expanding and allocating spooling buffers.
1378 if (!expand_and_ensure_spooling_space(promoInfo)) {
1379 return NULL;
1380 }
1381 }
1382 assert(promoInfo->has_spooling_space(), "Control point invariant");
1383 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1384 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1385 if (obj_ptr == NULL) {
1386 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1387 if (obj_ptr == NULL) {
1388 return NULL;
1389 }
1390 }
1391 oop obj = oop(obj_ptr);
1392 OrderAccess::storestore();
1393 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1394 assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size");
1395 // IMPORTANT: See note on object initialization for CMS above.
1396 // Otherwise, copy the object. Here we must be careful to insert the
1397 // klass pointer last, since this marks the block as an allocated object.
1398 // Except with compressed oops it's the mark word.
1399 HeapWord* old_ptr = (HeapWord*)old;
1400 // Restore the mark word copied above.
1401 obj->set_mark(m);
1402 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1403 assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size");
1404 OrderAccess::storestore();
1405
1406 if (UseCompressedOops) {
1407 // Copy gap missed by (aligned) header size calculation below
1408 obj->set_klass_gap(old->klass_gap());
1409 }
1410 if (word_sz > (size_t)oopDesc::header_size()) {
1411 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1412 obj_ptr + oopDesc::header_size(),
1413 word_sz - oopDesc::header_size());
1414 }
1415
1416 // Now we can track the promoted object, if necessary. We take care
1417 // to delay the transition from uninitialized to full object
1418 // (i.e., insertion of klass pointer) until after, so that it
1419 // atomically becomes a promoted object.
1420 if (promoInfo->tracking()) {
1421 promoInfo->track((PromotedObject*)obj, old->klass());
1422 }
1423 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1424 assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size");
1425 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1426
1427 // Finally, install the klass pointer (this should be volatile).
1428 OrderAccess::storestore();
1429 obj->set_klass(old->klass());
1430 // We should now be able to calculate the right size for this object
1431 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1432
1433 collector()->promoted(true, // parallel
1434 obj_ptr, old->is_objArray(), word_sz);
1435
1436 NOT_PRODUCT(
1437 Atomic::inc_ptr(&_numObjectsPromoted);
1438 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1439 )
1440
1441 return obj;
1442 }
1443
1444 void
1445 ConcurrentMarkSweepGeneration::
1446 par_promote_alloc_undo(int thread_num,
1447 HeapWord* obj, size_t word_sz) {
1448 // CMS does not support promotion undo.
1449 ShouldNotReachHere();
1450 }
1451
1452 void
1453 ConcurrentMarkSweepGeneration::
1454 par_promote_alloc_done(int thread_num) {
1455 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1456 ps->lab.retire(thread_num);
1457 }
1458
1459 void
1460 ConcurrentMarkSweepGeneration::
1461 par_oop_since_save_marks_iterate_done(int thread_num) {
1462 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1463 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1464 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1465 }
1466
1467 // XXXPERM
1468 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1469 size_t size,
1470 bool tlab)
1471 {
1472 // We allow a STW collection only if a full
1473 // collection was requested.
1474 return full || should_allocate(size, tlab); // FIX ME !!!
1475 // This and promotion failure handling are connected at the
1476 // hip and should be fixed by untying them.
1477 }
1478
1479 bool CMSCollector::shouldConcurrentCollect() {
1480 if (_full_gc_requested) {
1481 if (Verbose && PrintGCDetails) {
1482 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1483 " gc request (or gc_locker)");
1484 }
1485 return true;
1486 }
1487
1488 // For debugging purposes, change the type of collection.
1489 // If the rotation is not on the concurrent collection
1490 // type, don't start a concurrent collection.
1491 NOT_PRODUCT(
1492 if (RotateCMSCollectionTypes &&
1493 (_cmsGen->debug_collection_type() !=
1494 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
1495 assert(_cmsGen->debug_collection_type() !=
1496 ConcurrentMarkSweepGeneration::Unknown_collection_type,
1497 "Bad cms collection type");
1498 return false;
1499 }
1500 )
1501
1502 FreelistLocker x(this);
1503 // ------------------------------------------------------------------
1504 // Print out lots of information which affects the initiation of
1505 // a collection.
1506 if (PrintCMSInitiationStatistics && stats().valid()) {
1507 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1508 gclog_or_tty->stamp();
1509 gclog_or_tty->print_cr("");
1510 stats().print_on(gclog_or_tty);
1511 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1512 stats().time_until_cms_gen_full());
1513 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1514 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1515 _cmsGen->contiguous_available());
1516 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1517 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1518 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1519 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1520 gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy());
1521 }
1522 // ------------------------------------------------------------------
1523
1524 // If the estimated time to complete a cms collection (cms_duration())
1525 // is less than the estimated time remaining until the cms generation
1526 // is full, start a collection.
1527 if (!UseCMSInitiatingOccupancyOnly) {
1528 if (stats().valid()) {
1529 if (stats().time_until_cms_start() == 0.0) {
1530 return true;
1531 }
1532 } else {
1533 // We want to conservatively collect somewhat early in order
1534 // to try and "bootstrap" our CMS/promotion statistics;
1535 // this branch will not fire after the first successful CMS
1536 // collection because the stats should then be valid.
1537 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1538 if (Verbose && PrintGCDetails) {
1539 gclog_or_tty->print_cr(
1540 " CMSCollector: collect for bootstrapping statistics:"
1541 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1542 _bootstrap_occupancy);
1543 }
1544 return true;
1545 }
1546 }
1547 }
1548
1549 // Otherwise, we start a collection cycle if either the perm gen or
1550 // old gen want a collection cycle started. Each may use
1551 // an appropriate criterion for making this decision.
1552 // XXX We need to make sure that the gen expansion
1553 // criterion dovetails well with this. XXX NEED TO FIX THIS
1554 if (_cmsGen->should_concurrent_collect()) {
1555 if (Verbose && PrintGCDetails) {
1556 gclog_or_tty->print_cr("CMS old gen initiated");
1557 }
1558 return true;
1559 }
1560
1561 // We start a collection if we believe an incremental collection may fail;
1562 // this is not likely to be productive in practice because it's probably too
1563 // late anyway.
1564 GenCollectedHeap* gch = GenCollectedHeap::heap();
1565 assert(gch->collector_policy()->is_two_generation_policy(),
1566 "You may want to check the correctness of the following");
1567 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1568 if (Verbose && PrintGCDetails) {
1569 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1570 }
1571 return true;
1572 }
1573
1574 if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) {
1575 bool res = update_should_unload_classes();
1576 if (res) {
1577 if (Verbose && PrintGCDetails) {
1578 gclog_or_tty->print_cr("CMS perm gen initiated");
1579 }
1580 return true;
1581 }
1582 }
1583 return false;
1584 }
1585
1586 // Clear _expansion_cause fields of constituent generations
1587 void CMSCollector::clear_expansion_cause() {
1588 _cmsGen->clear_expansion_cause();
1589 _permGen->clear_expansion_cause();
1590 }
1591
1592 // We should be conservative in starting a collection cycle. To
1593 // start too eagerly runs the risk of collecting too often in the
1594 // extreme. To collect too rarely falls back on full collections,
1595 // which works, even if not optimum in terms of concurrent work.
1596 // As a work around for too eagerly collecting, use the flag
1597 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1598 // giving the user an easily understandable way of controlling the
1599 // collections.
1600 // We want to start a new collection cycle if any of the following
1601 // conditions hold:
1602 // . our current occupancy exceeds the configured initiating occupancy
1603 // for this generation, or
1604 // . we recently needed to expand this space and have not, since that
1605 // expansion, done a collection of this generation, or
1606 // . the underlying space believes that it may be a good idea to initiate
1607 // a concurrent collection (this may be based on criteria such as the
1608 // following: the space uses linear allocation and linear allocation is
1609 // going to fail, or there is believed to be excessive fragmentation in
1610 // the generation, etc... or ...
1611 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1612 // the case of the old generation, not the perm generation; see CR 6543076):
1613 // we may be approaching a point at which allocation requests may fail because
1614 // we will be out of sufficient free space given allocation rate estimates.]
1615 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1616
1617 assert_lock_strong(freelistLock());
1618 if (occupancy() > initiating_occupancy()) {
1619 if (PrintGCDetails && Verbose) {
1620 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1621 short_name(), occupancy(), initiating_occupancy());
1622 }
1623 return true;
1624 }
1625 if (UseCMSInitiatingOccupancyOnly) {
1626 return false;
1627 }
1628 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1629 if (PrintGCDetails && Verbose) {
1630 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1631 short_name());
1632 }
1633 return true;
1634 }
1635 if (_cmsSpace->should_concurrent_collect()) {
1636 if (PrintGCDetails && Verbose) {
1637 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1638 short_name());
1639 }
1640 return true;
1641 }
1642 return false;
1643 }
1644
1645 void ConcurrentMarkSweepGeneration::collect(bool full,
1646 bool clear_all_soft_refs,
1647 size_t size,
1648 bool tlab)
1649 {
1650 collector()->collect(full, clear_all_soft_refs, size, tlab);
1651 }
1652
1653 void CMSCollector::collect(bool full,
1654 bool clear_all_soft_refs,
1655 size_t size,
1656 bool tlab)
1657 {
1658 if (!UseCMSCollectionPassing && _collectorState > Idling) {
1659 // For debugging purposes skip the collection if the state
1660 // is not currently idle
1661 if (TraceCMSState) {
1662 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
1663 Thread::current(), full, _collectorState);
1664 }
1665 return;
1666 }
1667
1668 // The following "if" branch is present for defensive reasons.
1669 // In the current uses of this interface, it can be replaced with:
1670 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1671 // But I am not placing that assert here to allow future
1672 // generality in invoking this interface.
1673 if (GC_locker::is_active()) {
1674 // A consistency test for GC_locker
1675 assert(GC_locker::needs_gc(), "Should have been set already");
1676 // Skip this foreground collection, instead
1677 // expanding the heap if necessary.
1678 // Need the free list locks for the call to free() in compute_new_size()
1679 compute_new_size();
1680 return;
1681 }
1682 acquire_control_and_collect(full, clear_all_soft_refs);
1683 _full_gcs_since_conc_gc++;
1684
1685 }
1686
1687 void CMSCollector::request_full_gc(unsigned int full_gc_count) {
1688 GenCollectedHeap* gch = GenCollectedHeap::heap();
1689 unsigned int gc_count = gch->total_full_collections();
1690 if (gc_count == full_gc_count) {
1691 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1692 _full_gc_requested = true;
1693 CGC_lock->notify(); // nudge CMS thread
1694 } else {
1695 assert(gc_count > full_gc_count, "Error: causal loop");
1696 }
1697 }
1698
1699
1700 // The foreground and background collectors need to coordinate in order
1701 // to make sure that they do not mutually interfere with CMS collections.
1702 // When a background collection is active,
1703 // the foreground collector may need to take over (preempt) and
1704 // synchronously complete an ongoing collection. Depending on the
1705 // frequency of the background collections and the heap usage
1706 // of the application, this preemption can be seldom or frequent.
1707 // There are only certain
1708 // points in the background collection that the "collection-baton"
1709 // can be passed to the foreground collector.
1710 //
1711 // The foreground collector will wait for the baton before
1712 // starting any part of the collection. The foreground collector
1713 // will only wait at one location.
1714 //
1715 // The background collector will yield the baton before starting a new
1716 // phase of the collection (e.g., before initial marking, marking from roots,
1717 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1718 // of the loop which switches the phases. The background collector does some
1719 // of the phases (initial mark, final re-mark) with the world stopped.
1720 // Because of locking involved in stopping the world,
1721 // the foreground collector should not block waiting for the background
1722 // collector when it is doing a stop-the-world phase. The background
1723 // collector will yield the baton at an additional point just before
1724 // it enters a stop-the-world phase. Once the world is stopped, the
1725 // background collector checks the phase of the collection. If the
1726 // phase has not changed, it proceeds with the collection. If the
1727 // phase has changed, it skips that phase of the collection. See
1728 // the comments on the use of the Heap_lock in collect_in_background().
1729 //
1730 // Variable used in baton passing.
1731 // _foregroundGCIsActive - Set to true by the foreground collector when
1732 // it wants the baton. The foreground clears it when it has finished
1733 // the collection.
1734 // _foregroundGCShouldWait - Set to true by the background collector
1735 // when it is running. The foreground collector waits while
1736 // _foregroundGCShouldWait is true.
1737 // CGC_lock - monitor used to protect access to the above variables
1738 // and to notify the foreground and background collectors.
1739 // _collectorState - current state of the CMS collection.
1740 //
1741 // The foreground collector
1742 // acquires the CGC_lock
1743 // sets _foregroundGCIsActive
1744 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1745 // various locks acquired in preparation for the collection
1746 // are released so as not to block the background collector
1747 // that is in the midst of a collection
1748 // proceeds with the collection
1749 // clears _foregroundGCIsActive
1750 // returns
1751 //
1752 // The background collector in a loop iterating on the phases of the
1753 // collection
1754 // acquires the CGC_lock
1755 // sets _foregroundGCShouldWait
1756 // if _foregroundGCIsActive is set
1757 // clears _foregroundGCShouldWait, notifies _CGC_lock
1758 // waits on _CGC_lock for _foregroundGCIsActive to become false
1759 // and exits the loop.
1760 // otherwise
1761 // proceed with that phase of the collection
1762 // if the phase is a stop-the-world phase,
1763 // yield the baton once more just before enqueueing
1764 // the stop-world CMS operation (executed by the VM thread).
1765 // returns after all phases of the collection are done
1766 //
1767
1768 void CMSCollector::acquire_control_and_collect(bool full,
1769 bool clear_all_soft_refs) {
1770 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1771 assert(!Thread::current()->is_ConcurrentGC_thread(),
1772 "shouldn't try to acquire control from self!");
1773
1774 // Start the protocol for acquiring control of the
1775 // collection from the background collector (aka CMS thread).
1776 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1777 "VM thread should have CMS token");
1778 // Remember the possibly interrupted state of an ongoing
1779 // concurrent collection
1780 CollectorState first_state = _collectorState;
1781
1782 // Signal to a possibly ongoing concurrent collection that
1783 // we want to do a foreground collection.
1784 _foregroundGCIsActive = true;
1785
1786 // Disable incremental mode during a foreground collection.
1787 ICMSDisabler icms_disabler;
1788
1789 // release locks and wait for a notify from the background collector
1790 // releasing the locks in only necessary for phases which
1791 // do yields to improve the granularity of the collection.
1792 assert_lock_strong(bitMapLock());
1793 // We need to lock the Free list lock for the space that we are
1794 // currently collecting.
1795 assert(haveFreelistLocks(), "Must be holding free list locks");
1796 bitMapLock()->unlock();
1797 releaseFreelistLocks();
1798 {
1799 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1800 if (_foregroundGCShouldWait) {
1801 // We are going to be waiting for action for the CMS thread;
1802 // it had better not be gone (for instance at shutdown)!
1803 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1804 "CMS thread must be running");
1805 // Wait here until the background collector gives us the go-ahead
1806 ConcurrentMarkSweepThread::clear_CMS_flag(
1807 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1808 // Get a possibly blocked CMS thread going:
1809 // Note that we set _foregroundGCIsActive true above,
1810 // without protection of the CGC_lock.
1811 CGC_lock->notify();
1812 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1813 "Possible deadlock");
1814 while (_foregroundGCShouldWait) {
1815 // wait for notification
1816 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1817 // Possibility of delay/starvation here, since CMS token does
1818 // not know to give priority to VM thread? Actually, i think
1819 // there wouldn't be any delay/starvation, but the proof of
1820 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1821 }
1822 ConcurrentMarkSweepThread::set_CMS_flag(
1823 ConcurrentMarkSweepThread::CMS_vm_has_token);
1824 }
1825 }
1826 // The CMS_token is already held. Get back the other locks.
1827 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1828 "VM thread should have CMS token");
1829 getFreelistLocks();
1830 bitMapLock()->lock_without_safepoint_check();
1831 if (TraceCMSState) {
1832 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1833 INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1834 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1835 }
1836
1837 // Check if we need to do a compaction, or if not, whether
1838 // we need to start the mark-sweep from scratch.
1839 bool should_compact = false;
1840 bool should_start_over = false;
1841 decide_foreground_collection_type(clear_all_soft_refs,
1842 &should_compact, &should_start_over);
1843
1844 NOT_PRODUCT(
1845 if (RotateCMSCollectionTypes) {
1846 if (_cmsGen->debug_collection_type() ==
1847 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1848 should_compact = true;
1849 } else if (_cmsGen->debug_collection_type() ==
1850 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1851 should_compact = false;
1852 }
1853 }
1854 )
1855
1856 if (PrintGCDetails && first_state > Idling) {
1857 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1858 if (GCCause::is_user_requested_gc(cause) ||
1859 GCCause::is_serviceability_requested_gc(cause)) {
1860 gclog_or_tty->print(" (concurrent mode interrupted)");
1861 } else {
1862 gclog_or_tty->print(" (concurrent mode failure)");
1863 }
1864 }
1865
1866 if (should_compact) {
1867 // If the collection is being acquired from the background
1868 // collector, there may be references on the discovered
1869 // references lists that have NULL referents (being those
1870 // that were concurrently cleared by a mutator) or
1871 // that are no longer active (having been enqueued concurrently
1872 // by the mutator).
1873 // Scrub the list of those references because Mark-Sweep-Compact
1874 // code assumes referents are not NULL and that all discovered
1875 // Reference objects are active.
1876 ref_processor()->clean_up_discovered_references();
1877
1878 do_compaction_work(clear_all_soft_refs);
1879
1880 // Has the GC time limit been exceeded?
1881 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration();
1882 size_t max_eden_size = young_gen->max_capacity() -
1883 young_gen->to()->capacity() -
1884 young_gen->from()->capacity();
1885 GenCollectedHeap* gch = GenCollectedHeap::heap();
1886 GCCause::Cause gc_cause = gch->gc_cause();
1887 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1888 young_gen->eden()->used(),
1889 _cmsGen->max_capacity(),
1890 max_eden_size,
1891 full,
1892 gc_cause,
1893 gch->collector_policy());
1894 } else {
1895 do_mark_sweep_work(clear_all_soft_refs, first_state,
1896 should_start_over);
1897 }
1898 // Reset the expansion cause, now that we just completed
1899 // a collection cycle.
1900 clear_expansion_cause();
1901 _foregroundGCIsActive = false;
1902 return;
1903 }
1904
1905 // Resize the perm generation and the tenured generation
1906 // after obtaining the free list locks for the
1907 // two generations.
1908 void CMSCollector::compute_new_size() {
1909 assert_locked_or_safepoint(Heap_lock);
1910 FreelistLocker z(this);
1911 _permGen->compute_new_size();
1912 _cmsGen->compute_new_size();
1913 }
1914
1915 // A work method used by foreground collection to determine
1916 // what type of collection (compacting or not, continuing or fresh)
1917 // it should do.
1918 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1919 // and CMSCompactWhenClearAllSoftRefs the default in the future
1920 // and do away with the flags after a suitable period.
1921 void CMSCollector::decide_foreground_collection_type(
1922 bool clear_all_soft_refs, bool* should_compact,
1923 bool* should_start_over) {
1924 // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1925 // flag is set, and we have either requested a System.gc() or
1926 // the number of full gc's since the last concurrent cycle
1927 // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1928 // or if an incremental collection has failed
1929 GenCollectedHeap* gch = GenCollectedHeap::heap();
1930 assert(gch->collector_policy()->is_two_generation_policy(),
1931 "You may want to check the correctness of the following");
1932 // Inform cms gen if this was due to partial collection failing.
1933 // The CMS gen may use this fact to determine its expansion policy.
1934 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1935 assert(!_cmsGen->incremental_collection_failed(),
1936 "Should have been noticed, reacted to and cleared");
1937 _cmsGen->set_incremental_collection_failed();
1938 }
1939 *should_compact =
1940 UseCMSCompactAtFullCollection &&
1941 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1942 GCCause::is_user_requested_gc(gch->gc_cause()) ||
1943 gch->incremental_collection_will_fail(true /* consult_young */));
1944 *should_start_over = false;
1945 if (clear_all_soft_refs && !*should_compact) {
1946 // We are about to do a last ditch collection attempt
1947 // so it would normally make sense to do a compaction
1948 // to reclaim as much space as possible.
1949 if (CMSCompactWhenClearAllSoftRefs) {
1950 // Default: The rationale is that in this case either
1951 // we are past the final marking phase, in which case
1952 // we'd have to start over, or so little has been done
1953 // that there's little point in saving that work. Compaction
1954 // appears to be the sensible choice in either case.
1955 *should_compact = true;
1956 } else {
1957 // We have been asked to clear all soft refs, but not to
1958 // compact. Make sure that we aren't past the final checkpoint
1959 // phase, for that is where we process soft refs. If we are already
1960 // past that phase, we'll need to redo the refs discovery phase and
1961 // if necessary clear soft refs that weren't previously
1962 // cleared. We do so by remembering the phase in which
1963 // we came in, and if we are past the refs processing
1964 // phase, we'll choose to just redo the mark-sweep
1965 // collection from scratch.
1966 if (_collectorState > FinalMarking) {
1967 // We are past the refs processing phase;
1968 // start over and do a fresh synchronous CMS cycle
1969 _collectorState = Resetting; // skip to reset to start new cycle
1970 reset(false /* == !asynch */);
1971 *should_start_over = true;
1972 } // else we can continue a possibly ongoing current cycle
1973 }
1974 }
1975 }
1976
1977 // A work method used by the foreground collector to do
1978 // a mark-sweep-compact.
1979 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1980 GenCollectedHeap* gch = GenCollectedHeap::heap();
1981 TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty);
1982 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
1983 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
1984 "collections passed to foreground collector", _full_gcs_since_conc_gc);
1985 }
1986
1987 // Sample collection interval time and reset for collection pause.
1988 if (UseAdaptiveSizePolicy) {
1989 size_policy()->msc_collection_begin();
1990 }
1991
1992 // Temporarily widen the span of the weak reference processing to
1993 // the entire heap.
1994 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1995 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1996 // Temporarily, clear the "is_alive_non_header" field of the
1997 // reference processor.
1998 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1999 // Temporarily make reference _processing_ single threaded (non-MT).
2000 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
2001 // Temporarily make refs discovery atomic
2002 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
2003 // Temporarily make reference _discovery_ single threaded (non-MT)
2004 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
2005
2006 ref_processor()->set_enqueuing_is_done(false);
2007 ref_processor()->enable_discovery();
2008 ref_processor()->setup_policy(clear_all_soft_refs);
2009 // If an asynchronous collection finishes, the _modUnionTable is
2010 // all clear. If we are assuming the collection from an asynchronous
2011 // collection, clear the _modUnionTable.
2012 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
2013 "_modUnionTable should be clear if the baton was not passed");
2014 _modUnionTable.clear_all();
2015
2016 // We must adjust the allocation statistics being maintained
2017 // in the free list space. We do so by reading and clearing
2018 // the sweep timer and updating the block flux rate estimates below.
2019 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
2020 if (_inter_sweep_timer.is_active()) {
2021 _inter_sweep_timer.stop();
2022 // Note that we do not use this sample to update the _inter_sweep_estimate.
2023 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
2024 _inter_sweep_estimate.padded_average(),
2025 _intra_sweep_estimate.padded_average());
2026 }
2027
2028 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
2029 ref_processor(), clear_all_soft_refs);
2030 #ifdef ASSERT
2031 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
2032 size_t free_size = cms_space->free();
2033 assert(free_size ==
2034 pointer_delta(cms_space->end(), cms_space->compaction_top())
2035 * HeapWordSize,
2036 "All the free space should be compacted into one chunk at top");
2037 assert(cms_space->dictionary()->totalChunkSize(
2038 debug_only(cms_space->freelistLock())) == 0 ||
2039 cms_space->totalSizeInIndexedFreeLists() == 0,
2040 "All the free space should be in a single chunk");
2041 size_t num = cms_space->totalCount();
2042 assert((free_size == 0 && num == 0) ||
2043 (free_size > 0 && (num == 1 || num == 2)),
2044 "There should be at most 2 free chunks after compaction");
2045 #endif // ASSERT
2046 _collectorState = Resetting;
2047 assert(_restart_addr == NULL,
2048 "Should have been NULL'd before baton was passed");
2049 reset(false /* == !asynch */);
2050 _cmsGen->reset_after_compaction();
2051 _concurrent_cycles_since_last_unload = 0;
2052
2053 if (verifying() && !should_unload_classes()) {
2054 perm_gen_verify_bit_map()->clear_all();
2055 }
2056
2057 // Clear any data recorded in the PLAB chunk arrays.
2058 if (_survivor_plab_array != NULL) {
2059 reset_survivor_plab_arrays();
2060 }
2061
2062 // Adjust the per-size allocation stats for the next epoch.
2063 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
2064 // Restart the "inter sweep timer" for the next epoch.
2065 _inter_sweep_timer.reset();
2066 _inter_sweep_timer.start();
2067
2068 // Sample collection pause time and reset for collection interval.
2069 if (UseAdaptiveSizePolicy) {
2070 size_policy()->msc_collection_end(gch->gc_cause());
2071 }
2072
2073 // For a mark-sweep-compact, compute_new_size() will be called
2074 // in the heap's do_collection() method.
2075 }
2076
2077 // A work method used by the foreground collector to do
2078 // a mark-sweep, after taking over from a possibly on-going
2079 // concurrent mark-sweep collection.
2080 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2081 CollectorState first_state, bool should_start_over) {
2082 if (PrintGC && Verbose) {
2083 gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2084 "collector with count %d",
2085 _full_gcs_since_conc_gc);
2086 }
2087 switch (_collectorState) {
2088 case Idling:
2089 if (first_state == Idling || should_start_over) {
2090 // The background GC was not active, or should
2091 // restarted from scratch; start the cycle.
2092 _collectorState = InitialMarking;
2093 }
2094 // If first_state was not Idling, then a background GC
2095 // was in progress and has now finished. No need to do it
2096 // again. Leave the state as Idling.
2097 break;
2098 case Precleaning:
2099 // In the foreground case don't do the precleaning since
2100 // it is not done concurrently and there is extra work
2101 // required.
2102 _collectorState = FinalMarking;
2103 }
2104 if (PrintGCDetails &&
2105 (_collectorState > Idling ||
2106 !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) {
2107 gclog_or_tty->print(" (concurrent mode failure)");
2108 }
2109 collect_in_foreground(clear_all_soft_refs);
2110
2111 // For a mark-sweep, compute_new_size() will be called
2112 // in the heap's do_collection() method.
2113 }
2114
2115
2116 void CMSCollector::getFreelistLocks() const {
2117 // Get locks for all free lists in all generations that this
2118 // collector is responsible for
2119 _cmsGen->freelistLock()->lock_without_safepoint_check();
2120 _permGen->freelistLock()->lock_without_safepoint_check();
2121 }
2122
2123 void CMSCollector::releaseFreelistLocks() const {
2124 // Release locks for all free lists in all generations that this
2125 // collector is responsible for
2126 _cmsGen->freelistLock()->unlock();
2127 _permGen->freelistLock()->unlock();
2128 }
2129
2130 bool CMSCollector::haveFreelistLocks() const {
2131 // Check locks for all free lists in all generations that this
2132 // collector is responsible for
2133 assert_lock_strong(_cmsGen->freelistLock());
2134 assert_lock_strong(_permGen->freelistLock());
2135 PRODUCT_ONLY(ShouldNotReachHere());
2136 return true;
2137 }
2138
2139 // A utility class that is used by the CMS collector to
2140 // temporarily "release" the foreground collector from its
2141 // usual obligation to wait for the background collector to
2142 // complete an ongoing phase before proceeding.
2143 class ReleaseForegroundGC: public StackObj {
2144 private:
2145 CMSCollector* _c;
2146 public:
2147 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2148 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2149 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2150 // allow a potentially blocked foreground collector to proceed
2151 _c->_foregroundGCShouldWait = false;
2152 if (_c->_foregroundGCIsActive) {
2153 CGC_lock->notify();
2154 }
2155 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2156 "Possible deadlock");
2157 }
2158
2159 ~ReleaseForegroundGC() {
2160 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2161 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2162 _c->_foregroundGCShouldWait = true;
2163 }
2164 };
2165
2166 // There are separate collect_in_background and collect_in_foreground because of
2167 // the different locking requirements of the background collector and the
2168 // foreground collector. There was originally an attempt to share
2169 // one "collect" method between the background collector and the foreground
2170 // collector but the if-then-else required made it cleaner to have
2171 // separate methods.
2172 void CMSCollector::collect_in_background(bool clear_all_soft_refs) {
2173 assert(Thread::current()->is_ConcurrentGC_thread(),
2174 "A CMS asynchronous collection is only allowed on a CMS thread.");
2175
2176 GenCollectedHeap* gch = GenCollectedHeap::heap();
2177 {
2178 bool safepoint_check = Mutex::_no_safepoint_check_flag;
2179 MutexLockerEx hl(Heap_lock, safepoint_check);
2180 FreelistLocker fll(this);
2181 MutexLockerEx x(CGC_lock, safepoint_check);
2182 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2183 // The foreground collector is active or we're
2184 // not using asynchronous collections. Skip this
2185 // background collection.
2186 assert(!_foregroundGCShouldWait, "Should be clear");
2187 return;
2188 } else {
2189 assert(_collectorState == Idling, "Should be idling before start.");
2190 _collectorState = InitialMarking;
2191 // Reset the expansion cause, now that we are about to begin
2192 // a new cycle.
2193 clear_expansion_cause();
2194 }
2195 // Decide if we want to enable class unloading as part of the
2196 // ensuing concurrent GC cycle.
2197 update_should_unload_classes();
2198 _full_gc_requested = false; // acks all outstanding full gc requests
2199 // Signal that we are about to start a collection
2200 gch->increment_total_full_collections(); // ... starting a collection cycle
2201 _collection_count_start = gch->total_full_collections();
2202 }
2203
2204 // Used for PrintGC
2205 size_t prev_used;
2206 if (PrintGC && Verbose) {
2207 prev_used = _cmsGen->used(); // XXXPERM
2208 }
2209
2210 // The change of the collection state is normally done at this level;
2211 // the exceptions are phases that are executed while the world is
2212 // stopped. For those phases the change of state is done while the
2213 // world is stopped. For baton passing purposes this allows the
2214 // background collector to finish the phase and change state atomically.
2215 // The foreground collector cannot wait on a phase that is done
2216 // while the world is stopped because the foreground collector already
2217 // has the world stopped and would deadlock.
2218 while (_collectorState != Idling) {
2219 if (TraceCMSState) {
2220 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2221 Thread::current(), _collectorState);
2222 }
2223 // The foreground collector
2224 // holds the Heap_lock throughout its collection.
2225 // holds the CMS token (but not the lock)
2226 // except while it is waiting for the background collector to yield.
2227 //
2228 // The foreground collector should be blocked (not for long)
2229 // if the background collector is about to start a phase
2230 // executed with world stopped. If the background
2231 // collector has already started such a phase, the
2232 // foreground collector is blocked waiting for the
2233 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
2234 // are executed in the VM thread.
2235 //
2236 // The locking order is
2237 // PendingListLock (PLL) -- if applicable (FinalMarking)
2238 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
2239 // CMS token (claimed in
2240 // stop_world_and_do() -->
2241 // safepoint_synchronize() -->
2242 // CMSThread::synchronize())
2243
2244 {
2245 // Check if the FG collector wants us to yield.
2246 CMSTokenSync x(true); // is cms thread
2247 if (waitForForegroundGC()) {
2248 // We yielded to a foreground GC, nothing more to be
2249 // done this round.
2250 assert(_foregroundGCShouldWait == false, "We set it to false in "
2251 "waitForForegroundGC()");
2252 if (TraceCMSState) {
2253 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2254 " exiting collection CMS state %d",
2255 Thread::current(), _collectorState);
2256 }
2257 return;
2258 } else {
2259 // The background collector can run but check to see if the
2260 // foreground collector has done a collection while the
2261 // background collector was waiting to get the CGC_lock
2262 // above. If yes, break so that _foregroundGCShouldWait
2263 // is cleared before returning.
2264 if (_collectorState == Idling) {
2265 break;
2266 }
2267 }
2268 }
2269
2270 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2271 "should be waiting");
2272
2273 switch (_collectorState) {
2274 case InitialMarking:
2275 {
2276 ReleaseForegroundGC x(this);
2277 stats().record_cms_begin();
2278
2279 VM_CMS_Initial_Mark initial_mark_op(this);
2280 VMThread::execute(&initial_mark_op);
2281 }
2282 // The collector state may be any legal state at this point
2283 // since the background collector may have yielded to the
2284 // foreground collector.
2285 break;
2286 case Marking:
2287 // initial marking in checkpointRootsInitialWork has been completed
2288 if (markFromRoots(true)) { // we were successful
2289 assert(_collectorState == Precleaning, "Collector state should "
2290 "have changed");
2291 } else {
2292 assert(_foregroundGCIsActive, "Internal state inconsistency");
2293 }
2294 break;
2295 case Precleaning:
2296 if (UseAdaptiveSizePolicy) {
2297 size_policy()->concurrent_precleaning_begin();
2298 }
2299 // marking from roots in markFromRoots has been completed
2300 preclean();
2301 if (UseAdaptiveSizePolicy) {
2302 size_policy()->concurrent_precleaning_end();
2303 }
2304 assert(_collectorState == AbortablePreclean ||
2305 _collectorState == FinalMarking,
2306 "Collector state should have changed");
2307 break;
2308 case AbortablePreclean:
2309 if (UseAdaptiveSizePolicy) {
2310 size_policy()->concurrent_phases_resume();
2311 }
2312 abortable_preclean();
2313 if (UseAdaptiveSizePolicy) {
2314 size_policy()->concurrent_precleaning_end();
2315 }
2316 assert(_collectorState == FinalMarking, "Collector state should "
2317 "have changed");
2318 break;
2319 case FinalMarking:
2320 {
2321 ReleaseForegroundGC x(this);
2322
2323 VM_CMS_Final_Remark final_remark_op(this);
2324 VMThread::execute(&final_remark_op);
2325 }
2326 assert(_foregroundGCShouldWait, "block post-condition");
2327 break;
2328 case Sweeping:
2329 if (UseAdaptiveSizePolicy) {
2330 size_policy()->concurrent_sweeping_begin();
2331 }
2332 // final marking in checkpointRootsFinal has been completed
2333 sweep(true);
2334 assert(_collectorState == Resizing, "Collector state change "
2335 "to Resizing must be done under the free_list_lock");
2336 _full_gcs_since_conc_gc = 0;
2337
2338 // Stop the timers for adaptive size policy for the concurrent phases
2339 if (UseAdaptiveSizePolicy) {
2340 size_policy()->concurrent_sweeping_end();
2341 size_policy()->concurrent_phases_end(gch->gc_cause(),
2342 gch->prev_gen(_cmsGen)->capacity(),
2343 _cmsGen->free());
2344 }
2345
2346 case Resizing: {
2347 // Sweeping has been completed...
2348 // At this point the background collection has completed.
2349 // Don't move the call to compute_new_size() down
2350 // into code that might be executed if the background
2351 // collection was preempted.
2352 {
2353 ReleaseForegroundGC x(this); // unblock FG collection
2354 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
2355 CMSTokenSync z(true); // not strictly needed.
2356 if (_collectorState == Resizing) {
2357 compute_new_size();
2358 _collectorState = Resetting;
2359 } else {
2360 assert(_collectorState == Idling, "The state should only change"
2361 " because the foreground collector has finished the collection");
2362 }
2363 }
2364 break;
2365 }
2366 case Resetting:
2367 // CMS heap resizing has been completed
2368 reset(true);
2369 assert(_collectorState == Idling, "Collector state should "
2370 "have changed");
2371 stats().record_cms_end();
2372 // Don't move the concurrent_phases_end() and compute_new_size()
2373 // calls to here because a preempted background collection
2374 // has it's state set to "Resetting".
2375 break;
2376 case Idling:
2377 default:
2378 ShouldNotReachHere();
2379 break;
2380 }
2381 if (TraceCMSState) {
2382 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2383 Thread::current(), _collectorState);
2384 }
2385 assert(_foregroundGCShouldWait, "block post-condition");
2386 }
2387
2388 // Should this be in gc_epilogue?
2389 collector_policy()->counters()->update_counters();
2390
2391 {
2392 // Clear _foregroundGCShouldWait and, in the event that the
2393 // foreground collector is waiting, notify it, before
2394 // returning.
2395 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2396 _foregroundGCShouldWait = false;
2397 if (_foregroundGCIsActive) {
2398 CGC_lock->notify();
2399 }
2400 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2401 "Possible deadlock");
2402 }
2403 if (TraceCMSState) {
2404 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2405 " exiting collection CMS state %d",
2406 Thread::current(), _collectorState);
2407 }
2408 if (PrintGC && Verbose) {
2409 _cmsGen->print_heap_change(prev_used);
2410 }
2411 }
2412
2413 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) {
2414 assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2415 "Foreground collector should be waiting, not executing");
2416 assert(Thread::current()->is_VM_thread(), "A foreground collection"
2417 "may only be done by the VM Thread with the world stopped");
2418 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2419 "VM thread should have CMS token");
2420
2421 NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2422 true, gclog_or_tty);)
2423 if (UseAdaptiveSizePolicy) {
2424 size_policy()->ms_collection_begin();
2425 }
2426 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2427
2428 HandleMark hm; // Discard invalid handles created during verification
2429
2430 if (VerifyBeforeGC &&
2431 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2432 Universe::verify(true);
2433 }
2434
2435 // Snapshot the soft reference policy to be used in this collection cycle.
2436 ref_processor()->setup_policy(clear_all_soft_refs);
2437
2438 bool init_mark_was_synchronous = false; // until proven otherwise
2439 while (_collectorState != Idling) {
2440 if (TraceCMSState) {
2441 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2442 Thread::current(), _collectorState);
2443 }
2444 switch (_collectorState) {
2445 case InitialMarking:
2446 init_mark_was_synchronous = true; // fact to be exploited in re-mark
2447 checkpointRootsInitial(false);
2448 assert(_collectorState == Marking, "Collector state should have changed"
2449 " within checkpointRootsInitial()");
2450 break;
2451 case Marking:
2452 // initial marking in checkpointRootsInitialWork has been completed
2453 if (VerifyDuringGC &&
2454 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2455 gclog_or_tty->print("Verify before initial mark: ");
2456 Universe::verify(true);
2457 }
2458 {
2459 bool res = markFromRoots(false);
2460 assert(res && _collectorState == FinalMarking, "Collector state should "
2461 "have changed");
2462 break;
2463 }
2464 case FinalMarking:
2465 if (VerifyDuringGC &&
2466 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2467 gclog_or_tty->print("Verify before re-mark: ");
2468 Universe::verify(true);
2469 }
2470 checkpointRootsFinal(false, clear_all_soft_refs,
2471 init_mark_was_synchronous);
2472 assert(_collectorState == Sweeping, "Collector state should not "
2473 "have changed within checkpointRootsFinal()");
2474 break;
2475 case Sweeping:
2476 // final marking in checkpointRootsFinal has been completed
2477 if (VerifyDuringGC &&
2478 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2479 gclog_or_tty->print("Verify before sweep: ");
2480 Universe::verify(true);
2481 }
2482 sweep(false);
2483 assert(_collectorState == Resizing, "Incorrect state");
2484 break;
2485 case Resizing: {
2486 // Sweeping has been completed; the actual resize in this case
2487 // is done separately; nothing to be done in this state.
2488 _collectorState = Resetting;
2489 break;
2490 }
2491 case Resetting:
2492 // The heap has been resized.
2493 if (VerifyDuringGC &&
2494 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2495 gclog_or_tty->print("Verify before reset: ");
2496 Universe::verify(true);
2497 }
2498 reset(false);
2499 assert(_collectorState == Idling, "Collector state should "
2500 "have changed");
2501 break;
2502 case Precleaning:
2503 case AbortablePreclean:
2504 // Elide the preclean phase
2505 _collectorState = FinalMarking;
2506 break;
2507 default:
2508 ShouldNotReachHere();
2509 }
2510 if (TraceCMSState) {
2511 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2512 Thread::current(), _collectorState);
2513 }
2514 }
2515
2516 if (UseAdaptiveSizePolicy) {
2517 GenCollectedHeap* gch = GenCollectedHeap::heap();
2518 size_policy()->ms_collection_end(gch->gc_cause());
2519 }
2520
2521 if (VerifyAfterGC &&
2522 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2523 Universe::verify(true);
2524 }
2525 if (TraceCMSState) {
2526 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2527 " exiting collection CMS state %d",
2528 Thread::current(), _collectorState);
2529 }
2530 }
2531
2532 bool CMSCollector::waitForForegroundGC() {
2533 bool res = false;
2534 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2535 "CMS thread should have CMS token");
2536 // Block the foreground collector until the
2537 // background collectors decides whether to
2538 // yield.
2539 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2540 _foregroundGCShouldWait = true;
2541 if (_foregroundGCIsActive) {
2542 // The background collector yields to the
2543 // foreground collector and returns a value
2544 // indicating that it has yielded. The foreground
2545 // collector can proceed.
2546 res = true;
2547 _foregroundGCShouldWait = false;
2548 ConcurrentMarkSweepThread::clear_CMS_flag(
2549 ConcurrentMarkSweepThread::CMS_cms_has_token);
2550 ConcurrentMarkSweepThread::set_CMS_flag(
2551 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2552 // Get a possibly blocked foreground thread going
2553 CGC_lock->notify();
2554 if (TraceCMSState) {
2555 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2556 Thread::current(), _collectorState);
2557 }
2558 while (_foregroundGCIsActive) {
2559 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2560 }
2561 ConcurrentMarkSweepThread::set_CMS_flag(
2562 ConcurrentMarkSweepThread::CMS_cms_has_token);
2563 ConcurrentMarkSweepThread::clear_CMS_flag(
2564 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2565 }
2566 if (TraceCMSState) {
2567 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2568 Thread::current(), _collectorState);
2569 }
2570 return res;
2571 }
2572
2573 // Because of the need to lock the free lists and other structures in
2574 // the collector, common to all the generations that the collector is
2575 // collecting, we need the gc_prologues of individual CMS generations
2576 // delegate to their collector. It may have been simpler had the
2577 // current infrastructure allowed one to call a prologue on a
2578 // collector. In the absence of that we have the generation's
2579 // prologue delegate to the collector, which delegates back
2580 // some "local" work to a worker method in the individual generations
2581 // that it's responsible for collecting, while itself doing any
2582 // work common to all generations it's responsible for. A similar
2583 // comment applies to the gc_epilogue()'s.
2584 // The role of the varaible _between_prologue_and_epilogue is to
2585 // enforce the invocation protocol.
2586 void CMSCollector::gc_prologue(bool full) {
2587 // Call gc_prologue_work() for each CMSGen and PermGen that
2588 // we are responsible for.
2589
2590 // The following locking discipline assumes that we are only called
2591 // when the world is stopped.
2592 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2593
2594 // The CMSCollector prologue must call the gc_prologues for the
2595 // "generations" (including PermGen if any) that it's responsible
2596 // for.
2597
2598 assert( Thread::current()->is_VM_thread()
2599 || ( CMSScavengeBeforeRemark
2600 && Thread::current()->is_ConcurrentGC_thread()),
2601 "Incorrect thread type for prologue execution");
2602
2603 if (_between_prologue_and_epilogue) {
2604 // We have already been invoked; this is a gc_prologue delegation
2605 // from yet another CMS generation that we are responsible for, just
2606 // ignore it since all relevant work has already been done.
2607 return;
2608 }
2609
2610 // set a bit saying prologue has been called; cleared in epilogue
2611 _between_prologue_and_epilogue = true;
2612 // Claim locks for common data structures, then call gc_prologue_work()
2613 // for each CMSGen and PermGen that we are responsible for.
2614
2615 getFreelistLocks(); // gets free list locks on constituent spaces
2616 bitMapLock()->lock_without_safepoint_check();
2617
2618 // Should call gc_prologue_work() for all cms gens we are responsible for
2619 bool registerClosure = _collectorState >= Marking
2620 && _collectorState < Sweeping;
2621 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ?
2622 &_modUnionClosurePar
2623 : &_modUnionClosure;
2624 _cmsGen->gc_prologue_work(full, registerClosure, muc);
2625 _permGen->gc_prologue_work(full, registerClosure, muc);
2626
2627 if (!full) {
2628 stats().record_gc0_begin();
2629 }
2630 }
2631
2632 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2633 // Delegate to CMScollector which knows how to coordinate between
2634 // this and any other CMS generations that it is responsible for
2635 // collecting.
2636 collector()->gc_prologue(full);
2637 }
2638
2639 // This is a "private" interface for use by this generation's CMSCollector.
2640 // Not to be called directly by any other entity (for instance,
2641 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2642 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2643 bool registerClosure, ModUnionClosure* modUnionClosure) {
2644 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2645 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2646 "Should be NULL");
2647 if (registerClosure) {
2648 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2649 }
2650 cmsSpace()->gc_prologue();
2651 // Clear stat counters
2652 NOT_PRODUCT(
2653 assert(_numObjectsPromoted == 0, "check");
2654 assert(_numWordsPromoted == 0, "check");
2655 if (Verbose && PrintGC) {
2656 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2657 SIZE_FORMAT" bytes concurrently",
2658 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2659 }
2660 _numObjectsAllocated = 0;
2661 _numWordsAllocated = 0;
2662 )
2663 }
2664
2665 void CMSCollector::gc_epilogue(bool full) {
2666 // The following locking discipline assumes that we are only called
2667 // when the world is stopped.
2668 assert(SafepointSynchronize::is_at_safepoint(),
2669 "world is stopped assumption");
2670
2671 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2672 // if linear allocation blocks need to be appropriately marked to allow the
2673 // the blocks to be parsable. We also check here whether we need to nudge the
2674 // CMS collector thread to start a new cycle (if it's not already active).
2675 assert( Thread::current()->is_VM_thread()
2676 || ( CMSScavengeBeforeRemark
2677 && Thread::current()->is_ConcurrentGC_thread()),
2678 "Incorrect thread type for epilogue execution");
2679
2680 if (!_between_prologue_and_epilogue) {
2681 // We have already been invoked; this is a gc_epilogue delegation
2682 // from yet another CMS generation that we are responsible for, just
2683 // ignore it since all relevant work has already been done.
2684 return;
2685 }
2686 assert(haveFreelistLocks(), "must have freelist locks");
2687 assert_lock_strong(bitMapLock());
2688
2689 _cmsGen->gc_epilogue_work(full);
2690 _permGen->gc_epilogue_work(full);
2691
2692 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2693 // in case sampling was not already enabled, enable it
2694 _start_sampling = true;
2695 }
2696 // reset _eden_chunk_array so sampling starts afresh
2697 _eden_chunk_index = 0;
2698
2699 size_t cms_used = _cmsGen->cmsSpace()->used();
2700 size_t perm_used = _permGen->cmsSpace()->used();
2701
2702 // update performance counters - this uses a special version of
2703 // update_counters() that allows the utilization to be passed as a
2704 // parameter, avoiding multiple calls to used().
2705 //
2706 _cmsGen->update_counters(cms_used);
2707 _permGen->update_counters(perm_used);
2708
2709 if (CMSIncrementalMode) {
2710 icms_update_allocation_limits();
2711 }
2712
2713 bitMapLock()->unlock();
2714 releaseFreelistLocks();
2715
2716 if (!CleanChunkPoolAsync) {
2717 Chunk::clean_chunk_pool();
2718 }
2719
2720 _between_prologue_and_epilogue = false; // ready for next cycle
2721 }
2722
2723 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2724 collector()->gc_epilogue(full);
2725
2726 // Also reset promotion tracking in par gc thread states.
2727 if (CollectedHeap::use_parallel_gc_threads()) {
2728 for (uint i = 0; i < ParallelGCThreads; i++) {
2729 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2730 }
2731 }
2732 }
2733
2734 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2735 assert(!incremental_collection_failed(), "Should have been cleared");
2736 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2737 cmsSpace()->gc_epilogue();
2738 // Print stat counters
2739 NOT_PRODUCT(
2740 assert(_numObjectsAllocated == 0, "check");
2741 assert(_numWordsAllocated == 0, "check");
2742 if (Verbose && PrintGC) {
2743 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2744 SIZE_FORMAT" bytes",
2745 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2746 }
2747 _numObjectsPromoted = 0;
2748 _numWordsPromoted = 0;
2749 )
2750
2751 if (PrintGC && Verbose) {
2752 // Call down the chain in contiguous_available needs the freelistLock
2753 // so print this out before releasing the freeListLock.
2754 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2755 contiguous_available());
2756 }
2757 }
2758
2759 #ifndef PRODUCT
2760 bool CMSCollector::have_cms_token() {
2761 Thread* thr = Thread::current();
2762 if (thr->is_VM_thread()) {
2763 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2764 } else if (thr->is_ConcurrentGC_thread()) {
2765 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2766 } else if (thr->is_GC_task_thread()) {
2767 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2768 ParGCRareEvent_lock->owned_by_self();
2769 }
2770 return false;
2771 }
2772 #endif
2773
2774 // Check reachability of the given heap address in CMS generation,
2775 // treating all other generations as roots.
2776 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2777 // We could "guarantee" below, rather than assert, but i'll
2778 // leave these as "asserts" so that an adventurous debugger
2779 // could try this in the product build provided some subset of
2780 // the conditions were met, provided they were intersted in the
2781 // results and knew that the computation below wouldn't interfere
2782 // with other concurrent computations mutating the structures
2783 // being read or written.
2784 assert(SafepointSynchronize::is_at_safepoint(),
2785 "Else mutations in object graph will make answer suspect");
2786 assert(have_cms_token(), "Should hold cms token");
2787 assert(haveFreelistLocks(), "must hold free list locks");
2788 assert_lock_strong(bitMapLock());
2789
2790 // Clear the marking bit map array before starting, but, just
2791 // for kicks, first report if the given address is already marked
2792 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
2793 _markBitMap.isMarked(addr) ? "" : " not");
2794
2795 if (verify_after_remark()) {
2796 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2797 bool result = verification_mark_bm()->isMarked(addr);
2798 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
2799 result ? "IS" : "is NOT");
2800 return result;
2801 } else {
2802 gclog_or_tty->print_cr("Could not compute result");
2803 return false;
2804 }
2805 }
2806
2807 ////////////////////////////////////////////////////////
2808 // CMS Verification Support
2809 ////////////////////////////////////////////////////////
2810 // Following the remark phase, the following invariant
2811 // should hold -- each object in the CMS heap which is
2812 // marked in markBitMap() should be marked in the verification_mark_bm().
2813
2814 class VerifyMarkedClosure: public BitMapClosure {
2815 CMSBitMap* _marks;
2816 bool _failed;
2817
2818 public:
2819 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2820
2821 bool do_bit(size_t offset) {
2822 HeapWord* addr = _marks->offsetToHeapWord(offset);
2823 if (!_marks->isMarked(addr)) {
2824 oop(addr)->print_on(gclog_or_tty);
2825 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2826 _failed = true;
2827 }
2828 return true;
2829 }
2830
2831 bool failed() { return _failed; }
2832 };
2833
2834 bool CMSCollector::verify_after_remark() {
2835 gclog_or_tty->print(" [Verifying CMS Marking... ");
2836 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2837 static bool init = false;
2838
2839 assert(SafepointSynchronize::is_at_safepoint(),
2840 "Else mutations in object graph will make answer suspect");
2841 assert(have_cms_token(),
2842 "Else there may be mutual interference in use of "
2843 " verification data structures");
2844 assert(_collectorState > Marking && _collectorState <= Sweeping,
2845 "Else marking info checked here may be obsolete");
2846 assert(haveFreelistLocks(), "must hold free list locks");
2847 assert_lock_strong(bitMapLock());
2848
2849
2850 // Allocate marking bit map if not already allocated
2851 if (!init) { // first time
2852 if (!verification_mark_bm()->allocate(_span)) {
2853 return false;
2854 }
2855 init = true;
2856 }
2857
2858 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2859
2860 // Turn off refs discovery -- so we will be tracing through refs.
2861 // This is as intended, because by this time
2862 // GC must already have cleared any refs that need to be cleared,
2863 // and traced those that need to be marked; moreover,
2864 // the marking done here is not going to intefere in any
2865 // way with the marking information used by GC.
2866 NoRefDiscovery no_discovery(ref_processor());
2867
2868 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2869
2870 // Clear any marks from a previous round
2871 verification_mark_bm()->clear_all();
2872 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2873 verify_work_stacks_empty();
2874
2875 GenCollectedHeap* gch = GenCollectedHeap::heap();
2876 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2877 // Update the saved marks which may affect the root scans.
2878 gch->save_marks();
2879
2880 if (CMSRemarkVerifyVariant == 1) {
2881 // In this first variant of verification, we complete
2882 // all marking, then check if the new marks-verctor is
2883 // a subset of the CMS marks-vector.
2884 verify_after_remark_work_1();
2885 } else if (CMSRemarkVerifyVariant == 2) {
2886 // In this second variant of verification, we flag an error
2887 // (i.e. an object reachable in the new marks-vector not reachable
2888 // in the CMS marks-vector) immediately, also indicating the
2889 // identify of an object (A) that references the unmarked object (B) --
2890 // presumably, a mutation to A failed to be picked up by preclean/remark?
2891 verify_after_remark_work_2();
2892 } else {
2893 warning("Unrecognized value %d for CMSRemarkVerifyVariant",
2894 CMSRemarkVerifyVariant);
2895 }
2896 gclog_or_tty->print(" done] ");
2897 return true;
2898 }
2899
2900 void CMSCollector::verify_after_remark_work_1() {
2901 ResourceMark rm;
2902 HandleMark hm;
2903 GenCollectedHeap* gch = GenCollectedHeap::heap();
2904
2905 // Mark from roots one level into CMS
2906 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2907 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2908
2909 gch->gen_process_strong_roots(_cmsGen->level(),
2910 true, // younger gens are roots
2911 true, // activate StrongRootsScope
2912 true, // collecting perm gen
2913 SharedHeap::ScanningOption(roots_scanning_options()),
2914 ¬Older,
2915 true, // walk code active on stacks
2916 NULL);
2917
2918 // Now mark from the roots
2919 assert(_revisitStack.isEmpty(), "Should be empty");
2920 MarkFromRootsClosure markFromRootsClosure(this, _span,
2921 verification_mark_bm(), verification_mark_stack(), &_revisitStack,
2922 false /* don't yield */, true /* verifying */);
2923 assert(_restart_addr == NULL, "Expected pre-condition");
2924 verification_mark_bm()->iterate(&markFromRootsClosure);
2925 while (_restart_addr != NULL) {
2926 // Deal with stack overflow: by restarting at the indicated
2927 // address.
2928 HeapWord* ra = _restart_addr;
2929 markFromRootsClosure.reset(ra);
2930 _restart_addr = NULL;
2931 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2932 }
2933 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2934 verify_work_stacks_empty();
2935 // Should reset the revisit stack above, since no class tree
2936 // surgery is forthcoming.
2937 _revisitStack.reset(); // throwing away all contents
2938
2939 // Marking completed -- now verify that each bit marked in
2940 // verification_mark_bm() is also marked in markBitMap(); flag all
2941 // errors by printing corresponding objects.
2942 VerifyMarkedClosure vcl(markBitMap());
2943 verification_mark_bm()->iterate(&vcl);
2944 if (vcl.failed()) {
2945 gclog_or_tty->print("Verification failed");
2946 Universe::heap()->print_on(gclog_or_tty);
2947 fatal("CMS: failed marking verification after remark");
2948 }
2949 }
2950
2951 void CMSCollector::verify_after_remark_work_2() {
2952 ResourceMark rm;
2953 HandleMark hm;
2954 GenCollectedHeap* gch = GenCollectedHeap::heap();
2955
2956 // Mark from roots one level into CMS
2957 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2958 markBitMap());
2959 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2960 gch->gen_process_strong_roots(_cmsGen->level(),
2961 true, // younger gens are roots
2962 true, // activate StrongRootsScope
2963 true, // collecting perm gen
2964 SharedHeap::ScanningOption(roots_scanning_options()),
2965 ¬Older,
2966 true, // walk code active on stacks
2967 NULL);
2968
2969 // Now mark from the roots
2970 assert(_revisitStack.isEmpty(), "Should be empty");
2971 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2972 verification_mark_bm(), markBitMap(), verification_mark_stack());
2973 assert(_restart_addr == NULL, "Expected pre-condition");
2974 verification_mark_bm()->iterate(&markFromRootsClosure);
2975 while (_restart_addr != NULL) {
2976 // Deal with stack overflow: by restarting at the indicated
2977 // address.
2978 HeapWord* ra = _restart_addr;
2979 markFromRootsClosure.reset(ra);
2980 _restart_addr = NULL;
2981 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2982 }
2983 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2984 verify_work_stacks_empty();
2985 // Should reset the revisit stack above, since no class tree
2986 // surgery is forthcoming.
2987 _revisitStack.reset(); // throwing away all contents
2988
2989 // Marking completed -- now verify that each bit marked in
2990 // verification_mark_bm() is also marked in markBitMap(); flag all
2991 // errors by printing corresponding objects.
2992 VerifyMarkedClosure vcl(markBitMap());
2993 verification_mark_bm()->iterate(&vcl);
2994 assert(!vcl.failed(), "Else verification above should not have succeeded");
2995 }
2996
2997 void ConcurrentMarkSweepGeneration::save_marks() {
2998 // delegate to CMS space
2999 cmsSpace()->save_marks();
3000 for (uint i = 0; i < ParallelGCThreads; i++) {
3001 _par_gc_thread_states[i]->promo.startTrackingPromotions();
3002 }
3003 }
3004
3005 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
3006 return cmsSpace()->no_allocs_since_save_marks();
3007 }
3008
3009 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
3010 \
3011 void ConcurrentMarkSweepGeneration:: \
3012 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
3013 cl->set_generation(this); \
3014 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
3015 cl->reset_generation(); \
3016 save_marks(); \
3017 }
3018
3019 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
3020
3021 void
3022 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk)
3023 {
3024 // Not currently implemented; need to do the following. -- ysr.
3025 // dld -- I think that is used for some sort of allocation profiler. So it
3026 // really means the objects allocated by the mutator since the last
3027 // GC. We could potentially implement this cheaply by recording only
3028 // the direct allocations in a side data structure.
3029 //
3030 // I think we probably ought not to be required to support these
3031 // iterations at any arbitrary point; I think there ought to be some
3032 // call to enable/disable allocation profiling in a generation/space,
3033 // and the iterator ought to return the objects allocated in the
3034 // gen/space since the enable call, or the last iterator call (which
3035 // will probably be at a GC.) That way, for gens like CM&S that would
3036 // require some extra data structure to support this, we only pay the
3037 // cost when it's in use...
3038 cmsSpace()->object_iterate_since_last_GC(blk);
3039 }
3040
3041 void
3042 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
3043 cl->set_generation(this);
3044 younger_refs_in_space_iterate(_cmsSpace, cl);
3045 cl->reset_generation();
3046 }
3047
3048 void
3049 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) {
3050 if (freelistLock()->owned_by_self()) {
3051 Generation::oop_iterate(mr, cl);
3052 } else {
3053 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3054 Generation::oop_iterate(mr, cl);
3055 }
3056 }
3057
3058 void
3059 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) {
3060 if (freelistLock()->owned_by_self()) {
3061 Generation::oop_iterate(cl);
3062 } else {
3063 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3064 Generation::oop_iterate(cl);
3065 }
3066 }
3067
3068 void
3069 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
3070 if (freelistLock()->owned_by_self()) {
3071 Generation::object_iterate(cl);
3072 } else {
3073 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3074 Generation::object_iterate(cl);
3075 }
3076 }
3077
3078 void
3079 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
3080 if (freelistLock()->owned_by_self()) {
3081 Generation::safe_object_iterate(cl);
3082 } else {
3083 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3084 Generation::safe_object_iterate(cl);
3085 }
3086 }
3087
3088 void
3089 ConcurrentMarkSweepGeneration::pre_adjust_pointers() {
3090 }
3091
3092 void
3093 ConcurrentMarkSweepGeneration::post_compact() {
3094 }
3095
3096 void
3097 ConcurrentMarkSweepGeneration::prepare_for_verify() {
3098 // Fix the linear allocation blocks to look like free blocks.
3099
3100 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3101 // are not called when the heap is verified during universe initialization and
3102 // at vm shutdown.
3103 if (freelistLock()->owned_by_self()) {
3104 cmsSpace()->prepare_for_verify();
3105 } else {
3106 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3107 cmsSpace()->prepare_for_verify();
3108 }
3109 }
3110
3111 void
3112 ConcurrentMarkSweepGeneration::verify(bool allow_dirty /* ignored */) {
3113 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3114 // are not called when the heap is verified during universe initialization and
3115 // at vm shutdown.
3116 if (freelistLock()->owned_by_self()) {
3117 cmsSpace()->verify(false /* ignored */);
3118 } else {
3119 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3120 cmsSpace()->verify(false /* ignored */);
3121 }
3122 }
3123
3124 void CMSCollector::verify(bool allow_dirty /* ignored */) {
3125 _cmsGen->verify(allow_dirty);
3126 _permGen->verify(allow_dirty);
3127 }
3128
3129 #ifndef PRODUCT
3130 bool CMSCollector::overflow_list_is_empty() const {
3131 assert(_num_par_pushes >= 0, "Inconsistency");
3132 if (_overflow_list == NULL) {
3133 assert(_num_par_pushes == 0, "Inconsistency");
3134 }
3135 return _overflow_list == NULL;
3136 }
3137
3138 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3139 // merely consolidate assertion checks that appear to occur together frequently.
3140 void CMSCollector::verify_work_stacks_empty() const {
3141 assert(_markStack.isEmpty(), "Marking stack should be empty");
3142 assert(overflow_list_is_empty(), "Overflow list should be empty");
3143 }
3144
3145 void CMSCollector::verify_overflow_empty() const {
3146 assert(overflow_list_is_empty(), "Overflow list should be empty");
3147 assert(no_preserved_marks(), "No preserved marks");
3148 }
3149 #endif // PRODUCT
3150
3151 // Decide if we want to enable class unloading as part of the
3152 // ensuing concurrent GC cycle. We will collect the perm gen and
3153 // unload classes if it's the case that:
3154 // (1) an explicit gc request has been made and the flag
3155 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3156 // (2) (a) class unloading is enabled at the command line, and
3157 // (b) (i) perm gen threshold has been crossed, or
3158 // (ii) old gen is getting really full, or
3159 // (iii) the previous N CMS collections did not collect the
3160 // perm gen
3161 // NOTE: Provided there is no change in the state of the heap between
3162 // calls to this method, it should have idempotent results. Moreover,
3163 // its results should be monotonically increasing (i.e. going from 0 to 1,
3164 // but not 1 to 0) between successive calls between which the heap was
3165 // not collected. For the implementation below, it must thus rely on
3166 // the property that concurrent_cycles_since_last_unload()
3167 // will not decrease unless a collection cycle happened and that
3168 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are
3169 // themselves also monotonic in that sense. See check_monotonicity()
3170 // below.
3171 bool CMSCollector::update_should_unload_classes() {
3172 _should_unload_classes = false;
3173 // Condition 1 above
3174 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3175 _should_unload_classes = true;
3176 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3177 // Disjuncts 2.b.(i,ii,iii) above
3178 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3179 CMSClassUnloadingMaxInterval)
3180 || _permGen->should_concurrent_collect()
3181 || _cmsGen->is_too_full();
3182 }
3183 return _should_unload_classes;
3184 }
3185
3186 bool ConcurrentMarkSweepGeneration::is_too_full() const {
3187 bool res = should_concurrent_collect();
3188 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
3189 return res;
3190 }
3191
3192 void CMSCollector::setup_cms_unloading_and_verification_state() {
3193 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3194 || VerifyBeforeExit;
3195 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
3196
3197 if (should_unload_classes()) { // Should unload classes this cycle
3198 remove_root_scanning_option(rso); // Shrink the root set appropriately
3199 set_verifying(should_verify); // Set verification state for this cycle
3200 return; // Nothing else needs to be done at this time
3201 }
3202
3203 // Not unloading classes this cycle
3204 assert(!should_unload_classes(), "Inconsitency!");
3205 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3206 // We were not verifying, or we _were_ unloading classes in the last cycle,
3207 // AND some verification options are enabled this cycle; in this case,
3208 // we must make sure that the deadness map is allocated if not already so,
3209 // and cleared (if already allocated previously --
3210 // CMSBitMap::sizeInBits() is used to determine if it's allocated).
3211 if (perm_gen_verify_bit_map()->sizeInBits() == 0) {
3212 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) {
3213 warning("Failed to allocate permanent generation verification CMS Bit Map;\n"
3214 "permanent generation verification disabled");
3215 return; // Note that we leave verification disabled, so we'll retry this
3216 // allocation next cycle. We _could_ remember this failure
3217 // and skip further attempts and permanently disable verification
3218 // attempts if that is considered more desirable.
3219 }
3220 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()),
3221 "_perm_gen_ver_bit_map inconsistency?");
3222 } else {
3223 perm_gen_verify_bit_map()->clear_all();
3224 }
3225 // Include symbols, strings and code cache elements to prevent their resurrection.
3226 add_root_scanning_option(rso);
3227 set_verifying(true);
3228 } else if (verifying() && !should_verify) {
3229 // We were verifying, but some verification flags got disabled.
3230 set_verifying(false);
3231 // Exclude symbols, strings and code cache elements from root scanning to
3232 // reduce IM and RM pauses.
3233 remove_root_scanning_option(rso);
3234 }
3235 }
3236
3237
3238 #ifndef PRODUCT
3239 HeapWord* CMSCollector::block_start(const void* p) const {
3240 const HeapWord* addr = (HeapWord*)p;
3241 if (_span.contains(p)) {
3242 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3243 return _cmsGen->cmsSpace()->block_start(p);
3244 } else {
3245 assert(_permGen->cmsSpace()->is_in_reserved(addr),
3246 "Inconsistent _span?");
3247 return _permGen->cmsSpace()->block_start(p);
3248 }
3249 }
3250 return NULL;
3251 }
3252 #endif
3253
3254 HeapWord*
3255 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3256 bool tlab,
3257 bool parallel) {
3258 CMSSynchronousYieldRequest yr;
3259 assert(!tlab, "Can't deal with TLAB allocation");
3260 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3261 expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3262 CMSExpansionCause::_satisfy_allocation);
3263 if (GCExpandToAllocateDelayMillis > 0) {
3264 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3265 }
3266 return have_lock_and_allocate(word_size, tlab);
3267 }
3268
3269 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3270 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3271 // to CardGeneration and share it...
3272 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) {
3273 return CardGeneration::expand(bytes, expand_bytes);
3274 }
3275
3276 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3277 CMSExpansionCause::Cause cause)
3278 {
3279
3280 bool success = expand(bytes, expand_bytes);
3281
3282 // remember why we expanded; this information is used
3283 // by shouldConcurrentCollect() when making decisions on whether to start
3284 // a new CMS cycle.
3285 if (success) {
3286 set_expansion_cause(cause);
3287 if (PrintGCDetails && Verbose) {
3288 gclog_or_tty->print_cr("Expanded CMS gen for %s",
3289 CMSExpansionCause::to_string(cause));
3290 }
3291 }
3292 }
3293
3294 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3295 HeapWord* res = NULL;
3296 MutexLocker x(ParGCRareEvent_lock);
3297 while (true) {
3298 // Expansion by some other thread might make alloc OK now:
3299 res = ps->lab.alloc(word_sz);
3300 if (res != NULL) return res;
3301 // If there's not enough expansion space available, give up.
3302 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3303 return NULL;
3304 }
3305 // Otherwise, we try expansion.
3306 expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3307 CMSExpansionCause::_allocate_par_lab);
3308 // Now go around the loop and try alloc again;
3309 // A competing par_promote might beat us to the expansion space,
3310 // so we may go around the loop again if promotion fails agaion.
3311 if (GCExpandToAllocateDelayMillis > 0) {
3312 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3313 }
3314 }
3315 }
3316
3317
3318 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3319 PromotionInfo* promo) {
3320 MutexLocker x(ParGCRareEvent_lock);
3321 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3322 while (true) {
3323 // Expansion by some other thread might make alloc OK now:
3324 if (promo->ensure_spooling_space()) {
3325 assert(promo->has_spooling_space(),
3326 "Post-condition of successful ensure_spooling_space()");
3327 return true;
3328 }
3329 // If there's not enough expansion space available, give up.
3330 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3331 return false;
3332 }
3333 // Otherwise, we try expansion.
3334 expand(refill_size_bytes, MinHeapDeltaBytes,
3335 CMSExpansionCause::_allocate_par_spooling_space);
3336 // Now go around the loop and try alloc again;
3337 // A competing allocation might beat us to the expansion space,
3338 // so we may go around the loop again if allocation fails again.
3339 if (GCExpandToAllocateDelayMillis > 0) {
3340 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3341 }
3342 }
3343 }
3344
3345
3346
3347 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3348 assert_locked_or_safepoint(Heap_lock);
3349 size_t size = ReservedSpace::page_align_size_down(bytes);
3350 if (size > 0) {
3351 shrink_by(size);
3352 }
3353 }
3354
3355 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3356 assert_locked_or_safepoint(Heap_lock);
3357 bool result = _virtual_space.expand_by(bytes);
3358 if (result) {
3359 HeapWord* old_end = _cmsSpace->end();
3360 size_t new_word_size =
3361 heap_word_size(_virtual_space.committed_size());
3362 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3363 _bts->resize(new_word_size); // resize the block offset shared array
3364 Universe::heap()->barrier_set()->resize_covered_region(mr);
3365 // Hmmmm... why doesn't CFLS::set_end verify locking?
3366 // This is quite ugly; FIX ME XXX
3367 _cmsSpace->assert_locked(freelistLock());
3368 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3369
3370 // update the space and generation capacity counters
3371 if (UsePerfData) {
3372 _space_counters->update_capacity();
3373 _gen_counters->update_all();
3374 }
3375
3376 if (Verbose && PrintGC) {
3377 size_t new_mem_size = _virtual_space.committed_size();
3378 size_t old_mem_size = new_mem_size - bytes;
3379 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK",
3380 name(), old_mem_size/K, bytes/K, new_mem_size/K);
3381 }
3382 }
3383 return result;
3384 }
3385
3386 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3387 assert_locked_or_safepoint(Heap_lock);
3388 bool success = true;
3389 const size_t remaining_bytes = _virtual_space.uncommitted_size();
3390 if (remaining_bytes > 0) {
3391 success = grow_by(remaining_bytes);
3392 DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3393 }
3394 return success;
3395 }
3396
3397 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3398 assert_locked_or_safepoint(Heap_lock);
3399 assert_lock_strong(freelistLock());
3400 // XXX Fix when compaction is implemented.
3401 warning("Shrinking of CMS not yet implemented");
3402 return;
3403 }
3404
3405
3406 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3407 // phases.
3408 class CMSPhaseAccounting: public StackObj {
3409 public:
3410 CMSPhaseAccounting(CMSCollector *collector,
3411 const char *phase,
3412 bool print_cr = true);
3413 ~CMSPhaseAccounting();
3414
3415 private:
3416 CMSCollector *_collector;
3417 const char *_phase;
3418 elapsedTimer _wallclock;
3419 bool _print_cr;
3420
3421 public:
3422 // Not MT-safe; so do not pass around these StackObj's
3423 // where they may be accessed by other threads.
3424 jlong wallclock_millis() {
3425 assert(_wallclock.is_active(), "Wall clock should not stop");
3426 _wallclock.stop(); // to record time
3427 jlong ret = _wallclock.milliseconds();
3428 _wallclock.start(); // restart
3429 return ret;
3430 }
3431 };
3432
3433 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3434 const char *phase,
3435 bool print_cr) :
3436 _collector(collector), _phase(phase), _print_cr(print_cr) {
3437
3438 if (PrintCMSStatistics != 0) {
3439 _collector->resetYields();
3440 }
3441 if (PrintGCDetails && PrintGCTimeStamps) {
3442 gclog_or_tty->date_stamp(PrintGCDateStamps);
3443 gclog_or_tty->stamp();
3444 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]",
3445 _collector->cmsGen()->short_name(), _phase);
3446 }
3447 _collector->resetTimer();
3448 _wallclock.start();
3449 _collector->startTimer();
3450 }
3451
3452 CMSPhaseAccounting::~CMSPhaseAccounting() {
3453 assert(_wallclock.is_active(), "Wall clock should not have stopped");
3454 _collector->stopTimer();
3455 _wallclock.stop();
3456 if (PrintGCDetails) {
3457 gclog_or_tty->date_stamp(PrintGCDateStamps);
3458 if (PrintGCTimeStamps) {
3459 gclog_or_tty->stamp();
3460 gclog_or_tty->print(": ");
3461 }
3462 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3463 _collector->cmsGen()->short_name(),
3464 _phase, _collector->timerValue(), _wallclock.seconds());
3465 if (_print_cr) {
3466 gclog_or_tty->print_cr("");
3467 }
3468 if (PrintCMSStatistics != 0) {
3469 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3470 _collector->yields());
3471 }
3472 }
3473 }
3474
3475 // CMS work
3476
3477 // Checkpoint the roots into this generation from outside
3478 // this generation. [Note this initial checkpoint need only
3479 // be approximate -- we'll do a catch up phase subsequently.]
3480 void CMSCollector::checkpointRootsInitial(bool asynch) {
3481 assert(_collectorState == InitialMarking, "Wrong collector state");
3482 check_correct_thread_executing();
3483 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
3484
3485 ReferenceProcessor* rp = ref_processor();
3486 SpecializationStats::clear();
3487 assert(_restart_addr == NULL, "Control point invariant");
3488 if (asynch) {
3489 // acquire locks for subsequent manipulations
3490 MutexLockerEx x(bitMapLock(),
3491 Mutex::_no_safepoint_check_flag);
3492 checkpointRootsInitialWork(asynch);
3493 rp->verify_no_references_recorded();
3494 rp->enable_discovery(); // enable ("weak") refs discovery
3495 _collectorState = Marking;
3496 } else {
3497 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3498 // which recognizes if we are a CMS generation, and doesn't try to turn on
3499 // discovery; verify that they aren't meddling.
3500 assert(!rp->discovery_is_atomic(),
3501 "incorrect setting of discovery predicate");
3502 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3503 "ref discovery for this generation kind");
3504 // already have locks
3505 checkpointRootsInitialWork(asynch);
3506 rp->enable_discovery(); // now enable ("weak") refs discovery
3507 _collectorState = Marking;
3508 }
3509 SpecializationStats::print();
3510 }
3511
3512 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3513 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3514 assert(_collectorState == InitialMarking, "just checking");
3515
3516 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3517 // precede our marking with a collection of all
3518 // younger generations to keep floating garbage to a minimum.
3519 // XXX: we won't do this for now -- it's an optimization to be done later.
3520
3521 // already have locks
3522 assert_lock_strong(bitMapLock());
3523 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3524
3525 // Setup the verification and class unloading state for this
3526 // CMS collection cycle.
3527 setup_cms_unloading_and_verification_state();
3528
3529 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork",
3530 PrintGCDetails && Verbose, true, gclog_or_tty);)
3531 if (UseAdaptiveSizePolicy) {
3532 size_policy()->checkpoint_roots_initial_begin();
3533 }
3534
3535 // Reset all the PLAB chunk arrays if necessary.
3536 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3537 reset_survivor_plab_arrays();
3538 }
3539
3540 ResourceMark rm;
3541 HandleMark hm;
3542
3543 FalseClosure falseClosure;
3544 // In the case of a synchronous collection, we will elide the
3545 // remark step, so it's important to catch all the nmethod oops
3546 // in this step.
3547 // The final 'true' flag to gen_process_strong_roots will ensure this.
3548 // If 'async' is true, we can relax the nmethod tracing.
3549 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
3550 GenCollectedHeap* gch = GenCollectedHeap::heap();
3551
3552 verify_work_stacks_empty();
3553 verify_overflow_empty();
3554
3555 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3556 // Update the saved marks which may affect the root scans.
3557 gch->save_marks();
3558
3559 // weak reference processing has not started yet.
3560 ref_processor()->set_enqueuing_is_done(false);
3561
3562 {
3563 // This is not needed. DEBUG_ONLY(RememberKlassesChecker imx(true);)
3564 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3565 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3566 gch->gen_process_strong_roots(_cmsGen->level(),
3567 true, // younger gens are roots
3568 true, // activate StrongRootsScope
3569 true, // collecting perm gen
3570 SharedHeap::ScanningOption(roots_scanning_options()),
3571 ¬Older,
3572 true, // walk all of code cache if (so & SO_CodeCache)
3573 NULL);
3574 }
3575
3576 // Clear mod-union table; it will be dirtied in the prologue of
3577 // CMS generation per each younger generation collection.
3578
3579 assert(_modUnionTable.isAllClear(),
3580 "Was cleared in most recent final checkpoint phase"
3581 " or no bits are set in the gc_prologue before the start of the next "
3582 "subsequent marking phase.");
3583
3584 // Temporarily disabled, since pre/post-consumption closures don't
3585 // care about precleaned cards
3586 #if 0
3587 {
3588 MemRegion mr = MemRegion((HeapWord*)_virtual_space.low(),
3589 (HeapWord*)_virtual_space.high());
3590 _ct->ct_bs()->preclean_dirty_cards(mr);
3591 }
3592 #endif
3593
3594 // Save the end of the used_region of the constituent generations
3595 // to be used to limit the extent of sweep in each generation.
3596 save_sweep_limits();
3597 if (UseAdaptiveSizePolicy) {
3598 size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3599 }
3600 verify_overflow_empty();
3601 }
3602
3603 bool CMSCollector::markFromRoots(bool asynch) {
3604 // we might be tempted to assert that:
3605 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3606 // "inconsistent argument?");
3607 // However that wouldn't be right, because it's possible that
3608 // a safepoint is indeed in progress as a younger generation
3609 // stop-the-world GC happens even as we mark in this generation.
3610 assert(_collectorState == Marking, "inconsistent state?");
3611 check_correct_thread_executing();
3612 verify_overflow_empty();
3613
3614 bool res;
3615 if (asynch) {
3616
3617 // Start the timers for adaptive size policy for the concurrent phases
3618 // Do it here so that the foreground MS can use the concurrent
3619 // timer since a foreground MS might has the sweep done concurrently
3620 // or STW.
3621 if (UseAdaptiveSizePolicy) {
3622 size_policy()->concurrent_marking_begin();
3623 }
3624
3625 // Weak ref discovery note: We may be discovering weak
3626 // refs in this generation concurrent (but interleaved) with
3627 // weak ref discovery by a younger generation collector.
3628
3629 CMSTokenSyncWithLocks ts(true, bitMapLock());
3630 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3631 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3632 res = markFromRootsWork(asynch);
3633 if (res) {
3634 _collectorState = Precleaning;
3635 } else { // We failed and a foreground collection wants to take over
3636 assert(_foregroundGCIsActive, "internal state inconsistency");
3637 assert(_restart_addr == NULL, "foreground will restart from scratch");
3638 if (PrintGCDetails) {
3639 gclog_or_tty->print_cr("bailing out to foreground collection");
3640 }
3641 }
3642 if (UseAdaptiveSizePolicy) {
3643 size_policy()->concurrent_marking_end();
3644 }
3645 } else {
3646 assert(SafepointSynchronize::is_at_safepoint(),
3647 "inconsistent with asynch == false");
3648 if (UseAdaptiveSizePolicy) {
3649 size_policy()->ms_collection_marking_begin();
3650 }
3651 // already have locks
3652 res = markFromRootsWork(asynch);
3653 _collectorState = FinalMarking;
3654 if (UseAdaptiveSizePolicy) {
3655 GenCollectedHeap* gch = GenCollectedHeap::heap();
3656 size_policy()->ms_collection_marking_end(gch->gc_cause());
3657 }
3658 }
3659 verify_overflow_empty();
3660 return res;
3661 }
3662
3663 bool CMSCollector::markFromRootsWork(bool asynch) {
3664 // iterate over marked bits in bit map, doing a full scan and mark
3665 // from these roots using the following algorithm:
3666 // . if oop is to the right of the current scan pointer,
3667 // mark corresponding bit (we'll process it later)
3668 // . else (oop is to left of current scan pointer)
3669 // push oop on marking stack
3670 // . drain the marking stack
3671
3672 // Note that when we do a marking step we need to hold the
3673 // bit map lock -- recall that direct allocation (by mutators)
3674 // and promotion (by younger generation collectors) is also
3675 // marking the bit map. [the so-called allocate live policy.]
3676 // Because the implementation of bit map marking is not
3677 // robust wrt simultaneous marking of bits in the same word,
3678 // we need to make sure that there is no such interference
3679 // between concurrent such updates.
3680
3681 // already have locks
3682 assert_lock_strong(bitMapLock());
3683
3684 // Clear the revisit stack, just in case there are any
3685 // obsolete contents from a short-circuited previous CMS cycle.
3686 _revisitStack.reset();
3687 verify_work_stacks_empty();
3688 verify_overflow_empty();
3689 assert(_revisitStack.isEmpty(), "tabula rasa");
3690 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());)
3691 bool result = false;
3692 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3693 result = do_marking_mt(asynch);
3694 } else {
3695 result = do_marking_st(asynch);
3696 }
3697 return result;
3698 }
3699
3700 // Forward decl
3701 class CMSConcMarkingTask;
3702
3703 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3704 CMSCollector* _collector;
3705 CMSConcMarkingTask* _task;
3706 public:
3707 virtual void yield();
3708
3709 // "n_threads" is the number of threads to be terminated.
3710 // "queue_set" is a set of work queues of other threads.
3711 // "collector" is the CMS collector associated with this task terminator.
3712 // "yield" indicates whether we need the gang as a whole to yield.
3713 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3714 ParallelTaskTerminator(n_threads, queue_set),
3715 _collector(collector) { }
3716
3717 void set_task(CMSConcMarkingTask* task) {
3718 _task = task;
3719 }
3720 };
3721
3722 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3723 CMSConcMarkingTask* _task;
3724 public:
3725 bool should_exit_termination();
3726 void set_task(CMSConcMarkingTask* task) {
3727 _task = task;
3728 }
3729 };
3730
3731 // MT Concurrent Marking Task
3732 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3733 CMSCollector* _collector;
3734 int _n_workers; // requested/desired # workers
3735 bool _asynch;
3736 bool _result;
3737 CompactibleFreeListSpace* _cms_space;
3738 CompactibleFreeListSpace* _perm_space;
3739 char _pad_front[64]; // padding to ...
3740 HeapWord* _global_finger; // ... avoid sharing cache line
3741 char _pad_back[64];
3742 HeapWord* _restart_addr;
3743
3744 // Exposed here for yielding support
3745 Mutex* const _bit_map_lock;
3746
3747 // The per thread work queues, available here for stealing
3748 OopTaskQueueSet* _task_queues;
3749
3750 // Termination (and yielding) support
3751 CMSConcMarkingTerminator _term;
3752 CMSConcMarkingTerminatorTerminator _term_term;
3753
3754 public:
3755 CMSConcMarkingTask(CMSCollector* collector,
3756 CompactibleFreeListSpace* cms_space,
3757 CompactibleFreeListSpace* perm_space,
3758 bool asynch,
3759 YieldingFlexibleWorkGang* workers,
3760 OopTaskQueueSet* task_queues):
3761 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3762 _collector(collector),
3763 _cms_space(cms_space),
3764 _perm_space(perm_space),
3765 _asynch(asynch), _n_workers(0), _result(true),
3766 _task_queues(task_queues),
3767 _term(_n_workers, task_queues, _collector),
3768 _bit_map_lock(collector->bitMapLock())
3769 {
3770 _requested_size = _n_workers;
3771 _term.set_task(this);
3772 _term_term.set_task(this);
3773 assert(_cms_space->bottom() < _perm_space->bottom(),
3774 "Finger incorrectly initialized below");
3775 _restart_addr = _global_finger = _cms_space->bottom();
3776 }
3777
3778
3779 OopTaskQueueSet* task_queues() { return _task_queues; }
3780
3781 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3782
3783 HeapWord** global_finger_addr() { return &_global_finger; }
3784
3785 CMSConcMarkingTerminator* terminator() { return &_term; }
3786
3787 virtual void set_for_termination(int active_workers) {
3788 terminator()->reset_for_reuse(active_workers);
3789 }
3790
3791 void work(int i);
3792 bool should_yield() {
3793 return ConcurrentMarkSweepThread::should_yield()
3794 && !_collector->foregroundGCIsActive()
3795 && _asynch;
3796 }
3797
3798 virtual void coordinator_yield(); // stuff done by coordinator
3799 bool result() { return _result; }
3800
3801 void reset(HeapWord* ra) {
3802 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3803 assert(_global_finger >= _perm_space->end(), "Postcondition of ::work(i)");
3804 assert(ra < _perm_space->end(), "ra too large");
3805 _restart_addr = _global_finger = ra;
3806 _term.reset_for_reuse();
3807 }
3808
3809 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3810 OopTaskQueue* work_q);
3811
3812 private:
3813 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3814 void do_work_steal(int i);
3815 void bump_global_finger(HeapWord* f);
3816 };
3817
3818 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3819 assert(_task != NULL, "Error");
3820 return _task->yielding();
3821 // Note that we do not need the disjunct || _task->should_yield() above
3822 // because we want terminating threads to yield only if the task
3823 // is already in the midst of yielding, which happens only after at least one
3824 // thread has yielded.
3825 }
3826
3827 void CMSConcMarkingTerminator::yield() {
3828 if (_task->should_yield()) {
3829 _task->yield();
3830 } else {
3831 ParallelTaskTerminator::yield();
3832 }
3833 }
3834
3835 ////////////////////////////////////////////////////////////////
3836 // Concurrent Marking Algorithm Sketch
3837 ////////////////////////////////////////////////////////////////
3838 // Until all tasks exhausted (both spaces):
3839 // -- claim next available chunk
3840 // -- bump global finger via CAS
3841 // -- find first object that starts in this chunk
3842 // and start scanning bitmap from that position
3843 // -- scan marked objects for oops
3844 // -- CAS-mark target, and if successful:
3845 // . if target oop is above global finger (volatile read)
3846 // nothing to do
3847 // . if target oop is in chunk and above local finger
3848 // then nothing to do
3849 // . else push on work-queue
3850 // -- Deal with possible overflow issues:
3851 // . local work-queue overflow causes stuff to be pushed on
3852 // global (common) overflow queue
3853 // . always first empty local work queue
3854 // . then get a batch of oops from global work queue if any
3855 // . then do work stealing
3856 // -- When all tasks claimed (both spaces)
3857 // and local work queue empty,
3858 // then in a loop do:
3859 // . check global overflow stack; steal a batch of oops and trace
3860 // . try to steal from other threads oif GOS is empty
3861 // . if neither is available, offer termination
3862 // -- Terminate and return result
3863 //
3864 void CMSConcMarkingTask::work(int i) {
3865 elapsedTimer _timer;
3866 ResourceMark rm;
3867 HandleMark hm;
3868
3869 DEBUG_ONLY(_collector->verify_overflow_empty();)
3870
3871 // Before we begin work, our work queue should be empty
3872 assert(work_queue(i)->size() == 0, "Expected to be empty");
3873 // Scan the bitmap covering _cms_space, tracing through grey objects.
3874 _timer.start();
3875 do_scan_and_mark(i, _cms_space);
3876 _timer.stop();
3877 if (PrintCMSStatistics != 0) {
3878 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3879 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3880 }
3881
3882 // ... do the same for the _perm_space
3883 _timer.reset();
3884 _timer.start();
3885 do_scan_and_mark(i, _perm_space);
3886 _timer.stop();
3887 if (PrintCMSStatistics != 0) {
3888 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec",
3889 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3890 }
3891
3892 // ... do work stealing
3893 _timer.reset();
3894 _timer.start();
3895 do_work_steal(i);
3896 _timer.stop();
3897 if (PrintCMSStatistics != 0) {
3898 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3899 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3900 }
3901 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3902 assert(work_queue(i)->size() == 0, "Should have been emptied");
3903 // Note that under the current task protocol, the
3904 // following assertion is true even of the spaces
3905 // expanded since the completion of the concurrent
3906 // marking. XXX This will likely change under a strict
3907 // ABORT semantics.
3908 assert(_global_finger > _cms_space->end() &&
3909 _global_finger >= _perm_space->end(),
3910 "All tasks have been completed");
3911 DEBUG_ONLY(_collector->verify_overflow_empty();)
3912 }
3913
3914 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3915 HeapWord* read = _global_finger;
3916 HeapWord* cur = read;
3917 while (f > read) {
3918 cur = read;
3919 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3920 if (cur == read) {
3921 // our cas succeeded
3922 assert(_global_finger >= f, "protocol consistency");
3923 break;
3924 }
3925 }
3926 }
3927
3928 // This is really inefficient, and should be redone by
3929 // using (not yet available) block-read and -write interfaces to the
3930 // stack and the work_queue. XXX FIX ME !!!
3931 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3932 OopTaskQueue* work_q) {
3933 // Fast lock-free check
3934 if (ovflw_stk->length() == 0) {
3935 return false;
3936 }
3937 assert(work_q->size() == 0, "Shouldn't steal");
3938 MutexLockerEx ml(ovflw_stk->par_lock(),
3939 Mutex::_no_safepoint_check_flag);
3940 // Grab up to 1/4 the size of the work queue
3941 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3942 (size_t)ParGCDesiredObjsFromOverflowList);
3943 num = MIN2(num, ovflw_stk->length());
3944 for (int i = (int) num; i > 0; i--) {
3945 oop cur = ovflw_stk->pop();
3946 assert(cur != NULL, "Counted wrong?");
3947 work_q->push(cur);
3948 }
3949 return num > 0;
3950 }
3951
3952 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3953 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3954 int n_tasks = pst->n_tasks();
3955 // We allow that there may be no tasks to do here because
3956 // we are restarting after a stack overflow.
3957 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3958 int nth_task = 0;
3959
3960 HeapWord* aligned_start = sp->bottom();
3961 if (sp->used_region().contains(_restart_addr)) {
3962 // Align down to a card boundary for the start of 0th task
3963 // for this space.
3964 aligned_start =
3965 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3966 CardTableModRefBS::card_size);
3967 }
3968
3969 size_t chunk_size = sp->marking_task_size();
3970 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3971 // Having claimed the nth task in this space,
3972 // compute the chunk that it corresponds to:
3973 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3974 aligned_start + (nth_task+1)*chunk_size);
3975 // Try and bump the global finger via a CAS;
3976 // note that we need to do the global finger bump
3977 // _before_ taking the intersection below, because
3978 // the task corresponding to that region will be
3979 // deemed done even if the used_region() expands
3980 // because of allocation -- as it almost certainly will
3981 // during start-up while the threads yield in the
3982 // closure below.
3983 HeapWord* finger = span.end();
3984 bump_global_finger(finger); // atomically
3985 // There are null tasks here corresponding to chunks
3986 // beyond the "top" address of the space.
3987 span = span.intersection(sp->used_region());
3988 if (!span.is_empty()) { // Non-null task
3989 HeapWord* prev_obj;
3990 assert(!span.contains(_restart_addr) || nth_task == 0,
3991 "Inconsistency");
3992 if (nth_task == 0) {
3993 // For the 0th task, we'll not need to compute a block_start.
3994 if (span.contains(_restart_addr)) {
3995 // In the case of a restart because of stack overflow,
3996 // we might additionally skip a chunk prefix.
3997 prev_obj = _restart_addr;
3998 } else {
3999 prev_obj = span.start();
4000 }
4001 } else {
4002 // We want to skip the first object because
4003 // the protocol is to scan any object in its entirety
4004 // that _starts_ in this span; a fortiori, any
4005 // object starting in an earlier span is scanned
4006 // as part of an earlier claimed task.
4007 // Below we use the "careful" version of block_start
4008 // so we do not try to navigate uninitialized objects.
4009 prev_obj = sp->block_start_careful(span.start());
4010 // Below we use a variant of block_size that uses the
4011 // Printezis bits to avoid waiting for allocated
4012 // objects to become initialized/parsable.
4013 while (prev_obj < span.start()) {
4014 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
4015 if (sz > 0) {
4016 prev_obj += sz;
4017 } else {
4018 // In this case we may end up doing a bit of redundant
4019 // scanning, but that appears unavoidable, short of
4020 // locking the free list locks; see bug 6324141.
4021 break;
4022 }
4023 }
4024 }
4025 if (prev_obj < span.end()) {
4026 MemRegion my_span = MemRegion(prev_obj, span.end());
4027 // Do the marking work within a non-empty span --
4028 // the last argument to the constructor indicates whether the
4029 // iteration should be incremental with periodic yields.
4030 Par_MarkFromRootsClosure cl(this, _collector, my_span,
4031 &_collector->_markBitMap,
4032 work_queue(i),
4033 &_collector->_markStack,
4034 &_collector->_revisitStack,
4035 _asynch);
4036 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
4037 } // else nothing to do for this task
4038 } // else nothing to do for this task
4039 }
4040 // We'd be tempted to assert here that since there are no
4041 // more tasks left to claim in this space, the global_finger
4042 // must exceed space->top() and a fortiori space->end(). However,
4043 // that would not quite be correct because the bumping of
4044 // global_finger occurs strictly after the claiming of a task,
4045 // so by the time we reach here the global finger may not yet
4046 // have been bumped up by the thread that claimed the last
4047 // task.
4048 pst->all_tasks_completed();
4049 }
4050
4051 class Par_ConcMarkingClosure: public Par_KlassRememberingOopClosure {
4052 private:
4053 CMSConcMarkingTask* _task;
4054 MemRegion _span;
4055 CMSBitMap* _bit_map;
4056 CMSMarkStack* _overflow_stack;
4057 OopTaskQueue* _work_queue;
4058 protected:
4059 DO_OOP_WORK_DEFN
4060 public:
4061 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
4062 CMSBitMap* bit_map, CMSMarkStack* overflow_stack,
4063 CMSMarkStack* revisit_stack):
4064 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
4065 _task(task),
4066 _span(collector->_span),
4067 _work_queue(work_queue),
4068 _bit_map(bit_map),
4069 _overflow_stack(overflow_stack)
4070 { }
4071 virtual void do_oop(oop* p);
4072 virtual void do_oop(narrowOop* p);
4073 void trim_queue(size_t max);
4074 void handle_stack_overflow(HeapWord* lost);
4075 void do_yield_check() {
4076 if (_task->should_yield()) {
4077 _task->yield();
4078 }
4079 }
4080 };
4081
4082 // Grey object scanning during work stealing phase --
4083 // the salient assumption here is that any references
4084 // that are in these stolen objects being scanned must
4085 // already have been initialized (else they would not have
4086 // been published), so we do not need to check for
4087 // uninitialized objects before pushing here.
4088 void Par_ConcMarkingClosure::do_oop(oop obj) {
4089 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
4090 HeapWord* addr = (HeapWord*)obj;
4091 // Check if oop points into the CMS generation
4092 // and is not marked
4093 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
4094 // a white object ...
4095 // If we manage to "claim" the object, by being the
4096 // first thread to mark it, then we push it on our
4097 // marking stack
4098 if (_bit_map->par_mark(addr)) { // ... now grey
4099 // push on work queue (grey set)
4100 bool simulate_overflow = false;
4101 NOT_PRODUCT(
4102 if (CMSMarkStackOverflowALot &&
4103 _collector->simulate_overflow()) {
4104 // simulate a stack overflow
4105 simulate_overflow = true;
4106 }
4107 )
4108 if (simulate_overflow ||
4109 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
4110 // stack overflow
4111 if (PrintCMSStatistics != 0) {
4112 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
4113 SIZE_FORMAT, _overflow_stack->capacity());
4114 }
4115 // We cannot assert that the overflow stack is full because
4116 // it may have been emptied since.
4117 assert(simulate_overflow ||
4118 _work_queue->size() == _work_queue->max_elems(),
4119 "Else push should have succeeded");
4120 handle_stack_overflow(addr);
4121 }
4122 } // Else, some other thread got there first
4123 do_yield_check();
4124 }
4125 }
4126
4127 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4128 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4129
4130 void Par_ConcMarkingClosure::trim_queue(size_t max) {
4131 while (_work_queue->size() > max) {
4132 oop new_oop;
4133 if (_work_queue->pop_local(new_oop)) {
4134 assert(new_oop->is_oop(), "Should be an oop");
4135 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
4136 assert(_span.contains((HeapWord*)new_oop), "Not in span");
4137 assert(new_oop->is_parsable(), "Should be parsable");
4138 new_oop->oop_iterate(this); // do_oop() above
4139 do_yield_check();
4140 }
4141 }
4142 }
4143
4144 // Upon stack overflow, we discard (part of) the stack,
4145 // remembering the least address amongst those discarded
4146 // in CMSCollector's _restart_address.
4147 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
4148 // We need to do this under a mutex to prevent other
4149 // workers from interfering with the work done below.
4150 MutexLockerEx ml(_overflow_stack->par_lock(),
4151 Mutex::_no_safepoint_check_flag);
4152 // Remember the least grey address discarded
4153 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
4154 _collector->lower_restart_addr(ra);
4155 _overflow_stack->reset(); // discard stack contents
4156 _overflow_stack->expand(); // expand the stack if possible
4157 }
4158
4159
4160 void CMSConcMarkingTask::do_work_steal(int i) {
4161 OopTaskQueue* work_q = work_queue(i);
4162 oop obj_to_scan;
4163 CMSBitMap* bm = &(_collector->_markBitMap);
4164 CMSMarkStack* ovflw = &(_collector->_markStack);
4165 CMSMarkStack* revisit = &(_collector->_revisitStack);
4166 int* seed = _collector->hash_seed(i);
4167 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw, revisit);
4168 while (true) {
4169 cl.trim_queue(0);
4170 assert(work_q->size() == 0, "Should have been emptied above");
4171 if (get_work_from_overflow_stack(ovflw, work_q)) {
4172 // Can't assert below because the work obtained from the
4173 // overflow stack may already have been stolen from us.
4174 // assert(work_q->size() > 0, "Work from overflow stack");
4175 continue;
4176 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4177 assert(obj_to_scan->is_oop(), "Should be an oop");
4178 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
4179 obj_to_scan->oop_iterate(&cl);
4180 } else if (terminator()->offer_termination(&_term_term)) {
4181 assert(work_q->size() == 0, "Impossible!");
4182 break;
4183 } else if (yielding() || should_yield()) {
4184 yield();
4185 }
4186 }
4187 }
4188
4189 // This is run by the CMS (coordinator) thread.
4190 void CMSConcMarkingTask::coordinator_yield() {
4191 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4192 "CMS thread should hold CMS token");
4193 DEBUG_ONLY(RememberKlassesChecker mux(false);)
4194 // First give up the locks, then yield, then re-lock
4195 // We should probably use a constructor/destructor idiom to
4196 // do this unlock/lock or modify the MutexUnlocker class to
4197 // serve our purpose. XXX
4198 assert_lock_strong(_bit_map_lock);
4199 _bit_map_lock->unlock();
4200 ConcurrentMarkSweepThread::desynchronize(true);
4201 ConcurrentMarkSweepThread::acknowledge_yield_request();
4202 _collector->stopTimer();
4203 if (PrintCMSStatistics != 0) {
4204 _collector->incrementYields();
4205 }
4206 _collector->icms_wait();
4207
4208 // It is possible for whichever thread initiated the yield request
4209 // not to get a chance to wake up and take the bitmap lock between
4210 // this thread releasing it and reacquiring it. So, while the
4211 // should_yield() flag is on, let's sleep for a bit to give the
4212 // other thread a chance to wake up. The limit imposed on the number
4213 // of iterations is defensive, to avoid any unforseen circumstances
4214 // putting us into an infinite loop. Since it's always been this
4215 // (coordinator_yield()) method that was observed to cause the
4216 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4217 // which is by default non-zero. For the other seven methods that
4218 // also perform the yield operation, as are using a different
4219 // parameter (CMSYieldSleepCount) which is by default zero. This way we
4220 // can enable the sleeping for those methods too, if necessary.
4221 // See 6442774.
4222 //
4223 // We really need to reconsider the synchronization between the GC
4224 // thread and the yield-requesting threads in the future and we
4225 // should really use wait/notify, which is the recommended
4226 // way of doing this type of interaction. Additionally, we should
4227 // consolidate the eight methods that do the yield operation and they
4228 // are almost identical into one for better maintenability and
4229 // readability. See 6445193.
4230 //
4231 // Tony 2006.06.29
4232 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4233 ConcurrentMarkSweepThread::should_yield() &&
4234 !CMSCollector::foregroundGCIsActive(); ++i) {
4235 os::sleep(Thread::current(), 1, false);
4236 ConcurrentMarkSweepThread::acknowledge_yield_request();
4237 }
4238
4239 ConcurrentMarkSweepThread::synchronize(true);
4240 _bit_map_lock->lock_without_safepoint_check();
4241 _collector->startTimer();
4242 }
4243
4244 bool CMSCollector::do_marking_mt(bool asynch) {
4245 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
4246 // In the future this would be determined ergonomically, based
4247 // on #cpu's, # active mutator threads (and load), and mutation rate.
4248 int num_workers = ConcGCThreads;
4249
4250 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4251 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
4252
4253 CMSConcMarkingTask tsk(this,
4254 cms_space,
4255 perm_space,
4256 asynch,
4257 conc_workers(),
4258 task_queues());
4259
4260 // Since the actual number of workers we get may be different
4261 // from the number we requested above, do we need to do anything different
4262 // below? In particular, may be we need to subclass the SequantialSubTasksDone
4263 // class?? XXX
4264 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4265 perm_space->initialize_sequential_subtasks_for_marking(num_workers);
4266
4267 // Refs discovery is already non-atomic.
4268 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4269 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
4270 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());)
4271 conc_workers()->start_task(&tsk);
4272 while (tsk.yielded()) {
4273 tsk.coordinator_yield();
4274 conc_workers()->continue_task(&tsk);
4275 }
4276 // If the task was aborted, _restart_addr will be non-NULL
4277 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4278 while (_restart_addr != NULL) {
4279 // XXX For now we do not make use of ABORTED state and have not
4280 // yet implemented the right abort semantics (even in the original
4281 // single-threaded CMS case). That needs some more investigation
4282 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4283 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4284 // If _restart_addr is non-NULL, a marking stack overflow
4285 // occurred; we need to do a fresh marking iteration from the
4286 // indicated restart address.
4287 if (_foregroundGCIsActive && asynch) {
4288 // We may be running into repeated stack overflows, having
4289 // reached the limit of the stack size, while making very
4290 // slow forward progress. It may be best to bail out and
4291 // let the foreground collector do its job.
4292 // Clear _restart_addr, so that foreground GC
4293 // works from scratch. This avoids the headache of
4294 // a "rescan" which would otherwise be needed because
4295 // of the dirty mod union table & card table.
4296 _restart_addr = NULL;
4297 return false;
4298 }
4299 // Adjust the task to restart from _restart_addr
4300 tsk.reset(_restart_addr);
4301 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4302 _restart_addr);
4303 perm_space->initialize_sequential_subtasks_for_marking(num_workers,
4304 _restart_addr);
4305 _restart_addr = NULL;
4306 // Get the workers going again
4307 conc_workers()->start_task(&tsk);
4308 while (tsk.yielded()) {
4309 tsk.coordinator_yield();
4310 conc_workers()->continue_task(&tsk);
4311 }
4312 }
4313 assert(tsk.completed(), "Inconsistency");
4314 assert(tsk.result() == true, "Inconsistency");
4315 return true;
4316 }
4317
4318 bool CMSCollector::do_marking_st(bool asynch) {
4319 ResourceMark rm;
4320 HandleMark hm;
4321
4322 // Temporarily make refs discovery single threaded (non-MT)
4323 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
4324 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4325 &_markStack, &_revisitStack, CMSYield && asynch);
4326 // the last argument to iterate indicates whether the iteration
4327 // should be incremental with periodic yields.
4328 _markBitMap.iterate(&markFromRootsClosure);
4329 // If _restart_addr is non-NULL, a marking stack overflow
4330 // occurred; we need to do a fresh iteration from the
4331 // indicated restart address.
4332 while (_restart_addr != NULL) {
4333 if (_foregroundGCIsActive && asynch) {
4334 // We may be running into repeated stack overflows, having
4335 // reached the limit of the stack size, while making very
4336 // slow forward progress. It may be best to bail out and
4337 // let the foreground collector do its job.
4338 // Clear _restart_addr, so that foreground GC
4339 // works from scratch. This avoids the headache of
4340 // a "rescan" which would otherwise be needed because
4341 // of the dirty mod union table & card table.
4342 _restart_addr = NULL;
4343 return false; // indicating failure to complete marking
4344 }
4345 // Deal with stack overflow:
4346 // we restart marking from _restart_addr
4347 HeapWord* ra = _restart_addr;
4348 markFromRootsClosure.reset(ra);
4349 _restart_addr = NULL;
4350 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4351 }
4352 return true;
4353 }
4354
4355 void CMSCollector::preclean() {
4356 check_correct_thread_executing();
4357 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4358 verify_work_stacks_empty();
4359 verify_overflow_empty();
4360 _abort_preclean = false;
4361 if (CMSPrecleaningEnabled) {
4362 _eden_chunk_index = 0;
4363 size_t used = get_eden_used();
4364 size_t capacity = get_eden_capacity();
4365 // Don't start sampling unless we will get sufficiently
4366 // many samples.
4367 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4368 * CMSScheduleRemarkEdenPenetration)) {
4369 _start_sampling = true;
4370 } else {
4371 _start_sampling = false;
4372 }
4373 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4374 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
4375 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4376 }
4377 CMSTokenSync x(true); // is cms thread
4378 if (CMSPrecleaningEnabled) {
4379 sample_eden();
4380 _collectorState = AbortablePreclean;
4381 } else {
4382 _collectorState = FinalMarking;
4383 }
4384 verify_work_stacks_empty();
4385 verify_overflow_empty();
4386 }
4387
4388 // Try and schedule the remark such that young gen
4389 // occupancy is CMSScheduleRemarkEdenPenetration %.
4390 void CMSCollector::abortable_preclean() {
4391 check_correct_thread_executing();
4392 assert(CMSPrecleaningEnabled, "Inconsistent control state");
4393 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4394
4395 // If Eden's current occupancy is below this threshold,
4396 // immediately schedule the remark; else preclean
4397 // past the next scavenge in an effort to
4398 // schedule the pause as described avove. By choosing
4399 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4400 // we will never do an actual abortable preclean cycle.
4401 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4402 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4403 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
4404 // We need more smarts in the abortable preclean
4405 // loop below to deal with cases where allocation
4406 // in young gen is very very slow, and our precleaning
4407 // is running a losing race against a horde of
4408 // mutators intent on flooding us with CMS updates
4409 // (dirty cards).
4410 // One, admittedly dumb, strategy is to give up
4411 // after a certain number of abortable precleaning loops
4412 // or after a certain maximum time. We want to make
4413 // this smarter in the next iteration.
4414 // XXX FIX ME!!! YSR
4415 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4416 while (!(should_abort_preclean() ||
4417 ConcurrentMarkSweepThread::should_terminate())) {
4418 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4419 cumworkdone += workdone;
4420 loops++;
4421 // Voluntarily terminate abortable preclean phase if we have
4422 // been at it for too long.
4423 if ((CMSMaxAbortablePrecleanLoops != 0) &&
4424 loops >= CMSMaxAbortablePrecleanLoops) {
4425 if (PrintGCDetails) {
4426 gclog_or_tty->print(" CMS: abort preclean due to loops ");
4427 }
4428 break;
4429 }
4430 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4431 if (PrintGCDetails) {
4432 gclog_or_tty->print(" CMS: abort preclean due to time ");
4433 }
4434 break;
4435 }
4436 // If we are doing little work each iteration, we should
4437 // take a short break.
4438 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4439 // Sleep for some time, waiting for work to accumulate
4440 stopTimer();
4441 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4442 startTimer();
4443 waited++;
4444 }
4445 }
4446 if (PrintCMSStatistics > 0) {
4447 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4448 loops, waited, cumworkdone);
4449 }
4450 }
4451 CMSTokenSync x(true); // is cms thread
4452 if (_collectorState != Idling) {
4453 assert(_collectorState == AbortablePreclean,
4454 "Spontaneous state transition?");
4455 _collectorState = FinalMarking;
4456 } // Else, a foreground collection completed this CMS cycle.
4457 return;
4458 }
4459
4460 // Respond to an Eden sampling opportunity
4461 void CMSCollector::sample_eden() {
4462 // Make sure a young gc cannot sneak in between our
4463 // reading and recording of a sample.
4464 assert(Thread::current()->is_ConcurrentGC_thread(),
4465 "Only the cms thread may collect Eden samples");
4466 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4467 "Should collect samples while holding CMS token");
4468 if (!_start_sampling) {
4469 return;
4470 }
4471 if (_eden_chunk_array) {
4472 if (_eden_chunk_index < _eden_chunk_capacity) {
4473 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
4474 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4475 "Unexpected state of Eden");
4476 // We'd like to check that what we just sampled is an oop-start address;
4477 // however, we cannot do that here since the object may not yet have been
4478 // initialized. So we'll instead do the check when we _use_ this sample
4479 // later.
4480 if (_eden_chunk_index == 0 ||
4481 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4482 _eden_chunk_array[_eden_chunk_index-1])
4483 >= CMSSamplingGrain)) {
4484 _eden_chunk_index++; // commit sample
4485 }
4486 }
4487 }
4488 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4489 size_t used = get_eden_used();
4490 size_t capacity = get_eden_capacity();
4491 assert(used <= capacity, "Unexpected state of Eden");
4492 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4493 _abort_preclean = true;
4494 }
4495 }
4496 }
4497
4498
4499 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4500 assert(_collectorState == Precleaning ||
4501 _collectorState == AbortablePreclean, "incorrect state");
4502 ResourceMark rm;
4503 HandleMark hm;
4504
4505 // Precleaning is currently not MT but the reference processor
4506 // may be set for MT. Disable it temporarily here.
4507 ReferenceProcessor* rp = ref_processor();
4508 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
4509
4510 // Do one pass of scrubbing the discovered reference lists
4511 // to remove any reference objects with strongly-reachable
4512 // referents.
4513 if (clean_refs) {
4514 CMSPrecleanRefsYieldClosure yield_cl(this);
4515 assert(rp->span().equals(_span), "Spans should be equal");
4516 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4517 &_markStack, &_revisitStack,
4518 true /* preclean */);
4519 CMSDrainMarkingStackClosure complete_trace(this,
4520 _span, &_markBitMap, &_markStack,
4521 &keep_alive, true /* preclean */);
4522
4523 // We don't want this step to interfere with a young
4524 // collection because we don't want to take CPU
4525 // or memory bandwidth away from the young GC threads
4526 // (which may be as many as there are CPUs).
4527 // Note that we don't need to protect ourselves from
4528 // interference with mutators because they can't
4529 // manipulate the discovered reference lists nor affect
4530 // the computed reachability of the referents, the
4531 // only properties manipulated by the precleaning
4532 // of these reference lists.
4533 stopTimer();
4534 CMSTokenSyncWithLocks x(true /* is cms thread */,
4535 bitMapLock());
4536 startTimer();
4537 sample_eden();
4538
4539 // The following will yield to allow foreground
4540 // collection to proceed promptly. XXX YSR:
4541 // The code in this method may need further
4542 // tweaking for better performance and some restructuring
4543 // for cleaner interfaces.
4544 rp->preclean_discovered_references(
4545 rp->is_alive_non_header(), &keep_alive, &complete_trace,
4546 &yield_cl, should_unload_classes());
4547 }
4548
4549 if (clean_survivor) { // preclean the active survivor space(s)
4550 assert(_young_gen->kind() == Generation::DefNew ||
4551 _young_gen->kind() == Generation::ParNew ||
4552 _young_gen->kind() == Generation::ASParNew,
4553 "incorrect type for cast");
4554 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4555 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4556 &_markBitMap, &_modUnionTable,
4557 &_markStack, &_revisitStack,
4558 true /* precleaning phase */);
4559 stopTimer();
4560 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4561 bitMapLock());
4562 startTimer();
4563 unsigned int before_count =
4564 GenCollectedHeap::heap()->total_collections();
4565 SurvivorSpacePrecleanClosure
4566 sss_cl(this, _span, &_markBitMap, &_markStack,
4567 &pam_cl, before_count, CMSYield);
4568 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4569 dng->from()->object_iterate_careful(&sss_cl);
4570 dng->to()->object_iterate_careful(&sss_cl);
4571 }
4572 MarkRefsIntoAndScanClosure
4573 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4574 &_markStack, &_revisitStack, this, CMSYield,
4575 true /* precleaning phase */);
4576 // CAUTION: The following closure has persistent state that may need to
4577 // be reset upon a decrease in the sequence of addresses it
4578 // processes.
4579 ScanMarkedObjectsAgainCarefullyClosure
4580 smoac_cl(this, _span,
4581 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield);
4582
4583 // Preclean dirty cards in ModUnionTable and CardTable using
4584 // appropriate convergence criterion;
4585 // repeat CMSPrecleanIter times unless we find that
4586 // we are losing.
4587 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4588 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4589 "Bad convergence multiplier");
4590 assert(CMSPrecleanThreshold >= 100,
4591 "Unreasonably low CMSPrecleanThreshold");
4592
4593 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4594 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4595 numIter < CMSPrecleanIter;
4596 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4597 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4598 if (CMSPermGenPrecleaningEnabled) {
4599 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl);
4600 }
4601 if (Verbose && PrintGCDetails) {
4602 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4603 }
4604 // Either there are very few dirty cards, so re-mark
4605 // pause will be small anyway, or our pre-cleaning isn't
4606 // that much faster than the rate at which cards are being
4607 // dirtied, so we might as well stop and re-mark since
4608 // precleaning won't improve our re-mark time by much.
4609 if (curNumCards <= CMSPrecleanThreshold ||
4610 (numIter > 0 &&
4611 (curNumCards * CMSPrecleanDenominator >
4612 lastNumCards * CMSPrecleanNumerator))) {
4613 numIter++;
4614 cumNumCards += curNumCards;
4615 break;
4616 }
4617 }
4618 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4619 if (CMSPermGenPrecleaningEnabled) {
4620 curNumCards += preclean_card_table(_permGen, &smoac_cl);
4621 }
4622 cumNumCards += curNumCards;
4623 if (PrintGCDetails && PrintCMSStatistics != 0) {
4624 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4625 curNumCards, cumNumCards, numIter);
4626 }
4627 return cumNumCards; // as a measure of useful work done
4628 }
4629
4630 // PRECLEANING NOTES:
4631 // Precleaning involves:
4632 // . reading the bits of the modUnionTable and clearing the set bits.
4633 // . For the cards corresponding to the set bits, we scan the
4634 // objects on those cards. This means we need the free_list_lock
4635 // so that we can safely iterate over the CMS space when scanning
4636 // for oops.
4637 // . When we scan the objects, we'll be both reading and setting
4638 // marks in the marking bit map, so we'll need the marking bit map.
4639 // . For protecting _collector_state transitions, we take the CGC_lock.
4640 // Note that any races in the reading of of card table entries by the
4641 // CMS thread on the one hand and the clearing of those entries by the
4642 // VM thread or the setting of those entries by the mutator threads on the
4643 // other are quite benign. However, for efficiency it makes sense to keep
4644 // the VM thread from racing with the CMS thread while the latter is
4645 // dirty card info to the modUnionTable. We therefore also use the
4646 // CGC_lock to protect the reading of the card table and the mod union
4647 // table by the CM thread.
4648 // . We run concurrently with mutator updates, so scanning
4649 // needs to be done carefully -- we should not try to scan
4650 // potentially uninitialized objects.
4651 //
4652 // Locking strategy: While holding the CGC_lock, we scan over and
4653 // reset a maximal dirty range of the mod union / card tables, then lock
4654 // the free_list_lock and bitmap lock to do a full marking, then
4655 // release these locks; and repeat the cycle. This allows for a
4656 // certain amount of fairness in the sharing of these locks between
4657 // the CMS collector on the one hand, and the VM thread and the
4658 // mutators on the other.
4659
4660 // NOTE: preclean_mod_union_table() and preclean_card_table()
4661 // further below are largely identical; if you need to modify
4662 // one of these methods, please check the other method too.
4663
4664 size_t CMSCollector::preclean_mod_union_table(
4665 ConcurrentMarkSweepGeneration* gen,
4666 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4667 verify_work_stacks_empty();
4668 verify_overflow_empty();
4669
4670 // Turn off checking for this method but turn it back on
4671 // selectively. There are yield points in this method
4672 // but it is difficult to turn the checking off just around
4673 // the yield points. It is simpler to selectively turn
4674 // it on.
4675 DEBUG_ONLY(RememberKlassesChecker mux(false);)
4676
4677 // strategy: starting with the first card, accumulate contiguous
4678 // ranges of dirty cards; clear these cards, then scan the region
4679 // covered by these cards.
4680
4681 // Since all of the MUT is committed ahead, we can just use
4682 // that, in case the generations expand while we are precleaning.
4683 // It might also be fine to just use the committed part of the
4684 // generation, but we might potentially miss cards when the
4685 // generation is rapidly expanding while we are in the midst
4686 // of precleaning.
4687 HeapWord* startAddr = gen->reserved().start();
4688 HeapWord* endAddr = gen->reserved().end();
4689
4690 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4691
4692 size_t numDirtyCards, cumNumDirtyCards;
4693 HeapWord *nextAddr, *lastAddr;
4694 for (cumNumDirtyCards = numDirtyCards = 0,
4695 nextAddr = lastAddr = startAddr;
4696 nextAddr < endAddr;
4697 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4698
4699 ResourceMark rm;
4700 HandleMark hm;
4701
4702 MemRegion dirtyRegion;
4703 {
4704 stopTimer();
4705 // Potential yield point
4706 CMSTokenSync ts(true);
4707 startTimer();
4708 sample_eden();
4709 // Get dirty region starting at nextOffset (inclusive),
4710 // simultaneously clearing it.
4711 dirtyRegion =
4712 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4713 assert(dirtyRegion.start() >= nextAddr,
4714 "returned region inconsistent?");
4715 }
4716 // Remember where the next search should begin.
4717 // The returned region (if non-empty) is a right open interval,
4718 // so lastOffset is obtained from the right end of that
4719 // interval.
4720 lastAddr = dirtyRegion.end();
4721 // Should do something more transparent and less hacky XXX
4722 numDirtyCards =
4723 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4724
4725 // We'll scan the cards in the dirty region (with periodic
4726 // yields for foreground GC as needed).
4727 if (!dirtyRegion.is_empty()) {
4728 assert(numDirtyCards > 0, "consistency check");
4729 HeapWord* stop_point = NULL;
4730 stopTimer();
4731 // Potential yield point
4732 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4733 bitMapLock());
4734 startTimer();
4735 {
4736 verify_work_stacks_empty();
4737 verify_overflow_empty();
4738 sample_eden();
4739 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4740 stop_point =
4741 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4742 }
4743 if (stop_point != NULL) {
4744 // The careful iteration stopped early either because it found an
4745 // uninitialized object, or because we were in the midst of an
4746 // "abortable preclean", which should now be aborted. Redirty
4747 // the bits corresponding to the partially-scanned or unscanned
4748 // cards. We'll either restart at the next block boundary or
4749 // abort the preclean.
4750 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) ||
4751 (_collectorState == AbortablePreclean && should_abort_preclean()),
4752 "Unparsable objects should only be in perm gen.");
4753 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4754 if (should_abort_preclean()) {
4755 break; // out of preclean loop
4756 } else {
4757 // Compute the next address at which preclean should pick up;
4758 // might need bitMapLock in order to read P-bits.
4759 lastAddr = next_card_start_after_block(stop_point);
4760 }
4761 }
4762 } else {
4763 assert(lastAddr == endAddr, "consistency check");
4764 assert(numDirtyCards == 0, "consistency check");
4765 break;
4766 }
4767 }
4768 verify_work_stacks_empty();
4769 verify_overflow_empty();
4770 return cumNumDirtyCards;
4771 }
4772
4773 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4774 // below are largely identical; if you need to modify
4775 // one of these methods, please check the other method too.
4776
4777 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4778 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4779 // strategy: it's similar to precleamModUnionTable above, in that
4780 // we accumulate contiguous ranges of dirty cards, mark these cards
4781 // precleaned, then scan the region covered by these cards.
4782 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4783 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4784
4785 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4786
4787 size_t numDirtyCards, cumNumDirtyCards;
4788 HeapWord *lastAddr, *nextAddr;
4789
4790 for (cumNumDirtyCards = numDirtyCards = 0,
4791 nextAddr = lastAddr = startAddr;
4792 nextAddr < endAddr;
4793 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4794
4795 ResourceMark rm;
4796 HandleMark hm;
4797
4798 MemRegion dirtyRegion;
4799 {
4800 // See comments in "Precleaning notes" above on why we
4801 // do this locking. XXX Could the locking overheads be
4802 // too high when dirty cards are sparse? [I don't think so.]
4803 stopTimer();
4804 CMSTokenSync x(true); // is cms thread
4805 startTimer();
4806 sample_eden();
4807 // Get and clear dirty region from card table
4808 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4809 MemRegion(nextAddr, endAddr),
4810 true,
4811 CardTableModRefBS::precleaned_card_val());
4812
4813 assert(dirtyRegion.start() >= nextAddr,
4814 "returned region inconsistent?");
4815 }
4816 lastAddr = dirtyRegion.end();
4817 numDirtyCards =
4818 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4819
4820 if (!dirtyRegion.is_empty()) {
4821 stopTimer();
4822 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4823 startTimer();
4824 sample_eden();
4825 verify_work_stacks_empty();
4826 verify_overflow_empty();
4827 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4828 HeapWord* stop_point =
4829 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4830 if (stop_point != NULL) {
4831 // The careful iteration stopped early because it found an
4832 // uninitialized object. Redirty the bits corresponding to the
4833 // partially-scanned or unscanned cards, and start again at the
4834 // next block boundary.
4835 assert(CMSPermGenPrecleaningEnabled ||
4836 (_collectorState == AbortablePreclean && should_abort_preclean()),
4837 "Unparsable objects should only be in perm gen.");
4838 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4839 if (should_abort_preclean()) {
4840 break; // out of preclean loop
4841 } else {
4842 // Compute the next address at which preclean should pick up.
4843 lastAddr = next_card_start_after_block(stop_point);
4844 }
4845 }
4846 } else {
4847 break;
4848 }
4849 }
4850 verify_work_stacks_empty();
4851 verify_overflow_empty();
4852 return cumNumDirtyCards;
4853 }
4854
4855 void CMSCollector::checkpointRootsFinal(bool asynch,
4856 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4857 assert(_collectorState == FinalMarking, "incorrect state transition?");
4858 check_correct_thread_executing();
4859 // world is stopped at this checkpoint
4860 assert(SafepointSynchronize::is_at_safepoint(),
4861 "world should be stopped");
4862 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4863
4864 verify_work_stacks_empty();
4865 verify_overflow_empty();
4866
4867 SpecializationStats::clear();
4868 if (PrintGCDetails) {
4869 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4870 _young_gen->used() / K,
4871 _young_gen->capacity() / K);
4872 }
4873 if (asynch) {
4874 if (CMSScavengeBeforeRemark) {
4875 GenCollectedHeap* gch = GenCollectedHeap::heap();
4876 // Temporarily set flag to false, GCH->do_collection will
4877 // expect it to be false and set to true
4878 FlagSetting fl(gch->_is_gc_active, false);
4879 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark",
4880 PrintGCDetails && Verbose, true, gclog_or_tty);)
4881 int level = _cmsGen->level() - 1;
4882 if (level >= 0) {
4883 gch->do_collection(true, // full (i.e. force, see below)
4884 false, // !clear_all_soft_refs
4885 0, // size
4886 false, // is_tlab
4887 level // max_level
4888 );
4889 }
4890 }
4891 FreelistLocker x(this);
4892 MutexLockerEx y(bitMapLock(),
4893 Mutex::_no_safepoint_check_flag);
4894 assert(!init_mark_was_synchronous, "but that's impossible!");
4895 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
4896 } else {
4897 // already have all the locks
4898 checkpointRootsFinalWork(asynch, clear_all_soft_refs,
4899 init_mark_was_synchronous);
4900 }
4901 verify_work_stacks_empty();
4902 verify_overflow_empty();
4903 SpecializationStats::print();
4904 }
4905
4906 void CMSCollector::checkpointRootsFinalWork(bool asynch,
4907 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4908
4909 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);)
4910
4911 assert(haveFreelistLocks(), "must have free list locks");
4912 assert_lock_strong(bitMapLock());
4913
4914 if (UseAdaptiveSizePolicy) {
4915 size_policy()->checkpoint_roots_final_begin();
4916 }
4917
4918 ResourceMark rm;
4919 HandleMark hm;
4920
4921 GenCollectedHeap* gch = GenCollectedHeap::heap();
4922
4923 if (should_unload_classes()) {
4924 CodeCache::gc_prologue();
4925 }
4926 assert(haveFreelistLocks(), "must have free list locks");
4927 assert_lock_strong(bitMapLock());
4928
4929 DEBUG_ONLY(RememberKlassesChecker fmx(should_unload_classes());)
4930 if (!init_mark_was_synchronous) {
4931 // We might assume that we need not fill TLAB's when
4932 // CMSScavengeBeforeRemark is set, because we may have just done
4933 // a scavenge which would have filled all TLAB's -- and besides
4934 // Eden would be empty. This however may not always be the case --
4935 // for instance although we asked for a scavenge, it may not have
4936 // happened because of a JNI critical section. We probably need
4937 // a policy for deciding whether we can in that case wait until
4938 // the critical section releases and then do the remark following
4939 // the scavenge, and skip it here. In the absence of that policy,
4940 // or of an indication of whether the scavenge did indeed occur,
4941 // we cannot rely on TLAB's having been filled and must do
4942 // so here just in case a scavenge did not happen.
4943 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4944 // Update the saved marks which may affect the root scans.
4945 gch->save_marks();
4946
4947 {
4948 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4949
4950 // Note on the role of the mod union table:
4951 // Since the marker in "markFromRoots" marks concurrently with
4952 // mutators, it is possible for some reachable objects not to have been
4953 // scanned. For instance, an only reference to an object A was
4954 // placed in object B after the marker scanned B. Unless B is rescanned,
4955 // A would be collected. Such updates to references in marked objects
4956 // are detected via the mod union table which is the set of all cards
4957 // dirtied since the first checkpoint in this GC cycle and prior to
4958 // the most recent young generation GC, minus those cleaned up by the
4959 // concurrent precleaning.
4960 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
4961 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty);
4962 do_remark_parallel();
4963 } else {
4964 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4965 gclog_or_tty);
4966 do_remark_non_parallel();
4967 }
4968 }
4969 } else {
4970 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
4971 // The initial mark was stop-world, so there's no rescanning to
4972 // do; go straight on to the next step below.
4973 }
4974 verify_work_stacks_empty();
4975 verify_overflow_empty();
4976
4977 {
4978 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);)
4979 refProcessingWork(asynch, clear_all_soft_refs);
4980 }
4981 verify_work_stacks_empty();
4982 verify_overflow_empty();
4983
4984 if (should_unload_classes()) {
4985 CodeCache::gc_epilogue();
4986 }
4987 JvmtiExport::gc_epilogue();
4988
4989 // If we encountered any (marking stack / work queue) overflow
4990 // events during the current CMS cycle, take appropriate
4991 // remedial measures, where possible, so as to try and avoid
4992 // recurrence of that condition.
4993 assert(_markStack.isEmpty(), "No grey objects");
4994 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4995 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4996 if (ser_ovflw > 0) {
4997 if (PrintCMSStatistics != 0) {
4998 gclog_or_tty->print_cr("Marking stack overflow (benign) "
4999 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
5000 ", kac_preclean="SIZE_FORMAT")",
5001 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
5002 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
5003 }
5004 _markStack.expand();
5005 _ser_pmc_remark_ovflw = 0;
5006 _ser_pmc_preclean_ovflw = 0;
5007 _ser_kac_preclean_ovflw = 0;
5008 _ser_kac_ovflw = 0;
5009 }
5010 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
5011 if (PrintCMSStatistics != 0) {
5012 gclog_or_tty->print_cr("Work queue overflow (benign) "
5013 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
5014 _par_pmc_remark_ovflw, _par_kac_ovflw);
5015 }
5016 _par_pmc_remark_ovflw = 0;
5017 _par_kac_ovflw = 0;
5018 }
5019 if (PrintCMSStatistics != 0) {
5020 if (_markStack._hit_limit > 0) {
5021 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
5022 _markStack._hit_limit);
5023 }
5024 if (_markStack._failed_double > 0) {
5025 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
5026 " current capacity "SIZE_FORMAT,
5027 _markStack._failed_double,
5028 _markStack.capacity());
5029 }
5030 }
5031 _markStack._hit_limit = 0;
5032 _markStack._failed_double = 0;
5033
5034 // Check that all the klasses have been checked
5035 assert(_revisitStack.isEmpty(), "Not all klasses revisited");
5036
5037 if ((VerifyAfterGC || VerifyDuringGC) &&
5038 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5039 verify_after_remark();
5040 }
5041
5042 // Change under the freelistLocks.
5043 _collectorState = Sweeping;
5044 // Call isAllClear() under bitMapLock
5045 assert(_modUnionTable.isAllClear(), "Should be clear by end of the"
5046 " final marking");
5047 if (UseAdaptiveSizePolicy) {
5048 size_policy()->checkpoint_roots_final_end(gch->gc_cause());
5049 }
5050 }
5051
5052 // Parallel remark task
5053 class CMSParRemarkTask: public AbstractGangTask {
5054 CMSCollector* _collector;
5055 int _n_workers;
5056 CompactibleFreeListSpace* _cms_space;
5057 CompactibleFreeListSpace* _perm_space;
5058
5059 // The per-thread work queues, available here for stealing.
5060 OopTaskQueueSet* _task_queues;
5061 ParallelTaskTerminator _term;
5062
5063 public:
5064 CMSParRemarkTask(CMSCollector* collector,
5065 CompactibleFreeListSpace* cms_space,
5066 CompactibleFreeListSpace* perm_space,
5067 int n_workers, FlexibleWorkGang* workers,
5068 OopTaskQueueSet* task_queues):
5069 AbstractGangTask("Rescan roots and grey objects in parallel"),
5070 _collector(collector),
5071 _cms_space(cms_space), _perm_space(perm_space),
5072 _n_workers(n_workers),
5073 _task_queues(task_queues),
5074 _term(n_workers, task_queues) { }
5075
5076 OopTaskQueueSet* task_queues() { return _task_queues; }
5077
5078 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5079
5080 ParallelTaskTerminator* terminator() { return &_term; }
5081 int n_workers() { return _n_workers; }
5082
5083 void work(int i);
5084
5085 private:
5086 // Work method in support of parallel rescan ... of young gen spaces
5087 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl,
5088 ContiguousSpace* space,
5089 HeapWord** chunk_array, size_t chunk_top);
5090
5091 // ... of dirty cards in old space
5092 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
5093 Par_MarkRefsIntoAndScanClosure* cl);
5094
5095 // ... work stealing for the above
5096 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
5097 };
5098
5099 // work_queue(i) is passed to the closure
5100 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
5101 // also is passed to do_dirty_card_rescan_tasks() and to
5102 // do_work_steal() to select the i-th task_queue.
5103
5104 void CMSParRemarkTask::work(int i) {
5105 elapsedTimer _timer;
5106 ResourceMark rm;
5107 HandleMark hm;
5108
5109 // ---------- rescan from roots --------------
5110 _timer.start();
5111 GenCollectedHeap* gch = GenCollectedHeap::heap();
5112 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
5113 _collector->_span, _collector->ref_processor(),
5114 &(_collector->_markBitMap),
5115 work_queue(i), &(_collector->_revisitStack));
5116
5117 // Rescan young gen roots first since these are likely
5118 // coarsely partitioned and may, on that account, constitute
5119 // the critical path; thus, it's best to start off that
5120 // work first.
5121 // ---------- young gen roots --------------
5122 {
5123 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
5124 EdenSpace* eden_space = dng->eden();
5125 ContiguousSpace* from_space = dng->from();
5126 ContiguousSpace* to_space = dng->to();
5127
5128 HeapWord** eca = _collector->_eden_chunk_array;
5129 size_t ect = _collector->_eden_chunk_index;
5130 HeapWord** sca = _collector->_survivor_chunk_array;
5131 size_t sct = _collector->_survivor_chunk_index;
5132
5133 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
5134 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
5135
5136 do_young_space_rescan(i, &par_mrias_cl, to_space, NULL, 0);
5137 do_young_space_rescan(i, &par_mrias_cl, from_space, sca, sct);
5138 do_young_space_rescan(i, &par_mrias_cl, eden_space, eca, ect);
5139
5140 _timer.stop();
5141 if (PrintCMSStatistics != 0) {
5142 gclog_or_tty->print_cr(
5143 "Finished young gen rescan work in %dth thread: %3.3f sec",
5144 i, _timer.seconds());
5145 }
5146 }
5147
5148 // ---------- remaining roots --------------
5149 _timer.reset();
5150 _timer.start();
5151 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
5152 false, // yg was scanned above
5153 false, // this is parallel code
5154 true, // collecting perm gen
5155 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5156 &par_mrias_cl,
5157 true, // walk all of code cache if (so & SO_CodeCache)
5158 NULL);
5159 assert(_collector->should_unload_classes()
5160 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache),
5161 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5162 _timer.stop();
5163 if (PrintCMSStatistics != 0) {
5164 gclog_or_tty->print_cr(
5165 "Finished remaining root rescan work in %dth thread: %3.3f sec",
5166 i, _timer.seconds());
5167 }
5168
5169 // ---------- rescan dirty cards ------------
5170 _timer.reset();
5171 _timer.start();
5172
5173 // Do the rescan tasks for each of the two spaces
5174 // (cms_space and perm_space) in turn.
5175 // "i" is passed to select the "i-th" task_queue
5176 do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl);
5177 do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl);
5178 _timer.stop();
5179 if (PrintCMSStatistics != 0) {
5180 gclog_or_tty->print_cr(
5181 "Finished dirty card rescan work in %dth thread: %3.3f sec",
5182 i, _timer.seconds());
5183 }
5184
5185 // ---------- steal work from other threads ...
5186 // ---------- ... and drain overflow list.
5187 _timer.reset();
5188 _timer.start();
5189 do_work_steal(i, &par_mrias_cl, _collector->hash_seed(i));
5190 _timer.stop();
5191 if (PrintCMSStatistics != 0) {
5192 gclog_or_tty->print_cr(
5193 "Finished work stealing in %dth thread: %3.3f sec",
5194 i, _timer.seconds());
5195 }
5196 }
5197
5198 // Note that parameter "i" is not used.
5199 void
5200 CMSParRemarkTask::do_young_space_rescan(int i,
5201 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space,
5202 HeapWord** chunk_array, size_t chunk_top) {
5203 // Until all tasks completed:
5204 // . claim an unclaimed task
5205 // . compute region boundaries corresponding to task claimed
5206 // using chunk_array
5207 // . par_oop_iterate(cl) over that region
5208
5209 ResourceMark rm;
5210 HandleMark hm;
5211
5212 SequentialSubTasksDone* pst = space->par_seq_tasks();
5213 assert(pst->valid(), "Uninitialized use?");
5214
5215 int nth_task = 0;
5216 int n_tasks = pst->n_tasks();
5217
5218 HeapWord *start, *end;
5219 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5220 // We claimed task # nth_task; compute its boundaries.
5221 if (chunk_top == 0) { // no samples were taken
5222 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
5223 start = space->bottom();
5224 end = space->top();
5225 } else if (nth_task == 0) {
5226 start = space->bottom();
5227 end = chunk_array[nth_task];
5228 } else if (nth_task < (jint)chunk_top) {
5229 assert(nth_task >= 1, "Control point invariant");
5230 start = chunk_array[nth_task - 1];
5231 end = chunk_array[nth_task];
5232 } else {
5233 assert(nth_task == (jint)chunk_top, "Control point invariant");
5234 start = chunk_array[chunk_top - 1];
5235 end = space->top();
5236 }
5237 MemRegion mr(start, end);
5238 // Verify that mr is in space
5239 assert(mr.is_empty() || space->used_region().contains(mr),
5240 "Should be in space");
5241 // Verify that "start" is an object boundary
5242 assert(mr.is_empty() || oop(mr.start())->is_oop(),
5243 "Should be an oop");
5244 space->par_oop_iterate(mr, cl);
5245 }
5246 pst->all_tasks_completed();
5247 }
5248
5249 void
5250 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5251 CompactibleFreeListSpace* sp, int i,
5252 Par_MarkRefsIntoAndScanClosure* cl) {
5253 // Until all tasks completed:
5254 // . claim an unclaimed task
5255 // . compute region boundaries corresponding to task claimed
5256 // . transfer dirty bits ct->mut for that region
5257 // . apply rescanclosure to dirty mut bits for that region
5258
5259 ResourceMark rm;
5260 HandleMark hm;
5261
5262 OopTaskQueue* work_q = work_queue(i);
5263 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5264 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5265 // CAUTION: This closure has state that persists across calls to
5266 // the work method dirty_range_iterate_clear() in that it has
5267 // imbedded in it a (subtype of) UpwardsObjectClosure. The
5268 // use of that state in the imbedded UpwardsObjectClosure instance
5269 // assumes that the cards are always iterated (even if in parallel
5270 // by several threads) in monotonically increasing order per each
5271 // thread. This is true of the implementation below which picks
5272 // card ranges (chunks) in monotonically increasing order globally
5273 // and, a-fortiori, in monotonically increasing order per thread
5274 // (the latter order being a subsequence of the former).
5275 // If the work code below is ever reorganized into a more chaotic
5276 // work-partitioning form than the current "sequential tasks"
5277 // paradigm, the use of that persistent state will have to be
5278 // revisited and modified appropriately. See also related
5279 // bug 4756801 work on which should examine this code to make
5280 // sure that the changes there do not run counter to the
5281 // assumptions made here and necessary for correctness and
5282 // efficiency. Note also that this code might yield inefficient
5283 // behaviour in the case of very large objects that span one or
5284 // more work chunks. Such objects would potentially be scanned
5285 // several times redundantly. Work on 4756801 should try and
5286 // address that performance anomaly if at all possible. XXX
5287 MemRegion full_span = _collector->_span;
5288 CMSBitMap* bm = &(_collector->_markBitMap); // shared
5289 CMSMarkStack* rs = &(_collector->_revisitStack); // shared
5290 MarkFromDirtyCardsClosure
5291 greyRescanClosure(_collector, full_span, // entire span of interest
5292 sp, bm, work_q, rs, cl);
5293
5294 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5295 assert(pst->valid(), "Uninitialized use?");
5296 int nth_task = 0;
5297 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5298 MemRegion span = sp->used_region();
5299 HeapWord* start_addr = span.start();
5300 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5301 alignment);
5302 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5303 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5304 start_addr, "Check alignment");
5305 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5306 chunk_size, "Check alignment");
5307
5308 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5309 // Having claimed the nth_task, compute corresponding mem-region,
5310 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
5311 // The alignment restriction ensures that we do not need any
5312 // synchronization with other gang-workers while setting or
5313 // clearing bits in thus chunk of the MUT.
5314 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5315 start_addr + (nth_task+1)*chunk_size);
5316 // The last chunk's end might be way beyond end of the
5317 // used region. In that case pull back appropriately.
5318 if (this_span.end() > end_addr) {
5319 this_span.set_end(end_addr);
5320 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5321 }
5322 // Iterate over the dirty cards covering this chunk, marking them
5323 // precleaned, and setting the corresponding bits in the mod union
5324 // table. Since we have been careful to partition at Card and MUT-word
5325 // boundaries no synchronization is needed between parallel threads.
5326 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5327 &modUnionClosure);
5328
5329 // Having transferred these marks into the modUnionTable,
5330 // rescan the marked objects on the dirty cards in the modUnionTable.
5331 // Even if this is at a synchronous collection, the initial marking
5332 // may have been done during an asynchronous collection so there
5333 // may be dirty bits in the mod-union table.
5334 _collector->_modUnionTable.dirty_range_iterate_clear(
5335 this_span, &greyRescanClosure);
5336 _collector->_modUnionTable.verifyNoOneBitsInRange(
5337 this_span.start(),
5338 this_span.end());
5339 }
5340 pst->all_tasks_completed(); // declare that i am done
5341 }
5342
5343 // . see if we can share work_queues with ParNew? XXX
5344 void
5345 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5346 int* seed) {
5347 OopTaskQueue* work_q = work_queue(i);
5348 NOT_PRODUCT(int num_steals = 0;)
5349 oop obj_to_scan;
5350 CMSBitMap* bm = &(_collector->_markBitMap);
5351
5352 while (true) {
5353 // Completely finish any left over work from (an) earlier round(s)
5354 cl->trim_queue(0);
5355 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5356 (size_t)ParGCDesiredObjsFromOverflowList);
5357 // Now check if there's any work in the overflow list
5358 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5359 // only affects the number of attempts made to get work from the
5360 // overflow list and does not affect the number of workers. Just
5361 // pass ParallelGCThreads so this behavior is unchanged.
5362 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5363 work_q,
5364 ParallelGCThreads)) {
5365 // found something in global overflow list;
5366 // not yet ready to go stealing work from others.
5367 // We'd like to assert(work_q->size() != 0, ...)
5368 // because we just took work from the overflow list,
5369 // but of course we can't since all of that could have
5370 // been already stolen from us.
5371 // "He giveth and He taketh away."
5372 continue;
5373 }
5374 // Verify that we have no work before we resort to stealing
5375 assert(work_q->size() == 0, "Have work, shouldn't steal");
5376 // Try to steal from other queues that have work
5377 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5378 NOT_PRODUCT(num_steals++;)
5379 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5380 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5381 // Do scanning work
5382 obj_to_scan->oop_iterate(cl);
5383 // Loop around, finish this work, and try to steal some more
5384 } else if (terminator()->offer_termination()) {
5385 break; // nirvana from the infinite cycle
5386 }
5387 }
5388 NOT_PRODUCT(
5389 if (PrintCMSStatistics != 0) {
5390 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5391 }
5392 )
5393 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5394 "Else our work is not yet done");
5395 }
5396
5397 // Return a thread-local PLAB recording array, as appropriate.
5398 void* CMSCollector::get_data_recorder(int thr_num) {
5399 if (_survivor_plab_array != NULL &&
5400 (CMSPLABRecordAlways ||
5401 (_collectorState > Marking && _collectorState < FinalMarking))) {
5402 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5403 ChunkArray* ca = &_survivor_plab_array[thr_num];
5404 ca->reset(); // clear it so that fresh data is recorded
5405 return (void*) ca;
5406 } else {
5407 return NULL;
5408 }
5409 }
5410
5411 // Reset all the thread-local PLAB recording arrays
5412 void CMSCollector::reset_survivor_plab_arrays() {
5413 for (uint i = 0; i < ParallelGCThreads; i++) {
5414 _survivor_plab_array[i].reset();
5415 }
5416 }
5417
5418 // Merge the per-thread plab arrays into the global survivor chunk
5419 // array which will provide the partitioning of the survivor space
5420 // for CMS rescan.
5421 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
5422 int no_of_gc_threads) {
5423 assert(_survivor_plab_array != NULL, "Error");
5424 assert(_survivor_chunk_array != NULL, "Error");
5425 assert(_collectorState == FinalMarking, "Error");
5426 for (int j = 0; j < no_of_gc_threads; j++) {
5427 _cursor[j] = 0;
5428 }
5429 HeapWord* top = surv->top();
5430 size_t i;
5431 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
5432 HeapWord* min_val = top; // Higher than any PLAB address
5433 uint min_tid = 0; // position of min_val this round
5434 for (int j = 0; j < no_of_gc_threads; j++) {
5435 ChunkArray* cur_sca = &_survivor_plab_array[j];
5436 if (_cursor[j] == cur_sca->end()) {
5437 continue;
5438 }
5439 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5440 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5441 assert(surv->used_region().contains(cur_val), "Out of bounds value");
5442 if (cur_val < min_val) {
5443 min_tid = j;
5444 min_val = cur_val;
5445 } else {
5446 assert(cur_val < top, "All recorded addresses should be less");
5447 }
5448 }
5449 // At this point min_val and min_tid are respectively
5450 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5451 // and the thread (j) that witnesses that address.
5452 // We record this address in the _survivor_chunk_array[i]
5453 // and increment _cursor[min_tid] prior to the next round i.
5454 if (min_val == top) {
5455 break;
5456 }
5457 _survivor_chunk_array[i] = min_val;
5458 _cursor[min_tid]++;
5459 }
5460 // We are all done; record the size of the _survivor_chunk_array
5461 _survivor_chunk_index = i; // exclusive: [0, i)
5462 if (PrintCMSStatistics > 0) {
5463 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5464 }
5465 // Verify that we used up all the recorded entries
5466 #ifdef ASSERT
5467 size_t total = 0;
5468 for (int j = 0; j < no_of_gc_threads; j++) {
5469 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5470 total += _cursor[j];
5471 }
5472 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5473 // Check that the merged array is in sorted order
5474 if (total > 0) {
5475 for (size_t i = 0; i < total - 1; i++) {
5476 if (PrintCMSStatistics > 0) {
5477 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5478 i, _survivor_chunk_array[i]);
5479 }
5480 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5481 "Not sorted");
5482 }
5483 }
5484 #endif // ASSERT
5485 }
5486
5487 // Set up the space's par_seq_tasks structure for work claiming
5488 // for parallel rescan of young gen.
5489 // See ParRescanTask where this is currently used.
5490 void
5491 CMSCollector::
5492 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5493 assert(n_threads > 0, "Unexpected n_threads argument");
5494 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5495
5496 // Eden space
5497 {
5498 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5499 assert(!pst->valid(), "Clobbering existing data?");
5500 // Each valid entry in [0, _eden_chunk_index) represents a task.
5501 size_t n_tasks = _eden_chunk_index + 1;
5502 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5503 // Sets the condition for completion of the subtask (how many threads
5504 // need to finish in order to be done).
5505 pst->set_n_threads(n_threads);
5506 pst->set_n_tasks((int)n_tasks);
5507 }
5508
5509 // Merge the survivor plab arrays into _survivor_chunk_array
5510 if (_survivor_plab_array != NULL) {
5511 merge_survivor_plab_arrays(dng->from(), n_threads);
5512 } else {
5513 assert(_survivor_chunk_index == 0, "Error");
5514 }
5515
5516 // To space
5517 {
5518 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5519 assert(!pst->valid(), "Clobbering existing data?");
5520 // Sets the condition for completion of the subtask (how many threads
5521 // need to finish in order to be done).
5522 pst->set_n_threads(n_threads);
5523 pst->set_n_tasks(1);
5524 assert(pst->valid(), "Error");
5525 }
5526
5527 // From space
5528 {
5529 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5530 assert(!pst->valid(), "Clobbering existing data?");
5531 size_t n_tasks = _survivor_chunk_index + 1;
5532 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5533 // Sets the condition for completion of the subtask (how many threads
5534 // need to finish in order to be done).
5535 pst->set_n_threads(n_threads);
5536 pst->set_n_tasks((int)n_tasks);
5537 assert(pst->valid(), "Error");
5538 }
5539 }
5540
5541 // Parallel version of remark
5542 void CMSCollector::do_remark_parallel() {
5543 GenCollectedHeap* gch = GenCollectedHeap::heap();
5544 FlexibleWorkGang* workers = gch->workers();
5545 assert(workers != NULL, "Need parallel worker threads.");
5546 int n_workers = workers->total_workers();
5547 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5548 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
5549
5550 CMSParRemarkTask tsk(this,
5551 cms_space, perm_space,
5552 n_workers, workers, task_queues());
5553
5554 // Set up for parallel process_strong_roots work.
5555 gch->set_par_threads(n_workers);
5556 // We won't be iterating over the cards in the card table updating
5557 // the younger_gen cards, so we shouldn't call the following else
5558 // the verification code as well as subsequent younger_refs_iterate
5559 // code would get confused. XXX
5560 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5561
5562 // The young gen rescan work will not be done as part of
5563 // process_strong_roots (which currently doesn't knw how to
5564 // parallelize such a scan), but rather will be broken up into
5565 // a set of parallel tasks (via the sampling that the [abortable]
5566 // preclean phase did of EdenSpace, plus the [two] tasks of
5567 // scanning the [two] survivor spaces. Further fine-grain
5568 // parallelization of the scanning of the survivor spaces
5569 // themselves, and of precleaning of the younger gen itself
5570 // is deferred to the future.
5571 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5572
5573 // The dirty card rescan work is broken up into a "sequence"
5574 // of parallel tasks (per constituent space) that are dynamically
5575 // claimed by the parallel threads.
5576 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5577 perm_space->initialize_sequential_subtasks_for_rescan(n_workers);
5578
5579 // It turns out that even when we're using 1 thread, doing the work in a
5580 // separate thread causes wide variance in run times. We can't help this
5581 // in the multi-threaded case, but we special-case n=1 here to get
5582 // repeatable measurements of the 1-thread overhead of the parallel code.
5583 if (n_workers > 1) {
5584 // Make refs discovery MT-safe, if it isn't already: it may not
5585 // necessarily be so, since it's possible that we are doing
5586 // ST marking.
5587 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5588 GenCollectedHeap::StrongRootsScope srs(gch);
5589 workers->run_task(&tsk);
5590 } else {
5591 GenCollectedHeap::StrongRootsScope srs(gch);
5592 tsk.work(0);
5593 }
5594 gch->set_par_threads(0); // 0 ==> non-parallel.
5595 // restore, single-threaded for now, any preserved marks
5596 // as a result of work_q overflow
5597 restore_preserved_marks_if_any();
5598 }
5599
5600 // Non-parallel version of remark
5601 void CMSCollector::do_remark_non_parallel() {
5602 ResourceMark rm;
5603 HandleMark hm;
5604 GenCollectedHeap* gch = GenCollectedHeap::heap();
5605 MarkRefsIntoAndScanClosure
5606 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
5607 &_markStack, &_revisitStack, this,
5608 false /* should_yield */, false /* not precleaning */);
5609 MarkFromDirtyCardsClosure
5610 markFromDirtyCardsClosure(this, _span,
5611 NULL, // space is set further below
5612 &_markBitMap, &_markStack, &_revisitStack,
5613 &mrias_cl);
5614 {
5615 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty);
5616 // Iterate over the dirty cards, setting the corresponding bits in the
5617 // mod union table.
5618 {
5619 ModUnionClosure modUnionClosure(&_modUnionTable);
5620 _ct->ct_bs()->dirty_card_iterate(
5621 _cmsGen->used_region(),
5622 &modUnionClosure);
5623 _ct->ct_bs()->dirty_card_iterate(
5624 _permGen->used_region(),
5625 &modUnionClosure);
5626 }
5627 // Having transferred these marks into the modUnionTable, we just need
5628 // to rescan the marked objects on the dirty cards in the modUnionTable.
5629 // The initial marking may have been done during an asynchronous
5630 // collection so there may be dirty bits in the mod-union table.
5631 const int alignment =
5632 CardTableModRefBS::card_size * BitsPerWord;
5633 {
5634 // ... First handle dirty cards in CMS gen
5635 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5636 MemRegion ur = _cmsGen->used_region();
5637 HeapWord* lb = ur.start();
5638 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5639 MemRegion cms_span(lb, ub);
5640 _modUnionTable.dirty_range_iterate_clear(cms_span,
5641 &markFromDirtyCardsClosure);
5642 verify_work_stacks_empty();
5643 if (PrintCMSStatistics != 0) {
5644 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5645 markFromDirtyCardsClosure.num_dirty_cards());
5646 }
5647 }
5648 {
5649 // .. and then repeat for dirty cards in perm gen
5650 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace());
5651 MemRegion ur = _permGen->used_region();
5652 HeapWord* lb = ur.start();
5653 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5654 MemRegion perm_span(lb, ub);
5655 _modUnionTable.dirty_range_iterate_clear(perm_span,
5656 &markFromDirtyCardsClosure);
5657 verify_work_stacks_empty();
5658 if (PrintCMSStatistics != 0) {
5659 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ",
5660 markFromDirtyCardsClosure.num_dirty_cards());
5661 }
5662 }
5663 }
5664 if (VerifyDuringGC &&
5665 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5666 HandleMark hm; // Discard invalid handles created during verification
5667 Universe::verify(true);
5668 }
5669 {
5670 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty);
5671
5672 verify_work_stacks_empty();
5673
5674 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5675 GenCollectedHeap::StrongRootsScope srs(gch);
5676 gch->gen_process_strong_roots(_cmsGen->level(),
5677 true, // younger gens as roots
5678 false, // use the local StrongRootsScope
5679 true, // collecting perm gen
5680 SharedHeap::ScanningOption(roots_scanning_options()),
5681 &mrias_cl,
5682 true, // walk code active on stacks
5683 NULL);
5684 assert(should_unload_classes()
5685 || (roots_scanning_options() & SharedHeap::SO_CodeCache),
5686 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5687 }
5688 verify_work_stacks_empty();
5689 // Restore evacuated mark words, if any, used for overflow list links
5690 if (!CMSOverflowEarlyRestoration) {
5691 restore_preserved_marks_if_any();
5692 }
5693 verify_overflow_empty();
5694 }
5695
5696 ////////////////////////////////////////////////////////
5697 // Parallel Reference Processing Task Proxy Class
5698 ////////////////////////////////////////////////////////
5699 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5700 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5701 CMSCollector* _collector;
5702 CMSBitMap* _mark_bit_map;
5703 const MemRegion _span;
5704 ProcessTask& _task;
5705
5706 public:
5707 CMSRefProcTaskProxy(ProcessTask& task,
5708 CMSCollector* collector,
5709 const MemRegion& span,
5710 CMSBitMap* mark_bit_map,
5711 AbstractWorkGang* workers,
5712 OopTaskQueueSet* task_queues):
5713 // XXX Should superclass AGTWOQ also know about AWG since it knows
5714 // about the task_queues used by the AWG? Then it could initialize
5715 // the terminator() object. See 6984287. The set_for_termination()
5716 // below is a temporary band-aid for the regression in 6984287.
5717 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5718 task_queues),
5719 _task(task),
5720 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5721 {
5722 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5723 "Inconsistency in _span");
5724 set_for_termination(workers->active_workers());
5725 }
5726
5727 OopTaskQueueSet* task_queues() { return queues(); }
5728
5729 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5730
5731 void do_work_steal(int i,
5732 CMSParDrainMarkingStackClosure* drain,
5733 CMSParKeepAliveClosure* keep_alive,
5734 int* seed);
5735
5736 virtual void work(int i);
5737 };
5738
5739 void CMSRefProcTaskProxy::work(int i) {
5740 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5741 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5742 _mark_bit_map,
5743 &_collector->_revisitStack,
5744 work_queue(i));
5745 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5746 _mark_bit_map,
5747 &_collector->_revisitStack,
5748 work_queue(i));
5749 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5750 _task.work(i, is_alive_closure, par_keep_alive, par_drain_stack);
5751 if (_task.marks_oops_alive()) {
5752 do_work_steal(i, &par_drain_stack, &par_keep_alive,
5753 _collector->hash_seed(i));
5754 }
5755 assert(work_queue(i)->size() == 0, "work_queue should be empty");
5756 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5757 }
5758
5759 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5760 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5761 EnqueueTask& _task;
5762
5763 public:
5764 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5765 : AbstractGangTask("Enqueue reference objects in parallel"),
5766 _task(task)
5767 { }
5768
5769 virtual void work(int i)
5770 {
5771 _task.work(i);
5772 }
5773 };
5774
5775 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5776 MemRegion span, CMSBitMap* bit_map, CMSMarkStack* revisit_stack,
5777 OopTaskQueue* work_queue):
5778 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
5779 _span(span),
5780 _bit_map(bit_map),
5781 _work_queue(work_queue),
5782 _mark_and_push(collector, span, bit_map, revisit_stack, work_queue),
5783 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5784 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5785 { }
5786
5787 // . see if we can share work_queues with ParNew? XXX
5788 void CMSRefProcTaskProxy::do_work_steal(int i,
5789 CMSParDrainMarkingStackClosure* drain,
5790 CMSParKeepAliveClosure* keep_alive,
5791 int* seed) {
5792 OopTaskQueue* work_q = work_queue(i);
5793 NOT_PRODUCT(int num_steals = 0;)
5794 oop obj_to_scan;
5795
5796 while (true) {
5797 // Completely finish any left over work from (an) earlier round(s)
5798 drain->trim_queue(0);
5799 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5800 (size_t)ParGCDesiredObjsFromOverflowList);
5801 // Now check if there's any work in the overflow list
5802 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5803 // only affects the number of attempts made to get work from the
5804 // overflow list and does not affect the number of workers. Just
5805 // pass ParallelGCThreads so this behavior is unchanged.
5806 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5807 work_q,
5808 ParallelGCThreads)) {
5809 // Found something in global overflow list;
5810 // not yet ready to go stealing work from others.
5811 // We'd like to assert(work_q->size() != 0, ...)
5812 // because we just took work from the overflow list,
5813 // but of course we can't, since all of that might have
5814 // been already stolen from us.
5815 continue;
5816 }
5817 // Verify that we have no work before we resort to stealing
5818 assert(work_q->size() == 0, "Have work, shouldn't steal");
5819 // Try to steal from other queues that have work
5820 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5821 NOT_PRODUCT(num_steals++;)
5822 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5823 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5824 // Do scanning work
5825 obj_to_scan->oop_iterate(keep_alive);
5826 // Loop around, finish this work, and try to steal some more
5827 } else if (terminator()->offer_termination()) {
5828 break; // nirvana from the infinite cycle
5829 }
5830 }
5831 NOT_PRODUCT(
5832 if (PrintCMSStatistics != 0) {
5833 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5834 }
5835 )
5836 }
5837
5838 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5839 {
5840 GenCollectedHeap* gch = GenCollectedHeap::heap();
5841 FlexibleWorkGang* workers = gch->workers();
5842 assert(workers != NULL, "Need parallel worker threads.");
5843 CMSRefProcTaskProxy rp_task(task, &_collector,
5844 _collector.ref_processor()->span(),
5845 _collector.markBitMap(),
5846 workers, _collector.task_queues());
5847 workers->run_task(&rp_task);
5848 }
5849
5850 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5851 {
5852
5853 GenCollectedHeap* gch = GenCollectedHeap::heap();
5854 FlexibleWorkGang* workers = gch->workers();
5855 assert(workers != NULL, "Need parallel worker threads.");
5856 CMSRefEnqueueTaskProxy enq_task(task);
5857 workers->run_task(&enq_task);
5858 }
5859
5860 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
5861
5862 ResourceMark rm;
5863 HandleMark hm;
5864
5865 ReferenceProcessor* rp = ref_processor();
5866 assert(rp->span().equals(_span), "Spans should be equal");
5867 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5868 // Process weak references.
5869 rp->setup_policy(clear_all_soft_refs);
5870 verify_work_stacks_empty();
5871
5872 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5873 &_markStack, &_revisitStack,
5874 false /* !preclean */);
5875 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5876 _span, &_markBitMap, &_markStack,
5877 &cmsKeepAliveClosure, false /* !preclean */);
5878 {
5879 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty);
5880 if (rp->processing_is_mt()) {
5881 // Set the degree of MT here. If the discovery is done MT, there
5882 // may have been a different number of threads doing the discovery
5883 // and a different number of discovered lists may have Ref objects.
5884 // That is OK as long as the Reference lists are balanced (see
5885 // balance_all_queues() and balance_queues()).
5886
5887 rp->set_active_mt_degree(ParallelGCThreads);
5888 CMSRefProcTaskExecutor task_executor(*this);
5889 rp->process_discovered_references(&_is_alive_closure,
5890 &cmsKeepAliveClosure,
5891 &cmsDrainMarkingStackClosure,
5892 &task_executor);
5893 } else {
5894 rp->process_discovered_references(&_is_alive_closure,
5895 &cmsKeepAliveClosure,
5896 &cmsDrainMarkingStackClosure,
5897 NULL);
5898 }
5899 verify_work_stacks_empty();
5900 }
5901
5902 if (should_unload_classes()) {
5903 {
5904 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty);
5905
5906 // Follow SystemDictionary roots and unload classes
5907 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5908
5909 // Follow CodeCache roots and unload any methods marked for unloading
5910 CodeCache::do_unloading(&_is_alive_closure,
5911 &cmsKeepAliveClosure,
5912 purged_class);
5913
5914 cmsDrainMarkingStackClosure.do_void();
5915 verify_work_stacks_empty();
5916
5917 // Update subklass/sibling/implementor links in KlassKlass descendants
5918 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty");
5919 oop k;
5920 while ((k = _revisitStack.pop()) != NULL) {
5921 ((Klass*)(oopDesc*)k)->follow_weak_klass_links(
5922 &_is_alive_closure,
5923 &cmsKeepAliveClosure);
5924 }
5925 assert(!ClassUnloading ||
5926 (_markStack.isEmpty() && overflow_list_is_empty()),
5927 "Should not have found new reachable objects");
5928 assert(_revisitStack.isEmpty(), "revisit stack should have been drained");
5929 cmsDrainMarkingStackClosure.do_void();
5930 verify_work_stacks_empty();
5931 }
5932
5933 {
5934 TraceTime t("scrub symbol table", PrintGCDetails, false, gclog_or_tty);
5935 // Clean up unreferenced symbols in symbol table.
5936 SymbolTable::unlink();
5937 }
5938 }
5939
5940 if (should_unload_classes() || !JavaObjectsInPerm) {
5941 TraceTime t("scrub string table", PrintGCDetails, false, gclog_or_tty);
5942 // Now clean up stale oops in StringTable
5943 StringTable::unlink(&_is_alive_closure);
5944 }
5945
5946 verify_work_stacks_empty();
5947 // Restore any preserved marks as a result of mark stack or
5948 // work queue overflow
5949 restore_preserved_marks_if_any(); // done single-threaded for now
5950
5951 rp->set_enqueuing_is_done(true);
5952 if (rp->processing_is_mt()) {
5953 rp->balance_all_queues();
5954 CMSRefProcTaskExecutor task_executor(*this);
5955 rp->enqueue_discovered_references(&task_executor);
5956 } else {
5957 rp->enqueue_discovered_references(NULL);
5958 }
5959 rp->verify_no_references_recorded();
5960 assert(!rp->discovery_enabled(), "should have been disabled");
5961 }
5962
5963 #ifndef PRODUCT
5964 void CMSCollector::check_correct_thread_executing() {
5965 Thread* t = Thread::current();
5966 // Only the VM thread or the CMS thread should be here.
5967 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5968 "Unexpected thread type");
5969 // If this is the vm thread, the foreground process
5970 // should not be waiting. Note that _foregroundGCIsActive is
5971 // true while the foreground collector is waiting.
5972 if (_foregroundGCShouldWait) {
5973 // We cannot be the VM thread
5974 assert(t->is_ConcurrentGC_thread(),
5975 "Should be CMS thread");
5976 } else {
5977 // We can be the CMS thread only if we are in a stop-world
5978 // phase of CMS collection.
5979 if (t->is_ConcurrentGC_thread()) {
5980 assert(_collectorState == InitialMarking ||
5981 _collectorState == FinalMarking,
5982 "Should be a stop-world phase");
5983 // The CMS thread should be holding the CMS_token.
5984 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5985 "Potential interference with concurrently "
5986 "executing VM thread");
5987 }
5988 }
5989 }
5990 #endif
5991
5992 void CMSCollector::sweep(bool asynch) {
5993 assert(_collectorState == Sweeping, "just checking");
5994 check_correct_thread_executing();
5995 verify_work_stacks_empty();
5996 verify_overflow_empty();
5997 increment_sweep_count();
5998 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5999
6000 _inter_sweep_timer.stop();
6001 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
6002 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
6003
6004 // PermGen verification support: If perm gen sweeping is disabled in
6005 // this cycle, we preserve the perm gen object "deadness" information
6006 // in the perm_gen_verify_bit_map. In order to do that we traverse
6007 // all blocks in perm gen and mark all dead objects.
6008 if (verifying() && !should_unload_classes()) {
6009 assert(perm_gen_verify_bit_map()->sizeInBits() != 0,
6010 "Should have already been allocated");
6011 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(),
6012 markBitMap(), perm_gen_verify_bit_map());
6013 if (asynch) {
6014 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
6015 bitMapLock());
6016 _permGen->cmsSpace()->blk_iterate(&mdo);
6017 } else {
6018 // In the case of synchronous sweep, we already have
6019 // the requisite locks/tokens.
6020 _permGen->cmsSpace()->blk_iterate(&mdo);
6021 }
6022 }
6023
6024 assert(!_intra_sweep_timer.is_active(), "Should not be active");
6025 _intra_sweep_timer.reset();
6026 _intra_sweep_timer.start();
6027 if (asynch) {
6028 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6029 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
6030 // First sweep the old gen then the perm gen
6031 {
6032 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6033 bitMapLock());
6034 sweepWork(_cmsGen, asynch);
6035 }
6036
6037 // Now repeat for perm gen
6038 if (should_unload_classes()) {
6039 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
6040 bitMapLock());
6041 sweepWork(_permGen, asynch);
6042 }
6043
6044 // Update Universe::_heap_*_at_gc figures.
6045 // We need all the free list locks to make the abstract state
6046 // transition from Sweeping to Resetting. See detailed note
6047 // further below.
6048 {
6049 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6050 _permGen->freelistLock());
6051 // Update heap occupancy information which is used as
6052 // input to soft ref clearing policy at the next gc.
6053 Universe::update_heap_info_at_gc();
6054 _collectorState = Resizing;
6055 }
6056 } else {
6057 // already have needed locks
6058 sweepWork(_cmsGen, asynch);
6059
6060 if (should_unload_classes()) {
6061 sweepWork(_permGen, asynch);
6062 }
6063 // Update heap occupancy information which is used as
6064 // input to soft ref clearing policy at the next gc.
6065 Universe::update_heap_info_at_gc();
6066 _collectorState = Resizing;
6067 }
6068 verify_work_stacks_empty();
6069 verify_overflow_empty();
6070
6071 _intra_sweep_timer.stop();
6072 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
6073
6074 _inter_sweep_timer.reset();
6075 _inter_sweep_timer.start();
6076
6077 update_time_of_last_gc(os::javaTimeMillis());
6078
6079 // NOTE on abstract state transitions:
6080 // Mutators allocate-live and/or mark the mod-union table dirty
6081 // based on the state of the collection. The former is done in
6082 // the interval [Marking, Sweeping] and the latter in the interval
6083 // [Marking, Sweeping). Thus the transitions into the Marking state
6084 // and out of the Sweeping state must be synchronously visible
6085 // globally to the mutators.
6086 // The transition into the Marking state happens with the world
6087 // stopped so the mutators will globally see it. Sweeping is
6088 // done asynchronously by the background collector so the transition
6089 // from the Sweeping state to the Resizing state must be done
6090 // under the freelistLock (as is the check for whether to
6091 // allocate-live and whether to dirty the mod-union table).
6092 assert(_collectorState == Resizing, "Change of collector state to"
6093 " Resizing must be done under the freelistLocks (plural)");
6094
6095 // Now that sweeping has been completed, we clear
6096 // the incremental_collection_failed flag,
6097 // thus inviting a younger gen collection to promote into
6098 // this generation. If such a promotion may still fail,
6099 // the flag will be set again when a young collection is
6100 // attempted.
6101 GenCollectedHeap* gch = GenCollectedHeap::heap();
6102 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
6103 gch->update_full_collections_completed(_collection_count_start);
6104 }
6105
6106 // FIX ME!!! Looks like this belongs in CFLSpace, with
6107 // CMSGen merely delegating to it.
6108 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
6109 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
6110 HeapWord* minAddr = _cmsSpace->bottom();
6111 HeapWord* largestAddr =
6112 (HeapWord*) _cmsSpace->dictionary()->findLargestDict();
6113 if (largestAddr == NULL) {
6114 // The dictionary appears to be empty. In this case
6115 // try to coalesce at the end of the heap.
6116 largestAddr = _cmsSpace->end();
6117 }
6118 size_t largestOffset = pointer_delta(largestAddr, minAddr);
6119 size_t nearLargestOffset =
6120 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
6121 if (PrintFLSStatistics != 0) {
6122 gclog_or_tty->print_cr(
6123 "CMS: Large Block: " PTR_FORMAT ";"
6124 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
6125 largestAddr,
6126 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset);
6127 }
6128 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
6129 }
6130
6131 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
6132 return addr >= _cmsSpace->nearLargestChunk();
6133 }
6134
6135 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
6136 return _cmsSpace->find_chunk_at_end();
6137 }
6138
6139 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
6140 bool full) {
6141 // The next lower level has been collected. Gather any statistics
6142 // that are of interest at this point.
6143 if (!full && (current_level + 1) == level()) {
6144 // Gather statistics on the young generation collection.
6145 collector()->stats().record_gc0_end(used());
6146 }
6147 }
6148
6149 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
6150 GenCollectedHeap* gch = GenCollectedHeap::heap();
6151 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
6152 "Wrong type of heap");
6153 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
6154 gch->gen_policy()->size_policy();
6155 assert(sp->is_gc_cms_adaptive_size_policy(),
6156 "Wrong type of size policy");
6157 return sp;
6158 }
6159
6160 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
6161 if (PrintGCDetails && Verbose) {
6162 gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
6163 }
6164 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
6165 _debug_collection_type =
6166 (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
6167 if (PrintGCDetails && Verbose) {
6168 gclog_or_tty->print_cr("to %d ", _debug_collection_type);
6169 }
6170 }
6171
6172 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
6173 bool asynch) {
6174 // We iterate over the space(s) underlying this generation,
6175 // checking the mark bit map to see if the bits corresponding
6176 // to specific blocks are marked or not. Blocks that are
6177 // marked are live and are not swept up. All remaining blocks
6178 // are swept up, with coalescing on-the-fly as we sweep up
6179 // contiguous free and/or garbage blocks:
6180 // We need to ensure that the sweeper synchronizes with allocators
6181 // and stop-the-world collectors. In particular, the following
6182 // locks are used:
6183 // . CMS token: if this is held, a stop the world collection cannot occur
6184 // . freelistLock: if this is held no allocation can occur from this
6185 // generation by another thread
6186 // . bitMapLock: if this is held, no other thread can access or update
6187 //
6188
6189 // Note that we need to hold the freelistLock if we use
6190 // block iterate below; else the iterator might go awry if
6191 // a mutator (or promotion) causes block contents to change
6192 // (for instance if the allocator divvies up a block).
6193 // If we hold the free list lock, for all practical purposes
6194 // young generation GC's can't occur (they'll usually need to
6195 // promote), so we might as well prevent all young generation
6196 // GC's while we do a sweeping step. For the same reason, we might
6197 // as well take the bit map lock for the entire duration
6198
6199 // check that we hold the requisite locks
6200 assert(have_cms_token(), "Should hold cms token");
6201 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
6202 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
6203 "Should possess CMS token to sweep");
6204 assert_lock_strong(gen->freelistLock());
6205 assert_lock_strong(bitMapLock());
6206
6207 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
6208 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
6209 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
6210 _inter_sweep_estimate.padded_average(),
6211 _intra_sweep_estimate.padded_average());
6212 gen->setNearLargestChunk();
6213
6214 {
6215 SweepClosure sweepClosure(this, gen, &_markBitMap,
6216 CMSYield && asynch);
6217 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
6218 // We need to free-up/coalesce garbage/blocks from a
6219 // co-terminal free run. This is done in the SweepClosure
6220 // destructor; so, do not remove this scope, else the
6221 // end-of-sweep-census below will be off by a little bit.
6222 }
6223 gen->cmsSpace()->sweep_completed();
6224 gen->cmsSpace()->endSweepFLCensus(sweep_count());
6225 if (should_unload_classes()) { // unloaded classes this cycle,
6226 _concurrent_cycles_since_last_unload = 0; // ... reset count
6227 } else { // did not unload classes,
6228 _concurrent_cycles_since_last_unload++; // ... increment count
6229 }
6230 }
6231
6232 // Reset CMS data structures (for now just the marking bit map)
6233 // preparatory for the next cycle.
6234 void CMSCollector::reset(bool asynch) {
6235 GenCollectedHeap* gch = GenCollectedHeap::heap();
6236 CMSAdaptiveSizePolicy* sp = size_policy();
6237 AdaptiveSizePolicyOutput(sp, gch->total_collections());
6238 if (asynch) {
6239 CMSTokenSyncWithLocks ts(true, bitMapLock());
6240
6241 // If the state is not "Resetting", the foreground thread
6242 // has done a collection and the resetting.
6243 if (_collectorState != Resetting) {
6244 assert(_collectorState == Idling, "The state should only change"
6245 " because the foreground collector has finished the collection");
6246 return;
6247 }
6248
6249 // Clear the mark bitmap (no grey objects to start with)
6250 // for the next cycle.
6251 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6252 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
6253
6254 HeapWord* curAddr = _markBitMap.startWord();
6255 while (curAddr < _markBitMap.endWord()) {
6256 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
6257 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
6258 _markBitMap.clear_large_range(chunk);
6259 if (ConcurrentMarkSweepThread::should_yield() &&
6260 !foregroundGCIsActive() &&
6261 CMSYield) {
6262 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6263 "CMS thread should hold CMS token");
6264 assert_lock_strong(bitMapLock());
6265 bitMapLock()->unlock();
6266 ConcurrentMarkSweepThread::desynchronize(true);
6267 ConcurrentMarkSweepThread::acknowledge_yield_request();
6268 stopTimer();
6269 if (PrintCMSStatistics != 0) {
6270 incrementYields();
6271 }
6272 icms_wait();
6273
6274 // See the comment in coordinator_yield()
6275 for (unsigned i = 0; i < CMSYieldSleepCount &&
6276 ConcurrentMarkSweepThread::should_yield() &&
6277 !CMSCollector::foregroundGCIsActive(); ++i) {
6278 os::sleep(Thread::current(), 1, false);
6279 ConcurrentMarkSweepThread::acknowledge_yield_request();
6280 }
6281
6282 ConcurrentMarkSweepThread::synchronize(true);
6283 bitMapLock()->lock_without_safepoint_check();
6284 startTimer();
6285 }
6286 curAddr = chunk.end();
6287 }
6288 // A successful mostly concurrent collection has been done.
6289 // Because only the full (i.e., concurrent mode failure) collections
6290 // are being measured for gc overhead limits, clean the "near" flag
6291 // and count.
6292 sp->reset_gc_overhead_limit_count();
6293 _collectorState = Idling;
6294 } else {
6295 // already have the lock
6296 assert(_collectorState == Resetting, "just checking");
6297 assert_lock_strong(bitMapLock());
6298 _markBitMap.clear_all();
6299 _collectorState = Idling;
6300 }
6301
6302 // Stop incremental mode after a cycle completes, so that any future cycles
6303 // are triggered by allocation.
6304 stop_icms();
6305
6306 NOT_PRODUCT(
6307 if (RotateCMSCollectionTypes) {
6308 _cmsGen->rotate_debug_collection_type();
6309 }
6310 )
6311 }
6312
6313 void CMSCollector::do_CMS_operation(CMS_op_type op) {
6314 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
6315 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6316 TraceTime t("GC", PrintGC, !PrintGCDetails, gclog_or_tty);
6317 TraceCollectorStats tcs(counters());
6318
6319 switch (op) {
6320 case CMS_op_checkpointRootsInitial: {
6321 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6322 checkpointRootsInitial(true); // asynch
6323 if (PrintGC) {
6324 _cmsGen->printOccupancy("initial-mark");
6325 }
6326 break;
6327 }
6328 case CMS_op_checkpointRootsFinal: {
6329 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6330 checkpointRootsFinal(true, // asynch
6331 false, // !clear_all_soft_refs
6332 false); // !init_mark_was_synchronous
6333 if (PrintGC) {
6334 _cmsGen->printOccupancy("remark");
6335 }
6336 break;
6337 }
6338 default:
6339 fatal("No such CMS_op");
6340 }
6341 }
6342
6343 #ifndef PRODUCT
6344 size_t const CMSCollector::skip_header_HeapWords() {
6345 return FreeChunk::header_size();
6346 }
6347
6348 // Try and collect here conditions that should hold when
6349 // CMS thread is exiting. The idea is that the foreground GC
6350 // thread should not be blocked if it wants to terminate
6351 // the CMS thread and yet continue to run the VM for a while
6352 // after that.
6353 void CMSCollector::verify_ok_to_terminate() const {
6354 assert(Thread::current()->is_ConcurrentGC_thread(),
6355 "should be called by CMS thread");
6356 assert(!_foregroundGCShouldWait, "should be false");
6357 // We could check here that all the various low-level locks
6358 // are not held by the CMS thread, but that is overkill; see
6359 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6360 // is checked.
6361 }
6362 #endif
6363
6364 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6365 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6366 "missing Printezis mark?");
6367 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6368 size_t size = pointer_delta(nextOneAddr + 1, addr);
6369 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6370 "alignment problem");
6371 assert(size >= 3, "Necessary for Printezis marks to work");
6372 return size;
6373 }
6374
6375 // A variant of the above (block_size_using_printezis_bits()) except
6376 // that we return 0 if the P-bits are not yet set.
6377 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6378 if (_markBitMap.isMarked(addr + 1)) {
6379 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
6380 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6381 size_t size = pointer_delta(nextOneAddr + 1, addr);
6382 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6383 "alignment problem");
6384 assert(size >= 3, "Necessary for Printezis marks to work");
6385 return size;
6386 }
6387 return 0;
6388 }
6389
6390 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6391 size_t sz = 0;
6392 oop p = (oop)addr;
6393 if (p->klass_or_null() != NULL && p->is_parsable()) {
6394 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6395 } else {
6396 sz = block_size_using_printezis_bits(addr);
6397 }
6398 assert(sz > 0, "size must be nonzero");
6399 HeapWord* next_block = addr + sz;
6400 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
6401 CardTableModRefBS::card_size);
6402 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
6403 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6404 "must be different cards");
6405 return next_card;
6406 }
6407
6408
6409 // CMS Bit Map Wrapper /////////////////////////////////////////
6410
6411 // Construct a CMS bit map infrastructure, but don't create the
6412 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6413 // further below.
6414 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6415 _bm(),
6416 _shifter(shifter),
6417 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6418 {
6419 _bmStartWord = 0;
6420 _bmWordSize = 0;
6421 }
6422
6423 bool CMSBitMap::allocate(MemRegion mr) {
6424 _bmStartWord = mr.start();
6425 _bmWordSize = mr.word_size();
6426 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6427 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6428 if (!brs.is_reserved()) {
6429 warning("CMS bit map allocation failure");
6430 return false;
6431 }
6432 // For now we'll just commit all of the bit map up fromt.
6433 // Later on we'll try to be more parsimonious with swap.
6434 if (!_virtual_space.initialize(brs, brs.size())) {
6435 warning("CMS bit map backing store failure");
6436 return false;
6437 }
6438 assert(_virtual_space.committed_size() == brs.size(),
6439 "didn't reserve backing store for all of CMS bit map?");
6440 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
6441 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6442 _bmWordSize, "inconsistency in bit map sizing");
6443 _bm.set_size(_bmWordSize >> _shifter);
6444
6445 // bm.clear(); // can we rely on getting zero'd memory? verify below
6446 assert(isAllClear(),
6447 "Expected zero'd memory from ReservedSpace constructor");
6448 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6449 "consistency check");
6450 return true;
6451 }
6452
6453 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6454 HeapWord *next_addr, *end_addr, *last_addr;
6455 assert_locked();
6456 assert(covers(mr), "out-of-range error");
6457 // XXX assert that start and end are appropriately aligned
6458 for (next_addr = mr.start(), end_addr = mr.end();
6459 next_addr < end_addr; next_addr = last_addr) {
6460 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6461 last_addr = dirty_region.end();
6462 if (!dirty_region.is_empty()) {
6463 cl->do_MemRegion(dirty_region);
6464 } else {
6465 assert(last_addr == end_addr, "program logic");
6466 return;
6467 }
6468 }
6469 }
6470
6471 #ifndef PRODUCT
6472 void CMSBitMap::assert_locked() const {
6473 CMSLockVerifier::assert_locked(lock());
6474 }
6475
6476 bool CMSBitMap::covers(MemRegion mr) const {
6477 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6478 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6479 "size inconsistency");
6480 return (mr.start() >= _bmStartWord) &&
6481 (mr.end() <= endWord());
6482 }
6483
6484 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6485 return (start >= _bmStartWord && (start + size) <= endWord());
6486 }
6487
6488 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6489 // verify that there are no 1 bits in the interval [left, right)
6490 FalseBitMapClosure falseBitMapClosure;
6491 iterate(&falseBitMapClosure, left, right);
6492 }
6493
6494 void CMSBitMap::region_invariant(MemRegion mr)
6495 {
6496 assert_locked();
6497 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6498 assert(!mr.is_empty(), "unexpected empty region");
6499 assert(covers(mr), "mr should be covered by bit map");
6500 // convert address range into offset range
6501 size_t start_ofs = heapWordToOffset(mr.start());
6502 // Make sure that end() is appropriately aligned
6503 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6504 (1 << (_shifter+LogHeapWordSize))),
6505 "Misaligned mr.end()");
6506 size_t end_ofs = heapWordToOffset(mr.end());
6507 assert(end_ofs > start_ofs, "Should mark at least one bit");
6508 }
6509
6510 #endif
6511
6512 bool CMSMarkStack::allocate(size_t size) {
6513 // allocate a stack of the requisite depth
6514 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6515 size * sizeof(oop)));
6516 if (!rs.is_reserved()) {
6517 warning("CMSMarkStack allocation failure");
6518 return false;
6519 }
6520 if (!_virtual_space.initialize(rs, rs.size())) {
6521 warning("CMSMarkStack backing store failure");
6522 return false;
6523 }
6524 assert(_virtual_space.committed_size() == rs.size(),
6525 "didn't reserve backing store for all of CMS stack?");
6526 _base = (oop*)(_virtual_space.low());
6527 _index = 0;
6528 _capacity = size;
6529 NOT_PRODUCT(_max_depth = 0);
6530 return true;
6531 }
6532
6533 // XXX FIX ME !!! In the MT case we come in here holding a
6534 // leaf lock. For printing we need to take a further lock
6535 // which has lower rank. We need to recallibrate the two
6536 // lock-ranks involved in order to be able to rpint the
6537 // messages below. (Or defer the printing to the caller.
6538 // For now we take the expedient path of just disabling the
6539 // messages for the problematic case.)
6540 void CMSMarkStack::expand() {
6541 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6542 if (_capacity == MarkStackSizeMax) {
6543 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6544 // We print a warning message only once per CMS cycle.
6545 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6546 }
6547 return;
6548 }
6549 // Double capacity if possible
6550 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6551 // Do not give up existing stack until we have managed to
6552 // get the double capacity that we desired.
6553 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6554 new_capacity * sizeof(oop)));
6555 if (rs.is_reserved()) {
6556 // Release the backing store associated with old stack
6557 _virtual_space.release();
6558 // Reinitialize virtual space for new stack
6559 if (!_virtual_space.initialize(rs, rs.size())) {
6560 fatal("Not enough swap for expanded marking stack");
6561 }
6562 _base = (oop*)(_virtual_space.low());
6563 _index = 0;
6564 _capacity = new_capacity;
6565 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6566 // Failed to double capacity, continue;
6567 // we print a detail message only once per CMS cycle.
6568 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6569 SIZE_FORMAT"K",
6570 _capacity / K, new_capacity / K);
6571 }
6572 }
6573
6574
6575 // Closures
6576 // XXX: there seems to be a lot of code duplication here;
6577 // should refactor and consolidate common code.
6578
6579 // This closure is used to mark refs into the CMS generation in
6580 // the CMS bit map. Called at the first checkpoint. This closure
6581 // assumes that we do not need to re-mark dirty cards; if the CMS
6582 // generation on which this is used is not an oldest (modulo perm gen)
6583 // generation then this will lose younger_gen cards!
6584
6585 MarkRefsIntoClosure::MarkRefsIntoClosure(
6586 MemRegion span, CMSBitMap* bitMap):
6587 _span(span),
6588 _bitMap(bitMap)
6589 {
6590 assert(_ref_processor == NULL, "deliberately left NULL");
6591 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6592 }
6593
6594 void MarkRefsIntoClosure::do_oop(oop obj) {
6595 // if p points into _span, then mark corresponding bit in _markBitMap
6596 assert(obj->is_oop(), "expected an oop");
6597 HeapWord* addr = (HeapWord*)obj;
6598 if (_span.contains(addr)) {
6599 // this should be made more efficient
6600 _bitMap->mark(addr);
6601 }
6602 }
6603
6604 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6605 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6606
6607 // A variant of the above, used for CMS marking verification.
6608 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6609 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6610 _span(span),
6611 _verification_bm(verification_bm),
6612 _cms_bm(cms_bm)
6613 {
6614 assert(_ref_processor == NULL, "deliberately left NULL");
6615 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6616 }
6617
6618 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6619 // if p points into _span, then mark corresponding bit in _markBitMap
6620 assert(obj->is_oop(), "expected an oop");
6621 HeapWord* addr = (HeapWord*)obj;
6622 if (_span.contains(addr)) {
6623 _verification_bm->mark(addr);
6624 if (!_cms_bm->isMarked(addr)) {
6625 oop(addr)->print();
6626 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6627 fatal("... aborting");
6628 }
6629 }
6630 }
6631
6632 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6633 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6634
6635 //////////////////////////////////////////////////
6636 // MarkRefsIntoAndScanClosure
6637 //////////////////////////////////////////////////
6638
6639 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6640 ReferenceProcessor* rp,
6641 CMSBitMap* bit_map,
6642 CMSBitMap* mod_union_table,
6643 CMSMarkStack* mark_stack,
6644 CMSMarkStack* revisit_stack,
6645 CMSCollector* collector,
6646 bool should_yield,
6647 bool concurrent_precleaning):
6648 _collector(collector),
6649 _span(span),
6650 _bit_map(bit_map),
6651 _mark_stack(mark_stack),
6652 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6653 mark_stack, revisit_stack, concurrent_precleaning),
6654 _yield(should_yield),
6655 _concurrent_precleaning(concurrent_precleaning),
6656 _freelistLock(NULL)
6657 {
6658 _ref_processor = rp;
6659 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6660 }
6661
6662 // This closure is used to mark refs into the CMS generation at the
6663 // second (final) checkpoint, and to scan and transitively follow
6664 // the unmarked oops. It is also used during the concurrent precleaning
6665 // phase while scanning objects on dirty cards in the CMS generation.
6666 // The marks are made in the marking bit map and the marking stack is
6667 // used for keeping the (newly) grey objects during the scan.
6668 // The parallel version (Par_...) appears further below.
6669 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6670 if (obj != NULL) {
6671 assert(obj->is_oop(), "expected an oop");
6672 HeapWord* addr = (HeapWord*)obj;
6673 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6674 assert(_collector->overflow_list_is_empty(),
6675 "overflow list should be empty");
6676 if (_span.contains(addr) &&
6677 !_bit_map->isMarked(addr)) {
6678 // mark bit map (object is now grey)
6679 _bit_map->mark(addr);
6680 // push on marking stack (stack should be empty), and drain the
6681 // stack by applying this closure to the oops in the oops popped
6682 // from the stack (i.e. blacken the grey objects)
6683 bool res = _mark_stack->push(obj);
6684 assert(res, "Should have space to push on empty stack");
6685 do {
6686 oop new_oop = _mark_stack->pop();
6687 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6688 assert(new_oop->is_parsable(), "Found unparsable oop");
6689 assert(_bit_map->isMarked((HeapWord*)new_oop),
6690 "only grey objects on this stack");
6691 // iterate over the oops in this oop, marking and pushing
6692 // the ones in CMS heap (i.e. in _span).
6693 new_oop->oop_iterate(&_pushAndMarkClosure);
6694 // check if it's time to yield
6695 do_yield_check();
6696 } while (!_mark_stack->isEmpty() ||
6697 (!_concurrent_precleaning && take_from_overflow_list()));
6698 // if marking stack is empty, and we are not doing this
6699 // during precleaning, then check the overflow list
6700 }
6701 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6702 assert(_collector->overflow_list_is_empty(),
6703 "overflow list was drained above");
6704 // We could restore evacuated mark words, if any, used for
6705 // overflow list links here because the overflow list is
6706 // provably empty here. That would reduce the maximum
6707 // size requirements for preserved_{oop,mark}_stack.
6708 // But we'll just postpone it until we are all done
6709 // so we can just stream through.
6710 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6711 _collector->restore_preserved_marks_if_any();
6712 assert(_collector->no_preserved_marks(), "No preserved marks");
6713 }
6714 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6715 "All preserved marks should have been restored above");
6716 }
6717 }
6718
6719 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6720 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6721
6722 void MarkRefsIntoAndScanClosure::do_yield_work() {
6723 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6724 "CMS thread should hold CMS token");
6725 assert_lock_strong(_freelistLock);
6726 assert_lock_strong(_bit_map->lock());
6727 // relinquish the free_list_lock and bitMaplock()
6728 DEBUG_ONLY(RememberKlassesChecker mux(false);)
6729 _bit_map->lock()->unlock();
6730 _freelistLock->unlock();
6731 ConcurrentMarkSweepThread::desynchronize(true);
6732 ConcurrentMarkSweepThread::acknowledge_yield_request();
6733 _collector->stopTimer();
6734 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6735 if (PrintCMSStatistics != 0) {
6736 _collector->incrementYields();
6737 }
6738 _collector->icms_wait();
6739
6740 // See the comment in coordinator_yield()
6741 for (unsigned i = 0;
6742 i < CMSYieldSleepCount &&
6743 ConcurrentMarkSweepThread::should_yield() &&
6744 !CMSCollector::foregroundGCIsActive();
6745 ++i) {
6746 os::sleep(Thread::current(), 1, false);
6747 ConcurrentMarkSweepThread::acknowledge_yield_request();
6748 }
6749
6750 ConcurrentMarkSweepThread::synchronize(true);
6751 _freelistLock->lock_without_safepoint_check();
6752 _bit_map->lock()->lock_without_safepoint_check();
6753 _collector->startTimer();
6754 }
6755
6756 ///////////////////////////////////////////////////////////
6757 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6758 // MarkRefsIntoAndScanClosure
6759 ///////////////////////////////////////////////////////////
6760 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6761 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6762 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack):
6763 _span(span),
6764 _bit_map(bit_map),
6765 _work_queue(work_queue),
6766 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6767 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6768 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue,
6769 revisit_stack)
6770 {
6771 _ref_processor = rp;
6772 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6773 }
6774
6775 // This closure is used to mark refs into the CMS generation at the
6776 // second (final) checkpoint, and to scan and transitively follow
6777 // the unmarked oops. The marks are made in the marking bit map and
6778 // the work_queue is used for keeping the (newly) grey objects during
6779 // the scan phase whence they are also available for stealing by parallel
6780 // threads. Since the marking bit map is shared, updates are
6781 // synchronized (via CAS).
6782 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6783 if (obj != NULL) {
6784 // Ignore mark word because this could be an already marked oop
6785 // that may be chained at the end of the overflow list.
6786 assert(obj->is_oop(true), "expected an oop");
6787 HeapWord* addr = (HeapWord*)obj;
6788 if (_span.contains(addr) &&
6789 !_bit_map->isMarked(addr)) {
6790 // mark bit map (object will become grey):
6791 // It is possible for several threads to be
6792 // trying to "claim" this object concurrently;
6793 // the unique thread that succeeds in marking the
6794 // object first will do the subsequent push on
6795 // to the work queue (or overflow list).
6796 if (_bit_map->par_mark(addr)) {
6797 // push on work_queue (which may not be empty), and trim the
6798 // queue to an appropriate length by applying this closure to
6799 // the oops in the oops popped from the stack (i.e. blacken the
6800 // grey objects)
6801 bool res = _work_queue->push(obj);
6802 assert(res, "Low water mark should be less than capacity?");
6803 trim_queue(_low_water_mark);
6804 } // Else, another thread claimed the object
6805 }
6806 }
6807 }
6808
6809 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6810 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6811
6812 // This closure is used to rescan the marked objects on the dirty cards
6813 // in the mod union table and the card table proper.
6814 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6815 oop p, MemRegion mr) {
6816
6817 size_t size = 0;
6818 HeapWord* addr = (HeapWord*)p;
6819 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6820 assert(_span.contains(addr), "we are scanning the CMS generation");
6821 // check if it's time to yield
6822 if (do_yield_check()) {
6823 // We yielded for some foreground stop-world work,
6824 // and we have been asked to abort this ongoing preclean cycle.
6825 return 0;
6826 }
6827 if (_bitMap->isMarked(addr)) {
6828 // it's marked; is it potentially uninitialized?
6829 if (p->klass_or_null() != NULL) {
6830 // If is_conc_safe is false, the object may be undergoing
6831 // change by the VM outside a safepoint. Don't try to
6832 // scan it, but rather leave it for the remark phase.
6833 if (CMSPermGenPrecleaningEnabled &&
6834 (!p->is_conc_safe() || !p->is_parsable())) {
6835 // Signal precleaning to redirty the card since
6836 // the klass pointer is already installed.
6837 assert(size == 0, "Initial value");
6838 } else {
6839 assert(p->is_parsable(), "must be parsable.");
6840 // an initialized object; ignore mark word in verification below
6841 // since we are running concurrent with mutators
6842 assert(p->is_oop(true), "should be an oop");
6843 if (p->is_objArray()) {
6844 // objArrays are precisely marked; restrict scanning
6845 // to dirty cards only.
6846 size = CompactibleFreeListSpace::adjustObjectSize(
6847 p->oop_iterate(_scanningClosure, mr));
6848 } else {
6849 // A non-array may have been imprecisely marked; we need
6850 // to scan object in its entirety.
6851 size = CompactibleFreeListSpace::adjustObjectSize(
6852 p->oop_iterate(_scanningClosure));
6853 }
6854 #ifdef DEBUG
6855 size_t direct_size =
6856 CompactibleFreeListSpace::adjustObjectSize(p->size());
6857 assert(size == direct_size, "Inconsistency in size");
6858 assert(size >= 3, "Necessary for Printezis marks to work");
6859 if (!_bitMap->isMarked(addr+1)) {
6860 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6861 } else {
6862 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6863 assert(_bitMap->isMarked(addr+size-1),
6864 "inconsistent Printezis mark");
6865 }
6866 #endif // DEBUG
6867 }
6868 } else {
6869 // an unitialized object
6870 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6871 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6872 size = pointer_delta(nextOneAddr + 1, addr);
6873 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6874 "alignment problem");
6875 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6876 // will dirty the card when the klass pointer is installed in the
6877 // object (signalling the completion of initialization).
6878 }
6879 } else {
6880 // Either a not yet marked object or an uninitialized object
6881 if (p->klass_or_null() == NULL || !p->is_parsable()) {
6882 // An uninitialized object, skip to the next card, since
6883 // we may not be able to read its P-bits yet.
6884 assert(size == 0, "Initial value");
6885 } else {
6886 // An object not (yet) reached by marking: we merely need to
6887 // compute its size so as to go look at the next block.
6888 assert(p->is_oop(true), "should be an oop");
6889 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6890 }
6891 }
6892 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6893 return size;
6894 }
6895
6896 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6897 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6898 "CMS thread should hold CMS token");
6899 assert_lock_strong(_freelistLock);
6900 assert_lock_strong(_bitMap->lock());
6901 DEBUG_ONLY(RememberKlassesChecker mux(false);)
6902 // relinquish the free_list_lock and bitMaplock()
6903 _bitMap->lock()->unlock();
6904 _freelistLock->unlock();
6905 ConcurrentMarkSweepThread::desynchronize(true);
6906 ConcurrentMarkSweepThread::acknowledge_yield_request();
6907 _collector->stopTimer();
6908 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6909 if (PrintCMSStatistics != 0) {
6910 _collector->incrementYields();
6911 }
6912 _collector->icms_wait();
6913
6914 // See the comment in coordinator_yield()
6915 for (unsigned i = 0; i < CMSYieldSleepCount &&
6916 ConcurrentMarkSweepThread::should_yield() &&
6917 !CMSCollector::foregroundGCIsActive(); ++i) {
6918 os::sleep(Thread::current(), 1, false);
6919 ConcurrentMarkSweepThread::acknowledge_yield_request();
6920 }
6921
6922 ConcurrentMarkSweepThread::synchronize(true);
6923 _freelistLock->lock_without_safepoint_check();
6924 _bitMap->lock()->lock_without_safepoint_check();
6925 _collector->startTimer();
6926 }
6927
6928
6929 //////////////////////////////////////////////////////////////////
6930 // SurvivorSpacePrecleanClosure
6931 //////////////////////////////////////////////////////////////////
6932 // This (single-threaded) closure is used to preclean the oops in
6933 // the survivor spaces.
6934 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6935
6936 HeapWord* addr = (HeapWord*)p;
6937 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6938 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6939 assert(p->klass_or_null() != NULL, "object should be initializd");
6940 assert(p->is_parsable(), "must be parsable.");
6941 // an initialized object; ignore mark word in verification below
6942 // since we are running concurrent with mutators
6943 assert(p->is_oop(true), "should be an oop");
6944 // Note that we do not yield while we iterate over
6945 // the interior oops of p, pushing the relevant ones
6946 // on our marking stack.
6947 size_t size = p->oop_iterate(_scanning_closure);
6948 do_yield_check();
6949 // Observe that below, we do not abandon the preclean
6950 // phase as soon as we should; rather we empty the
6951 // marking stack before returning. This is to satisfy
6952 // some existing assertions. In general, it may be a
6953 // good idea to abort immediately and complete the marking
6954 // from the grey objects at a later time.
6955 while (!_mark_stack->isEmpty()) {
6956 oop new_oop = _mark_stack->pop();
6957 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6958 assert(new_oop->is_parsable(), "Found unparsable oop");
6959 assert(_bit_map->isMarked((HeapWord*)new_oop),
6960 "only grey objects on this stack");
6961 // iterate over the oops in this oop, marking and pushing
6962 // the ones in CMS heap (i.e. in _span).
6963 new_oop->oop_iterate(_scanning_closure);
6964 // check if it's time to yield
6965 do_yield_check();
6966 }
6967 unsigned int after_count =
6968 GenCollectedHeap::heap()->total_collections();
6969 bool abort = (_before_count != after_count) ||
6970 _collector->should_abort_preclean();
6971 return abort ? 0 : size;
6972 }
6973
6974 void SurvivorSpacePrecleanClosure::do_yield_work() {
6975 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6976 "CMS thread should hold CMS token");
6977 assert_lock_strong(_bit_map->lock());
6978 DEBUG_ONLY(RememberKlassesChecker smx(false);)
6979 // Relinquish the bit map lock
6980 _bit_map->lock()->unlock();
6981 ConcurrentMarkSweepThread::desynchronize(true);
6982 ConcurrentMarkSweepThread::acknowledge_yield_request();
6983 _collector->stopTimer();
6984 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6985 if (PrintCMSStatistics != 0) {
6986 _collector->incrementYields();
6987 }
6988 _collector->icms_wait();
6989
6990 // See the comment in coordinator_yield()
6991 for (unsigned i = 0; i < CMSYieldSleepCount &&
6992 ConcurrentMarkSweepThread::should_yield() &&
6993 !CMSCollector::foregroundGCIsActive(); ++i) {
6994 os::sleep(Thread::current(), 1, false);
6995 ConcurrentMarkSweepThread::acknowledge_yield_request();
6996 }
6997
6998 ConcurrentMarkSweepThread::synchronize(true);
6999 _bit_map->lock()->lock_without_safepoint_check();
7000 _collector->startTimer();
7001 }
7002
7003 // This closure is used to rescan the marked objects on the dirty cards
7004 // in the mod union table and the card table proper. In the parallel
7005 // case, although the bitMap is shared, we do a single read so the
7006 // isMarked() query is "safe".
7007 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
7008 // Ignore mark word because we are running concurrent with mutators
7009 assert(p->is_oop_or_null(true), "expected an oop or null");
7010 HeapWord* addr = (HeapWord*)p;
7011 assert(_span.contains(addr), "we are scanning the CMS generation");
7012 bool is_obj_array = false;
7013 #ifdef DEBUG
7014 if (!_parallel) {
7015 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
7016 assert(_collector->overflow_list_is_empty(),
7017 "overflow list should be empty");
7018
7019 }
7020 #endif // DEBUG
7021 if (_bit_map->isMarked(addr)) {
7022 // Obj arrays are precisely marked, non-arrays are not;
7023 // so we scan objArrays precisely and non-arrays in their
7024 // entirety.
7025 if (p->is_objArray()) {
7026 is_obj_array = true;
7027 if (_parallel) {
7028 p->oop_iterate(_par_scan_closure, mr);
7029 } else {
7030 p->oop_iterate(_scan_closure, mr);
7031 }
7032 } else {
7033 if (_parallel) {
7034 p->oop_iterate(_par_scan_closure);
7035 } else {
7036 p->oop_iterate(_scan_closure);
7037 }
7038 }
7039 }
7040 #ifdef DEBUG
7041 if (!_parallel) {
7042 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
7043 assert(_collector->overflow_list_is_empty(),
7044 "overflow list should be empty");
7045
7046 }
7047 #endif // DEBUG
7048 return is_obj_array;
7049 }
7050
7051 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
7052 MemRegion span,
7053 CMSBitMap* bitMap, CMSMarkStack* markStack,
7054 CMSMarkStack* revisitStack,
7055 bool should_yield, bool verifying):
7056 _collector(collector),
7057 _span(span),
7058 _bitMap(bitMap),
7059 _mut(&collector->_modUnionTable),
7060 _markStack(markStack),
7061 _revisitStack(revisitStack),
7062 _yield(should_yield),
7063 _skipBits(0)
7064 {
7065 assert(_markStack->isEmpty(), "stack should be empty");
7066 _finger = _bitMap->startWord();
7067 _threshold = _finger;
7068 assert(_collector->_restart_addr == NULL, "Sanity check");
7069 assert(_span.contains(_finger), "Out of bounds _finger?");
7070 DEBUG_ONLY(_verifying = verifying;)
7071 }
7072
7073 void MarkFromRootsClosure::reset(HeapWord* addr) {
7074 assert(_markStack->isEmpty(), "would cause duplicates on stack");
7075 assert(_span.contains(addr), "Out of bounds _finger?");
7076 _finger = addr;
7077 _threshold = (HeapWord*)round_to(
7078 (intptr_t)_finger, CardTableModRefBS::card_size);
7079 }
7080
7081 // Should revisit to see if this should be restructured for
7082 // greater efficiency.
7083 bool MarkFromRootsClosure::do_bit(size_t offset) {
7084 if (_skipBits > 0) {
7085 _skipBits--;
7086 return true;
7087 }
7088 // convert offset into a HeapWord*
7089 HeapWord* addr = _bitMap->startWord() + offset;
7090 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
7091 "address out of range");
7092 assert(_bitMap->isMarked(addr), "tautology");
7093 if (_bitMap->isMarked(addr+1)) {
7094 // this is an allocated but not yet initialized object
7095 assert(_skipBits == 0, "tautology");
7096 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
7097 oop p = oop(addr);
7098 if (p->klass_or_null() == NULL || !p->is_parsable()) {
7099 DEBUG_ONLY(if (!_verifying) {)
7100 // We re-dirty the cards on which this object lies and increase
7101 // the _threshold so that we'll come back to scan this object
7102 // during the preclean or remark phase. (CMSCleanOnEnter)
7103 if (CMSCleanOnEnter) {
7104 size_t sz = _collector->block_size_using_printezis_bits(addr);
7105 HeapWord* end_card_addr = (HeapWord*)round_to(
7106 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7107 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7108 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7109 // Bump _threshold to end_card_addr; note that
7110 // _threshold cannot possibly exceed end_card_addr, anyhow.
7111 // This prevents future clearing of the card as the scan proceeds
7112 // to the right.
7113 assert(_threshold <= end_card_addr,
7114 "Because we are just scanning into this object");
7115 if (_threshold < end_card_addr) {
7116 _threshold = end_card_addr;
7117 }
7118 if (p->klass_or_null() != NULL) {
7119 // Redirty the range of cards...
7120 _mut->mark_range(redirty_range);
7121 } // ...else the setting of klass will dirty the card anyway.
7122 }
7123 DEBUG_ONLY(})
7124 return true;
7125 }
7126 }
7127 scanOopsInOop(addr);
7128 return true;
7129 }
7130
7131 // We take a break if we've been at this for a while,
7132 // so as to avoid monopolizing the locks involved.
7133 void MarkFromRootsClosure::do_yield_work() {
7134 // First give up the locks, then yield, then re-lock
7135 // We should probably use a constructor/destructor idiom to
7136 // do this unlock/lock or modify the MutexUnlocker class to
7137 // serve our purpose. XXX
7138 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7139 "CMS thread should hold CMS token");
7140 assert_lock_strong(_bitMap->lock());
7141 DEBUG_ONLY(RememberKlassesChecker mux(false);)
7142 _bitMap->lock()->unlock();
7143 ConcurrentMarkSweepThread::desynchronize(true);
7144 ConcurrentMarkSweepThread::acknowledge_yield_request();
7145 _collector->stopTimer();
7146 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7147 if (PrintCMSStatistics != 0) {
7148 _collector->incrementYields();
7149 }
7150 _collector->icms_wait();
7151
7152 // See the comment in coordinator_yield()
7153 for (unsigned i = 0; i < CMSYieldSleepCount &&
7154 ConcurrentMarkSweepThread::should_yield() &&
7155 !CMSCollector::foregroundGCIsActive(); ++i) {
7156 os::sleep(Thread::current(), 1, false);
7157 ConcurrentMarkSweepThread::acknowledge_yield_request();
7158 }
7159
7160 ConcurrentMarkSweepThread::synchronize(true);
7161 _bitMap->lock()->lock_without_safepoint_check();
7162 _collector->startTimer();
7163 }
7164
7165 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
7166 assert(_bitMap->isMarked(ptr), "expected bit to be set");
7167 assert(_markStack->isEmpty(),
7168 "should drain stack to limit stack usage");
7169 // convert ptr to an oop preparatory to scanning
7170 oop obj = oop(ptr);
7171 // Ignore mark word in verification below, since we
7172 // may be running concurrent with mutators.
7173 assert(obj->is_oop(true), "should be an oop");
7174 assert(_finger <= ptr, "_finger runneth ahead");
7175 // advance the finger to right end of this object
7176 _finger = ptr + obj->size();
7177 assert(_finger > ptr, "we just incremented it above");
7178 // On large heaps, it may take us some time to get through
7179 // the marking phase (especially if running iCMS). During
7180 // this time it's possible that a lot of mutations have
7181 // accumulated in the card table and the mod union table --
7182 // these mutation records are redundant until we have
7183 // actually traced into the corresponding card.
7184 // Here, we check whether advancing the finger would make
7185 // us cross into a new card, and if so clear corresponding
7186 // cards in the MUT (preclean them in the card-table in the
7187 // future).
7188
7189 DEBUG_ONLY(if (!_verifying) {)
7190 // The clean-on-enter optimization is disabled by default,
7191 // until we fix 6178663.
7192 if (CMSCleanOnEnter && (_finger > _threshold)) {
7193 // [_threshold, _finger) represents the interval
7194 // of cards to be cleared in MUT (or precleaned in card table).
7195 // The set of cards to be cleared is all those that overlap
7196 // with the interval [_threshold, _finger); note that
7197 // _threshold is always kept card-aligned but _finger isn't
7198 // always card-aligned.
7199 HeapWord* old_threshold = _threshold;
7200 assert(old_threshold == (HeapWord*)round_to(
7201 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7202 "_threshold should always be card-aligned");
7203 _threshold = (HeapWord*)round_to(
7204 (intptr_t)_finger, CardTableModRefBS::card_size);
7205 MemRegion mr(old_threshold, _threshold);
7206 assert(!mr.is_empty(), "Control point invariant");
7207 assert(_span.contains(mr), "Should clear within span");
7208 // XXX When _finger crosses from old gen into perm gen
7209 // we may be doing unnecessary cleaning; do better in the
7210 // future by detecting that condition and clearing fewer
7211 // MUT/CT entries.
7212 _mut->clear_range(mr);
7213 }
7214 DEBUG_ONLY(})
7215 // Note: the finger doesn't advance while we drain
7216 // the stack below.
7217 PushOrMarkClosure pushOrMarkClosure(_collector,
7218 _span, _bitMap, _markStack,
7219 _revisitStack,
7220 _finger, this);
7221 bool res = _markStack->push(obj);
7222 assert(res, "Empty non-zero size stack should have space for single push");
7223 while (!_markStack->isEmpty()) {
7224 oop new_oop = _markStack->pop();
7225 // Skip verifying header mark word below because we are
7226 // running concurrent with mutators.
7227 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7228 // now scan this oop's oops
7229 new_oop->oop_iterate(&pushOrMarkClosure);
7230 do_yield_check();
7231 }
7232 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
7233 }
7234
7235 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
7236 CMSCollector* collector, MemRegion span,
7237 CMSBitMap* bit_map,
7238 OopTaskQueue* work_queue,
7239 CMSMarkStack* overflow_stack,
7240 CMSMarkStack* revisit_stack,
7241 bool should_yield):
7242 _collector(collector),
7243 _whole_span(collector->_span),
7244 _span(span),
7245 _bit_map(bit_map),
7246 _mut(&collector->_modUnionTable),
7247 _work_queue(work_queue),
7248 _overflow_stack(overflow_stack),
7249 _revisit_stack(revisit_stack),
7250 _yield(should_yield),
7251 _skip_bits(0),
7252 _task(task)
7253 {
7254 assert(_work_queue->size() == 0, "work_queue should be empty");
7255 _finger = span.start();
7256 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
7257 assert(_span.contains(_finger), "Out of bounds _finger?");
7258 }
7259
7260 // Should revisit to see if this should be restructured for
7261 // greater efficiency.
7262 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
7263 if (_skip_bits > 0) {
7264 _skip_bits--;
7265 return true;
7266 }
7267 // convert offset into a HeapWord*
7268 HeapWord* addr = _bit_map->startWord() + offset;
7269 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
7270 "address out of range");
7271 assert(_bit_map->isMarked(addr), "tautology");
7272 if (_bit_map->isMarked(addr+1)) {
7273 // this is an allocated object that might not yet be initialized
7274 assert(_skip_bits == 0, "tautology");
7275 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
7276 oop p = oop(addr);
7277 if (p->klass_or_null() == NULL || !p->is_parsable()) {
7278 // in the case of Clean-on-Enter optimization, redirty card
7279 // and avoid clearing card by increasing the threshold.
7280 return true;
7281 }
7282 }
7283 scan_oops_in_oop(addr);
7284 return true;
7285 }
7286
7287 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
7288 assert(_bit_map->isMarked(ptr), "expected bit to be set");
7289 // Should we assert that our work queue is empty or
7290 // below some drain limit?
7291 assert(_work_queue->size() == 0,
7292 "should drain stack to limit stack usage");
7293 // convert ptr to an oop preparatory to scanning
7294 oop obj = oop(ptr);
7295 // Ignore mark word in verification below, since we
7296 // may be running concurrent with mutators.
7297 assert(obj->is_oop(true), "should be an oop");
7298 assert(_finger <= ptr, "_finger runneth ahead");
7299 // advance the finger to right end of this object
7300 _finger = ptr + obj->size();
7301 assert(_finger > ptr, "we just incremented it above");
7302 // On large heaps, it may take us some time to get through
7303 // the marking phase (especially if running iCMS). During
7304 // this time it's possible that a lot of mutations have
7305 // accumulated in the card table and the mod union table --
7306 // these mutation records are redundant until we have
7307 // actually traced into the corresponding card.
7308 // Here, we check whether advancing the finger would make
7309 // us cross into a new card, and if so clear corresponding
7310 // cards in the MUT (preclean them in the card-table in the
7311 // future).
7312
7313 // The clean-on-enter optimization is disabled by default,
7314 // until we fix 6178663.
7315 if (CMSCleanOnEnter && (_finger > _threshold)) {
7316 // [_threshold, _finger) represents the interval
7317 // of cards to be cleared in MUT (or precleaned in card table).
7318 // The set of cards to be cleared is all those that overlap
7319 // with the interval [_threshold, _finger); note that
7320 // _threshold is always kept card-aligned but _finger isn't
7321 // always card-aligned.
7322 HeapWord* old_threshold = _threshold;
7323 assert(old_threshold == (HeapWord*)round_to(
7324 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7325 "_threshold should always be card-aligned");
7326 _threshold = (HeapWord*)round_to(
7327 (intptr_t)_finger, CardTableModRefBS::card_size);
7328 MemRegion mr(old_threshold, _threshold);
7329 assert(!mr.is_empty(), "Control point invariant");
7330 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7331 // XXX When _finger crosses from old gen into perm gen
7332 // we may be doing unnecessary cleaning; do better in the
7333 // future by detecting that condition and clearing fewer
7334 // MUT/CT entries.
7335 _mut->clear_range(mr);
7336 }
7337
7338 // Note: the local finger doesn't advance while we drain
7339 // the stack below, but the global finger sure can and will.
7340 HeapWord** gfa = _task->global_finger_addr();
7341 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7342 _span, _bit_map,
7343 _work_queue,
7344 _overflow_stack,
7345 _revisit_stack,
7346 _finger,
7347 gfa, this);
7348 bool res = _work_queue->push(obj); // overflow could occur here
7349 assert(res, "Will hold once we use workqueues");
7350 while (true) {
7351 oop new_oop;
7352 if (!_work_queue->pop_local(new_oop)) {
7353 // We emptied our work_queue; check if there's stuff that can
7354 // be gotten from the overflow stack.
7355 if (CMSConcMarkingTask::get_work_from_overflow_stack(
7356 _overflow_stack, _work_queue)) {
7357 do_yield_check();
7358 continue;
7359 } else { // done
7360 break;
7361 }
7362 }
7363 // Skip verifying header mark word below because we are
7364 // running concurrent with mutators.
7365 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7366 // now scan this oop's oops
7367 new_oop->oop_iterate(&pushOrMarkClosure);
7368 do_yield_check();
7369 }
7370 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7371 }
7372
7373 // Yield in response to a request from VM Thread or
7374 // from mutators.
7375 void Par_MarkFromRootsClosure::do_yield_work() {
7376 assert(_task != NULL, "sanity");
7377 _task->yield();
7378 }
7379
7380 // A variant of the above used for verifying CMS marking work.
7381 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7382 MemRegion span,
7383 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7384 CMSMarkStack* mark_stack):
7385 _collector(collector),
7386 _span(span),
7387 _verification_bm(verification_bm),
7388 _cms_bm(cms_bm),
7389 _mark_stack(mark_stack),
7390 _pam_verify_closure(collector, span, verification_bm, cms_bm,
7391 mark_stack)
7392 {
7393 assert(_mark_stack->isEmpty(), "stack should be empty");
7394 _finger = _verification_bm->startWord();
7395 assert(_collector->_restart_addr == NULL, "Sanity check");
7396 assert(_span.contains(_finger), "Out of bounds _finger?");
7397 }
7398
7399 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7400 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7401 assert(_span.contains(addr), "Out of bounds _finger?");
7402 _finger = addr;
7403 }
7404
7405 // Should revisit to see if this should be restructured for
7406 // greater efficiency.
7407 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7408 // convert offset into a HeapWord*
7409 HeapWord* addr = _verification_bm->startWord() + offset;
7410 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7411 "address out of range");
7412 assert(_verification_bm->isMarked(addr), "tautology");
7413 assert(_cms_bm->isMarked(addr), "tautology");
7414
7415 assert(_mark_stack->isEmpty(),
7416 "should drain stack to limit stack usage");
7417 // convert addr to an oop preparatory to scanning
7418 oop obj = oop(addr);
7419 assert(obj->is_oop(), "should be an oop");
7420 assert(_finger <= addr, "_finger runneth ahead");
7421 // advance the finger to right end of this object
7422 _finger = addr + obj->size();
7423 assert(_finger > addr, "we just incremented it above");
7424 // Note: the finger doesn't advance while we drain
7425 // the stack below.
7426 bool res = _mark_stack->push(obj);
7427 assert(res, "Empty non-zero size stack should have space for single push");
7428 while (!_mark_stack->isEmpty()) {
7429 oop new_oop = _mark_stack->pop();
7430 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7431 // now scan this oop's oops
7432 new_oop->oop_iterate(&_pam_verify_closure);
7433 }
7434 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7435 return true;
7436 }
7437
7438 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7439 CMSCollector* collector, MemRegion span,
7440 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7441 CMSMarkStack* mark_stack):
7442 OopClosure(collector->ref_processor()),
7443 _collector(collector),
7444 _span(span),
7445 _verification_bm(verification_bm),
7446 _cms_bm(cms_bm),
7447 _mark_stack(mark_stack)
7448 { }
7449
7450 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7451 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7452
7453 // Upon stack overflow, we discard (part of) the stack,
7454 // remembering the least address amongst those discarded
7455 // in CMSCollector's _restart_address.
7456 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7457 // Remember the least grey address discarded
7458 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7459 _collector->lower_restart_addr(ra);
7460 _mark_stack->reset(); // discard stack contents
7461 _mark_stack->expand(); // expand the stack if possible
7462 }
7463
7464 void PushAndMarkVerifyClosure::do_oop(oop obj) {
7465 assert(obj->is_oop_or_null(), "expected an oop or NULL");
7466 HeapWord* addr = (HeapWord*)obj;
7467 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7468 // Oop lies in _span and isn't yet grey or black
7469 _verification_bm->mark(addr); // now grey
7470 if (!_cms_bm->isMarked(addr)) {
7471 oop(addr)->print();
7472 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
7473 addr);
7474 fatal("... aborting");
7475 }
7476
7477 if (!_mark_stack->push(obj)) { // stack overflow
7478 if (PrintCMSStatistics != 0) {
7479 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7480 SIZE_FORMAT, _mark_stack->capacity());
7481 }
7482 assert(_mark_stack->isFull(), "Else push should have succeeded");
7483 handle_stack_overflow(addr);
7484 }
7485 // anything including and to the right of _finger
7486 // will be scanned as we iterate over the remainder of the
7487 // bit map
7488 }
7489 }
7490
7491 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7492 MemRegion span,
7493 CMSBitMap* bitMap, CMSMarkStack* markStack,
7494 CMSMarkStack* revisitStack,
7495 HeapWord* finger, MarkFromRootsClosure* parent) :
7496 KlassRememberingOopClosure(collector, collector->ref_processor(), revisitStack),
7497 _span(span),
7498 _bitMap(bitMap),
7499 _markStack(markStack),
7500 _finger(finger),
7501 _parent(parent)
7502 { }
7503
7504 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7505 MemRegion span,
7506 CMSBitMap* bit_map,
7507 OopTaskQueue* work_queue,
7508 CMSMarkStack* overflow_stack,
7509 CMSMarkStack* revisit_stack,
7510 HeapWord* finger,
7511 HeapWord** global_finger_addr,
7512 Par_MarkFromRootsClosure* parent) :
7513 Par_KlassRememberingOopClosure(collector,
7514 collector->ref_processor(),
7515 revisit_stack),
7516 _whole_span(collector->_span),
7517 _span(span),
7518 _bit_map(bit_map),
7519 _work_queue(work_queue),
7520 _overflow_stack(overflow_stack),
7521 _finger(finger),
7522 _global_finger_addr(global_finger_addr),
7523 _parent(parent)
7524 { }
7525
7526 // Assumes thread-safe access by callers, who are
7527 // responsible for mutual exclusion.
7528 void CMSCollector::lower_restart_addr(HeapWord* low) {
7529 assert(_span.contains(low), "Out of bounds addr");
7530 if (_restart_addr == NULL) {
7531 _restart_addr = low;
7532 } else {
7533 _restart_addr = MIN2(_restart_addr, low);
7534 }
7535 }
7536
7537 // Upon stack overflow, we discard (part of) the stack,
7538 // remembering the least address amongst those discarded
7539 // in CMSCollector's _restart_address.
7540 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7541 // Remember the least grey address discarded
7542 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7543 _collector->lower_restart_addr(ra);
7544 _markStack->reset(); // discard stack contents
7545 _markStack->expand(); // expand the stack if possible
7546 }
7547
7548 // Upon stack overflow, we discard (part of) the stack,
7549 // remembering the least address amongst those discarded
7550 // in CMSCollector's _restart_address.
7551 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7552 // We need to do this under a mutex to prevent other
7553 // workers from interfering with the work done below.
7554 MutexLockerEx ml(_overflow_stack->par_lock(),
7555 Mutex::_no_safepoint_check_flag);
7556 // Remember the least grey address discarded
7557 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7558 _collector->lower_restart_addr(ra);
7559 _overflow_stack->reset(); // discard stack contents
7560 _overflow_stack->expand(); // expand the stack if possible
7561 }
7562
7563 void PushOrMarkClosure::do_oop(oop obj) {
7564 // Ignore mark word because we are running concurrent with mutators.
7565 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7566 HeapWord* addr = (HeapWord*)obj;
7567 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7568 // Oop lies in _span and isn't yet grey or black
7569 _bitMap->mark(addr); // now grey
7570 if (addr < _finger) {
7571 // the bit map iteration has already either passed, or
7572 // sampled, this bit in the bit map; we'll need to
7573 // use the marking stack to scan this oop's oops.
7574 bool simulate_overflow = false;
7575 NOT_PRODUCT(
7576 if (CMSMarkStackOverflowALot &&
7577 _collector->simulate_overflow()) {
7578 // simulate a stack overflow
7579 simulate_overflow = true;
7580 }
7581 )
7582 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7583 if (PrintCMSStatistics != 0) {
7584 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7585 SIZE_FORMAT, _markStack->capacity());
7586 }
7587 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7588 handle_stack_overflow(addr);
7589 }
7590 }
7591 // anything including and to the right of _finger
7592 // will be scanned as we iterate over the remainder of the
7593 // bit map
7594 do_yield_check();
7595 }
7596 }
7597
7598 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7599 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7600
7601 void Par_PushOrMarkClosure::do_oop(oop obj) {
7602 // Ignore mark word because we are running concurrent with mutators.
7603 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7604 HeapWord* addr = (HeapWord*)obj;
7605 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7606 // Oop lies in _span and isn't yet grey or black
7607 // We read the global_finger (volatile read) strictly after marking oop
7608 bool res = _bit_map->par_mark(addr); // now grey
7609 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7610 // Should we push this marked oop on our stack?
7611 // -- if someone else marked it, nothing to do
7612 // -- if target oop is above global finger nothing to do
7613 // -- if target oop is in chunk and above local finger
7614 // then nothing to do
7615 // -- else push on work queue
7616 if ( !res // someone else marked it, they will deal with it
7617 || (addr >= *gfa) // will be scanned in a later task
7618 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7619 return;
7620 }
7621 // the bit map iteration has already either passed, or
7622 // sampled, this bit in the bit map; we'll need to
7623 // use the marking stack to scan this oop's oops.
7624 bool simulate_overflow = false;
7625 NOT_PRODUCT(
7626 if (CMSMarkStackOverflowALot &&
7627 _collector->simulate_overflow()) {
7628 // simulate a stack overflow
7629 simulate_overflow = true;
7630 }
7631 )
7632 if (simulate_overflow ||
7633 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7634 // stack overflow
7635 if (PrintCMSStatistics != 0) {
7636 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7637 SIZE_FORMAT, _overflow_stack->capacity());
7638 }
7639 // We cannot assert that the overflow stack is full because
7640 // it may have been emptied since.
7641 assert(simulate_overflow ||
7642 _work_queue->size() == _work_queue->max_elems(),
7643 "Else push should have succeeded");
7644 handle_stack_overflow(addr);
7645 }
7646 do_yield_check();
7647 }
7648 }
7649
7650 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7651 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7652
7653 KlassRememberingOopClosure::KlassRememberingOopClosure(CMSCollector* collector,
7654 ReferenceProcessor* rp,
7655 CMSMarkStack* revisit_stack) :
7656 OopClosure(rp),
7657 _collector(collector),
7658 _revisit_stack(revisit_stack),
7659 _should_remember_klasses(collector->should_unload_classes()) {}
7660
7661 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7662 MemRegion span,
7663 ReferenceProcessor* rp,
7664 CMSBitMap* bit_map,
7665 CMSBitMap* mod_union_table,
7666 CMSMarkStack* mark_stack,
7667 CMSMarkStack* revisit_stack,
7668 bool concurrent_precleaning):
7669 KlassRememberingOopClosure(collector, rp, revisit_stack),
7670 _span(span),
7671 _bit_map(bit_map),
7672 _mod_union_table(mod_union_table),
7673 _mark_stack(mark_stack),
7674 _concurrent_precleaning(concurrent_precleaning)
7675 {
7676 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7677 }
7678
7679 // Grey object rescan during pre-cleaning and second checkpoint phases --
7680 // the non-parallel version (the parallel version appears further below.)
7681 void PushAndMarkClosure::do_oop(oop obj) {
7682 // Ignore mark word verification. If during concurrent precleaning,
7683 // the object monitor may be locked. If during the checkpoint
7684 // phases, the object may already have been reached by a different
7685 // path and may be at the end of the global overflow list (so
7686 // the mark word may be NULL).
7687 assert(obj->is_oop_or_null(true /* ignore mark word */),
7688 "expected an oop or NULL");
7689 HeapWord* addr = (HeapWord*)obj;
7690 // Check if oop points into the CMS generation
7691 // and is not marked
7692 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7693 // a white object ...
7694 _bit_map->mark(addr); // ... now grey
7695 // push on the marking stack (grey set)
7696 bool simulate_overflow = false;
7697 NOT_PRODUCT(
7698 if (CMSMarkStackOverflowALot &&
7699 _collector->simulate_overflow()) {
7700 // simulate a stack overflow
7701 simulate_overflow = true;
7702 }
7703 )
7704 if (simulate_overflow || !_mark_stack->push(obj)) {
7705 if (_concurrent_precleaning) {
7706 // During precleaning we can just dirty the appropriate card(s)
7707 // in the mod union table, thus ensuring that the object remains
7708 // in the grey set and continue. In the case of object arrays
7709 // we need to dirty all of the cards that the object spans,
7710 // since the rescan of object arrays will be limited to the
7711 // dirty cards.
7712 // Note that no one can be intefering with us in this action
7713 // of dirtying the mod union table, so no locking or atomics
7714 // are required.
7715 if (obj->is_objArray()) {
7716 size_t sz = obj->size();
7717 HeapWord* end_card_addr = (HeapWord*)round_to(
7718 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7719 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7720 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7721 _mod_union_table->mark_range(redirty_range);
7722 } else {
7723 _mod_union_table->mark(addr);
7724 }
7725 _collector->_ser_pmc_preclean_ovflw++;
7726 } else {
7727 // During the remark phase, we need to remember this oop
7728 // in the overflow list.
7729 _collector->push_on_overflow_list(obj);
7730 _collector->_ser_pmc_remark_ovflw++;
7731 }
7732 }
7733 }
7734 }
7735
7736 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7737 MemRegion span,
7738 ReferenceProcessor* rp,
7739 CMSBitMap* bit_map,
7740 OopTaskQueue* work_queue,
7741 CMSMarkStack* revisit_stack):
7742 Par_KlassRememberingOopClosure(collector, rp, revisit_stack),
7743 _span(span),
7744 _bit_map(bit_map),
7745 _work_queue(work_queue)
7746 {
7747 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7748 }
7749
7750 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7751 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7752
7753 // Grey object rescan during second checkpoint phase --
7754 // the parallel version.
7755 void Par_PushAndMarkClosure::do_oop(oop obj) {
7756 // In the assert below, we ignore the mark word because
7757 // this oop may point to an already visited object that is
7758 // on the overflow stack (in which case the mark word has
7759 // been hijacked for chaining into the overflow stack --
7760 // if this is the last object in the overflow stack then
7761 // its mark word will be NULL). Because this object may
7762 // have been subsequently popped off the global overflow
7763 // stack, and the mark word possibly restored to the prototypical
7764 // value, by the time we get to examined this failing assert in
7765 // the debugger, is_oop_or_null(false) may subsequently start
7766 // to hold.
7767 assert(obj->is_oop_or_null(true),
7768 "expected an oop or NULL");
7769 HeapWord* addr = (HeapWord*)obj;
7770 // Check if oop points into the CMS generation
7771 // and is not marked
7772 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7773 // a white object ...
7774 // If we manage to "claim" the object, by being the
7775 // first thread to mark it, then we push it on our
7776 // marking stack
7777 if (_bit_map->par_mark(addr)) { // ... now grey
7778 // push on work queue (grey set)
7779 bool simulate_overflow = false;
7780 NOT_PRODUCT(
7781 if (CMSMarkStackOverflowALot &&
7782 _collector->par_simulate_overflow()) {
7783 // simulate a stack overflow
7784 simulate_overflow = true;
7785 }
7786 )
7787 if (simulate_overflow || !_work_queue->push(obj)) {
7788 _collector->par_push_on_overflow_list(obj);
7789 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7790 }
7791 } // Else, some other thread got there first
7792 }
7793 }
7794
7795 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7796 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7797
7798 void PushAndMarkClosure::remember_mdo(DataLayout* v) {
7799 // TBD
7800 }
7801
7802 void Par_PushAndMarkClosure::remember_mdo(DataLayout* v) {
7803 // TBD
7804 }
7805
7806 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7807 DEBUG_ONLY(RememberKlassesChecker mux(false);)
7808 Mutex* bml = _collector->bitMapLock();
7809 assert_lock_strong(bml);
7810 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7811 "CMS thread should hold CMS token");
7812
7813 bml->unlock();
7814 ConcurrentMarkSweepThread::desynchronize(true);
7815
7816 ConcurrentMarkSweepThread::acknowledge_yield_request();
7817
7818 _collector->stopTimer();
7819 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7820 if (PrintCMSStatistics != 0) {
7821 _collector->incrementYields();
7822 }
7823 _collector->icms_wait();
7824
7825 // See the comment in coordinator_yield()
7826 for (unsigned i = 0; i < CMSYieldSleepCount &&
7827 ConcurrentMarkSweepThread::should_yield() &&
7828 !CMSCollector::foregroundGCIsActive(); ++i) {
7829 os::sleep(Thread::current(), 1, false);
7830 ConcurrentMarkSweepThread::acknowledge_yield_request();
7831 }
7832
7833 ConcurrentMarkSweepThread::synchronize(true);
7834 bml->lock();
7835
7836 _collector->startTimer();
7837 }
7838
7839 bool CMSPrecleanRefsYieldClosure::should_return() {
7840 if (ConcurrentMarkSweepThread::should_yield()) {
7841 do_yield_work();
7842 }
7843 return _collector->foregroundGCIsActive();
7844 }
7845
7846 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7847 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7848 "mr should be aligned to start at a card boundary");
7849 // We'd like to assert:
7850 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7851 // "mr should be a range of cards");
7852 // However, that would be too strong in one case -- the last
7853 // partition ends at _unallocated_block which, in general, can be
7854 // an arbitrary boundary, not necessarily card aligned.
7855 if (PrintCMSStatistics != 0) {
7856 _num_dirty_cards +=
7857 mr.word_size()/CardTableModRefBS::card_size_in_words;
7858 }
7859 _space->object_iterate_mem(mr, &_scan_cl);
7860 }
7861
7862 SweepClosure::SweepClosure(CMSCollector* collector,
7863 ConcurrentMarkSweepGeneration* g,
7864 CMSBitMap* bitMap, bool should_yield) :
7865 _collector(collector),
7866 _g(g),
7867 _sp(g->cmsSpace()),
7868 _limit(_sp->sweep_limit()),
7869 _freelistLock(_sp->freelistLock()),
7870 _bitMap(bitMap),
7871 _yield(should_yield),
7872 _inFreeRange(false), // No free range at beginning of sweep
7873 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7874 _lastFreeRangeCoalesced(false),
7875 _freeFinger(g->used_region().start())
7876 {
7877 NOT_PRODUCT(
7878 _numObjectsFreed = 0;
7879 _numWordsFreed = 0;
7880 _numObjectsLive = 0;
7881 _numWordsLive = 0;
7882 _numObjectsAlreadyFree = 0;
7883 _numWordsAlreadyFree = 0;
7884 _last_fc = NULL;
7885
7886 _sp->initializeIndexedFreeListArrayReturnedBytes();
7887 _sp->dictionary()->initializeDictReturnedBytes();
7888 )
7889 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7890 "sweep _limit out of bounds");
7891 if (CMSTraceSweeper) {
7892 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
7893 _limit);
7894 }
7895 }
7896
7897 void SweepClosure::print_on(outputStream* st) const {
7898 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7899 _sp->bottom(), _sp->end());
7900 tty->print_cr("_limit = " PTR_FORMAT, _limit);
7901 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger);
7902 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);)
7903 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7904 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7905 }
7906
7907 #ifndef PRODUCT
7908 // Assertion checking only: no useful work in product mode --
7909 // however, if any of the flags below become product flags,
7910 // you may need to review this code to see if it needs to be
7911 // enabled in product mode.
7912 SweepClosure::~SweepClosure() {
7913 assert_lock_strong(_freelistLock);
7914 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7915 "sweep _limit out of bounds");
7916 if (inFreeRange()) {
7917 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
7918 print();
7919 ShouldNotReachHere();
7920 }
7921 if (Verbose && PrintGC) {
7922 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes",
7923 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7924 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
7925 SIZE_FORMAT" bytes "
7926 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
7927 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7928 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7929 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
7930 * sizeof(HeapWord);
7931 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
7932
7933 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7934 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7935 size_t dictReturnedBytes = _sp->dictionary()->sumDictReturnedBytes();
7936 size_t returnedBytes = indexListReturnedBytes + dictReturnedBytes;
7937 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returnedBytes);
7938 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
7939 indexListReturnedBytes);
7940 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
7941 dictReturnedBytes);
7942 }
7943 }
7944 if (CMSTraceSweeper) {
7945 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
7946 _limit);
7947 }
7948 }
7949 #endif // PRODUCT
7950
7951 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7952 bool freeRangeInFreeLists) {
7953 if (CMSTraceSweeper) {
7954 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n",
7955 freeFinger, freeRangeInFreeLists);
7956 }
7957 assert(!inFreeRange(), "Trampling existing free range");
7958 set_inFreeRange(true);
7959 set_lastFreeRangeCoalesced(false);
7960
7961 set_freeFinger(freeFinger);
7962 set_freeRangeInFreeLists(freeRangeInFreeLists);
7963 if (CMSTestInFreeList) {
7964 if (freeRangeInFreeLists) {
7965 FreeChunk* fc = (FreeChunk*) freeFinger;
7966 assert(fc->isFree(), "A chunk on the free list should be free.");
7967 assert(fc->size() > 0, "Free range should have a size");
7968 assert(_sp->verifyChunkInFreeLists(fc), "Chunk is not in free lists");
7969 }
7970 }
7971 }
7972
7973 // Note that the sweeper runs concurrently with mutators. Thus,
7974 // it is possible for direct allocation in this generation to happen
7975 // in the middle of the sweep. Note that the sweeper also coalesces
7976 // contiguous free blocks. Thus, unless the sweeper and the allocator
7977 // synchronize appropriately freshly allocated blocks may get swept up.
7978 // This is accomplished by the sweeper locking the free lists while
7979 // it is sweeping. Thus blocks that are determined to be free are
7980 // indeed free. There is however one additional complication:
7981 // blocks that have been allocated since the final checkpoint and
7982 // mark, will not have been marked and so would be treated as
7983 // unreachable and swept up. To prevent this, the allocator marks
7984 // the bit map when allocating during the sweep phase. This leads,
7985 // however, to a further complication -- objects may have been allocated
7986 // but not yet initialized -- in the sense that the header isn't yet
7987 // installed. The sweeper can not then determine the size of the block
7988 // in order to skip over it. To deal with this case, we use a technique
7989 // (due to Printezis) to encode such uninitialized block sizes in the
7990 // bit map. Since the bit map uses a bit per every HeapWord, but the
7991 // CMS generation has a minimum object size of 3 HeapWords, it follows
7992 // that "normal marks" won't be adjacent in the bit map (there will
7993 // always be at least two 0 bits between successive 1 bits). We make use
7994 // of these "unused" bits to represent uninitialized blocks -- the bit
7995 // corresponding to the start of the uninitialized object and the next
7996 // bit are both set. Finally, a 1 bit marks the end of the object that
7997 // started with the two consecutive 1 bits to indicate its potentially
7998 // uninitialized state.
7999
8000 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
8001 FreeChunk* fc = (FreeChunk*)addr;
8002 size_t res;
8003
8004 // Check if we are done sweeping. Below we check "addr >= _limit" rather
8005 // than "addr == _limit" because although _limit was a block boundary when
8006 // we started the sweep, it may no longer be one because heap expansion
8007 // may have caused us to coalesce the block ending at the address _limit
8008 // with a newly expanded chunk (this happens when _limit was set to the
8009 // previous _end of the space), so we may have stepped past _limit:
8010 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
8011 if (addr >= _limit) { // we have swept up to or past the limit: finish up
8012 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8013 "sweep _limit out of bounds");
8014 assert(addr < _sp->end(), "addr out of bounds");
8015 // Flush any free range we might be holding as a single
8016 // coalesced chunk to the appropriate free list.
8017 if (inFreeRange()) {
8018 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
8019 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger()));
8020 flush_cur_free_chunk(freeFinger(),
8021 pointer_delta(addr, freeFinger()));
8022 if (CMSTraceSweeper) {
8023 gclog_or_tty->print("Sweep: last chunk: ");
8024 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") "
8025 "[coalesced:"SIZE_FORMAT"]\n",
8026 freeFinger(), pointer_delta(addr, freeFinger()),
8027 lastFreeRangeCoalesced());
8028 }
8029 }
8030
8031 // help the iterator loop finish
8032 return pointer_delta(_sp->end(), addr);
8033 }
8034
8035 assert(addr < _limit, "sweep invariant");
8036 // check if we should yield
8037 do_yield_check(addr);
8038 if (fc->isFree()) {
8039 // Chunk that is already free
8040 res = fc->size();
8041 do_already_free_chunk(fc);
8042 debug_only(_sp->verifyFreeLists());
8043 // If we flush the chunk at hand in lookahead_and_flush()
8044 // and it's coalesced with a preceding chunk, then the
8045 // process of "mangling" the payload of the coalesced block
8046 // will cause erasure of the size information from the
8047 // (erstwhile) header of all the coalesced blocks but the
8048 // first, so the first disjunct in the assert will not hold
8049 // in that specific case (in which case the second disjunct
8050 // will hold).
8051 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
8052 "Otherwise the size info doesn't change at this step");
8053 NOT_PRODUCT(
8054 _numObjectsAlreadyFree++;
8055 _numWordsAlreadyFree += res;
8056 )
8057 NOT_PRODUCT(_last_fc = fc;)
8058 } else if (!_bitMap->isMarked(addr)) {
8059 // Chunk is fresh garbage
8060 res = do_garbage_chunk(fc);
8061 debug_only(_sp->verifyFreeLists());
8062 NOT_PRODUCT(
8063 _numObjectsFreed++;
8064 _numWordsFreed += res;
8065 )
8066 } else {
8067 // Chunk that is alive.
8068 res = do_live_chunk(fc);
8069 debug_only(_sp->verifyFreeLists());
8070 NOT_PRODUCT(
8071 _numObjectsLive++;
8072 _numWordsLive += res;
8073 )
8074 }
8075 return res;
8076 }
8077
8078 // For the smart allocation, record following
8079 // split deaths - a free chunk is removed from its free list because
8080 // it is being split into two or more chunks.
8081 // split birth - a free chunk is being added to its free list because
8082 // a larger free chunk has been split and resulted in this free chunk.
8083 // coal death - a free chunk is being removed from its free list because
8084 // it is being coalesced into a large free chunk.
8085 // coal birth - a free chunk is being added to its free list because
8086 // it was created when two or more free chunks where coalesced into
8087 // this free chunk.
8088 //
8089 // These statistics are used to determine the desired number of free
8090 // chunks of a given size. The desired number is chosen to be relative
8091 // to the end of a CMS sweep. The desired number at the end of a sweep
8092 // is the
8093 // count-at-end-of-previous-sweep (an amount that was enough)
8094 // - count-at-beginning-of-current-sweep (the excess)
8095 // + split-births (gains in this size during interval)
8096 // - split-deaths (demands on this size during interval)
8097 // where the interval is from the end of one sweep to the end of the
8098 // next.
8099 //
8100 // When sweeping the sweeper maintains an accumulated chunk which is
8101 // the chunk that is made up of chunks that have been coalesced. That
8102 // will be termed the left-hand chunk. A new chunk of garbage that
8103 // is being considered for coalescing will be referred to as the
8104 // right-hand chunk.
8105 //
8106 // When making a decision on whether to coalesce a right-hand chunk with
8107 // the current left-hand chunk, the current count vs. the desired count
8108 // of the left-hand chunk is considered. Also if the right-hand chunk
8109 // is near the large chunk at the end of the heap (see
8110 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
8111 // left-hand chunk is coalesced.
8112 //
8113 // When making a decision about whether to split a chunk, the desired count
8114 // vs. the current count of the candidate to be split is also considered.
8115 // If the candidate is underpopulated (currently fewer chunks than desired)
8116 // a chunk of an overpopulated (currently more chunks than desired) size may
8117 // be chosen. The "hint" associated with a free list, if non-null, points
8118 // to a free list which may be overpopulated.
8119 //
8120
8121 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
8122 const size_t size = fc->size();
8123 // Chunks that cannot be coalesced are not in the
8124 // free lists.
8125 if (CMSTestInFreeList && !fc->cantCoalesce()) {
8126 assert(_sp->verifyChunkInFreeLists(fc),
8127 "free chunk should be in free lists");
8128 }
8129 // a chunk that is already free, should not have been
8130 // marked in the bit map
8131 HeapWord* const addr = (HeapWord*) fc;
8132 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
8133 // Verify that the bit map has no bits marked between
8134 // addr and purported end of this block.
8135 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8136
8137 // Some chunks cannot be coalesced under any circumstances.
8138 // See the definition of cantCoalesce().
8139 if (!fc->cantCoalesce()) {
8140 // This chunk can potentially be coalesced.
8141 if (_sp->adaptive_freelists()) {
8142 // All the work is done in
8143 do_post_free_or_garbage_chunk(fc, size);
8144 } else { // Not adaptive free lists
8145 // this is a free chunk that can potentially be coalesced by the sweeper;
8146 if (!inFreeRange()) {
8147 // if the next chunk is a free block that can't be coalesced
8148 // it doesn't make sense to remove this chunk from the free lists
8149 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
8150 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
8151 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
8152 nextChunk->isFree() && // ... which is free...
8153 nextChunk->cantCoalesce()) { // ... but can't be coalesced
8154 // nothing to do
8155 } else {
8156 // Potentially the start of a new free range:
8157 // Don't eagerly remove it from the free lists.
8158 // No need to remove it if it will just be put
8159 // back again. (Also from a pragmatic point of view
8160 // if it is a free block in a region that is beyond
8161 // any allocated blocks, an assertion will fail)
8162 // Remember the start of a free run.
8163 initialize_free_range(addr, true);
8164 // end - can coalesce with next chunk
8165 }
8166 } else {
8167 // the midst of a free range, we are coalescing
8168 print_free_block_coalesced(fc);
8169 if (CMSTraceSweeper) {
8170 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size);
8171 }
8172 // remove it from the free lists
8173 _sp->removeFreeChunkFromFreeLists(fc);
8174 set_lastFreeRangeCoalesced(true);
8175 // If the chunk is being coalesced and the current free range is
8176 // in the free lists, remove the current free range so that it
8177 // will be returned to the free lists in its entirety - all
8178 // the coalesced pieces included.
8179 if (freeRangeInFreeLists()) {
8180 FreeChunk* ffc = (FreeChunk*) freeFinger();
8181 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8182 "Size of free range is inconsistent with chunk size.");
8183 if (CMSTestInFreeList) {
8184 assert(_sp->verifyChunkInFreeLists(ffc),
8185 "free range is not in free lists");
8186 }
8187 _sp->removeFreeChunkFromFreeLists(ffc);
8188 set_freeRangeInFreeLists(false);
8189 }
8190 }
8191 }
8192 // Note that if the chunk is not coalescable (the else arm
8193 // below), we unconditionally flush, without needing to do
8194 // a "lookahead," as we do below.
8195 if (inFreeRange()) lookahead_and_flush(fc, size);
8196 } else {
8197 // Code path common to both original and adaptive free lists.
8198
8199 // cant coalesce with previous block; this should be treated
8200 // as the end of a free run if any
8201 if (inFreeRange()) {
8202 // we kicked some butt; time to pick up the garbage
8203 assert(freeFinger() < addr, "freeFinger points too high");
8204 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8205 }
8206 // else, nothing to do, just continue
8207 }
8208 }
8209
8210 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
8211 // This is a chunk of garbage. It is not in any free list.
8212 // Add it to a free list or let it possibly be coalesced into
8213 // a larger chunk.
8214 HeapWord* const addr = (HeapWord*) fc;
8215 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8216
8217 if (_sp->adaptive_freelists()) {
8218 // Verify that the bit map has no bits marked between
8219 // addr and purported end of just dead object.
8220 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8221
8222 do_post_free_or_garbage_chunk(fc, size);
8223 } else {
8224 if (!inFreeRange()) {
8225 // start of a new free range
8226 assert(size > 0, "A free range should have a size");
8227 initialize_free_range(addr, false);
8228 } else {
8229 // this will be swept up when we hit the end of the
8230 // free range
8231 if (CMSTraceSweeper) {
8232 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size);
8233 }
8234 // If the chunk is being coalesced and the current free range is
8235 // in the free lists, remove the current free range so that it
8236 // will be returned to the free lists in its entirety - all
8237 // the coalesced pieces included.
8238 if (freeRangeInFreeLists()) {
8239 FreeChunk* ffc = (FreeChunk*)freeFinger();
8240 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8241 "Size of free range is inconsistent with chunk size.");
8242 if (CMSTestInFreeList) {
8243 assert(_sp->verifyChunkInFreeLists(ffc),
8244 "free range is not in free lists");
8245 }
8246 _sp->removeFreeChunkFromFreeLists(ffc);
8247 set_freeRangeInFreeLists(false);
8248 }
8249 set_lastFreeRangeCoalesced(true);
8250 }
8251 // this will be swept up when we hit the end of the free range
8252
8253 // Verify that the bit map has no bits marked between
8254 // addr and purported end of just dead object.
8255 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8256 }
8257 assert(_limit >= addr + size,
8258 "A freshly garbage chunk can't possibly straddle over _limit");
8259 if (inFreeRange()) lookahead_and_flush(fc, size);
8260 return size;
8261 }
8262
8263 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
8264 HeapWord* addr = (HeapWord*) fc;
8265 // The sweeper has just found a live object. Return any accumulated
8266 // left hand chunk to the free lists.
8267 if (inFreeRange()) {
8268 assert(freeFinger() < addr, "freeFinger points too high");
8269 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8270 }
8271
8272 // This object is live: we'd normally expect this to be
8273 // an oop, and like to assert the following:
8274 // assert(oop(addr)->is_oop(), "live block should be an oop");
8275 // However, as we commented above, this may be an object whose
8276 // header hasn't yet been initialized.
8277 size_t size;
8278 assert(_bitMap->isMarked(addr), "Tautology for this control point");
8279 if (_bitMap->isMarked(addr + 1)) {
8280 // Determine the size from the bit map, rather than trying to
8281 // compute it from the object header.
8282 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
8283 size = pointer_delta(nextOneAddr + 1, addr);
8284 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
8285 "alignment problem");
8286
8287 #ifdef DEBUG
8288 if (oop(addr)->klass_or_null() != NULL &&
8289 ( !_collector->should_unload_classes()
8290 || (oop(addr)->is_parsable()) &&
8291 oop(addr)->is_conc_safe())) {
8292 // Ignore mark word because we are running concurrent with mutators
8293 assert(oop(addr)->is_oop(true), "live block should be an oop");
8294 // is_conc_safe is checked before performing this assertion
8295 // because an object that is not is_conc_safe may yet have
8296 // the return from size() correct.
8297 assert(size ==
8298 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
8299 "P-mark and computed size do not agree");
8300 }
8301 #endif
8302
8303 } else {
8304 // This should be an initialized object that's alive.
8305 assert(oop(addr)->klass_or_null() != NULL &&
8306 (!_collector->should_unload_classes()
8307 || oop(addr)->is_parsable()),
8308 "Should be an initialized object");
8309 // Note that there are objects used during class redefinition,
8310 // e.g. merge_cp in VM_RedefineClasses::merge_cp_and_rewrite(),
8311 // which are discarded with their is_conc_safe state still
8312 // false. These object may be floating garbage so may be
8313 // seen here. If they are floating garbage their size
8314 // should be attainable from their klass. Do not that
8315 // is_conc_safe() is true for oop(addr).
8316 // Ignore mark word because we are running concurrent with mutators
8317 assert(oop(addr)->is_oop(true), "live block should be an oop");
8318 // Verify that the bit map has no bits marked between
8319 // addr and purported end of this block.
8320 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8321 assert(size >= 3, "Necessary for Printezis marks to work");
8322 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
8323 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
8324 }
8325 return size;
8326 }
8327
8328 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
8329 size_t chunkSize) {
8330 // do_post_free_or_garbage_chunk() should only be called in the case
8331 // of the adaptive free list allocator.
8332 const bool fcInFreeLists = fc->isFree();
8333 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
8334 assert((HeapWord*)fc <= _limit, "sweep invariant");
8335 if (CMSTestInFreeList && fcInFreeLists) {
8336 assert(_sp->verifyChunkInFreeLists(fc), "free chunk is not in free lists");
8337 }
8338
8339 if (CMSTraceSweeper) {
8340 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
8341 }
8342
8343 HeapWord* const fc_addr = (HeapWord*) fc;
8344
8345 bool coalesce;
8346 const size_t left = pointer_delta(fc_addr, freeFinger());
8347 const size_t right = chunkSize;
8348 switch (FLSCoalescePolicy) {
8349 // numeric value forms a coalition aggressiveness metric
8350 case 0: { // never coalesce
8351 coalesce = false;
8352 break;
8353 }
8354 case 1: { // coalesce if left & right chunks on overpopulated lists
8355 coalesce = _sp->coalOverPopulated(left) &&
8356 _sp->coalOverPopulated(right);
8357 break;
8358 }
8359 case 2: { // coalesce if left chunk on overpopulated list (default)
8360 coalesce = _sp->coalOverPopulated(left);
8361 break;
8362 }
8363 case 3: { // coalesce if left OR right chunk on overpopulated list
8364 coalesce = _sp->coalOverPopulated(left) ||
8365 _sp->coalOverPopulated(right);
8366 break;
8367 }
8368 case 4: { // always coalesce
8369 coalesce = true;
8370 break;
8371 }
8372 default:
8373 ShouldNotReachHere();
8374 }
8375
8376 // Should the current free range be coalesced?
8377 // If the chunk is in a free range and either we decided to coalesce above
8378 // or the chunk is near the large block at the end of the heap
8379 // (isNearLargestChunk() returns true), then coalesce this chunk.
8380 const bool doCoalesce = inFreeRange()
8381 && (coalesce || _g->isNearLargestChunk(fc_addr));
8382 if (doCoalesce) {
8383 // Coalesce the current free range on the left with the new
8384 // chunk on the right. If either is on a free list,
8385 // it must be removed from the list and stashed in the closure.
8386 if (freeRangeInFreeLists()) {
8387 FreeChunk* const ffc = (FreeChunk*)freeFinger();
8388 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
8389 "Size of free range is inconsistent with chunk size.");
8390 if (CMSTestInFreeList) {
8391 assert(_sp->verifyChunkInFreeLists(ffc),
8392 "Chunk is not in free lists");
8393 }
8394 _sp->coalDeath(ffc->size());
8395 _sp->removeFreeChunkFromFreeLists(ffc);
8396 set_freeRangeInFreeLists(false);
8397 }
8398 if (fcInFreeLists) {
8399 _sp->coalDeath(chunkSize);
8400 assert(fc->size() == chunkSize,
8401 "The chunk has the wrong size or is not in the free lists");
8402 _sp->removeFreeChunkFromFreeLists(fc);
8403 }
8404 set_lastFreeRangeCoalesced(true);
8405 print_free_block_coalesced(fc);
8406 } else { // not in a free range and/or should not coalesce
8407 // Return the current free range and start a new one.
8408 if (inFreeRange()) {
8409 // In a free range but cannot coalesce with the right hand chunk.
8410 // Put the current free range into the free lists.
8411 flush_cur_free_chunk(freeFinger(),
8412 pointer_delta(fc_addr, freeFinger()));
8413 }
8414 // Set up for new free range. Pass along whether the right hand
8415 // chunk is in the free lists.
8416 initialize_free_range((HeapWord*)fc, fcInFreeLists);
8417 }
8418 }
8419
8420 // Lookahead flush:
8421 // If we are tracking a free range, and this is the last chunk that
8422 // we'll look at because its end crosses past _limit, we'll preemptively
8423 // flush it along with any free range we may be holding on to. Note that
8424 // this can be the case only for an already free or freshly garbage
8425 // chunk. If this block is an object, it can never straddle
8426 // over _limit. The "straddling" occurs when _limit is set at
8427 // the previous end of the space when this cycle started, and
8428 // a subsequent heap expansion caused the previously co-terminal
8429 // free block to be coalesced with the newly expanded portion,
8430 // thus rendering _limit a non-block-boundary making it dangerous
8431 // for the sweeper to step over and examine.
8432 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
8433 assert(inFreeRange(), "Should only be called if currently in a free range.");
8434 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
8435 assert(_sp->used_region().contains(eob - 1),
8436 err_msg("eob = " PTR_FORMAT " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
8437 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
8438 _limit, _sp->bottom(), _sp->end(), fc, chunk_size));
8439 if (eob >= _limit) {
8440 assert(eob == _limit || fc->isFree(), "Only a free chunk should allow us to cross over the limit");
8441 if (CMSTraceSweeper) {
8442 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
8443 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
8444 "[" PTR_FORMAT "," PTR_FORMAT ")",
8445 _limit, fc, eob, _sp->bottom(), _sp->end());
8446 }
8447 // Return the storage we are tracking back into the free lists.
8448 if (CMSTraceSweeper) {
8449 gclog_or_tty->print_cr("Flushing ... ");
8450 }
8451 assert(freeFinger() < eob, "Error");
8452 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
8453 }
8454 }
8455
8456 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
8457 assert(inFreeRange(), "Should only be called if currently in a free range.");
8458 assert(size > 0,
8459 "A zero sized chunk cannot be added to the free lists.");
8460 if (!freeRangeInFreeLists()) {
8461 if (CMSTestInFreeList) {
8462 FreeChunk* fc = (FreeChunk*) chunk;
8463 fc->setSize(size);
8464 assert(!_sp->verifyChunkInFreeLists(fc),
8465 "chunk should not be in free lists yet");
8466 }
8467 if (CMSTraceSweeper) {
8468 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8469 chunk, size);
8470 }
8471 // A new free range is going to be starting. The current
8472 // free range has not been added to the free lists yet or
8473 // was removed so add it back.
8474 // If the current free range was coalesced, then the death
8475 // of the free range was recorded. Record a birth now.
8476 if (lastFreeRangeCoalesced()) {
8477 _sp->coalBirth(size);
8478 }
8479 _sp->addChunkAndRepairOffsetTable(chunk, size,
8480 lastFreeRangeCoalesced());
8481 } else if (CMSTraceSweeper) {
8482 gclog_or_tty->print_cr("Already in free list: nothing to flush");
8483 }
8484 set_inFreeRange(false);
8485 set_freeRangeInFreeLists(false);
8486 }
8487
8488 // We take a break if we've been at this for a while,
8489 // so as to avoid monopolizing the locks involved.
8490 void SweepClosure::do_yield_work(HeapWord* addr) {
8491 // Return current free chunk being used for coalescing (if any)
8492 // to the appropriate freelist. After yielding, the next
8493 // free block encountered will start a coalescing range of
8494 // free blocks. If the next free block is adjacent to the
8495 // chunk just flushed, they will need to wait for the next
8496 // sweep to be coalesced.
8497 if (inFreeRange()) {
8498 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8499 }
8500
8501 // First give up the locks, then yield, then re-lock.
8502 // We should probably use a constructor/destructor idiom to
8503 // do this unlock/lock or modify the MutexUnlocker class to
8504 // serve our purpose. XXX
8505 assert_lock_strong(_bitMap->lock());
8506 assert_lock_strong(_freelistLock);
8507 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8508 "CMS thread should hold CMS token");
8509 _bitMap->lock()->unlock();
8510 _freelistLock->unlock();
8511 ConcurrentMarkSweepThread::desynchronize(true);
8512 ConcurrentMarkSweepThread::acknowledge_yield_request();
8513 _collector->stopTimer();
8514 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8515 if (PrintCMSStatistics != 0) {
8516 _collector->incrementYields();
8517 }
8518 _collector->icms_wait();
8519
8520 // See the comment in coordinator_yield()
8521 for (unsigned i = 0; i < CMSYieldSleepCount &&
8522 ConcurrentMarkSweepThread::should_yield() &&
8523 !CMSCollector::foregroundGCIsActive(); ++i) {
8524 os::sleep(Thread::current(), 1, false);
8525 ConcurrentMarkSweepThread::acknowledge_yield_request();
8526 }
8527
8528 ConcurrentMarkSweepThread::synchronize(true);
8529 _freelistLock->lock();
8530 _bitMap->lock()->lock_without_safepoint_check();
8531 _collector->startTimer();
8532 }
8533
8534 #ifndef PRODUCT
8535 // This is actually very useful in a product build if it can
8536 // be called from the debugger. Compile it into the product
8537 // as needed.
8538 bool debug_verifyChunkInFreeLists(FreeChunk* fc) {
8539 return debug_cms_space->verifyChunkInFreeLists(fc);
8540 }
8541 #endif
8542
8543 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
8544 if (CMSTraceSweeper) {
8545 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
8546 fc, fc->size());
8547 }
8548 }
8549
8550 // CMSIsAliveClosure
8551 bool CMSIsAliveClosure::do_object_b(oop obj) {
8552 HeapWord* addr = (HeapWord*)obj;
8553 return addr != NULL &&
8554 (!_span.contains(addr) || _bit_map->isMarked(addr));
8555 }
8556
8557 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
8558 MemRegion span,
8559 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
8560 CMSMarkStack* revisit_stack, bool cpc):
8561 KlassRememberingOopClosure(collector, NULL, revisit_stack),
8562 _span(span),
8563 _bit_map(bit_map),
8564 _mark_stack(mark_stack),
8565 _concurrent_precleaning(cpc) {
8566 assert(!_span.is_empty(), "Empty span could spell trouble");
8567 }
8568
8569
8570 // CMSKeepAliveClosure: the serial version
8571 void CMSKeepAliveClosure::do_oop(oop obj) {
8572 HeapWord* addr = (HeapWord*)obj;
8573 if (_span.contains(addr) &&
8574 !_bit_map->isMarked(addr)) {
8575 _bit_map->mark(addr);
8576 bool simulate_overflow = false;
8577 NOT_PRODUCT(
8578 if (CMSMarkStackOverflowALot &&
8579 _collector->simulate_overflow()) {
8580 // simulate a stack overflow
8581 simulate_overflow = true;
8582 }
8583 )
8584 if (simulate_overflow || !_mark_stack->push(obj)) {
8585 if (_concurrent_precleaning) {
8586 // We dirty the overflown object and let the remark
8587 // phase deal with it.
8588 assert(_collector->overflow_list_is_empty(), "Error");
8589 // In the case of object arrays, we need to dirty all of
8590 // the cards that the object spans. No locking or atomics
8591 // are needed since no one else can be mutating the mod union
8592 // table.
8593 if (obj->is_objArray()) {
8594 size_t sz = obj->size();
8595 HeapWord* end_card_addr =
8596 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
8597 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8598 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8599 _collector->_modUnionTable.mark_range(redirty_range);
8600 } else {
8601 _collector->_modUnionTable.mark(addr);
8602 }
8603 _collector->_ser_kac_preclean_ovflw++;
8604 } else {
8605 _collector->push_on_overflow_list(obj);
8606 _collector->_ser_kac_ovflw++;
8607 }
8608 }
8609 }
8610 }
8611
8612 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8613 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8614
8615 // CMSParKeepAliveClosure: a parallel version of the above.
8616 // The work queues are private to each closure (thread),
8617 // but (may be) available for stealing by other threads.
8618 void CMSParKeepAliveClosure::do_oop(oop obj) {
8619 HeapWord* addr = (HeapWord*)obj;
8620 if (_span.contains(addr) &&
8621 !_bit_map->isMarked(addr)) {
8622 // In general, during recursive tracing, several threads
8623 // may be concurrently getting here; the first one to
8624 // "tag" it, claims it.
8625 if (_bit_map->par_mark(addr)) {
8626 bool res = _work_queue->push(obj);
8627 assert(res, "Low water mark should be much less than capacity");
8628 // Do a recursive trim in the hope that this will keep
8629 // stack usage lower, but leave some oops for potential stealers
8630 trim_queue(_low_water_mark);
8631 } // Else, another thread got there first
8632 }
8633 }
8634
8635 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8636 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8637
8638 void CMSParKeepAliveClosure::trim_queue(uint max) {
8639 while (_work_queue->size() > max) {
8640 oop new_oop;
8641 if (_work_queue->pop_local(new_oop)) {
8642 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8643 assert(_bit_map->isMarked((HeapWord*)new_oop),
8644 "no white objects on this stack!");
8645 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8646 // iterate over the oops in this oop, marking and pushing
8647 // the ones in CMS heap (i.e. in _span).
8648 new_oop->oop_iterate(&_mark_and_push);
8649 }
8650 }
8651 }
8652
8653 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8654 CMSCollector* collector,
8655 MemRegion span, CMSBitMap* bit_map,
8656 CMSMarkStack* revisit_stack,
8657 OopTaskQueue* work_queue):
8658 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
8659 _span(span),
8660 _bit_map(bit_map),
8661 _work_queue(work_queue) { }
8662
8663 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8664 HeapWord* addr = (HeapWord*)obj;
8665 if (_span.contains(addr) &&
8666 !_bit_map->isMarked(addr)) {
8667 if (_bit_map->par_mark(addr)) {
8668 bool simulate_overflow = false;
8669 NOT_PRODUCT(
8670 if (CMSMarkStackOverflowALot &&
8671 _collector->par_simulate_overflow()) {
8672 // simulate a stack overflow
8673 simulate_overflow = true;
8674 }
8675 )
8676 if (simulate_overflow || !_work_queue->push(obj)) {
8677 _collector->par_push_on_overflow_list(obj);
8678 _collector->_par_kac_ovflw++;
8679 }
8680 } // Else another thread got there already
8681 }
8682 }
8683
8684 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8685 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8686
8687 //////////////////////////////////////////////////////////////////
8688 // CMSExpansionCause /////////////////////////////
8689 //////////////////////////////////////////////////////////////////
8690 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8691 switch (cause) {
8692 case _no_expansion:
8693 return "No expansion";
8694 case _satisfy_free_ratio:
8695 return "Free ratio";
8696 case _satisfy_promotion:
8697 return "Satisfy promotion";
8698 case _satisfy_allocation:
8699 return "allocation";
8700 case _allocate_par_lab:
8701 return "Par LAB";
8702 case _allocate_par_spooling_space:
8703 return "Par Spooling Space";
8704 case _adaptive_size_policy:
8705 return "Ergonomics";
8706 default:
8707 return "unknown";
8708 }
8709 }
8710
8711 void CMSDrainMarkingStackClosure::do_void() {
8712 // the max number to take from overflow list at a time
8713 const size_t num = _mark_stack->capacity()/4;
8714 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8715 "Overflow list should be NULL during concurrent phases");
8716 while (!_mark_stack->isEmpty() ||
8717 // if stack is empty, check the overflow list
8718 _collector->take_from_overflow_list(num, _mark_stack)) {
8719 oop obj = _mark_stack->pop();
8720 HeapWord* addr = (HeapWord*)obj;
8721 assert(_span.contains(addr), "Should be within span");
8722 assert(_bit_map->isMarked(addr), "Should be marked");
8723 assert(obj->is_oop(), "Should be an oop");
8724 obj->oop_iterate(_keep_alive);
8725 }
8726 }
8727
8728 void CMSParDrainMarkingStackClosure::do_void() {
8729 // drain queue
8730 trim_queue(0);
8731 }
8732
8733 // Trim our work_queue so its length is below max at return
8734 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8735 while (_work_queue->size() > max) {
8736 oop new_oop;
8737 if (_work_queue->pop_local(new_oop)) {
8738 assert(new_oop->is_oop(), "Expected an oop");
8739 assert(_bit_map->isMarked((HeapWord*)new_oop),
8740 "no white objects on this stack!");
8741 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8742 // iterate over the oops in this oop, marking and pushing
8743 // the ones in CMS heap (i.e. in _span).
8744 new_oop->oop_iterate(&_mark_and_push);
8745 }
8746 }
8747 }
8748
8749 ////////////////////////////////////////////////////////////////////
8750 // Support for Marking Stack Overflow list handling and related code
8751 ////////////////////////////////////////////////////////////////////
8752 // Much of the following code is similar in shape and spirit to the
8753 // code used in ParNewGC. We should try and share that code
8754 // as much as possible in the future.
8755
8756 #ifndef PRODUCT
8757 // Debugging support for CMSStackOverflowALot
8758
8759 // It's OK to call this multi-threaded; the worst thing
8760 // that can happen is that we'll get a bunch of closely
8761 // spaced simulated oveflows, but that's OK, in fact
8762 // probably good as it would exercise the overflow code
8763 // under contention.
8764 bool CMSCollector::simulate_overflow() {
8765 if (_overflow_counter-- <= 0) { // just being defensive
8766 _overflow_counter = CMSMarkStackOverflowInterval;
8767 return true;
8768 } else {
8769 return false;
8770 }
8771 }
8772
8773 bool CMSCollector::par_simulate_overflow() {
8774 return simulate_overflow();
8775 }
8776 #endif
8777
8778 // Single-threaded
8779 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8780 assert(stack->isEmpty(), "Expected precondition");
8781 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8782 size_t i = num;
8783 oop cur = _overflow_list;
8784 const markOop proto = markOopDesc::prototype();
8785 NOT_PRODUCT(ssize_t n = 0;)
8786 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8787 next = oop(cur->mark());
8788 cur->set_mark(proto); // until proven otherwise
8789 assert(cur->is_oop(), "Should be an oop");
8790 bool res = stack->push(cur);
8791 assert(res, "Bit off more than can chew?");
8792 NOT_PRODUCT(n++;)
8793 }
8794 _overflow_list = cur;
8795 #ifndef PRODUCT
8796 assert(_num_par_pushes >= n, "Too many pops?");
8797 _num_par_pushes -=n;
8798 #endif
8799 return !stack->isEmpty();
8800 }
8801
8802 #define BUSY (oop(0x1aff1aff))
8803 // (MT-safe) Get a prefix of at most "num" from the list.
8804 // The overflow list is chained through the mark word of
8805 // each object in the list. We fetch the entire list,
8806 // break off a prefix of the right size and return the
8807 // remainder. If other threads try to take objects from
8808 // the overflow list at that time, they will wait for
8809 // some time to see if data becomes available. If (and
8810 // only if) another thread places one or more object(s)
8811 // on the global list before we have returned the suffix
8812 // to the global list, we will walk down our local list
8813 // to find its end and append the global list to
8814 // our suffix before returning it. This suffix walk can
8815 // prove to be expensive (quadratic in the amount of traffic)
8816 // when there are many objects in the overflow list and
8817 // there is much producer-consumer contention on the list.
8818 // *NOTE*: The overflow list manipulation code here and
8819 // in ParNewGeneration:: are very similar in shape,
8820 // except that in the ParNew case we use the old (from/eden)
8821 // copy of the object to thread the list via its klass word.
8822 // Because of the common code, if you make any changes in
8823 // the code below, please check the ParNew version to see if
8824 // similar changes might be needed.
8825 // CR 6797058 has been filed to consolidate the common code.
8826 bool CMSCollector::par_take_from_overflow_list(size_t num,
8827 OopTaskQueue* work_q,
8828 int no_of_gc_threads) {
8829 assert(work_q->size() == 0, "First empty local work queue");
8830 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8831 if (_overflow_list == NULL) {
8832 return false;
8833 }
8834 // Grab the entire list; we'll put back a suffix
8835 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
8836 Thread* tid = Thread::current();
8837 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
8838 // set to ParallelGCThreads.
8839 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
8840 size_t sleep_time_millis = MAX2((size_t)1, num/100);
8841 // If the list is busy, we spin for a short while,
8842 // sleeping between attempts to get the list.
8843 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
8844 os::sleep(tid, sleep_time_millis, false);
8845 if (_overflow_list == NULL) {
8846 // Nothing left to take
8847 return false;
8848 } else if (_overflow_list != BUSY) {
8849 // Try and grab the prefix
8850 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
8851 }
8852 }
8853 // If the list was found to be empty, or we spun long
8854 // enough, we give up and return empty-handed. If we leave
8855 // the list in the BUSY state below, it must be the case that
8856 // some other thread holds the overflow list and will set it
8857 // to a non-BUSY state in the future.
8858 if (prefix == NULL || prefix == BUSY) {
8859 // Nothing to take or waited long enough
8860 if (prefix == NULL) {
8861 // Write back the NULL in case we overwrote it with BUSY above
8862 // and it is still the same value.
8863 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8864 }
8865 return false;
8866 }
8867 assert(prefix != NULL && prefix != BUSY, "Error");
8868 size_t i = num;
8869 oop cur = prefix;
8870 // Walk down the first "num" objects, unless we reach the end.
8871 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8872 if (cur->mark() == NULL) {
8873 // We have "num" or fewer elements in the list, so there
8874 // is nothing to return to the global list.
8875 // Write back the NULL in lieu of the BUSY we wrote
8876 // above, if it is still the same value.
8877 if (_overflow_list == BUSY) {
8878 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8879 }
8880 } else {
8881 // Chop off the suffix and rerturn it to the global list.
8882 assert(cur->mark() != BUSY, "Error");
8883 oop suffix_head = cur->mark(); // suffix will be put back on global list
8884 cur->set_mark(NULL); // break off suffix
8885 // It's possible that the list is still in the empty(busy) state
8886 // we left it in a short while ago; in that case we may be
8887 // able to place back the suffix without incurring the cost
8888 // of a walk down the list.
8889 oop observed_overflow_list = _overflow_list;
8890 oop cur_overflow_list = observed_overflow_list;
8891 bool attached = false;
8892 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
8893 observed_overflow_list =
8894 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8895 if (cur_overflow_list == observed_overflow_list) {
8896 attached = true;
8897 break;
8898 } else cur_overflow_list = observed_overflow_list;
8899 }
8900 if (!attached) {
8901 // Too bad, someone else sneaked in (at least) an element; we'll need
8902 // to do a splice. Find tail of suffix so we can prepend suffix to global
8903 // list.
8904 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8905 oop suffix_tail = cur;
8906 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8907 "Tautology");
8908 observed_overflow_list = _overflow_list;
8909 do {
8910 cur_overflow_list = observed_overflow_list;
8911 if (cur_overflow_list != BUSY) {
8912 // Do the splice ...
8913 suffix_tail->set_mark(markOop(cur_overflow_list));
8914 } else { // cur_overflow_list == BUSY
8915 suffix_tail->set_mark(NULL);
8916 }
8917 // ... and try to place spliced list back on overflow_list ...
8918 observed_overflow_list =
8919 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8920 } while (cur_overflow_list != observed_overflow_list);
8921 // ... until we have succeeded in doing so.
8922 }
8923 }
8924
8925 // Push the prefix elements on work_q
8926 assert(prefix != NULL, "control point invariant");
8927 const markOop proto = markOopDesc::prototype();
8928 oop next;
8929 NOT_PRODUCT(ssize_t n = 0;)
8930 for (cur = prefix; cur != NULL; cur = next) {
8931 next = oop(cur->mark());
8932 cur->set_mark(proto); // until proven otherwise
8933 assert(cur->is_oop(), "Should be an oop");
8934 bool res = work_q->push(cur);
8935 assert(res, "Bit off more than we can chew?");
8936 NOT_PRODUCT(n++;)
8937 }
8938 #ifndef PRODUCT
8939 assert(_num_par_pushes >= n, "Too many pops?");
8940 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8941 #endif
8942 return true;
8943 }
8944
8945 // Single-threaded
8946 void CMSCollector::push_on_overflow_list(oop p) {
8947 NOT_PRODUCT(_num_par_pushes++;)
8948 assert(p->is_oop(), "Not an oop");
8949 preserve_mark_if_necessary(p);
8950 p->set_mark((markOop)_overflow_list);
8951 _overflow_list = p;
8952 }
8953
8954 // Multi-threaded; use CAS to prepend to overflow list
8955 void CMSCollector::par_push_on_overflow_list(oop p) {
8956 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8957 assert(p->is_oop(), "Not an oop");
8958 par_preserve_mark_if_necessary(p);
8959 oop observed_overflow_list = _overflow_list;
8960 oop cur_overflow_list;
8961 do {
8962 cur_overflow_list = observed_overflow_list;
8963 if (cur_overflow_list != BUSY) {
8964 p->set_mark(markOop(cur_overflow_list));
8965 } else {
8966 p->set_mark(NULL);
8967 }
8968 observed_overflow_list =
8969 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8970 } while (cur_overflow_list != observed_overflow_list);
8971 }
8972 #undef BUSY
8973
8974 // Single threaded
8975 // General Note on GrowableArray: pushes may silently fail
8976 // because we are (temporarily) out of C-heap for expanding
8977 // the stack. The problem is quite ubiquitous and affects
8978 // a lot of code in the JVM. The prudent thing for GrowableArray
8979 // to do (for now) is to exit with an error. However, that may
8980 // be too draconian in some cases because the caller may be
8981 // able to recover without much harm. For such cases, we
8982 // should probably introduce a "soft_push" method which returns
8983 // an indication of success or failure with the assumption that
8984 // the caller may be able to recover from a failure; code in
8985 // the VM can then be changed, incrementally, to deal with such
8986 // failures where possible, thus, incrementally hardening the VM
8987 // in such low resource situations.
8988 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8989 _preserved_oop_stack.push(p);
8990 _preserved_mark_stack.push(m);
8991 assert(m == p->mark(), "Mark word changed");
8992 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8993 "bijection");
8994 }
8995
8996 // Single threaded
8997 void CMSCollector::preserve_mark_if_necessary(oop p) {
8998 markOop m = p->mark();
8999 if (m->must_be_preserved(p)) {
9000 preserve_mark_work(p, m);
9001 }
9002 }
9003
9004 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
9005 markOop m = p->mark();
9006 if (m->must_be_preserved(p)) {
9007 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
9008 // Even though we read the mark word without holding
9009 // the lock, we are assured that it will not change
9010 // because we "own" this oop, so no other thread can
9011 // be trying to push it on the overflow list; see
9012 // the assertion in preserve_mark_work() that checks
9013 // that m == p->mark().
9014 preserve_mark_work(p, m);
9015 }
9016 }
9017
9018 // We should be able to do this multi-threaded,
9019 // a chunk of stack being a task (this is
9020 // correct because each oop only ever appears
9021 // once in the overflow list. However, it's
9022 // not very easy to completely overlap this with
9023 // other operations, so will generally not be done
9024 // until all work's been completed. Because we
9025 // expect the preserved oop stack (set) to be small,
9026 // it's probably fine to do this single-threaded.
9027 // We can explore cleverer concurrent/overlapped/parallel
9028 // processing of preserved marks if we feel the
9029 // need for this in the future. Stack overflow should
9030 // be so rare in practice and, when it happens, its
9031 // effect on performance so great that this will
9032 // likely just be in the noise anyway.
9033 void CMSCollector::restore_preserved_marks_if_any() {
9034 assert(SafepointSynchronize::is_at_safepoint(),
9035 "world should be stopped");
9036 assert(Thread::current()->is_ConcurrentGC_thread() ||
9037 Thread::current()->is_VM_thread(),
9038 "should be single-threaded");
9039 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9040 "bijection");
9041
9042 while (!_preserved_oop_stack.is_empty()) {
9043 oop p = _preserved_oop_stack.pop();
9044 assert(p->is_oop(), "Should be an oop");
9045 assert(_span.contains(p), "oop should be in _span");
9046 assert(p->mark() == markOopDesc::prototype(),
9047 "Set when taken from overflow list");
9048 markOop m = _preserved_mark_stack.pop();
9049 p->set_mark(m);
9050 }
9051 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
9052 "stacks were cleared above");
9053 }
9054
9055 #ifndef PRODUCT
9056 bool CMSCollector::no_preserved_marks() const {
9057 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
9058 }
9059 #endif
9060
9061 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
9062 {
9063 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
9064 CMSAdaptiveSizePolicy* size_policy =
9065 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
9066 assert(size_policy->is_gc_cms_adaptive_size_policy(),
9067 "Wrong type for size policy");
9068 return size_policy;
9069 }
9070
9071 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
9072 size_t desired_promo_size) {
9073 if (cur_promo_size < desired_promo_size) {
9074 size_t expand_bytes = desired_promo_size - cur_promo_size;
9075 if (PrintAdaptiveSizePolicy && Verbose) {
9076 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9077 "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
9078 expand_bytes);
9079 }
9080 expand(expand_bytes,
9081 MinHeapDeltaBytes,
9082 CMSExpansionCause::_adaptive_size_policy);
9083 } else if (desired_promo_size < cur_promo_size) {
9084 size_t shrink_bytes = cur_promo_size - desired_promo_size;
9085 if (PrintAdaptiveSizePolicy && Verbose) {
9086 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9087 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
9088 shrink_bytes);
9089 }
9090 shrink(shrink_bytes);
9091 }
9092 }
9093
9094 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
9095 GenCollectedHeap* gch = GenCollectedHeap::heap();
9096 CMSGCAdaptivePolicyCounters* counters =
9097 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
9098 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
9099 "Wrong kind of counters");
9100 return counters;
9101 }
9102
9103
9104 void ASConcurrentMarkSweepGeneration::update_counters() {
9105 if (UsePerfData) {
9106 _space_counters->update_all();
9107 _gen_counters->update_all();
9108 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9109 GenCollectedHeap* gch = GenCollectedHeap::heap();
9110 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9111 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9112 "Wrong gc statistics type");
9113 counters->update_counters(gc_stats_l);
9114 }
9115 }
9116
9117 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
9118 if (UsePerfData) {
9119 _space_counters->update_used(used);
9120 _space_counters->update_capacity();
9121 _gen_counters->update_all();
9122
9123 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9124 GenCollectedHeap* gch = GenCollectedHeap::heap();
9125 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9126 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9127 "Wrong gc statistics type");
9128 counters->update_counters(gc_stats_l);
9129 }
9130 }
9131
9132 // The desired expansion delta is computed so that:
9133 // . desired free percentage or greater is used
9134 void ASConcurrentMarkSweepGeneration::compute_new_size() {
9135 assert_locked_or_safepoint(Heap_lock);
9136
9137 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
9138
9139 // If incremental collection failed, we just want to expand
9140 // to the limit.
9141 if (incremental_collection_failed()) {
9142 clear_incremental_collection_failed();
9143 grow_to_reserved();
9144 return;
9145 }
9146
9147 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing");
9148
9149 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
9150 "Wrong type of heap");
9151 int prev_level = level() - 1;
9152 assert(prev_level >= 0, "The cms generation is the lowest generation");
9153 Generation* prev_gen = gch->get_gen(prev_level);
9154 assert(prev_gen->kind() == Generation::ASParNew,
9155 "Wrong type of young generation");
9156 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen;
9157 size_t cur_eden = younger_gen->eden()->capacity();
9158 CMSAdaptiveSizePolicy* size_policy = cms_size_policy();
9159 size_t cur_promo = free();
9160 size_policy->compute_tenured_generation_free_space(cur_promo,
9161 max_available(),
9162 cur_eden);
9163 resize(cur_promo, size_policy->promo_size());
9164
9165 // Record the new size of the space in the cms generation
9166 // that is available for promotions. This is temporary.
9167 // It should be the desired promo size.
9168 size_policy->avg_cms_promo()->sample(free());
9169 size_policy->avg_old_live()->sample(used());
9170
9171 if (UsePerfData) {
9172 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9173 counters->update_cms_capacity_counter(capacity());
9174 }
9175 }
9176
9177 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
9178 assert_locked_or_safepoint(Heap_lock);
9179 assert_lock_strong(freelistLock());
9180 HeapWord* old_end = _cmsSpace->end();
9181 HeapWord* unallocated_start = _cmsSpace->unallocated_block();
9182 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
9183 FreeChunk* chunk_at_end = find_chunk_at_end();
9184 if (chunk_at_end == NULL) {
9185 // No room to shrink
9186 if (PrintGCDetails && Verbose) {
9187 gclog_or_tty->print_cr("No room to shrink: old_end "
9188 PTR_FORMAT " unallocated_start " PTR_FORMAT
9189 " chunk_at_end " PTR_FORMAT,
9190 old_end, unallocated_start, chunk_at_end);
9191 }
9192 return;
9193 } else {
9194
9195 // Find the chunk at the end of the space and determine
9196 // how much it can be shrunk.
9197 size_t shrinkable_size_in_bytes = chunk_at_end->size();
9198 size_t aligned_shrinkable_size_in_bytes =
9199 align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
9200 assert(unallocated_start <= chunk_at_end->end(),
9201 "Inconsistent chunk at end of space");
9202 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
9203 size_t word_size_before = heap_word_size(_virtual_space.committed_size());
9204
9205 // Shrink the underlying space
9206 _virtual_space.shrink_by(bytes);
9207 if (PrintGCDetails && Verbose) {
9208 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
9209 " desired_bytes " SIZE_FORMAT
9210 " shrinkable_size_in_bytes " SIZE_FORMAT
9211 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
9212 " bytes " SIZE_FORMAT,
9213 desired_bytes, shrinkable_size_in_bytes,
9214 aligned_shrinkable_size_in_bytes, bytes);
9215 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT
9216 " unallocated_start " SIZE_FORMAT,
9217 old_end, unallocated_start);
9218 }
9219
9220 // If the space did shrink (shrinking is not guaranteed),
9221 // shrink the chunk at the end by the appropriate amount.
9222 if (((HeapWord*)_virtual_space.high()) < old_end) {
9223 size_t new_word_size =
9224 heap_word_size(_virtual_space.committed_size());
9225
9226 // Have to remove the chunk from the dictionary because it is changing
9227 // size and might be someplace elsewhere in the dictionary.
9228
9229 // Get the chunk at end, shrink it, and put it
9230 // back.
9231 _cmsSpace->removeChunkFromDictionary(chunk_at_end);
9232 size_t word_size_change = word_size_before - new_word_size;
9233 size_t chunk_at_end_old_size = chunk_at_end->size();
9234 assert(chunk_at_end_old_size >= word_size_change,
9235 "Shrink is too large");
9236 chunk_at_end->setSize(chunk_at_end_old_size -
9237 word_size_change);
9238 _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
9239 word_size_change);
9240
9241 _cmsSpace->returnChunkToDictionary(chunk_at_end);
9242
9243 MemRegion mr(_cmsSpace->bottom(), new_word_size);
9244 _bts->resize(new_word_size); // resize the block offset shared array
9245 Universe::heap()->barrier_set()->resize_covered_region(mr);
9246 _cmsSpace->assert_locked();
9247 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
9248
9249 NOT_PRODUCT(_cmsSpace->dictionary()->verify());
9250
9251 // update the space and generation capacity counters
9252 if (UsePerfData) {
9253 _space_counters->update_capacity();
9254 _gen_counters->update_all();
9255 }
9256
9257 if (Verbose && PrintGCDetails) {
9258 size_t new_mem_size = _virtual_space.committed_size();
9259 size_t old_mem_size = new_mem_size + bytes;
9260 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK",
9261 name(), old_mem_size/K, bytes/K, new_mem_size/K);
9262 }
9263 }
9264
9265 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
9266 "Inconsistency at end of space");
9267 assert(chunk_at_end->end() == _cmsSpace->end(),
9268 "Shrinking is inconsistent");
9269 return;
9270 }
9271 }
9272
9273 // Transfer some number of overflown objects to usual marking
9274 // stack. Return true if some objects were transferred.
9275 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
9276 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
9277 (size_t)ParGCDesiredObjsFromOverflowList);
9278
9279 bool res = _collector->take_from_overflow_list(num, _mark_stack);
9280 assert(_collector->overflow_list_is_empty() || res,
9281 "If list is not empty, we should have taken something");
9282 assert(!res || !_mark_stack->isEmpty(),
9283 "If we took something, it should now be on our stack");
9284 return res;
9285 }
9286
9287 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
9288 size_t res = _sp->block_size_no_stall(addr, _collector);
9289 if (_sp->block_is_obj(addr)) {
9290 if (_live_bit_map->isMarked(addr)) {
9291 // It can't have been dead in a previous cycle
9292 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
9293 } else {
9294 _dead_bit_map->mark(addr); // mark the dead object
9295 }
9296 }
9297 // Could be 0, if the block size could not be computed without stalling.
9298 return res;
9299 }
9300
9301 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
9302
9303 switch (phase) {
9304 case CMSCollector::InitialMarking:
9305 initialize(true /* fullGC */ ,
9306 cause /* cause of the GC */,
9307 true /* recordGCBeginTime */,
9308 true /* recordPreGCUsage */,
9309 false /* recordPeakUsage */,
9310 false /* recordPostGCusage */,
9311 true /* recordAccumulatedGCTime */,
9312 false /* recordGCEndTime */,
9313 false /* countCollection */ );
9314 break;
9315
9316 case CMSCollector::FinalMarking:
9317 initialize(true /* fullGC */ ,
9318 cause /* cause of the GC */,
9319 false /* recordGCBeginTime */,
9320 false /* recordPreGCUsage */,
9321 false /* recordPeakUsage */,
9322 false /* recordPostGCusage */,
9323 true /* recordAccumulatedGCTime */,
9324 false /* recordGCEndTime */,
9325 false /* countCollection */ );
9326 break;
9327
9328 case CMSCollector::Sweeping:
9329 initialize(true /* fullGC */ ,
9330 cause /* cause of the GC */,
9331 false /* recordGCBeginTime */,
9332 false /* recordPreGCUsage */,
9333 true /* recordPeakUsage */,
9334 true /* recordPostGCusage */,
9335 false /* recordAccumulatedGCTime */,
9336 true /* recordGCEndTime */,
9337 true /* countCollection */ );
9338 break;
9339
9340 default:
9341 ShouldNotReachHere();
9342 }
9343 }
9344