1 /*
2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/symbolTable.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "code/codeCache.hpp"
29 #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp"
30 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp"
31 #include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.hpp"
32 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp"
33 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp"
34 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp"
35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp"
36 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp"
37 #include "gc_implementation/parNew/parNewGeneration.hpp"
38 #include "gc_implementation/shared/collectorCounters.hpp"
39 #include "gc_implementation/shared/gcTimer.hpp"
40 #include "gc_implementation/shared/gcTrace.hpp"
41 #include "gc_implementation/shared/gcTraceTime.hpp"
42 #include "gc_implementation/shared/isGCActiveMark.hpp"
43 #include "gc_interface/collectedHeap.inline.hpp"
44 #include "memory/allocation.hpp"
45 #include "memory/cardTableRS.hpp"
46 #include "memory/collectorPolicy.hpp"
47 #include "memory/gcLocker.inline.hpp"
48 #include "memory/genCollectedHeap.hpp"
49 #include "memory/genMarkSweep.hpp"
50 #include "memory/genOopClosures.inline.hpp"
51 #include "memory/iterator.hpp"
52 #include "memory/referencePolicy.hpp"
53 #include "memory/resourceArea.hpp"
54 #include "oops/oop.inline.hpp"
55 #include "prims/jvmtiExport.hpp"
56 #include "runtime/globals_extension.hpp"
57 #include "runtime/handles.inline.hpp"
58 #include "runtime/java.hpp"
59 #include "runtime/vmThread.hpp"
60 #include "services/memoryService.hpp"
61 #include "services/runtimeService.hpp"
62
63 // statics
64 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
65 bool CMSCollector::_full_gc_requested = false;
66 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
67
68 //////////////////////////////////////////////////////////////////
69 // In support of CMS/VM thread synchronization
70 //////////////////////////////////////////////////////////////////
71 // We split use of the CGC_lock into 2 "levels".
72 // The low-level locking is of the usual CGC_lock monitor. We introduce
73 // a higher level "token" (hereafter "CMS token") built on top of the
74 // low level monitor (hereafter "CGC lock").
75 // The token-passing protocol gives priority to the VM thread. The
76 // CMS-lock doesn't provide any fairness guarantees, but clients
77 // should ensure that it is only held for very short, bounded
78 // durations.
79 //
80 // When either of the CMS thread or the VM thread is involved in
81 // collection operations during which it does not want the other
82 // thread to interfere, it obtains the CMS token.
83 //
84 // If either thread tries to get the token while the other has
85 // it, that thread waits. However, if the VM thread and CMS thread
86 // both want the token, then the VM thread gets priority while the
87 // CMS thread waits. This ensures, for instance, that the "concurrent"
88 // phases of the CMS thread's work do not block out the VM thread
89 // for long periods of time as the CMS thread continues to hog
90 // the token. (See bug 4616232).
91 //
92 // The baton-passing functions are, however, controlled by the
93 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
94 // and here the low-level CMS lock, not the high level token,
95 // ensures mutual exclusion.
96 //
97 // Two important conditions that we have to satisfy:
98 // 1. if a thread does a low-level wait on the CMS lock, then it
99 // relinquishes the CMS token if it were holding that token
100 // when it acquired the low-level CMS lock.
101 // 2. any low-level notifications on the low-level lock
102 // should only be sent when a thread has relinquished the token.
103 //
104 // In the absence of either property, we'd have potential deadlock.
105 //
106 // We protect each of the CMS (concurrent and sequential) phases
107 // with the CMS _token_, not the CMS _lock_.
108 //
109 // The only code protected by CMS lock is the token acquisition code
110 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
111 // baton-passing code.
112 //
113 // Unfortunately, i couldn't come up with a good abstraction to factor and
114 // hide the naked CGC_lock manipulation in the baton-passing code
115 // further below. That's something we should try to do. Also, the proof
116 // of correctness of this 2-level locking scheme is far from obvious,
117 // and potentially quite slippery. We have an uneasy supsicion, for instance,
118 // that there may be a theoretical possibility of delay/starvation in the
119 // low-level lock/wait/notify scheme used for the baton-passing because of
120 // potential intereference with the priority scheme embodied in the
121 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
122 // invocation further below and marked with "XXX 20011219YSR".
123 // Indeed, as we note elsewhere, this may become yet more slippery
124 // in the presence of multiple CMS and/or multiple VM threads. XXX
125
126 class CMSTokenSync: public StackObj {
127 private:
128 bool _is_cms_thread;
129 public:
130 CMSTokenSync(bool is_cms_thread):
131 _is_cms_thread(is_cms_thread) {
132 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
133 "Incorrect argument to constructor");
134 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
135 }
136
137 ~CMSTokenSync() {
138 assert(_is_cms_thread ?
139 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
140 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
141 "Incorrect state");
142 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
143 }
144 };
145
146 // Convenience class that does a CMSTokenSync, and then acquires
147 // upto three locks.
148 class CMSTokenSyncWithLocks: public CMSTokenSync {
149 private:
150 // Note: locks are acquired in textual declaration order
151 // and released in the opposite order
152 MutexLockerEx _locker1, _locker2, _locker3;
153 public:
154 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
155 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
156 CMSTokenSync(is_cms_thread),
157 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
158 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
159 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
160 { }
161 };
162
163
164 // Wrapper class to temporarily disable icms during a foreground cms collection.
165 class ICMSDisabler: public StackObj {
166 public:
167 // The ctor disables icms and wakes up the thread so it notices the change;
168 // the dtor re-enables icms. Note that the CMSCollector methods will check
169 // CMSIncrementalMode.
170 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); }
171 ~ICMSDisabler() { CMSCollector::enable_icms(); }
172 };
173
174 //////////////////////////////////////////////////////////////////
175 // Concurrent Mark-Sweep Generation /////////////////////////////
176 //////////////////////////////////////////////////////////////////
177
178 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
179
180 // This struct contains per-thread things necessary to support parallel
181 // young-gen collection.
182 class CMSParGCThreadState: public CHeapObj<mtGC> {
183 public:
184 CFLS_LAB lab;
185 PromotionInfo promo;
186
187 // Constructor.
188 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
189 promo.setSpace(cfls);
190 }
191 };
192
193 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
194 ReservedSpace rs, size_t initial_byte_size, int level,
195 CardTableRS* ct, bool use_adaptive_freelists,
196 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) :
197 CardGeneration(rs, initial_byte_size, level, ct),
198 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
199 _debug_collection_type(Concurrent_collection_type)
200 {
201 HeapWord* bottom = (HeapWord*) _virtual_space.low();
202 HeapWord* end = (HeapWord*) _virtual_space.high();
203
204 _direct_allocated_words = 0;
205 NOT_PRODUCT(
206 _numObjectsPromoted = 0;
207 _numWordsPromoted = 0;
208 _numObjectsAllocated = 0;
209 _numWordsAllocated = 0;
210 )
211
212 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
213 use_adaptive_freelists,
214 dictionaryChoice);
215 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
216 if (_cmsSpace == NULL) {
217 vm_exit_during_initialization(
218 "CompactibleFreeListSpace allocation failure");
219 }
220 _cmsSpace->_gen = this;
221
222 _gc_stats = new CMSGCStats();
223
224 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
225 // offsets match. The ability to tell free chunks from objects
226 // depends on this property.
227 debug_only(
228 FreeChunk* junk = NULL;
229 assert(UseCompressedOops ||
230 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
231 "Offset of FreeChunk::_prev within FreeChunk must match"
232 " that of OopDesc::_klass within OopDesc");
233 )
234 if (CollectedHeap::use_parallel_gc_threads()) {
235 typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
236 _par_gc_thread_states =
237 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC);
238 if (_par_gc_thread_states == NULL) {
239 vm_exit_during_initialization("Could not allocate par gc structs");
240 }
241 for (uint i = 0; i < ParallelGCThreads; i++) {
242 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
243 if (_par_gc_thread_states[i] == NULL) {
244 vm_exit_during_initialization("Could not allocate par gc structs");
245 }
246 }
247 } else {
248 _par_gc_thread_states = NULL;
249 }
250 _incremental_collection_failed = false;
251 // The "dilatation_factor" is the expansion that can occur on
252 // account of the fact that the minimum object size in the CMS
253 // generation may be larger than that in, say, a contiguous young
254 // generation.
255 // Ideally, in the calculation below, we'd compute the dilatation
256 // factor as: MinChunkSize/(promoting_gen's min object size)
257 // Since we do not have such a general query interface for the
258 // promoting generation, we'll instead just use the mimimum
259 // object size (which today is a header's worth of space);
260 // note that all arithmetic is in units of HeapWords.
261 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
262 assert(_dilatation_factor >= 1.0, "from previous assert");
263 }
264
265
266 // The field "_initiating_occupancy" represents the occupancy percentage
267 // at which we trigger a new collection cycle. Unless explicitly specified
268 // via CMSInitiating[Perm]OccupancyFraction (argument "io" below), it
269 // is calculated by:
270 //
271 // Let "f" be MinHeapFreeRatio in
272 //
273 // _intiating_occupancy = 100-f +
274 // f * (CMSTrigger[Perm]Ratio/100)
275 // where CMSTrigger[Perm]Ratio is the argument "tr" below.
276 //
277 // That is, if we assume the heap is at its desired maximum occupancy at the
278 // end of a collection, we let CMSTrigger[Perm]Ratio of the (purported) free
279 // space be allocated before initiating a new collection cycle.
280 //
281 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) {
282 assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments");
283 if (io >= 0) {
284 _initiating_occupancy = (double)io / 100.0;
285 } else {
286 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
287 (double)(tr * MinHeapFreeRatio) / 100.0)
288 / 100.0;
289 }
290 }
291
292 void ConcurrentMarkSweepGeneration::ref_processor_init() {
293 assert(collector() != NULL, "no collector");
294 collector()->ref_processor_init();
295 }
296
297 void CMSCollector::ref_processor_init() {
298 if (_ref_processor == NULL) {
299 // Allocate and initialize a reference processor
300 _ref_processor =
301 new ReferenceProcessor(_span, // span
302 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
303 (int) ParallelGCThreads, // mt processing degree
304 _cmsGen->refs_discovery_is_mt(), // mt discovery
305 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
306 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
307 &_is_alive_closure, // closure for liveness info
308 false); // next field updates do not need write barrier
309 // Initialize the _ref_processor field of CMSGen
310 _cmsGen->set_ref_processor(_ref_processor);
311
312 // Allocate a dummy ref processor for perm gen.
313 ReferenceProcessor* rp2 = new ReferenceProcessor();
314 if (rp2 == NULL) {
315 vm_exit_during_initialization("Could not allocate ReferenceProcessor object");
316 }
317 _permGen->set_ref_processor(rp2);
318 }
319 }
320
321 CMSAdaptiveSizePolicy* CMSCollector::size_policy() {
322 GenCollectedHeap* gch = GenCollectedHeap::heap();
323 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
324 "Wrong type of heap");
325 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
326 gch->gen_policy()->size_policy();
327 assert(sp->is_gc_cms_adaptive_size_policy(),
328 "Wrong type of size policy");
329 return sp;
330 }
331
332 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() {
333 CMSGCAdaptivePolicyCounters* results =
334 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters();
335 assert(
336 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
337 "Wrong gc policy counter kind");
338 return results;
339 }
340
341
342 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
343
344 const char* gen_name = "old";
345
346 // Generation Counters - generation 1, 1 subspace
347 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space);
348
349 _space_counters = new GSpaceCounters(gen_name, 0,
350 _virtual_space.reserved_size(),
351 this, _gen_counters);
352 }
353
354 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
355 _cms_gen(cms_gen)
356 {
357 assert(alpha <= 100, "bad value");
358 _saved_alpha = alpha;
359
360 // Initialize the alphas to the bootstrap value of 100.
361 _gc0_alpha = _cms_alpha = 100;
362
363 _cms_begin_time.update();
364 _cms_end_time.update();
365
366 _gc0_duration = 0.0;
367 _gc0_period = 0.0;
368 _gc0_promoted = 0;
369
370 _cms_duration = 0.0;
371 _cms_period = 0.0;
372 _cms_allocated = 0;
373
374 _cms_used_at_gc0_begin = 0;
375 _cms_used_at_gc0_end = 0;
376 _allow_duty_cycle_reduction = false;
377 _valid_bits = 0;
378 _icms_duty_cycle = CMSIncrementalDutyCycle;
379 }
380
381 double CMSStats::cms_free_adjustment_factor(size_t free) const {
382 // TBD: CR 6909490
383 return 1.0;
384 }
385
386 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
387 }
388
389 // If promotion failure handling is on use
390 // the padded average size of the promotion for each
391 // young generation collection.
392 double CMSStats::time_until_cms_gen_full() const {
393 size_t cms_free = _cms_gen->cmsSpace()->free();
394 GenCollectedHeap* gch = GenCollectedHeap::heap();
395 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(),
396 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
397 if (cms_free > expected_promotion) {
398 // Start a cms collection if there isn't enough space to promote
399 // for the next minor collection. Use the padded average as
400 // a safety factor.
401 cms_free -= expected_promotion;
402
403 // Adjust by the safety factor.
404 double cms_free_dbl = (double)cms_free;
405 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
406 // Apply a further correction factor which tries to adjust
407 // for recent occurance of concurrent mode failures.
408 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
409 cms_free_dbl = cms_free_dbl * cms_adjustment;
410
411 if (PrintGCDetails && Verbose) {
412 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
413 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
414 cms_free, expected_promotion);
415 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
416 cms_free_dbl, cms_consumption_rate() + 1.0);
417 }
418 // Add 1 in case the consumption rate goes to zero.
419 return cms_free_dbl / (cms_consumption_rate() + 1.0);
420 }
421 return 0.0;
422 }
423
424 // Compare the duration of the cms collection to the
425 // time remaining before the cms generation is empty.
426 // Note that the time from the start of the cms collection
427 // to the start of the cms sweep (less than the total
428 // duration of the cms collection) can be used. This
429 // has been tried and some applications experienced
430 // promotion failures early in execution. This was
431 // possibly because the averages were not accurate
432 // enough at the beginning.
433 double CMSStats::time_until_cms_start() const {
434 // We add "gc0_period" to the "work" calculation
435 // below because this query is done (mostly) at the
436 // end of a scavenge, so we need to conservatively
437 // account for that much possible delay
438 // in the query so as to avoid concurrent mode failures
439 // due to starting the collection just a wee bit too
440 // late.
441 double work = cms_duration() + gc0_period();
442 double deadline = time_until_cms_gen_full();
443 // If a concurrent mode failure occurred recently, we want to be
444 // more conservative and halve our expected time_until_cms_gen_full()
445 if (work > deadline) {
446 if (Verbose && PrintGCDetails) {
447 gclog_or_tty->print(
448 " CMSCollector: collect because of anticipated promotion "
449 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
450 gc0_period(), time_until_cms_gen_full());
451 }
452 return 0.0;
453 }
454 return work - deadline;
455 }
456
457 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the
458 // amount of change to prevent wild oscillation.
459 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle,
460 unsigned int new_duty_cycle) {
461 assert(old_duty_cycle <= 100, "bad input value");
462 assert(new_duty_cycle <= 100, "bad input value");
463
464 // Note: use subtraction with caution since it may underflow (values are
465 // unsigned). Addition is safe since we're in the range 0-100.
466 unsigned int damped_duty_cycle = new_duty_cycle;
467 if (new_duty_cycle < old_duty_cycle) {
468 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U);
469 if (new_duty_cycle + largest_delta < old_duty_cycle) {
470 damped_duty_cycle = old_duty_cycle - largest_delta;
471 }
472 } else if (new_duty_cycle > old_duty_cycle) {
473 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U);
474 if (new_duty_cycle > old_duty_cycle + largest_delta) {
475 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U);
476 }
477 }
478 assert(damped_duty_cycle <= 100, "invalid duty cycle computed");
479
480 if (CMSTraceIncrementalPacing) {
481 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ",
482 old_duty_cycle, new_duty_cycle, damped_duty_cycle);
483 }
484 return damped_duty_cycle;
485 }
486
487 unsigned int CMSStats::icms_update_duty_cycle_impl() {
488 assert(CMSIncrementalPacing && valid(),
489 "should be handled in icms_update_duty_cycle()");
490
491 double cms_time_so_far = cms_timer().seconds();
492 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M;
493 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far);
494
495 // Avoid division by 0.
496 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01);
497 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full;
498
499 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U);
500 if (new_duty_cycle > _icms_duty_cycle) {
501 // Avoid very small duty cycles (1 or 2); 0 is allowed.
502 if (new_duty_cycle > 2) {
503 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle,
504 new_duty_cycle);
505 }
506 } else if (_allow_duty_cycle_reduction) {
507 // The duty cycle is reduced only once per cms cycle (see record_cms_end()).
508 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle);
509 // Respect the minimum duty cycle.
510 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin;
511 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle);
512 }
513
514 if (PrintGCDetails || CMSTraceIncrementalPacing) {
515 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle);
516 }
517
518 _allow_duty_cycle_reduction = false;
519 return _icms_duty_cycle;
520 }
521
522 #ifndef PRODUCT
523 void CMSStats::print_on(outputStream *st) const {
524 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
525 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
526 gc0_duration(), gc0_period(), gc0_promoted());
527 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
528 cms_duration(), cms_duration_per_mb(),
529 cms_period(), cms_allocated());
530 st->print(",cms_since_beg=%g,cms_since_end=%g",
531 cms_time_since_begin(), cms_time_since_end());
532 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
533 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
534 if (CMSIncrementalMode) {
535 st->print(",dc=%d", icms_duty_cycle());
536 }
537
538 if (valid()) {
539 st->print(",promo_rate=%g,cms_alloc_rate=%g",
540 promotion_rate(), cms_allocation_rate());
541 st->print(",cms_consumption_rate=%g,time_until_full=%g",
542 cms_consumption_rate(), time_until_cms_gen_full());
543 }
544 st->print(" ");
545 }
546 #endif // #ifndef PRODUCT
547
548 CMSCollector::CollectorState CMSCollector::_collectorState =
549 CMSCollector::Idling;
550 bool CMSCollector::_foregroundGCIsActive = false;
551 bool CMSCollector::_foregroundGCShouldWait = false;
552
553 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
554 ConcurrentMarkSweepGeneration* permGen,
555 CardTableRS* ct,
556 ConcurrentMarkSweepPolicy* cp):
557 _cmsGen(cmsGen),
558 _permGen(permGen),
559 _ct(ct),
560 _ref_processor(NULL), // will be set later
561 _conc_workers(NULL), // may be set later
562 _abort_preclean(false),
563 _start_sampling(false),
564 _between_prologue_and_epilogue(false),
565 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
566 _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"),
567 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
568 -1 /* lock-free */, "No_lock" /* dummy */),
569 _modUnionClosure(&_modUnionTable),
570 _modUnionClosurePar(&_modUnionTable),
571 // Adjust my span to cover old (cms) gen and perm gen
572 _span(cmsGen->reserved()._union(permGen->reserved())),
573 // Construct the is_alive_closure with _span & markBitMap
574 _is_alive_closure(_span, &_markBitMap),
575 _restart_addr(NULL),
576 _overflow_list(NULL),
577 _stats(cmsGen),
578 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true)),
579 _eden_chunk_array(NULL), // may be set in ctor body
580 _eden_chunk_capacity(0), // -- ditto --
581 _eden_chunk_index(0), // -- ditto --
582 _survivor_plab_array(NULL), // -- ditto --
583 _survivor_chunk_array(NULL), // -- ditto --
584 _survivor_chunk_capacity(0), // -- ditto --
585 _survivor_chunk_index(0), // -- ditto --
586 _ser_pmc_preclean_ovflw(0),
587 _ser_kac_preclean_ovflw(0),
588 _ser_pmc_remark_ovflw(0),
589 _par_pmc_remark_ovflw(0),
590 _ser_kac_ovflw(0),
591 _par_kac_ovflw(0),
592 #ifndef PRODUCT
593 _num_par_pushes(0),
594 #endif
595 _collection_count_start(0),
596 _verifying(false),
597 _icms_start_limit(NULL),
598 _icms_stop_limit(NULL),
599 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
600 _completed_initialization(false),
601 _collector_policy(cp),
602 _should_unload_classes(false),
603 _concurrent_cycles_since_last_unload(0),
604 _roots_scanning_options(0),
605 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
606 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
607 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
608 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
609 _cms_start_registered(false)
610 {
611 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
612 ExplicitGCInvokesConcurrent = true;
613 }
614 // Now expand the span and allocate the collection support structures
615 // (MUT, marking bit map etc.) to cover both generations subject to
616 // collection.
617
618 // First check that _permGen is adjacent to _cmsGen and above it.
619 assert( _cmsGen->reserved().word_size() > 0
620 && _permGen->reserved().word_size() > 0,
621 "generations should not be of zero size");
622 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(),
623 "_cmsGen and _permGen should not overlap");
624 assert(_cmsGen->reserved().end() == _permGen->reserved().start(),
625 "_cmsGen->end() different from _permGen->start()");
626
627 // For use by dirty card to oop closures.
628 _cmsGen->cmsSpace()->set_collector(this);
629 _permGen->cmsSpace()->set_collector(this);
630
631 // Allocate MUT and marking bit map
632 {
633 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
634 if (!_markBitMap.allocate(_span)) {
635 warning("Failed to allocate CMS Bit Map");
636 return;
637 }
638 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
639 }
640 {
641 _modUnionTable.allocate(_span);
642 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
643 }
644
645 if (!_markStack.allocate(MarkStackSize)) {
646 warning("Failed to allocate CMS Marking Stack");
647 return;
648 }
649 if (!_revisitStack.allocate(CMSRevisitStackSize)) {
650 warning("Failed to allocate CMS Revisit Stack");
651 return;
652 }
653
654 // Support for multi-threaded concurrent phases
655 if (CMSConcurrentMTEnabled) {
656 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
657 // just for now
658 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
659 }
660 if (ConcGCThreads > 1) {
661 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
662 ConcGCThreads, true);
663 if (_conc_workers == NULL) {
664 warning("GC/CMS: _conc_workers allocation failure: "
665 "forcing -CMSConcurrentMTEnabled");
666 CMSConcurrentMTEnabled = false;
667 } else {
668 _conc_workers->initialize_workers();
669 }
670 } else {
671 CMSConcurrentMTEnabled = false;
672 }
673 }
674 if (!CMSConcurrentMTEnabled) {
675 ConcGCThreads = 0;
676 } else {
677 // Turn off CMSCleanOnEnter optimization temporarily for
678 // the MT case where it's not fixed yet; see 6178663.
679 CMSCleanOnEnter = false;
680 }
681 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
682 "Inconsistency");
683
684 // Parallel task queues; these are shared for the
685 // concurrent and stop-world phases of CMS, but
686 // are not shared with parallel scavenge (ParNew).
687 {
688 uint i;
689 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
690
691 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
692 || ParallelRefProcEnabled)
693 && num_queues > 0) {
694 _task_queues = new OopTaskQueueSet(num_queues);
695 if (_task_queues == NULL) {
696 warning("task_queues allocation failure.");
697 return;
698 }
699 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
700 if (_hash_seed == NULL) {
701 warning("_hash_seed array allocation failure");
702 return;
703 }
704
705 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
706 for (i = 0; i < num_queues; i++) {
707 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
708 if (q == NULL) {
709 warning("work_queue allocation failure.");
710 return;
711 }
712 _task_queues->register_queue(i, q);
713 }
714 for (i = 0; i < num_queues; i++) {
715 _task_queues->queue(i)->initialize();
716 _hash_seed[i] = 17; // copied from ParNew
717 }
718 }
719 }
720
721 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
722 _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio);
723
724 // Clip CMSBootstrapOccupancy between 0 and 100.
725 _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy)))
726 /(double)100;
727
728 _full_gcs_since_conc_gc = 0;
729
730 // Now tell CMS generations the identity of their collector
731 ConcurrentMarkSweepGeneration::set_collector(this);
732
733 // Create & start a CMS thread for this CMS collector
734 _cmsThread = ConcurrentMarkSweepThread::start(this);
735 assert(cmsThread() != NULL, "CMS Thread should have been created");
736 assert(cmsThread()->collector() == this,
737 "CMS Thread should refer to this gen");
738 assert(CGC_lock != NULL, "Where's the CGC_lock?");
739
740 // Support for parallelizing young gen rescan
741 GenCollectedHeap* gch = GenCollectedHeap::heap();
742 _young_gen = gch->prev_gen(_cmsGen);
743 if (gch->supports_inline_contig_alloc()) {
744 _top_addr = gch->top_addr();
745 _end_addr = gch->end_addr();
746 assert(_young_gen != NULL, "no _young_gen");
747 _eden_chunk_index = 0;
748 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
749 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
750 if (_eden_chunk_array == NULL) {
751 _eden_chunk_capacity = 0;
752 warning("GC/CMS: _eden_chunk_array allocation failure");
753 }
754 }
755 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
756
757 // Support for parallelizing survivor space rescan
758 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
759 const size_t max_plab_samples =
760 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize;
761
762 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
763 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC);
764 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
765 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
766 || _cursor == NULL) {
767 warning("Failed to allocate survivor plab/chunk array");
768 if (_survivor_plab_array != NULL) {
769 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC);
770 _survivor_plab_array = NULL;
771 }
772 if (_survivor_chunk_array != NULL) {
773 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC);
774 _survivor_chunk_array = NULL;
775 }
776 if (_cursor != NULL) {
777 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC);
778 _cursor = NULL;
779 }
780 } else {
781 _survivor_chunk_capacity = 2*max_plab_samples;
782 for (uint i = 0; i < ParallelGCThreads; i++) {
783 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
784 if (vec == NULL) {
785 warning("Failed to allocate survivor plab array");
786 for (int j = i; j > 0; j--) {
787 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC);
788 }
789 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC);
790 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC);
791 _survivor_plab_array = NULL;
792 _survivor_chunk_array = NULL;
793 _survivor_chunk_capacity = 0;
794 break;
795 } else {
796 ChunkArray* cur =
797 ::new (&_survivor_plab_array[i]) ChunkArray(vec,
798 max_plab_samples);
799 assert(cur->end() == 0, "Should be 0");
800 assert(cur->array() == vec, "Should be vec");
801 assert(cur->capacity() == max_plab_samples, "Error");
802 }
803 }
804 }
805 }
806 assert( ( _survivor_plab_array != NULL
807 && _survivor_chunk_array != NULL)
808 || ( _survivor_chunk_capacity == 0
809 && _survivor_chunk_index == 0),
810 "Error");
811
812 // Choose what strong roots should be scanned depending on verification options
813 // and perm gen collection mode.
814 if (!CMSClassUnloadingEnabled) {
815 // If class unloading is disabled we want to include all classes into the root set.
816 add_root_scanning_option(SharedHeap::SO_AllClasses);
817 } else {
818 add_root_scanning_option(SharedHeap::SO_SystemClasses);
819 }
820
821 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
822 _gc_counters = new CollectorCounters("CMS", 1);
823 _completed_initialization = true;
824 _inter_sweep_timer.start(); // start of time
825 }
826
827 const char* ConcurrentMarkSweepGeneration::name() const {
828 return "concurrent mark-sweep generation";
829 }
830 void ConcurrentMarkSweepGeneration::update_counters() {
831 if (UsePerfData) {
832 _space_counters->update_all();
833 _gen_counters->update_all();
834 }
835 }
836
837 // this is an optimized version of update_counters(). it takes the
838 // used value as a parameter rather than computing it.
839 //
840 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
841 if (UsePerfData) {
842 _space_counters->update_used(used);
843 _space_counters->update_capacity();
844 _gen_counters->update_all();
845 }
846 }
847
848 void ConcurrentMarkSweepGeneration::print() const {
849 Generation::print();
850 cmsSpace()->print();
851 }
852
853 #ifndef PRODUCT
854 void ConcurrentMarkSweepGeneration::print_statistics() {
855 cmsSpace()->printFLCensus(0);
856 }
857 #endif
858
859 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
860 GenCollectedHeap* gch = GenCollectedHeap::heap();
861 if (PrintGCDetails) {
862 if (Verbose) {
863 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
864 level(), short_name(), s, used(), capacity());
865 } else {
866 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
867 level(), short_name(), s, used() / K, capacity() / K);
868 }
869 }
870 if (Verbose) {
871 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
872 gch->used(), gch->capacity());
873 } else {
874 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
875 gch->used() / K, gch->capacity() / K);
876 }
877 }
878
879 size_t
880 ConcurrentMarkSweepGeneration::contiguous_available() const {
881 // dld proposes an improvement in precision here. If the committed
882 // part of the space ends in a free block we should add that to
883 // uncommitted size in the calculation below. Will make this
884 // change later, staying with the approximation below for the
885 // time being. -- ysr.
886 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
887 }
888
889 size_t
890 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
891 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
892 }
893
894 size_t ConcurrentMarkSweepGeneration::max_available() const {
895 return free() + _virtual_space.uncommitted_size();
896 }
897
898 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
899 size_t available = max_available();
900 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
901 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
902 if (Verbose && PrintGCDetails) {
903 gclog_or_tty->print_cr(
904 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
905 "max_promo("SIZE_FORMAT")",
906 res? "":" not", available, res? ">=":"<",
907 av_promo, max_promotion_in_bytes);
908 }
909 return res;
910 }
911
912 // At a promotion failure dump information on block layout in heap
913 // (cms old generation).
914 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
915 if (CMSDumpAtPromotionFailure) {
916 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
917 }
918 }
919
920 CompactibleSpace*
921 ConcurrentMarkSweepGeneration::first_compaction_space() const {
922 return _cmsSpace;
923 }
924
925 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
926 // Clear the promotion information. These pointers can be adjusted
927 // along with all the other pointers into the heap but
928 // compaction is expected to be a rare event with
929 // a heap using cms so don't do it without seeing the need.
930 if (CollectedHeap::use_parallel_gc_threads()) {
931 for (uint i = 0; i < ParallelGCThreads; i++) {
932 _par_gc_thread_states[i]->promo.reset();
933 }
934 }
935 }
936
937 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
938 blk->do_space(_cmsSpace);
939 }
940
941 void ConcurrentMarkSweepGeneration::compute_new_size() {
942 assert_locked_or_safepoint(Heap_lock);
943
944 // If incremental collection failed, we just want to expand
945 // to the limit.
946 if (incremental_collection_failed()) {
947 clear_incremental_collection_failed();
948 grow_to_reserved();
949 return;
950 }
951
952 size_t expand_bytes = 0;
953 double free_percentage = ((double) free()) / capacity();
954 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
955 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
956
957 // compute expansion delta needed for reaching desired free percentage
958 if (free_percentage < desired_free_percentage) {
959 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
960 assert(desired_capacity >= capacity(), "invalid expansion size");
961 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
962 }
963 if (expand_bytes > 0) {
964 if (PrintGCDetails && Verbose) {
965 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
966 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
967 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
968 gclog_or_tty->print_cr(" Desired free fraction %f",
969 desired_free_percentage);
970 gclog_or_tty->print_cr(" Maximum free fraction %f",
971 maximum_free_percentage);
972 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000);
973 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
974 desired_capacity/1000);
975 int prev_level = level() - 1;
976 if (prev_level >= 0) {
977 size_t prev_size = 0;
978 GenCollectedHeap* gch = GenCollectedHeap::heap();
979 Generation* prev_gen = gch->_gens[prev_level];
980 prev_size = prev_gen->capacity();
981 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
982 prev_size/1000);
983 }
984 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
985 unsafe_max_alloc_nogc()/1000);
986 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
987 contiguous_available()/1000);
988 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
989 expand_bytes);
990 }
991 // safe if expansion fails
992 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
993 if (PrintGCDetails && Verbose) {
994 gclog_or_tty->print_cr(" Expanded free fraction %f",
995 ((double) free()) / capacity());
996 }
997 }
998 }
999
1000 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
1001 return cmsSpace()->freelistLock();
1002 }
1003
1004 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
1005 bool tlab) {
1006 CMSSynchronousYieldRequest yr;
1007 MutexLockerEx x(freelistLock(),
1008 Mutex::_no_safepoint_check_flag);
1009 return have_lock_and_allocate(size, tlab);
1010 }
1011
1012 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
1013 bool tlab /* ignored */) {
1014 assert_lock_strong(freelistLock());
1015 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
1016 HeapWord* res = cmsSpace()->allocate(adjustedSize);
1017 // Allocate the object live (grey) if the background collector has
1018 // started marking. This is necessary because the marker may
1019 // have passed this address and consequently this object will
1020 // not otherwise be greyed and would be incorrectly swept up.
1021 // Note that if this object contains references, the writing
1022 // of those references will dirty the card containing this object
1023 // allowing the object to be blackened (and its references scanned)
1024 // either during a preclean phase or at the final checkpoint.
1025 if (res != NULL) {
1026 // We may block here with an uninitialized object with
1027 // its mark-bit or P-bits not yet set. Such objects need
1028 // to be safely navigable by block_start().
1029 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
1030 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
1031 collector()->direct_allocated(res, adjustedSize);
1032 _direct_allocated_words += adjustedSize;
1033 // allocation counters
1034 NOT_PRODUCT(
1035 _numObjectsAllocated++;
1036 _numWordsAllocated += (int)adjustedSize;
1037 )
1038 }
1039 return res;
1040 }
1041
1042 // In the case of direct allocation by mutators in a generation that
1043 // is being concurrently collected, the object must be allocated
1044 // live (grey) if the background collector has started marking.
1045 // This is necessary because the marker may
1046 // have passed this address and consequently this object will
1047 // not otherwise be greyed and would be incorrectly swept up.
1048 // Note that if this object contains references, the writing
1049 // of those references will dirty the card containing this object
1050 // allowing the object to be blackened (and its references scanned)
1051 // either during a preclean phase or at the final checkpoint.
1052 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
1053 assert(_markBitMap.covers(start, size), "Out of bounds");
1054 if (_collectorState >= Marking) {
1055 MutexLockerEx y(_markBitMap.lock(),
1056 Mutex::_no_safepoint_check_flag);
1057 // [see comments preceding SweepClosure::do_blk() below for details]
1058 // 1. need to mark the object as live so it isn't collected
1059 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
1060 // 3. need to mark the end of the object so marking, precleaning or sweeping
1061 // can skip over uninitialized or unparsable objects. An allocated
1062 // object is considered uninitialized for our purposes as long as
1063 // its klass word is NULL. (Unparsable objects are those which are
1064 // initialized in the sense just described, but whose sizes can still
1065 // not be correctly determined. Note that the class of unparsable objects
1066 // can only occur in the perm gen. All old gen objects are parsable
1067 // as soon as they are initialized.)
1068 _markBitMap.mark(start); // object is live
1069 _markBitMap.mark(start + 1); // object is potentially uninitialized?
1070 _markBitMap.mark(start + size - 1);
1071 // mark end of object
1072 }
1073 // check that oop looks uninitialized
1074 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
1075 }
1076
1077 void CMSCollector::promoted(bool par, HeapWord* start,
1078 bool is_obj_array, size_t obj_size) {
1079 assert(_markBitMap.covers(start), "Out of bounds");
1080 // See comment in direct_allocated() about when objects should
1081 // be allocated live.
1082 if (_collectorState >= Marking) {
1083 // we already hold the marking bit map lock, taken in
1084 // the prologue
1085 if (par) {
1086 _markBitMap.par_mark(start);
1087 } else {
1088 _markBitMap.mark(start);
1089 }
1090 // We don't need to mark the object as uninitialized (as
1091 // in direct_allocated above) because this is being done with the
1092 // world stopped and the object will be initialized by the
1093 // time the marking, precleaning or sweeping get to look at it.
1094 // But see the code for copying objects into the CMS generation,
1095 // where we need to ensure that concurrent readers of the
1096 // block offset table are able to safely navigate a block that
1097 // is in flux from being free to being allocated (and in
1098 // transition while being copied into) and subsequently
1099 // becoming a bona-fide object when the copy/promotion is complete.
1100 assert(SafepointSynchronize::is_at_safepoint(),
1101 "expect promotion only at safepoints");
1102
1103 if (_collectorState < Sweeping) {
1104 // Mark the appropriate cards in the modUnionTable, so that
1105 // this object gets scanned before the sweep. If this is
1106 // not done, CMS generation references in the object might
1107 // not get marked.
1108 // For the case of arrays, which are otherwise precisely
1109 // marked, we need to dirty the entire array, not just its head.
1110 if (is_obj_array) {
1111 // The [par_]mark_range() method expects mr.end() below to
1112 // be aligned to the granularity of a bit's representation
1113 // in the heap. In the case of the MUT below, that's a
1114 // card size.
1115 MemRegion mr(start,
1116 (HeapWord*)round_to((intptr_t)(start + obj_size),
1117 CardTableModRefBS::card_size /* bytes */));
1118 if (par) {
1119 _modUnionTable.par_mark_range(mr);
1120 } else {
1121 _modUnionTable.mark_range(mr);
1122 }
1123 } else { // not an obj array; we can just mark the head
1124 if (par) {
1125 _modUnionTable.par_mark(start);
1126 } else {
1127 _modUnionTable.mark(start);
1128 }
1129 }
1130 }
1131 }
1132 }
1133
1134 static inline size_t percent_of_space(Space* space, HeapWord* addr)
1135 {
1136 size_t delta = pointer_delta(addr, space->bottom());
1137 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
1138 }
1139
1140 void CMSCollector::icms_update_allocation_limits()
1141 {
1142 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
1143 EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
1144
1145 const unsigned int duty_cycle = stats().icms_update_duty_cycle();
1146 if (CMSTraceIncrementalPacing) {
1147 stats().print();
1148 }
1149
1150 assert(duty_cycle <= 100, "invalid duty cycle");
1151 if (duty_cycle != 0) {
1152 // The duty_cycle is a percentage between 0 and 100; convert to words and
1153 // then compute the offset from the endpoints of the space.
1154 size_t free_words = eden->free() / HeapWordSize;
1155 double free_words_dbl = (double)free_words;
1156 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
1157 size_t offset_words = (free_words - duty_cycle_words) / 2;
1158
1159 _icms_start_limit = eden->top() + offset_words;
1160 _icms_stop_limit = eden->end() - offset_words;
1161
1162 // The limits may be adjusted (shifted to the right) by
1163 // CMSIncrementalOffset, to allow the application more mutator time after a
1164 // young gen gc (when all mutators were stopped) and before CMS starts and
1165 // takes away one or more cpus.
1166 if (CMSIncrementalOffset != 0) {
1167 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
1168 size_t adjustment = (size_t)adjustment_dbl;
1169 HeapWord* tmp_stop = _icms_stop_limit + adjustment;
1170 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
1171 _icms_start_limit += adjustment;
1172 _icms_stop_limit = tmp_stop;
1173 }
1174 }
1175 }
1176 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
1177 _icms_start_limit = _icms_stop_limit = eden->end();
1178 }
1179
1180 // Install the new start limit.
1181 eden->set_soft_end(_icms_start_limit);
1182
1183 if (CMSTraceIncrementalMode) {
1184 gclog_or_tty->print(" icms alloc limits: "
1185 PTR_FORMAT "," PTR_FORMAT
1186 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
1187 _icms_start_limit, _icms_stop_limit,
1188 percent_of_space(eden, _icms_start_limit),
1189 percent_of_space(eden, _icms_stop_limit));
1190 if (Verbose) {
1191 gclog_or_tty->print("eden: ");
1192 eden->print_on(gclog_or_tty);
1193 }
1194 }
1195 }
1196
1197 // Any changes here should try to maintain the invariant
1198 // that if this method is called with _icms_start_limit
1199 // and _icms_stop_limit both NULL, then it should return NULL
1200 // and not notify the icms thread.
1201 HeapWord*
1202 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
1203 size_t word_size)
1204 {
1205 // A start_limit equal to end() means the duty cycle is 0, so treat that as a
1206 // nop.
1207 if (CMSIncrementalMode && _icms_start_limit != space->end()) {
1208 if (top <= _icms_start_limit) {
1209 if (CMSTraceIncrementalMode) {
1210 space->print_on(gclog_or_tty);
1211 gclog_or_tty->stamp();
1212 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
1213 ", new limit=" PTR_FORMAT
1214 " (" SIZE_FORMAT "%%)",
1215 top, _icms_stop_limit,
1216 percent_of_space(space, _icms_stop_limit));
1217 }
1218 ConcurrentMarkSweepThread::start_icms();
1219 assert(top < _icms_stop_limit, "Tautology");
1220 if (word_size < pointer_delta(_icms_stop_limit, top)) {
1221 return _icms_stop_limit;
1222 }
1223
1224 // The allocation will cross both the _start and _stop limits, so do the
1225 // stop notification also and return end().
1226 if (CMSTraceIncrementalMode) {
1227 space->print_on(gclog_or_tty);
1228 gclog_or_tty->stamp();
1229 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
1230 ", new limit=" PTR_FORMAT
1231 " (" SIZE_FORMAT "%%)",
1232 top, space->end(),
1233 percent_of_space(space, space->end()));
1234 }
1235 ConcurrentMarkSweepThread::stop_icms();
1236 return space->end();
1237 }
1238
1239 if (top <= _icms_stop_limit) {
1240 if (CMSTraceIncrementalMode) {
1241 space->print_on(gclog_or_tty);
1242 gclog_or_tty->stamp();
1243 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
1244 ", new limit=" PTR_FORMAT
1245 " (" SIZE_FORMAT "%%)",
1246 top, space->end(),
1247 percent_of_space(space, space->end()));
1248 }
1249 ConcurrentMarkSweepThread::stop_icms();
1250 return space->end();
1251 }
1252
1253 if (CMSTraceIncrementalMode) {
1254 space->print_on(gclog_or_tty);
1255 gclog_or_tty->stamp();
1256 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
1257 ", new limit=" PTR_FORMAT,
1258 top, NULL);
1259 }
1260 }
1261
1262 return NULL;
1263 }
1264
1265 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
1266 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1267 // allocate, copy and if necessary update promoinfo --
1268 // delegate to underlying space.
1269 assert_lock_strong(freelistLock());
1270
1271 #ifndef PRODUCT
1272 if (Universe::heap()->promotion_should_fail()) {
1273 return NULL;
1274 }
1275 #endif // #ifndef PRODUCT
1276
1277 oop res = _cmsSpace->promote(obj, obj_size);
1278 if (res == NULL) {
1279 // expand and retry
1280 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
1281 expand(s*HeapWordSize, MinHeapDeltaBytes,
1282 CMSExpansionCause::_satisfy_promotion);
1283 // Since there's currently no next generation, we don't try to promote
1284 // into a more senior generation.
1285 assert(next_gen() == NULL, "assumption, based upon which no attempt "
1286 "is made to pass on a possibly failing "
1287 "promotion to next generation");
1288 res = _cmsSpace->promote(obj, obj_size);
1289 }
1290 if (res != NULL) {
1291 // See comment in allocate() about when objects should
1292 // be allocated live.
1293 assert(obj->is_oop(), "Will dereference klass pointer below");
1294 collector()->promoted(false, // Not parallel
1295 (HeapWord*)res, obj->is_objArray(), obj_size);
1296 // promotion counters
1297 NOT_PRODUCT(
1298 _numObjectsPromoted++;
1299 _numWordsPromoted +=
1300 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1301 )
1302 }
1303 return res;
1304 }
1305
1306
1307 HeapWord*
1308 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
1309 HeapWord* top,
1310 size_t word_sz)
1311 {
1312 return collector()->allocation_limit_reached(space, top, word_sz);
1313 }
1314
1315 // IMPORTANT: Notes on object size recognition in CMS.
1316 // ---------------------------------------------------
1317 // A block of storage in the CMS generation is always in
1318 // one of three states. A free block (FREE), an allocated
1319 // object (OBJECT) whose size() method reports the correct size,
1320 // and an intermediate state (TRANSIENT) in which its size cannot
1321 // be accurately determined.
1322 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1323 // -----------------------------------------------------
1324 // FREE: klass_word & 1 == 1; mark_word holds block size
1325 //
1326 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1327 // obj->size() computes correct size
1328 // [Perm Gen objects needs to be "parsable" before they can be navigated]
1329 //
1330 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1331 //
1332 // STATE IDENTIFICATION: (64 bit+COOPS)
1333 // ------------------------------------
1334 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1335 //
1336 // OBJECT: klass_word installed; klass_word != 0;
1337 // obj->size() computes correct size
1338 // [Perm Gen comment above continues to hold]
1339 //
1340 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1341 //
1342 //
1343 // STATE TRANSITION DIAGRAM
1344 //
1345 // mut / parnew mut / parnew
1346 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1347 // ^ |
1348 // |------------------------ DEAD <------------------------------------|
1349 // sweep mut
1350 //
1351 // While a block is in TRANSIENT state its size cannot be determined
1352 // so readers will either need to come back later or stall until
1353 // the size can be determined. Note that for the case of direct
1354 // allocation, P-bits, when available, may be used to determine the
1355 // size of an object that may not yet have been initialized.
1356
1357 // Things to support parallel young-gen collection.
1358 oop
1359 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1360 oop old, markOop m,
1361 size_t word_sz) {
1362 #ifndef PRODUCT
1363 if (Universe::heap()->promotion_should_fail()) {
1364 return NULL;
1365 }
1366 #endif // #ifndef PRODUCT
1367
1368 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1369 PromotionInfo* promoInfo = &ps->promo;
1370 // if we are tracking promotions, then first ensure space for
1371 // promotion (including spooling space for saving header if necessary).
1372 // then allocate and copy, then track promoted info if needed.
1373 // When tracking (see PromotionInfo::track()), the mark word may
1374 // be displaced and in this case restoration of the mark word
1375 // occurs in the (oop_since_save_marks_)iterate phase.
1376 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1377 // Out of space for allocating spooling buffers;
1378 // try expanding and allocating spooling buffers.
1379 if (!expand_and_ensure_spooling_space(promoInfo)) {
1380 return NULL;
1381 }
1382 }
1383 assert(promoInfo->has_spooling_space(), "Control point invariant");
1384 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1385 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1386 if (obj_ptr == NULL) {
1387 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1388 if (obj_ptr == NULL) {
1389 return NULL;
1390 }
1391 }
1392 oop obj = oop(obj_ptr);
1393 OrderAccess::storestore();
1394 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1395 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1396 // IMPORTANT: See note on object initialization for CMS above.
1397 // Otherwise, copy the object. Here we must be careful to insert the
1398 // klass pointer last, since this marks the block as an allocated object.
1399 // Except with compressed oops it's the mark word.
1400 HeapWord* old_ptr = (HeapWord*)old;
1401 // Restore the mark word copied above.
1402 obj->set_mark(m);
1403 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1404 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1405 OrderAccess::storestore();
1406
1407 if (UseCompressedOops) {
1408 // Copy gap missed by (aligned) header size calculation below
1409 obj->set_klass_gap(old->klass_gap());
1410 }
1411 if (word_sz > (size_t)oopDesc::header_size()) {
1412 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1413 obj_ptr + oopDesc::header_size(),
1414 word_sz - oopDesc::header_size());
1415 }
1416
1417 // Now we can track the promoted object, if necessary. We take care
1418 // to delay the transition from uninitialized to full object
1419 // (i.e., insertion of klass pointer) until after, so that it
1420 // atomically becomes a promoted object.
1421 if (promoInfo->tracking()) {
1422 promoInfo->track((PromotedObject*)obj, old->klass());
1423 }
1424 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1425 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1426 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1427
1428 // Finally, install the klass pointer (this should be volatile).
1429 OrderAccess::storestore();
1430 obj->set_klass(old->klass());
1431 // We should now be able to calculate the right size for this object
1432 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1433
1434 collector()->promoted(true, // parallel
1435 obj_ptr, old->is_objArray(), word_sz);
1436
1437 NOT_PRODUCT(
1438 Atomic::inc_ptr(&_numObjectsPromoted);
1439 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1440 )
1441
1442 return obj;
1443 }
1444
1445 void
1446 ConcurrentMarkSweepGeneration::
1447 par_promote_alloc_undo(int thread_num,
1448 HeapWord* obj, size_t word_sz) {
1449 // CMS does not support promotion undo.
1450 ShouldNotReachHere();
1451 }
1452
1453 void
1454 ConcurrentMarkSweepGeneration::
1455 par_promote_alloc_done(int thread_num) {
1456 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1457 ps->lab.retire(thread_num);
1458 }
1459
1460 void
1461 ConcurrentMarkSweepGeneration::
1462 par_oop_since_save_marks_iterate_done(int thread_num) {
1463 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1464 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1465 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1466 }
1467
1468 // XXXPERM
1469 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1470 size_t size,
1471 bool tlab)
1472 {
1473 // We allow a STW collection only if a full
1474 // collection was requested.
1475 return full || should_allocate(size, tlab); // FIX ME !!!
1476 // This and promotion failure handling are connected at the
1477 // hip and should be fixed by untying them.
1478 }
1479
1480 bool CMSCollector::shouldConcurrentCollect() {
1481 if (_full_gc_requested) {
1482 if (Verbose && PrintGCDetails) {
1483 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1484 " gc request (or gc_locker)");
1485 }
1486 return true;
1487 }
1488
1489 // For debugging purposes, change the type of collection.
1490 // If the rotation is not on the concurrent collection
1491 // type, don't start a concurrent collection.
1492 NOT_PRODUCT(
1493 if (RotateCMSCollectionTypes &&
1494 (_cmsGen->debug_collection_type() !=
1495 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
1496 assert(_cmsGen->debug_collection_type() !=
1497 ConcurrentMarkSweepGeneration::Unknown_collection_type,
1498 "Bad cms collection type");
1499 return false;
1500 }
1501 )
1502
1503 FreelistLocker x(this);
1504 // ------------------------------------------------------------------
1505 // Print out lots of information which affects the initiation of
1506 // a collection.
1507 if (PrintCMSInitiationStatistics && stats().valid()) {
1508 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1509 gclog_or_tty->stamp();
1510 gclog_or_tty->print_cr("");
1511 stats().print_on(gclog_or_tty);
1512 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1513 stats().time_until_cms_gen_full());
1514 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1515 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1516 _cmsGen->contiguous_available());
1517 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1518 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1519 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1520 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1521 gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy());
1522 }
1523 // ------------------------------------------------------------------
1524
1525 // If the estimated time to complete a cms collection (cms_duration())
1526 // is less than the estimated time remaining until the cms generation
1527 // is full, start a collection.
1528 if (!UseCMSInitiatingOccupancyOnly) {
1529 if (stats().valid()) {
1530 if (stats().time_until_cms_start() == 0.0) {
1531 return true;
1532 }
1533 } else {
1534 // We want to conservatively collect somewhat early in order
1535 // to try and "bootstrap" our CMS/promotion statistics;
1536 // this branch will not fire after the first successful CMS
1537 // collection because the stats should then be valid.
1538 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1539 if (Verbose && PrintGCDetails) {
1540 gclog_or_tty->print_cr(
1541 " CMSCollector: collect for bootstrapping statistics:"
1542 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1543 _bootstrap_occupancy);
1544 }
1545 return true;
1546 }
1547 }
1548 }
1549
1550 // Otherwise, we start a collection cycle if either the perm gen or
1551 // old gen want a collection cycle started. Each may use
1552 // an appropriate criterion for making this decision.
1553 // XXX We need to make sure that the gen expansion
1554 // criterion dovetails well with this. XXX NEED TO FIX THIS
1555 if (_cmsGen->should_concurrent_collect()) {
1556 if (Verbose && PrintGCDetails) {
1557 gclog_or_tty->print_cr("CMS old gen initiated");
1558 }
1559 return true;
1560 }
1561
1562 // We start a collection if we believe an incremental collection may fail;
1563 // this is not likely to be productive in practice because it's probably too
1564 // late anyway.
1565 GenCollectedHeap* gch = GenCollectedHeap::heap();
1566 assert(gch->collector_policy()->is_two_generation_policy(),
1567 "You may want to check the correctness of the following");
1568 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1569 if (Verbose && PrintGCDetails) {
1570 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1571 }
1572 return true;
1573 }
1574
1575 if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) {
1576 bool res = update_should_unload_classes();
1577 if (res) {
1578 if (Verbose && PrintGCDetails) {
1579 gclog_or_tty->print_cr("CMS perm gen initiated");
1580 }
1581 return true;
1582 }
1583 }
1584 return false;
1585 }
1586
1587 // Clear _expansion_cause fields of constituent generations
1588 void CMSCollector::clear_expansion_cause() {
1589 _cmsGen->clear_expansion_cause();
1590 _permGen->clear_expansion_cause();
1591 }
1592
1593 // We should be conservative in starting a collection cycle. To
1594 // start too eagerly runs the risk of collecting too often in the
1595 // extreme. To collect too rarely falls back on full collections,
1596 // which works, even if not optimum in terms of concurrent work.
1597 // As a work around for too eagerly collecting, use the flag
1598 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1599 // giving the user an easily understandable way of controlling the
1600 // collections.
1601 // We want to start a new collection cycle if any of the following
1602 // conditions hold:
1603 // . our current occupancy exceeds the configured initiating occupancy
1604 // for this generation, or
1605 // . we recently needed to expand this space and have not, since that
1606 // expansion, done a collection of this generation, or
1607 // . the underlying space believes that it may be a good idea to initiate
1608 // a concurrent collection (this may be based on criteria such as the
1609 // following: the space uses linear allocation and linear allocation is
1610 // going to fail, or there is believed to be excessive fragmentation in
1611 // the generation, etc... or ...
1612 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1613 // the case of the old generation, not the perm generation; see CR 6543076):
1614 // we may be approaching a point at which allocation requests may fail because
1615 // we will be out of sufficient free space given allocation rate estimates.]
1616 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1617
1618 assert_lock_strong(freelistLock());
1619 if (occupancy() > initiating_occupancy()) {
1620 if (PrintGCDetails && Verbose) {
1621 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1622 short_name(), occupancy(), initiating_occupancy());
1623 }
1624 return true;
1625 }
1626 if (UseCMSInitiatingOccupancyOnly) {
1627 return false;
1628 }
1629 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1630 if (PrintGCDetails && Verbose) {
1631 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1632 short_name());
1633 }
1634 return true;
1635 }
1636 if (_cmsSpace->should_concurrent_collect()) {
1637 if (PrintGCDetails && Verbose) {
1638 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1639 short_name());
1640 }
1641 return true;
1642 }
1643 return false;
1644 }
1645
1646 void ConcurrentMarkSweepGeneration::collect(bool full,
1647 bool clear_all_soft_refs,
1648 size_t size,
1649 bool tlab)
1650 {
1651 collector()->collect(full, clear_all_soft_refs, size, tlab);
1652 }
1653
1654 void CMSCollector::collect(bool full,
1655 bool clear_all_soft_refs,
1656 size_t size,
1657 bool tlab)
1658 {
1659 if (!UseCMSCollectionPassing && _collectorState > Idling) {
1660 // For debugging purposes skip the collection if the state
1661 // is not currently idle
1662 if (TraceCMSState) {
1663 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
1664 Thread::current(), full, _collectorState);
1665 }
1666 return;
1667 }
1668
1669 // The following "if" branch is present for defensive reasons.
1670 // In the current uses of this interface, it can be replaced with:
1671 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1672 // But I am not placing that assert here to allow future
1673 // generality in invoking this interface.
1674 if (GC_locker::is_active()) {
1675 // A consistency test for GC_locker
1676 assert(GC_locker::needs_gc(), "Should have been set already");
1677 // Skip this foreground collection, instead
1678 // expanding the heap if necessary.
1679 // Need the free list locks for the call to free() in compute_new_size()
1680 compute_new_size();
1681 return;
1682 }
1683 acquire_control_and_collect(full, clear_all_soft_refs);
1684 _full_gcs_since_conc_gc++;
1685
1686 }
1687
1688 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1689 GenCollectedHeap* gch = GenCollectedHeap::heap();
1690 unsigned int gc_count = gch->total_full_collections();
1691 if (gc_count == full_gc_count) {
1692 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1693 _full_gc_requested = true;
1694 _full_gc_cause = cause;
1695 CGC_lock->notify(); // nudge CMS thread
1696 } else {
1697 assert(gc_count > full_gc_count, "Error: causal loop");
1698 }
1699 }
1700
1701 bool CMSCollector::is_external_interruption() {
1702 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1703 return GCCause::is_user_requested_gc(cause) ||
1704 GCCause::is_serviceability_requested_gc(cause);
1705 }
1706
1707 void CMSCollector::report_concurrent_mode_interruption() {
1708 if (is_external_interruption()) {
1709 if (PrintGCDetails) {
1710 gclog_or_tty->print(" (concurrent mode interrupted)");
1711 }
1712 } else {
1713 if (PrintGCDetails) {
1714 gclog_or_tty->print(" (concurrent mode failure)");
1715 }
1716 _gc_tracer_cm->report_concurrent_mode_failure();
1717 }
1718 }
1719
1720 // The foreground and background collectors need to coordinate in order
1721 // to make sure that they do not mutually interfere with CMS collections.
1722 // When a background collection is active,
1723 // the foreground collector may need to take over (preempt) and
1724 // synchronously complete an ongoing collection. Depending on the
1725 // frequency of the background collections and the heap usage
1726 // of the application, this preemption can be seldom or frequent.
1727 // There are only certain
1728 // points in the background collection that the "collection-baton"
1729 // can be passed to the foreground collector.
1730 //
1731 // The foreground collector will wait for the baton before
1732 // starting any part of the collection. The foreground collector
1733 // will only wait at one location.
1734 //
1735 // The background collector will yield the baton before starting a new
1736 // phase of the collection (e.g., before initial marking, marking from roots,
1737 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1738 // of the loop which switches the phases. The background collector does some
1739 // of the phases (initial mark, final re-mark) with the world stopped.
1740 // Because of locking involved in stopping the world,
1741 // the foreground collector should not block waiting for the background
1742 // collector when it is doing a stop-the-world phase. The background
1743 // collector will yield the baton at an additional point just before
1744 // it enters a stop-the-world phase. Once the world is stopped, the
1745 // background collector checks the phase of the collection. If the
1746 // phase has not changed, it proceeds with the collection. If the
1747 // phase has changed, it skips that phase of the collection. See
1748 // the comments on the use of the Heap_lock in collect_in_background().
1749 //
1750 // Variable used in baton passing.
1751 // _foregroundGCIsActive - Set to true by the foreground collector when
1752 // it wants the baton. The foreground clears it when it has finished
1753 // the collection.
1754 // _foregroundGCShouldWait - Set to true by the background collector
1755 // when it is running. The foreground collector waits while
1756 // _foregroundGCShouldWait is true.
1757 // CGC_lock - monitor used to protect access to the above variables
1758 // and to notify the foreground and background collectors.
1759 // _collectorState - current state of the CMS collection.
1760 //
1761 // The foreground collector
1762 // acquires the CGC_lock
1763 // sets _foregroundGCIsActive
1764 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1765 // various locks acquired in preparation for the collection
1766 // are released so as not to block the background collector
1767 // that is in the midst of a collection
1768 // proceeds with the collection
1769 // clears _foregroundGCIsActive
1770 // returns
1771 //
1772 // The background collector in a loop iterating on the phases of the
1773 // collection
1774 // acquires the CGC_lock
1775 // sets _foregroundGCShouldWait
1776 // if _foregroundGCIsActive is set
1777 // clears _foregroundGCShouldWait, notifies _CGC_lock
1778 // waits on _CGC_lock for _foregroundGCIsActive to become false
1779 // and exits the loop.
1780 // otherwise
1781 // proceed with that phase of the collection
1782 // if the phase is a stop-the-world phase,
1783 // yield the baton once more just before enqueueing
1784 // the stop-world CMS operation (executed by the VM thread).
1785 // returns after all phases of the collection are done
1786 //
1787
1788 void CMSCollector::acquire_control_and_collect(bool full,
1789 bool clear_all_soft_refs) {
1790 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1791 assert(!Thread::current()->is_ConcurrentGC_thread(),
1792 "shouldn't try to acquire control from self!");
1793
1794 // Start the protocol for acquiring control of the
1795 // collection from the background collector (aka CMS thread).
1796 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1797 "VM thread should have CMS token");
1798 // Remember the possibly interrupted state of an ongoing
1799 // concurrent collection
1800 CollectorState first_state = _collectorState;
1801
1802 // Signal to a possibly ongoing concurrent collection that
1803 // we want to do a foreground collection.
1804 _foregroundGCIsActive = true;
1805
1806 // Disable incremental mode during a foreground collection.
1807 ICMSDisabler icms_disabler;
1808
1809 // release locks and wait for a notify from the background collector
1810 // releasing the locks in only necessary for phases which
1811 // do yields to improve the granularity of the collection.
1812 assert_lock_strong(bitMapLock());
1813 // We need to lock the Free list lock for the space that we are
1814 // currently collecting.
1815 assert(haveFreelistLocks(), "Must be holding free list locks");
1816 bitMapLock()->unlock();
1817 releaseFreelistLocks();
1818 {
1819 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1820 if (_foregroundGCShouldWait) {
1821 // We are going to be waiting for action for the CMS thread;
1822 // it had better not be gone (for instance at shutdown)!
1823 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1824 "CMS thread must be running");
1825 // Wait here until the background collector gives us the go-ahead
1826 ConcurrentMarkSweepThread::clear_CMS_flag(
1827 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1828 // Get a possibly blocked CMS thread going:
1829 // Note that we set _foregroundGCIsActive true above,
1830 // without protection of the CGC_lock.
1831 CGC_lock->notify();
1832 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1833 "Possible deadlock");
1834 while (_foregroundGCShouldWait) {
1835 // wait for notification
1836 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1837 // Possibility of delay/starvation here, since CMS token does
1838 // not know to give priority to VM thread? Actually, i think
1839 // there wouldn't be any delay/starvation, but the proof of
1840 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1841 }
1842 ConcurrentMarkSweepThread::set_CMS_flag(
1843 ConcurrentMarkSweepThread::CMS_vm_has_token);
1844 }
1845 }
1846 // The CMS_token is already held. Get back the other locks.
1847 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1848 "VM thread should have CMS token");
1849 getFreelistLocks();
1850 bitMapLock()->lock_without_safepoint_check();
1851 if (TraceCMSState) {
1852 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1853 INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1854 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1855 }
1856
1857 // Check if we need to do a compaction, or if not, whether
1858 // we need to start the mark-sweep from scratch.
1859 bool should_compact = false;
1860 bool should_start_over = false;
1861 decide_foreground_collection_type(clear_all_soft_refs,
1862 &should_compact, &should_start_over);
1863
1864 NOT_PRODUCT(
1865 if (RotateCMSCollectionTypes) {
1866 if (_cmsGen->debug_collection_type() ==
1867 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1868 should_compact = true;
1869 } else if (_cmsGen->debug_collection_type() ==
1870 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1871 should_compact = false;
1872 }
1873 }
1874 )
1875
1876 if (first_state > Idling) {
1877 report_concurrent_mode_interruption();
1878 }
1879
1880 if (should_compact) {
1881 // If the collection is being acquired from the background
1882 // collector, there may be references on the discovered
1883 // references lists that have NULL referents (being those
1884 // that were concurrently cleared by a mutator) or
1885 // that are no longer active (having been enqueued concurrently
1886 // by the mutator).
1887 // Scrub the list of those references because Mark-Sweep-Compact
1888 // code assumes referents are not NULL and that all discovered
1889 // Reference objects are active.
1890 ref_processor()->clean_up_discovered_references();
1891
1892 if (first_state > Idling) {
1893 save_heap_summary();
1894 }
1895
1896 do_compaction_work(clear_all_soft_refs);
1897
1898 // Has the GC time limit been exceeded?
1899 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration();
1900 size_t max_eden_size = young_gen->max_capacity() -
1901 young_gen->to()->capacity() -
1902 young_gen->from()->capacity();
1903 GenCollectedHeap* gch = GenCollectedHeap::heap();
1904 GCCause::Cause gc_cause = gch->gc_cause();
1905 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1906 young_gen->eden()->used(),
1907 _cmsGen->max_capacity(),
1908 max_eden_size,
1909 full,
1910 gc_cause,
1911 gch->collector_policy());
1912 } else {
1913 do_mark_sweep_work(clear_all_soft_refs, first_state,
1914 should_start_over);
1915 }
1916 // Reset the expansion cause, now that we just completed
1917 // a collection cycle.
1918 clear_expansion_cause();
1919 _foregroundGCIsActive = false;
1920 return;
1921 }
1922
1923 // Resize the perm generation and the tenured generation
1924 // after obtaining the free list locks for the
1925 // two generations.
1926 void CMSCollector::compute_new_size() {
1927 assert_locked_or_safepoint(Heap_lock);
1928 FreelistLocker z(this);
1929 _permGen->compute_new_size();
1930 _cmsGen->compute_new_size();
1931 }
1932
1933 // A work method used by foreground collection to determine
1934 // what type of collection (compacting or not, continuing or fresh)
1935 // it should do.
1936 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1937 // and CMSCompactWhenClearAllSoftRefs the default in the future
1938 // and do away with the flags after a suitable period.
1939 void CMSCollector::decide_foreground_collection_type(
1940 bool clear_all_soft_refs, bool* should_compact,
1941 bool* should_start_over) {
1942 // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1943 // flag is set, and we have either requested a System.gc() or
1944 // the number of full gc's since the last concurrent cycle
1945 // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1946 // or if an incremental collection has failed
1947 GenCollectedHeap* gch = GenCollectedHeap::heap();
1948 assert(gch->collector_policy()->is_two_generation_policy(),
1949 "You may want to check the correctness of the following");
1950 // Inform cms gen if this was due to partial collection failing.
1951 // The CMS gen may use this fact to determine its expansion policy.
1952 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1953 assert(!_cmsGen->incremental_collection_failed(),
1954 "Should have been noticed, reacted to and cleared");
1955 _cmsGen->set_incremental_collection_failed();
1956 }
1957 *should_compact =
1958 UseCMSCompactAtFullCollection &&
1959 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1960 GCCause::is_user_requested_gc(gch->gc_cause()) ||
1961 gch->incremental_collection_will_fail(true /* consult_young */));
1962 *should_start_over = false;
1963 if (clear_all_soft_refs && !*should_compact) {
1964 // We are about to do a last ditch collection attempt
1965 // so it would normally make sense to do a compaction
1966 // to reclaim as much space as possible.
1967 if (CMSCompactWhenClearAllSoftRefs) {
1968 // Default: The rationale is that in this case either
1969 // we are past the final marking phase, in which case
1970 // we'd have to start over, or so little has been done
1971 // that there's little point in saving that work. Compaction
1972 // appears to be the sensible choice in either case.
1973 *should_compact = true;
1974 } else {
1975 // We have been asked to clear all soft refs, but not to
1976 // compact. Make sure that we aren't past the final checkpoint
1977 // phase, for that is where we process soft refs. If we are already
1978 // past that phase, we'll need to redo the refs discovery phase and
1979 // if necessary clear soft refs that weren't previously
1980 // cleared. We do so by remembering the phase in which
1981 // we came in, and if we are past the refs processing
1982 // phase, we'll choose to just redo the mark-sweep
1983 // collection from scratch.
1984 if (_collectorState > FinalMarking) {
1985 // We are past the refs processing phase;
1986 // start over and do a fresh synchronous CMS cycle
1987 _collectorState = Resetting; // skip to reset to start new cycle
1988 reset(false /* == !asynch */);
1989 *should_start_over = true;
1990 } // else we can continue a possibly ongoing current cycle
1991 }
1992 }
1993 }
1994
1995 // A work method used by the foreground collector to do
1996 // a mark-sweep-compact.
1997 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1998 GenCollectedHeap* gch = GenCollectedHeap::heap();
1999
2000 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
2001 gc_timer->register_gc_start(os::elapsed_counter());
2002
2003 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
2004 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
2005
2006 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL);
2007 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
2008 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
2009 "collections passed to foreground collector", _full_gcs_since_conc_gc);
2010 }
2011
2012 // Sample collection interval time and reset for collection pause.
2013 if (UseAdaptiveSizePolicy) {
2014 size_policy()->msc_collection_begin();
2015 }
2016
2017 // Temporarily widen the span of the weak reference processing to
2018 // the entire heap.
2019 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
2020 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
2021 // Temporarily, clear the "is_alive_non_header" field of the
2022 // reference processor.
2023 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
2024 // Temporarily make reference _processing_ single threaded (non-MT).
2025 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
2026 // Temporarily make refs discovery atomic
2027 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
2028 // Temporarily make reference _discovery_ single threaded (non-MT)
2029 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
2030
2031 ref_processor()->set_enqueuing_is_done(false);
2032 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/);
2033 ref_processor()->setup_policy(clear_all_soft_refs);
2034 // If an asynchronous collection finishes, the _modUnionTable is
2035 // all clear. If we are assuming the collection from an asynchronous
2036 // collection, clear the _modUnionTable.
2037 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
2038 "_modUnionTable should be clear if the baton was not passed");
2039 _modUnionTable.clear_all();
2040
2041 // We must adjust the allocation statistics being maintained
2042 // in the free list space. We do so by reading and clearing
2043 // the sweep timer and updating the block flux rate estimates below.
2044 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
2045 if (_inter_sweep_timer.is_active()) {
2046 _inter_sweep_timer.stop();
2047 // Note that we do not use this sample to update the _inter_sweep_estimate.
2048 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
2049 _inter_sweep_estimate.padded_average(),
2050 _intra_sweep_estimate.padded_average());
2051 }
2052
2053 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
2054 ref_processor(), clear_all_soft_refs);
2055 #ifdef ASSERT
2056 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
2057 size_t free_size = cms_space->free();
2058 assert(free_size ==
2059 pointer_delta(cms_space->end(), cms_space->compaction_top())
2060 * HeapWordSize,
2061 "All the free space should be compacted into one chunk at top");
2062 assert(cms_space->dictionary()->total_chunk_size(
2063 debug_only(cms_space->freelistLock())) == 0 ||
2064 cms_space->totalSizeInIndexedFreeLists() == 0,
2065 "All the free space should be in a single chunk");
2066 size_t num = cms_space->totalCount();
2067 assert((free_size == 0 && num == 0) ||
2068 (free_size > 0 && (num == 1 || num == 2)),
2069 "There should be at most 2 free chunks after compaction");
2070 #endif // ASSERT
2071 _collectorState = Resetting;
2072 assert(_restart_addr == NULL,
2073 "Should have been NULL'd before baton was passed");
2074 reset(false /* == !asynch */);
2075 _cmsGen->reset_after_compaction();
2076 _concurrent_cycles_since_last_unload = 0;
2077
2078 if (verifying() && !should_unload_classes()) {
2079 perm_gen_verify_bit_map()->clear_all();
2080 }
2081
2082 // Clear any data recorded in the PLAB chunk arrays.
2083 if (_survivor_plab_array != NULL) {
2084 reset_survivor_plab_arrays();
2085 }
2086
2087 // Adjust the per-size allocation stats for the next epoch.
2088 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
2089 // Restart the "inter sweep timer" for the next epoch.
2090 _inter_sweep_timer.reset();
2091 _inter_sweep_timer.start();
2092
2093 // Sample collection pause time and reset for collection interval.
2094 if (UseAdaptiveSizePolicy) {
2095 size_policy()->msc_collection_end(gch->gc_cause());
2096 }
2097
2098 gc_timer->register_gc_end(os::elapsed_counter());
2099
2100 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
2101
2102 // For a mark-sweep-compact, compute_new_size() will be called
2103 // in the heap's do_collection() method.
2104 }
2105
2106 // A work method used by the foreground collector to do
2107 // a mark-sweep, after taking over from a possibly on-going
2108 // concurrent mark-sweep collection.
2109 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2110 CollectorState first_state, bool should_start_over) {
2111 if (PrintGC && Verbose) {
2112 gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2113 "collector with count %d",
2114 _full_gcs_since_conc_gc);
2115 }
2116 switch (_collectorState) {
2117 case Idling:
2118 if (first_state == Idling || should_start_over) {
2119 // The background GC was not active, or should
2120 // restarted from scratch; start the cycle.
2121 _collectorState = InitialMarking;
2122 }
2123 // If first_state was not Idling, then a background GC
2124 // was in progress and has now finished. No need to do it
2125 // again. Leave the state as Idling.
2126 break;
2127 case Precleaning:
2128 // In the foreground case don't do the precleaning since
2129 // it is not done concurrently and there is extra work
2130 // required.
2131 _collectorState = FinalMarking;
2132 }
2133 collect_in_foreground(clear_all_soft_refs, GenCollectedHeap::heap()->gc_cause());
2134
2135 // For a mark-sweep, compute_new_size() will be called
2136 // in the heap's do_collection() method.
2137 }
2138
2139
2140 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
2141 DefNewGeneration* dng = _young_gen->as_DefNewGeneration();
2142 EdenSpace* eden_space = dng->eden();
2143 ContiguousSpace* from_space = dng->from();
2144 ContiguousSpace* to_space = dng->to();
2145 // Eden
2146 if (_eden_chunk_array != NULL) {
2147 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
2148 eden_space->bottom(), eden_space->top(),
2149 eden_space->end(), eden_space->capacity());
2150 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", "
2151 "_eden_chunk_capacity=" SIZE_FORMAT,
2152 _eden_chunk_index, _eden_chunk_capacity);
2153 for (size_t i = 0; i < _eden_chunk_index; i++) {
2154 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
2155 i, _eden_chunk_array[i]);
2156 }
2157 }
2158 // Survivor
2159 if (_survivor_chunk_array != NULL) {
2160 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
2161 from_space->bottom(), from_space->top(),
2162 from_space->end(), from_space->capacity());
2163 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", "
2164 "_survivor_chunk_capacity=" SIZE_FORMAT,
2165 _survivor_chunk_index, _survivor_chunk_capacity);
2166 for (size_t i = 0; i < _survivor_chunk_index; i++) {
2167 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
2168 i, _survivor_chunk_array[i]);
2169 }
2170 }
2171 }
2172
2173 void CMSCollector::getFreelistLocks() const {
2174 // Get locks for all free lists in all generations that this
2175 // collector is responsible for
2176 _cmsGen->freelistLock()->lock_without_safepoint_check();
2177 _permGen->freelistLock()->lock_without_safepoint_check();
2178 }
2179
2180 void CMSCollector::releaseFreelistLocks() const {
2181 // Release locks for all free lists in all generations that this
2182 // collector is responsible for
2183 _cmsGen->freelistLock()->unlock();
2184 _permGen->freelistLock()->unlock();
2185 }
2186
2187 bool CMSCollector::haveFreelistLocks() const {
2188 // Check locks for all free lists in all generations that this
2189 // collector is responsible for
2190 assert_lock_strong(_cmsGen->freelistLock());
2191 assert_lock_strong(_permGen->freelistLock());
2192 PRODUCT_ONLY(ShouldNotReachHere());
2193 return true;
2194 }
2195
2196 // A utility class that is used by the CMS collector to
2197 // temporarily "release" the foreground collector from its
2198 // usual obligation to wait for the background collector to
2199 // complete an ongoing phase before proceeding.
2200 class ReleaseForegroundGC: public StackObj {
2201 private:
2202 CMSCollector* _c;
2203 public:
2204 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2205 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2206 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2207 // allow a potentially blocked foreground collector to proceed
2208 _c->_foregroundGCShouldWait = false;
2209 if (_c->_foregroundGCIsActive) {
2210 CGC_lock->notify();
2211 }
2212 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2213 "Possible deadlock");
2214 }
2215
2216 ~ReleaseForegroundGC() {
2217 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2218 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2219 _c->_foregroundGCShouldWait = true;
2220 }
2221 };
2222
2223 // There are separate collect_in_background and collect_in_foreground because of
2224 // the different locking requirements of the background collector and the
2225 // foreground collector. There was originally an attempt to share
2226 // one "collect" method between the background collector and the foreground
2227 // collector but the if-then-else required made it cleaner to have
2228 // separate methods.
2229 void CMSCollector::collect_in_background(bool clear_all_soft_refs, GCCause::Cause cause) {
2230 assert(Thread::current()->is_ConcurrentGC_thread(),
2231 "A CMS asynchronous collection is only allowed on a CMS thread.");
2232
2233 GenCollectedHeap* gch = GenCollectedHeap::heap();
2234 {
2235 bool safepoint_check = Mutex::_no_safepoint_check_flag;
2236 MutexLockerEx hl(Heap_lock, safepoint_check);
2237 FreelistLocker fll(this);
2238 MutexLockerEx x(CGC_lock, safepoint_check);
2239 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2240 // The foreground collector is active or we're
2241 // not using asynchronous collections. Skip this
2242 // background collection.
2243 assert(!_foregroundGCShouldWait, "Should be clear");
2244 return;
2245 } else {
2246 assert(_collectorState == Idling, "Should be idling before start.");
2247 _collectorState = InitialMarking;
2248 register_gc_start(cause);
2249 // Reset the expansion cause, now that we are about to begin
2250 // a new cycle.
2251 clear_expansion_cause();
2252 }
2253 // Decide if we want to enable class unloading as part of the
2254 // ensuing concurrent GC cycle.
2255 update_should_unload_classes();
2256 _full_gc_requested = false; // acks all outstanding full gc requests
2257 _full_gc_cause = GCCause::_no_gc;
2258 // Signal that we are about to start a collection
2259 gch->increment_total_full_collections(); // ... starting a collection cycle
2260 _collection_count_start = gch->total_full_collections();
2261 }
2262
2263 // Used for PrintGC
2264 size_t prev_used;
2265 if (PrintGC && Verbose) {
2266 prev_used = _cmsGen->used(); // XXXPERM
2267 }
2268
2269 // The change of the collection state is normally done at this level;
2270 // the exceptions are phases that are executed while the world is
2271 // stopped. For those phases the change of state is done while the
2272 // world is stopped. For baton passing purposes this allows the
2273 // background collector to finish the phase and change state atomically.
2274 // The foreground collector cannot wait on a phase that is done
2275 // while the world is stopped because the foreground collector already
2276 // has the world stopped and would deadlock.
2277 while (_collectorState != Idling) {
2278 if (TraceCMSState) {
2279 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2280 Thread::current(), _collectorState);
2281 }
2282 // The foreground collector
2283 // holds the Heap_lock throughout its collection.
2284 // holds the CMS token (but not the lock)
2285 // except while it is waiting for the background collector to yield.
2286 //
2287 // The foreground collector should be blocked (not for long)
2288 // if the background collector is about to start a phase
2289 // executed with world stopped. If the background
2290 // collector has already started such a phase, the
2291 // foreground collector is blocked waiting for the
2292 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
2293 // are executed in the VM thread.
2294 //
2295 // The locking order is
2296 // PendingListLock (PLL) -- if applicable (FinalMarking)
2297 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
2298 // CMS token (claimed in
2299 // stop_world_and_do() -->
2300 // safepoint_synchronize() -->
2301 // CMSThread::synchronize())
2302
2303 {
2304 // Check if the FG collector wants us to yield.
2305 CMSTokenSync x(true); // is cms thread
2306 if (waitForForegroundGC()) {
2307 // We yielded to a foreground GC, nothing more to be
2308 // done this round.
2309 assert(_foregroundGCShouldWait == false, "We set it to false in "
2310 "waitForForegroundGC()");
2311 if (TraceCMSState) {
2312 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2313 " exiting collection CMS state %d",
2314 Thread::current(), _collectorState);
2315 }
2316 return;
2317 } else {
2318 // The background collector can run but check to see if the
2319 // foreground collector has done a collection while the
2320 // background collector was waiting to get the CGC_lock
2321 // above. If yes, break so that _foregroundGCShouldWait
2322 // is cleared before returning.
2323 if (_collectorState == Idling) {
2324 break;
2325 }
2326 }
2327 }
2328
2329 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2330 "should be waiting");
2331
2332 switch (_collectorState) {
2333 case InitialMarking:
2334 {
2335 ReleaseForegroundGC x(this);
2336 stats().record_cms_begin();
2337 VM_CMS_Initial_Mark initial_mark_op(this);
2338 VMThread::execute(&initial_mark_op);
2339 }
2340 // The collector state may be any legal state at this point
2341 // since the background collector may have yielded to the
2342 // foreground collector.
2343 break;
2344 case Marking:
2345 // initial marking in checkpointRootsInitialWork has been completed
2346 if (markFromRoots(true)) { // we were successful
2347 assert(_collectorState == Precleaning, "Collector state should "
2348 "have changed");
2349 } else {
2350 assert(_foregroundGCIsActive, "Internal state inconsistency");
2351 }
2352 break;
2353 case Precleaning:
2354 if (UseAdaptiveSizePolicy) {
2355 size_policy()->concurrent_precleaning_begin();
2356 }
2357 // marking from roots in markFromRoots has been completed
2358 preclean();
2359 if (UseAdaptiveSizePolicy) {
2360 size_policy()->concurrent_precleaning_end();
2361 }
2362 assert(_collectorState == AbortablePreclean ||
2363 _collectorState == FinalMarking,
2364 "Collector state should have changed");
2365 break;
2366 case AbortablePreclean:
2367 if (UseAdaptiveSizePolicy) {
2368 size_policy()->concurrent_phases_resume();
2369 }
2370 abortable_preclean();
2371 if (UseAdaptiveSizePolicy) {
2372 size_policy()->concurrent_precleaning_end();
2373 }
2374 assert(_collectorState == FinalMarking, "Collector state should "
2375 "have changed");
2376 break;
2377 case FinalMarking:
2378 {
2379 ReleaseForegroundGC x(this);
2380
2381 VM_CMS_Final_Remark final_remark_op(this);
2382 VMThread::execute(&final_remark_op);
2383 }
2384 assert(_foregroundGCShouldWait, "block post-condition");
2385 break;
2386 case Sweeping:
2387 if (UseAdaptiveSizePolicy) {
2388 size_policy()->concurrent_sweeping_begin();
2389 }
2390 // final marking in checkpointRootsFinal has been completed
2391 sweep(true);
2392 assert(_collectorState == Resizing, "Collector state change "
2393 "to Resizing must be done under the free_list_lock");
2394 _full_gcs_since_conc_gc = 0;
2395
2396 // Stop the timers for adaptive size policy for the concurrent phases
2397 if (UseAdaptiveSizePolicy) {
2398 size_policy()->concurrent_sweeping_end();
2399 size_policy()->concurrent_phases_end(gch->gc_cause(),
2400 gch->prev_gen(_cmsGen)->capacity(),
2401 _cmsGen->free());
2402 }
2403
2404 case Resizing: {
2405 // Sweeping has been completed...
2406 // At this point the background collection has completed.
2407 // Don't move the call to compute_new_size() down
2408 // into code that might be executed if the background
2409 // collection was preempted.
2410 {
2411 ReleaseForegroundGC x(this); // unblock FG collection
2412 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
2413 CMSTokenSync z(true); // not strictly needed.
2414 if (_collectorState == Resizing) {
2415 compute_new_size();
2416 save_heap_summary();
2417 _collectorState = Resetting;
2418 } else {
2419 assert(_collectorState == Idling, "The state should only change"
2420 " because the foreground collector has finished the collection");
2421 }
2422 }
2423 break;
2424 }
2425 case Resetting:
2426 // CMS heap resizing has been completed
2427 reset(true);
2428 assert(_collectorState == Idling, "Collector state should "
2429 "have changed");
2430 stats().record_cms_end();
2431 // Don't move the concurrent_phases_end() and compute_new_size()
2432 // calls to here because a preempted background collection
2433 // has it's state set to "Resetting".
2434 break;
2435 case Idling:
2436 default:
2437 ShouldNotReachHere();
2438 break;
2439 }
2440 if (TraceCMSState) {
2441 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2442 Thread::current(), _collectorState);
2443 }
2444 assert(_foregroundGCShouldWait, "block post-condition");
2445 }
2446
2447 // Should this be in gc_epilogue?
2448 collector_policy()->counters()->update_counters();
2449
2450 {
2451 // Clear _foregroundGCShouldWait and, in the event that the
2452 // foreground collector is waiting, notify it, before
2453 // returning.
2454 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2455 _foregroundGCShouldWait = false;
2456 if (_foregroundGCIsActive) {
2457 CGC_lock->notify();
2458 }
2459 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2460 "Possible deadlock");
2461 }
2462 if (TraceCMSState) {
2463 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2464 " exiting collection CMS state %d",
2465 Thread::current(), _collectorState);
2466 }
2467 if (PrintGC && Verbose) {
2468 _cmsGen->print_heap_change(prev_used);
2469 }
2470 }
2471
2472 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) {
2473 if (!_cms_start_registered) {
2474 register_gc_start(cause);
2475 }
2476 }
2477
2478 void CMSCollector::register_gc_start(GCCause::Cause cause) {
2479 _cms_start_registered = true;
2480 _gc_timer_cm->register_gc_start(os::elapsed_counter());
2481 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
2482 }
2483
2484 void CMSCollector::register_gc_end() {
2485 if (_cms_start_registered) {
2486 report_heap_summary(GCWhen::AfterGC);
2487
2488 _gc_timer_cm->register_gc_end(os::elapsed_counter());
2489 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
2490 _cms_start_registered = false;
2491 }
2492 }
2493
2494 void CMSCollector::save_heap_summary() {
2495 GenCollectedHeap* gch = GenCollectedHeap::heap();
2496 _last_heap_summary = gch->create_heap_summary();
2497 _last_perm_gen_summary = gch->create_perm_gen_summary();
2498 }
2499
2500 void CMSCollector::report_heap_summary(GCWhen::Type when) {
2501 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary, _last_perm_gen_summary);
2502 }
2503
2504 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) {
2505 assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2506 "Foreground collector should be waiting, not executing");
2507 assert(Thread::current()->is_VM_thread(), "A foreground collection"
2508 "may only be done by the VM Thread with the world stopped");
2509 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2510 "VM thread should have CMS token");
2511
2512 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2513 true, NULL);)
2514 if (UseAdaptiveSizePolicy) {
2515 size_policy()->ms_collection_begin();
2516 }
2517 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2518
2519 HandleMark hm; // Discard invalid handles created during verification
2520
2521 if (VerifyBeforeGC &&
2522 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2523 Universe::verify();
2524 }
2525
2526 // Snapshot the soft reference policy to be used in this collection cycle.
2527 ref_processor()->setup_policy(clear_all_soft_refs);
2528
2529 bool init_mark_was_synchronous = false; // until proven otherwise
2530 while (_collectorState != Idling) {
2531 if (TraceCMSState) {
2532 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2533 Thread::current(), _collectorState);
2534 }
2535 switch (_collectorState) {
2536 case InitialMarking:
2537 register_foreground_gc_start(cause);
2538 init_mark_was_synchronous = true; // fact to be exploited in re-mark
2539 checkpointRootsInitial(false);
2540 assert(_collectorState == Marking, "Collector state should have changed"
2541 " within checkpointRootsInitial()");
2542 break;
2543 case Marking:
2544 // initial marking in checkpointRootsInitialWork has been completed
2545 if (VerifyDuringGC &&
2546 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2547 gclog_or_tty->print("Verify before initial mark: ");
2548 Universe::verify();
2549 }
2550 {
2551 bool res = markFromRoots(false);
2552 assert(res && _collectorState == FinalMarking, "Collector state should "
2553 "have changed");
2554 break;
2555 }
2556 case FinalMarking:
2557 if (VerifyDuringGC &&
2558 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2559 gclog_or_tty->print("Verify before re-mark: ");
2560 Universe::verify();
2561 }
2562 checkpointRootsFinal(false, clear_all_soft_refs,
2563 init_mark_was_synchronous);
2564 assert(_collectorState == Sweeping, "Collector state should not "
2565 "have changed within checkpointRootsFinal()");
2566 break;
2567 case Sweeping:
2568 // final marking in checkpointRootsFinal has been completed
2569 if (VerifyDuringGC &&
2570 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2571 gclog_or_tty->print("Verify before sweep: ");
2572 Universe::verify();
2573 }
2574 sweep(false);
2575 assert(_collectorState == Resizing, "Incorrect state");
2576 break;
2577 case Resizing: {
2578 // Sweeping has been completed; the actual resize in this case
2579 // is done separately; nothing to be done in this state.
2580 _collectorState = Resetting;
2581 break;
2582 }
2583 case Resetting:
2584 // The heap has been resized.
2585 if (VerifyDuringGC &&
2586 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2587 gclog_or_tty->print("Verify before reset: ");
2588 Universe::verify();
2589 }
2590 save_heap_summary();
2591 reset(false);
2592 assert(_collectorState == Idling, "Collector state should "
2593 "have changed");
2594 break;
2595 case Precleaning:
2596 case AbortablePreclean:
2597 // Elide the preclean phase
2598 _collectorState = FinalMarking;
2599 break;
2600 default:
2601 ShouldNotReachHere();
2602 }
2603 if (TraceCMSState) {
2604 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2605 Thread::current(), _collectorState);
2606 }
2607 }
2608
2609 if (UseAdaptiveSizePolicy) {
2610 GenCollectedHeap* gch = GenCollectedHeap::heap();
2611 size_policy()->ms_collection_end(gch->gc_cause());
2612 }
2613
2614 if (VerifyAfterGC &&
2615 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2616 Universe::verify();
2617 }
2618 if (TraceCMSState) {
2619 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2620 " exiting collection CMS state %d",
2621 Thread::current(), _collectorState);
2622 }
2623 }
2624
2625 bool CMSCollector::waitForForegroundGC() {
2626 bool res = false;
2627 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2628 "CMS thread should have CMS token");
2629 // Block the foreground collector until the
2630 // background collectors decides whether to
2631 // yield.
2632 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2633 _foregroundGCShouldWait = true;
2634 if (_foregroundGCIsActive) {
2635 // The background collector yields to the
2636 // foreground collector and returns a value
2637 // indicating that it has yielded. The foreground
2638 // collector can proceed.
2639 res = true;
2640 _foregroundGCShouldWait = false;
2641 ConcurrentMarkSweepThread::clear_CMS_flag(
2642 ConcurrentMarkSweepThread::CMS_cms_has_token);
2643 ConcurrentMarkSweepThread::set_CMS_flag(
2644 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2645 // Get a possibly blocked foreground thread going
2646 CGC_lock->notify();
2647 if (TraceCMSState) {
2648 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2649 Thread::current(), _collectorState);
2650 }
2651 while (_foregroundGCIsActive) {
2652 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2653 }
2654 ConcurrentMarkSweepThread::set_CMS_flag(
2655 ConcurrentMarkSweepThread::CMS_cms_has_token);
2656 ConcurrentMarkSweepThread::clear_CMS_flag(
2657 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2658 }
2659 if (TraceCMSState) {
2660 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2661 Thread::current(), _collectorState);
2662 }
2663 return res;
2664 }
2665
2666 // Because of the need to lock the free lists and other structures in
2667 // the collector, common to all the generations that the collector is
2668 // collecting, we need the gc_prologues of individual CMS generations
2669 // delegate to their collector. It may have been simpler had the
2670 // current infrastructure allowed one to call a prologue on a
2671 // collector. In the absence of that we have the generation's
2672 // prologue delegate to the collector, which delegates back
2673 // some "local" work to a worker method in the individual generations
2674 // that it's responsible for collecting, while itself doing any
2675 // work common to all generations it's responsible for. A similar
2676 // comment applies to the gc_epilogue()'s.
2677 // The role of the varaible _between_prologue_and_epilogue is to
2678 // enforce the invocation protocol.
2679 void CMSCollector::gc_prologue(bool full) {
2680 // Call gc_prologue_work() for each CMSGen and PermGen that
2681 // we are responsible for.
2682
2683 // The following locking discipline assumes that we are only called
2684 // when the world is stopped.
2685 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2686
2687 // The CMSCollector prologue must call the gc_prologues for the
2688 // "generations" (including PermGen if any) that it's responsible
2689 // for.
2690
2691 assert( Thread::current()->is_VM_thread()
2692 || ( CMSScavengeBeforeRemark
2693 && Thread::current()->is_ConcurrentGC_thread()),
2694 "Incorrect thread type for prologue execution");
2695
2696 if (_between_prologue_and_epilogue) {
2697 // We have already been invoked; this is a gc_prologue delegation
2698 // from yet another CMS generation that we are responsible for, just
2699 // ignore it since all relevant work has already been done.
2700 return;
2701 }
2702
2703 // set a bit saying prologue has been called; cleared in epilogue
2704 _between_prologue_and_epilogue = true;
2705 // Claim locks for common data structures, then call gc_prologue_work()
2706 // for each CMSGen and PermGen that we are responsible for.
2707
2708 getFreelistLocks(); // gets free list locks on constituent spaces
2709 bitMapLock()->lock_without_safepoint_check();
2710
2711 // Should call gc_prologue_work() for all cms gens we are responsible for
2712 bool registerClosure = _collectorState >= Marking
2713 && _collectorState < Sweeping;
2714 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ?
2715 &_modUnionClosurePar
2716 : &_modUnionClosure;
2717 _cmsGen->gc_prologue_work(full, registerClosure, muc);
2718 _permGen->gc_prologue_work(full, registerClosure, muc);
2719
2720 if (!full) {
2721 stats().record_gc0_begin();
2722 }
2723 }
2724
2725 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2726 // Delegate to CMScollector which knows how to coordinate between
2727 // this and any other CMS generations that it is responsible for
2728 // collecting.
2729 collector()->gc_prologue(full);
2730 }
2731
2732 // This is a "private" interface for use by this generation's CMSCollector.
2733 // Not to be called directly by any other entity (for instance,
2734 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2735 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2736 bool registerClosure, ModUnionClosure* modUnionClosure) {
2737 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2738 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2739 "Should be NULL");
2740 if (registerClosure) {
2741 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2742 }
2743 cmsSpace()->gc_prologue();
2744 // Clear stat counters
2745 NOT_PRODUCT(
2746 assert(_numObjectsPromoted == 0, "check");
2747 assert(_numWordsPromoted == 0, "check");
2748 if (Verbose && PrintGC) {
2749 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2750 SIZE_FORMAT" bytes concurrently",
2751 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2752 }
2753 _numObjectsAllocated = 0;
2754 _numWordsAllocated = 0;
2755 )
2756 }
2757
2758 void CMSCollector::gc_epilogue(bool full) {
2759 // The following locking discipline assumes that we are only called
2760 // when the world is stopped.
2761 assert(SafepointSynchronize::is_at_safepoint(),
2762 "world is stopped assumption");
2763
2764 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2765 // if linear allocation blocks need to be appropriately marked to allow the
2766 // the blocks to be parsable. We also check here whether we need to nudge the
2767 // CMS collector thread to start a new cycle (if it's not already active).
2768 assert( Thread::current()->is_VM_thread()
2769 || ( CMSScavengeBeforeRemark
2770 && Thread::current()->is_ConcurrentGC_thread()),
2771 "Incorrect thread type for epilogue execution");
2772
2773 if (!_between_prologue_and_epilogue) {
2774 // We have already been invoked; this is a gc_epilogue delegation
2775 // from yet another CMS generation that we are responsible for, just
2776 // ignore it since all relevant work has already been done.
2777 return;
2778 }
2779 assert(haveFreelistLocks(), "must have freelist locks");
2780 assert_lock_strong(bitMapLock());
2781
2782 _cmsGen->gc_epilogue_work(full);
2783 _permGen->gc_epilogue_work(full);
2784
2785 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2786 // in case sampling was not already enabled, enable it
2787 _start_sampling = true;
2788 }
2789 // reset _eden_chunk_array so sampling starts afresh
2790 _eden_chunk_index = 0;
2791
2792 size_t cms_used = _cmsGen->cmsSpace()->used();
2793 size_t perm_used = _permGen->cmsSpace()->used();
2794
2795 // update performance counters - this uses a special version of
2796 // update_counters() that allows the utilization to be passed as a
2797 // parameter, avoiding multiple calls to used().
2798 //
2799 _cmsGen->update_counters(cms_used);
2800 _permGen->update_counters(perm_used);
2801
2802 if (CMSIncrementalMode) {
2803 icms_update_allocation_limits();
2804 }
2805
2806 bitMapLock()->unlock();
2807 releaseFreelistLocks();
2808
2809 if (!CleanChunkPoolAsync) {
2810 Chunk::clean_chunk_pool();
2811 }
2812
2813 _between_prologue_and_epilogue = false; // ready for next cycle
2814 }
2815
2816 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2817 collector()->gc_epilogue(full);
2818
2819 // Also reset promotion tracking in par gc thread states.
2820 if (CollectedHeap::use_parallel_gc_threads()) {
2821 for (uint i = 0; i < ParallelGCThreads; i++) {
2822 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2823 }
2824 }
2825 }
2826
2827 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2828 assert(!incremental_collection_failed(), "Should have been cleared");
2829 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2830 cmsSpace()->gc_epilogue();
2831 // Print stat counters
2832 NOT_PRODUCT(
2833 assert(_numObjectsAllocated == 0, "check");
2834 assert(_numWordsAllocated == 0, "check");
2835 if (Verbose && PrintGC) {
2836 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2837 SIZE_FORMAT" bytes",
2838 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2839 }
2840 _numObjectsPromoted = 0;
2841 _numWordsPromoted = 0;
2842 )
2843
2844 if (PrintGC && Verbose) {
2845 // Call down the chain in contiguous_available needs the freelistLock
2846 // so print this out before releasing the freeListLock.
2847 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2848 contiguous_available());
2849 }
2850 }
2851
2852 #ifndef PRODUCT
2853 bool CMSCollector::have_cms_token() {
2854 Thread* thr = Thread::current();
2855 if (thr->is_VM_thread()) {
2856 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2857 } else if (thr->is_ConcurrentGC_thread()) {
2858 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2859 } else if (thr->is_GC_task_thread()) {
2860 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2861 ParGCRareEvent_lock->owned_by_self();
2862 }
2863 return false;
2864 }
2865 #endif
2866
2867 // Check reachability of the given heap address in CMS generation,
2868 // treating all other generations as roots.
2869 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2870 // We could "guarantee" below, rather than assert, but i'll
2871 // leave these as "asserts" so that an adventurous debugger
2872 // could try this in the product build provided some subset of
2873 // the conditions were met, provided they were intersted in the
2874 // results and knew that the computation below wouldn't interfere
2875 // with other concurrent computations mutating the structures
2876 // being read or written.
2877 assert(SafepointSynchronize::is_at_safepoint(),
2878 "Else mutations in object graph will make answer suspect");
2879 assert(have_cms_token(), "Should hold cms token");
2880 assert(haveFreelistLocks(), "must hold free list locks");
2881 assert_lock_strong(bitMapLock());
2882
2883 // Clear the marking bit map array before starting, but, just
2884 // for kicks, first report if the given address is already marked
2885 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
2886 _markBitMap.isMarked(addr) ? "" : " not");
2887
2888 if (verify_after_remark()) {
2889 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2890 bool result = verification_mark_bm()->isMarked(addr);
2891 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
2892 result ? "IS" : "is NOT");
2893 return result;
2894 } else {
2895 gclog_or_tty->print_cr("Could not compute result");
2896 return false;
2897 }
2898 }
2899
2900 ////////////////////////////////////////////////////////
2901 // CMS Verification Support
2902 ////////////////////////////////////////////////////////
2903 // Following the remark phase, the following invariant
2904 // should hold -- each object in the CMS heap which is
2905 // marked in markBitMap() should be marked in the verification_mark_bm().
2906
2907 class VerifyMarkedClosure: public BitMapClosure {
2908 CMSBitMap* _marks;
2909 bool _failed;
2910
2911 public:
2912 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2913
2914 bool do_bit(size_t offset) {
2915 HeapWord* addr = _marks->offsetToHeapWord(offset);
2916 if (!_marks->isMarked(addr)) {
2917 oop(addr)->print_on(gclog_or_tty);
2918 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2919 _failed = true;
2920 }
2921 return true;
2922 }
2923
2924 bool failed() { return _failed; }
2925 };
2926
2927 bool CMSCollector::verify_after_remark() {
2928 gclog_or_tty->print(" [Verifying CMS Marking... ");
2929 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2930 static bool init = false;
2931
2932 assert(SafepointSynchronize::is_at_safepoint(),
2933 "Else mutations in object graph will make answer suspect");
2934 assert(have_cms_token(),
2935 "Else there may be mutual interference in use of "
2936 " verification data structures");
2937 assert(_collectorState > Marking && _collectorState <= Sweeping,
2938 "Else marking info checked here may be obsolete");
2939 assert(haveFreelistLocks(), "must hold free list locks");
2940 assert_lock_strong(bitMapLock());
2941
2942
2943 // Allocate marking bit map if not already allocated
2944 if (!init) { // first time
2945 if (!verification_mark_bm()->allocate(_span)) {
2946 return false;
2947 }
2948 init = true;
2949 }
2950
2951 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2952
2953 // Turn off refs discovery -- so we will be tracing through refs.
2954 // This is as intended, because by this time
2955 // GC must already have cleared any refs that need to be cleared,
2956 // and traced those that need to be marked; moreover,
2957 // the marking done here is not going to intefere in any
2958 // way with the marking information used by GC.
2959 NoRefDiscovery no_discovery(ref_processor());
2960
2961 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2962
2963 // Clear any marks from a previous round
2964 verification_mark_bm()->clear_all();
2965 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2966 verify_work_stacks_empty();
2967
2968 GenCollectedHeap* gch = GenCollectedHeap::heap();
2969 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2970 // Update the saved marks which may affect the root scans.
2971 gch->save_marks();
2972
2973 if (CMSRemarkVerifyVariant == 1) {
2974 // In this first variant of verification, we complete
2975 // all marking, then check if the new marks-verctor is
2976 // a subset of the CMS marks-vector.
2977 verify_after_remark_work_1();
2978 } else if (CMSRemarkVerifyVariant == 2) {
2979 // In this second variant of verification, we flag an error
2980 // (i.e. an object reachable in the new marks-vector not reachable
2981 // in the CMS marks-vector) immediately, also indicating the
2982 // identify of an object (A) that references the unmarked object (B) --
2983 // presumably, a mutation to A failed to be picked up by preclean/remark?
2984 verify_after_remark_work_2();
2985 } else {
2986 warning("Unrecognized value %d for CMSRemarkVerifyVariant",
2987 CMSRemarkVerifyVariant);
2988 }
2989 gclog_or_tty->print(" done] ");
2990 return true;
2991 }
2992
2993 void CMSCollector::verify_after_remark_work_1() {
2994 ResourceMark rm;
2995 HandleMark hm;
2996 GenCollectedHeap* gch = GenCollectedHeap::heap();
2997
2998 // Mark from roots one level into CMS
2999 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
3000 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3001
3002 gch->gen_process_strong_roots(_cmsGen->level(),
3003 true, // younger gens are roots
3004 true, // activate StrongRootsScope
3005 true, // collecting perm gen
3006 SharedHeap::ScanningOption(roots_scanning_options()),
3007 ¬Older,
3008 true, // walk code active on stacks
3009 NULL);
3010
3011 // Now mark from the roots
3012 assert(_revisitStack.isEmpty(), "Should be empty");
3013 MarkFromRootsClosure markFromRootsClosure(this, _span,
3014 verification_mark_bm(), verification_mark_stack(), &_revisitStack,
3015 false /* don't yield */, true /* verifying */);
3016 assert(_restart_addr == NULL, "Expected pre-condition");
3017 verification_mark_bm()->iterate(&markFromRootsClosure);
3018 while (_restart_addr != NULL) {
3019 // Deal with stack overflow: by restarting at the indicated
3020 // address.
3021 HeapWord* ra = _restart_addr;
3022 markFromRootsClosure.reset(ra);
3023 _restart_addr = NULL;
3024 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
3025 }
3026 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
3027 verify_work_stacks_empty();
3028 // Should reset the revisit stack above, since no class tree
3029 // surgery is forthcoming.
3030 _revisitStack.reset(); // throwing away all contents
3031
3032 // Marking completed -- now verify that each bit marked in
3033 // verification_mark_bm() is also marked in markBitMap(); flag all
3034 // errors by printing corresponding objects.
3035 VerifyMarkedClosure vcl(markBitMap());
3036 verification_mark_bm()->iterate(&vcl);
3037 if (vcl.failed()) {
3038 gclog_or_tty->print("Verification failed");
3039 Universe::heap()->print_on(gclog_or_tty);
3040 fatal("CMS: failed marking verification after remark");
3041 }
3042 }
3043
3044 void CMSCollector::verify_after_remark_work_2() {
3045 ResourceMark rm;
3046 HandleMark hm;
3047 GenCollectedHeap* gch = GenCollectedHeap::heap();
3048
3049 // Mark from roots one level into CMS
3050 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
3051 markBitMap());
3052 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3053 gch->gen_process_strong_roots(_cmsGen->level(),
3054 true, // younger gens are roots
3055 true, // activate StrongRootsScope
3056 true, // collecting perm gen
3057 SharedHeap::ScanningOption(roots_scanning_options()),
3058 ¬Older,
3059 true, // walk code active on stacks
3060 NULL);
3061
3062 // Now mark from the roots
3063 assert(_revisitStack.isEmpty(), "Should be empty");
3064 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
3065 verification_mark_bm(), markBitMap(), verification_mark_stack());
3066 assert(_restart_addr == NULL, "Expected pre-condition");
3067 verification_mark_bm()->iterate(&markFromRootsClosure);
3068 while (_restart_addr != NULL) {
3069 // Deal with stack overflow: by restarting at the indicated
3070 // address.
3071 HeapWord* ra = _restart_addr;
3072 markFromRootsClosure.reset(ra);
3073 _restart_addr = NULL;
3074 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
3075 }
3076 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
3077 verify_work_stacks_empty();
3078 // Should reset the revisit stack above, since no class tree
3079 // surgery is forthcoming.
3080 _revisitStack.reset(); // throwing away all contents
3081
3082 // Marking completed -- now verify that each bit marked in
3083 // verification_mark_bm() is also marked in markBitMap(); flag all
3084 // errors by printing corresponding objects.
3085 VerifyMarkedClosure vcl(markBitMap());
3086 verification_mark_bm()->iterate(&vcl);
3087 assert(!vcl.failed(), "Else verification above should not have succeeded");
3088 }
3089
3090 void ConcurrentMarkSweepGeneration::save_marks() {
3091 // delegate to CMS space
3092 cmsSpace()->save_marks();
3093 for (uint i = 0; i < ParallelGCThreads; i++) {
3094 _par_gc_thread_states[i]->promo.startTrackingPromotions();
3095 }
3096 }
3097
3098 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
3099 return cmsSpace()->no_allocs_since_save_marks();
3100 }
3101
3102 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
3103 \
3104 void ConcurrentMarkSweepGeneration:: \
3105 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
3106 cl->set_generation(this); \
3107 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
3108 cl->reset_generation(); \
3109 save_marks(); \
3110 }
3111
3112 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
3113
3114 void
3115 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk)
3116 {
3117 // Not currently implemented; need to do the following. -- ysr.
3118 // dld -- I think that is used for some sort of allocation profiler. So it
3119 // really means the objects allocated by the mutator since the last
3120 // GC. We could potentially implement this cheaply by recording only
3121 // the direct allocations in a side data structure.
3122 //
3123 // I think we probably ought not to be required to support these
3124 // iterations at any arbitrary point; I think there ought to be some
3125 // call to enable/disable allocation profiling in a generation/space,
3126 // and the iterator ought to return the objects allocated in the
3127 // gen/space since the enable call, or the last iterator call (which
3128 // will probably be at a GC.) That way, for gens like CM&S that would
3129 // require some extra data structure to support this, we only pay the
3130 // cost when it's in use...
3131 cmsSpace()->object_iterate_since_last_GC(blk);
3132 }
3133
3134 void
3135 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
3136 cl->set_generation(this);
3137 younger_refs_in_space_iterate(_cmsSpace, cl);
3138 cl->reset_generation();
3139 }
3140
3141 void
3142 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) {
3143 if (freelistLock()->owned_by_self()) {
3144 Generation::oop_iterate(mr, cl);
3145 } else {
3146 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3147 Generation::oop_iterate(mr, cl);
3148 }
3149 }
3150
3151 void
3152 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) {
3153 if (freelistLock()->owned_by_self()) {
3154 Generation::oop_iterate(cl);
3155 } else {
3156 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3157 Generation::oop_iterate(cl);
3158 }
3159 }
3160
3161 void
3162 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
3163 if (freelistLock()->owned_by_self()) {
3164 Generation::object_iterate(cl);
3165 } else {
3166 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3167 Generation::object_iterate(cl);
3168 }
3169 }
3170
3171 void
3172 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
3173 if (freelistLock()->owned_by_self()) {
3174 Generation::safe_object_iterate(cl);
3175 } else {
3176 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3177 Generation::safe_object_iterate(cl);
3178 }
3179 }
3180
3181 void
3182 ConcurrentMarkSweepGeneration::pre_adjust_pointers() {
3183 }
3184
3185 void
3186 ConcurrentMarkSweepGeneration::post_compact() {
3187 }
3188
3189 void
3190 ConcurrentMarkSweepGeneration::prepare_for_verify() {
3191 // Fix the linear allocation blocks to look like free blocks.
3192
3193 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3194 // are not called when the heap is verified during universe initialization and
3195 // at vm shutdown.
3196 if (freelistLock()->owned_by_self()) {
3197 cmsSpace()->prepare_for_verify();
3198 } else {
3199 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3200 cmsSpace()->prepare_for_verify();
3201 }
3202 }
3203
3204 void
3205 ConcurrentMarkSweepGeneration::verify() {
3206 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3207 // are not called when the heap is verified during universe initialization and
3208 // at vm shutdown.
3209 if (freelistLock()->owned_by_self()) {
3210 cmsSpace()->verify();
3211 } else {
3212 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3213 cmsSpace()->verify();
3214 }
3215 }
3216
3217 void CMSCollector::verify() {
3218 _cmsGen->verify();
3219 _permGen->verify();
3220 }
3221
3222 #ifndef PRODUCT
3223 bool CMSCollector::overflow_list_is_empty() const {
3224 assert(_num_par_pushes >= 0, "Inconsistency");
3225 if (_overflow_list == NULL) {
3226 assert(_num_par_pushes == 0, "Inconsistency");
3227 }
3228 return _overflow_list == NULL;
3229 }
3230
3231 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3232 // merely consolidate assertion checks that appear to occur together frequently.
3233 void CMSCollector::verify_work_stacks_empty() const {
3234 assert(_markStack.isEmpty(), "Marking stack should be empty");
3235 assert(overflow_list_is_empty(), "Overflow list should be empty");
3236 }
3237
3238 void CMSCollector::verify_overflow_empty() const {
3239 assert(overflow_list_is_empty(), "Overflow list should be empty");
3240 assert(no_preserved_marks(), "No preserved marks");
3241 }
3242 #endif // PRODUCT
3243
3244 // Decide if we want to enable class unloading as part of the
3245 // ensuing concurrent GC cycle. We will collect the perm gen and
3246 // unload classes if it's the case that:
3247 // (1) an explicit gc request has been made and the flag
3248 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3249 // (2) (a) class unloading is enabled at the command line, and
3250 // (b) (i) perm gen threshold has been crossed, or
3251 // (ii) old gen is getting really full, or
3252 // (iii) the previous N CMS collections did not collect the
3253 // perm gen
3254 // NOTE: Provided there is no change in the state of the heap between
3255 // calls to this method, it should have idempotent results. Moreover,
3256 // its results should be monotonically increasing (i.e. going from 0 to 1,
3257 // but not 1 to 0) between successive calls between which the heap was
3258 // not collected. For the implementation below, it must thus rely on
3259 // the property that concurrent_cycles_since_last_unload()
3260 // will not decrease unless a collection cycle happened and that
3261 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are
3262 // themselves also monotonic in that sense. See check_monotonicity()
3263 // below.
3264 bool CMSCollector::update_should_unload_classes() {
3265 _should_unload_classes = false;
3266 // Condition 1 above
3267 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3268 _should_unload_classes = true;
3269 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3270 // Disjuncts 2.b.(i,ii,iii) above
3271 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3272 CMSClassUnloadingMaxInterval)
3273 || _permGen->should_concurrent_collect()
3274 || _cmsGen->is_too_full();
3275 }
3276 return _should_unload_classes;
3277 }
3278
3279 bool ConcurrentMarkSweepGeneration::is_too_full() const {
3280 bool res = should_concurrent_collect();
3281 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
3282 return res;
3283 }
3284
3285 void CMSCollector::setup_cms_unloading_and_verification_state() {
3286 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3287 || VerifyBeforeExit;
3288 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
3289
3290 if (should_unload_classes()) { // Should unload classes this cycle
3291 remove_root_scanning_option(rso); // Shrink the root set appropriately
3292 set_verifying(should_verify); // Set verification state for this cycle
3293 return; // Nothing else needs to be done at this time
3294 }
3295
3296 // Not unloading classes this cycle
3297 assert(!should_unload_classes(), "Inconsitency!");
3298 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3299 // We were not verifying, or we _were_ unloading classes in the last cycle,
3300 // AND some verification options are enabled this cycle; in this case,
3301 // we must make sure that the deadness map is allocated if not already so,
3302 // and cleared (if already allocated previously --
3303 // CMSBitMap::sizeInBits() is used to determine if it's allocated).
3304 if (perm_gen_verify_bit_map()->sizeInBits() == 0) {
3305 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) {
3306 warning("Failed to allocate permanent generation verification CMS Bit Map;\n"
3307 "permanent generation verification disabled");
3308 return; // Note that we leave verification disabled, so we'll retry this
3309 // allocation next cycle. We _could_ remember this failure
3310 // and skip further attempts and permanently disable verification
3311 // attempts if that is considered more desirable.
3312 }
3313 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()),
3314 "_perm_gen_ver_bit_map inconsistency?");
3315 } else {
3316 perm_gen_verify_bit_map()->clear_all();
3317 }
3318 // Include symbols, strings and code cache elements to prevent their resurrection.
3319 add_root_scanning_option(rso);
3320 set_verifying(true);
3321 } else if (verifying() && !should_verify) {
3322 // We were verifying, but some verification flags got disabled.
3323 set_verifying(false);
3324 // Exclude symbols, strings and code cache elements from root scanning to
3325 // reduce IM and RM pauses.
3326 remove_root_scanning_option(rso);
3327 }
3328 }
3329
3330
3331 #ifndef PRODUCT
3332 HeapWord* CMSCollector::block_start(const void* p) const {
3333 const HeapWord* addr = (HeapWord*)p;
3334 if (_span.contains(p)) {
3335 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3336 return _cmsGen->cmsSpace()->block_start(p);
3337 } else {
3338 assert(_permGen->cmsSpace()->is_in_reserved(addr),
3339 "Inconsistent _span?");
3340 return _permGen->cmsSpace()->block_start(p);
3341 }
3342 }
3343 return NULL;
3344 }
3345 #endif
3346
3347 HeapWord*
3348 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3349 bool tlab,
3350 bool parallel) {
3351 CMSSynchronousYieldRequest yr;
3352 assert(!tlab, "Can't deal with TLAB allocation");
3353 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3354 expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3355 CMSExpansionCause::_satisfy_allocation);
3356 if (GCExpandToAllocateDelayMillis > 0) {
3357 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3358 }
3359 return have_lock_and_allocate(word_size, tlab);
3360 }
3361
3362 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3363 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3364 // to CardGeneration and share it...
3365 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) {
3366 return CardGeneration::expand(bytes, expand_bytes);
3367 }
3368
3369 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3370 CMSExpansionCause::Cause cause)
3371 {
3372
3373 bool success = expand(bytes, expand_bytes);
3374
3375 // remember why we expanded; this information is used
3376 // by shouldConcurrentCollect() when making decisions on whether to start
3377 // a new CMS cycle.
3378 if (success) {
3379 set_expansion_cause(cause);
3380 if (PrintGCDetails && Verbose) {
3381 gclog_or_tty->print_cr("Expanded CMS gen for %s",
3382 CMSExpansionCause::to_string(cause));
3383 }
3384 }
3385 }
3386
3387 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3388 HeapWord* res = NULL;
3389 MutexLocker x(ParGCRareEvent_lock);
3390 while (true) {
3391 // Expansion by some other thread might make alloc OK now:
3392 res = ps->lab.alloc(word_sz);
3393 if (res != NULL) return res;
3394 // If there's not enough expansion space available, give up.
3395 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3396 return NULL;
3397 }
3398 // Otherwise, we try expansion.
3399 expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3400 CMSExpansionCause::_allocate_par_lab);
3401 // Now go around the loop and try alloc again;
3402 // A competing par_promote might beat us to the expansion space,
3403 // so we may go around the loop again if promotion fails agaion.
3404 if (GCExpandToAllocateDelayMillis > 0) {
3405 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3406 }
3407 }
3408 }
3409
3410
3411 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3412 PromotionInfo* promo) {
3413 MutexLocker x(ParGCRareEvent_lock);
3414 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3415 while (true) {
3416 // Expansion by some other thread might make alloc OK now:
3417 if (promo->ensure_spooling_space()) {
3418 assert(promo->has_spooling_space(),
3419 "Post-condition of successful ensure_spooling_space()");
3420 return true;
3421 }
3422 // If there's not enough expansion space available, give up.
3423 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3424 return false;
3425 }
3426 // Otherwise, we try expansion.
3427 expand(refill_size_bytes, MinHeapDeltaBytes,
3428 CMSExpansionCause::_allocate_par_spooling_space);
3429 // Now go around the loop and try alloc again;
3430 // A competing allocation might beat us to the expansion space,
3431 // so we may go around the loop again if allocation fails again.
3432 if (GCExpandToAllocateDelayMillis > 0) {
3433 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3434 }
3435 }
3436 }
3437
3438
3439
3440 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3441 assert_locked_or_safepoint(Heap_lock);
3442 size_t size = ReservedSpace::page_align_size_down(bytes);
3443 if (size > 0) {
3444 shrink_by(size);
3445 }
3446 }
3447
3448 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3449 assert_locked_or_safepoint(Heap_lock);
3450 bool result = _virtual_space.expand_by(bytes);
3451 if (result) {
3452 HeapWord* old_end = _cmsSpace->end();
3453 size_t new_word_size =
3454 heap_word_size(_virtual_space.committed_size());
3455 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3456 _bts->resize(new_word_size); // resize the block offset shared array
3457 Universe::heap()->barrier_set()->resize_covered_region(mr);
3458 // Hmmmm... why doesn't CFLS::set_end verify locking?
3459 // This is quite ugly; FIX ME XXX
3460 _cmsSpace->assert_locked(freelistLock());
3461 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3462
3463 // update the space and generation capacity counters
3464 if (UsePerfData) {
3465 _space_counters->update_capacity();
3466 _gen_counters->update_all();
3467 }
3468
3469 if (Verbose && PrintGC) {
3470 size_t new_mem_size = _virtual_space.committed_size();
3471 size_t old_mem_size = new_mem_size - bytes;
3472 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK",
3473 name(), old_mem_size/K, bytes/K, new_mem_size/K);
3474 }
3475 }
3476 return result;
3477 }
3478
3479 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3480 assert_locked_or_safepoint(Heap_lock);
3481 bool success = true;
3482 const size_t remaining_bytes = _virtual_space.uncommitted_size();
3483 if (remaining_bytes > 0) {
3484 success = grow_by(remaining_bytes);
3485 DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3486 }
3487 return success;
3488 }
3489
3490 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3491 assert_locked_or_safepoint(Heap_lock);
3492 assert_lock_strong(freelistLock());
3493 // XXX Fix when compaction is implemented.
3494 warning("Shrinking of CMS not yet implemented");
3495 return;
3496 }
3497
3498
3499 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3500 // phases.
3501 class CMSPhaseAccounting: public StackObj {
3502 public:
3503 CMSPhaseAccounting(CMSCollector *collector,
3504 const char *phase,
3505 bool print_cr = true);
3506 ~CMSPhaseAccounting();
3507
3508 private:
3509 CMSCollector *_collector;
3510 const char *_phase;
3511 elapsedTimer _wallclock;
3512 bool _print_cr;
3513
3514 public:
3515 // Not MT-safe; so do not pass around these StackObj's
3516 // where they may be accessed by other threads.
3517 jlong wallclock_millis() {
3518 assert(_wallclock.is_active(), "Wall clock should not stop");
3519 _wallclock.stop(); // to record time
3520 jlong ret = _wallclock.milliseconds();
3521 _wallclock.start(); // restart
3522 return ret;
3523 }
3524 };
3525
3526 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3527 const char *phase,
3528 bool print_cr) :
3529 _collector(collector), _phase(phase), _print_cr(print_cr) {
3530
3531 if (PrintCMSStatistics != 0) {
3532 _collector->resetYields();
3533 }
3534 if (PrintGCDetails && PrintGCTimeStamps) {
3535 gclog_or_tty->date_stamp(PrintGCDateStamps);
3536 gclog_or_tty->stamp();
3537 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]",
3538 _collector->cmsGen()->short_name(), _phase);
3539 }
3540 _collector->resetTimer();
3541 _wallclock.start();
3542 _collector->startTimer();
3543 }
3544
3545 CMSPhaseAccounting::~CMSPhaseAccounting() {
3546 assert(_wallclock.is_active(), "Wall clock should not have stopped");
3547 _collector->stopTimer();
3548 _wallclock.stop();
3549 if (PrintGCDetails) {
3550 gclog_or_tty->date_stamp(PrintGCDateStamps);
3551 gclog_or_tty->stamp(PrintGCTimeStamps);
3552 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3553 _collector->cmsGen()->short_name(),
3554 _phase, _collector->timerValue(), _wallclock.seconds());
3555 if (_print_cr) {
3556 gclog_or_tty->print_cr("");
3557 }
3558 if (PrintCMSStatistics != 0) {
3559 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3560 _collector->yields());
3561 }
3562 }
3563 }
3564
3565 // CMS work
3566
3567 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
3568 class CMSParMarkTask : public AbstractGangTask {
3569 protected:
3570 CMSCollector* _collector;
3571 int _n_workers;
3572 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) :
3573 AbstractGangTask(name),
3574 _collector(collector),
3575 _n_workers(n_workers) {}
3576 // Work method in support of parallel rescan ... of young gen spaces
3577 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
3578 ContiguousSpace* space,
3579 HeapWord** chunk_array, size_t chunk_top);
3580 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
3581 };
3582
3583 // Parallel initial mark task
3584 class CMSParInitialMarkTask: public CMSParMarkTask {
3585 public:
3586 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) :
3587 CMSParMarkTask("Scan roots and young gen for initial mark in parallel",
3588 collector, n_workers) {}
3589 void work(uint worker_id);
3590 };
3591
3592 // Checkpoint the roots into this generation from outside
3593 // this generation. [Note this initial checkpoint need only
3594 // be approximate -- we'll do a catch up phase subsequently.]
3595 void CMSCollector::checkpointRootsInitial(bool asynch) {
3596 assert(_collectorState == InitialMarking, "Wrong collector state");
3597 check_correct_thread_executing();
3598 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
3599
3600 save_heap_summary();
3601 report_heap_summary(GCWhen::BeforeGC);
3602
3603 ReferenceProcessor* rp = ref_processor();
3604 SpecializationStats::clear();
3605 assert(_restart_addr == NULL, "Control point invariant");
3606 if (asynch) {
3607 // acquire locks for subsequent manipulations
3608 MutexLockerEx x(bitMapLock(),
3609 Mutex::_no_safepoint_check_flag);
3610 checkpointRootsInitialWork(asynch);
3611 // enable ("weak") refs discovery
3612 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/);
3613 _collectorState = Marking;
3614 } else {
3615 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3616 // which recognizes if we are a CMS generation, and doesn't try to turn on
3617 // discovery; verify that they aren't meddling.
3618 assert(!rp->discovery_is_atomic(),
3619 "incorrect setting of discovery predicate");
3620 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3621 "ref discovery for this generation kind");
3622 // already have locks
3623 checkpointRootsInitialWork(asynch);
3624 // now enable ("weak") refs discovery
3625 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/);
3626 _collectorState = Marking;
3627 }
3628 SpecializationStats::print();
3629 }
3630
3631 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3632 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3633 assert(_collectorState == InitialMarking, "just checking");
3634
3635 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3636 // precede our marking with a collection of all
3637 // younger generations to keep floating garbage to a minimum.
3638 // XXX: we won't do this for now -- it's an optimization to be done later.
3639
3640 // already have locks
3641 assert_lock_strong(bitMapLock());
3642 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3643
3644 // Setup the verification and class unloading state for this
3645 // CMS collection cycle.
3646 setup_cms_unloading_and_verification_state();
3647
3648 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
3649 PrintGCDetails && Verbose, true, _gc_timer_cm);)
3650 if (UseAdaptiveSizePolicy) {
3651 size_policy()->checkpoint_roots_initial_begin();
3652 }
3653
3654 // Reset all the PLAB chunk arrays if necessary.
3655 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3656 reset_survivor_plab_arrays();
3657 }
3658
3659 ResourceMark rm;
3660 HandleMark hm;
3661
3662 FalseClosure falseClosure;
3663 // In the case of a synchronous collection, we will elide the
3664 // remark step, so it's important to catch all the nmethod oops
3665 // in this step.
3666 // The final 'true' flag to gen_process_strong_roots will ensure this.
3667 // If 'async' is true, we can relax the nmethod tracing.
3668 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
3669 GenCollectedHeap* gch = GenCollectedHeap::heap();
3670
3671 verify_work_stacks_empty();
3672 verify_overflow_empty();
3673
3674 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3675 // Update the saved marks which may affect the root scans.
3676 gch->save_marks();
3677
3678 // weak reference processing has not started yet.
3679 ref_processor()->set_enqueuing_is_done(false);
3680
3681 if (CMSPrintEdenSurvivorChunks) {
3682 print_eden_and_survivor_chunk_arrays();
3683 }
3684
3685 {
3686 // This is not needed. DEBUG_ONLY(RememberKlassesChecker imx(true);)
3687 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3688 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
3689 // The parallel version.
3690 FlexibleWorkGang* workers = gch->workers();
3691 assert(workers != NULL, "Need parallel worker threads.");
3692 int n_workers = workers->active_workers();
3693 CMSParInitialMarkTask tsk(this, n_workers);
3694 gch->set_par_threads(n_workers);
3695 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
3696 if (n_workers > 1) {
3697 GenCollectedHeap::StrongRootsScope srs(gch);
3698 workers->run_task(&tsk);
3699 } else {
3700 GenCollectedHeap::StrongRootsScope srs(gch);
3701 tsk.work(0);
3702 }
3703 gch->set_par_threads(0);
3704 } else {
3705 // The serial version.
3706 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3707 gch->gen_process_strong_roots(_cmsGen->level(),
3708 true, // younger gens are roots
3709 true, // activate StrongRootsScope
3710 true, // collecting perm gen
3711 SharedHeap::ScanningOption(roots_scanning_options()),
3712 ¬Older,
3713 true, // walk all of code cache if (so & SO_CodeCache)
3714 NULL);
3715 }
3716 }
3717 // Clear mod-union table; it will be dirtied in the prologue of
3718 // CMS generation per each younger generation collection.
3719
3720 assert(_modUnionTable.isAllClear(),
3721 "Was cleared in most recent final checkpoint phase"
3722 " or no bits are set in the gc_prologue before the start of the next "
3723 "subsequent marking phase.");
3724
3725 // Save the end of the used_region of the constituent generations
3726 // to be used to limit the extent of sweep in each generation.
3727 save_sweep_limits();
3728 if (UseAdaptiveSizePolicy) {
3729 size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3730 }
3731 verify_overflow_empty();
3732 }
3733
3734 bool CMSCollector::markFromRoots(bool asynch) {
3735 // we might be tempted to assert that:
3736 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3737 // "inconsistent argument?");
3738 // However that wouldn't be right, because it's possible that
3739 // a safepoint is indeed in progress as a younger generation
3740 // stop-the-world GC happens even as we mark in this generation.
3741 assert(_collectorState == Marking, "inconsistent state?");
3742 check_correct_thread_executing();
3743 verify_overflow_empty();
3744
3745 bool res;
3746 if (asynch) {
3747
3748 // Start the timers for adaptive size policy for the concurrent phases
3749 // Do it here so that the foreground MS can use the concurrent
3750 // timer since a foreground MS might has the sweep done concurrently
3751 // or STW.
3752 if (UseAdaptiveSizePolicy) {
3753 size_policy()->concurrent_marking_begin();
3754 }
3755
3756 // Weak ref discovery note: We may be discovering weak
3757 // refs in this generation concurrent (but interleaved) with
3758 // weak ref discovery by a younger generation collector.
3759
3760 CMSTokenSyncWithLocks ts(true, bitMapLock());
3761 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3762 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3763 res = markFromRootsWork(asynch);
3764 if (res) {
3765 _collectorState = Precleaning;
3766 } else { // We failed and a foreground collection wants to take over
3767 assert(_foregroundGCIsActive, "internal state inconsistency");
3768 assert(_restart_addr == NULL, "foreground will restart from scratch");
3769 if (PrintGCDetails) {
3770 gclog_or_tty->print_cr("bailing out to foreground collection");
3771 }
3772 }
3773 if (UseAdaptiveSizePolicy) {
3774 size_policy()->concurrent_marking_end();
3775 }
3776 } else {
3777 assert(SafepointSynchronize::is_at_safepoint(),
3778 "inconsistent with asynch == false");
3779 if (UseAdaptiveSizePolicy) {
3780 size_policy()->ms_collection_marking_begin();
3781 }
3782 // already have locks
3783 res = markFromRootsWork(asynch);
3784 _collectorState = FinalMarking;
3785 if (UseAdaptiveSizePolicy) {
3786 GenCollectedHeap* gch = GenCollectedHeap::heap();
3787 size_policy()->ms_collection_marking_end(gch->gc_cause());
3788 }
3789 }
3790 verify_overflow_empty();
3791 return res;
3792 }
3793
3794 bool CMSCollector::markFromRootsWork(bool asynch) {
3795 // iterate over marked bits in bit map, doing a full scan and mark
3796 // from these roots using the following algorithm:
3797 // . if oop is to the right of the current scan pointer,
3798 // mark corresponding bit (we'll process it later)
3799 // . else (oop is to left of current scan pointer)
3800 // push oop on marking stack
3801 // . drain the marking stack
3802
3803 // Note that when we do a marking step we need to hold the
3804 // bit map lock -- recall that direct allocation (by mutators)
3805 // and promotion (by younger generation collectors) is also
3806 // marking the bit map. [the so-called allocate live policy.]
3807 // Because the implementation of bit map marking is not
3808 // robust wrt simultaneous marking of bits in the same word,
3809 // we need to make sure that there is no such interference
3810 // between concurrent such updates.
3811
3812 // already have locks
3813 assert_lock_strong(bitMapLock());
3814
3815 // Clear the revisit stack, just in case there are any
3816 // obsolete contents from a short-circuited previous CMS cycle.
3817 _revisitStack.reset();
3818 verify_work_stacks_empty();
3819 verify_overflow_empty();
3820 assert(_revisitStack.isEmpty(), "tabula rasa");
3821 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());)
3822 bool result = false;
3823 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3824 result = do_marking_mt(asynch);
3825 } else {
3826 result = do_marking_st(asynch);
3827 }
3828 return result;
3829 }
3830
3831 // Forward decl
3832 class CMSConcMarkingTask;
3833
3834 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3835 CMSCollector* _collector;
3836 CMSConcMarkingTask* _task;
3837 public:
3838 virtual void yield();
3839
3840 // "n_threads" is the number of threads to be terminated.
3841 // "queue_set" is a set of work queues of other threads.
3842 // "collector" is the CMS collector associated with this task terminator.
3843 // "yield" indicates whether we need the gang as a whole to yield.
3844 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3845 ParallelTaskTerminator(n_threads, queue_set),
3846 _collector(collector) { }
3847
3848 void set_task(CMSConcMarkingTask* task) {
3849 _task = task;
3850 }
3851 };
3852
3853 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3854 CMSConcMarkingTask* _task;
3855 public:
3856 bool should_exit_termination();
3857 void set_task(CMSConcMarkingTask* task) {
3858 _task = task;
3859 }
3860 };
3861
3862 // MT Concurrent Marking Task
3863 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3864 CMSCollector* _collector;
3865 int _n_workers; // requested/desired # workers
3866 bool _asynch;
3867 bool _result;
3868 CompactibleFreeListSpace* _cms_space;
3869 CompactibleFreeListSpace* _perm_space;
3870 char _pad_front[64]; // padding to ...
3871 HeapWord* _global_finger; // ... avoid sharing cache line
3872 char _pad_back[64];
3873 HeapWord* _restart_addr;
3874
3875 // Exposed here for yielding support
3876 Mutex* const _bit_map_lock;
3877
3878 // The per thread work queues, available here for stealing
3879 OopTaskQueueSet* _task_queues;
3880
3881 // Termination (and yielding) support
3882 CMSConcMarkingTerminator _term;
3883 CMSConcMarkingTerminatorTerminator _term_term;
3884
3885 public:
3886 CMSConcMarkingTask(CMSCollector* collector,
3887 CompactibleFreeListSpace* cms_space,
3888 CompactibleFreeListSpace* perm_space,
3889 bool asynch,
3890 YieldingFlexibleWorkGang* workers,
3891 OopTaskQueueSet* task_queues):
3892 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3893 _collector(collector),
3894 _cms_space(cms_space),
3895 _perm_space(perm_space),
3896 _asynch(asynch), _n_workers(0), _result(true),
3897 _task_queues(task_queues),
3898 _term(_n_workers, task_queues, _collector),
3899 _bit_map_lock(collector->bitMapLock())
3900 {
3901 _requested_size = _n_workers;
3902 _term.set_task(this);
3903 _term_term.set_task(this);
3904 assert(_cms_space->bottom() < _perm_space->bottom(),
3905 "Finger incorrectly initialized below");
3906 _restart_addr = _global_finger = _cms_space->bottom();
3907 }
3908
3909
3910 OopTaskQueueSet* task_queues() { return _task_queues; }
3911
3912 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3913
3914 HeapWord** global_finger_addr() { return &_global_finger; }
3915
3916 CMSConcMarkingTerminator* terminator() { return &_term; }
3917
3918 virtual void set_for_termination(int active_workers) {
3919 terminator()->reset_for_reuse(active_workers);
3920 }
3921
3922 void work(uint worker_id);
3923 bool should_yield() {
3924 return ConcurrentMarkSweepThread::should_yield()
3925 && !_collector->foregroundGCIsActive()
3926 && _asynch;
3927 }
3928
3929 virtual void coordinator_yield(); // stuff done by coordinator
3930 bool result() { return _result; }
3931
3932 void reset(HeapWord* ra) {
3933 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3934 assert(_global_finger >= _perm_space->end(), "Postcondition of ::work(i)");
3935 assert(ra < _perm_space->end(), "ra too large");
3936 _restart_addr = _global_finger = ra;
3937 _term.reset_for_reuse();
3938 }
3939
3940 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3941 OopTaskQueue* work_q);
3942
3943 private:
3944 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3945 void do_work_steal(int i);
3946 void bump_global_finger(HeapWord* f);
3947 };
3948
3949 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3950 assert(_task != NULL, "Error");
3951 return _task->yielding();
3952 // Note that we do not need the disjunct || _task->should_yield() above
3953 // because we want terminating threads to yield only if the task
3954 // is already in the midst of yielding, which happens only after at least one
3955 // thread has yielded.
3956 }
3957
3958 void CMSConcMarkingTerminator::yield() {
3959 if (_task->should_yield()) {
3960 _task->yield();
3961 } else {
3962 ParallelTaskTerminator::yield();
3963 }
3964 }
3965
3966 ////////////////////////////////////////////////////////////////
3967 // Concurrent Marking Algorithm Sketch
3968 ////////////////////////////////////////////////////////////////
3969 // Until all tasks exhausted (both spaces):
3970 // -- claim next available chunk
3971 // -- bump global finger via CAS
3972 // -- find first object that starts in this chunk
3973 // and start scanning bitmap from that position
3974 // -- scan marked objects for oops
3975 // -- CAS-mark target, and if successful:
3976 // . if target oop is above global finger (volatile read)
3977 // nothing to do
3978 // . if target oop is in chunk and above local finger
3979 // then nothing to do
3980 // . else push on work-queue
3981 // -- Deal with possible overflow issues:
3982 // . local work-queue overflow causes stuff to be pushed on
3983 // global (common) overflow queue
3984 // . always first empty local work queue
3985 // . then get a batch of oops from global work queue if any
3986 // . then do work stealing
3987 // -- When all tasks claimed (both spaces)
3988 // and local work queue empty,
3989 // then in a loop do:
3990 // . check global overflow stack; steal a batch of oops and trace
3991 // . try to steal from other threads oif GOS is empty
3992 // . if neither is available, offer termination
3993 // -- Terminate and return result
3994 //
3995 void CMSConcMarkingTask::work(uint worker_id) {
3996 elapsedTimer _timer;
3997 ResourceMark rm;
3998 HandleMark hm;
3999
4000 DEBUG_ONLY(_collector->verify_overflow_empty();)
4001
4002 // Before we begin work, our work queue should be empty
4003 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
4004 // Scan the bitmap covering _cms_space, tracing through grey objects.
4005 _timer.start();
4006 do_scan_and_mark(worker_id, _cms_space);
4007 _timer.stop();
4008 if (PrintCMSStatistics != 0) {
4009 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
4010 worker_id, _timer.seconds());
4011 // XXX: need xxx/xxx type of notation, two timers
4012 }
4013
4014 // ... do the same for the _perm_space
4015 _timer.reset();
4016 _timer.start();
4017 do_scan_and_mark(worker_id, _perm_space);
4018 _timer.stop();
4019 if (PrintCMSStatistics != 0) {
4020 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec",
4021 worker_id, _timer.seconds());
4022 // XXX: need xxx/xxx type of notation, two timers
4023 }
4024
4025 // ... do work stealing
4026 _timer.reset();
4027 _timer.start();
4028 do_work_steal(worker_id);
4029 _timer.stop();
4030 if (PrintCMSStatistics != 0) {
4031 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
4032 worker_id, _timer.seconds());
4033 // XXX: need xxx/xxx type of notation, two timers
4034 }
4035 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
4036 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
4037 // Note that under the current task protocol, the
4038 // following assertion is true even of the spaces
4039 // expanded since the completion of the concurrent
4040 // marking. XXX This will likely change under a strict
4041 // ABORT semantics.
4042 assert(_global_finger > _cms_space->end() &&
4043 _global_finger >= _perm_space->end(),
4044 "All tasks have been completed");
4045 DEBUG_ONLY(_collector->verify_overflow_empty();)
4046 }
4047
4048 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
4049 HeapWord* read = _global_finger;
4050 HeapWord* cur = read;
4051 while (f > read) {
4052 cur = read;
4053 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
4054 if (cur == read) {
4055 // our cas succeeded
4056 assert(_global_finger >= f, "protocol consistency");
4057 break;
4058 }
4059 }
4060 }
4061
4062 // This is really inefficient, and should be redone by
4063 // using (not yet available) block-read and -write interfaces to the
4064 // stack and the work_queue. XXX FIX ME !!!
4065 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
4066 OopTaskQueue* work_q) {
4067 // Fast lock-free check
4068 if (ovflw_stk->length() == 0) {
4069 return false;
4070 }
4071 assert(work_q->size() == 0, "Shouldn't steal");
4072 MutexLockerEx ml(ovflw_stk->par_lock(),
4073 Mutex::_no_safepoint_check_flag);
4074 // Grab up to 1/4 the size of the work queue
4075 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4076 (size_t)ParGCDesiredObjsFromOverflowList);
4077 num = MIN2(num, ovflw_stk->length());
4078 for (int i = (int) num; i > 0; i--) {
4079 oop cur = ovflw_stk->pop();
4080 assert(cur != NULL, "Counted wrong?");
4081 work_q->push(cur);
4082 }
4083 return num > 0;
4084 }
4085
4086 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
4087 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4088 int n_tasks = pst->n_tasks();
4089 // We allow that there may be no tasks to do here because
4090 // we are restarting after a stack overflow.
4091 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
4092 uint nth_task = 0;
4093
4094 HeapWord* aligned_start = sp->bottom();
4095 if (sp->used_region().contains(_restart_addr)) {
4096 // Align down to a card boundary for the start of 0th task
4097 // for this space.
4098 aligned_start =
4099 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
4100 CardTableModRefBS::card_size);
4101 }
4102
4103 size_t chunk_size = sp->marking_task_size();
4104 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4105 // Having claimed the nth task in this space,
4106 // compute the chunk that it corresponds to:
4107 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
4108 aligned_start + (nth_task+1)*chunk_size);
4109 // Try and bump the global finger via a CAS;
4110 // note that we need to do the global finger bump
4111 // _before_ taking the intersection below, because
4112 // the task corresponding to that region will be
4113 // deemed done even if the used_region() expands
4114 // because of allocation -- as it almost certainly will
4115 // during start-up while the threads yield in the
4116 // closure below.
4117 HeapWord* finger = span.end();
4118 bump_global_finger(finger); // atomically
4119 // There are null tasks here corresponding to chunks
4120 // beyond the "top" address of the space.
4121 span = span.intersection(sp->used_region());
4122 if (!span.is_empty()) { // Non-null task
4123 HeapWord* prev_obj;
4124 assert(!span.contains(_restart_addr) || nth_task == 0,
4125 "Inconsistency");
4126 if (nth_task == 0) {
4127 // For the 0th task, we'll not need to compute a block_start.
4128 if (span.contains(_restart_addr)) {
4129 // In the case of a restart because of stack overflow,
4130 // we might additionally skip a chunk prefix.
4131 prev_obj = _restart_addr;
4132 } else {
4133 prev_obj = span.start();
4134 }
4135 } else {
4136 // We want to skip the first object because
4137 // the protocol is to scan any object in its entirety
4138 // that _starts_ in this span; a fortiori, any
4139 // object starting in an earlier span is scanned
4140 // as part of an earlier claimed task.
4141 // Below we use the "careful" version of block_start
4142 // so we do not try to navigate uninitialized objects.
4143 prev_obj = sp->block_start_careful(span.start());
4144 // Below we use a variant of block_size that uses the
4145 // Printezis bits to avoid waiting for allocated
4146 // objects to become initialized/parsable.
4147 while (prev_obj < span.start()) {
4148 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
4149 if (sz > 0) {
4150 prev_obj += sz;
4151 } else {
4152 // In this case we may end up doing a bit of redundant
4153 // scanning, but that appears unavoidable, short of
4154 // locking the free list locks; see bug 6324141.
4155 break;
4156 }
4157 }
4158 }
4159 if (prev_obj < span.end()) {
4160 MemRegion my_span = MemRegion(prev_obj, span.end());
4161 // Do the marking work within a non-empty span --
4162 // the last argument to the constructor indicates whether the
4163 // iteration should be incremental with periodic yields.
4164 Par_MarkFromRootsClosure cl(this, _collector, my_span,
4165 &_collector->_markBitMap,
4166 work_queue(i),
4167 &_collector->_markStack,
4168 &_collector->_revisitStack,
4169 _asynch);
4170 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
4171 } // else nothing to do for this task
4172 } // else nothing to do for this task
4173 }
4174 // We'd be tempted to assert here that since there are no
4175 // more tasks left to claim in this space, the global_finger
4176 // must exceed space->top() and a fortiori space->end(). However,
4177 // that would not quite be correct because the bumping of
4178 // global_finger occurs strictly after the claiming of a task,
4179 // so by the time we reach here the global finger may not yet
4180 // have been bumped up by the thread that claimed the last
4181 // task.
4182 pst->all_tasks_completed();
4183 }
4184
4185 class Par_ConcMarkingClosure: public Par_KlassRememberingOopClosure {
4186 private:
4187 CMSConcMarkingTask* _task;
4188 MemRegion _span;
4189 CMSBitMap* _bit_map;
4190 CMSMarkStack* _overflow_stack;
4191 OopTaskQueue* _work_queue;
4192 protected:
4193 DO_OOP_WORK_DEFN
4194 public:
4195 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
4196 CMSBitMap* bit_map, CMSMarkStack* overflow_stack,
4197 CMSMarkStack* revisit_stack):
4198 Par_KlassRememberingOopClosure(collector, collector->ref_processor(), revisit_stack),
4199 _task(task),
4200 _span(collector->_span),
4201 _work_queue(work_queue),
4202 _bit_map(bit_map),
4203 _overflow_stack(overflow_stack)
4204 { }
4205 virtual void do_oop(oop* p);
4206 virtual void do_oop(narrowOop* p);
4207 void trim_queue(size_t max);
4208 void handle_stack_overflow(HeapWord* lost);
4209 void do_yield_check() {
4210 if (_task->should_yield()) {
4211 _task->yield();
4212 }
4213 }
4214 };
4215
4216 // Grey object scanning during work stealing phase --
4217 // the salient assumption here is that any references
4218 // that are in these stolen objects being scanned must
4219 // already have been initialized (else they would not have
4220 // been published), so we do not need to check for
4221 // uninitialized objects before pushing here.
4222 void Par_ConcMarkingClosure::do_oop(oop obj) {
4223 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
4224 HeapWord* addr = (HeapWord*)obj;
4225 // Check if oop points into the CMS generation
4226 // and is not marked
4227 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
4228 // a white object ...
4229 // If we manage to "claim" the object, by being the
4230 // first thread to mark it, then we push it on our
4231 // marking stack
4232 if (_bit_map->par_mark(addr)) { // ... now grey
4233 // push on work queue (grey set)
4234 bool simulate_overflow = false;
4235 NOT_PRODUCT(
4236 if (CMSMarkStackOverflowALot &&
4237 _collector->simulate_overflow()) {
4238 // simulate a stack overflow
4239 simulate_overflow = true;
4240 }
4241 )
4242 if (simulate_overflow ||
4243 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
4244 // stack overflow
4245 if (PrintCMSStatistics != 0) {
4246 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
4247 SIZE_FORMAT, _overflow_stack->capacity());
4248 }
4249 // We cannot assert that the overflow stack is full because
4250 // it may have been emptied since.
4251 assert(simulate_overflow ||
4252 _work_queue->size() == _work_queue->max_elems(),
4253 "Else push should have succeeded");
4254 handle_stack_overflow(addr);
4255 }
4256 } // Else, some other thread got there first
4257 do_yield_check();
4258 }
4259 }
4260
4261 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4262 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4263
4264 void Par_ConcMarkingClosure::trim_queue(size_t max) {
4265 while (_work_queue->size() > max) {
4266 oop new_oop;
4267 if (_work_queue->pop_local(new_oop)) {
4268 assert(new_oop->is_oop(), "Should be an oop");
4269 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
4270 assert(_span.contains((HeapWord*)new_oop), "Not in span");
4271 assert(new_oop->is_parsable(), "Should be parsable");
4272 new_oop->oop_iterate(this); // do_oop() above
4273 do_yield_check();
4274 }
4275 }
4276 }
4277
4278 // Upon stack overflow, we discard (part of) the stack,
4279 // remembering the least address amongst those discarded
4280 // in CMSCollector's _restart_address.
4281 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
4282 // We need to do this under a mutex to prevent other
4283 // workers from interfering with the work done below.
4284 MutexLockerEx ml(_overflow_stack->par_lock(),
4285 Mutex::_no_safepoint_check_flag);
4286 // Remember the least grey address discarded
4287 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
4288 _collector->lower_restart_addr(ra);
4289 _overflow_stack->reset(); // discard stack contents
4290 _overflow_stack->expand(); // expand the stack if possible
4291 }
4292
4293
4294 void CMSConcMarkingTask::do_work_steal(int i) {
4295 OopTaskQueue* work_q = work_queue(i);
4296 oop obj_to_scan;
4297 CMSBitMap* bm = &(_collector->_markBitMap);
4298 CMSMarkStack* ovflw = &(_collector->_markStack);
4299 CMSMarkStack* revisit = &(_collector->_revisitStack);
4300 int* seed = _collector->hash_seed(i);
4301 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw, revisit);
4302 while (true) {
4303 cl.trim_queue(0);
4304 assert(work_q->size() == 0, "Should have been emptied above");
4305 if (get_work_from_overflow_stack(ovflw, work_q)) {
4306 // Can't assert below because the work obtained from the
4307 // overflow stack may already have been stolen from us.
4308 // assert(work_q->size() > 0, "Work from overflow stack");
4309 continue;
4310 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4311 assert(obj_to_scan->is_oop(), "Should be an oop");
4312 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
4313 obj_to_scan->oop_iterate(&cl);
4314 } else if (terminator()->offer_termination(&_term_term)) {
4315 assert(work_q->size() == 0, "Impossible!");
4316 break;
4317 } else if (yielding() || should_yield()) {
4318 yield();
4319 }
4320 }
4321 }
4322
4323 // This is run by the CMS (coordinator) thread.
4324 void CMSConcMarkingTask::coordinator_yield() {
4325 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4326 "CMS thread should hold CMS token");
4327 DEBUG_ONLY(RememberKlassesChecker mux(false);)
4328 // First give up the locks, then yield, then re-lock
4329 // We should probably use a constructor/destructor idiom to
4330 // do this unlock/lock or modify the MutexUnlocker class to
4331 // serve our purpose. XXX
4332 assert_lock_strong(_bit_map_lock);
4333 _bit_map_lock->unlock();
4334 ConcurrentMarkSweepThread::desynchronize(true);
4335 ConcurrentMarkSweepThread::acknowledge_yield_request();
4336 _collector->stopTimer();
4337 if (PrintCMSStatistics != 0) {
4338 _collector->incrementYields();
4339 }
4340 _collector->icms_wait();
4341
4342 // It is possible for whichever thread initiated the yield request
4343 // not to get a chance to wake up and take the bitmap lock between
4344 // this thread releasing it and reacquiring it. So, while the
4345 // should_yield() flag is on, let's sleep for a bit to give the
4346 // other thread a chance to wake up. The limit imposed on the number
4347 // of iterations is defensive, to avoid any unforseen circumstances
4348 // putting us into an infinite loop. Since it's always been this
4349 // (coordinator_yield()) method that was observed to cause the
4350 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4351 // which is by default non-zero. For the other seven methods that
4352 // also perform the yield operation, as are using a different
4353 // parameter (CMSYieldSleepCount) which is by default zero. This way we
4354 // can enable the sleeping for those methods too, if necessary.
4355 // See 6442774.
4356 //
4357 // We really need to reconsider the synchronization between the GC
4358 // thread and the yield-requesting threads in the future and we
4359 // should really use wait/notify, which is the recommended
4360 // way of doing this type of interaction. Additionally, we should
4361 // consolidate the eight methods that do the yield operation and they
4362 // are almost identical into one for better maintenability and
4363 // readability. See 6445193.
4364 //
4365 // Tony 2006.06.29
4366 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4367 ConcurrentMarkSweepThread::should_yield() &&
4368 !CMSCollector::foregroundGCIsActive(); ++i) {
4369 os::sleep(Thread::current(), 1, false);
4370 ConcurrentMarkSweepThread::acknowledge_yield_request();
4371 }
4372
4373 ConcurrentMarkSweepThread::synchronize(true);
4374 _bit_map_lock->lock_without_safepoint_check();
4375 _collector->startTimer();
4376 }
4377
4378 bool CMSCollector::do_marking_mt(bool asynch) {
4379 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
4380 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers(
4381 conc_workers()->total_workers(),
4382 conc_workers()->active_workers(),
4383 Threads::number_of_non_daemon_threads());
4384 conc_workers()->set_active_workers(num_workers);
4385
4386 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4387 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
4388
4389 CMSConcMarkingTask tsk(this,
4390 cms_space,
4391 perm_space,
4392 asynch,
4393 conc_workers(),
4394 task_queues());
4395
4396 // Since the actual number of workers we get may be different
4397 // from the number we requested above, do we need to do anything different
4398 // below? In particular, may be we need to subclass the SequantialSubTasksDone
4399 // class?? XXX
4400 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4401 perm_space->initialize_sequential_subtasks_for_marking(num_workers);
4402
4403 // Refs discovery is already non-atomic.
4404 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4405 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
4406 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());)
4407 conc_workers()->start_task(&tsk);
4408 while (tsk.yielded()) {
4409 tsk.coordinator_yield();
4410 conc_workers()->continue_task(&tsk);
4411 }
4412 // If the task was aborted, _restart_addr will be non-NULL
4413 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4414 while (_restart_addr != NULL) {
4415 // XXX For now we do not make use of ABORTED state and have not
4416 // yet implemented the right abort semantics (even in the original
4417 // single-threaded CMS case). That needs some more investigation
4418 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4419 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4420 // If _restart_addr is non-NULL, a marking stack overflow
4421 // occurred; we need to do a fresh marking iteration from the
4422 // indicated restart address.
4423 if (_foregroundGCIsActive && asynch) {
4424 // We may be running into repeated stack overflows, having
4425 // reached the limit of the stack size, while making very
4426 // slow forward progress. It may be best to bail out and
4427 // let the foreground collector do its job.
4428 // Clear _restart_addr, so that foreground GC
4429 // works from scratch. This avoids the headache of
4430 // a "rescan" which would otherwise be needed because
4431 // of the dirty mod union table & card table.
4432 _restart_addr = NULL;
4433 return false;
4434 }
4435 // Adjust the task to restart from _restart_addr
4436 tsk.reset(_restart_addr);
4437 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4438 _restart_addr);
4439 perm_space->initialize_sequential_subtasks_for_marking(num_workers,
4440 _restart_addr);
4441 _restart_addr = NULL;
4442 // Get the workers going again
4443 conc_workers()->start_task(&tsk);
4444 while (tsk.yielded()) {
4445 tsk.coordinator_yield();
4446 conc_workers()->continue_task(&tsk);
4447 }
4448 }
4449 assert(tsk.completed(), "Inconsistency");
4450 assert(tsk.result() == true, "Inconsistency");
4451 return true;
4452 }
4453
4454 bool CMSCollector::do_marking_st(bool asynch) {
4455 ResourceMark rm;
4456 HandleMark hm;
4457
4458 // Temporarily make refs discovery single threaded (non-MT)
4459 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
4460 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4461 &_markStack, &_revisitStack, CMSYield && asynch);
4462 // the last argument to iterate indicates whether the iteration
4463 // should be incremental with periodic yields.
4464 _markBitMap.iterate(&markFromRootsClosure);
4465 // If _restart_addr is non-NULL, a marking stack overflow
4466 // occurred; we need to do a fresh iteration from the
4467 // indicated restart address.
4468 while (_restart_addr != NULL) {
4469 if (_foregroundGCIsActive && asynch) {
4470 // We may be running into repeated stack overflows, having
4471 // reached the limit of the stack size, while making very
4472 // slow forward progress. It may be best to bail out and
4473 // let the foreground collector do its job.
4474 // Clear _restart_addr, so that foreground GC
4475 // works from scratch. This avoids the headache of
4476 // a "rescan" which would otherwise be needed because
4477 // of the dirty mod union table & card table.
4478 _restart_addr = NULL;
4479 return false; // indicating failure to complete marking
4480 }
4481 // Deal with stack overflow:
4482 // we restart marking from _restart_addr
4483 HeapWord* ra = _restart_addr;
4484 markFromRootsClosure.reset(ra);
4485 _restart_addr = NULL;
4486 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4487 }
4488 return true;
4489 }
4490
4491 void CMSCollector::preclean() {
4492 check_correct_thread_executing();
4493 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4494 verify_work_stacks_empty();
4495 verify_overflow_empty();
4496 _abort_preclean = false;
4497 if (CMSPrecleaningEnabled) {
4498 if (!CMSEdenChunksRecordAlways) {
4499 _eden_chunk_index = 0;
4500 }
4501 size_t used = get_eden_used();
4502 size_t capacity = get_eden_capacity();
4503 // Don't start sampling unless we will get sufficiently
4504 // many samples.
4505 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4506 * CMSScheduleRemarkEdenPenetration)) {
4507 _start_sampling = true;
4508 } else {
4509 _start_sampling = false;
4510 }
4511 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4512 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
4513 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4514 }
4515 CMSTokenSync x(true); // is cms thread
4516 if (CMSPrecleaningEnabled) {
4517 sample_eden();
4518 _collectorState = AbortablePreclean;
4519 } else {
4520 _collectorState = FinalMarking;
4521 }
4522 verify_work_stacks_empty();
4523 verify_overflow_empty();
4524 }
4525
4526 // Try and schedule the remark such that young gen
4527 // occupancy is CMSScheduleRemarkEdenPenetration %.
4528 void CMSCollector::abortable_preclean() {
4529 check_correct_thread_executing();
4530 assert(CMSPrecleaningEnabled, "Inconsistent control state");
4531 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4532
4533 // If Eden's current occupancy is below this threshold,
4534 // immediately schedule the remark; else preclean
4535 // past the next scavenge in an effort to
4536 // schedule the pause as described avove. By choosing
4537 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4538 // we will never do an actual abortable preclean cycle.
4539 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4540 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4541 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
4542 // We need more smarts in the abortable preclean
4543 // loop below to deal with cases where allocation
4544 // in young gen is very very slow, and our precleaning
4545 // is running a losing race against a horde of
4546 // mutators intent on flooding us with CMS updates
4547 // (dirty cards).
4548 // One, admittedly dumb, strategy is to give up
4549 // after a certain number of abortable precleaning loops
4550 // or after a certain maximum time. We want to make
4551 // this smarter in the next iteration.
4552 // XXX FIX ME!!! YSR
4553 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4554 while (!(should_abort_preclean() ||
4555 ConcurrentMarkSweepThread::should_terminate())) {
4556 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4557 cumworkdone += workdone;
4558 loops++;
4559 // Voluntarily terminate abortable preclean phase if we have
4560 // been at it for too long.
4561 if ((CMSMaxAbortablePrecleanLoops != 0) &&
4562 loops >= CMSMaxAbortablePrecleanLoops) {
4563 if (PrintGCDetails) {
4564 gclog_or_tty->print(" CMS: abort preclean due to loops ");
4565 }
4566 break;
4567 }
4568 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4569 if (PrintGCDetails) {
4570 gclog_or_tty->print(" CMS: abort preclean due to time ");
4571 }
4572 break;
4573 }
4574 // If we are doing little work each iteration, we should
4575 // take a short break.
4576 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4577 // Sleep for some time, waiting for work to accumulate
4578 stopTimer();
4579 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4580 startTimer();
4581 waited++;
4582 }
4583 }
4584 if (PrintCMSStatistics > 0) {
4585 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4586 loops, waited, cumworkdone);
4587 }
4588 }
4589 CMSTokenSync x(true); // is cms thread
4590 if (_collectorState != Idling) {
4591 assert(_collectorState == AbortablePreclean,
4592 "Spontaneous state transition?");
4593 _collectorState = FinalMarking;
4594 } // Else, a foreground collection completed this CMS cycle.
4595 return;
4596 }
4597
4598 // Respond to an Eden sampling opportunity
4599 void CMSCollector::sample_eden() {
4600 // Make sure a young gc cannot sneak in between our
4601 // reading and recording of a sample.
4602 assert(Thread::current()->is_ConcurrentGC_thread(),
4603 "Only the cms thread may collect Eden samples");
4604 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4605 "Should collect samples while holding CMS token");
4606 if (!_start_sampling) {
4607 return;
4608 }
4609 // When CMSEdenChunksRecordAlways is true, the eden chunk array
4610 // is populated by the young generation.
4611 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
4612 if (_eden_chunk_index < _eden_chunk_capacity) {
4613 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
4614 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4615 "Unexpected state of Eden");
4616 // We'd like to check that what we just sampled is an oop-start address;
4617 // however, we cannot do that here since the object may not yet have been
4618 // initialized. So we'll instead do the check when we _use_ this sample
4619 // later.
4620 if (_eden_chunk_index == 0 ||
4621 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4622 _eden_chunk_array[_eden_chunk_index-1])
4623 >= CMSSamplingGrain)) {
4624 _eden_chunk_index++; // commit sample
4625 }
4626 }
4627 }
4628 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4629 size_t used = get_eden_used();
4630 size_t capacity = get_eden_capacity();
4631 assert(used <= capacity, "Unexpected state of Eden");
4632 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4633 _abort_preclean = true;
4634 }
4635 }
4636 }
4637
4638
4639 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4640 assert(_collectorState == Precleaning ||
4641 _collectorState == AbortablePreclean, "incorrect state");
4642 ResourceMark rm;
4643 HandleMark hm;
4644
4645 // Precleaning is currently not MT but the reference processor
4646 // may be set for MT. Disable it temporarily here.
4647 ReferenceProcessor* rp = ref_processor();
4648 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
4649
4650 // Do one pass of scrubbing the discovered reference lists
4651 // to remove any reference objects with strongly-reachable
4652 // referents.
4653 if (clean_refs) {
4654 CMSPrecleanRefsYieldClosure yield_cl(this);
4655 assert(rp->span().equals(_span), "Spans should be equal");
4656 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4657 &_markStack, &_revisitStack,
4658 true /* preclean */);
4659 CMSDrainMarkingStackClosure complete_trace(this,
4660 _span, &_markBitMap, &_markStack,
4661 &keep_alive, true /* preclean */);
4662
4663 // We don't want this step to interfere with a young
4664 // collection because we don't want to take CPU
4665 // or memory bandwidth away from the young GC threads
4666 // (which may be as many as there are CPUs).
4667 // Note that we don't need to protect ourselves from
4668 // interference with mutators because they can't
4669 // manipulate the discovered reference lists nor affect
4670 // the computed reachability of the referents, the
4671 // only properties manipulated by the precleaning
4672 // of these reference lists.
4673 stopTimer();
4674 CMSTokenSyncWithLocks x(true /* is cms thread */,
4675 bitMapLock());
4676 startTimer();
4677 sample_eden();
4678
4679 // The following will yield to allow foreground
4680 // collection to proceed promptly. XXX YSR:
4681 // The code in this method may need further
4682 // tweaking for better performance and some restructuring
4683 // for cleaner interfaces.
4684 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
4685 rp->preclean_discovered_references(
4686 rp->is_alive_non_header(), &keep_alive, &complete_trace,
4687 &yield_cl, should_unload_classes(), gc_timer);
4688 }
4689
4690 if (clean_survivor) { // preclean the active survivor space(s)
4691 assert(_young_gen->kind() == Generation::DefNew ||
4692 _young_gen->kind() == Generation::ParNew ||
4693 _young_gen->kind() == Generation::ASParNew,
4694 "incorrect type for cast");
4695 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4696 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4697 &_markBitMap, &_modUnionTable,
4698 &_markStack, &_revisitStack,
4699 true /* precleaning phase */);
4700 stopTimer();
4701 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4702 bitMapLock());
4703 startTimer();
4704 unsigned int before_count =
4705 GenCollectedHeap::heap()->total_collections();
4706 SurvivorSpacePrecleanClosure
4707 sss_cl(this, _span, &_markBitMap, &_markStack,
4708 &pam_cl, before_count, CMSYield);
4709 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4710 dng->from()->object_iterate_careful(&sss_cl);
4711 dng->to()->object_iterate_careful(&sss_cl);
4712 }
4713 MarkRefsIntoAndScanClosure
4714 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4715 &_markStack, &_revisitStack, this, CMSYield,
4716 true /* precleaning phase */);
4717 // CAUTION: The following closure has persistent state that may need to
4718 // be reset upon a decrease in the sequence of addresses it
4719 // processes.
4720 ScanMarkedObjectsAgainCarefullyClosure
4721 smoac_cl(this, _span,
4722 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield);
4723
4724 // Preclean dirty cards in ModUnionTable and CardTable using
4725 // appropriate convergence criterion;
4726 // repeat CMSPrecleanIter times unless we find that
4727 // we are losing.
4728 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4729 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4730 "Bad convergence multiplier");
4731 assert(CMSPrecleanThreshold >= 100,
4732 "Unreasonably low CMSPrecleanThreshold");
4733
4734 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4735 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4736 numIter < CMSPrecleanIter;
4737 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4738 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4739 if (CMSPermGenPrecleaningEnabled) {
4740 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl);
4741 }
4742 if (Verbose && PrintGCDetails) {
4743 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4744 }
4745 // Either there are very few dirty cards, so re-mark
4746 // pause will be small anyway, or our pre-cleaning isn't
4747 // that much faster than the rate at which cards are being
4748 // dirtied, so we might as well stop and re-mark since
4749 // precleaning won't improve our re-mark time by much.
4750 if (curNumCards <= CMSPrecleanThreshold ||
4751 (numIter > 0 &&
4752 (curNumCards * CMSPrecleanDenominator >
4753 lastNumCards * CMSPrecleanNumerator))) {
4754 numIter++;
4755 cumNumCards += curNumCards;
4756 break;
4757 }
4758 }
4759 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4760 if (CMSPermGenPrecleaningEnabled) {
4761 curNumCards += preclean_card_table(_permGen, &smoac_cl);
4762 }
4763 cumNumCards += curNumCards;
4764 if (PrintGCDetails && PrintCMSStatistics != 0) {
4765 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4766 curNumCards, cumNumCards, numIter);
4767 }
4768 return cumNumCards; // as a measure of useful work done
4769 }
4770
4771 // PRECLEANING NOTES:
4772 // Precleaning involves:
4773 // . reading the bits of the modUnionTable and clearing the set bits.
4774 // . For the cards corresponding to the set bits, we scan the
4775 // objects on those cards. This means we need the free_list_lock
4776 // so that we can safely iterate over the CMS space when scanning
4777 // for oops.
4778 // . When we scan the objects, we'll be both reading and setting
4779 // marks in the marking bit map, so we'll need the marking bit map.
4780 // . For protecting _collector_state transitions, we take the CGC_lock.
4781 // Note that any races in the reading of of card table entries by the
4782 // CMS thread on the one hand and the clearing of those entries by the
4783 // VM thread or the setting of those entries by the mutator threads on the
4784 // other are quite benign. However, for efficiency it makes sense to keep
4785 // the VM thread from racing with the CMS thread while the latter is
4786 // dirty card info to the modUnionTable. We therefore also use the
4787 // CGC_lock to protect the reading of the card table and the mod union
4788 // table by the CM thread.
4789 // . We run concurrently with mutator updates, so scanning
4790 // needs to be done carefully -- we should not try to scan
4791 // potentially uninitialized objects.
4792 //
4793 // Locking strategy: While holding the CGC_lock, we scan over and
4794 // reset a maximal dirty range of the mod union / card tables, then lock
4795 // the free_list_lock and bitmap lock to do a full marking, then
4796 // release these locks; and repeat the cycle. This allows for a
4797 // certain amount of fairness in the sharing of these locks between
4798 // the CMS collector on the one hand, and the VM thread and the
4799 // mutators on the other.
4800
4801 // NOTE: preclean_mod_union_table() and preclean_card_table()
4802 // further below are largely identical; if you need to modify
4803 // one of these methods, please check the other method too.
4804
4805 size_t CMSCollector::preclean_mod_union_table(
4806 ConcurrentMarkSweepGeneration* gen,
4807 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4808 verify_work_stacks_empty();
4809 verify_overflow_empty();
4810
4811 // Turn off checking for this method but turn it back on
4812 // selectively. There are yield points in this method
4813 // but it is difficult to turn the checking off just around
4814 // the yield points. It is simpler to selectively turn
4815 // it on.
4816 DEBUG_ONLY(RememberKlassesChecker mux(false);)
4817
4818 // strategy: starting with the first card, accumulate contiguous
4819 // ranges of dirty cards; clear these cards, then scan the region
4820 // covered by these cards.
4821
4822 // Since all of the MUT is committed ahead, we can just use
4823 // that, in case the generations expand while we are precleaning.
4824 // It might also be fine to just use the committed part of the
4825 // generation, but we might potentially miss cards when the
4826 // generation is rapidly expanding while we are in the midst
4827 // of precleaning.
4828 HeapWord* startAddr = gen->reserved().start();
4829 HeapWord* endAddr = gen->reserved().end();
4830
4831 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4832
4833 size_t numDirtyCards, cumNumDirtyCards;
4834 HeapWord *nextAddr, *lastAddr;
4835 for (cumNumDirtyCards = numDirtyCards = 0,
4836 nextAddr = lastAddr = startAddr;
4837 nextAddr < endAddr;
4838 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4839
4840 ResourceMark rm;
4841 HandleMark hm;
4842
4843 MemRegion dirtyRegion;
4844 {
4845 stopTimer();
4846 // Potential yield point
4847 CMSTokenSync ts(true);
4848 startTimer();
4849 sample_eden();
4850 // Get dirty region starting at nextOffset (inclusive),
4851 // simultaneously clearing it.
4852 dirtyRegion =
4853 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4854 assert(dirtyRegion.start() >= nextAddr,
4855 "returned region inconsistent?");
4856 }
4857 // Remember where the next search should begin.
4858 // The returned region (if non-empty) is a right open interval,
4859 // so lastOffset is obtained from the right end of that
4860 // interval.
4861 lastAddr = dirtyRegion.end();
4862 // Should do something more transparent and less hacky XXX
4863 numDirtyCards =
4864 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4865
4866 // We'll scan the cards in the dirty region (with periodic
4867 // yields for foreground GC as needed).
4868 if (!dirtyRegion.is_empty()) {
4869 assert(numDirtyCards > 0, "consistency check");
4870 HeapWord* stop_point = NULL;
4871 stopTimer();
4872 // Potential yield point
4873 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4874 bitMapLock());
4875 startTimer();
4876 {
4877 verify_work_stacks_empty();
4878 verify_overflow_empty();
4879 sample_eden();
4880 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4881 stop_point =
4882 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4883 }
4884 if (stop_point != NULL) {
4885 // The careful iteration stopped early either because it found an
4886 // uninitialized object, or because we were in the midst of an
4887 // "abortable preclean", which should now be aborted. Redirty
4888 // the bits corresponding to the partially-scanned or unscanned
4889 // cards. We'll either restart at the next block boundary or
4890 // abort the preclean.
4891 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) ||
4892 (_collectorState == AbortablePreclean && should_abort_preclean()),
4893 "Unparsable objects should only be in perm gen.");
4894 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4895 if (should_abort_preclean()) {
4896 break; // out of preclean loop
4897 } else {
4898 // Compute the next address at which preclean should pick up;
4899 // might need bitMapLock in order to read P-bits.
4900 lastAddr = next_card_start_after_block(stop_point);
4901 }
4902 }
4903 } else {
4904 assert(lastAddr == endAddr, "consistency check");
4905 assert(numDirtyCards == 0, "consistency check");
4906 break;
4907 }
4908 }
4909 verify_work_stacks_empty();
4910 verify_overflow_empty();
4911 return cumNumDirtyCards;
4912 }
4913
4914 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4915 // below are largely identical; if you need to modify
4916 // one of these methods, please check the other method too.
4917
4918 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4919 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4920 // strategy: it's similar to precleamModUnionTable above, in that
4921 // we accumulate contiguous ranges of dirty cards, mark these cards
4922 // precleaned, then scan the region covered by these cards.
4923 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4924 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4925
4926 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4927
4928 size_t numDirtyCards, cumNumDirtyCards;
4929 HeapWord *lastAddr, *nextAddr;
4930
4931 for (cumNumDirtyCards = numDirtyCards = 0,
4932 nextAddr = lastAddr = startAddr;
4933 nextAddr < endAddr;
4934 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4935
4936 ResourceMark rm;
4937 HandleMark hm;
4938
4939 MemRegion dirtyRegion;
4940 {
4941 // See comments in "Precleaning notes" above on why we
4942 // do this locking. XXX Could the locking overheads be
4943 // too high when dirty cards are sparse? [I don't think so.]
4944 stopTimer();
4945 CMSTokenSync x(true); // is cms thread
4946 startTimer();
4947 sample_eden();
4948 // Get and clear dirty region from card table
4949 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4950 MemRegion(nextAddr, endAddr),
4951 true,
4952 CardTableModRefBS::precleaned_card_val());
4953
4954 assert(dirtyRegion.start() >= nextAddr,
4955 "returned region inconsistent?");
4956 }
4957 lastAddr = dirtyRegion.end();
4958 numDirtyCards =
4959 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4960
4961 if (!dirtyRegion.is_empty()) {
4962 stopTimer();
4963 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4964 startTimer();
4965 sample_eden();
4966 verify_work_stacks_empty();
4967 verify_overflow_empty();
4968 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4969 HeapWord* stop_point =
4970 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4971 if (stop_point != NULL) {
4972 // The careful iteration stopped early because it found an
4973 // uninitialized object. Redirty the bits corresponding to the
4974 // partially-scanned or unscanned cards, and start again at the
4975 // next block boundary.
4976 assert(CMSPermGenPrecleaningEnabled ||
4977 (_collectorState == AbortablePreclean && should_abort_preclean()),
4978 "Unparsable objects should only be in perm gen.");
4979 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4980 if (should_abort_preclean()) {
4981 break; // out of preclean loop
4982 } else {
4983 // Compute the next address at which preclean should pick up.
4984 lastAddr = next_card_start_after_block(stop_point);
4985 }
4986 }
4987 } else {
4988 break;
4989 }
4990 }
4991 verify_work_stacks_empty();
4992 verify_overflow_empty();
4993 return cumNumDirtyCards;
4994 }
4995
4996 void CMSCollector::checkpointRootsFinal(bool asynch,
4997 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4998 assert(_collectorState == FinalMarking, "incorrect state transition?");
4999 check_correct_thread_executing();
5000 // world is stopped at this checkpoint
5001 assert(SafepointSynchronize::is_at_safepoint(),
5002 "world should be stopped");
5003 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5004
5005 verify_work_stacks_empty();
5006 verify_overflow_empty();
5007
5008 SpecializationStats::clear();
5009 if (PrintGCDetails) {
5010 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
5011 _young_gen->used() / K,
5012 _young_gen->capacity() / K);
5013 }
5014 if (asynch) {
5015 if (CMSScavengeBeforeRemark) {
5016 GenCollectedHeap* gch = GenCollectedHeap::heap();
5017 // Temporarily set flag to false, GCH->do_collection will
5018 // expect it to be false and set to true
5019 FlagSetting fl(gch->_is_gc_active, false);
5020 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
5021 PrintGCDetails && Verbose, true, _gc_timer_cm);)
5022 int level = _cmsGen->level() - 1;
5023 if (level >= 0) {
5024 gch->do_collection(true, // full (i.e. force, see below)
5025 false, // !clear_all_soft_refs
5026 0, // size
5027 false, // is_tlab
5028 level // max_level
5029 );
5030 }
5031 }
5032 FreelistLocker x(this);
5033 MutexLockerEx y(bitMapLock(),
5034 Mutex::_no_safepoint_check_flag);
5035 assert(!init_mark_was_synchronous, "but that's impossible!");
5036 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
5037 } else {
5038 // already have all the locks
5039 checkpointRootsFinalWork(asynch, clear_all_soft_refs,
5040 init_mark_was_synchronous);
5041 }
5042 verify_work_stacks_empty();
5043 verify_overflow_empty();
5044 SpecializationStats::print();
5045 }
5046
5047 void CMSCollector::checkpointRootsFinalWork(bool asynch,
5048 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
5049
5050 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);)
5051
5052 assert(haveFreelistLocks(), "must have free list locks");
5053 assert_lock_strong(bitMapLock());
5054
5055 if (UseAdaptiveSizePolicy) {
5056 size_policy()->checkpoint_roots_final_begin();
5057 }
5058
5059 ResourceMark rm;
5060 HandleMark hm;
5061
5062 GenCollectedHeap* gch = GenCollectedHeap::heap();
5063
5064 if (should_unload_classes()) {
5065 CodeCache::gc_prologue();
5066 }
5067 assert(haveFreelistLocks(), "must have free list locks");
5068 assert_lock_strong(bitMapLock());
5069
5070 DEBUG_ONLY(RememberKlassesChecker fmx(should_unload_classes());)
5071 if (!init_mark_was_synchronous) {
5072 // We might assume that we need not fill TLAB's when
5073 // CMSScavengeBeforeRemark is set, because we may have just done
5074 // a scavenge which would have filled all TLAB's -- and besides
5075 // Eden would be empty. This however may not always be the case --
5076 // for instance although we asked for a scavenge, it may not have
5077 // happened because of a JNI critical section. We probably need
5078 // a policy for deciding whether we can in that case wait until
5079 // the critical section releases and then do the remark following
5080 // the scavenge, and skip it here. In the absence of that policy,
5081 // or of an indication of whether the scavenge did indeed occur,
5082 // we cannot rely on TLAB's having been filled and must do
5083 // so here just in case a scavenge did not happen.
5084 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
5085 // Update the saved marks which may affect the root scans.
5086 gch->save_marks();
5087
5088 if (CMSPrintEdenSurvivorChunks) {
5089 print_eden_and_survivor_chunk_arrays();
5090 }
5091
5092 {
5093 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
5094
5095 // Note on the role of the mod union table:
5096 // Since the marker in "markFromRoots" marks concurrently with
5097 // mutators, it is possible for some reachable objects not to have been
5098 // scanned. For instance, an only reference to an object A was
5099 // placed in object B after the marker scanned B. Unless B is rescanned,
5100 // A would be collected. Such updates to references in marked objects
5101 // are detected via the mod union table which is the set of all cards
5102 // dirtied since the first checkpoint in this GC cycle and prior to
5103 // the most recent young generation GC, minus those cleaned up by the
5104 // concurrent precleaning.
5105 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
5106 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm);
5107 do_remark_parallel();
5108 } else {
5109 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
5110 _gc_timer_cm);
5111 do_remark_non_parallel();
5112 }
5113 }
5114 } else {
5115 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
5116 // The initial mark was stop-world, so there's no rescanning to
5117 // do; go straight on to the next step below.
5118 }
5119 verify_work_stacks_empty();
5120 verify_overflow_empty();
5121
5122 {
5123 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);)
5124 refProcessingWork(asynch, clear_all_soft_refs);
5125 }
5126 verify_work_stacks_empty();
5127 verify_overflow_empty();
5128
5129 if (should_unload_classes()) {
5130 CodeCache::gc_epilogue();
5131 }
5132 JvmtiExport::gc_epilogue();
5133
5134 // If we encountered any (marking stack / work queue) overflow
5135 // events during the current CMS cycle, take appropriate
5136 // remedial measures, where possible, so as to try and avoid
5137 // recurrence of that condition.
5138 assert(_markStack.isEmpty(), "No grey objects");
5139 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
5140 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
5141 if (ser_ovflw > 0) {
5142 if (PrintCMSStatistics != 0) {
5143 gclog_or_tty->print_cr("Marking stack overflow (benign) "
5144 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
5145 ", kac_preclean="SIZE_FORMAT")",
5146 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
5147 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
5148 }
5149 _markStack.expand();
5150 _ser_pmc_remark_ovflw = 0;
5151 _ser_pmc_preclean_ovflw = 0;
5152 _ser_kac_preclean_ovflw = 0;
5153 _ser_kac_ovflw = 0;
5154 }
5155 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
5156 if (PrintCMSStatistics != 0) {
5157 gclog_or_tty->print_cr("Work queue overflow (benign) "
5158 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
5159 _par_pmc_remark_ovflw, _par_kac_ovflw);
5160 }
5161 _par_pmc_remark_ovflw = 0;
5162 _par_kac_ovflw = 0;
5163 }
5164 if (PrintCMSStatistics != 0) {
5165 if (_markStack._hit_limit > 0) {
5166 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
5167 _markStack._hit_limit);
5168 }
5169 if (_markStack._failed_double > 0) {
5170 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
5171 " current capacity "SIZE_FORMAT,
5172 _markStack._failed_double,
5173 _markStack.capacity());
5174 }
5175 }
5176 _markStack._hit_limit = 0;
5177 _markStack._failed_double = 0;
5178
5179 // Check that all the klasses have been checked
5180 assert(_revisitStack.isEmpty(), "Not all klasses revisited");
5181
5182 if ((VerifyAfterGC || VerifyDuringGC) &&
5183 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5184 verify_after_remark();
5185 }
5186
5187 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
5188
5189 // Change under the freelistLocks.
5190 _collectorState = Sweeping;
5191 // Call isAllClear() under bitMapLock
5192 assert(_modUnionTable.isAllClear(), "Should be clear by end of the"
5193 " final marking");
5194 if (UseAdaptiveSizePolicy) {
5195 size_policy()->checkpoint_roots_final_end(gch->gc_cause());
5196 }
5197 }
5198
5199 void CMSParInitialMarkTask::work(uint worker_id) {
5200 elapsedTimer _timer;
5201 ResourceMark rm;
5202 HandleMark hm;
5203
5204 // ---------- scan from roots --------------
5205 _timer.start();
5206 GenCollectedHeap* gch = GenCollectedHeap::heap();
5207 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
5208
5209 // ---------- young gen roots --------------
5210 {
5211 work_on_young_gen_roots(worker_id, &par_mri_cl);
5212 _timer.stop();
5213 if (PrintCMSStatistics != 0) {
5214 gclog_or_tty->print_cr(
5215 "Finished young gen initial mark scan work in %dth thread: %3.3f sec",
5216 worker_id, _timer.seconds());
5217 }
5218 }
5219
5220 // ---------- remaining roots --------------
5221 _timer.reset();
5222 _timer.start();
5223 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
5224 false, // yg was scanned above
5225 false, // this is parallel code
5226 true, // collecting perm gen
5227 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5228 &par_mri_cl,
5229 true, // walk all of code cache if (so & SO_CodeCache)
5230 NULL);
5231 assert(_collector->should_unload_classes()
5232 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache),
5233 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5234 _timer.stop();
5235 if (PrintCMSStatistics != 0) {
5236 gclog_or_tty->print_cr(
5237 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec",
5238 worker_id, _timer.seconds());
5239 }
5240 }
5241
5242 // Parallel remark task
5243 class CMSParRemarkTask: public CMSParMarkTask {
5244 CompactibleFreeListSpace* _cms_space;
5245 CompactibleFreeListSpace* _perm_space;
5246
5247 // The per-thread work queues, available here for stealing.
5248 OopTaskQueueSet* _task_queues;
5249 ParallelTaskTerminator _term;
5250
5251 public:
5252 // A value of 0 passed to n_workers will cause the number of
5253 // workers to be taken from the active workers in the work gang.
5254 CMSParRemarkTask(CMSCollector* collector,
5255 CompactibleFreeListSpace* cms_space,
5256 CompactibleFreeListSpace* perm_space,
5257 int n_workers, FlexibleWorkGang* workers,
5258 OopTaskQueueSet* task_queues):
5259 CMSParMarkTask("Rescan roots and grey objects in parallel",
5260 collector, n_workers),
5261 _cms_space(cms_space), _perm_space(perm_space),
5262 _task_queues(task_queues),
5263 _term(n_workers, task_queues) { }
5264
5265 OopTaskQueueSet* task_queues() { return _task_queues; }
5266
5267 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5268
5269 ParallelTaskTerminator* terminator() { return &_term; }
5270 int n_workers() { return _n_workers; }
5271
5272 void work(uint worker_id);
5273
5274 private:
5275 // ... of dirty cards in old space
5276 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
5277 Par_MarkRefsIntoAndScanClosure* cl);
5278
5279 // ... work stealing for the above
5280 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
5281 };
5282
5283 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
5284 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
5285 EdenSpace* eden_space = dng->eden();
5286 ContiguousSpace* from_space = dng->from();
5287 ContiguousSpace* to_space = dng->to();
5288
5289 HeapWord** eca = _collector->_eden_chunk_array;
5290 size_t ect = _collector->_eden_chunk_index;
5291 HeapWord** sca = _collector->_survivor_chunk_array;
5292 size_t sct = _collector->_survivor_chunk_index;
5293
5294 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
5295 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
5296
5297 do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
5298 do_young_space_rescan(worker_id, cl, from_space, sca, sct);
5299 do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
5300 }
5301
5302 // work_queue(i) is passed to the closure
5303 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
5304 // also is passed to do_dirty_card_rescan_tasks() and to
5305 // do_work_steal() to select the i-th task_queue.
5306
5307 void CMSParRemarkTask::work(uint worker_id) {
5308 elapsedTimer _timer;
5309 ResourceMark rm;
5310 HandleMark hm;
5311
5312 // ---------- rescan from roots --------------
5313 _timer.start();
5314 GenCollectedHeap* gch = GenCollectedHeap::heap();
5315 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
5316 _collector->_span, _collector->ref_processor(),
5317 &(_collector->_markBitMap),
5318 work_queue(worker_id), &(_collector->_revisitStack));
5319
5320 // Rescan young gen roots first since these are likely
5321 // coarsely partitioned and may, on that account, constitute
5322 // the critical path; thus, it's best to start off that
5323 // work first.
5324 // ---------- young gen roots --------------
5325 {
5326 work_on_young_gen_roots(worker_id, &par_mrias_cl);
5327 _timer.stop();
5328 if (PrintCMSStatistics != 0) {
5329 gclog_or_tty->print_cr(
5330 "Finished young gen rescan work in %dth thread: %3.3f sec",
5331 worker_id, _timer.seconds());
5332 }
5333 }
5334
5335 // ---------- remaining roots --------------
5336 _timer.reset();
5337 _timer.start();
5338 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
5339 false, // yg was scanned above
5340 false, // this is parallel code
5341 true, // collecting perm gen
5342 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5343 &par_mrias_cl,
5344 true, // walk all of code cache if (so & SO_CodeCache)
5345 NULL);
5346 assert(_collector->should_unload_classes()
5347 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache),
5348 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5349 _timer.stop();
5350 if (PrintCMSStatistics != 0) {
5351 gclog_or_tty->print_cr(
5352 "Finished remaining root rescan work in %dth thread: %3.3f sec",
5353 worker_id, _timer.seconds());
5354 }
5355
5356 // ---------- rescan dirty cards ------------
5357 _timer.reset();
5358 _timer.start();
5359
5360 // Do the rescan tasks for each of the two spaces
5361 // (cms_space and perm_space) in turn.
5362 // "worker_id" is passed to select the task_queue for "worker_id"
5363 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
5364 do_dirty_card_rescan_tasks(_perm_space, worker_id, &par_mrias_cl);
5365 _timer.stop();
5366 if (PrintCMSStatistics != 0) {
5367 gclog_or_tty->print_cr(
5368 "Finished dirty card rescan work in %dth thread: %3.3f sec",
5369 worker_id, _timer.seconds());
5370 }
5371
5372 // ---------- steal work from other threads ...
5373 // ---------- ... and drain overflow list.
5374 _timer.reset();
5375 _timer.start();
5376 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
5377 _timer.stop();
5378 if (PrintCMSStatistics != 0) {
5379 gclog_or_tty->print_cr(
5380 "Finished work stealing in %dth thread: %3.3f sec",
5381 worker_id, _timer.seconds());
5382 }
5383 }
5384
5385 // Note that parameter "i" is not used.
5386 void
5387 CMSParMarkTask::do_young_space_rescan(uint worker_id,
5388 OopsInGenClosure* cl, ContiguousSpace* space,
5389 HeapWord** chunk_array, size_t chunk_top) {
5390 // Until all tasks completed:
5391 // . claim an unclaimed task
5392 // . compute region boundaries corresponding to task claimed
5393 // using chunk_array
5394 // . par_oop_iterate(cl) over that region
5395
5396 ResourceMark rm;
5397 HandleMark hm;
5398
5399 SequentialSubTasksDone* pst = space->par_seq_tasks();
5400 assert(pst->valid(), "Uninitialized use?");
5401
5402 uint nth_task = 0;
5403 uint n_tasks = pst->n_tasks();
5404
5405 HeapWord *start, *end;
5406 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5407 // We claimed task # nth_task; compute its boundaries.
5408 if (chunk_top == 0) { // no samples were taken
5409 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
5410 start = space->bottom();
5411 end = space->top();
5412 } else if (nth_task == 0) {
5413 start = space->bottom();
5414 end = chunk_array[nth_task];
5415 } else if (nth_task < (uint)chunk_top) {
5416 assert(nth_task >= 1, "Control point invariant");
5417 start = chunk_array[nth_task - 1];
5418 end = chunk_array[nth_task];
5419 } else {
5420 assert(nth_task == (uint)chunk_top, "Control point invariant");
5421 start = chunk_array[chunk_top - 1];
5422 end = space->top();
5423 }
5424 MemRegion mr(start, end);
5425 // Verify that mr is in space
5426 assert(mr.is_empty() || space->used_region().contains(mr),
5427 "Should be in space");
5428 // Verify that "start" is an object boundary
5429 assert(mr.is_empty() || oop(mr.start())->is_oop(),
5430 "Should be an oop");
5431 space->par_oop_iterate(mr, cl);
5432 }
5433 pst->all_tasks_completed();
5434 }
5435
5436 void
5437 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5438 CompactibleFreeListSpace* sp, int i,
5439 Par_MarkRefsIntoAndScanClosure* cl) {
5440 // Until all tasks completed:
5441 // . claim an unclaimed task
5442 // . compute region boundaries corresponding to task claimed
5443 // . transfer dirty bits ct->mut for that region
5444 // . apply rescanclosure to dirty mut bits for that region
5445
5446 ResourceMark rm;
5447 HandleMark hm;
5448
5449 OopTaskQueue* work_q = work_queue(i);
5450 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5451 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5452 // CAUTION: This closure has state that persists across calls to
5453 // the work method dirty_range_iterate_clear() in that it has
5454 // imbedded in it a (subtype of) UpwardsObjectClosure. The
5455 // use of that state in the imbedded UpwardsObjectClosure instance
5456 // assumes that the cards are always iterated (even if in parallel
5457 // by several threads) in monotonically increasing order per each
5458 // thread. This is true of the implementation below which picks
5459 // card ranges (chunks) in monotonically increasing order globally
5460 // and, a-fortiori, in monotonically increasing order per thread
5461 // (the latter order being a subsequence of the former).
5462 // If the work code below is ever reorganized into a more chaotic
5463 // work-partitioning form than the current "sequential tasks"
5464 // paradigm, the use of that persistent state will have to be
5465 // revisited and modified appropriately. See also related
5466 // bug 4756801 work on which should examine this code to make
5467 // sure that the changes there do not run counter to the
5468 // assumptions made here and necessary for correctness and
5469 // efficiency. Note also that this code might yield inefficient
5470 // behaviour in the case of very large objects that span one or
5471 // more work chunks. Such objects would potentially be scanned
5472 // several times redundantly. Work on 4756801 should try and
5473 // address that performance anomaly if at all possible. XXX
5474 MemRegion full_span = _collector->_span;
5475 CMSBitMap* bm = &(_collector->_markBitMap); // shared
5476 CMSMarkStack* rs = &(_collector->_revisitStack); // shared
5477 MarkFromDirtyCardsClosure
5478 greyRescanClosure(_collector, full_span, // entire span of interest
5479 sp, bm, work_q, rs, cl);
5480
5481 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5482 assert(pst->valid(), "Uninitialized use?");
5483 uint nth_task = 0;
5484 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5485 MemRegion span = sp->used_region();
5486 HeapWord* start_addr = span.start();
5487 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5488 alignment);
5489 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5490 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5491 start_addr, "Check alignment");
5492 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5493 chunk_size, "Check alignment");
5494
5495 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5496 // Having claimed the nth_task, compute corresponding mem-region,
5497 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
5498 // The alignment restriction ensures that we do not need any
5499 // synchronization with other gang-workers while setting or
5500 // clearing bits in thus chunk of the MUT.
5501 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5502 start_addr + (nth_task+1)*chunk_size);
5503 // The last chunk's end might be way beyond end of the
5504 // used region. In that case pull back appropriately.
5505 if (this_span.end() > end_addr) {
5506 this_span.set_end(end_addr);
5507 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5508 }
5509 // Iterate over the dirty cards covering this chunk, marking them
5510 // precleaned, and setting the corresponding bits in the mod union
5511 // table. Since we have been careful to partition at Card and MUT-word
5512 // boundaries no synchronization is needed between parallel threads.
5513 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5514 &modUnionClosure);
5515
5516 // Having transferred these marks into the modUnionTable,
5517 // rescan the marked objects on the dirty cards in the modUnionTable.
5518 // Even if this is at a synchronous collection, the initial marking
5519 // may have been done during an asynchronous collection so there
5520 // may be dirty bits in the mod-union table.
5521 _collector->_modUnionTable.dirty_range_iterate_clear(
5522 this_span, &greyRescanClosure);
5523 _collector->_modUnionTable.verifyNoOneBitsInRange(
5524 this_span.start(),
5525 this_span.end());
5526 }
5527 pst->all_tasks_completed(); // declare that i am done
5528 }
5529
5530 // . see if we can share work_queues with ParNew? XXX
5531 void
5532 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5533 int* seed) {
5534 OopTaskQueue* work_q = work_queue(i);
5535 NOT_PRODUCT(int num_steals = 0;)
5536 oop obj_to_scan;
5537 CMSBitMap* bm = &(_collector->_markBitMap);
5538
5539 while (true) {
5540 // Completely finish any left over work from (an) earlier round(s)
5541 cl->trim_queue(0);
5542 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5543 (size_t)ParGCDesiredObjsFromOverflowList);
5544 // Now check if there's any work in the overflow list
5545 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5546 // only affects the number of attempts made to get work from the
5547 // overflow list and does not affect the number of workers. Just
5548 // pass ParallelGCThreads so this behavior is unchanged.
5549 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5550 work_q,
5551 ParallelGCThreads)) {
5552 // found something in global overflow list;
5553 // not yet ready to go stealing work from others.
5554 // We'd like to assert(work_q->size() != 0, ...)
5555 // because we just took work from the overflow list,
5556 // but of course we can't since all of that could have
5557 // been already stolen from us.
5558 // "He giveth and He taketh away."
5559 continue;
5560 }
5561 // Verify that we have no work before we resort to stealing
5562 assert(work_q->size() == 0, "Have work, shouldn't steal");
5563 // Try to steal from other queues that have work
5564 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5565 NOT_PRODUCT(num_steals++;)
5566 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5567 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5568 // Do scanning work
5569 obj_to_scan->oop_iterate(cl);
5570 // Loop around, finish this work, and try to steal some more
5571 } else if (terminator()->offer_termination()) {
5572 break; // nirvana from the infinite cycle
5573 }
5574 }
5575 NOT_PRODUCT(
5576 if (PrintCMSStatistics != 0) {
5577 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5578 }
5579 )
5580 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5581 "Else our work is not yet done");
5582 }
5583
5584 // Record object boundaries in _eden_chunk_array by sampling the eden
5585 // top in the slow-path eden object allocation code path and record
5586 // the boundaries, if CMSEdenChunksRecordAlways is true. If
5587 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
5588 // sampling in sample_eden() that activates during the part of the
5589 // preclean phase.
5590 void CMSCollector::sample_eden_chunk() {
5591 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
5592 if (_eden_chunk_lock->try_lock()) {
5593 // Record a sample. This is the critical section. The contents
5594 // of the _eden_chunk_array have to be non-decreasing in the
5595 // address order.
5596 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
5597 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
5598 "Unexpected state of Eden");
5599 if (_eden_chunk_index == 0 ||
5600 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
5601 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
5602 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
5603 _eden_chunk_index++; // commit sample
5604 }
5605 _eden_chunk_lock->unlock();
5606 }
5607 }
5608 }
5609
5610 // Return a thread-local PLAB recording array, as appropriate.
5611 void* CMSCollector::get_data_recorder(int thr_num) {
5612 if (_survivor_plab_array != NULL &&
5613 (CMSPLABRecordAlways ||
5614 (_collectorState > Marking && _collectorState < FinalMarking))) {
5615 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5616 ChunkArray* ca = &_survivor_plab_array[thr_num];
5617 ca->reset(); // clear it so that fresh data is recorded
5618 return (void*) ca;
5619 } else {
5620 return NULL;
5621 }
5622 }
5623
5624 // Reset all the thread-local PLAB recording arrays
5625 void CMSCollector::reset_survivor_plab_arrays() {
5626 for (uint i = 0; i < ParallelGCThreads; i++) {
5627 _survivor_plab_array[i].reset();
5628 }
5629 }
5630
5631 // Merge the per-thread plab arrays into the global survivor chunk
5632 // array which will provide the partitioning of the survivor space
5633 // for CMS initial scan and rescan.
5634 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
5635 int no_of_gc_threads) {
5636 assert(_survivor_plab_array != NULL, "Error");
5637 assert(_survivor_chunk_array != NULL, "Error");
5638 assert(_collectorState == FinalMarking ||
5639 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
5640 for (int j = 0; j < no_of_gc_threads; j++) {
5641 _cursor[j] = 0;
5642 }
5643 HeapWord* top = surv->top();
5644 size_t i;
5645 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
5646 HeapWord* min_val = top; // Higher than any PLAB address
5647 uint min_tid = 0; // position of min_val this round
5648 for (int j = 0; j < no_of_gc_threads; j++) {
5649 ChunkArray* cur_sca = &_survivor_plab_array[j];
5650 if (_cursor[j] == cur_sca->end()) {
5651 continue;
5652 }
5653 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5654 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5655 assert(surv->used_region().contains(cur_val), "Out of bounds value");
5656 if (cur_val < min_val) {
5657 min_tid = j;
5658 min_val = cur_val;
5659 } else {
5660 assert(cur_val < top, "All recorded addresses should be less");
5661 }
5662 }
5663 // At this point min_val and min_tid are respectively
5664 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5665 // and the thread (j) that witnesses that address.
5666 // We record this address in the _survivor_chunk_array[i]
5667 // and increment _cursor[min_tid] prior to the next round i.
5668 if (min_val == top) {
5669 break;
5670 }
5671 _survivor_chunk_array[i] = min_val;
5672 _cursor[min_tid]++;
5673 }
5674 // We are all done; record the size of the _survivor_chunk_array
5675 _survivor_chunk_index = i; // exclusive: [0, i)
5676 if (PrintCMSStatistics > 0) {
5677 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5678 }
5679 // Verify that we used up all the recorded entries
5680 #ifdef ASSERT
5681 size_t total = 0;
5682 for (int j = 0; j < no_of_gc_threads; j++) {
5683 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5684 total += _cursor[j];
5685 }
5686 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5687 // Check that the merged array is in sorted order
5688 if (total > 0) {
5689 for (size_t i = 0; i < total - 1; i++) {
5690 if (PrintCMSStatistics > 0) {
5691 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5692 i, _survivor_chunk_array[i]);
5693 }
5694 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5695 "Not sorted");
5696 }
5697 }
5698 #endif // ASSERT
5699 }
5700
5701 // Set up the space's par_seq_tasks structure for work claiming
5702 // for parallel initial scan and rescan of young gen.
5703 // See ParRescanTask where this is currently used.
5704 void
5705 CMSCollector::
5706 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5707 assert(n_threads > 0, "Unexpected n_threads argument");
5708 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5709
5710 // Eden space
5711 {
5712 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5713 assert(!pst->valid(), "Clobbering existing data?");
5714 // Each valid entry in [0, _eden_chunk_index) represents a task.
5715 size_t n_tasks = _eden_chunk_index + 1;
5716 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5717 // Sets the condition for completion of the subtask (how many threads
5718 // need to finish in order to be done).
5719 pst->set_n_threads(n_threads);
5720 pst->set_n_tasks((int)n_tasks);
5721 }
5722
5723 // Merge the survivor plab arrays into _survivor_chunk_array
5724 if (_survivor_plab_array != NULL) {
5725 merge_survivor_plab_arrays(dng->from(), n_threads);
5726 } else {
5727 assert(_survivor_chunk_index == 0, "Error");
5728 }
5729
5730 // To space
5731 {
5732 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5733 assert(!pst->valid(), "Clobbering existing data?");
5734 // Sets the condition for completion of the subtask (how many threads
5735 // need to finish in order to be done).
5736 pst->set_n_threads(n_threads);
5737 pst->set_n_tasks(1);
5738 assert(pst->valid(), "Error");
5739 }
5740
5741 // From space
5742 {
5743 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5744 assert(!pst->valid(), "Clobbering existing data?");
5745 size_t n_tasks = _survivor_chunk_index + 1;
5746 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5747 // Sets the condition for completion of the subtask (how many threads
5748 // need to finish in order to be done).
5749 pst->set_n_threads(n_threads);
5750 pst->set_n_tasks((int)n_tasks);
5751 assert(pst->valid(), "Error");
5752 }
5753 }
5754
5755 // Parallel version of remark
5756 void CMSCollector::do_remark_parallel() {
5757 GenCollectedHeap* gch = GenCollectedHeap::heap();
5758 FlexibleWorkGang* workers = gch->workers();
5759 assert(workers != NULL, "Need parallel worker threads.");
5760 // Choose to use the number of GC workers most recently set
5761 // into "active_workers". If active_workers is not set, set it
5762 // to ParallelGCThreads.
5763 int n_workers = workers->active_workers();
5764 if (n_workers == 0) {
5765 assert(n_workers > 0, "Should have been set during scavenge");
5766 n_workers = ParallelGCThreads;
5767 workers->set_active_workers(n_workers);
5768 }
5769 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5770 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
5771
5772 CMSParRemarkTask tsk(this,
5773 cms_space, perm_space,
5774 n_workers, workers, task_queues());
5775
5776 // Set up for parallel process_strong_roots work.
5777 gch->set_par_threads(n_workers);
5778 // We won't be iterating over the cards in the card table updating
5779 // the younger_gen cards, so we shouldn't call the following else
5780 // the verification code as well as subsequent younger_refs_iterate
5781 // code would get confused. XXX
5782 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5783
5784 // The young gen rescan work will not be done as part of
5785 // process_strong_roots (which currently doesn't knw how to
5786 // parallelize such a scan), but rather will be broken up into
5787 // a set of parallel tasks (via the sampling that the [abortable]
5788 // preclean phase did of EdenSpace, plus the [two] tasks of
5789 // scanning the [two] survivor spaces. Further fine-grain
5790 // parallelization of the scanning of the survivor spaces
5791 // themselves, and of precleaning of the younger gen itself
5792 // is deferred to the future.
5793 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5794
5795 // The dirty card rescan work is broken up into a "sequence"
5796 // of parallel tasks (per constituent space) that are dynamically
5797 // claimed by the parallel threads.
5798 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5799 perm_space->initialize_sequential_subtasks_for_rescan(n_workers);
5800
5801 // It turns out that even when we're using 1 thread, doing the work in a
5802 // separate thread causes wide variance in run times. We can't help this
5803 // in the multi-threaded case, but we special-case n=1 here to get
5804 // repeatable measurements of the 1-thread overhead of the parallel code.
5805 if (n_workers > 1) {
5806 // Make refs discovery MT-safe, if it isn't already: it may not
5807 // necessarily be so, since it's possible that we are doing
5808 // ST marking.
5809 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5810 GenCollectedHeap::StrongRootsScope srs(gch);
5811 workers->run_task(&tsk);
5812 } else {
5813 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5814 GenCollectedHeap::StrongRootsScope srs(gch);
5815 tsk.work(0);
5816 }
5817 gch->set_par_threads(0); // 0 ==> non-parallel.
5818 // restore, single-threaded for now, any preserved marks
5819 // as a result of work_q overflow
5820 restore_preserved_marks_if_any();
5821 }
5822
5823 // Non-parallel version of remark
5824 void CMSCollector::do_remark_non_parallel() {
5825 ResourceMark rm;
5826 HandleMark hm;
5827 GenCollectedHeap* gch = GenCollectedHeap::heap();
5828 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5829
5830 MarkRefsIntoAndScanClosure
5831 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
5832 &_markStack, &_revisitStack, this,
5833 false /* should_yield */, false /* not precleaning */);
5834 MarkFromDirtyCardsClosure
5835 markFromDirtyCardsClosure(this, _span,
5836 NULL, // space is set further below
5837 &_markBitMap, &_markStack, &_revisitStack,
5838 &mrias_cl);
5839 {
5840 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm);
5841 // Iterate over the dirty cards, setting the corresponding bits in the
5842 // mod union table.
5843 {
5844 ModUnionClosure modUnionClosure(&_modUnionTable);
5845 _ct->ct_bs()->dirty_card_iterate(
5846 _cmsGen->used_region(),
5847 &modUnionClosure);
5848 _ct->ct_bs()->dirty_card_iterate(
5849 _permGen->used_region(),
5850 &modUnionClosure);
5851 }
5852 // Having transferred these marks into the modUnionTable, we just need
5853 // to rescan the marked objects on the dirty cards in the modUnionTable.
5854 // The initial marking may have been done during an asynchronous
5855 // collection so there may be dirty bits in the mod-union table.
5856 const int alignment =
5857 CardTableModRefBS::card_size * BitsPerWord;
5858 {
5859 // ... First handle dirty cards in CMS gen
5860 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5861 MemRegion ur = _cmsGen->used_region();
5862 HeapWord* lb = ur.start();
5863 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5864 MemRegion cms_span(lb, ub);
5865 _modUnionTable.dirty_range_iterate_clear(cms_span,
5866 &markFromDirtyCardsClosure);
5867 verify_work_stacks_empty();
5868 if (PrintCMSStatistics != 0) {
5869 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5870 markFromDirtyCardsClosure.num_dirty_cards());
5871 }
5872 }
5873 {
5874 // .. and then repeat for dirty cards in perm gen
5875 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace());
5876 MemRegion ur = _permGen->used_region();
5877 HeapWord* lb = ur.start();
5878 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5879 MemRegion perm_span(lb, ub);
5880 _modUnionTable.dirty_range_iterate_clear(perm_span,
5881 &markFromDirtyCardsClosure);
5882 verify_work_stacks_empty();
5883 if (PrintCMSStatistics != 0) {
5884 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ",
5885 markFromDirtyCardsClosure.num_dirty_cards());
5886 }
5887 }
5888 }
5889 if (VerifyDuringGC &&
5890 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5891 HandleMark hm; // Discard invalid handles created during verification
5892 Universe::verify();
5893 }
5894 {
5895 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm);
5896
5897 verify_work_stacks_empty();
5898
5899 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5900 GenCollectedHeap::StrongRootsScope srs(gch);
5901 gch->gen_process_strong_roots(_cmsGen->level(),
5902 true, // younger gens as roots
5903 false, // use the local StrongRootsScope
5904 true, // collecting perm gen
5905 SharedHeap::ScanningOption(roots_scanning_options()),
5906 &mrias_cl,
5907 true, // walk code active on stacks
5908 NULL);
5909 assert(should_unload_classes()
5910 || (roots_scanning_options() & SharedHeap::SO_CodeCache),
5911 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5912 }
5913 verify_work_stacks_empty();
5914 // Restore evacuated mark words, if any, used for overflow list links
5915 if (!CMSOverflowEarlyRestoration) {
5916 restore_preserved_marks_if_any();
5917 }
5918 verify_overflow_empty();
5919 }
5920
5921 ////////////////////////////////////////////////////////
5922 // Parallel Reference Processing Task Proxy Class
5923 ////////////////////////////////////////////////////////
5924 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5925 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5926 CMSCollector* _collector;
5927 CMSBitMap* _mark_bit_map;
5928 const MemRegion _span;
5929 ProcessTask& _task;
5930
5931 public:
5932 CMSRefProcTaskProxy(ProcessTask& task,
5933 CMSCollector* collector,
5934 const MemRegion& span,
5935 CMSBitMap* mark_bit_map,
5936 AbstractWorkGang* workers,
5937 OopTaskQueueSet* task_queues):
5938 // XXX Should superclass AGTWOQ also know about AWG since it knows
5939 // about the task_queues used by the AWG? Then it could initialize
5940 // the terminator() object. See 6984287. The set_for_termination()
5941 // below is a temporary band-aid for the regression in 6984287.
5942 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5943 task_queues),
5944 _task(task),
5945 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5946 {
5947 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5948 "Inconsistency in _span");
5949 set_for_termination(workers->active_workers());
5950 }
5951
5952 OopTaskQueueSet* task_queues() { return queues(); }
5953
5954 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5955
5956 void do_work_steal(int i,
5957 CMSParDrainMarkingStackClosure* drain,
5958 CMSParKeepAliveClosure* keep_alive,
5959 int* seed);
5960
5961 virtual void work(uint worker_id);
5962 };
5963
5964 void CMSRefProcTaskProxy::work(uint worker_id) {
5965 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5966 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5967 _mark_bit_map,
5968 &_collector->_revisitStack,
5969 work_queue(worker_id));
5970 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5971 _mark_bit_map,
5972 &_collector->_revisitStack,
5973 work_queue(worker_id));
5974 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5975 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5976 if (_task.marks_oops_alive()) {
5977 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5978 _collector->hash_seed(worker_id));
5979 }
5980 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5981 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5982 }
5983
5984 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5985 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5986 EnqueueTask& _task;
5987
5988 public:
5989 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5990 : AbstractGangTask("Enqueue reference objects in parallel"),
5991 _task(task)
5992 { }
5993
5994 virtual void work(uint worker_id)
5995 {
5996 _task.work(worker_id);
5997 }
5998 };
5999
6000 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
6001 MemRegion span, CMSBitMap* bit_map, CMSMarkStack* revisit_stack,
6002 OopTaskQueue* work_queue):
6003 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
6004 _span(span),
6005 _bit_map(bit_map),
6006 _work_queue(work_queue),
6007 _mark_and_push(collector, span, bit_map, revisit_stack, work_queue),
6008 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6009 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
6010 { }
6011
6012 // . see if we can share work_queues with ParNew? XXX
6013 void CMSRefProcTaskProxy::do_work_steal(int i,
6014 CMSParDrainMarkingStackClosure* drain,
6015 CMSParKeepAliveClosure* keep_alive,
6016 int* seed) {
6017 OopTaskQueue* work_q = work_queue(i);
6018 NOT_PRODUCT(int num_steals = 0;)
6019 oop obj_to_scan;
6020
6021 while (true) {
6022 // Completely finish any left over work from (an) earlier round(s)
6023 drain->trim_queue(0);
6024 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
6025 (size_t)ParGCDesiredObjsFromOverflowList);
6026 // Now check if there's any work in the overflow list
6027 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
6028 // only affects the number of attempts made to get work from the
6029 // overflow list and does not affect the number of workers. Just
6030 // pass ParallelGCThreads so this behavior is unchanged.
6031 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
6032 work_q,
6033 ParallelGCThreads)) {
6034 // Found something in global overflow list;
6035 // not yet ready to go stealing work from others.
6036 // We'd like to assert(work_q->size() != 0, ...)
6037 // because we just took work from the overflow list,
6038 // but of course we can't, since all of that might have
6039 // been already stolen from us.
6040 continue;
6041 }
6042 // Verify that we have no work before we resort to stealing
6043 assert(work_q->size() == 0, "Have work, shouldn't steal");
6044 // Try to steal from other queues that have work
6045 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
6046 NOT_PRODUCT(num_steals++;)
6047 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
6048 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
6049 // Do scanning work
6050 obj_to_scan->oop_iterate(keep_alive);
6051 // Loop around, finish this work, and try to steal some more
6052 } else if (terminator()->offer_termination()) {
6053 break; // nirvana from the infinite cycle
6054 }
6055 }
6056 NOT_PRODUCT(
6057 if (PrintCMSStatistics != 0) {
6058 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
6059 }
6060 )
6061 }
6062
6063 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
6064 {
6065 GenCollectedHeap* gch = GenCollectedHeap::heap();
6066 FlexibleWorkGang* workers = gch->workers();
6067 assert(workers != NULL, "Need parallel worker threads.");
6068 CMSRefProcTaskProxy rp_task(task, &_collector,
6069 _collector.ref_processor()->span(),
6070 _collector.markBitMap(),
6071 workers, _collector.task_queues());
6072 workers->run_task(&rp_task);
6073 }
6074
6075 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
6076 {
6077
6078 GenCollectedHeap* gch = GenCollectedHeap::heap();
6079 FlexibleWorkGang* workers = gch->workers();
6080 assert(workers != NULL, "Need parallel worker threads.");
6081 CMSRefEnqueueTaskProxy enq_task(task);
6082 workers->run_task(&enq_task);
6083 }
6084
6085 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
6086
6087 ResourceMark rm;
6088 HandleMark hm;
6089
6090 ReferenceProcessor* rp = ref_processor();
6091 assert(rp->span().equals(_span), "Spans should be equal");
6092 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
6093 // Process weak references.
6094 rp->setup_policy(clear_all_soft_refs);
6095 verify_work_stacks_empty();
6096
6097 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
6098 &_markStack, &_revisitStack,
6099 false /* !preclean */);
6100 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
6101 _span, &_markBitMap, &_markStack,
6102 &cmsKeepAliveClosure, false /* !preclean */);
6103 {
6104 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm);
6105
6106 ReferenceProcessorStats stats;
6107 if (rp->processing_is_mt()) {
6108 // Set the degree of MT here. If the discovery is done MT, there
6109 // may have been a different number of threads doing the discovery
6110 // and a different number of discovered lists may have Ref objects.
6111 // That is OK as long as the Reference lists are balanced (see
6112 // balance_all_queues() and balance_queues()).
6113 GenCollectedHeap* gch = GenCollectedHeap::heap();
6114 int active_workers = ParallelGCThreads;
6115 FlexibleWorkGang* workers = gch->workers();
6116 if (workers != NULL) {
6117 active_workers = workers->active_workers();
6118 // The expectation is that active_workers will have already
6119 // been set to a reasonable value. If it has not been set,
6120 // investigate.
6121 assert(active_workers > 0, "Should have been set during scavenge");
6122 }
6123 rp->set_active_mt_degree(active_workers);
6124 CMSRefProcTaskExecutor task_executor(*this);
6125 stats = rp->process_discovered_references(&_is_alive_closure,
6126 &cmsKeepAliveClosure,
6127 &cmsDrainMarkingStackClosure,
6128 &task_executor,
6129 _gc_timer_cm);
6130 } else {
6131 stats = rp->process_discovered_references(&_is_alive_closure,
6132 &cmsKeepAliveClosure,
6133 &cmsDrainMarkingStackClosure,
6134 NULL,
6135 _gc_timer_cm);
6136 }
6137 _gc_tracer_cm->report_gc_reference_stats(stats);
6138
6139 verify_work_stacks_empty();
6140 }
6141
6142 if (should_unload_classes()) {
6143 {
6144 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm);
6145
6146 // Follow SystemDictionary roots and unload classes
6147 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
6148
6149 // Follow CodeCache roots and unload any methods marked for unloading
6150 CodeCache::do_unloading(&_is_alive_closure,
6151 &cmsKeepAliveClosure,
6152 purged_class);
6153
6154 cmsDrainMarkingStackClosure.do_void();
6155 verify_work_stacks_empty();
6156
6157 // Update subklass/sibling/implementor links in KlassKlass descendants
6158 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty");
6159 oop k;
6160 while ((k = _revisitStack.pop()) != NULL) {
6161 ((Klass*)(oopDesc*)k)->follow_weak_klass_links(
6162 &_is_alive_closure,
6163 &cmsKeepAliveClosure);
6164 }
6165 assert(!ClassUnloading ||
6166 (_markStack.isEmpty() && overflow_list_is_empty()),
6167 "Should not have found new reachable objects");
6168 assert(_revisitStack.isEmpty(), "revisit stack should have been drained");
6169 cmsDrainMarkingStackClosure.do_void();
6170 verify_work_stacks_empty();
6171 }
6172
6173 {
6174 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm);
6175 // Clean up unreferenced symbols in symbol table.
6176 SymbolTable::unlink();
6177 }
6178 }
6179
6180 if (should_unload_classes() || !JavaObjectsInPerm) {
6181 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm);
6182 // Now clean up stale oops in StringTable
6183 StringTable::unlink(&_is_alive_closure);
6184 }
6185
6186 verify_work_stacks_empty();
6187 // Restore any preserved marks as a result of mark stack or
6188 // work queue overflow
6189 restore_preserved_marks_if_any(); // done single-threaded for now
6190
6191 rp->set_enqueuing_is_done(true);
6192 if (rp->processing_is_mt()) {
6193 rp->balance_all_queues();
6194 CMSRefProcTaskExecutor task_executor(*this);
6195 rp->enqueue_discovered_references(&task_executor);
6196 } else {
6197 rp->enqueue_discovered_references(NULL);
6198 }
6199 rp->verify_no_references_recorded();
6200 assert(!rp->discovery_enabled(), "should have been disabled");
6201 }
6202
6203 #ifndef PRODUCT
6204 void CMSCollector::check_correct_thread_executing() {
6205 Thread* t = Thread::current();
6206 // Only the VM thread or the CMS thread should be here.
6207 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
6208 "Unexpected thread type");
6209 // If this is the vm thread, the foreground process
6210 // should not be waiting. Note that _foregroundGCIsActive is
6211 // true while the foreground collector is waiting.
6212 if (_foregroundGCShouldWait) {
6213 // We cannot be the VM thread
6214 assert(t->is_ConcurrentGC_thread(),
6215 "Should be CMS thread");
6216 } else {
6217 // We can be the CMS thread only if we are in a stop-world
6218 // phase of CMS collection.
6219 if (t->is_ConcurrentGC_thread()) {
6220 assert(_collectorState == InitialMarking ||
6221 _collectorState == FinalMarking,
6222 "Should be a stop-world phase");
6223 // The CMS thread should be holding the CMS_token.
6224 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6225 "Potential interference with concurrently "
6226 "executing VM thread");
6227 }
6228 }
6229 }
6230 #endif
6231
6232 void CMSCollector::sweep(bool asynch) {
6233 assert(_collectorState == Sweeping, "just checking");
6234 check_correct_thread_executing();
6235 verify_work_stacks_empty();
6236 verify_overflow_empty();
6237 increment_sweep_count();
6238 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
6239
6240 _inter_sweep_timer.stop();
6241 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
6242 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
6243
6244 // PermGen verification support: If perm gen sweeping is disabled in
6245 // this cycle, we preserve the perm gen object "deadness" information
6246 // in the perm_gen_verify_bit_map. In order to do that we traverse
6247 // all blocks in perm gen and mark all dead objects.
6248 if (verifying() && !should_unload_classes()) {
6249 assert(perm_gen_verify_bit_map()->sizeInBits() != 0,
6250 "Should have already been allocated");
6251 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(),
6252 markBitMap(), perm_gen_verify_bit_map());
6253 if (asynch) {
6254 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
6255 bitMapLock());
6256 _permGen->cmsSpace()->blk_iterate(&mdo);
6257 } else {
6258 // In the case of synchronous sweep, we already have
6259 // the requisite locks/tokens.
6260 _permGen->cmsSpace()->blk_iterate(&mdo);
6261 }
6262 }
6263
6264 assert(!_intra_sweep_timer.is_active(), "Should not be active");
6265 _intra_sweep_timer.reset();
6266 _intra_sweep_timer.start();
6267 if (asynch) {
6268 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6269 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
6270 // First sweep the old gen then the perm gen
6271 {
6272 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6273 bitMapLock());
6274 sweepWork(_cmsGen, asynch);
6275 }
6276
6277 // Now repeat for perm gen
6278 if (should_unload_classes()) {
6279 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
6280 bitMapLock());
6281 sweepWork(_permGen, asynch);
6282 }
6283
6284 // Update Universe::_heap_*_at_gc figures.
6285 // We need all the free list locks to make the abstract state
6286 // transition from Sweeping to Resetting. See detailed note
6287 // further below.
6288 {
6289 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6290 _permGen->freelistLock());
6291 // Update heap occupancy information which is used as
6292 // input to soft ref clearing policy at the next gc.
6293 Universe::update_heap_info_at_gc();
6294 _collectorState = Resizing;
6295 }
6296 } else {
6297 // already have needed locks
6298 sweepWork(_cmsGen, asynch);
6299
6300 if (should_unload_classes()) {
6301 sweepWork(_permGen, asynch);
6302 }
6303 // Update heap occupancy information which is used as
6304 // input to soft ref clearing policy at the next gc.
6305 Universe::update_heap_info_at_gc();
6306 _collectorState = Resizing;
6307 }
6308 verify_work_stacks_empty();
6309 verify_overflow_empty();
6310
6311 _intra_sweep_timer.stop();
6312 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
6313
6314 _inter_sweep_timer.reset();
6315 _inter_sweep_timer.start();
6316
6317 // We need to use a monotonically non-deccreasing time in ms
6318 // or we will see time-warp warnings and os::javaTimeMillis()
6319 // does not guarantee monotonicity.
6320 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
6321 update_time_of_last_gc(now);
6322
6323 // NOTE on abstract state transitions:
6324 // Mutators allocate-live and/or mark the mod-union table dirty
6325 // based on the state of the collection. The former is done in
6326 // the interval [Marking, Sweeping] and the latter in the interval
6327 // [Marking, Sweeping). Thus the transitions into the Marking state
6328 // and out of the Sweeping state must be synchronously visible
6329 // globally to the mutators.
6330 // The transition into the Marking state happens with the world
6331 // stopped so the mutators will globally see it. Sweeping is
6332 // done asynchronously by the background collector so the transition
6333 // from the Sweeping state to the Resizing state must be done
6334 // under the freelistLock (as is the check for whether to
6335 // allocate-live and whether to dirty the mod-union table).
6336 assert(_collectorState == Resizing, "Change of collector state to"
6337 " Resizing must be done under the freelistLocks (plural)");
6338
6339 // Now that sweeping has been completed, we clear
6340 // the incremental_collection_failed flag,
6341 // thus inviting a younger gen collection to promote into
6342 // this generation. If such a promotion may still fail,
6343 // the flag will be set again when a young collection is
6344 // attempted.
6345 GenCollectedHeap* gch = GenCollectedHeap::heap();
6346 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
6347 gch->update_full_collections_completed(_collection_count_start);
6348 }
6349
6350 // FIX ME!!! Looks like this belongs in CFLSpace, with
6351 // CMSGen merely delegating to it.
6352 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
6353 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
6354 HeapWord* minAddr = _cmsSpace->bottom();
6355 HeapWord* largestAddr =
6356 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
6357 if (largestAddr == NULL) {
6358 // The dictionary appears to be empty. In this case
6359 // try to coalesce at the end of the heap.
6360 largestAddr = _cmsSpace->end();
6361 }
6362 size_t largestOffset = pointer_delta(largestAddr, minAddr);
6363 size_t nearLargestOffset =
6364 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
6365 if (PrintFLSStatistics != 0) {
6366 gclog_or_tty->print_cr(
6367 "CMS: Large Block: " PTR_FORMAT ";"
6368 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
6369 largestAddr,
6370 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset);
6371 }
6372 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
6373 }
6374
6375 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
6376 return addr >= _cmsSpace->nearLargestChunk();
6377 }
6378
6379 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
6380 return _cmsSpace->find_chunk_at_end();
6381 }
6382
6383 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
6384 bool full) {
6385 // The next lower level has been collected. Gather any statistics
6386 // that are of interest at this point.
6387 if (!full && (current_level + 1) == level()) {
6388 // Gather statistics on the young generation collection.
6389 collector()->stats().record_gc0_end(used());
6390 }
6391 }
6392
6393 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
6394 GenCollectedHeap* gch = GenCollectedHeap::heap();
6395 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
6396 "Wrong type of heap");
6397 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
6398 gch->gen_policy()->size_policy();
6399 assert(sp->is_gc_cms_adaptive_size_policy(),
6400 "Wrong type of size policy");
6401 return sp;
6402 }
6403
6404 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
6405 if (PrintGCDetails && Verbose) {
6406 gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
6407 }
6408 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
6409 _debug_collection_type =
6410 (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
6411 if (PrintGCDetails && Verbose) {
6412 gclog_or_tty->print_cr("to %d ", _debug_collection_type);
6413 }
6414 }
6415
6416 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
6417 bool asynch) {
6418 // We iterate over the space(s) underlying this generation,
6419 // checking the mark bit map to see if the bits corresponding
6420 // to specific blocks are marked or not. Blocks that are
6421 // marked are live and are not swept up. All remaining blocks
6422 // are swept up, with coalescing on-the-fly as we sweep up
6423 // contiguous free and/or garbage blocks:
6424 // We need to ensure that the sweeper synchronizes with allocators
6425 // and stop-the-world collectors. In particular, the following
6426 // locks are used:
6427 // . CMS token: if this is held, a stop the world collection cannot occur
6428 // . freelistLock: if this is held no allocation can occur from this
6429 // generation by another thread
6430 // . bitMapLock: if this is held, no other thread can access or update
6431 //
6432
6433 // Note that we need to hold the freelistLock if we use
6434 // block iterate below; else the iterator might go awry if
6435 // a mutator (or promotion) causes block contents to change
6436 // (for instance if the allocator divvies up a block).
6437 // If we hold the free list lock, for all practical purposes
6438 // young generation GC's can't occur (they'll usually need to
6439 // promote), so we might as well prevent all young generation
6440 // GC's while we do a sweeping step. For the same reason, we might
6441 // as well take the bit map lock for the entire duration
6442
6443 // check that we hold the requisite locks
6444 assert(have_cms_token(), "Should hold cms token");
6445 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
6446 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
6447 "Should possess CMS token to sweep");
6448 assert_lock_strong(gen->freelistLock());
6449 assert_lock_strong(bitMapLock());
6450
6451 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
6452 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
6453 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
6454 _inter_sweep_estimate.padded_average(),
6455 _intra_sweep_estimate.padded_average());
6456 gen->setNearLargestChunk();
6457
6458 {
6459 SweepClosure sweepClosure(this, gen, &_markBitMap,
6460 CMSYield && asynch);
6461 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
6462 // We need to free-up/coalesce garbage/blocks from a
6463 // co-terminal free run. This is done in the SweepClosure
6464 // destructor; so, do not remove this scope, else the
6465 // end-of-sweep-census below will be off by a little bit.
6466 }
6467 gen->cmsSpace()->sweep_completed();
6468 gen->cmsSpace()->endSweepFLCensus(sweep_count());
6469 if (should_unload_classes()) { // unloaded classes this cycle,
6470 _concurrent_cycles_since_last_unload = 0; // ... reset count
6471 } else { // did not unload classes,
6472 _concurrent_cycles_since_last_unload++; // ... increment count
6473 }
6474 }
6475
6476 // Reset CMS data structures (for now just the marking bit map)
6477 // preparatory for the next cycle.
6478 void CMSCollector::reset(bool asynch) {
6479 GenCollectedHeap* gch = GenCollectedHeap::heap();
6480 CMSAdaptiveSizePolicy* sp = size_policy();
6481 AdaptiveSizePolicyOutput(sp, gch->total_collections());
6482 if (asynch) {
6483 CMSTokenSyncWithLocks ts(true, bitMapLock());
6484
6485 // If the state is not "Resetting", the foreground thread
6486 // has done a collection and the resetting.
6487 if (_collectorState != Resetting) {
6488 assert(_collectorState == Idling, "The state should only change"
6489 " because the foreground collector has finished the collection");
6490 return;
6491 }
6492
6493 // Clear the mark bitmap (no grey objects to start with)
6494 // for the next cycle.
6495 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6496 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
6497
6498 HeapWord* curAddr = _markBitMap.startWord();
6499 while (curAddr < _markBitMap.endWord()) {
6500 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
6501 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
6502 _markBitMap.clear_large_range(chunk);
6503 if (ConcurrentMarkSweepThread::should_yield() &&
6504 !foregroundGCIsActive() &&
6505 CMSYield) {
6506 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6507 "CMS thread should hold CMS token");
6508 assert_lock_strong(bitMapLock());
6509 bitMapLock()->unlock();
6510 ConcurrentMarkSweepThread::desynchronize(true);
6511 ConcurrentMarkSweepThread::acknowledge_yield_request();
6512 stopTimer();
6513 if (PrintCMSStatistics != 0) {
6514 incrementYields();
6515 }
6516 icms_wait();
6517
6518 // See the comment in coordinator_yield()
6519 for (unsigned i = 0; i < CMSYieldSleepCount &&
6520 ConcurrentMarkSweepThread::should_yield() &&
6521 !CMSCollector::foregroundGCIsActive(); ++i) {
6522 os::sleep(Thread::current(), 1, false);
6523 ConcurrentMarkSweepThread::acknowledge_yield_request();
6524 }
6525
6526 ConcurrentMarkSweepThread::synchronize(true);
6527 bitMapLock()->lock_without_safepoint_check();
6528 startTimer();
6529 }
6530 curAddr = chunk.end();
6531 }
6532 // A successful mostly concurrent collection has been done.
6533 // Because only the full (i.e., concurrent mode failure) collections
6534 // are being measured for gc overhead limits, clean the "near" flag
6535 // and count.
6536 sp->reset_gc_overhead_limit_count();
6537 _collectorState = Idling;
6538 } else {
6539 // already have the lock
6540 assert(_collectorState == Resetting, "just checking");
6541 assert_lock_strong(bitMapLock());
6542 _markBitMap.clear_all();
6543 _collectorState = Idling;
6544 }
6545
6546 // Stop incremental mode after a cycle completes, so that any future cycles
6547 // are triggered by allocation.
6548 stop_icms();
6549
6550 NOT_PRODUCT(
6551 if (RotateCMSCollectionTypes) {
6552 _cmsGen->rotate_debug_collection_type();
6553 }
6554 )
6555
6556 register_gc_end();
6557 }
6558
6559 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
6560 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6561 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL);
6562 TraceCollectorStats tcs(counters());
6563
6564 switch (op) {
6565 case CMS_op_checkpointRootsInitial: {
6566 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6567 checkpointRootsInitial(true); // asynch
6568 if (PrintGC) {
6569 _cmsGen->printOccupancy("initial-mark");
6570 }
6571 break;
6572 }
6573 case CMS_op_checkpointRootsFinal: {
6574 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6575 checkpointRootsFinal(true, // asynch
6576 false, // !clear_all_soft_refs
6577 false); // !init_mark_was_synchronous
6578 if (PrintGC) {
6579 _cmsGen->printOccupancy("remark");
6580 }
6581 break;
6582 }
6583 default:
6584 fatal("No such CMS_op");
6585 }
6586 }
6587
6588 #ifndef PRODUCT
6589 size_t const CMSCollector::skip_header_HeapWords() {
6590 return FreeChunk::header_size();
6591 }
6592
6593 // Try and collect here conditions that should hold when
6594 // CMS thread is exiting. The idea is that the foreground GC
6595 // thread should not be blocked if it wants to terminate
6596 // the CMS thread and yet continue to run the VM for a while
6597 // after that.
6598 void CMSCollector::verify_ok_to_terminate() const {
6599 assert(Thread::current()->is_ConcurrentGC_thread(),
6600 "should be called by CMS thread");
6601 assert(!_foregroundGCShouldWait, "should be false");
6602 // We could check here that all the various low-level locks
6603 // are not held by the CMS thread, but that is overkill; see
6604 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6605 // is checked.
6606 }
6607 #endif
6608
6609 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6610 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6611 "missing Printezis mark?");
6612 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6613 size_t size = pointer_delta(nextOneAddr + 1, addr);
6614 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6615 "alignment problem");
6616 assert(size >= 3, "Necessary for Printezis marks to work");
6617 return size;
6618 }
6619
6620 // A variant of the above (block_size_using_printezis_bits()) except
6621 // that we return 0 if the P-bits are not yet set.
6622 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6623 if (_markBitMap.isMarked(addr + 1)) {
6624 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
6625 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6626 size_t size = pointer_delta(nextOneAddr + 1, addr);
6627 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6628 "alignment problem");
6629 assert(size >= 3, "Necessary for Printezis marks to work");
6630 return size;
6631 }
6632 return 0;
6633 }
6634
6635 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6636 size_t sz = 0;
6637 oop p = (oop)addr;
6638 if (p->klass_or_null() != NULL && p->is_parsable()) {
6639 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6640 } else {
6641 sz = block_size_using_printezis_bits(addr);
6642 }
6643 assert(sz > 0, "size must be nonzero");
6644 HeapWord* next_block = addr + sz;
6645 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
6646 CardTableModRefBS::card_size);
6647 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
6648 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6649 "must be different cards");
6650 return next_card;
6651 }
6652
6653
6654 // CMS Bit Map Wrapper /////////////////////////////////////////
6655
6656 // Construct a CMS bit map infrastructure, but don't create the
6657 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6658 // further below.
6659 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6660 _bm(),
6661 _shifter(shifter),
6662 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6663 {
6664 _bmStartWord = 0;
6665 _bmWordSize = 0;
6666 }
6667
6668 bool CMSBitMap::allocate(MemRegion mr) {
6669 _bmStartWord = mr.start();
6670 _bmWordSize = mr.word_size();
6671 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6672 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6673 if (!brs.is_reserved()) {
6674 warning("CMS bit map allocation failure");
6675 return false;
6676 }
6677 // For now we'll just commit all of the bit map up fromt.
6678 // Later on we'll try to be more parsimonious with swap.
6679 if (!_virtual_space.initialize(brs, brs.size())) {
6680 warning("CMS bit map backing store failure");
6681 return false;
6682 }
6683 assert(_virtual_space.committed_size() == brs.size(),
6684 "didn't reserve backing store for all of CMS bit map?");
6685 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
6686 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6687 _bmWordSize, "inconsistency in bit map sizing");
6688 _bm.set_size(_bmWordSize >> _shifter);
6689
6690 // bm.clear(); // can we rely on getting zero'd memory? verify below
6691 assert(isAllClear(),
6692 "Expected zero'd memory from ReservedSpace constructor");
6693 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6694 "consistency check");
6695 return true;
6696 }
6697
6698 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6699 HeapWord *next_addr, *end_addr, *last_addr;
6700 assert_locked();
6701 assert(covers(mr), "out-of-range error");
6702 // XXX assert that start and end are appropriately aligned
6703 for (next_addr = mr.start(), end_addr = mr.end();
6704 next_addr < end_addr; next_addr = last_addr) {
6705 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6706 last_addr = dirty_region.end();
6707 if (!dirty_region.is_empty()) {
6708 cl->do_MemRegion(dirty_region);
6709 } else {
6710 assert(last_addr == end_addr, "program logic");
6711 return;
6712 }
6713 }
6714 }
6715
6716 #ifndef PRODUCT
6717 void CMSBitMap::assert_locked() const {
6718 CMSLockVerifier::assert_locked(lock());
6719 }
6720
6721 bool CMSBitMap::covers(MemRegion mr) const {
6722 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6723 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6724 "size inconsistency");
6725 return (mr.start() >= _bmStartWord) &&
6726 (mr.end() <= endWord());
6727 }
6728
6729 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6730 return (start >= _bmStartWord && (start + size) <= endWord());
6731 }
6732
6733 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6734 // verify that there are no 1 bits in the interval [left, right)
6735 FalseBitMapClosure falseBitMapClosure;
6736 iterate(&falseBitMapClosure, left, right);
6737 }
6738
6739 void CMSBitMap::region_invariant(MemRegion mr)
6740 {
6741 assert_locked();
6742 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6743 assert(!mr.is_empty(), "unexpected empty region");
6744 assert(covers(mr), "mr should be covered by bit map");
6745 // convert address range into offset range
6746 size_t start_ofs = heapWordToOffset(mr.start());
6747 // Make sure that end() is appropriately aligned
6748 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6749 (1 << (_shifter+LogHeapWordSize))),
6750 "Misaligned mr.end()");
6751 size_t end_ofs = heapWordToOffset(mr.end());
6752 assert(end_ofs > start_ofs, "Should mark at least one bit");
6753 }
6754
6755 #endif
6756
6757 bool CMSMarkStack::allocate(size_t size) {
6758 // allocate a stack of the requisite depth
6759 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6760 size * sizeof(oop)));
6761 if (!rs.is_reserved()) {
6762 warning("CMSMarkStack allocation failure");
6763 return false;
6764 }
6765 if (!_virtual_space.initialize(rs, rs.size())) {
6766 warning("CMSMarkStack backing store failure");
6767 return false;
6768 }
6769 assert(_virtual_space.committed_size() == rs.size(),
6770 "didn't reserve backing store for all of CMS stack?");
6771 _base = (oop*)(_virtual_space.low());
6772 _index = 0;
6773 _capacity = size;
6774 NOT_PRODUCT(_max_depth = 0);
6775 return true;
6776 }
6777
6778 // XXX FIX ME !!! In the MT case we come in here holding a
6779 // leaf lock. For printing we need to take a further lock
6780 // which has lower rank. We need to recallibrate the two
6781 // lock-ranks involved in order to be able to rpint the
6782 // messages below. (Or defer the printing to the caller.
6783 // For now we take the expedient path of just disabling the
6784 // messages for the problematic case.)
6785 void CMSMarkStack::expand() {
6786 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6787 if (_capacity == MarkStackSizeMax) {
6788 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6789 // We print a warning message only once per CMS cycle.
6790 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6791 }
6792 return;
6793 }
6794 // Double capacity if possible
6795 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6796 // Do not give up existing stack until we have managed to
6797 // get the double capacity that we desired.
6798 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6799 new_capacity * sizeof(oop)));
6800 if (rs.is_reserved()) {
6801 // Release the backing store associated with old stack
6802 _virtual_space.release();
6803 // Reinitialize virtual space for new stack
6804 if (!_virtual_space.initialize(rs, rs.size())) {
6805 fatal("Not enough swap for expanded marking stack");
6806 }
6807 _base = (oop*)(_virtual_space.low());
6808 _index = 0;
6809 _capacity = new_capacity;
6810 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6811 // Failed to double capacity, continue;
6812 // we print a detail message only once per CMS cycle.
6813 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6814 SIZE_FORMAT"K",
6815 _capacity / K, new_capacity / K);
6816 }
6817 }
6818
6819
6820 // Closures
6821 // XXX: there seems to be a lot of code duplication here;
6822 // should refactor and consolidate common code.
6823
6824 // This closure is used to mark refs into the CMS generation in
6825 // the CMS bit map. Called at the first checkpoint. This closure
6826 // assumes that we do not need to re-mark dirty cards; if the CMS
6827 // generation on which this is used is not an oldest (modulo perm gen)
6828 // generation then this will lose younger_gen cards!
6829
6830 MarkRefsIntoClosure::MarkRefsIntoClosure(
6831 MemRegion span, CMSBitMap* bitMap):
6832 _span(span),
6833 _bitMap(bitMap)
6834 {
6835 assert(_ref_processor == NULL, "deliberately left NULL");
6836 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6837 }
6838
6839 void MarkRefsIntoClosure::do_oop(oop obj) {
6840 // if p points into _span, then mark corresponding bit in _markBitMap
6841 assert(obj->is_oop(), "expected an oop");
6842 HeapWord* addr = (HeapWord*)obj;
6843 if (_span.contains(addr)) {
6844 // this should be made more efficient
6845 _bitMap->mark(addr);
6846 }
6847 }
6848
6849 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6850 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6851
6852 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
6853 MemRegion span, CMSBitMap* bitMap):
6854 _span(span),
6855 _bitMap(bitMap)
6856 {
6857 assert(_ref_processor == NULL, "deliberately left NULL");
6858 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6859 }
6860
6861 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
6862 // if p points into _span, then mark corresponding bit in _markBitMap
6863 assert(obj->is_oop(), "expected an oop");
6864 HeapWord* addr = (HeapWord*)obj;
6865 if (_span.contains(addr)) {
6866 // this should be made more efficient
6867 _bitMap->par_mark(addr);
6868 }
6869 }
6870
6871 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6872 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6873
6874 // A variant of the above, used for CMS marking verification.
6875 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6876 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6877 _span(span),
6878 _verification_bm(verification_bm),
6879 _cms_bm(cms_bm)
6880 {
6881 assert(_ref_processor == NULL, "deliberately left NULL");
6882 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6883 }
6884
6885 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6886 // if p points into _span, then mark corresponding bit in _markBitMap
6887 assert(obj->is_oop(), "expected an oop");
6888 HeapWord* addr = (HeapWord*)obj;
6889 if (_span.contains(addr)) {
6890 _verification_bm->mark(addr);
6891 if (!_cms_bm->isMarked(addr)) {
6892 oop(addr)->print();
6893 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6894 fatal("... aborting");
6895 }
6896 }
6897 }
6898
6899 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6900 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6901
6902 //////////////////////////////////////////////////
6903 // MarkRefsIntoAndScanClosure
6904 //////////////////////////////////////////////////
6905
6906 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6907 ReferenceProcessor* rp,
6908 CMSBitMap* bit_map,
6909 CMSBitMap* mod_union_table,
6910 CMSMarkStack* mark_stack,
6911 CMSMarkStack* revisit_stack,
6912 CMSCollector* collector,
6913 bool should_yield,
6914 bool concurrent_precleaning):
6915 _collector(collector),
6916 _span(span),
6917 _bit_map(bit_map),
6918 _mark_stack(mark_stack),
6919 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6920 mark_stack, revisit_stack, concurrent_precleaning),
6921 _yield(should_yield),
6922 _concurrent_precleaning(concurrent_precleaning),
6923 _freelistLock(NULL)
6924 {
6925 _ref_processor = rp;
6926 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6927 }
6928
6929 // This closure is used to mark refs into the CMS generation at the
6930 // second (final) checkpoint, and to scan and transitively follow
6931 // the unmarked oops. It is also used during the concurrent precleaning
6932 // phase while scanning objects on dirty cards in the CMS generation.
6933 // The marks are made in the marking bit map and the marking stack is
6934 // used for keeping the (newly) grey objects during the scan.
6935 // The parallel version (Par_...) appears further below.
6936 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6937 if (obj != NULL) {
6938 assert(obj->is_oop(), "expected an oop");
6939 HeapWord* addr = (HeapWord*)obj;
6940 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6941 assert(_collector->overflow_list_is_empty(),
6942 "overflow list should be empty");
6943 if (_span.contains(addr) &&
6944 !_bit_map->isMarked(addr)) {
6945 // mark bit map (object is now grey)
6946 _bit_map->mark(addr);
6947 // push on marking stack (stack should be empty), and drain the
6948 // stack by applying this closure to the oops in the oops popped
6949 // from the stack (i.e. blacken the grey objects)
6950 bool res = _mark_stack->push(obj);
6951 assert(res, "Should have space to push on empty stack");
6952 do {
6953 oop new_oop = _mark_stack->pop();
6954 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6955 assert(new_oop->is_parsable(), "Found unparsable oop");
6956 assert(_bit_map->isMarked((HeapWord*)new_oop),
6957 "only grey objects on this stack");
6958 // iterate over the oops in this oop, marking and pushing
6959 // the ones in CMS heap (i.e. in _span).
6960 new_oop->oop_iterate(&_pushAndMarkClosure);
6961 // check if it's time to yield
6962 do_yield_check();
6963 } while (!_mark_stack->isEmpty() ||
6964 (!_concurrent_precleaning && take_from_overflow_list()));
6965 // if marking stack is empty, and we are not doing this
6966 // during precleaning, then check the overflow list
6967 }
6968 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6969 assert(_collector->overflow_list_is_empty(),
6970 "overflow list was drained above");
6971 // We could restore evacuated mark words, if any, used for
6972 // overflow list links here because the overflow list is
6973 // provably empty here. That would reduce the maximum
6974 // size requirements for preserved_{oop,mark}_stack.
6975 // But we'll just postpone it until we are all done
6976 // so we can just stream through.
6977 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6978 _collector->restore_preserved_marks_if_any();
6979 assert(_collector->no_preserved_marks(), "No preserved marks");
6980 }
6981 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6982 "All preserved marks should have been restored above");
6983 }
6984 }
6985
6986 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6987 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6988
6989 void MarkRefsIntoAndScanClosure::do_yield_work() {
6990 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6991 "CMS thread should hold CMS token");
6992 assert_lock_strong(_freelistLock);
6993 assert_lock_strong(_bit_map->lock());
6994 // relinquish the free_list_lock and bitMaplock()
6995 DEBUG_ONLY(RememberKlassesChecker mux(false);)
6996 _bit_map->lock()->unlock();
6997 _freelistLock->unlock();
6998 ConcurrentMarkSweepThread::desynchronize(true);
6999 ConcurrentMarkSweepThread::acknowledge_yield_request();
7000 _collector->stopTimer();
7001 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7002 if (PrintCMSStatistics != 0) {
7003 _collector->incrementYields();
7004 }
7005 _collector->icms_wait();
7006
7007 // See the comment in coordinator_yield()
7008 for (unsigned i = 0;
7009 i < CMSYieldSleepCount &&
7010 ConcurrentMarkSweepThread::should_yield() &&
7011 !CMSCollector::foregroundGCIsActive();
7012 ++i) {
7013 os::sleep(Thread::current(), 1, false);
7014 ConcurrentMarkSweepThread::acknowledge_yield_request();
7015 }
7016
7017 ConcurrentMarkSweepThread::synchronize(true);
7018 _freelistLock->lock_without_safepoint_check();
7019 _bit_map->lock()->lock_without_safepoint_check();
7020 _collector->startTimer();
7021 }
7022
7023 ///////////////////////////////////////////////////////////
7024 // Par_MarkRefsIntoAndScanClosure: a parallel version of
7025 // MarkRefsIntoAndScanClosure
7026 ///////////////////////////////////////////////////////////
7027 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
7028 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
7029 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack):
7030 _span(span),
7031 _bit_map(bit_map),
7032 _work_queue(work_queue),
7033 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
7034 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
7035 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue,
7036 revisit_stack)
7037 {
7038 _ref_processor = rp;
7039 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7040 }
7041
7042 // This closure is used to mark refs into the CMS generation at the
7043 // second (final) checkpoint, and to scan and transitively follow
7044 // the unmarked oops. The marks are made in the marking bit map and
7045 // the work_queue is used for keeping the (newly) grey objects during
7046 // the scan phase whence they are also available for stealing by parallel
7047 // threads. Since the marking bit map is shared, updates are
7048 // synchronized (via CAS).
7049 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
7050 if (obj != NULL) {
7051 // Ignore mark word because this could be an already marked oop
7052 // that may be chained at the end of the overflow list.
7053 assert(obj->is_oop(true), "expected an oop");
7054 HeapWord* addr = (HeapWord*)obj;
7055 if (_span.contains(addr) &&
7056 !_bit_map->isMarked(addr)) {
7057 // mark bit map (object will become grey):
7058 // It is possible for several threads to be
7059 // trying to "claim" this object concurrently;
7060 // the unique thread that succeeds in marking the
7061 // object first will do the subsequent push on
7062 // to the work queue (or overflow list).
7063 if (_bit_map->par_mark(addr)) {
7064 // push on work_queue (which may not be empty), and trim the
7065 // queue to an appropriate length by applying this closure to
7066 // the oops in the oops popped from the stack (i.e. blacken the
7067 // grey objects)
7068 bool res = _work_queue->push(obj);
7069 assert(res, "Low water mark should be less than capacity?");
7070 trim_queue(_low_water_mark);
7071 } // Else, another thread claimed the object
7072 }
7073 }
7074 }
7075
7076 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
7077 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
7078
7079 // This closure is used to rescan the marked objects on the dirty cards
7080 // in the mod union table and the card table proper.
7081 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
7082 oop p, MemRegion mr) {
7083
7084 size_t size = 0;
7085 HeapWord* addr = (HeapWord*)p;
7086 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7087 assert(_span.contains(addr), "we are scanning the CMS generation");
7088 // check if it's time to yield
7089 if (do_yield_check()) {
7090 // We yielded for some foreground stop-world work,
7091 // and we have been asked to abort this ongoing preclean cycle.
7092 return 0;
7093 }
7094 if (_bitMap->isMarked(addr)) {
7095 // it's marked; is it potentially uninitialized?
7096 if (p->klass_or_null() != NULL) {
7097 // If is_conc_safe is false, the object may be undergoing
7098 // change by the VM outside a safepoint. Don't try to
7099 // scan it, but rather leave it for the remark phase.
7100 if (CMSPermGenPrecleaningEnabled &&
7101 (!p->is_conc_safe() || !p->is_parsable())) {
7102 // Signal precleaning to redirty the card since
7103 // the klass pointer is already installed.
7104 assert(size == 0, "Initial value");
7105 } else {
7106 assert(p->is_parsable(), "must be parsable.");
7107 // an initialized object; ignore mark word in verification below
7108 // since we are running concurrent with mutators
7109 assert(p->is_oop(true), "should be an oop");
7110 if (p->is_objArray()) {
7111 // objArrays are precisely marked; restrict scanning
7112 // to dirty cards only.
7113 size = CompactibleFreeListSpace::adjustObjectSize(
7114 p->oop_iterate(_scanningClosure, mr));
7115 } else {
7116 // A non-array may have been imprecisely marked; we need
7117 // to scan object in its entirety.
7118 size = CompactibleFreeListSpace::adjustObjectSize(
7119 p->oop_iterate(_scanningClosure));
7120 }
7121 #ifdef DEBUG
7122 size_t direct_size =
7123 CompactibleFreeListSpace::adjustObjectSize(p->size());
7124 assert(size == direct_size, "Inconsistency in size");
7125 assert(size >= 3, "Necessary for Printezis marks to work");
7126 if (!_bitMap->isMarked(addr+1)) {
7127 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
7128 } else {
7129 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
7130 assert(_bitMap->isMarked(addr+size-1),
7131 "inconsistent Printezis mark");
7132 }
7133 #endif // DEBUG
7134 }
7135 } else {
7136 // an unitialized object
7137 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
7138 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7139 size = pointer_delta(nextOneAddr + 1, addr);
7140 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7141 "alignment problem");
7142 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
7143 // will dirty the card when the klass pointer is installed in the
7144 // object (signalling the completion of initialization).
7145 }
7146 } else {
7147 // Either a not yet marked object or an uninitialized object
7148 if (p->klass_or_null() == NULL || !p->is_parsable()) {
7149 // An uninitialized object, skip to the next card, since
7150 // we may not be able to read its P-bits yet.
7151 assert(size == 0, "Initial value");
7152 } else {
7153 // An object not (yet) reached by marking: we merely need to
7154 // compute its size so as to go look at the next block.
7155 assert(p->is_oop(true), "should be an oop");
7156 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
7157 }
7158 }
7159 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7160 return size;
7161 }
7162
7163 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
7164 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7165 "CMS thread should hold CMS token");
7166 assert_lock_strong(_freelistLock);
7167 assert_lock_strong(_bitMap->lock());
7168 DEBUG_ONLY(RememberKlassesChecker mux(false);)
7169 // relinquish the free_list_lock and bitMaplock()
7170 _bitMap->lock()->unlock();
7171 _freelistLock->unlock();
7172 ConcurrentMarkSweepThread::desynchronize(true);
7173 ConcurrentMarkSweepThread::acknowledge_yield_request();
7174 _collector->stopTimer();
7175 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7176 if (PrintCMSStatistics != 0) {
7177 _collector->incrementYields();
7178 }
7179 _collector->icms_wait();
7180
7181 // See the comment in coordinator_yield()
7182 for (unsigned i = 0; i < CMSYieldSleepCount &&
7183 ConcurrentMarkSweepThread::should_yield() &&
7184 !CMSCollector::foregroundGCIsActive(); ++i) {
7185 os::sleep(Thread::current(), 1, false);
7186 ConcurrentMarkSweepThread::acknowledge_yield_request();
7187 }
7188
7189 ConcurrentMarkSweepThread::synchronize(true);
7190 _freelistLock->lock_without_safepoint_check();
7191 _bitMap->lock()->lock_without_safepoint_check();
7192 _collector->startTimer();
7193 }
7194
7195
7196 //////////////////////////////////////////////////////////////////
7197 // SurvivorSpacePrecleanClosure
7198 //////////////////////////////////////////////////////////////////
7199 // This (single-threaded) closure is used to preclean the oops in
7200 // the survivor spaces.
7201 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
7202
7203 HeapWord* addr = (HeapWord*)p;
7204 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7205 assert(!_span.contains(addr), "we are scanning the survivor spaces");
7206 assert(p->klass_or_null() != NULL, "object should be initializd");
7207 assert(p->is_parsable(), "must be parsable.");
7208 // an initialized object; ignore mark word in verification below
7209 // since we are running concurrent with mutators
7210 assert(p->is_oop(true), "should be an oop");
7211 // Note that we do not yield while we iterate over
7212 // the interior oops of p, pushing the relevant ones
7213 // on our marking stack.
7214 size_t size = p->oop_iterate(_scanning_closure);
7215 do_yield_check();
7216 // Observe that below, we do not abandon the preclean
7217 // phase as soon as we should; rather we empty the
7218 // marking stack before returning. This is to satisfy
7219 // some existing assertions. In general, it may be a
7220 // good idea to abort immediately and complete the marking
7221 // from the grey objects at a later time.
7222 while (!_mark_stack->isEmpty()) {
7223 oop new_oop = _mark_stack->pop();
7224 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
7225 assert(new_oop->is_parsable(), "Found unparsable oop");
7226 assert(_bit_map->isMarked((HeapWord*)new_oop),
7227 "only grey objects on this stack");
7228 // iterate over the oops in this oop, marking and pushing
7229 // the ones in CMS heap (i.e. in _span).
7230 new_oop->oop_iterate(_scanning_closure);
7231 // check if it's time to yield
7232 do_yield_check();
7233 }
7234 unsigned int after_count =
7235 GenCollectedHeap::heap()->total_collections();
7236 bool abort = (_before_count != after_count) ||
7237 _collector->should_abort_preclean();
7238 return abort ? 0 : size;
7239 }
7240
7241 void SurvivorSpacePrecleanClosure::do_yield_work() {
7242 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7243 "CMS thread should hold CMS token");
7244 assert_lock_strong(_bit_map->lock());
7245 DEBUG_ONLY(RememberKlassesChecker smx(false);)
7246 // Relinquish the bit map lock
7247 _bit_map->lock()->unlock();
7248 ConcurrentMarkSweepThread::desynchronize(true);
7249 ConcurrentMarkSweepThread::acknowledge_yield_request();
7250 _collector->stopTimer();
7251 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7252 if (PrintCMSStatistics != 0) {
7253 _collector->incrementYields();
7254 }
7255 _collector->icms_wait();
7256
7257 // See the comment in coordinator_yield()
7258 for (unsigned i = 0; i < CMSYieldSleepCount &&
7259 ConcurrentMarkSweepThread::should_yield() &&
7260 !CMSCollector::foregroundGCIsActive(); ++i) {
7261 os::sleep(Thread::current(), 1, false);
7262 ConcurrentMarkSweepThread::acknowledge_yield_request();
7263 }
7264
7265 ConcurrentMarkSweepThread::synchronize(true);
7266 _bit_map->lock()->lock_without_safepoint_check();
7267 _collector->startTimer();
7268 }
7269
7270 // This closure is used to rescan the marked objects on the dirty cards
7271 // in the mod union table and the card table proper. In the parallel
7272 // case, although the bitMap is shared, we do a single read so the
7273 // isMarked() query is "safe".
7274 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
7275 // Ignore mark word because we are running concurrent with mutators
7276 assert(p->is_oop_or_null(true), "expected an oop or null");
7277 HeapWord* addr = (HeapWord*)p;
7278 assert(_span.contains(addr), "we are scanning the CMS generation");
7279 bool is_obj_array = false;
7280 #ifdef DEBUG
7281 if (!_parallel) {
7282 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
7283 assert(_collector->overflow_list_is_empty(),
7284 "overflow list should be empty");
7285
7286 }
7287 #endif // DEBUG
7288 if (_bit_map->isMarked(addr)) {
7289 // Obj arrays are precisely marked, non-arrays are not;
7290 // so we scan objArrays precisely and non-arrays in their
7291 // entirety.
7292 if (p->is_objArray()) {
7293 is_obj_array = true;
7294 if (_parallel) {
7295 p->oop_iterate(_par_scan_closure, mr);
7296 } else {
7297 p->oop_iterate(_scan_closure, mr);
7298 }
7299 } else {
7300 if (_parallel) {
7301 p->oop_iterate(_par_scan_closure);
7302 } else {
7303 p->oop_iterate(_scan_closure);
7304 }
7305 }
7306 }
7307 #ifdef DEBUG
7308 if (!_parallel) {
7309 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
7310 assert(_collector->overflow_list_is_empty(),
7311 "overflow list should be empty");
7312
7313 }
7314 #endif // DEBUG
7315 return is_obj_array;
7316 }
7317
7318 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
7319 MemRegion span,
7320 CMSBitMap* bitMap, CMSMarkStack* markStack,
7321 CMSMarkStack* revisitStack,
7322 bool should_yield, bool verifying):
7323 _collector(collector),
7324 _span(span),
7325 _bitMap(bitMap),
7326 _mut(&collector->_modUnionTable),
7327 _markStack(markStack),
7328 _revisitStack(revisitStack),
7329 _yield(should_yield),
7330 _skipBits(0)
7331 {
7332 assert(_markStack->isEmpty(), "stack should be empty");
7333 _finger = _bitMap->startWord();
7334 _threshold = _finger;
7335 assert(_collector->_restart_addr == NULL, "Sanity check");
7336 assert(_span.contains(_finger), "Out of bounds _finger?");
7337 DEBUG_ONLY(_verifying = verifying;)
7338 }
7339
7340 void MarkFromRootsClosure::reset(HeapWord* addr) {
7341 assert(_markStack->isEmpty(), "would cause duplicates on stack");
7342 assert(_span.contains(addr), "Out of bounds _finger?");
7343 _finger = addr;
7344 _threshold = (HeapWord*)round_to(
7345 (intptr_t)_finger, CardTableModRefBS::card_size);
7346 }
7347
7348 // Should revisit to see if this should be restructured for
7349 // greater efficiency.
7350 bool MarkFromRootsClosure::do_bit(size_t offset) {
7351 if (_skipBits > 0) {
7352 _skipBits--;
7353 return true;
7354 }
7355 // convert offset into a HeapWord*
7356 HeapWord* addr = _bitMap->startWord() + offset;
7357 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
7358 "address out of range");
7359 assert(_bitMap->isMarked(addr), "tautology");
7360 if (_bitMap->isMarked(addr+1)) {
7361 // this is an allocated but not yet initialized object
7362 assert(_skipBits == 0, "tautology");
7363 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
7364 oop p = oop(addr);
7365 if (p->klass_or_null() == NULL || !p->is_parsable()) {
7366 DEBUG_ONLY(if (!_verifying) {)
7367 // We re-dirty the cards on which this object lies and increase
7368 // the _threshold so that we'll come back to scan this object
7369 // during the preclean or remark phase. (CMSCleanOnEnter)
7370 if (CMSCleanOnEnter) {
7371 size_t sz = _collector->block_size_using_printezis_bits(addr);
7372 HeapWord* end_card_addr = (HeapWord*)round_to(
7373 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7374 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7375 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7376 // Bump _threshold to end_card_addr; note that
7377 // _threshold cannot possibly exceed end_card_addr, anyhow.
7378 // This prevents future clearing of the card as the scan proceeds
7379 // to the right.
7380 assert(_threshold <= end_card_addr,
7381 "Because we are just scanning into this object");
7382 if (_threshold < end_card_addr) {
7383 _threshold = end_card_addr;
7384 }
7385 if (p->klass_or_null() != NULL) {
7386 // Redirty the range of cards...
7387 _mut->mark_range(redirty_range);
7388 } // ...else the setting of klass will dirty the card anyway.
7389 }
7390 DEBUG_ONLY(})
7391 return true;
7392 }
7393 }
7394 scanOopsInOop(addr);
7395 return true;
7396 }
7397
7398 // We take a break if we've been at this for a while,
7399 // so as to avoid monopolizing the locks involved.
7400 void MarkFromRootsClosure::do_yield_work() {
7401 // First give up the locks, then yield, then re-lock
7402 // We should probably use a constructor/destructor idiom to
7403 // do this unlock/lock or modify the MutexUnlocker class to
7404 // serve our purpose. XXX
7405 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7406 "CMS thread should hold CMS token");
7407 assert_lock_strong(_bitMap->lock());
7408 DEBUG_ONLY(RememberKlassesChecker mux(false);)
7409 _bitMap->lock()->unlock();
7410 ConcurrentMarkSweepThread::desynchronize(true);
7411 ConcurrentMarkSweepThread::acknowledge_yield_request();
7412 _collector->stopTimer();
7413 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7414 if (PrintCMSStatistics != 0) {
7415 _collector->incrementYields();
7416 }
7417 _collector->icms_wait();
7418
7419 // See the comment in coordinator_yield()
7420 for (unsigned i = 0; i < CMSYieldSleepCount &&
7421 ConcurrentMarkSweepThread::should_yield() &&
7422 !CMSCollector::foregroundGCIsActive(); ++i) {
7423 os::sleep(Thread::current(), 1, false);
7424 ConcurrentMarkSweepThread::acknowledge_yield_request();
7425 }
7426
7427 ConcurrentMarkSweepThread::synchronize(true);
7428 _bitMap->lock()->lock_without_safepoint_check();
7429 _collector->startTimer();
7430 }
7431
7432 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
7433 assert(_bitMap->isMarked(ptr), "expected bit to be set");
7434 assert(_markStack->isEmpty(),
7435 "should drain stack to limit stack usage");
7436 // convert ptr to an oop preparatory to scanning
7437 oop obj = oop(ptr);
7438 // Ignore mark word in verification below, since we
7439 // may be running concurrent with mutators.
7440 assert(obj->is_oop(true), "should be an oop");
7441 assert(_finger <= ptr, "_finger runneth ahead");
7442 // advance the finger to right end of this object
7443 _finger = ptr + obj->size();
7444 assert(_finger > ptr, "we just incremented it above");
7445 // On large heaps, it may take us some time to get through
7446 // the marking phase (especially if running iCMS). During
7447 // this time it's possible that a lot of mutations have
7448 // accumulated in the card table and the mod union table --
7449 // these mutation records are redundant until we have
7450 // actually traced into the corresponding card.
7451 // Here, we check whether advancing the finger would make
7452 // us cross into a new card, and if so clear corresponding
7453 // cards in the MUT (preclean them in the card-table in the
7454 // future).
7455
7456 DEBUG_ONLY(if (!_verifying) {)
7457 // The clean-on-enter optimization is disabled by default,
7458 // until we fix 6178663.
7459 if (CMSCleanOnEnter && (_finger > _threshold)) {
7460 // [_threshold, _finger) represents the interval
7461 // of cards to be cleared in MUT (or precleaned in card table).
7462 // The set of cards to be cleared is all those that overlap
7463 // with the interval [_threshold, _finger); note that
7464 // _threshold is always kept card-aligned but _finger isn't
7465 // always card-aligned.
7466 HeapWord* old_threshold = _threshold;
7467 assert(old_threshold == (HeapWord*)round_to(
7468 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7469 "_threshold should always be card-aligned");
7470 _threshold = (HeapWord*)round_to(
7471 (intptr_t)_finger, CardTableModRefBS::card_size);
7472 MemRegion mr(old_threshold, _threshold);
7473 assert(!mr.is_empty(), "Control point invariant");
7474 assert(_span.contains(mr), "Should clear within span");
7475 // XXX When _finger crosses from old gen into perm gen
7476 // we may be doing unnecessary cleaning; do better in the
7477 // future by detecting that condition and clearing fewer
7478 // MUT/CT entries.
7479 _mut->clear_range(mr);
7480 }
7481 DEBUG_ONLY(})
7482 // Note: the finger doesn't advance while we drain
7483 // the stack below.
7484 PushOrMarkClosure pushOrMarkClosure(_collector,
7485 _span, _bitMap, _markStack,
7486 _revisitStack,
7487 _finger, this);
7488 bool res = _markStack->push(obj);
7489 assert(res, "Empty non-zero size stack should have space for single push");
7490 while (!_markStack->isEmpty()) {
7491 oop new_oop = _markStack->pop();
7492 // Skip verifying header mark word below because we are
7493 // running concurrent with mutators.
7494 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7495 // now scan this oop's oops
7496 new_oop->oop_iterate(&pushOrMarkClosure);
7497 do_yield_check();
7498 }
7499 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
7500 }
7501
7502 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
7503 CMSCollector* collector, MemRegion span,
7504 CMSBitMap* bit_map,
7505 OopTaskQueue* work_queue,
7506 CMSMarkStack* overflow_stack,
7507 CMSMarkStack* revisit_stack,
7508 bool should_yield):
7509 _collector(collector),
7510 _whole_span(collector->_span),
7511 _span(span),
7512 _bit_map(bit_map),
7513 _mut(&collector->_modUnionTable),
7514 _work_queue(work_queue),
7515 _overflow_stack(overflow_stack),
7516 _revisit_stack(revisit_stack),
7517 _yield(should_yield),
7518 _skip_bits(0),
7519 _task(task)
7520 {
7521 assert(_work_queue->size() == 0, "work_queue should be empty");
7522 _finger = span.start();
7523 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
7524 assert(_span.contains(_finger), "Out of bounds _finger?");
7525 }
7526
7527 // Should revisit to see if this should be restructured for
7528 // greater efficiency.
7529 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
7530 if (_skip_bits > 0) {
7531 _skip_bits--;
7532 return true;
7533 }
7534 // convert offset into a HeapWord*
7535 HeapWord* addr = _bit_map->startWord() + offset;
7536 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
7537 "address out of range");
7538 assert(_bit_map->isMarked(addr), "tautology");
7539 if (_bit_map->isMarked(addr+1)) {
7540 // this is an allocated object that might not yet be initialized
7541 assert(_skip_bits == 0, "tautology");
7542 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
7543 oop p = oop(addr);
7544 if (p->klass_or_null() == NULL || !p->is_parsable()) {
7545 // in the case of Clean-on-Enter optimization, redirty card
7546 // and avoid clearing card by increasing the threshold.
7547 return true;
7548 }
7549 }
7550 scan_oops_in_oop(addr);
7551 return true;
7552 }
7553
7554 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
7555 assert(_bit_map->isMarked(ptr), "expected bit to be set");
7556 // Should we assert that our work queue is empty or
7557 // below some drain limit?
7558 assert(_work_queue->size() == 0,
7559 "should drain stack to limit stack usage");
7560 // convert ptr to an oop preparatory to scanning
7561 oop obj = oop(ptr);
7562 // Ignore mark word in verification below, since we
7563 // may be running concurrent with mutators.
7564 assert(obj->is_oop(true), "should be an oop");
7565 assert(_finger <= ptr, "_finger runneth ahead");
7566 // advance the finger to right end of this object
7567 _finger = ptr + obj->size();
7568 assert(_finger > ptr, "we just incremented it above");
7569 // On large heaps, it may take us some time to get through
7570 // the marking phase (especially if running iCMS). During
7571 // this time it's possible that a lot of mutations have
7572 // accumulated in the card table and the mod union table --
7573 // these mutation records are redundant until we have
7574 // actually traced into the corresponding card.
7575 // Here, we check whether advancing the finger would make
7576 // us cross into a new card, and if so clear corresponding
7577 // cards in the MUT (preclean them in the card-table in the
7578 // future).
7579
7580 // The clean-on-enter optimization is disabled by default,
7581 // until we fix 6178663.
7582 if (CMSCleanOnEnter && (_finger > _threshold)) {
7583 // [_threshold, _finger) represents the interval
7584 // of cards to be cleared in MUT (or precleaned in card table).
7585 // The set of cards to be cleared is all those that overlap
7586 // with the interval [_threshold, _finger); note that
7587 // _threshold is always kept card-aligned but _finger isn't
7588 // always card-aligned.
7589 HeapWord* old_threshold = _threshold;
7590 assert(old_threshold == (HeapWord*)round_to(
7591 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7592 "_threshold should always be card-aligned");
7593 _threshold = (HeapWord*)round_to(
7594 (intptr_t)_finger, CardTableModRefBS::card_size);
7595 MemRegion mr(old_threshold, _threshold);
7596 assert(!mr.is_empty(), "Control point invariant");
7597 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7598 // XXX When _finger crosses from old gen into perm gen
7599 // we may be doing unnecessary cleaning; do better in the
7600 // future by detecting that condition and clearing fewer
7601 // MUT/CT entries.
7602 _mut->clear_range(mr);
7603 }
7604
7605 // Note: the local finger doesn't advance while we drain
7606 // the stack below, but the global finger sure can and will.
7607 HeapWord** gfa = _task->global_finger_addr();
7608 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7609 _span, _bit_map,
7610 _work_queue,
7611 _overflow_stack,
7612 _revisit_stack,
7613 _finger,
7614 gfa, this);
7615 bool res = _work_queue->push(obj); // overflow could occur here
7616 assert(res, "Will hold once we use workqueues");
7617 while (true) {
7618 oop new_oop;
7619 if (!_work_queue->pop_local(new_oop)) {
7620 // We emptied our work_queue; check if there's stuff that can
7621 // be gotten from the overflow stack.
7622 if (CMSConcMarkingTask::get_work_from_overflow_stack(
7623 _overflow_stack, _work_queue)) {
7624 do_yield_check();
7625 continue;
7626 } else { // done
7627 break;
7628 }
7629 }
7630 // Skip verifying header mark word below because we are
7631 // running concurrent with mutators.
7632 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7633 // now scan this oop's oops
7634 new_oop->oop_iterate(&pushOrMarkClosure);
7635 do_yield_check();
7636 }
7637 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7638 }
7639
7640 // Yield in response to a request from VM Thread or
7641 // from mutators.
7642 void Par_MarkFromRootsClosure::do_yield_work() {
7643 assert(_task != NULL, "sanity");
7644 _task->yield();
7645 }
7646
7647 // A variant of the above used for verifying CMS marking work.
7648 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7649 MemRegion span,
7650 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7651 CMSMarkStack* mark_stack):
7652 _collector(collector),
7653 _span(span),
7654 _verification_bm(verification_bm),
7655 _cms_bm(cms_bm),
7656 _mark_stack(mark_stack),
7657 _pam_verify_closure(collector, span, verification_bm, cms_bm,
7658 mark_stack)
7659 {
7660 assert(_mark_stack->isEmpty(), "stack should be empty");
7661 _finger = _verification_bm->startWord();
7662 assert(_collector->_restart_addr == NULL, "Sanity check");
7663 assert(_span.contains(_finger), "Out of bounds _finger?");
7664 }
7665
7666 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7667 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7668 assert(_span.contains(addr), "Out of bounds _finger?");
7669 _finger = addr;
7670 }
7671
7672 // Should revisit to see if this should be restructured for
7673 // greater efficiency.
7674 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7675 // convert offset into a HeapWord*
7676 HeapWord* addr = _verification_bm->startWord() + offset;
7677 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7678 "address out of range");
7679 assert(_verification_bm->isMarked(addr), "tautology");
7680 assert(_cms_bm->isMarked(addr), "tautology");
7681
7682 assert(_mark_stack->isEmpty(),
7683 "should drain stack to limit stack usage");
7684 // convert addr to an oop preparatory to scanning
7685 oop obj = oop(addr);
7686 assert(obj->is_oop(), "should be an oop");
7687 assert(_finger <= addr, "_finger runneth ahead");
7688 // advance the finger to right end of this object
7689 _finger = addr + obj->size();
7690 assert(_finger > addr, "we just incremented it above");
7691 // Note: the finger doesn't advance while we drain
7692 // the stack below.
7693 bool res = _mark_stack->push(obj);
7694 assert(res, "Empty non-zero size stack should have space for single push");
7695 while (!_mark_stack->isEmpty()) {
7696 oop new_oop = _mark_stack->pop();
7697 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7698 // now scan this oop's oops
7699 new_oop->oop_iterate(&_pam_verify_closure);
7700 }
7701 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7702 return true;
7703 }
7704
7705 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7706 CMSCollector* collector, MemRegion span,
7707 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7708 CMSMarkStack* mark_stack):
7709 OopClosure(collector->ref_processor()),
7710 _collector(collector),
7711 _span(span),
7712 _verification_bm(verification_bm),
7713 _cms_bm(cms_bm),
7714 _mark_stack(mark_stack)
7715 { }
7716
7717 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7718 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7719
7720 // Upon stack overflow, we discard (part of) the stack,
7721 // remembering the least address amongst those discarded
7722 // in CMSCollector's _restart_address.
7723 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7724 // Remember the least grey address discarded
7725 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7726 _collector->lower_restart_addr(ra);
7727 _mark_stack->reset(); // discard stack contents
7728 _mark_stack->expand(); // expand the stack if possible
7729 }
7730
7731 void PushAndMarkVerifyClosure::do_oop(oop obj) {
7732 assert(obj->is_oop_or_null(), "expected an oop or NULL");
7733 HeapWord* addr = (HeapWord*)obj;
7734 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7735 // Oop lies in _span and isn't yet grey or black
7736 _verification_bm->mark(addr); // now grey
7737 if (!_cms_bm->isMarked(addr)) {
7738 oop(addr)->print();
7739 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
7740 addr);
7741 fatal("... aborting");
7742 }
7743
7744 if (!_mark_stack->push(obj)) { // stack overflow
7745 if (PrintCMSStatistics != 0) {
7746 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7747 SIZE_FORMAT, _mark_stack->capacity());
7748 }
7749 assert(_mark_stack->isFull(), "Else push should have succeeded");
7750 handle_stack_overflow(addr);
7751 }
7752 // anything including and to the right of _finger
7753 // will be scanned as we iterate over the remainder of the
7754 // bit map
7755 }
7756 }
7757
7758 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7759 MemRegion span,
7760 CMSBitMap* bitMap, CMSMarkStack* markStack,
7761 CMSMarkStack* revisitStack,
7762 HeapWord* finger, MarkFromRootsClosure* parent) :
7763 KlassRememberingOopClosure(collector, collector->ref_processor(), revisitStack),
7764 _span(span),
7765 _bitMap(bitMap),
7766 _markStack(markStack),
7767 _finger(finger),
7768 _parent(parent)
7769 { }
7770
7771 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7772 MemRegion span,
7773 CMSBitMap* bit_map,
7774 OopTaskQueue* work_queue,
7775 CMSMarkStack* overflow_stack,
7776 CMSMarkStack* revisit_stack,
7777 HeapWord* finger,
7778 HeapWord** global_finger_addr,
7779 Par_MarkFromRootsClosure* parent) :
7780 Par_KlassRememberingOopClosure(collector,
7781 collector->ref_processor(),
7782 revisit_stack),
7783 _whole_span(collector->_span),
7784 _span(span),
7785 _bit_map(bit_map),
7786 _work_queue(work_queue),
7787 _overflow_stack(overflow_stack),
7788 _finger(finger),
7789 _global_finger_addr(global_finger_addr),
7790 _parent(parent)
7791 { }
7792
7793 // Assumes thread-safe access by callers, who are
7794 // responsible for mutual exclusion.
7795 void CMSCollector::lower_restart_addr(HeapWord* low) {
7796 assert(_span.contains(low), "Out of bounds addr");
7797 if (_restart_addr == NULL) {
7798 _restart_addr = low;
7799 } else {
7800 _restart_addr = MIN2(_restart_addr, low);
7801 }
7802 }
7803
7804 // Upon stack overflow, we discard (part of) the stack,
7805 // remembering the least address amongst those discarded
7806 // in CMSCollector's _restart_address.
7807 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7808 // Remember the least grey address discarded
7809 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7810 _collector->lower_restart_addr(ra);
7811 _markStack->reset(); // discard stack contents
7812 _markStack->expand(); // expand the stack if possible
7813 }
7814
7815 // Upon stack overflow, we discard (part of) the stack,
7816 // remembering the least address amongst those discarded
7817 // in CMSCollector's _restart_address.
7818 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7819 // We need to do this under a mutex to prevent other
7820 // workers from interfering with the work done below.
7821 MutexLockerEx ml(_overflow_stack->par_lock(),
7822 Mutex::_no_safepoint_check_flag);
7823 // Remember the least grey address discarded
7824 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7825 _collector->lower_restart_addr(ra);
7826 _overflow_stack->reset(); // discard stack contents
7827 _overflow_stack->expand(); // expand the stack if possible
7828 }
7829
7830 void PushOrMarkClosure::do_oop(oop obj) {
7831 // Ignore mark word because we are running concurrent with mutators.
7832 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7833 HeapWord* addr = (HeapWord*)obj;
7834 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7835 // Oop lies in _span and isn't yet grey or black
7836 _bitMap->mark(addr); // now grey
7837 if (addr < _finger) {
7838 // the bit map iteration has already either passed, or
7839 // sampled, this bit in the bit map; we'll need to
7840 // use the marking stack to scan this oop's oops.
7841 bool simulate_overflow = false;
7842 NOT_PRODUCT(
7843 if (CMSMarkStackOverflowALot &&
7844 _collector->simulate_overflow()) {
7845 // simulate a stack overflow
7846 simulate_overflow = true;
7847 }
7848 )
7849 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7850 if (PrintCMSStatistics != 0) {
7851 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7852 SIZE_FORMAT, _markStack->capacity());
7853 }
7854 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7855 handle_stack_overflow(addr);
7856 }
7857 }
7858 // anything including and to the right of _finger
7859 // will be scanned as we iterate over the remainder of the
7860 // bit map
7861 do_yield_check();
7862 }
7863 }
7864
7865 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7866 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7867
7868 void Par_PushOrMarkClosure::do_oop(oop obj) {
7869 // Ignore mark word because we are running concurrent with mutators.
7870 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7871 HeapWord* addr = (HeapWord*)obj;
7872 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7873 // Oop lies in _span and isn't yet grey or black
7874 // We read the global_finger (volatile read) strictly after marking oop
7875 bool res = _bit_map->par_mark(addr); // now grey
7876 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7877 // Should we push this marked oop on our stack?
7878 // -- if someone else marked it, nothing to do
7879 // -- if target oop is above global finger nothing to do
7880 // -- if target oop is in chunk and above local finger
7881 // then nothing to do
7882 // -- else push on work queue
7883 if ( !res // someone else marked it, they will deal with it
7884 || (addr >= *gfa) // will be scanned in a later task
7885 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7886 return;
7887 }
7888 // the bit map iteration has already either passed, or
7889 // sampled, this bit in the bit map; we'll need to
7890 // use the marking stack to scan this oop's oops.
7891 bool simulate_overflow = false;
7892 NOT_PRODUCT(
7893 if (CMSMarkStackOverflowALot &&
7894 _collector->simulate_overflow()) {
7895 // simulate a stack overflow
7896 simulate_overflow = true;
7897 }
7898 )
7899 if (simulate_overflow ||
7900 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7901 // stack overflow
7902 if (PrintCMSStatistics != 0) {
7903 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7904 SIZE_FORMAT, _overflow_stack->capacity());
7905 }
7906 // We cannot assert that the overflow stack is full because
7907 // it may have been emptied since.
7908 assert(simulate_overflow ||
7909 _work_queue->size() == _work_queue->max_elems(),
7910 "Else push should have succeeded");
7911 handle_stack_overflow(addr);
7912 }
7913 do_yield_check();
7914 }
7915 }
7916
7917 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7918 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7919
7920 KlassRememberingOopClosure::KlassRememberingOopClosure(CMSCollector* collector,
7921 ReferenceProcessor* rp,
7922 CMSMarkStack* revisit_stack) :
7923 OopClosure(rp),
7924 _collector(collector),
7925 _revisit_stack(revisit_stack),
7926 _should_remember_klasses(collector->should_unload_classes()) {}
7927
7928 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7929 MemRegion span,
7930 ReferenceProcessor* rp,
7931 CMSBitMap* bit_map,
7932 CMSBitMap* mod_union_table,
7933 CMSMarkStack* mark_stack,
7934 CMSMarkStack* revisit_stack,
7935 bool concurrent_precleaning):
7936 KlassRememberingOopClosure(collector, rp, revisit_stack),
7937 _span(span),
7938 _bit_map(bit_map),
7939 _mod_union_table(mod_union_table),
7940 _mark_stack(mark_stack),
7941 _concurrent_precleaning(concurrent_precleaning)
7942 {
7943 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7944 }
7945
7946 // Grey object rescan during pre-cleaning and second checkpoint phases --
7947 // the non-parallel version (the parallel version appears further below.)
7948 void PushAndMarkClosure::do_oop(oop obj) {
7949 // Ignore mark word verification. If during concurrent precleaning,
7950 // the object monitor may be locked. If during the checkpoint
7951 // phases, the object may already have been reached by a different
7952 // path and may be at the end of the global overflow list (so
7953 // the mark word may be NULL).
7954 assert(obj->is_oop_or_null(true /* ignore mark word */),
7955 "expected an oop or NULL");
7956 HeapWord* addr = (HeapWord*)obj;
7957 // Check if oop points into the CMS generation
7958 // and is not marked
7959 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7960 // a white object ...
7961 _bit_map->mark(addr); // ... now grey
7962 // push on the marking stack (grey set)
7963 bool simulate_overflow = false;
7964 NOT_PRODUCT(
7965 if (CMSMarkStackOverflowALot &&
7966 _collector->simulate_overflow()) {
7967 // simulate a stack overflow
7968 simulate_overflow = true;
7969 }
7970 )
7971 if (simulate_overflow || !_mark_stack->push(obj)) {
7972 if (_concurrent_precleaning) {
7973 // During precleaning we can just dirty the appropriate card(s)
7974 // in the mod union table, thus ensuring that the object remains
7975 // in the grey set and continue. In the case of object arrays
7976 // we need to dirty all of the cards that the object spans,
7977 // since the rescan of object arrays will be limited to the
7978 // dirty cards.
7979 // Note that no one can be intefering with us in this action
7980 // of dirtying the mod union table, so no locking or atomics
7981 // are required.
7982 if (obj->is_objArray()) {
7983 size_t sz = obj->size();
7984 HeapWord* end_card_addr = (HeapWord*)round_to(
7985 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7986 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7987 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7988 _mod_union_table->mark_range(redirty_range);
7989 } else {
7990 _mod_union_table->mark(addr);
7991 }
7992 _collector->_ser_pmc_preclean_ovflw++;
7993 } else {
7994 // During the remark phase, we need to remember this oop
7995 // in the overflow list.
7996 _collector->push_on_overflow_list(obj);
7997 _collector->_ser_pmc_remark_ovflw++;
7998 }
7999 }
8000 }
8001 }
8002
8003 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
8004 MemRegion span,
8005 ReferenceProcessor* rp,
8006 CMSBitMap* bit_map,
8007 OopTaskQueue* work_queue,
8008 CMSMarkStack* revisit_stack):
8009 Par_KlassRememberingOopClosure(collector, rp, revisit_stack),
8010 _span(span),
8011 _bit_map(bit_map),
8012 _work_queue(work_queue)
8013 {
8014 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
8015 }
8016
8017 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
8018 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
8019
8020 // Grey object rescan during second checkpoint phase --
8021 // the parallel version.
8022 void Par_PushAndMarkClosure::do_oop(oop obj) {
8023 // In the assert below, we ignore the mark word because
8024 // this oop may point to an already visited object that is
8025 // on the overflow stack (in which case the mark word has
8026 // been hijacked for chaining into the overflow stack --
8027 // if this is the last object in the overflow stack then
8028 // its mark word will be NULL). Because this object may
8029 // have been subsequently popped off the global overflow
8030 // stack, and the mark word possibly restored to the prototypical
8031 // value, by the time we get to examined this failing assert in
8032 // the debugger, is_oop_or_null(false) may subsequently start
8033 // to hold.
8034 assert(obj->is_oop_or_null(true),
8035 "expected an oop or NULL");
8036 HeapWord* addr = (HeapWord*)obj;
8037 // Check if oop points into the CMS generation
8038 // and is not marked
8039 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
8040 // a white object ...
8041 // If we manage to "claim" the object, by being the
8042 // first thread to mark it, then we push it on our
8043 // marking stack
8044 if (_bit_map->par_mark(addr)) { // ... now grey
8045 // push on work queue (grey set)
8046 bool simulate_overflow = false;
8047 NOT_PRODUCT(
8048 if (CMSMarkStackOverflowALot &&
8049 _collector->par_simulate_overflow()) {
8050 // simulate a stack overflow
8051 simulate_overflow = true;
8052 }
8053 )
8054 if (simulate_overflow || !_work_queue->push(obj)) {
8055 _collector->par_push_on_overflow_list(obj);
8056 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
8057 }
8058 } // Else, some other thread got there first
8059 }
8060 }
8061
8062 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
8063 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
8064
8065 void PushAndMarkClosure::remember_mdo(DataLayout* v) {
8066 // TBD
8067 }
8068
8069 void Par_PushAndMarkClosure::remember_mdo(DataLayout* v) {
8070 // TBD
8071 }
8072
8073 void CMSPrecleanRefsYieldClosure::do_yield_work() {
8074 DEBUG_ONLY(RememberKlassesChecker mux(false);)
8075 Mutex* bml = _collector->bitMapLock();
8076 assert_lock_strong(bml);
8077 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8078 "CMS thread should hold CMS token");
8079
8080 bml->unlock();
8081 ConcurrentMarkSweepThread::desynchronize(true);
8082
8083 ConcurrentMarkSweepThread::acknowledge_yield_request();
8084
8085 _collector->stopTimer();
8086 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8087 if (PrintCMSStatistics != 0) {
8088 _collector->incrementYields();
8089 }
8090 _collector->icms_wait();
8091
8092 // See the comment in coordinator_yield()
8093 for (unsigned i = 0; i < CMSYieldSleepCount &&
8094 ConcurrentMarkSweepThread::should_yield() &&
8095 !CMSCollector::foregroundGCIsActive(); ++i) {
8096 os::sleep(Thread::current(), 1, false);
8097 ConcurrentMarkSweepThread::acknowledge_yield_request();
8098 }
8099
8100 ConcurrentMarkSweepThread::synchronize(true);
8101 bml->lock();
8102
8103 _collector->startTimer();
8104 }
8105
8106 bool CMSPrecleanRefsYieldClosure::should_return() {
8107 if (ConcurrentMarkSweepThread::should_yield()) {
8108 do_yield_work();
8109 }
8110 return _collector->foregroundGCIsActive();
8111 }
8112
8113 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
8114 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
8115 "mr should be aligned to start at a card boundary");
8116 // We'd like to assert:
8117 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
8118 // "mr should be a range of cards");
8119 // However, that would be too strong in one case -- the last
8120 // partition ends at _unallocated_block which, in general, can be
8121 // an arbitrary boundary, not necessarily card aligned.
8122 if (PrintCMSStatistics != 0) {
8123 _num_dirty_cards +=
8124 mr.word_size()/CardTableModRefBS::card_size_in_words;
8125 }
8126 _space->object_iterate_mem(mr, &_scan_cl);
8127 }
8128
8129 SweepClosure::SweepClosure(CMSCollector* collector,
8130 ConcurrentMarkSweepGeneration* g,
8131 CMSBitMap* bitMap, bool should_yield) :
8132 _collector(collector),
8133 _g(g),
8134 _sp(g->cmsSpace()),
8135 _limit(_sp->sweep_limit()),
8136 _freelistLock(_sp->freelistLock()),
8137 _bitMap(bitMap),
8138 _yield(should_yield),
8139 _inFreeRange(false), // No free range at beginning of sweep
8140 _freeRangeInFreeLists(false), // No free range at beginning of sweep
8141 _lastFreeRangeCoalesced(false),
8142 _freeFinger(g->used_region().start())
8143 {
8144 NOT_PRODUCT(
8145 _numObjectsFreed = 0;
8146 _numWordsFreed = 0;
8147 _numObjectsLive = 0;
8148 _numWordsLive = 0;
8149 _numObjectsAlreadyFree = 0;
8150 _numWordsAlreadyFree = 0;
8151 _last_fc = NULL;
8152
8153 _sp->initializeIndexedFreeListArrayReturnedBytes();
8154 _sp->dictionary()->initialize_dict_returned_bytes();
8155 )
8156 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8157 "sweep _limit out of bounds");
8158 if (CMSTraceSweeper) {
8159 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
8160 _limit);
8161 }
8162 }
8163
8164 void SweepClosure::print_on(outputStream* st) const {
8165 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
8166 _sp->bottom(), _sp->end());
8167 tty->print_cr("_limit = " PTR_FORMAT, _limit);
8168 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger);
8169 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);)
8170 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
8171 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
8172 }
8173
8174 #ifndef PRODUCT
8175 // Assertion checking only: no useful work in product mode --
8176 // however, if any of the flags below become product flags,
8177 // you may need to review this code to see if it needs to be
8178 // enabled in product mode.
8179 SweepClosure::~SweepClosure() {
8180 assert_lock_strong(_freelistLock);
8181 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8182 "sweep _limit out of bounds");
8183 if (inFreeRange()) {
8184 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
8185 print();
8186 ShouldNotReachHere();
8187 }
8188 if (Verbose && PrintGC) {
8189 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes",
8190 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
8191 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
8192 SIZE_FORMAT" bytes "
8193 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
8194 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
8195 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
8196 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
8197 * sizeof(HeapWord);
8198 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
8199
8200 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
8201 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
8202 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
8203 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
8204 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes);
8205 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
8206 indexListReturnedBytes);
8207 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
8208 dict_returned_bytes);
8209 }
8210 }
8211 if (CMSTraceSweeper) {
8212 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
8213 _limit);
8214 }
8215 }
8216 #endif // PRODUCT
8217
8218 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
8219 bool freeRangeInFreeLists) {
8220 if (CMSTraceSweeper) {
8221 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n",
8222 freeFinger, freeRangeInFreeLists);
8223 }
8224 assert(!inFreeRange(), "Trampling existing free range");
8225 set_inFreeRange(true);
8226 set_lastFreeRangeCoalesced(false);
8227
8228 set_freeFinger(freeFinger);
8229 set_freeRangeInFreeLists(freeRangeInFreeLists);
8230 if (CMSTestInFreeList) {
8231 if (freeRangeInFreeLists) {
8232 FreeChunk* fc = (FreeChunk*) freeFinger;
8233 assert(fc->is_free(), "A chunk on the free list should be free.");
8234 assert(fc->size() > 0, "Free range should have a size");
8235 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
8236 }
8237 }
8238 }
8239
8240 // Note that the sweeper runs concurrently with mutators. Thus,
8241 // it is possible for direct allocation in this generation to happen
8242 // in the middle of the sweep. Note that the sweeper also coalesces
8243 // contiguous free blocks. Thus, unless the sweeper and the allocator
8244 // synchronize appropriately freshly allocated blocks may get swept up.
8245 // This is accomplished by the sweeper locking the free lists while
8246 // it is sweeping. Thus blocks that are determined to be free are
8247 // indeed free. There is however one additional complication:
8248 // blocks that have been allocated since the final checkpoint and
8249 // mark, will not have been marked and so would be treated as
8250 // unreachable and swept up. To prevent this, the allocator marks
8251 // the bit map when allocating during the sweep phase. This leads,
8252 // however, to a further complication -- objects may have been allocated
8253 // but not yet initialized -- in the sense that the header isn't yet
8254 // installed. The sweeper can not then determine the size of the block
8255 // in order to skip over it. To deal with this case, we use a technique
8256 // (due to Printezis) to encode such uninitialized block sizes in the
8257 // bit map. Since the bit map uses a bit per every HeapWord, but the
8258 // CMS generation has a minimum object size of 3 HeapWords, it follows
8259 // that "normal marks" won't be adjacent in the bit map (there will
8260 // always be at least two 0 bits between successive 1 bits). We make use
8261 // of these "unused" bits to represent uninitialized blocks -- the bit
8262 // corresponding to the start of the uninitialized object and the next
8263 // bit are both set. Finally, a 1 bit marks the end of the object that
8264 // started with the two consecutive 1 bits to indicate its potentially
8265 // uninitialized state.
8266
8267 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
8268 FreeChunk* fc = (FreeChunk*)addr;
8269 size_t res;
8270
8271 // Check if we are done sweeping. Below we check "addr >= _limit" rather
8272 // than "addr == _limit" because although _limit was a block boundary when
8273 // we started the sweep, it may no longer be one because heap expansion
8274 // may have caused us to coalesce the block ending at the address _limit
8275 // with a newly expanded chunk (this happens when _limit was set to the
8276 // previous _end of the space), so we may have stepped past _limit:
8277 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
8278 if (addr >= _limit) { // we have swept up to or past the limit: finish up
8279 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8280 "sweep _limit out of bounds");
8281 assert(addr < _sp->end(), "addr out of bounds");
8282 // Flush any free range we might be holding as a single
8283 // coalesced chunk to the appropriate free list.
8284 if (inFreeRange()) {
8285 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
8286 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger()));
8287 flush_cur_free_chunk(freeFinger(),
8288 pointer_delta(addr, freeFinger()));
8289 if (CMSTraceSweeper) {
8290 gclog_or_tty->print("Sweep: last chunk: ");
8291 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") "
8292 "[coalesced:"SIZE_FORMAT"]\n",
8293 freeFinger(), pointer_delta(addr, freeFinger()),
8294 lastFreeRangeCoalesced());
8295 }
8296 }
8297
8298 // help the iterator loop finish
8299 return pointer_delta(_sp->end(), addr);
8300 }
8301
8302 assert(addr < _limit, "sweep invariant");
8303 // check if we should yield
8304 do_yield_check(addr);
8305 if (fc->is_free()) {
8306 // Chunk that is already free
8307 res = fc->size();
8308 do_already_free_chunk(fc);
8309 debug_only(_sp->verifyFreeLists());
8310 // If we flush the chunk at hand in lookahead_and_flush()
8311 // and it's coalesced with a preceding chunk, then the
8312 // process of "mangling" the payload of the coalesced block
8313 // will cause erasure of the size information from the
8314 // (erstwhile) header of all the coalesced blocks but the
8315 // first, so the first disjunct in the assert will not hold
8316 // in that specific case (in which case the second disjunct
8317 // will hold).
8318 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
8319 "Otherwise the size info doesn't change at this step");
8320 NOT_PRODUCT(
8321 _numObjectsAlreadyFree++;
8322 _numWordsAlreadyFree += res;
8323 )
8324 NOT_PRODUCT(_last_fc = fc;)
8325 } else if (!_bitMap->isMarked(addr)) {
8326 // Chunk is fresh garbage
8327 res = do_garbage_chunk(fc);
8328 debug_only(_sp->verifyFreeLists());
8329 NOT_PRODUCT(
8330 _numObjectsFreed++;
8331 _numWordsFreed += res;
8332 )
8333 } else {
8334 // Chunk that is alive.
8335 res = do_live_chunk(fc);
8336 debug_only(_sp->verifyFreeLists());
8337 NOT_PRODUCT(
8338 _numObjectsLive++;
8339 _numWordsLive += res;
8340 )
8341 }
8342 return res;
8343 }
8344
8345 // For the smart allocation, record following
8346 // split deaths - a free chunk is removed from its free list because
8347 // it is being split into two or more chunks.
8348 // split birth - a free chunk is being added to its free list because
8349 // a larger free chunk has been split and resulted in this free chunk.
8350 // coal death - a free chunk is being removed from its free list because
8351 // it is being coalesced into a large free chunk.
8352 // coal birth - a free chunk is being added to its free list because
8353 // it was created when two or more free chunks where coalesced into
8354 // this free chunk.
8355 //
8356 // These statistics are used to determine the desired number of free
8357 // chunks of a given size. The desired number is chosen to be relative
8358 // to the end of a CMS sweep. The desired number at the end of a sweep
8359 // is the
8360 // count-at-end-of-previous-sweep (an amount that was enough)
8361 // - count-at-beginning-of-current-sweep (the excess)
8362 // + split-births (gains in this size during interval)
8363 // - split-deaths (demands on this size during interval)
8364 // where the interval is from the end of one sweep to the end of the
8365 // next.
8366 //
8367 // When sweeping the sweeper maintains an accumulated chunk which is
8368 // the chunk that is made up of chunks that have been coalesced. That
8369 // will be termed the left-hand chunk. A new chunk of garbage that
8370 // is being considered for coalescing will be referred to as the
8371 // right-hand chunk.
8372 //
8373 // When making a decision on whether to coalesce a right-hand chunk with
8374 // the current left-hand chunk, the current count vs. the desired count
8375 // of the left-hand chunk is considered. Also if the right-hand chunk
8376 // is near the large chunk at the end of the heap (see
8377 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
8378 // left-hand chunk is coalesced.
8379 //
8380 // When making a decision about whether to split a chunk, the desired count
8381 // vs. the current count of the candidate to be split is also considered.
8382 // If the candidate is underpopulated (currently fewer chunks than desired)
8383 // a chunk of an overpopulated (currently more chunks than desired) size may
8384 // be chosen. The "hint" associated with a free list, if non-null, points
8385 // to a free list which may be overpopulated.
8386 //
8387
8388 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
8389 const size_t size = fc->size();
8390 // Chunks that cannot be coalesced are not in the
8391 // free lists.
8392 if (CMSTestInFreeList && !fc->cantCoalesce()) {
8393 assert(_sp->verify_chunk_in_free_list(fc),
8394 "free chunk should be in free lists");
8395 }
8396 // a chunk that is already free, should not have been
8397 // marked in the bit map
8398 HeapWord* const addr = (HeapWord*) fc;
8399 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
8400 // Verify that the bit map has no bits marked between
8401 // addr and purported end of this block.
8402 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8403
8404 // Some chunks cannot be coalesced under any circumstances.
8405 // See the definition of cantCoalesce().
8406 if (!fc->cantCoalesce()) {
8407 // This chunk can potentially be coalesced.
8408 if (_sp->adaptive_freelists()) {
8409 // All the work is done in
8410 do_post_free_or_garbage_chunk(fc, size);
8411 } else { // Not adaptive free lists
8412 // this is a free chunk that can potentially be coalesced by the sweeper;
8413 if (!inFreeRange()) {
8414 // if the next chunk is a free block that can't be coalesced
8415 // it doesn't make sense to remove this chunk from the free lists
8416 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
8417 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
8418 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
8419 nextChunk->is_free() && // ... which is free...
8420 nextChunk->cantCoalesce()) { // ... but can't be coalesced
8421 // nothing to do
8422 } else {
8423 // Potentially the start of a new free range:
8424 // Don't eagerly remove it from the free lists.
8425 // No need to remove it if it will just be put
8426 // back again. (Also from a pragmatic point of view
8427 // if it is a free block in a region that is beyond
8428 // any allocated blocks, an assertion will fail)
8429 // Remember the start of a free run.
8430 initialize_free_range(addr, true);
8431 // end - can coalesce with next chunk
8432 }
8433 } else {
8434 // the midst of a free range, we are coalescing
8435 print_free_block_coalesced(fc);
8436 if (CMSTraceSweeper) {
8437 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size);
8438 }
8439 // remove it from the free lists
8440 _sp->removeFreeChunkFromFreeLists(fc);
8441 set_lastFreeRangeCoalesced(true);
8442 // If the chunk is being coalesced and the current free range is
8443 // in the free lists, remove the current free range so that it
8444 // will be returned to the free lists in its entirety - all
8445 // the coalesced pieces included.
8446 if (freeRangeInFreeLists()) {
8447 FreeChunk* ffc = (FreeChunk*) freeFinger();
8448 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8449 "Size of free range is inconsistent with chunk size.");
8450 if (CMSTestInFreeList) {
8451 assert(_sp->verify_chunk_in_free_list(ffc),
8452 "free range is not in free lists");
8453 }
8454 _sp->removeFreeChunkFromFreeLists(ffc);
8455 set_freeRangeInFreeLists(false);
8456 }
8457 }
8458 }
8459 // Note that if the chunk is not coalescable (the else arm
8460 // below), we unconditionally flush, without needing to do
8461 // a "lookahead," as we do below.
8462 if (inFreeRange()) lookahead_and_flush(fc, size);
8463 } else {
8464 // Code path common to both original and adaptive free lists.
8465
8466 // cant coalesce with previous block; this should be treated
8467 // as the end of a free run if any
8468 if (inFreeRange()) {
8469 // we kicked some butt; time to pick up the garbage
8470 assert(freeFinger() < addr, "freeFinger points too high");
8471 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8472 }
8473 // else, nothing to do, just continue
8474 }
8475 }
8476
8477 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
8478 // This is a chunk of garbage. It is not in any free list.
8479 // Add it to a free list or let it possibly be coalesced into
8480 // a larger chunk.
8481 HeapWord* const addr = (HeapWord*) fc;
8482 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8483
8484 if (_sp->adaptive_freelists()) {
8485 // Verify that the bit map has no bits marked between
8486 // addr and purported end of just dead object.
8487 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8488
8489 do_post_free_or_garbage_chunk(fc, size);
8490 } else {
8491 if (!inFreeRange()) {
8492 // start of a new free range
8493 assert(size > 0, "A free range should have a size");
8494 initialize_free_range(addr, false);
8495 } else {
8496 // this will be swept up when we hit the end of the
8497 // free range
8498 if (CMSTraceSweeper) {
8499 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size);
8500 }
8501 // If the chunk is being coalesced and the current free range is
8502 // in the free lists, remove the current free range so that it
8503 // will be returned to the free lists in its entirety - all
8504 // the coalesced pieces included.
8505 if (freeRangeInFreeLists()) {
8506 FreeChunk* ffc = (FreeChunk*)freeFinger();
8507 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8508 "Size of free range is inconsistent with chunk size.");
8509 if (CMSTestInFreeList) {
8510 assert(_sp->verify_chunk_in_free_list(ffc),
8511 "free range is not in free lists");
8512 }
8513 _sp->removeFreeChunkFromFreeLists(ffc);
8514 set_freeRangeInFreeLists(false);
8515 }
8516 set_lastFreeRangeCoalesced(true);
8517 }
8518 // this will be swept up when we hit the end of the free range
8519
8520 // Verify that the bit map has no bits marked between
8521 // addr and purported end of just dead object.
8522 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8523 }
8524 assert(_limit >= addr + size,
8525 "A freshly garbage chunk can't possibly straddle over _limit");
8526 if (inFreeRange()) lookahead_and_flush(fc, size);
8527 return size;
8528 }
8529
8530 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
8531 HeapWord* addr = (HeapWord*) fc;
8532 // The sweeper has just found a live object. Return any accumulated
8533 // left hand chunk to the free lists.
8534 if (inFreeRange()) {
8535 assert(freeFinger() < addr, "freeFinger points too high");
8536 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8537 }
8538
8539 // This object is live: we'd normally expect this to be
8540 // an oop, and like to assert the following:
8541 // assert(oop(addr)->is_oop(), "live block should be an oop");
8542 // However, as we commented above, this may be an object whose
8543 // header hasn't yet been initialized.
8544 size_t size;
8545 assert(_bitMap->isMarked(addr), "Tautology for this control point");
8546 if (_bitMap->isMarked(addr + 1)) {
8547 // Determine the size from the bit map, rather than trying to
8548 // compute it from the object header.
8549 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
8550 size = pointer_delta(nextOneAddr + 1, addr);
8551 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
8552 "alignment problem");
8553
8554 #ifdef DEBUG
8555 if (oop(addr)->klass_or_null() != NULL &&
8556 ( !_collector->should_unload_classes()
8557 || (oop(addr)->is_parsable()) &&
8558 oop(addr)->is_conc_safe())) {
8559 // Ignore mark word because we are running concurrent with mutators
8560 assert(oop(addr)->is_oop(true), "live block should be an oop");
8561 // is_conc_safe is checked before performing this assertion
8562 // because an object that is not is_conc_safe may yet have
8563 // the return from size() correct.
8564 assert(size ==
8565 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
8566 "P-mark and computed size do not agree");
8567 }
8568 #endif
8569
8570 } else {
8571 // This should be an initialized object that's alive.
8572 assert(oop(addr)->klass_or_null() != NULL &&
8573 (!_collector->should_unload_classes()
8574 || oop(addr)->is_parsable()),
8575 "Should be an initialized object");
8576 // Note that there are objects used during class redefinition,
8577 // e.g. merge_cp in VM_RedefineClasses::merge_cp_and_rewrite(),
8578 // which are discarded with their is_conc_safe state still
8579 // false. These object may be floating garbage so may be
8580 // seen here. If they are floating garbage their size
8581 // should be attainable from their klass. Do not that
8582 // is_conc_safe() is true for oop(addr).
8583 // Ignore mark word because we are running concurrent with mutators
8584 assert(oop(addr)->is_oop(true), "live block should be an oop");
8585 // Verify that the bit map has no bits marked between
8586 // addr and purported end of this block.
8587 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8588 assert(size >= 3, "Necessary for Printezis marks to work");
8589 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
8590 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
8591 }
8592 return size;
8593 }
8594
8595 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
8596 size_t chunkSize) {
8597 // do_post_free_or_garbage_chunk() should only be called in the case
8598 // of the adaptive free list allocator.
8599 const bool fcInFreeLists = fc->is_free();
8600 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
8601 assert((HeapWord*)fc <= _limit, "sweep invariant");
8602 if (CMSTestInFreeList && fcInFreeLists) {
8603 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
8604 }
8605
8606 if (CMSTraceSweeper) {
8607 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
8608 }
8609
8610 HeapWord* const fc_addr = (HeapWord*) fc;
8611
8612 bool coalesce;
8613 const size_t left = pointer_delta(fc_addr, freeFinger());
8614 const size_t right = chunkSize;
8615 switch (FLSCoalescePolicy) {
8616 // numeric value forms a coalition aggressiveness metric
8617 case 0: { // never coalesce
8618 coalesce = false;
8619 break;
8620 }
8621 case 1: { // coalesce if left & right chunks on overpopulated lists
8622 coalesce = _sp->coalOverPopulated(left) &&
8623 _sp->coalOverPopulated(right);
8624 break;
8625 }
8626 case 2: { // coalesce if left chunk on overpopulated list (default)
8627 coalesce = _sp->coalOverPopulated(left);
8628 break;
8629 }
8630 case 3: { // coalesce if left OR right chunk on overpopulated list
8631 coalesce = _sp->coalOverPopulated(left) ||
8632 _sp->coalOverPopulated(right);
8633 break;
8634 }
8635 case 4: { // always coalesce
8636 coalesce = true;
8637 break;
8638 }
8639 default:
8640 ShouldNotReachHere();
8641 }
8642
8643 // Should the current free range be coalesced?
8644 // If the chunk is in a free range and either we decided to coalesce above
8645 // or the chunk is near the large block at the end of the heap
8646 // (isNearLargestChunk() returns true), then coalesce this chunk.
8647 const bool doCoalesce = inFreeRange()
8648 && (coalesce || _g->isNearLargestChunk(fc_addr));
8649 if (doCoalesce) {
8650 // Coalesce the current free range on the left with the new
8651 // chunk on the right. If either is on a free list,
8652 // it must be removed from the list and stashed in the closure.
8653 if (freeRangeInFreeLists()) {
8654 FreeChunk* const ffc = (FreeChunk*)freeFinger();
8655 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
8656 "Size of free range is inconsistent with chunk size.");
8657 if (CMSTestInFreeList) {
8658 assert(_sp->verify_chunk_in_free_list(ffc),
8659 "Chunk is not in free lists");
8660 }
8661 _sp->coalDeath(ffc->size());
8662 _sp->removeFreeChunkFromFreeLists(ffc);
8663 set_freeRangeInFreeLists(false);
8664 }
8665 if (fcInFreeLists) {
8666 _sp->coalDeath(chunkSize);
8667 assert(fc->size() == chunkSize,
8668 "The chunk has the wrong size or is not in the free lists");
8669 _sp->removeFreeChunkFromFreeLists(fc);
8670 }
8671 set_lastFreeRangeCoalesced(true);
8672 print_free_block_coalesced(fc);
8673 } else { // not in a free range and/or should not coalesce
8674 // Return the current free range and start a new one.
8675 if (inFreeRange()) {
8676 // In a free range but cannot coalesce with the right hand chunk.
8677 // Put the current free range into the free lists.
8678 flush_cur_free_chunk(freeFinger(),
8679 pointer_delta(fc_addr, freeFinger()));
8680 }
8681 // Set up for new free range. Pass along whether the right hand
8682 // chunk is in the free lists.
8683 initialize_free_range((HeapWord*)fc, fcInFreeLists);
8684 }
8685 }
8686
8687 // Lookahead flush:
8688 // If we are tracking a free range, and this is the last chunk that
8689 // we'll look at because its end crosses past _limit, we'll preemptively
8690 // flush it along with any free range we may be holding on to. Note that
8691 // this can be the case only for an already free or freshly garbage
8692 // chunk. If this block is an object, it can never straddle
8693 // over _limit. The "straddling" occurs when _limit is set at
8694 // the previous end of the space when this cycle started, and
8695 // a subsequent heap expansion caused the previously co-terminal
8696 // free block to be coalesced with the newly expanded portion,
8697 // thus rendering _limit a non-block-boundary making it dangerous
8698 // for the sweeper to step over and examine.
8699 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
8700 assert(inFreeRange(), "Should only be called if currently in a free range.");
8701 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
8702 assert(_sp->used_region().contains(eob - 1),
8703 err_msg("eob = " PTR_FORMAT " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
8704 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
8705 _limit, _sp->bottom(), _sp->end(), fc, chunk_size));
8706 if (eob >= _limit) {
8707 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
8708 if (CMSTraceSweeper) {
8709 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
8710 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
8711 "[" PTR_FORMAT "," PTR_FORMAT ")",
8712 _limit, fc, eob, _sp->bottom(), _sp->end());
8713 }
8714 // Return the storage we are tracking back into the free lists.
8715 if (CMSTraceSweeper) {
8716 gclog_or_tty->print_cr("Flushing ... ");
8717 }
8718 assert(freeFinger() < eob, "Error");
8719 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
8720 }
8721 }
8722
8723 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
8724 assert(inFreeRange(), "Should only be called if currently in a free range.");
8725 assert(size > 0,
8726 "A zero sized chunk cannot be added to the free lists.");
8727 if (!freeRangeInFreeLists()) {
8728 if (CMSTestInFreeList) {
8729 FreeChunk* fc = (FreeChunk*) chunk;
8730 fc->set_size(size);
8731 assert(!_sp->verify_chunk_in_free_list(fc),
8732 "chunk should not be in free lists yet");
8733 }
8734 if (CMSTraceSweeper) {
8735 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8736 chunk, size);
8737 }
8738 // A new free range is going to be starting. The current
8739 // free range has not been added to the free lists yet or
8740 // was removed so add it back.
8741 // If the current free range was coalesced, then the death
8742 // of the free range was recorded. Record a birth now.
8743 if (lastFreeRangeCoalesced()) {
8744 _sp->coalBirth(size);
8745 }
8746 _sp->addChunkAndRepairOffsetTable(chunk, size,
8747 lastFreeRangeCoalesced());
8748 } else if (CMSTraceSweeper) {
8749 gclog_or_tty->print_cr("Already in free list: nothing to flush");
8750 }
8751 set_inFreeRange(false);
8752 set_freeRangeInFreeLists(false);
8753 }
8754
8755 // We take a break if we've been at this for a while,
8756 // so as to avoid monopolizing the locks involved.
8757 void SweepClosure::do_yield_work(HeapWord* addr) {
8758 // Return current free chunk being used for coalescing (if any)
8759 // to the appropriate freelist. After yielding, the next
8760 // free block encountered will start a coalescing range of
8761 // free blocks. If the next free block is adjacent to the
8762 // chunk just flushed, they will need to wait for the next
8763 // sweep to be coalesced.
8764 if (inFreeRange()) {
8765 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8766 }
8767
8768 // First give up the locks, then yield, then re-lock.
8769 // We should probably use a constructor/destructor idiom to
8770 // do this unlock/lock or modify the MutexUnlocker class to
8771 // serve our purpose. XXX
8772 assert_lock_strong(_bitMap->lock());
8773 assert_lock_strong(_freelistLock);
8774 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8775 "CMS thread should hold CMS token");
8776 _bitMap->lock()->unlock();
8777 _freelistLock->unlock();
8778 ConcurrentMarkSweepThread::desynchronize(true);
8779 ConcurrentMarkSweepThread::acknowledge_yield_request();
8780 _collector->stopTimer();
8781 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8782 if (PrintCMSStatistics != 0) {
8783 _collector->incrementYields();
8784 }
8785 _collector->icms_wait();
8786
8787 // See the comment in coordinator_yield()
8788 for (unsigned i = 0; i < CMSYieldSleepCount &&
8789 ConcurrentMarkSweepThread::should_yield() &&
8790 !CMSCollector::foregroundGCIsActive(); ++i) {
8791 os::sleep(Thread::current(), 1, false);
8792 ConcurrentMarkSweepThread::acknowledge_yield_request();
8793 }
8794
8795 ConcurrentMarkSweepThread::synchronize(true);
8796 _freelistLock->lock();
8797 _bitMap->lock()->lock_without_safepoint_check();
8798 _collector->startTimer();
8799 }
8800
8801 #ifndef PRODUCT
8802 // This is actually very useful in a product build if it can
8803 // be called from the debugger. Compile it into the product
8804 // as needed.
8805 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
8806 return debug_cms_space->verify_chunk_in_free_list(fc);
8807 }
8808 #endif
8809
8810 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
8811 if (CMSTraceSweeper) {
8812 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
8813 fc, fc->size());
8814 }
8815 }
8816
8817 // CMSIsAliveClosure
8818 bool CMSIsAliveClosure::do_object_b(oop obj) {
8819 HeapWord* addr = (HeapWord*)obj;
8820 return addr != NULL &&
8821 (!_span.contains(addr) || _bit_map->isMarked(addr));
8822 }
8823
8824 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
8825 MemRegion span,
8826 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
8827 CMSMarkStack* revisit_stack, bool cpc):
8828 KlassRememberingOopClosure(collector, NULL, revisit_stack),
8829 _span(span),
8830 _bit_map(bit_map),
8831 _mark_stack(mark_stack),
8832 _concurrent_precleaning(cpc) {
8833 assert(!_span.is_empty(), "Empty span could spell trouble");
8834 }
8835
8836
8837 // CMSKeepAliveClosure: the serial version
8838 void CMSKeepAliveClosure::do_oop(oop obj) {
8839 HeapWord* addr = (HeapWord*)obj;
8840 if (_span.contains(addr) &&
8841 !_bit_map->isMarked(addr)) {
8842 _bit_map->mark(addr);
8843 bool simulate_overflow = false;
8844 NOT_PRODUCT(
8845 if (CMSMarkStackOverflowALot &&
8846 _collector->simulate_overflow()) {
8847 // simulate a stack overflow
8848 simulate_overflow = true;
8849 }
8850 )
8851 if (simulate_overflow || !_mark_stack->push(obj)) {
8852 if (_concurrent_precleaning) {
8853 // We dirty the overflown object and let the remark
8854 // phase deal with it.
8855 assert(_collector->overflow_list_is_empty(), "Error");
8856 // In the case of object arrays, we need to dirty all of
8857 // the cards that the object spans. No locking or atomics
8858 // are needed since no one else can be mutating the mod union
8859 // table.
8860 if (obj->is_objArray()) {
8861 size_t sz = obj->size();
8862 HeapWord* end_card_addr =
8863 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
8864 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8865 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8866 _collector->_modUnionTable.mark_range(redirty_range);
8867 } else {
8868 _collector->_modUnionTable.mark(addr);
8869 }
8870 _collector->_ser_kac_preclean_ovflw++;
8871 } else {
8872 _collector->push_on_overflow_list(obj);
8873 _collector->_ser_kac_ovflw++;
8874 }
8875 }
8876 }
8877 }
8878
8879 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8880 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8881
8882 // CMSParKeepAliveClosure: a parallel version of the above.
8883 // The work queues are private to each closure (thread),
8884 // but (may be) available for stealing by other threads.
8885 void CMSParKeepAliveClosure::do_oop(oop obj) {
8886 HeapWord* addr = (HeapWord*)obj;
8887 if (_span.contains(addr) &&
8888 !_bit_map->isMarked(addr)) {
8889 // In general, during recursive tracing, several threads
8890 // may be concurrently getting here; the first one to
8891 // "tag" it, claims it.
8892 if (_bit_map->par_mark(addr)) {
8893 bool res = _work_queue->push(obj);
8894 assert(res, "Low water mark should be much less than capacity");
8895 // Do a recursive trim in the hope that this will keep
8896 // stack usage lower, but leave some oops for potential stealers
8897 trim_queue(_low_water_mark);
8898 } // Else, another thread got there first
8899 }
8900 }
8901
8902 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8903 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8904
8905 void CMSParKeepAliveClosure::trim_queue(uint max) {
8906 while (_work_queue->size() > max) {
8907 oop new_oop;
8908 if (_work_queue->pop_local(new_oop)) {
8909 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8910 assert(_bit_map->isMarked((HeapWord*)new_oop),
8911 "no white objects on this stack!");
8912 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8913 // iterate over the oops in this oop, marking and pushing
8914 // the ones in CMS heap (i.e. in _span).
8915 new_oop->oop_iterate(&_mark_and_push);
8916 }
8917 }
8918 }
8919
8920 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8921 CMSCollector* collector,
8922 MemRegion span, CMSBitMap* bit_map,
8923 CMSMarkStack* revisit_stack,
8924 OopTaskQueue* work_queue):
8925 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
8926 _span(span),
8927 _bit_map(bit_map),
8928 _work_queue(work_queue) { }
8929
8930 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8931 HeapWord* addr = (HeapWord*)obj;
8932 if (_span.contains(addr) &&
8933 !_bit_map->isMarked(addr)) {
8934 if (_bit_map->par_mark(addr)) {
8935 bool simulate_overflow = false;
8936 NOT_PRODUCT(
8937 if (CMSMarkStackOverflowALot &&
8938 _collector->par_simulate_overflow()) {
8939 // simulate a stack overflow
8940 simulate_overflow = true;
8941 }
8942 )
8943 if (simulate_overflow || !_work_queue->push(obj)) {
8944 _collector->par_push_on_overflow_list(obj);
8945 _collector->_par_kac_ovflw++;
8946 }
8947 } // Else another thread got there already
8948 }
8949 }
8950
8951 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8952 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8953
8954 //////////////////////////////////////////////////////////////////
8955 // CMSExpansionCause /////////////////////////////
8956 //////////////////////////////////////////////////////////////////
8957 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8958 switch (cause) {
8959 case _no_expansion:
8960 return "No expansion";
8961 case _satisfy_free_ratio:
8962 return "Free ratio";
8963 case _satisfy_promotion:
8964 return "Satisfy promotion";
8965 case _satisfy_allocation:
8966 return "allocation";
8967 case _allocate_par_lab:
8968 return "Par LAB";
8969 case _allocate_par_spooling_space:
8970 return "Par Spooling Space";
8971 case _adaptive_size_policy:
8972 return "Ergonomics";
8973 default:
8974 return "unknown";
8975 }
8976 }
8977
8978 void CMSDrainMarkingStackClosure::do_void() {
8979 // the max number to take from overflow list at a time
8980 const size_t num = _mark_stack->capacity()/4;
8981 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8982 "Overflow list should be NULL during concurrent phases");
8983 while (!_mark_stack->isEmpty() ||
8984 // if stack is empty, check the overflow list
8985 _collector->take_from_overflow_list(num, _mark_stack)) {
8986 oop obj = _mark_stack->pop();
8987 HeapWord* addr = (HeapWord*)obj;
8988 assert(_span.contains(addr), "Should be within span");
8989 assert(_bit_map->isMarked(addr), "Should be marked");
8990 assert(obj->is_oop(), "Should be an oop");
8991 obj->oop_iterate(_keep_alive);
8992 }
8993 }
8994
8995 void CMSParDrainMarkingStackClosure::do_void() {
8996 // drain queue
8997 trim_queue(0);
8998 }
8999
9000 // Trim our work_queue so its length is below max at return
9001 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
9002 while (_work_queue->size() > max) {
9003 oop new_oop;
9004 if (_work_queue->pop_local(new_oop)) {
9005 assert(new_oop->is_oop(), "Expected an oop");
9006 assert(_bit_map->isMarked((HeapWord*)new_oop),
9007 "no white objects on this stack!");
9008 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
9009 // iterate over the oops in this oop, marking and pushing
9010 // the ones in CMS heap (i.e. in _span).
9011 new_oop->oop_iterate(&_mark_and_push);
9012 }
9013 }
9014 }
9015
9016 ////////////////////////////////////////////////////////////////////
9017 // Support for Marking Stack Overflow list handling and related code
9018 ////////////////////////////////////////////////////////////////////
9019 // Much of the following code is similar in shape and spirit to the
9020 // code used in ParNewGC. We should try and share that code
9021 // as much as possible in the future.
9022
9023 #ifndef PRODUCT
9024 // Debugging support for CMSStackOverflowALot
9025
9026 // It's OK to call this multi-threaded; the worst thing
9027 // that can happen is that we'll get a bunch of closely
9028 // spaced simulated oveflows, but that's OK, in fact
9029 // probably good as it would exercise the overflow code
9030 // under contention.
9031 bool CMSCollector::simulate_overflow() {
9032 if (_overflow_counter-- <= 0) { // just being defensive
9033 _overflow_counter = CMSMarkStackOverflowInterval;
9034 return true;
9035 } else {
9036 return false;
9037 }
9038 }
9039
9040 bool CMSCollector::par_simulate_overflow() {
9041 return simulate_overflow();
9042 }
9043 #endif
9044
9045 // Single-threaded
9046 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
9047 assert(stack->isEmpty(), "Expected precondition");
9048 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
9049 size_t i = num;
9050 oop cur = _overflow_list;
9051 const markOop proto = markOopDesc::prototype();
9052 NOT_PRODUCT(ssize_t n = 0;)
9053 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
9054 next = oop(cur->mark());
9055 cur->set_mark(proto); // until proven otherwise
9056 assert(cur->is_oop(), "Should be an oop");
9057 bool res = stack->push(cur);
9058 assert(res, "Bit off more than can chew?");
9059 NOT_PRODUCT(n++;)
9060 }
9061 _overflow_list = cur;
9062 #ifndef PRODUCT
9063 assert(_num_par_pushes >= n, "Too many pops?");
9064 _num_par_pushes -=n;
9065 #endif
9066 return !stack->isEmpty();
9067 }
9068
9069 #define BUSY (oop(0x1aff1aff))
9070 // (MT-safe) Get a prefix of at most "num" from the list.
9071 // The overflow list is chained through the mark word of
9072 // each object in the list. We fetch the entire list,
9073 // break off a prefix of the right size and return the
9074 // remainder. If other threads try to take objects from
9075 // the overflow list at that time, they will wait for
9076 // some time to see if data becomes available. If (and
9077 // only if) another thread places one or more object(s)
9078 // on the global list before we have returned the suffix
9079 // to the global list, we will walk down our local list
9080 // to find its end and append the global list to
9081 // our suffix before returning it. This suffix walk can
9082 // prove to be expensive (quadratic in the amount of traffic)
9083 // when there are many objects in the overflow list and
9084 // there is much producer-consumer contention on the list.
9085 // *NOTE*: The overflow list manipulation code here and
9086 // in ParNewGeneration:: are very similar in shape,
9087 // except that in the ParNew case we use the old (from/eden)
9088 // copy of the object to thread the list via its klass word.
9089 // Because of the common code, if you make any changes in
9090 // the code below, please check the ParNew version to see if
9091 // similar changes might be needed.
9092 // CR 6797058 has been filed to consolidate the common code.
9093 bool CMSCollector::par_take_from_overflow_list(size_t num,
9094 OopTaskQueue* work_q,
9095 int no_of_gc_threads) {
9096 assert(work_q->size() == 0, "First empty local work queue");
9097 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
9098 if (_overflow_list == NULL) {
9099 return false;
9100 }
9101 // Grab the entire list; we'll put back a suffix
9102 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
9103 Thread* tid = Thread::current();
9104 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
9105 // set to ParallelGCThreads.
9106 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
9107 size_t sleep_time_millis = MAX2((size_t)1, num/100);
9108 // If the list is busy, we spin for a short while,
9109 // sleeping between attempts to get the list.
9110 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
9111 os::sleep(tid, sleep_time_millis, false);
9112 if (_overflow_list == NULL) {
9113 // Nothing left to take
9114 return false;
9115 } else if (_overflow_list != BUSY) {
9116 // Try and grab the prefix
9117 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
9118 }
9119 }
9120 // If the list was found to be empty, or we spun long
9121 // enough, we give up and return empty-handed. If we leave
9122 // the list in the BUSY state below, it must be the case that
9123 // some other thread holds the overflow list and will set it
9124 // to a non-BUSY state in the future.
9125 if (prefix == NULL || prefix == BUSY) {
9126 // Nothing to take or waited long enough
9127 if (prefix == NULL) {
9128 // Write back the NULL in case we overwrote it with BUSY above
9129 // and it is still the same value.
9130 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
9131 }
9132 return false;
9133 }
9134 assert(prefix != NULL && prefix != BUSY, "Error");
9135 size_t i = num;
9136 oop cur = prefix;
9137 // Walk down the first "num" objects, unless we reach the end.
9138 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
9139 if (cur->mark() == NULL) {
9140 // We have "num" or fewer elements in the list, so there
9141 // is nothing to return to the global list.
9142 // Write back the NULL in lieu of the BUSY we wrote
9143 // above, if it is still the same value.
9144 if (_overflow_list == BUSY) {
9145 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
9146 }
9147 } else {
9148 // Chop off the suffix and rerturn it to the global list.
9149 assert(cur->mark() != BUSY, "Error");
9150 oop suffix_head = cur->mark(); // suffix will be put back on global list
9151 cur->set_mark(NULL); // break off suffix
9152 // It's possible that the list is still in the empty(busy) state
9153 // we left it in a short while ago; in that case we may be
9154 // able to place back the suffix without incurring the cost
9155 // of a walk down the list.
9156 oop observed_overflow_list = _overflow_list;
9157 oop cur_overflow_list = observed_overflow_list;
9158 bool attached = false;
9159 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
9160 observed_overflow_list =
9161 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
9162 if (cur_overflow_list == observed_overflow_list) {
9163 attached = true;
9164 break;
9165 } else cur_overflow_list = observed_overflow_list;
9166 }
9167 if (!attached) {
9168 // Too bad, someone else sneaked in (at least) an element; we'll need
9169 // to do a splice. Find tail of suffix so we can prepend suffix to global
9170 // list.
9171 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
9172 oop suffix_tail = cur;
9173 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
9174 "Tautology");
9175 observed_overflow_list = _overflow_list;
9176 do {
9177 cur_overflow_list = observed_overflow_list;
9178 if (cur_overflow_list != BUSY) {
9179 // Do the splice ...
9180 suffix_tail->set_mark(markOop(cur_overflow_list));
9181 } else { // cur_overflow_list == BUSY
9182 suffix_tail->set_mark(NULL);
9183 }
9184 // ... and try to place spliced list back on overflow_list ...
9185 observed_overflow_list =
9186 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
9187 } while (cur_overflow_list != observed_overflow_list);
9188 // ... until we have succeeded in doing so.
9189 }
9190 }
9191
9192 // Push the prefix elements on work_q
9193 assert(prefix != NULL, "control point invariant");
9194 const markOop proto = markOopDesc::prototype();
9195 oop next;
9196 NOT_PRODUCT(ssize_t n = 0;)
9197 for (cur = prefix; cur != NULL; cur = next) {
9198 next = oop(cur->mark());
9199 cur->set_mark(proto); // until proven otherwise
9200 assert(cur->is_oop(), "Should be an oop");
9201 bool res = work_q->push(cur);
9202 assert(res, "Bit off more than we can chew?");
9203 NOT_PRODUCT(n++;)
9204 }
9205 #ifndef PRODUCT
9206 assert(_num_par_pushes >= n, "Too many pops?");
9207 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
9208 #endif
9209 return true;
9210 }
9211
9212 // Single-threaded
9213 void CMSCollector::push_on_overflow_list(oop p) {
9214 NOT_PRODUCT(_num_par_pushes++;)
9215 assert(p->is_oop(), "Not an oop");
9216 preserve_mark_if_necessary(p);
9217 p->set_mark((markOop)_overflow_list);
9218 _overflow_list = p;
9219 }
9220
9221 // Multi-threaded; use CAS to prepend to overflow list
9222 void CMSCollector::par_push_on_overflow_list(oop p) {
9223 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
9224 assert(p->is_oop(), "Not an oop");
9225 par_preserve_mark_if_necessary(p);
9226 oop observed_overflow_list = _overflow_list;
9227 oop cur_overflow_list;
9228 do {
9229 cur_overflow_list = observed_overflow_list;
9230 if (cur_overflow_list != BUSY) {
9231 p->set_mark(markOop(cur_overflow_list));
9232 } else {
9233 p->set_mark(NULL);
9234 }
9235 observed_overflow_list =
9236 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
9237 } while (cur_overflow_list != observed_overflow_list);
9238 }
9239 #undef BUSY
9240
9241 // Single threaded
9242 // General Note on GrowableArray: pushes may silently fail
9243 // because we are (temporarily) out of C-heap for expanding
9244 // the stack. The problem is quite ubiquitous and affects
9245 // a lot of code in the JVM. The prudent thing for GrowableArray
9246 // to do (for now) is to exit with an error. However, that may
9247 // be too draconian in some cases because the caller may be
9248 // able to recover without much harm. For such cases, we
9249 // should probably introduce a "soft_push" method which returns
9250 // an indication of success or failure with the assumption that
9251 // the caller may be able to recover from a failure; code in
9252 // the VM can then be changed, incrementally, to deal with such
9253 // failures where possible, thus, incrementally hardening the VM
9254 // in such low resource situations.
9255 void CMSCollector::preserve_mark_work(oop p, markOop m) {
9256 _preserved_oop_stack.push(p);
9257 _preserved_mark_stack.push(m);
9258 assert(m == p->mark(), "Mark word changed");
9259 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9260 "bijection");
9261 }
9262
9263 // Single threaded
9264 void CMSCollector::preserve_mark_if_necessary(oop p) {
9265 markOop m = p->mark();
9266 if (m->must_be_preserved(p)) {
9267 preserve_mark_work(p, m);
9268 }
9269 }
9270
9271 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
9272 markOop m = p->mark();
9273 if (m->must_be_preserved(p)) {
9274 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
9275 // Even though we read the mark word without holding
9276 // the lock, we are assured that it will not change
9277 // because we "own" this oop, so no other thread can
9278 // be trying to push it on the overflow list; see
9279 // the assertion in preserve_mark_work() that checks
9280 // that m == p->mark().
9281 preserve_mark_work(p, m);
9282 }
9283 }
9284
9285 // We should be able to do this multi-threaded,
9286 // a chunk of stack being a task (this is
9287 // correct because each oop only ever appears
9288 // once in the overflow list. However, it's
9289 // not very easy to completely overlap this with
9290 // other operations, so will generally not be done
9291 // until all work's been completed. Because we
9292 // expect the preserved oop stack (set) to be small,
9293 // it's probably fine to do this single-threaded.
9294 // We can explore cleverer concurrent/overlapped/parallel
9295 // processing of preserved marks if we feel the
9296 // need for this in the future. Stack overflow should
9297 // be so rare in practice and, when it happens, its
9298 // effect on performance so great that this will
9299 // likely just be in the noise anyway.
9300 void CMSCollector::restore_preserved_marks_if_any() {
9301 assert(SafepointSynchronize::is_at_safepoint(),
9302 "world should be stopped");
9303 assert(Thread::current()->is_ConcurrentGC_thread() ||
9304 Thread::current()->is_VM_thread(),
9305 "should be single-threaded");
9306 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9307 "bijection");
9308
9309 while (!_preserved_oop_stack.is_empty()) {
9310 oop p = _preserved_oop_stack.pop();
9311 assert(p->is_oop(), "Should be an oop");
9312 assert(_span.contains(p), "oop should be in _span");
9313 assert(p->mark() == markOopDesc::prototype(),
9314 "Set when taken from overflow list");
9315 markOop m = _preserved_mark_stack.pop();
9316 p->set_mark(m);
9317 }
9318 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
9319 "stacks were cleared above");
9320 }
9321
9322 #ifndef PRODUCT
9323 bool CMSCollector::no_preserved_marks() const {
9324 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
9325 }
9326 #endif
9327
9328 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
9329 {
9330 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
9331 CMSAdaptiveSizePolicy* size_policy =
9332 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
9333 assert(size_policy->is_gc_cms_adaptive_size_policy(),
9334 "Wrong type for size policy");
9335 return size_policy;
9336 }
9337
9338 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
9339 size_t desired_promo_size) {
9340 if (cur_promo_size < desired_promo_size) {
9341 size_t expand_bytes = desired_promo_size - cur_promo_size;
9342 if (PrintAdaptiveSizePolicy && Verbose) {
9343 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9344 "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
9345 expand_bytes);
9346 }
9347 expand(expand_bytes,
9348 MinHeapDeltaBytes,
9349 CMSExpansionCause::_adaptive_size_policy);
9350 } else if (desired_promo_size < cur_promo_size) {
9351 size_t shrink_bytes = cur_promo_size - desired_promo_size;
9352 if (PrintAdaptiveSizePolicy && Verbose) {
9353 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9354 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
9355 shrink_bytes);
9356 }
9357 shrink(shrink_bytes);
9358 }
9359 }
9360
9361 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
9362 GenCollectedHeap* gch = GenCollectedHeap::heap();
9363 CMSGCAdaptivePolicyCounters* counters =
9364 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
9365 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
9366 "Wrong kind of counters");
9367 return counters;
9368 }
9369
9370
9371 void ASConcurrentMarkSweepGeneration::update_counters() {
9372 if (UsePerfData) {
9373 _space_counters->update_all();
9374 _gen_counters->update_all();
9375 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9376 GenCollectedHeap* gch = GenCollectedHeap::heap();
9377 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9378 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9379 "Wrong gc statistics type");
9380 counters->update_counters(gc_stats_l);
9381 }
9382 }
9383
9384 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
9385 if (UsePerfData) {
9386 _space_counters->update_used(used);
9387 _space_counters->update_capacity();
9388 _gen_counters->update_all();
9389
9390 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9391 GenCollectedHeap* gch = GenCollectedHeap::heap();
9392 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9393 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9394 "Wrong gc statistics type");
9395 counters->update_counters(gc_stats_l);
9396 }
9397 }
9398
9399 // The desired expansion delta is computed so that:
9400 // . desired free percentage or greater is used
9401 void ASConcurrentMarkSweepGeneration::compute_new_size() {
9402 assert_locked_or_safepoint(Heap_lock);
9403
9404 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
9405
9406 // If incremental collection failed, we just want to expand
9407 // to the limit.
9408 if (incremental_collection_failed()) {
9409 clear_incremental_collection_failed();
9410 grow_to_reserved();
9411 return;
9412 }
9413
9414 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing");
9415
9416 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
9417 "Wrong type of heap");
9418 int prev_level = level() - 1;
9419 assert(prev_level >= 0, "The cms generation is the lowest generation");
9420 Generation* prev_gen = gch->get_gen(prev_level);
9421 assert(prev_gen->kind() == Generation::ASParNew,
9422 "Wrong type of young generation");
9423 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen;
9424 size_t cur_eden = younger_gen->eden()->capacity();
9425 CMSAdaptiveSizePolicy* size_policy = cms_size_policy();
9426 size_t cur_promo = free();
9427 size_policy->compute_tenured_generation_free_space(cur_promo,
9428 max_available(),
9429 cur_eden);
9430 resize(cur_promo, size_policy->promo_size());
9431
9432 // Record the new size of the space in the cms generation
9433 // that is available for promotions. This is temporary.
9434 // It should be the desired promo size.
9435 size_policy->avg_cms_promo()->sample(free());
9436 size_policy->avg_old_live()->sample(used());
9437
9438 if (UsePerfData) {
9439 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9440 counters->update_cms_capacity_counter(capacity());
9441 }
9442 }
9443
9444 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
9445 assert_locked_or_safepoint(Heap_lock);
9446 assert_lock_strong(freelistLock());
9447 HeapWord* old_end = _cmsSpace->end();
9448 HeapWord* unallocated_start = _cmsSpace->unallocated_block();
9449 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
9450 FreeChunk* chunk_at_end = find_chunk_at_end();
9451 if (chunk_at_end == NULL) {
9452 // No room to shrink
9453 if (PrintGCDetails && Verbose) {
9454 gclog_or_tty->print_cr("No room to shrink: old_end "
9455 PTR_FORMAT " unallocated_start " PTR_FORMAT
9456 " chunk_at_end " PTR_FORMAT,
9457 old_end, unallocated_start, chunk_at_end);
9458 }
9459 return;
9460 } else {
9461
9462 // Find the chunk at the end of the space and determine
9463 // how much it can be shrunk.
9464 size_t shrinkable_size_in_bytes = chunk_at_end->size();
9465 size_t aligned_shrinkable_size_in_bytes =
9466 align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
9467 assert(unallocated_start <= chunk_at_end->end(),
9468 "Inconsistent chunk at end of space");
9469 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
9470 size_t word_size_before = heap_word_size(_virtual_space.committed_size());
9471
9472 // Shrink the underlying space
9473 _virtual_space.shrink_by(bytes);
9474 if (PrintGCDetails && Verbose) {
9475 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
9476 " desired_bytes " SIZE_FORMAT
9477 " shrinkable_size_in_bytes " SIZE_FORMAT
9478 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
9479 " bytes " SIZE_FORMAT,
9480 desired_bytes, shrinkable_size_in_bytes,
9481 aligned_shrinkable_size_in_bytes, bytes);
9482 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT
9483 " unallocated_start " SIZE_FORMAT,
9484 old_end, unallocated_start);
9485 }
9486
9487 // If the space did shrink (shrinking is not guaranteed),
9488 // shrink the chunk at the end by the appropriate amount.
9489 if (((HeapWord*)_virtual_space.high()) < old_end) {
9490 size_t new_word_size =
9491 heap_word_size(_virtual_space.committed_size());
9492
9493 // Have to remove the chunk from the dictionary because it is changing
9494 // size and might be someplace elsewhere in the dictionary.
9495
9496 // Get the chunk at end, shrink it, and put it
9497 // back.
9498 _cmsSpace->removeChunkFromDictionary(chunk_at_end);
9499 size_t word_size_change = word_size_before - new_word_size;
9500 size_t chunk_at_end_old_size = chunk_at_end->size();
9501 assert(chunk_at_end_old_size >= word_size_change,
9502 "Shrink is too large");
9503 chunk_at_end->set_size(chunk_at_end_old_size -
9504 word_size_change);
9505 _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
9506 word_size_change);
9507
9508 _cmsSpace->returnChunkToDictionary(chunk_at_end);
9509
9510 MemRegion mr(_cmsSpace->bottom(), new_word_size);
9511 _bts->resize(new_word_size); // resize the block offset shared array
9512 Universe::heap()->barrier_set()->resize_covered_region(mr);
9513 _cmsSpace->assert_locked();
9514 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
9515
9516 NOT_PRODUCT(_cmsSpace->dictionary()->verify());
9517
9518 // update the space and generation capacity counters
9519 if (UsePerfData) {
9520 _space_counters->update_capacity();
9521 _gen_counters->update_all();
9522 }
9523
9524 if (Verbose && PrintGCDetails) {
9525 size_t new_mem_size = _virtual_space.committed_size();
9526 size_t old_mem_size = new_mem_size + bytes;
9527 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK",
9528 name(), old_mem_size/K, bytes/K, new_mem_size/K);
9529 }
9530 }
9531
9532 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
9533 "Inconsistency at end of space");
9534 assert(chunk_at_end->end() == _cmsSpace->end(),
9535 "Shrinking is inconsistent");
9536 return;
9537 }
9538 }
9539 // Transfer some number of overflown objects to usual marking
9540 // stack. Return true if some objects were transferred.
9541 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
9542 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
9543 (size_t)ParGCDesiredObjsFromOverflowList);
9544
9545 bool res = _collector->take_from_overflow_list(num, _mark_stack);
9546 assert(_collector->overflow_list_is_empty() || res,
9547 "If list is not empty, we should have taken something");
9548 assert(!res || !_mark_stack->isEmpty(),
9549 "If we took something, it should now be on our stack");
9550 return res;
9551 }
9552
9553 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
9554 size_t res = _sp->block_size_no_stall(addr, _collector);
9555 if (_sp->block_is_obj(addr)) {
9556 if (_live_bit_map->isMarked(addr)) {
9557 // It can't have been dead in a previous cycle
9558 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
9559 } else {
9560 _dead_bit_map->mark(addr); // mark the dead object
9561 }
9562 }
9563 // Could be 0, if the block size could not be computed without stalling.
9564 return res;
9565 }
9566
9567 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
9568
9569 switch (phase) {
9570 case CMSCollector::InitialMarking:
9571 initialize(true /* fullGC */ ,
9572 cause /* cause of the GC */,
9573 true /* recordGCBeginTime */,
9574 true /* recordPreGCUsage */,
9575 false /* recordPeakUsage */,
9576 false /* recordPostGCusage */,
9577 true /* recordAccumulatedGCTime */,
9578 false /* recordGCEndTime */,
9579 false /* countCollection */ );
9580 break;
9581
9582 case CMSCollector::FinalMarking:
9583 initialize(true /* fullGC */ ,
9584 cause /* cause of the GC */,
9585 false /* recordGCBeginTime */,
9586 false /* recordPreGCUsage */,
9587 false /* recordPeakUsage */,
9588 false /* recordPostGCusage */,
9589 true /* recordAccumulatedGCTime */,
9590 false /* recordGCEndTime */,
9591 false /* countCollection */ );
9592 break;
9593
9594 case CMSCollector::Sweeping:
9595 initialize(true /* fullGC */ ,
9596 cause /* cause of the GC */,
9597 false /* recordGCBeginTime */,
9598 false /* recordPreGCUsage */,
9599 true /* recordPeakUsage */,
9600 true /* recordPostGCusage */,
9601 false /* recordAccumulatedGCTime */,
9602 true /* recordGCEndTime */,
9603 true /* countCollection */ );
9604 break;
9605
9606 default:
9607 ShouldNotReachHere();
9608 }
9609 }