1 /*
2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/classLoaderData.hpp"
27 #include "classfile/symbolTable.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "code/codeCache.hpp"
30 #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp"
31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp"
32 #include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.hpp"
33 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp"
34 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp"
35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp"
36 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp"
37 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp"
38 #include "gc_implementation/parNew/parNewGeneration.hpp"
39 #include "gc_implementation/shared/collectorCounters.hpp"
40 #include "gc_implementation/shared/gcTimer.hpp"
41 #include "gc_implementation/shared/gcTrace.hpp"
42 #include "gc_implementation/shared/gcTraceTime.hpp"
43 #include "gc_implementation/shared/isGCActiveMark.hpp"
44 #include "gc_interface/collectedHeap.inline.hpp"
45 #include "memory/allocation.hpp"
46 #include "memory/cardTableRS.hpp"
47 #include "memory/collectorPolicy.hpp"
48 #include "memory/gcLocker.inline.hpp"
49 #include "memory/genCollectedHeap.hpp"
50 #include "memory/genMarkSweep.hpp"
51 #include "memory/genOopClosures.inline.hpp"
52 #include "memory/iterator.inline.hpp"
53 #include "memory/padded.hpp"
54 #include "memory/referencePolicy.hpp"
55 #include "memory/resourceArea.hpp"
56 #include "memory/tenuredGeneration.hpp"
57 #include "oops/oop.inline.hpp"
58 #include "prims/jvmtiExport.hpp"
59 #include "runtime/globals_extension.hpp"
60 #include "runtime/handles.inline.hpp"
61 #include "runtime/java.hpp"
62 #include "runtime/orderAccess.inline.hpp"
63 #include "runtime/vmThread.hpp"
64 #include "services/memoryService.hpp"
65 #include "services/runtimeService.hpp"
66
67 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
68
69 // statics
70 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
71 bool CMSCollector::_full_gc_requested = false;
72 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
73
74 //////////////////////////////////////////////////////////////////
75 // In support of CMS/VM thread synchronization
76 //////////////////////////////////////////////////////////////////
77 // We split use of the CGC_lock into 2 "levels".
78 // The low-level locking is of the usual CGC_lock monitor. We introduce
79 // a higher level "token" (hereafter "CMS token") built on top of the
80 // low level monitor (hereafter "CGC lock").
81 // The token-passing protocol gives priority to the VM thread. The
82 // CMS-lock doesn't provide any fairness guarantees, but clients
83 // should ensure that it is only held for very short, bounded
84 // durations.
85 //
86 // When either of the CMS thread or the VM thread is involved in
87 // collection operations during which it does not want the other
88 // thread to interfere, it obtains the CMS token.
89 //
90 // If either thread tries to get the token while the other has
91 // it, that thread waits. However, if the VM thread and CMS thread
92 // both want the token, then the VM thread gets priority while the
93 // CMS thread waits. This ensures, for instance, that the "concurrent"
94 // phases of the CMS thread's work do not block out the VM thread
95 // for long periods of time as the CMS thread continues to hog
96 // the token. (See bug 4616232).
97 //
98 // The baton-passing functions are, however, controlled by the
99 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
100 // and here the low-level CMS lock, not the high level token,
101 // ensures mutual exclusion.
102 //
103 // Two important conditions that we have to satisfy:
104 // 1. if a thread does a low-level wait on the CMS lock, then it
105 // relinquishes the CMS token if it were holding that token
106 // when it acquired the low-level CMS lock.
107 // 2. any low-level notifications on the low-level lock
108 // should only be sent when a thread has relinquished the token.
109 //
110 // In the absence of either property, we'd have potential deadlock.
111 //
112 // We protect each of the CMS (concurrent and sequential) phases
113 // with the CMS _token_, not the CMS _lock_.
114 //
115 // The only code protected by CMS lock is the token acquisition code
116 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
117 // baton-passing code.
118 //
119 // Unfortunately, i couldn't come up with a good abstraction to factor and
120 // hide the naked CGC_lock manipulation in the baton-passing code
121 // further below. That's something we should try to do. Also, the proof
122 // of correctness of this 2-level locking scheme is far from obvious,
123 // and potentially quite slippery. We have an uneasy supsicion, for instance,
124 // that there may be a theoretical possibility of delay/starvation in the
125 // low-level lock/wait/notify scheme used for the baton-passing because of
126 // potential intereference with the priority scheme embodied in the
127 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
128 // invocation further below and marked with "XXX 20011219YSR".
129 // Indeed, as we note elsewhere, this may become yet more slippery
130 // in the presence of multiple CMS and/or multiple VM threads. XXX
131
132 class CMSTokenSync: public StackObj {
133 private:
134 bool _is_cms_thread;
135 public:
136 CMSTokenSync(bool is_cms_thread):
137 _is_cms_thread(is_cms_thread) {
138 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
139 "Incorrect argument to constructor");
140 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
141 }
142
143 ~CMSTokenSync() {
144 assert(_is_cms_thread ?
145 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
146 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
147 "Incorrect state");
148 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
149 }
150 };
151
152 // Convenience class that does a CMSTokenSync, and then acquires
153 // upto three locks.
154 class CMSTokenSyncWithLocks: public CMSTokenSync {
155 private:
156 // Note: locks are acquired in textual declaration order
157 // and released in the opposite order
158 MutexLockerEx _locker1, _locker2, _locker3;
159 public:
160 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
161 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
162 CMSTokenSync(is_cms_thread),
163 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
164 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
165 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
166 { }
167 };
168
169
170 // Wrapper class to temporarily disable icms during a foreground cms collection.
171 class ICMSDisabler: public StackObj {
172 public:
173 // The ctor disables icms and wakes up the thread so it notices the change;
174 // the dtor re-enables icms. Note that the CMSCollector methods will check
175 // CMSIncrementalMode.
176 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); }
177 ~ICMSDisabler() { CMSCollector::enable_icms(); }
178 };
179
180 //////////////////////////////////////////////////////////////////
181 // Concurrent Mark-Sweep Generation /////////////////////////////
182 //////////////////////////////////////////////////////////////////
183
184 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
185
186 // This struct contains per-thread things necessary to support parallel
187 // young-gen collection.
188 class CMSParGCThreadState: public CHeapObj<mtGC> {
189 public:
190 CFLS_LAB lab;
191 PromotionInfo promo;
192
193 // Constructor.
194 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
195 promo.setSpace(cfls);
196 }
197 };
198
199 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
200 ReservedSpace rs, size_t initial_byte_size, int level,
201 CardTableRS* ct, bool use_adaptive_freelists,
202 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) :
203 CardGeneration(rs, initial_byte_size, level, ct),
204 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
205 _debug_collection_type(Concurrent_collection_type),
206 _did_compact(false)
207 {
208 HeapWord* bottom = (HeapWord*) _virtual_space.low();
209 HeapWord* end = (HeapWord*) _virtual_space.high();
210
211 _direct_allocated_words = 0;
212 NOT_PRODUCT(
213 _numObjectsPromoted = 0;
214 _numWordsPromoted = 0;
215 _numObjectsAllocated = 0;
216 _numWordsAllocated = 0;
217 )
218
219 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
220 use_adaptive_freelists,
221 dictionaryChoice);
222 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
223 if (_cmsSpace == NULL) {
224 vm_exit_during_initialization(
225 "CompactibleFreeListSpace allocation failure");
226 }
227 _cmsSpace->_gen = this;
228
229 _gc_stats = new CMSGCStats();
230
231 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
232 // offsets match. The ability to tell free chunks from objects
233 // depends on this property.
234 debug_only(
235 FreeChunk* junk = NULL;
236 assert(UseCompressedClassPointers ||
237 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
238 "Offset of FreeChunk::_prev within FreeChunk must match"
239 " that of OopDesc::_klass within OopDesc");
240 )
241 if (CollectedHeap::use_parallel_gc_threads()) {
242 typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
243 _par_gc_thread_states =
244 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC);
245 if (_par_gc_thread_states == NULL) {
246 vm_exit_during_initialization("Could not allocate par gc structs");
247 }
248 for (uint i = 0; i < ParallelGCThreads; i++) {
249 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
250 if (_par_gc_thread_states[i] == NULL) {
251 vm_exit_during_initialization("Could not allocate par gc structs");
252 }
253 }
254 } else {
255 _par_gc_thread_states = NULL;
256 }
257 _incremental_collection_failed = false;
258 // The "dilatation_factor" is the expansion that can occur on
259 // account of the fact that the minimum object size in the CMS
260 // generation may be larger than that in, say, a contiguous young
261 // generation.
262 // Ideally, in the calculation below, we'd compute the dilatation
263 // factor as: MinChunkSize/(promoting_gen's min object size)
264 // Since we do not have such a general query interface for the
265 // promoting generation, we'll instead just use the mimimum
266 // object size (which today is a header's worth of space);
267 // note that all arithmetic is in units of HeapWords.
268 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
269 assert(_dilatation_factor >= 1.0, "from previous assert");
270 }
271
272
273 // The field "_initiating_occupancy" represents the occupancy percentage
274 // at which we trigger a new collection cycle. Unless explicitly specified
275 // via CMSInitiatingOccupancyFraction (argument "io" below), it
276 // is calculated by:
277 //
278 // Let "f" be MinHeapFreeRatio in
279 //
280 // _intiating_occupancy = 100-f +
281 // f * (CMSTriggerRatio/100)
282 // where CMSTriggerRatio is the argument "tr" below.
283 //
284 // That is, if we assume the heap is at its desired maximum occupancy at the
285 // end of a collection, we let CMSTriggerRatio of the (purported) free
286 // space be allocated before initiating a new collection cycle.
287 //
288 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
289 assert(io <= 100 && tr <= 100, "Check the arguments");
290 if (io >= 0) {
291 _initiating_occupancy = (double)io / 100.0;
292 } else {
293 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
294 (double)(tr * MinHeapFreeRatio) / 100.0)
295 / 100.0;
296 }
297 }
298
299 void ConcurrentMarkSweepGeneration::ref_processor_init() {
300 assert(collector() != NULL, "no collector");
301 collector()->ref_processor_init();
302 }
303
304 void CMSCollector::ref_processor_init() {
305 if (_ref_processor == NULL) {
306 // Allocate and initialize a reference processor
307 _ref_processor =
308 new ReferenceProcessor(_span, // span
309 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
310 (int) ParallelGCThreads, // mt processing degree
311 _cmsGen->refs_discovery_is_mt(), // mt discovery
312 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
313 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
314 &_is_alive_closure); // closure for liveness info
315 // Initialize the _ref_processor field of CMSGen
316 _cmsGen->set_ref_processor(_ref_processor);
317
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 CardTableRS* ct,
555 ConcurrentMarkSweepPolicy* cp):
556 _cmsGen(cmsGen),
557 _ct(ct),
558 _ref_processor(NULL), // will be set later
559 _conc_workers(NULL), // may be set later
560 _abort_preclean(false),
561 _start_sampling(false),
562 _between_prologue_and_epilogue(false),
563 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
564 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
565 -1 /* lock-free */, "No_lock" /* dummy */),
566 _modUnionClosure(&_modUnionTable),
567 _modUnionClosurePar(&_modUnionTable),
568 // Adjust my span to cover old (cms) gen
569 _span(cmsGen->reserved()),
570 // Construct the is_alive_closure with _span & markBitMap
571 _is_alive_closure(_span, &_markBitMap),
572 _restart_addr(NULL),
573 _overflow_list(NULL),
574 _stats(cmsGen),
575 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true)),
576 _eden_chunk_array(NULL), // may be set in ctor body
577 _eden_chunk_capacity(0), // -- ditto --
578 _eden_chunk_index(0), // -- ditto --
579 _survivor_plab_array(NULL), // -- ditto --
580 _survivor_chunk_array(NULL), // -- ditto --
581 _survivor_chunk_capacity(0), // -- ditto --
582 _survivor_chunk_index(0), // -- ditto --
583 _ser_pmc_preclean_ovflw(0),
584 _ser_kac_preclean_ovflw(0),
585 _ser_pmc_remark_ovflw(0),
586 _par_pmc_remark_ovflw(0),
587 _ser_kac_ovflw(0),
588 _par_kac_ovflw(0),
589 #ifndef PRODUCT
590 _num_par_pushes(0),
591 #endif
592 _collection_count_start(0),
593 _verifying(false),
594 _icms_start_limit(NULL),
595 _icms_stop_limit(NULL),
596 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
597 _completed_initialization(false),
598 _collector_policy(cp),
599 _should_unload_classes(CMSClassUnloadingEnabled),
600 _concurrent_cycles_since_last_unload(0),
601 _roots_scanning_options(GenCollectedHeap::SO_None),
602 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
603 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
604 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
605 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
606 _cms_start_registered(false)
607 {
608 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
609 ExplicitGCInvokesConcurrent = true;
610 }
611 // Now expand the span and allocate the collection support structures
612 // (MUT, marking bit map etc.) to cover both generations subject to
613 // collection.
614
615 // For use by dirty card to oop closures.
616 _cmsGen->cmsSpace()->set_collector(this);
617
618 // Allocate MUT and marking bit map
619 {
620 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
621 if (!_markBitMap.allocate(_span)) {
622 warning("Failed to allocate CMS Bit Map");
623 return;
624 }
625 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
626 }
627 {
628 _modUnionTable.allocate(_span);
629 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
630 }
631
632 if (!_markStack.allocate(MarkStackSize)) {
633 warning("Failed to allocate CMS Marking Stack");
634 return;
635 }
636
637 // Support for multi-threaded concurrent phases
638 if (CMSConcurrentMTEnabled) {
639 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
640 // just for now
641 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
642 }
643 if (ConcGCThreads > 1) {
644 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
645 ConcGCThreads, true);
646 if (_conc_workers == NULL) {
647 warning("GC/CMS: _conc_workers allocation failure: "
648 "forcing -CMSConcurrentMTEnabled");
649 CMSConcurrentMTEnabled = false;
650 } else {
651 _conc_workers->initialize_workers();
652 }
653 } else {
654 CMSConcurrentMTEnabled = false;
655 }
656 }
657 if (!CMSConcurrentMTEnabled) {
658 ConcGCThreads = 0;
659 } else {
660 // Turn off CMSCleanOnEnter optimization temporarily for
661 // the MT case where it's not fixed yet; see 6178663.
662 CMSCleanOnEnter = false;
663 }
664 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
665 "Inconsistency");
666
667 // Parallel task queues; these are shared for the
668 // concurrent and stop-world phases of CMS, but
669 // are not shared with parallel scavenge (ParNew).
670 {
671 uint i;
672 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
673
674 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
675 || ParallelRefProcEnabled)
676 && num_queues > 0) {
677 _task_queues = new OopTaskQueueSet(num_queues);
678 if (_task_queues == NULL) {
679 warning("task_queues allocation failure.");
680 return;
681 }
682 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
683 if (_hash_seed == NULL) {
684 warning("_hash_seed array allocation failure");
685 return;
686 }
687
688 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
689 for (i = 0; i < num_queues; i++) {
690 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
691 if (q == NULL) {
692 warning("work_queue allocation failure.");
693 return;
694 }
695 _task_queues->register_queue(i, q);
696 }
697 for (i = 0; i < num_queues; i++) {
698 _task_queues->queue(i)->initialize();
699 _hash_seed[i] = 17; // copied from ParNew
700 }
701 }
702 }
703
704 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
705
706 // Clip CMSBootstrapOccupancy between 0 and 100.
707 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100;
708
709 _full_gcs_since_conc_gc = 0;
710
711 // Now tell CMS generations the identity of their collector
712 ConcurrentMarkSweepGeneration::set_collector(this);
713
714 // Create & start a CMS thread for this CMS collector
715 _cmsThread = ConcurrentMarkSweepThread::start(this);
716 assert(cmsThread() != NULL, "CMS Thread should have been created");
717 assert(cmsThread()->collector() == this,
718 "CMS Thread should refer to this gen");
719 assert(CGC_lock != NULL, "Where's the CGC_lock?");
720
721 // Support for parallelizing young gen rescan
722 GenCollectedHeap* gch = GenCollectedHeap::heap();
723 _young_gen = gch->prev_gen(_cmsGen);
724 if (gch->supports_inline_contig_alloc()) {
725 _top_addr = gch->top_addr();
726 _end_addr = gch->end_addr();
727 assert(_young_gen != NULL, "no _young_gen");
728 _eden_chunk_index = 0;
729 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
730 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
731 if (_eden_chunk_array == NULL) {
732 _eden_chunk_capacity = 0;
733 warning("GC/CMS: _eden_chunk_array allocation failure");
734 }
735 }
736 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
737
738 // Support for parallelizing survivor space rescan
739 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
740 const size_t max_plab_samples =
741 ((DefNewGeneration*)_young_gen)->max_survivor_size() / plab_sample_minimum_size();
742
743 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
744 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC);
745 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
746 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
747 || _cursor == NULL) {
748 warning("Failed to allocate survivor plab/chunk array");
749 if (_survivor_plab_array != NULL) {
750 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC);
751 _survivor_plab_array = NULL;
752 }
753 if (_survivor_chunk_array != NULL) {
754 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC);
755 _survivor_chunk_array = NULL;
756 }
757 if (_cursor != NULL) {
758 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC);
759 _cursor = NULL;
760 }
761 } else {
762 _survivor_chunk_capacity = 2*max_plab_samples;
763 for (uint i = 0; i < ParallelGCThreads; i++) {
764 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
765 if (vec == NULL) {
766 warning("Failed to allocate survivor plab array");
767 for (int j = i; j > 0; j--) {
768 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC);
769 }
770 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC);
771 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC);
772 _survivor_plab_array = NULL;
773 _survivor_chunk_array = NULL;
774 _survivor_chunk_capacity = 0;
775 break;
776 } else {
777 ChunkArray* cur =
778 ::new (&_survivor_plab_array[i]) ChunkArray(vec,
779 max_plab_samples);
780 assert(cur->end() == 0, "Should be 0");
781 assert(cur->array() == vec, "Should be vec");
782 assert(cur->capacity() == max_plab_samples, "Error");
783 }
784 }
785 }
786 }
787 assert( ( _survivor_plab_array != NULL
788 && _survivor_chunk_array != NULL)
789 || ( _survivor_chunk_capacity == 0
790 && _survivor_chunk_index == 0),
791 "Error");
792
793 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
794 _gc_counters = new CollectorCounters("CMS", 1);
795 _completed_initialization = true;
796 _inter_sweep_timer.start(); // start of time
797 }
798
799 size_t CMSCollector::plab_sample_minimum_size() {
800 // The default value of MinTLABSize is 2k, but there is
801 // no way to get the default value if the flag has been overridden.
802 return MAX2(ThreadLocalAllocBuffer::min_size() * HeapWordSize, 2 * K);
803 }
804
805 const char* ConcurrentMarkSweepGeneration::name() const {
806 return "concurrent mark-sweep generation";
807 }
808 void ConcurrentMarkSweepGeneration::update_counters() {
809 if (UsePerfData) {
810 _space_counters->update_all();
811 _gen_counters->update_all();
812 }
813 }
814
815 // this is an optimized version of update_counters(). it takes the
816 // used value as a parameter rather than computing it.
817 //
818 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
819 if (UsePerfData) {
820 _space_counters->update_used(used);
821 _space_counters->update_capacity();
822 _gen_counters->update_all();
823 }
824 }
825
826 void ConcurrentMarkSweepGeneration::print() const {
827 Generation::print();
828 cmsSpace()->print();
829 }
830
831 #ifndef PRODUCT
832 void ConcurrentMarkSweepGeneration::print_statistics() {
833 cmsSpace()->printFLCensus(0);
834 }
835 #endif
836
837 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
838 GenCollectedHeap* gch = GenCollectedHeap::heap();
839 if (PrintGCDetails) {
840 if (Verbose) {
841 gclog_or_tty->print("[%d %s-%s: " SIZE_FORMAT "(" SIZE_FORMAT ")]",
842 level(), short_name(), s, used(), capacity());
843 } else {
844 gclog_or_tty->print("[%d %s-%s: " SIZE_FORMAT "K(" SIZE_FORMAT "K)]",
845 level(), short_name(), s, used() / K, capacity() / K);
846 }
847 }
848 if (Verbose) {
849 gclog_or_tty->print(" " SIZE_FORMAT "(" SIZE_FORMAT ")",
850 gch->used(), gch->capacity());
851 } else {
852 gclog_or_tty->print(" " SIZE_FORMAT "K(" SIZE_FORMAT "K)",
853 gch->used() / K, gch->capacity() / K);
854 }
855 }
856
857 size_t
858 ConcurrentMarkSweepGeneration::contiguous_available() const {
859 // dld proposes an improvement in precision here. If the committed
860 // part of the space ends in a free block we should add that to
861 // uncommitted size in the calculation below. Will make this
862 // change later, staying with the approximation below for the
863 // time being. -- ysr.
864 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
865 }
866
867 size_t
868 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
869 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
870 }
871
872 size_t ConcurrentMarkSweepGeneration::max_available() const {
873 return free() + _virtual_space.uncommitted_size();
874 }
875
876 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
877 size_t available = max_available();
878 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
879 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
880 if (Verbose && PrintGCDetails) {
881 gclog_or_tty->print_cr(
882 "CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "),"
883 "max_promo(" SIZE_FORMAT ")",
884 res? "":" not", available, res? ">=":"<",
885 av_promo, max_promotion_in_bytes);
886 }
887 return res;
888 }
889
890 // At a promotion failure dump information on block layout in heap
891 // (cms old generation).
892 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
893 if (CMSDumpAtPromotionFailure) {
894 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
895 }
896 }
897
898 CompactibleSpace*
899 ConcurrentMarkSweepGeneration::first_compaction_space() const {
900 return _cmsSpace;
901 }
902
903 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
904 // Clear the promotion information. These pointers can be adjusted
905 // along with all the other pointers into the heap but
906 // compaction is expected to be a rare event with
907 // a heap using cms so don't do it without seeing the need.
908 if (CollectedHeap::use_parallel_gc_threads()) {
909 for (uint i = 0; i < ParallelGCThreads; i++) {
910 _par_gc_thread_states[i]->promo.reset();
911 }
912 }
913 }
914
915 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
916 blk->do_space(_cmsSpace);
917 }
918
919 void ConcurrentMarkSweepGeneration::compute_new_size() {
920 assert_locked_or_safepoint(Heap_lock);
921
922 // If incremental collection failed, we just want to expand
923 // to the limit.
924 if (incremental_collection_failed()) {
925 clear_incremental_collection_failed();
926 grow_to_reserved();
927 return;
928 }
929
930 // The heap has been compacted but not reset yet.
931 // Any metric such as free() or used() will be incorrect.
932
933 CardGeneration::compute_new_size();
934
935 // Reset again after a possible resizing
936 if (did_compact()) {
937 cmsSpace()->reset_after_compaction();
938 }
939 }
940
941 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
942 assert_locked_or_safepoint(Heap_lock);
943
944 // If incremental collection failed, we just want to expand
945 // to the limit.
946 if (incremental_collection_failed()) {
947 clear_incremental_collection_failed();
948 grow_to_reserved();
949 return;
950 }
951
952 double free_percentage = ((double) free()) / capacity();
953 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
954 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
955
956 // compute expansion delta needed for reaching desired free percentage
957 if (free_percentage < desired_free_percentage) {
958 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
959 assert(desired_capacity >= capacity(), "invalid expansion size");
960 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
961 if (PrintGCDetails && Verbose) {
962 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
963 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
964 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
965 gclog_or_tty->print_cr(" Desired free fraction %f",
966 desired_free_percentage);
967 gclog_or_tty->print_cr(" Maximum free fraction %f",
968 maximum_free_percentage);
969 gclog_or_tty->print_cr(" Capactiy " SIZE_FORMAT, capacity()/1000);
970 gclog_or_tty->print_cr(" Desired capacity " SIZE_FORMAT,
971 desired_capacity/1000);
972 int prev_level = level() - 1;
973 if (prev_level >= 0) {
974 size_t prev_size = 0;
975 GenCollectedHeap* gch = GenCollectedHeap::heap();
976 Generation* prev_gen = gch->_gens[prev_level];
977 prev_size = prev_gen->capacity();
978 gclog_or_tty->print_cr(" Younger gen size " SIZE_FORMAT,
979 prev_size/1000);
980 }
981 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc " SIZE_FORMAT,
982 unsafe_max_alloc_nogc()/1000);
983 gclog_or_tty->print_cr(" contiguous available " SIZE_FORMAT,
984 contiguous_available()/1000);
985 gclog_or_tty->print_cr(" Expand by " SIZE_FORMAT " (bytes)",
986 expand_bytes);
987 }
988 // safe if expansion fails
989 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
990 if (PrintGCDetails && Verbose) {
991 gclog_or_tty->print_cr(" Expanded free fraction %f",
992 ((double) free()) / capacity());
993 }
994 } else {
995 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
996 assert(desired_capacity <= capacity(), "invalid expansion size");
997 size_t shrink_bytes = capacity() - desired_capacity;
998 // Don't shrink unless the delta is greater than the minimum shrink we want
999 if (shrink_bytes >= MinHeapDeltaBytes) {
1000 shrink_free_list_by(shrink_bytes);
1001 }
1002 }
1003 }
1004
1005 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
1006 return cmsSpace()->freelistLock();
1007 }
1008
1009 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
1010 bool tlab) {
1011 CMSSynchronousYieldRequest yr;
1012 MutexLockerEx x(freelistLock(),
1013 Mutex::_no_safepoint_check_flag);
1014 return have_lock_and_allocate(size, tlab);
1015 }
1016
1017 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
1018 bool tlab /* ignored */) {
1019 assert_lock_strong(freelistLock());
1020 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
1021 HeapWord* res = cmsSpace()->allocate(adjustedSize);
1022 // Allocate the object live (grey) if the background collector has
1023 // started marking. This is necessary because the marker may
1024 // have passed this address and consequently this object will
1025 // not otherwise be greyed and would be incorrectly swept up.
1026 // Note that if this object contains references, the writing
1027 // of those references will dirty the card containing this object
1028 // allowing the object to be blackened (and its references scanned)
1029 // either during a preclean phase or at the final checkpoint.
1030 if (res != NULL) {
1031 // We may block here with an uninitialized object with
1032 // its mark-bit or P-bits not yet set. Such objects need
1033 // to be safely navigable by block_start().
1034 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
1035 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
1036 collector()->direct_allocated(res, adjustedSize);
1037 _direct_allocated_words += adjustedSize;
1038 // allocation counters
1039 NOT_PRODUCT(
1040 _numObjectsAllocated++;
1041 _numWordsAllocated += (int)adjustedSize;
1042 )
1043 }
1044 return res;
1045 }
1046
1047 // In the case of direct allocation by mutators in a generation that
1048 // is being concurrently collected, the object must be allocated
1049 // live (grey) if the background collector has started marking.
1050 // This is necessary because the marker may
1051 // have passed this address and consequently this object will
1052 // not otherwise be greyed and would be incorrectly swept up.
1053 // Note that if this object contains references, the writing
1054 // of those references will dirty the card containing this object
1055 // allowing the object to be blackened (and its references scanned)
1056 // either during a preclean phase or at the final checkpoint.
1057 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
1058 assert(_markBitMap.covers(start, size), "Out of bounds");
1059 if (_collectorState >= Marking) {
1060 MutexLockerEx y(_markBitMap.lock(),
1061 Mutex::_no_safepoint_check_flag);
1062 // [see comments preceding SweepClosure::do_blk() below for details]
1063 //
1064 // Can the P-bits be deleted now? JJJ
1065 //
1066 // 1. need to mark the object as live so it isn't collected
1067 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
1068 // 3. need to mark the end of the object so marking, precleaning or sweeping
1069 // can skip over uninitialized or unparsable objects. An allocated
1070 // object is considered uninitialized for our purposes as long as
1071 // its klass word is NULL. All old gen objects are parsable
1072 // as soon as they are initialized.)
1073 _markBitMap.mark(start); // object is live
1074 _markBitMap.mark(start + 1); // object is potentially uninitialized?
1075 _markBitMap.mark(start + size - 1);
1076 // mark end of object
1077 }
1078 // check that oop looks uninitialized
1079 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
1080 }
1081
1082 void CMSCollector::promoted(bool par, HeapWord* start,
1083 bool is_obj_array, size_t obj_size) {
1084 assert(_markBitMap.covers(start), "Out of bounds");
1085 // See comment in direct_allocated() about when objects should
1086 // be allocated live.
1087 if (_collectorState >= Marking) {
1088 // we already hold the marking bit map lock, taken in
1089 // the prologue
1090 if (par) {
1091 _markBitMap.par_mark(start);
1092 } else {
1093 _markBitMap.mark(start);
1094 }
1095 // We don't need to mark the object as uninitialized (as
1096 // in direct_allocated above) because this is being done with the
1097 // world stopped and the object will be initialized by the
1098 // time the marking, precleaning or sweeping get to look at it.
1099 // But see the code for copying objects into the CMS generation,
1100 // where we need to ensure that concurrent readers of the
1101 // block offset table are able to safely navigate a block that
1102 // is in flux from being free to being allocated (and in
1103 // transition while being copied into) and subsequently
1104 // becoming a bona-fide object when the copy/promotion is complete.
1105 assert(SafepointSynchronize::is_at_safepoint(),
1106 "expect promotion only at safepoints");
1107
1108 if (_collectorState < Sweeping) {
1109 // Mark the appropriate cards in the modUnionTable, so that
1110 // this object gets scanned before the sweep. If this is
1111 // not done, CMS generation references in the object might
1112 // not get marked.
1113 // For the case of arrays, which are otherwise precisely
1114 // marked, we need to dirty the entire array, not just its head.
1115 if (is_obj_array) {
1116 // The [par_]mark_range() method expects mr.end() below to
1117 // be aligned to the granularity of a bit's representation
1118 // in the heap. In the case of the MUT below, that's a
1119 // card size.
1120 MemRegion mr(start,
1121 (HeapWord*)round_to((intptr_t)(start + obj_size),
1122 CardTableModRefBS::card_size /* bytes */));
1123 if (par) {
1124 _modUnionTable.par_mark_range(mr);
1125 } else {
1126 _modUnionTable.mark_range(mr);
1127 }
1128 } else { // not an obj array; we can just mark the head
1129 if (par) {
1130 _modUnionTable.par_mark(start);
1131 } else {
1132 _modUnionTable.mark(start);
1133 }
1134 }
1135 }
1136 }
1137 }
1138
1139 static inline size_t percent_of_space(Space* space, HeapWord* addr)
1140 {
1141 size_t delta = pointer_delta(addr, space->bottom());
1142 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
1143 }
1144
1145 void CMSCollector::icms_update_allocation_limits()
1146 {
1147 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
1148 EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
1149
1150 const unsigned int duty_cycle = stats().icms_update_duty_cycle();
1151 if (CMSTraceIncrementalPacing) {
1152 stats().print();
1153 }
1154
1155 assert(duty_cycle <= 100, "invalid duty cycle");
1156 if (duty_cycle != 0) {
1157 // The duty_cycle is a percentage between 0 and 100; convert to words and
1158 // then compute the offset from the endpoints of the space.
1159 size_t free_words = eden->free() / HeapWordSize;
1160 double free_words_dbl = (double)free_words;
1161 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
1162 size_t offset_words = (free_words - duty_cycle_words) / 2;
1163
1164 _icms_start_limit = eden->top() + offset_words;
1165 _icms_stop_limit = eden->end() - offset_words;
1166
1167 // The limits may be adjusted (shifted to the right) by
1168 // CMSIncrementalOffset, to allow the application more mutator time after a
1169 // young gen gc (when all mutators were stopped) and before CMS starts and
1170 // takes away one or more cpus.
1171 if (CMSIncrementalOffset != 0) {
1172 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
1173 size_t adjustment = (size_t)adjustment_dbl;
1174 HeapWord* tmp_stop = _icms_stop_limit + adjustment;
1175 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
1176 _icms_start_limit += adjustment;
1177 _icms_stop_limit = tmp_stop;
1178 }
1179 }
1180 }
1181 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
1182 _icms_start_limit = _icms_stop_limit = eden->end();
1183 }
1184
1185 // Install the new start limit.
1186 eden->set_soft_end(_icms_start_limit);
1187
1188 if (CMSTraceIncrementalMode) {
1189 gclog_or_tty->print(" icms alloc limits: "
1190 PTR_FORMAT "," PTR_FORMAT
1191 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
1192 p2i(_icms_start_limit), p2i(_icms_stop_limit),
1193 percent_of_space(eden, _icms_start_limit),
1194 percent_of_space(eden, _icms_stop_limit));
1195 if (Verbose) {
1196 gclog_or_tty->print("eden: ");
1197 eden->print_on(gclog_or_tty);
1198 }
1199 }
1200 }
1201
1202 // Any changes here should try to maintain the invariant
1203 // that if this method is called with _icms_start_limit
1204 // and _icms_stop_limit both NULL, then it should return NULL
1205 // and not notify the icms thread.
1206 HeapWord*
1207 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
1208 size_t word_size)
1209 {
1210 // A start_limit equal to end() means the duty cycle is 0, so treat that as a
1211 // nop.
1212 if (CMSIncrementalMode && _icms_start_limit != space->end()) {
1213 if (top <= _icms_start_limit) {
1214 if (CMSTraceIncrementalMode) {
1215 space->print_on(gclog_or_tty);
1216 gclog_or_tty->stamp();
1217 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
1218 ", new limit=" PTR_FORMAT
1219 " (" SIZE_FORMAT "%%)",
1220 p2i(top), p2i(_icms_stop_limit),
1221 percent_of_space(space, _icms_stop_limit));
1222 }
1223 ConcurrentMarkSweepThread::start_icms();
1224 assert(top < _icms_stop_limit, "Tautology");
1225 if (word_size < pointer_delta(_icms_stop_limit, top)) {
1226 return _icms_stop_limit;
1227 }
1228
1229 // The allocation will cross both the _start and _stop limits, so do the
1230 // stop notification also and return end().
1231 if (CMSTraceIncrementalMode) {
1232 space->print_on(gclog_or_tty);
1233 gclog_or_tty->stamp();
1234 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
1235 ", new limit=" PTR_FORMAT
1236 " (" SIZE_FORMAT "%%)",
1237 p2i(top), p2i(space->end()),
1238 percent_of_space(space, space->end()));
1239 }
1240 ConcurrentMarkSweepThread::stop_icms();
1241 return space->end();
1242 }
1243
1244 if (top <= _icms_stop_limit) {
1245 if (CMSTraceIncrementalMode) {
1246 space->print_on(gclog_or_tty);
1247 gclog_or_tty->stamp();
1248 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
1249 ", new limit=" PTR_FORMAT
1250 " (" SIZE_FORMAT "%%)",
1251 top, space->end(),
1252 percent_of_space(space, space->end()));
1253 }
1254 ConcurrentMarkSweepThread::stop_icms();
1255 return space->end();
1256 }
1257
1258 if (CMSTraceIncrementalMode) {
1259 space->print_on(gclog_or_tty);
1260 gclog_or_tty->stamp();
1261 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
1262 ", new limit=" PTR_FORMAT,
1263 top, NULL);
1264 }
1265 }
1266
1267 return NULL;
1268 }
1269
1270 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
1271 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1272 // allocate, copy and if necessary update promoinfo --
1273 // delegate to underlying space.
1274 assert_lock_strong(freelistLock());
1275
1276 #ifndef PRODUCT
1277 if (Universe::heap()->promotion_should_fail()) {
1278 return NULL;
1279 }
1280 #endif // #ifndef PRODUCT
1281
1282 oop res = _cmsSpace->promote(obj, obj_size);
1283 if (res == NULL) {
1284 // expand and retry
1285 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
1286 expand(s*HeapWordSize, MinHeapDeltaBytes,
1287 CMSExpansionCause::_satisfy_promotion);
1288 // Since there's currently no next generation, we don't try to promote
1289 // into a more senior generation.
1290 assert(next_gen() == NULL, "assumption, based upon which no attempt "
1291 "is made to pass on a possibly failing "
1292 "promotion to next generation");
1293 res = _cmsSpace->promote(obj, obj_size);
1294 }
1295 if (res != NULL) {
1296 // See comment in allocate() about when objects should
1297 // be allocated live.
1298 assert(obj->is_oop(), "Will dereference klass pointer below");
1299 collector()->promoted(false, // Not parallel
1300 (HeapWord*)res, obj->is_objArray(), obj_size);
1301 // promotion counters
1302 NOT_PRODUCT(
1303 _numObjectsPromoted++;
1304 _numWordsPromoted +=
1305 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1306 )
1307 }
1308 return res;
1309 }
1310
1311
1312 HeapWord*
1313 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
1314 HeapWord* top,
1315 size_t word_sz)
1316 {
1317 return collector()->allocation_limit_reached(space, top, word_sz);
1318 }
1319
1320 // IMPORTANT: Notes on object size recognition in CMS.
1321 // ---------------------------------------------------
1322 // A block of storage in the CMS generation is always in
1323 // one of three states. A free block (FREE), an allocated
1324 // object (OBJECT) whose size() method reports the correct size,
1325 // and an intermediate state (TRANSIENT) in which its size cannot
1326 // be accurately determined.
1327 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1328 // -----------------------------------------------------
1329 // FREE: klass_word & 1 == 1; mark_word holds block size
1330 //
1331 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1332 // obj->size() computes correct size
1333 //
1334 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1335 //
1336 // STATE IDENTIFICATION: (64 bit+COOPS)
1337 // ------------------------------------
1338 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1339 //
1340 // OBJECT: klass_word installed; klass_word != 0;
1341 // obj->size() computes correct size
1342 //
1343 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1344 //
1345 //
1346 // STATE TRANSITION DIAGRAM
1347 //
1348 // mut / parnew mut / parnew
1349 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1350 // ^ |
1351 // |------------------------ DEAD <------------------------------------|
1352 // sweep mut
1353 //
1354 // While a block is in TRANSIENT state its size cannot be determined
1355 // so readers will either need to come back later or stall until
1356 // the size can be determined. Note that for the case of direct
1357 // allocation, P-bits, when available, may be used to determine the
1358 // size of an object that may not yet have been initialized.
1359
1360 // Things to support parallel young-gen collection.
1361 oop
1362 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1363 oop old, markOop m,
1364 size_t word_sz) {
1365 #ifndef PRODUCT
1366 if (Universe::heap()->promotion_should_fail()) {
1367 return NULL;
1368 }
1369 #endif // #ifndef PRODUCT
1370
1371 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1372 PromotionInfo* promoInfo = &ps->promo;
1373 // if we are tracking promotions, then first ensure space for
1374 // promotion (including spooling space for saving header if necessary).
1375 // then allocate and copy, then track promoted info if needed.
1376 // When tracking (see PromotionInfo::track()), the mark word may
1377 // be displaced and in this case restoration of the mark word
1378 // occurs in the (oop_since_save_marks_)iterate phase.
1379 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1380 // Out of space for allocating spooling buffers;
1381 // try expanding and allocating spooling buffers.
1382 if (!expand_and_ensure_spooling_space(promoInfo)) {
1383 return NULL;
1384 }
1385 }
1386 assert(promoInfo->has_spooling_space(), "Control point invariant");
1387 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1388 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1389 if (obj_ptr == NULL) {
1390 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1391 if (obj_ptr == NULL) {
1392 return NULL;
1393 }
1394 }
1395 oop obj = oop(obj_ptr);
1396 OrderAccess::storestore();
1397 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1398 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1399 // IMPORTANT: See note on object initialization for CMS above.
1400 // Otherwise, copy the object. Here we must be careful to insert the
1401 // klass pointer last, since this marks the block as an allocated object.
1402 // Except with compressed oops it's the mark word.
1403 HeapWord* old_ptr = (HeapWord*)old;
1404 // Restore the mark word copied above.
1405 obj->set_mark(m);
1406 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1407 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1408 OrderAccess::storestore();
1409
1410 if (UseCompressedClassPointers) {
1411 // Copy gap missed by (aligned) header size calculation below
1412 obj->set_klass_gap(old->klass_gap());
1413 }
1414 if (word_sz > (size_t)oopDesc::header_size()) {
1415 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1416 obj_ptr + oopDesc::header_size(),
1417 word_sz - oopDesc::header_size());
1418 }
1419
1420 // Now we can track the promoted object, if necessary. We take care
1421 // to delay the transition from uninitialized to full object
1422 // (i.e., insertion of klass pointer) until after, so that it
1423 // atomically becomes a promoted object.
1424 if (promoInfo->tracking()) {
1425 promoInfo->track((PromotedObject*)obj, old->klass());
1426 }
1427 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1428 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1429 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1430
1431 // Finally, install the klass pointer (this should be volatile).
1432 OrderAccess::storestore();
1433 obj->set_klass(old->klass());
1434 // We should now be able to calculate the right size for this object
1435 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1436
1437 collector()->promoted(true, // parallel
1438 obj_ptr, old->is_objArray(), word_sz);
1439
1440 NOT_PRODUCT(
1441 Atomic::inc_ptr(&_numObjectsPromoted);
1442 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1443 )
1444
1445 return obj;
1446 }
1447
1448 void
1449 ConcurrentMarkSweepGeneration::
1450 par_promote_alloc_undo(int thread_num,
1451 HeapWord* obj, size_t word_sz) {
1452 // CMS does not support promotion undo.
1453 ShouldNotReachHere();
1454 }
1455
1456 void
1457 ConcurrentMarkSweepGeneration::
1458 par_promote_alloc_done(int thread_num) {
1459 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1460 ps->lab.retire(thread_num);
1461 }
1462
1463 void
1464 ConcurrentMarkSweepGeneration::
1465 par_oop_since_save_marks_iterate_done(int thread_num) {
1466 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1467 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1468 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1469 }
1470
1471 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1472 size_t size,
1473 bool tlab)
1474 {
1475 // We allow a STW collection only if a full
1476 // collection was requested.
1477 return full || should_allocate(size, tlab); // FIX ME !!!
1478 // This and promotion failure handling are connected at the
1479 // hip and should be fixed by untying them.
1480 }
1481
1482 bool CMSCollector::shouldConcurrentCollect() {
1483 if (_full_gc_requested) {
1484 if (Verbose && PrintGCDetails) {
1485 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1486 " gc request (or gc_locker)");
1487 }
1488 return true;
1489 }
1490
1491 // For debugging purposes, change the type of collection.
1492 // If the rotation is not on the concurrent collection
1493 // type, don't start a concurrent collection.
1494 NOT_PRODUCT(
1495 if (RotateCMSCollectionTypes &&
1496 (_cmsGen->debug_collection_type() !=
1497 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
1498 assert(_cmsGen->debug_collection_type() !=
1499 ConcurrentMarkSweepGeneration::Unknown_collection_type,
1500 "Bad cms collection type");
1501 return false;
1502 }
1503 )
1504
1505 FreelistLocker x(this);
1506 // ------------------------------------------------------------------
1507 // Print out lots of information which affects the initiation of
1508 // a collection.
1509 if (PrintCMSInitiationStatistics && stats().valid()) {
1510 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1511 gclog_or_tty->stamp();
1512 gclog_or_tty->cr();
1513 stats().print_on(gclog_or_tty);
1514 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1515 stats().time_until_cms_gen_full());
1516 gclog_or_tty->print_cr("free=" SIZE_FORMAT, _cmsGen->free());
1517 gclog_or_tty->print_cr("contiguous_available=" SIZE_FORMAT,
1518 _cmsGen->contiguous_available());
1519 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1520 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1521 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1522 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1523 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1524 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1525 gclog_or_tty->print_cr("metadata initialized %d",
1526 MetaspaceGC::should_concurrent_collect());
1527 }
1528 // ------------------------------------------------------------------
1529
1530 // If the estimated time to complete a cms collection (cms_duration())
1531 // is less than the estimated time remaining until the cms generation
1532 // is full, start a collection.
1533 if (!UseCMSInitiatingOccupancyOnly) {
1534 if (stats().valid()) {
1535 if (stats().time_until_cms_start() == 0.0) {
1536 return true;
1537 }
1538 } else {
1539 // We want to conservatively collect somewhat early in order
1540 // to try and "bootstrap" our CMS/promotion statistics;
1541 // this branch will not fire after the first successful CMS
1542 // collection because the stats should then be valid.
1543 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1544 if (Verbose && PrintGCDetails) {
1545 gclog_or_tty->print_cr(
1546 " CMSCollector: collect for bootstrapping statistics:"
1547 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1548 _bootstrap_occupancy);
1549 }
1550 return true;
1551 }
1552 }
1553 }
1554
1555 // Otherwise, we start a collection cycle if
1556 // old gen want a collection cycle started. Each may use
1557 // an appropriate criterion for making this decision.
1558 // XXX We need to make sure that the gen expansion
1559 // criterion dovetails well with this. XXX NEED TO FIX THIS
1560 if (_cmsGen->should_concurrent_collect()) {
1561 if (Verbose && PrintGCDetails) {
1562 gclog_or_tty->print_cr("CMS old gen initiated");
1563 }
1564 return true;
1565 }
1566
1567 // We start a collection if we believe an incremental collection may fail;
1568 // this is not likely to be productive in practice because it's probably too
1569 // late anyway.
1570 GenCollectedHeap* gch = GenCollectedHeap::heap();
1571 assert(gch->collector_policy()->is_two_generation_policy(),
1572 "You may want to check the correctness of the following");
1573 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1574 if (Verbose && PrintGCDetails) {
1575 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1576 }
1577 return true;
1578 }
1579
1580 if (MetaspaceGC::should_concurrent_collect()) {
1581 if (Verbose && PrintGCDetails) {
1582 gclog_or_tty->print("CMSCollector: collect for metadata allocation ");
1583 }
1584 return true;
1585 }
1586
1587 // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1588 if (CMSTriggerInterval >= 0) {
1589 if (CMSTriggerInterval == 0) {
1590 // Trigger always
1591 return true;
1592 }
1593
1594 // Check the CMS time since begin (we do not check the stats validity
1595 // as we want to be able to trigger the first CMS cycle as well)
1596 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1597 if (Verbose && PrintGCDetails) {
1598 if (stats().valid()) {
1599 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1600 stats().cms_time_since_begin());
1601 } else {
1602 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)");
1603 }
1604 }
1605 return true;
1606 }
1607 }
1608
1609 return false;
1610 }
1611
1612 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1613
1614 // Clear _expansion_cause fields of constituent generations
1615 void CMSCollector::clear_expansion_cause() {
1616 _cmsGen->clear_expansion_cause();
1617 }
1618
1619 // We should be conservative in starting a collection cycle. To
1620 // start too eagerly runs the risk of collecting too often in the
1621 // extreme. To collect too rarely falls back on full collections,
1622 // which works, even if not optimum in terms of concurrent work.
1623 // As a work around for too eagerly collecting, use the flag
1624 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1625 // giving the user an easily understandable way of controlling the
1626 // collections.
1627 // We want to start a new collection cycle if any of the following
1628 // conditions hold:
1629 // . our current occupancy exceeds the configured initiating occupancy
1630 // for this generation, or
1631 // . we recently needed to expand this space and have not, since that
1632 // expansion, done a collection of this generation, or
1633 // . the underlying space believes that it may be a good idea to initiate
1634 // a concurrent collection (this may be based on criteria such as the
1635 // following: the space uses linear allocation and linear allocation is
1636 // going to fail, or there is believed to be excessive fragmentation in
1637 // the generation, etc... or ...
1638 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1639 // the case of the old generation; see CR 6543076):
1640 // we may be approaching a point at which allocation requests may fail because
1641 // we will be out of sufficient free space given allocation rate estimates.]
1642 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1643
1644 assert_lock_strong(freelistLock());
1645 if (occupancy() > initiating_occupancy()) {
1646 if (PrintGCDetails && Verbose) {
1647 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1648 short_name(), occupancy(), initiating_occupancy());
1649 }
1650 return true;
1651 }
1652 if (UseCMSInitiatingOccupancyOnly) {
1653 return false;
1654 }
1655 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1656 if (PrintGCDetails && Verbose) {
1657 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1658 short_name());
1659 }
1660 return true;
1661 }
1662 if (_cmsSpace->should_concurrent_collect()) {
1663 if (PrintGCDetails && Verbose) {
1664 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1665 short_name());
1666 }
1667 return true;
1668 }
1669 return false;
1670 }
1671
1672 void ConcurrentMarkSweepGeneration::collect(bool full,
1673 bool clear_all_soft_refs,
1674 size_t size,
1675 bool tlab)
1676 {
1677 collector()->collect(full, clear_all_soft_refs, size, tlab);
1678 }
1679
1680 void CMSCollector::collect(bool full,
1681 bool clear_all_soft_refs,
1682 size_t size,
1683 bool tlab)
1684 {
1685 if (!UseCMSCollectionPassing && _collectorState > Idling) {
1686 // For debugging purposes skip the collection if the state
1687 // is not currently idle
1688 if (TraceCMSState) {
1689 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
1690 Thread::current(), full, _collectorState);
1691 }
1692 return;
1693 }
1694
1695 // The following "if" branch is present for defensive reasons.
1696 // In the current uses of this interface, it can be replaced with:
1697 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1698 // But I am not placing that assert here to allow future
1699 // generality in invoking this interface.
1700 if (GC_locker::is_active()) {
1701 // A consistency test for GC_locker
1702 assert(GC_locker::needs_gc(), "Should have been set already");
1703 // Skip this foreground collection, instead
1704 // expanding the heap if necessary.
1705 // Need the free list locks for the call to free() in compute_new_size()
1706 compute_new_size();
1707 return;
1708 }
1709 acquire_control_and_collect(full, clear_all_soft_refs);
1710 _full_gcs_since_conc_gc++;
1711 }
1712
1713 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1714 GenCollectedHeap* gch = GenCollectedHeap::heap();
1715 unsigned int gc_count = gch->total_full_collections();
1716 if (gc_count == full_gc_count) {
1717 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1718 _full_gc_requested = true;
1719 _full_gc_cause = cause;
1720 CGC_lock->notify(); // nudge CMS thread
1721 } else {
1722 assert(gc_count > full_gc_count, "Error: causal loop");
1723 }
1724 }
1725
1726 bool CMSCollector::is_external_interruption() {
1727 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1728 return GCCause::is_user_requested_gc(cause) ||
1729 GCCause::is_serviceability_requested_gc(cause);
1730 }
1731
1732 void CMSCollector::report_concurrent_mode_interruption() {
1733 if (is_external_interruption()) {
1734 if (PrintGCDetails) {
1735 gclog_or_tty->print(" (concurrent mode interrupted)");
1736 }
1737 } else {
1738 if (PrintGCDetails) {
1739 gclog_or_tty->print(" (concurrent mode failure)");
1740 }
1741 _gc_tracer_cm->report_concurrent_mode_failure();
1742 }
1743 }
1744
1745
1746 // The foreground and background collectors need to coordinate in order
1747 // to make sure that they do not mutually interfere with CMS collections.
1748 // When a background collection is active,
1749 // the foreground collector may need to take over (preempt) and
1750 // synchronously complete an ongoing collection. Depending on the
1751 // frequency of the background collections and the heap usage
1752 // of the application, this preemption can be seldom or frequent.
1753 // There are only certain
1754 // points in the background collection that the "collection-baton"
1755 // can be passed to the foreground collector.
1756 //
1757 // The foreground collector will wait for the baton before
1758 // starting any part of the collection. The foreground collector
1759 // will only wait at one location.
1760 //
1761 // The background collector will yield the baton before starting a new
1762 // phase of the collection (e.g., before initial marking, marking from roots,
1763 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1764 // of the loop which switches the phases. The background collector does some
1765 // of the phases (initial mark, final re-mark) with the world stopped.
1766 // Because of locking involved in stopping the world,
1767 // the foreground collector should not block waiting for the background
1768 // collector when it is doing a stop-the-world phase. The background
1769 // collector will yield the baton at an additional point just before
1770 // it enters a stop-the-world phase. Once the world is stopped, the
1771 // background collector checks the phase of the collection. If the
1772 // phase has not changed, it proceeds with the collection. If the
1773 // phase has changed, it skips that phase of the collection. See
1774 // the comments on the use of the Heap_lock in collect_in_background().
1775 //
1776 // Variable used in baton passing.
1777 // _foregroundGCIsActive - Set to true by the foreground collector when
1778 // it wants the baton. The foreground clears it when it has finished
1779 // the collection.
1780 // _foregroundGCShouldWait - Set to true by the background collector
1781 // when it is running. The foreground collector waits while
1782 // _foregroundGCShouldWait is true.
1783 // CGC_lock - monitor used to protect access to the above variables
1784 // and to notify the foreground and background collectors.
1785 // _collectorState - current state of the CMS collection.
1786 //
1787 // The foreground collector
1788 // acquires the CGC_lock
1789 // sets _foregroundGCIsActive
1790 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1791 // various locks acquired in preparation for the collection
1792 // are released so as not to block the background collector
1793 // that is in the midst of a collection
1794 // proceeds with the collection
1795 // clears _foregroundGCIsActive
1796 // returns
1797 //
1798 // The background collector in a loop iterating on the phases of the
1799 // collection
1800 // acquires the CGC_lock
1801 // sets _foregroundGCShouldWait
1802 // if _foregroundGCIsActive is set
1803 // clears _foregroundGCShouldWait, notifies _CGC_lock
1804 // waits on _CGC_lock for _foregroundGCIsActive to become false
1805 // and exits the loop.
1806 // otherwise
1807 // proceed with that phase of the collection
1808 // if the phase is a stop-the-world phase,
1809 // yield the baton once more just before enqueueing
1810 // the stop-world CMS operation (executed by the VM thread).
1811 // returns after all phases of the collection are done
1812 //
1813
1814 void CMSCollector::acquire_control_and_collect(bool full,
1815 bool clear_all_soft_refs) {
1816 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1817 assert(!Thread::current()->is_ConcurrentGC_thread(),
1818 "shouldn't try to acquire control from self!");
1819
1820 // Start the protocol for acquiring control of the
1821 // collection from the background collector (aka CMS thread).
1822 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1823 "VM thread should have CMS token");
1824 // Remember the possibly interrupted state of an ongoing
1825 // concurrent collection
1826 CollectorState first_state = _collectorState;
1827
1828 // Signal to a possibly ongoing concurrent collection that
1829 // we want to do a foreground collection.
1830 _foregroundGCIsActive = true;
1831
1832 // Disable incremental mode during a foreground collection.
1833 ICMSDisabler icms_disabler;
1834
1835 // release locks and wait for a notify from the background collector
1836 // releasing the locks in only necessary for phases which
1837 // do yields to improve the granularity of the collection.
1838 assert_lock_strong(bitMapLock());
1839 // We need to lock the Free list lock for the space that we are
1840 // currently collecting.
1841 assert(haveFreelistLocks(), "Must be holding free list locks");
1842 bitMapLock()->unlock();
1843 releaseFreelistLocks();
1844 {
1845 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1846 if (_foregroundGCShouldWait) {
1847 // We are going to be waiting for action for the CMS thread;
1848 // it had better not be gone (for instance at shutdown)!
1849 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1850 "CMS thread must be running");
1851 // Wait here until the background collector gives us the go-ahead
1852 ConcurrentMarkSweepThread::clear_CMS_flag(
1853 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1854 // Get a possibly blocked CMS thread going:
1855 // Note that we set _foregroundGCIsActive true above,
1856 // without protection of the CGC_lock.
1857 CGC_lock->notify();
1858 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1859 "Possible deadlock");
1860 while (_foregroundGCShouldWait) {
1861 // wait for notification
1862 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1863 // Possibility of delay/starvation here, since CMS token does
1864 // not know to give priority to VM thread? Actually, i think
1865 // there wouldn't be any delay/starvation, but the proof of
1866 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1867 }
1868 ConcurrentMarkSweepThread::set_CMS_flag(
1869 ConcurrentMarkSweepThread::CMS_vm_has_token);
1870 }
1871 }
1872 // The CMS_token is already held. Get back the other locks.
1873 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1874 "VM thread should have CMS token");
1875 getFreelistLocks();
1876 bitMapLock()->lock_without_safepoint_check();
1877 if (TraceCMSState) {
1878 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1879 INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1880 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1881 }
1882
1883 // Check if we need to do a compaction, or if not, whether
1884 // we need to start the mark-sweep from scratch.
1885 bool should_compact = false;
1886 bool should_start_over = false;
1887 decide_foreground_collection_type(clear_all_soft_refs,
1888 &should_compact, &should_start_over);
1889
1890 NOT_PRODUCT(
1891 if (RotateCMSCollectionTypes) {
1892 if (_cmsGen->debug_collection_type() ==
1893 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1894 should_compact = true;
1895 } else if (_cmsGen->debug_collection_type() ==
1896 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1897 should_compact = false;
1898 }
1899 }
1900 )
1901
1902 if (first_state > Idling) {
1903 report_concurrent_mode_interruption();
1904 }
1905
1906 set_did_compact(should_compact);
1907 if (should_compact) {
1908 // If the collection is being acquired from the background
1909 // collector, there may be references on the discovered
1910 // references lists that have NULL referents (being those
1911 // that were concurrently cleared by a mutator) or
1912 // that are no longer active (having been enqueued concurrently
1913 // by the mutator).
1914 // Scrub the list of those references because Mark-Sweep-Compact
1915 // code assumes referents are not NULL and that all discovered
1916 // Reference objects are active.
1917 ref_processor()->clean_up_discovered_references();
1918
1919 if (first_state > Idling) {
1920 save_heap_summary();
1921 }
1922
1923 do_compaction_work(clear_all_soft_refs);
1924
1925 // Has the GC time limit been exceeded?
1926 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration();
1927 size_t max_eden_size = young_gen->max_capacity() -
1928 young_gen->to()->capacity() -
1929 young_gen->from()->capacity();
1930 GenCollectedHeap* gch = GenCollectedHeap::heap();
1931 GCCause::Cause gc_cause = gch->gc_cause();
1932 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1933 young_gen->eden()->used(),
1934 _cmsGen->max_capacity(),
1935 max_eden_size,
1936 full,
1937 gc_cause,
1938 gch->collector_policy());
1939 } else {
1940 do_mark_sweep_work(clear_all_soft_refs, first_state,
1941 should_start_over);
1942 }
1943 // Reset the expansion cause, now that we just completed
1944 // a collection cycle.
1945 clear_expansion_cause();
1946 _foregroundGCIsActive = false;
1947 return;
1948 }
1949
1950 // Resize the tenured generation
1951 // after obtaining the free list locks for the
1952 // two generations.
1953 void CMSCollector::compute_new_size() {
1954 assert_locked_or_safepoint(Heap_lock);
1955 FreelistLocker z(this);
1956 MetaspaceGC::compute_new_size();
1957 _cmsGen->compute_new_size_free_list();
1958 }
1959
1960 // A work method used by foreground collection to determine
1961 // what type of collection (compacting or not, continuing or fresh)
1962 // it should do.
1963 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1964 // and CMSCompactWhenClearAllSoftRefs the default in the future
1965 // and do away with the flags after a suitable period.
1966 void CMSCollector::decide_foreground_collection_type(
1967 bool clear_all_soft_refs, bool* should_compact,
1968 bool* should_start_over) {
1969 // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1970 // flag is set, and we have either requested a System.gc() or
1971 // the number of full gc's since the last concurrent cycle
1972 // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1973 // or if an incremental collection has failed
1974 GenCollectedHeap* gch = GenCollectedHeap::heap();
1975 assert(gch->collector_policy()->is_two_generation_policy(),
1976 "You may want to check the correctness of the following");
1977 // Inform cms gen if this was due to partial collection failing.
1978 // The CMS gen may use this fact to determine its expansion policy.
1979 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1980 assert(!_cmsGen->incremental_collection_failed(),
1981 "Should have been noticed, reacted to and cleared");
1982 _cmsGen->set_incremental_collection_failed();
1983 }
1984 *should_compact =
1985 UseCMSCompactAtFullCollection &&
1986 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1987 GCCause::is_user_requested_gc(gch->gc_cause()) ||
1988 gch->incremental_collection_will_fail(true /* consult_young */));
1989 *should_start_over = false;
1990 if (clear_all_soft_refs && !*should_compact) {
1991 // We are about to do a last ditch collection attempt
1992 // so it would normally make sense to do a compaction
1993 // to reclaim as much space as possible.
1994 if (CMSCompactWhenClearAllSoftRefs) {
1995 // Default: The rationale is that in this case either
1996 // we are past the final marking phase, in which case
1997 // we'd have to start over, or so little has been done
1998 // that there's little point in saving that work. Compaction
1999 // appears to be the sensible choice in either case.
2000 *should_compact = true;
2001 } else {
2002 // We have been asked to clear all soft refs, but not to
2003 // compact. Make sure that we aren't past the final checkpoint
2004 // phase, for that is where we process soft refs. If we are already
2005 // past that phase, we'll need to redo the refs discovery phase and
2006 // if necessary clear soft refs that weren't previously
2007 // cleared. We do so by remembering the phase in which
2008 // we came in, and if we are past the refs processing
2009 // phase, we'll choose to just redo the mark-sweep
2010 // collection from scratch.
2011 if (_collectorState > FinalMarking) {
2012 // We are past the refs processing phase;
2013 // start over and do a fresh synchronous CMS cycle
2014 _collectorState = Resetting; // skip to reset to start new cycle
2015 reset(false /* == !asynch */);
2016 *should_start_over = true;
2017 } // else we can continue a possibly ongoing current cycle
2018 }
2019 }
2020 }
2021
2022 // A work method used by the foreground collector to do
2023 // a mark-sweep-compact.
2024 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
2025 GenCollectedHeap* gch = GenCollectedHeap::heap();
2026
2027 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
2028 gc_timer->register_gc_start();
2029
2030 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
2031 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
2032
2033 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id());
2034 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
2035 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
2036 "collections passed to foreground collector", _full_gcs_since_conc_gc);
2037 }
2038
2039 // Sample collection interval time and reset for collection pause.
2040 if (UseAdaptiveSizePolicy) {
2041 size_policy()->msc_collection_begin();
2042 }
2043
2044 // Temporarily widen the span of the weak reference processing to
2045 // the entire heap.
2046 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
2047 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
2048 // Temporarily, clear the "is_alive_non_header" field of the
2049 // reference processor.
2050 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
2051 // Temporarily make reference _processing_ single threaded (non-MT).
2052 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
2053 // Temporarily make refs discovery atomic
2054 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
2055 // Temporarily make reference _discovery_ single threaded (non-MT)
2056 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
2057
2058 ref_processor()->set_enqueuing_is_done(false);
2059 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/);
2060 ref_processor()->setup_policy(clear_all_soft_refs);
2061 // If an asynchronous collection finishes, the _modUnionTable is
2062 // all clear. If we are assuming the collection from an asynchronous
2063 // collection, clear the _modUnionTable.
2064 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
2065 "_modUnionTable should be clear if the baton was not passed");
2066 _modUnionTable.clear_all();
2067 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
2068 "mod union for klasses should be clear if the baton was passed");
2069 _ct->klass_rem_set()->clear_mod_union();
2070
2071 // We must adjust the allocation statistics being maintained
2072 // in the free list space. We do so by reading and clearing
2073 // the sweep timer and updating the block flux rate estimates below.
2074 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
2075 if (_inter_sweep_timer.is_active()) {
2076 _inter_sweep_timer.stop();
2077 // Note that we do not use this sample to update the _inter_sweep_estimate.
2078 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
2079 _inter_sweep_estimate.padded_average(),
2080 _intra_sweep_estimate.padded_average());
2081 }
2082
2083 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
2084 ref_processor(), clear_all_soft_refs);
2085 #ifdef ASSERT
2086 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
2087 size_t free_size = cms_space->free();
2088 assert(free_size ==
2089 pointer_delta(cms_space->end(), cms_space->compaction_top())
2090 * HeapWordSize,
2091 "All the free space should be compacted into one chunk at top");
2092 assert(cms_space->dictionary()->total_chunk_size(
2093 debug_only(cms_space->freelistLock())) == 0 ||
2094 cms_space->totalSizeInIndexedFreeLists() == 0,
2095 "All the free space should be in a single chunk");
2096 size_t num = cms_space->totalCount();
2097 assert((free_size == 0 && num == 0) ||
2098 (free_size > 0 && (num == 1 || num == 2)),
2099 "There should be at most 2 free chunks after compaction");
2100 #endif // ASSERT
2101 _collectorState = Resetting;
2102 assert(_restart_addr == NULL,
2103 "Should have been NULL'd before baton was passed");
2104 reset(false /* == !asynch */);
2105 _cmsGen->reset_after_compaction();
2106 _concurrent_cycles_since_last_unload = 0;
2107
2108 // Clear any data recorded in the PLAB chunk arrays.
2109 if (_survivor_plab_array != NULL) {
2110 reset_survivor_plab_arrays();
2111 }
2112
2113 // Adjust the per-size allocation stats for the next epoch.
2114 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
2115 // Restart the "inter sweep timer" for the next epoch.
2116 _inter_sweep_timer.reset();
2117 _inter_sweep_timer.start();
2118
2119 // Sample collection pause time and reset for collection interval.
2120 if (UseAdaptiveSizePolicy) {
2121 size_policy()->msc_collection_end(gch->gc_cause());
2122 }
2123
2124 gc_timer->register_gc_end();
2125
2126 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
2127
2128 // For a mark-sweep-compact, compute_new_size() will be called
2129 // in the heap's do_collection() method.
2130 }
2131
2132 // A work method used by the foreground collector to do
2133 // a mark-sweep, after taking over from a possibly on-going
2134 // concurrent mark-sweep collection.
2135 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2136 CollectorState first_state, bool should_start_over) {
2137 if (PrintGC && Verbose) {
2138 gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2139 "collector with count %d",
2140 _full_gcs_since_conc_gc);
2141 }
2142 switch (_collectorState) {
2143 case Idling:
2144 if (first_state == Idling || should_start_over) {
2145 // The background GC was not active, or should
2146 // restarted from scratch; start the cycle.
2147 _collectorState = InitialMarking;
2148 }
2149 // If first_state was not Idling, then a background GC
2150 // was in progress and has now finished. No need to do it
2151 // again. Leave the state as Idling.
2152 break;
2153 case Precleaning:
2154 // In the foreground case don't do the precleaning since
2155 // it is not done concurrently and there is extra work
2156 // required.
2157 _collectorState = FinalMarking;
2158 }
2159 collect_in_foreground(clear_all_soft_refs, GenCollectedHeap::heap()->gc_cause());
2160
2161 // For a mark-sweep, compute_new_size() will be called
2162 // in the heap's do_collection() method.
2163 }
2164
2165
2166 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
2167 DefNewGeneration* dng = _young_gen->as_DefNewGeneration();
2168 EdenSpace* eden_space = dng->eden();
2169 ContiguousSpace* from_space = dng->from();
2170 ContiguousSpace* to_space = dng->to();
2171 // Eden
2172 if (_eden_chunk_array != NULL) {
2173 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
2174 eden_space->bottom(), eden_space->top(),
2175 eden_space->end(), eden_space->capacity());
2176 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", "
2177 "_eden_chunk_capacity=" SIZE_FORMAT,
2178 _eden_chunk_index, _eden_chunk_capacity);
2179 for (size_t i = 0; i < _eden_chunk_index; i++) {
2180 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
2181 i, _eden_chunk_array[i]);
2182 }
2183 }
2184 // Survivor
2185 if (_survivor_chunk_array != NULL) {
2186 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
2187 from_space->bottom(), from_space->top(),
2188 from_space->end(), from_space->capacity());
2189 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", "
2190 "_survivor_chunk_capacity=" SIZE_FORMAT,
2191 _survivor_chunk_index, _survivor_chunk_capacity);
2192 for (size_t i = 0; i < _survivor_chunk_index; i++) {
2193 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
2194 i, _survivor_chunk_array[i]);
2195 }
2196 }
2197 }
2198
2199 void CMSCollector::getFreelistLocks() const {
2200 // Get locks for all free lists in all generations that this
2201 // collector is responsible for
2202 _cmsGen->freelistLock()->lock_without_safepoint_check();
2203 }
2204
2205 void CMSCollector::releaseFreelistLocks() const {
2206 // Release locks for all free lists in all generations that this
2207 // collector is responsible for
2208 _cmsGen->freelistLock()->unlock();
2209 }
2210
2211 bool CMSCollector::haveFreelistLocks() const {
2212 // Check locks for all free lists in all generations that this
2213 // collector is responsible for
2214 assert_lock_strong(_cmsGen->freelistLock());
2215 PRODUCT_ONLY(ShouldNotReachHere());
2216 return true;
2217 }
2218
2219 // A utility class that is used by the CMS collector to
2220 // temporarily "release" the foreground collector from its
2221 // usual obligation to wait for the background collector to
2222 // complete an ongoing phase before proceeding.
2223 class ReleaseForegroundGC: public StackObj {
2224 private:
2225 CMSCollector* _c;
2226 public:
2227 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2228 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2229 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2230 // allow a potentially blocked foreground collector to proceed
2231 _c->_foregroundGCShouldWait = false;
2232 if (_c->_foregroundGCIsActive) {
2233 CGC_lock->notify();
2234 }
2235 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2236 "Possible deadlock");
2237 }
2238
2239 ~ReleaseForegroundGC() {
2240 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2241 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2242 _c->_foregroundGCShouldWait = true;
2243 }
2244 };
2245
2246 // There are separate collect_in_background and collect_in_foreground because of
2247 // the different locking requirements of the background collector and the
2248 // foreground collector. There was originally an attempt to share
2249 // one "collect" method between the background collector and the foreground
2250 // collector but the if-then-else required made it cleaner to have
2251 // separate methods.
2252 void CMSCollector::collect_in_background(bool clear_all_soft_refs, GCCause::Cause cause) {
2253 assert(Thread::current()->is_ConcurrentGC_thread(),
2254 "A CMS asynchronous collection is only allowed on a CMS thread.");
2255
2256 GenCollectedHeap* gch = GenCollectedHeap::heap();
2257 {
2258 bool safepoint_check = Mutex::_no_safepoint_check_flag;
2259 MutexLockerEx hl(Heap_lock, safepoint_check);
2260 FreelistLocker fll(this);
2261 MutexLockerEx x(CGC_lock, safepoint_check);
2262 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2263 // The foreground collector is active or we're
2264 // not using asynchronous collections. Skip this
2265 // background collection.
2266 assert(!_foregroundGCShouldWait, "Should be clear");
2267 return;
2268 } else {
2269 assert(_collectorState == Idling, "Should be idling before start.");
2270 _collectorState = InitialMarking;
2271 register_gc_start(cause);
2272 // Reset the expansion cause, now that we are about to begin
2273 // a new cycle.
2274 clear_expansion_cause();
2275
2276 // Clear the MetaspaceGC flag since a concurrent collection
2277 // is starting but also clear it after the collection.
2278 MetaspaceGC::set_should_concurrent_collect(false);
2279 }
2280 // Decide if we want to enable class unloading as part of the
2281 // ensuing concurrent GC cycle.
2282 update_should_unload_classes();
2283 _full_gc_requested = false; // acks all outstanding full gc requests
2284 _full_gc_cause = GCCause::_no_gc;
2285 // Signal that we are about to start a collection
2286 gch->increment_total_full_collections(); // ... starting a collection cycle
2287 _collection_count_start = gch->total_full_collections();
2288 }
2289
2290 // Used for PrintGC
2291 size_t prev_used = 0;
2292 if (PrintGC && Verbose) {
2293 prev_used = _cmsGen->used(); // XXXPERM
2294 }
2295
2296 // The change of the collection state is normally done at this level;
2297 // the exceptions are phases that are executed while the world is
2298 // stopped. For those phases the change of state is done while the
2299 // world is stopped. For baton passing purposes this allows the
2300 // background collector to finish the phase and change state atomically.
2301 // The foreground collector cannot wait on a phase that is done
2302 // while the world is stopped because the foreground collector already
2303 // has the world stopped and would deadlock.
2304 while (_collectorState != Idling) {
2305 if (TraceCMSState) {
2306 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2307 Thread::current(), _collectorState);
2308 }
2309 // The foreground collector
2310 // holds the Heap_lock throughout its collection.
2311 // holds the CMS token (but not the lock)
2312 // except while it is waiting for the background collector to yield.
2313 //
2314 // The foreground collector should be blocked (not for long)
2315 // if the background collector is about to start a phase
2316 // executed with world stopped. If the background
2317 // collector has already started such a phase, the
2318 // foreground collector is blocked waiting for the
2319 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
2320 // are executed in the VM thread.
2321 //
2322 // The locking order is
2323 // PendingListLock (PLL) -- if applicable (FinalMarking)
2324 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
2325 // CMS token (claimed in
2326 // stop_world_and_do() -->
2327 // safepoint_synchronize() -->
2328 // CMSThread::synchronize())
2329
2330 {
2331 // Check if the FG collector wants us to yield.
2332 CMSTokenSync x(true); // is cms thread
2333 if (waitForForegroundGC()) {
2334 // We yielded to a foreground GC, nothing more to be
2335 // done this round.
2336 assert(_foregroundGCShouldWait == false, "We set it to false in "
2337 "waitForForegroundGC()");
2338 if (TraceCMSState) {
2339 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2340 " exiting collection CMS state %d",
2341 Thread::current(), _collectorState);
2342 }
2343 return;
2344 } else {
2345 // The background collector can run but check to see if the
2346 // foreground collector has done a collection while the
2347 // background collector was waiting to get the CGC_lock
2348 // above. If yes, break so that _foregroundGCShouldWait
2349 // is cleared before returning.
2350 if (_collectorState == Idling) {
2351 break;
2352 }
2353 }
2354 }
2355
2356 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2357 "should be waiting");
2358
2359 switch (_collectorState) {
2360 case InitialMarking:
2361 {
2362 ReleaseForegroundGC x(this);
2363 stats().record_cms_begin();
2364 VM_CMS_Initial_Mark initial_mark_op(this);
2365 VMThread::execute(&initial_mark_op);
2366 }
2367 // The collector state may be any legal state at this point
2368 // since the background collector may have yielded to the
2369 // foreground collector.
2370 break;
2371 case Marking:
2372 // initial marking in checkpointRootsInitialWork has been completed
2373 if (markFromRoots(true)) { // we were successful
2374 assert(_collectorState == Precleaning, "Collector state should "
2375 "have changed");
2376 } else {
2377 assert(_foregroundGCIsActive, "Internal state inconsistency");
2378 }
2379 break;
2380 case Precleaning:
2381 if (UseAdaptiveSizePolicy) {
2382 size_policy()->concurrent_precleaning_begin();
2383 }
2384 // marking from roots in markFromRoots has been completed
2385 preclean();
2386 if (UseAdaptiveSizePolicy) {
2387 size_policy()->concurrent_precleaning_end();
2388 }
2389 assert(_collectorState == AbortablePreclean ||
2390 _collectorState == FinalMarking,
2391 "Collector state should have changed");
2392 break;
2393 case AbortablePreclean:
2394 if (UseAdaptiveSizePolicy) {
2395 size_policy()->concurrent_phases_resume();
2396 }
2397 abortable_preclean();
2398 if (UseAdaptiveSizePolicy) {
2399 size_policy()->concurrent_precleaning_end();
2400 }
2401 assert(_collectorState == FinalMarking, "Collector state should "
2402 "have changed");
2403 break;
2404 case FinalMarking:
2405 {
2406 ReleaseForegroundGC x(this);
2407
2408 VM_CMS_Final_Remark final_remark_op(this);
2409 VMThread::execute(&final_remark_op);
2410 }
2411 assert(_foregroundGCShouldWait, "block post-condition");
2412 break;
2413 case Sweeping:
2414 if (UseAdaptiveSizePolicy) {
2415 size_policy()->concurrent_sweeping_begin();
2416 }
2417 // final marking in checkpointRootsFinal has been completed
2418 sweep(true);
2419 assert(_collectorState == Resizing, "Collector state change "
2420 "to Resizing must be done under the free_list_lock");
2421 _full_gcs_since_conc_gc = 0;
2422
2423 // Stop the timers for adaptive size policy for the concurrent phases
2424 if (UseAdaptiveSizePolicy) {
2425 size_policy()->concurrent_sweeping_end();
2426 size_policy()->concurrent_phases_end(gch->gc_cause(),
2427 gch->prev_gen(_cmsGen)->capacity(),
2428 _cmsGen->free());
2429 }
2430
2431 case Resizing: {
2432 // Sweeping has been completed...
2433 // At this point the background collection has completed.
2434 // Don't move the call to compute_new_size() down
2435 // into code that might be executed if the background
2436 // collection was preempted.
2437 {
2438 ReleaseForegroundGC x(this); // unblock FG collection
2439 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
2440 CMSTokenSync z(true); // not strictly needed.
2441 if (_collectorState == Resizing) {
2442 compute_new_size();
2443 save_heap_summary();
2444 _collectorState = Resetting;
2445 } else {
2446 assert(_collectorState == Idling, "The state should only change"
2447 " because the foreground collector has finished the collection");
2448 }
2449 }
2450 break;
2451 }
2452 case Resetting:
2453 // CMS heap resizing has been completed
2454 reset(true);
2455 assert(_collectorState == Idling, "Collector state should "
2456 "have changed");
2457
2458 MetaspaceGC::set_should_concurrent_collect(false);
2459
2460 stats().record_cms_end();
2461 // Don't move the concurrent_phases_end() and compute_new_size()
2462 // calls to here because a preempted background collection
2463 // has it's state set to "Resetting".
2464 break;
2465 case Idling:
2466 default:
2467 ShouldNotReachHere();
2468 break;
2469 }
2470 if (TraceCMSState) {
2471 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2472 Thread::current(), _collectorState);
2473 }
2474 assert(_foregroundGCShouldWait, "block post-condition");
2475 }
2476
2477 // Should this be in gc_epilogue?
2478 collector_policy()->counters()->update_counters();
2479
2480 {
2481 // Clear _foregroundGCShouldWait and, in the event that the
2482 // foreground collector is waiting, notify it, before
2483 // returning.
2484 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2485 _foregroundGCShouldWait = false;
2486 if (_foregroundGCIsActive) {
2487 CGC_lock->notify();
2488 }
2489 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2490 "Possible deadlock");
2491 }
2492 if (TraceCMSState) {
2493 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2494 " exiting collection CMS state %d",
2495 Thread::current(), _collectorState);
2496 }
2497 if (PrintGC && Verbose) {
2498 _cmsGen->print_heap_change(prev_used);
2499 }
2500 }
2501
2502 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) {
2503 if (!_cms_start_registered) {
2504 register_gc_start(cause);
2505 }
2506 }
2507
2508 void CMSCollector::register_gc_start(GCCause::Cause cause) {
2509 _cms_start_registered = true;
2510 _gc_timer_cm->register_gc_start();
2511 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
2512 }
2513
2514 void CMSCollector::register_gc_end() {
2515 if (_cms_start_registered) {
2516 report_heap_summary(GCWhen::AfterGC);
2517
2518 _gc_timer_cm->register_gc_end();
2519 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
2520 _cms_start_registered = false;
2521 }
2522 }
2523
2524 void CMSCollector::save_heap_summary() {
2525 GenCollectedHeap* gch = GenCollectedHeap::heap();
2526 _last_heap_summary = gch->create_heap_summary();
2527 _last_metaspace_summary = gch->create_metaspace_summary();
2528 }
2529
2530 void CMSCollector::report_heap_summary(GCWhen::Type when) {
2531 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
2532 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
2533 }
2534
2535 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) {
2536 assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2537 "Foreground collector should be waiting, not executing");
2538 assert(Thread::current()->is_VM_thread(), "A foreground collection"
2539 "may only be done by the VM Thread with the world stopped");
2540 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2541 "VM thread should have CMS token");
2542
2543 // The gc id is created in register_foreground_gc_start if this collection is synchronous
2544 const GCId gc_id = _collectorState == InitialMarking ? GCId::peek() : _gc_tracer_cm->gc_id();
2545 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2546 true, NULL, gc_id);)
2547 if (UseAdaptiveSizePolicy) {
2548 size_policy()->ms_collection_begin();
2549 }
2550 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2551
2552 HandleMark hm; // Discard invalid handles created during verification
2553
2554 if (VerifyBeforeGC &&
2555 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2556 Universe::verify();
2557 }
2558
2559 // Snapshot the soft reference policy to be used in this collection cycle.
2560 ref_processor()->setup_policy(clear_all_soft_refs);
2561
2562 // Decide if class unloading should be done
2563 update_should_unload_classes();
2564
2565 bool init_mark_was_synchronous = false; // until proven otherwise
2566 while (_collectorState != Idling) {
2567 if (TraceCMSState) {
2568 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2569 Thread::current(), _collectorState);
2570 }
2571 switch (_collectorState) {
2572 case InitialMarking:
2573 register_foreground_gc_start(cause);
2574 init_mark_was_synchronous = true; // fact to be exploited in re-mark
2575 checkpointRootsInitial(false);
2576 assert(_collectorState == Marking, "Collector state should have changed"
2577 " within checkpointRootsInitial()");
2578 break;
2579 case Marking:
2580 // initial marking in checkpointRootsInitialWork has been completed
2581 if (VerifyDuringGC &&
2582 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2583 Universe::verify("Verify before initial mark: ");
2584 }
2585 {
2586 bool res = markFromRoots(false);
2587 assert(res && _collectorState == FinalMarking, "Collector state should "
2588 "have changed");
2589 break;
2590 }
2591 case FinalMarking:
2592 if (VerifyDuringGC &&
2593 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2594 Universe::verify("Verify before re-mark: ");
2595 }
2596 checkpointRootsFinal(false, clear_all_soft_refs,
2597 init_mark_was_synchronous);
2598 assert(_collectorState == Sweeping, "Collector state should not "
2599 "have changed within checkpointRootsFinal()");
2600 break;
2601 case Sweeping:
2602 // final marking in checkpointRootsFinal has been completed
2603 if (VerifyDuringGC &&
2604 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2605 Universe::verify("Verify before sweep: ");
2606 }
2607 sweep(false);
2608 assert(_collectorState == Resizing, "Incorrect state");
2609 break;
2610 case Resizing: {
2611 // Sweeping has been completed; the actual resize in this case
2612 // is done separately; nothing to be done in this state.
2613 _collectorState = Resetting;
2614 break;
2615 }
2616 case Resetting:
2617 // The heap has been resized.
2618 if (VerifyDuringGC &&
2619 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2620 Universe::verify("Verify before reset: ");
2621 }
2622 save_heap_summary();
2623 reset(false);
2624 assert(_collectorState == Idling, "Collector state should "
2625 "have changed");
2626 break;
2627 case Precleaning:
2628 case AbortablePreclean:
2629 // Elide the preclean phase
2630 _collectorState = FinalMarking;
2631 break;
2632 default:
2633 ShouldNotReachHere();
2634 }
2635 if (TraceCMSState) {
2636 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2637 Thread::current(), _collectorState);
2638 }
2639 }
2640
2641 if (UseAdaptiveSizePolicy) {
2642 GenCollectedHeap* gch = GenCollectedHeap::heap();
2643 size_policy()->ms_collection_end(gch->gc_cause());
2644 }
2645
2646 if (VerifyAfterGC &&
2647 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2648 Universe::verify();
2649 }
2650 if (TraceCMSState) {
2651 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2652 " exiting collection CMS state %d",
2653 Thread::current(), _collectorState);
2654 }
2655 }
2656
2657 bool CMSCollector::waitForForegroundGC() {
2658 bool res = false;
2659 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2660 "CMS thread should have CMS token");
2661 // Block the foreground collector until the
2662 // background collectors decides whether to
2663 // yield.
2664 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2665 _foregroundGCShouldWait = true;
2666 if (_foregroundGCIsActive) {
2667 // The background collector yields to the
2668 // foreground collector and returns a value
2669 // indicating that it has yielded. The foreground
2670 // collector can proceed.
2671 res = true;
2672 _foregroundGCShouldWait = false;
2673 ConcurrentMarkSweepThread::clear_CMS_flag(
2674 ConcurrentMarkSweepThread::CMS_cms_has_token);
2675 ConcurrentMarkSweepThread::set_CMS_flag(
2676 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2677 // Get a possibly blocked foreground thread going
2678 CGC_lock->notify();
2679 if (TraceCMSState) {
2680 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2681 Thread::current(), _collectorState);
2682 }
2683 while (_foregroundGCIsActive) {
2684 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2685 }
2686 ConcurrentMarkSweepThread::set_CMS_flag(
2687 ConcurrentMarkSweepThread::CMS_cms_has_token);
2688 ConcurrentMarkSweepThread::clear_CMS_flag(
2689 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2690 }
2691 if (TraceCMSState) {
2692 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2693 Thread::current(), _collectorState);
2694 }
2695 return res;
2696 }
2697
2698 // Because of the need to lock the free lists and other structures in
2699 // the collector, common to all the generations that the collector is
2700 // collecting, we need the gc_prologues of individual CMS generations
2701 // delegate to their collector. It may have been simpler had the
2702 // current infrastructure allowed one to call a prologue on a
2703 // collector. In the absence of that we have the generation's
2704 // prologue delegate to the collector, which delegates back
2705 // some "local" work to a worker method in the individual generations
2706 // that it's responsible for collecting, while itself doing any
2707 // work common to all generations it's responsible for. A similar
2708 // comment applies to the gc_epilogue()'s.
2709 // The role of the varaible _between_prologue_and_epilogue is to
2710 // enforce the invocation protocol.
2711 void CMSCollector::gc_prologue(bool full) {
2712 // Call gc_prologue_work() for the CMSGen
2713 // we are responsible for.
2714
2715 // The following locking discipline assumes that we are only called
2716 // when the world is stopped.
2717 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2718
2719 // The CMSCollector prologue must call the gc_prologues for the
2720 // "generations" that it's responsible
2721 // for.
2722
2723 assert( Thread::current()->is_VM_thread()
2724 || ( CMSScavengeBeforeRemark
2725 && Thread::current()->is_ConcurrentGC_thread()),
2726 "Incorrect thread type for prologue execution");
2727
2728 if (_between_prologue_and_epilogue) {
2729 // We have already been invoked; this is a gc_prologue delegation
2730 // from yet another CMS generation that we are responsible for, just
2731 // ignore it since all relevant work has already been done.
2732 return;
2733 }
2734
2735 // set a bit saying prologue has been called; cleared in epilogue
2736 _between_prologue_and_epilogue = true;
2737 // Claim locks for common data structures, then call gc_prologue_work()
2738 // for each CMSGen.
2739
2740 getFreelistLocks(); // gets free list locks on constituent spaces
2741 bitMapLock()->lock_without_safepoint_check();
2742
2743 // Should call gc_prologue_work() for all cms gens we are responsible for
2744 bool duringMarking = _collectorState >= Marking
2745 && _collectorState < Sweeping;
2746
2747 // The young collections clear the modified oops state, which tells if
2748 // there are any modified oops in the class. The remark phase also needs
2749 // that information. Tell the young collection to save the union of all
2750 // modified klasses.
2751 if (duringMarking) {
2752 _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2753 }
2754
2755 bool registerClosure = duringMarking;
2756
2757 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ?
2758 &_modUnionClosurePar
2759 : &_modUnionClosure;
2760 _cmsGen->gc_prologue_work(full, registerClosure, muc);
2761
2762 if (!full) {
2763 stats().record_gc0_begin();
2764 }
2765 }
2766
2767 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2768
2769 _capacity_at_prologue = capacity();
2770 _used_at_prologue = used();
2771
2772 // Delegate to CMScollector which knows how to coordinate between
2773 // this and any other CMS generations that it is responsible for
2774 // collecting.
2775 collector()->gc_prologue(full);
2776 }
2777
2778 // This is a "private" interface for use by this generation's CMSCollector.
2779 // Not to be called directly by any other entity (for instance,
2780 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2781 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2782 bool registerClosure, ModUnionClosure* modUnionClosure) {
2783 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2784 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2785 "Should be NULL");
2786 if (registerClosure) {
2787 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2788 }
2789 cmsSpace()->gc_prologue();
2790 // Clear stat counters
2791 NOT_PRODUCT(
2792 assert(_numObjectsPromoted == 0, "check");
2793 assert(_numWordsPromoted == 0, "check");
2794 if (Verbose && PrintGC) {
2795 gclog_or_tty->print("Allocated " SIZE_FORMAT " objects, "
2796 SIZE_FORMAT " bytes concurrently",
2797 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2798 }
2799 _numObjectsAllocated = 0;
2800 _numWordsAllocated = 0;
2801 )
2802 }
2803
2804 void CMSCollector::gc_epilogue(bool full) {
2805 // The following locking discipline assumes that we are only called
2806 // when the world is stopped.
2807 assert(SafepointSynchronize::is_at_safepoint(),
2808 "world is stopped assumption");
2809
2810 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2811 // if linear allocation blocks need to be appropriately marked to allow the
2812 // the blocks to be parsable. We also check here whether we need to nudge the
2813 // CMS collector thread to start a new cycle (if it's not already active).
2814 assert( Thread::current()->is_VM_thread()
2815 || ( CMSScavengeBeforeRemark
2816 && Thread::current()->is_ConcurrentGC_thread()),
2817 "Incorrect thread type for epilogue execution");
2818
2819 if (!_between_prologue_and_epilogue) {
2820 // We have already been invoked; this is a gc_epilogue delegation
2821 // from yet another CMS generation that we are responsible for, just
2822 // ignore it since all relevant work has already been done.
2823 return;
2824 }
2825 assert(haveFreelistLocks(), "must have freelist locks");
2826 assert_lock_strong(bitMapLock());
2827
2828 _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2829
2830 _cmsGen->gc_epilogue_work(full);
2831
2832 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2833 // in case sampling was not already enabled, enable it
2834 _start_sampling = true;
2835 }
2836 // reset _eden_chunk_array so sampling starts afresh
2837 _eden_chunk_index = 0;
2838
2839 size_t cms_used = _cmsGen->cmsSpace()->used();
2840
2841 // update performance counters - this uses a special version of
2842 // update_counters() that allows the utilization to be passed as a
2843 // parameter, avoiding multiple calls to used().
2844 //
2845 _cmsGen->update_counters(cms_used);
2846
2847 if (CMSIncrementalMode) {
2848 icms_update_allocation_limits();
2849 }
2850
2851 bitMapLock()->unlock();
2852 releaseFreelistLocks();
2853
2854 if (!CleanChunkPoolAsync) {
2855 Chunk::clean_chunk_pool();
2856 }
2857
2858 set_did_compact(false);
2859 _between_prologue_and_epilogue = false; // ready for next cycle
2860 }
2861
2862 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2863 collector()->gc_epilogue(full);
2864
2865 // Also reset promotion tracking in par gc thread states.
2866 if (CollectedHeap::use_parallel_gc_threads()) {
2867 for (uint i = 0; i < ParallelGCThreads; i++) {
2868 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2869 }
2870 }
2871 }
2872
2873 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2874 assert(!incremental_collection_failed(), "Should have been cleared");
2875 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2876 cmsSpace()->gc_epilogue();
2877 // Print stat counters
2878 NOT_PRODUCT(
2879 assert(_numObjectsAllocated == 0, "check");
2880 assert(_numWordsAllocated == 0, "check");
2881 if (Verbose && PrintGC) {
2882 gclog_or_tty->print("Promoted " SIZE_FORMAT " objects, "
2883 SIZE_FORMAT " bytes",
2884 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2885 }
2886 _numObjectsPromoted = 0;
2887 _numWordsPromoted = 0;
2888 )
2889
2890 if (PrintGC && Verbose) {
2891 // Call down the chain in contiguous_available needs the freelistLock
2892 // so print this out before releasing the freeListLock.
2893 gclog_or_tty->print(" Contiguous available " SIZE_FORMAT " bytes ",
2894 contiguous_available());
2895 }
2896 }
2897
2898 #ifndef PRODUCT
2899 bool CMSCollector::have_cms_token() {
2900 Thread* thr = Thread::current();
2901 if (thr->is_VM_thread()) {
2902 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2903 } else if (thr->is_ConcurrentGC_thread()) {
2904 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2905 } else if (thr->is_GC_task_thread()) {
2906 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2907 ParGCRareEvent_lock->owned_by_self();
2908 }
2909 return false;
2910 }
2911 #endif
2912
2913 // Check reachability of the given heap address in CMS generation,
2914 // treating all other generations as roots.
2915 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2916 // We could "guarantee" below, rather than assert, but i'll
2917 // leave these as "asserts" so that an adventurous debugger
2918 // could try this in the product build provided some subset of
2919 // the conditions were met, provided they were intersted in the
2920 // results and knew that the computation below wouldn't interfere
2921 // with other concurrent computations mutating the structures
2922 // being read or written.
2923 assert(SafepointSynchronize::is_at_safepoint(),
2924 "Else mutations in object graph will make answer suspect");
2925 assert(have_cms_token(), "Should hold cms token");
2926 assert(haveFreelistLocks(), "must hold free list locks");
2927 assert_lock_strong(bitMapLock());
2928
2929 // Clear the marking bit map array before starting, but, just
2930 // for kicks, first report if the given address is already marked
2931 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
2932 _markBitMap.isMarked(addr) ? "" : " not");
2933
2934 if (verify_after_remark()) {
2935 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2936 bool result = verification_mark_bm()->isMarked(addr);
2937 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
2938 result ? "IS" : "is NOT");
2939 return result;
2940 } else {
2941 gclog_or_tty->print_cr("Could not compute result");
2942 return false;
2943 }
2944 }
2945
2946
2947 void
2948 CMSCollector::print_on_error(outputStream* st) {
2949 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2950 if (collector != NULL) {
2951 CMSBitMap* bitmap = &collector->_markBitMap;
2952 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap);
2953 bitmap->print_on_error(st, " Bits: ");
2954
2955 st->cr();
2956
2957 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2958 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap);
2959 mut_bitmap->print_on_error(st, " Bits: ");
2960 }
2961 }
2962
2963 ////////////////////////////////////////////////////////
2964 // CMS Verification Support
2965 ////////////////////////////////////////////////////////
2966 // Following the remark phase, the following invariant
2967 // should hold -- each object in the CMS heap which is
2968 // marked in markBitMap() should be marked in the verification_mark_bm().
2969
2970 class VerifyMarkedClosure: public BitMapClosure {
2971 CMSBitMap* _marks;
2972 bool _failed;
2973
2974 public:
2975 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2976
2977 bool do_bit(size_t offset) {
2978 HeapWord* addr = _marks->offsetToHeapWord(offset);
2979 if (!_marks->isMarked(addr)) {
2980 oop(addr)->print_on(gclog_or_tty);
2981 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
2982 _failed = true;
2983 }
2984 return true;
2985 }
2986
2987 bool failed() { return _failed; }
2988 };
2989
2990 bool CMSCollector::verify_after_remark(bool silent) {
2991 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... ");
2992 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2993 static bool init = false;
2994
2995 assert(SafepointSynchronize::is_at_safepoint(),
2996 "Else mutations in object graph will make answer suspect");
2997 assert(have_cms_token(),
2998 "Else there may be mutual interference in use of "
2999 " verification data structures");
3000 assert(_collectorState > Marking && _collectorState <= Sweeping,
3001 "Else marking info checked here may be obsolete");
3002 assert(haveFreelistLocks(), "must hold free list locks");
3003 assert_lock_strong(bitMapLock());
3004
3005
3006 // Allocate marking bit map if not already allocated
3007 if (!init) { // first time
3008 if (!verification_mark_bm()->allocate(_span)) {
3009 return false;
3010 }
3011 init = true;
3012 }
3013
3014 assert(verification_mark_stack()->isEmpty(), "Should be empty");
3015
3016 // Turn off refs discovery -- so we will be tracing through refs.
3017 // This is as intended, because by this time
3018 // GC must already have cleared any refs that need to be cleared,
3019 // and traced those that need to be marked; moreover,
3020 // the marking done here is not going to intefere in any
3021 // way with the marking information used by GC.
3022 NoRefDiscovery no_discovery(ref_processor());
3023
3024 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3025
3026 // Clear any marks from a previous round
3027 verification_mark_bm()->clear_all();
3028 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
3029 verify_work_stacks_empty();
3030
3031 GenCollectedHeap* gch = GenCollectedHeap::heap();
3032 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3033 // Update the saved marks which may affect the root scans.
3034 gch->save_marks();
3035
3036 if (CMSRemarkVerifyVariant == 1) {
3037 // In this first variant of verification, we complete
3038 // all marking, then check if the new marks-verctor is
3039 // a subset of the CMS marks-vector.
3040 verify_after_remark_work_1();
3041 } else if (CMSRemarkVerifyVariant == 2) {
3042 // In this second variant of verification, we flag an error
3043 // (i.e. an object reachable in the new marks-vector not reachable
3044 // in the CMS marks-vector) immediately, also indicating the
3045 // identify of an object (A) that references the unmarked object (B) --
3046 // presumably, a mutation to A failed to be picked up by preclean/remark?
3047 verify_after_remark_work_2();
3048 } else {
3049 warning("Unrecognized value %d for CMSRemarkVerifyVariant",
3050 CMSRemarkVerifyVariant);
3051 }
3052 if (!silent) gclog_or_tty->print(" done] ");
3053 return true;
3054 }
3055
3056 void CMSCollector::verify_after_remark_work_1() {
3057 ResourceMark rm;
3058 HandleMark hm;
3059 GenCollectedHeap* gch = GenCollectedHeap::heap();
3060
3061 // Get a clear set of claim bits for the roots processing to work with.
3062 ClassLoaderDataGraph::clear_claimed_marks();
3063
3064 // Mark from roots one level into CMS
3065 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
3066 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3067
3068 gch->gen_process_roots(_cmsGen->level(),
3069 true, // younger gens are roots
3070 true, // activate StrongRootsScope
3071 GenCollectedHeap::ScanningOption(roots_scanning_options()),
3072 should_unload_classes(),
3073 ¬Older,
3074 NULL,
3075 NULL); // SSS: Provide correct closure
3076
3077 // Now mark from the roots
3078 MarkFromRootsClosure markFromRootsClosure(this, _span,
3079 verification_mark_bm(), verification_mark_stack(),
3080 false /* don't yield */, true /* verifying */);
3081 assert(_restart_addr == NULL, "Expected pre-condition");
3082 verification_mark_bm()->iterate(&markFromRootsClosure);
3083 while (_restart_addr != NULL) {
3084 // Deal with stack overflow: by restarting at the indicated
3085 // address.
3086 HeapWord* ra = _restart_addr;
3087 markFromRootsClosure.reset(ra);
3088 _restart_addr = NULL;
3089 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
3090 }
3091 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
3092 verify_work_stacks_empty();
3093
3094 // Marking completed -- now verify that each bit marked in
3095 // verification_mark_bm() is also marked in markBitMap(); flag all
3096 // errors by printing corresponding objects.
3097 VerifyMarkedClosure vcl(markBitMap());
3098 verification_mark_bm()->iterate(&vcl);
3099 if (vcl.failed()) {
3100 gclog_or_tty->print("Verification failed");
3101 Universe::heap()->print_on(gclog_or_tty);
3102 fatal("CMS: failed marking verification after remark");
3103 }
3104 }
3105
3106 class VerifyKlassOopsKlassClosure : public KlassClosure {
3107 class VerifyKlassOopsClosure : public OopClosure {
3108 CMSBitMap* _bitmap;
3109 public:
3110 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
3111 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
3112 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
3113 } _oop_closure;
3114 public:
3115 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
3116 void do_klass(Klass* k) {
3117 k->oops_do(&_oop_closure);
3118 }
3119 };
3120
3121 void CMSCollector::verify_after_remark_work_2() {
3122 ResourceMark rm;
3123 HandleMark hm;
3124 GenCollectedHeap* gch = GenCollectedHeap::heap();
3125
3126 // Get a clear set of claim bits for the roots processing to work with.
3127 ClassLoaderDataGraph::clear_claimed_marks();
3128
3129 // Mark from roots one level into CMS
3130 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
3131 markBitMap());
3132 CLDToOopClosure cld_closure(¬Older, true);
3133
3134 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3135
3136 gch->gen_process_roots(_cmsGen->level(),
3137 true, // younger gens are roots
3138 true, // activate StrongRootsScope
3139 GenCollectedHeap::ScanningOption(roots_scanning_options()),
3140 should_unload_classes(),
3141 ¬Older,
3142 NULL,
3143 &cld_closure);
3144
3145 // Now mark from the roots
3146 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
3147 verification_mark_bm(), markBitMap(), verification_mark_stack());
3148 assert(_restart_addr == NULL, "Expected pre-condition");
3149 verification_mark_bm()->iterate(&markFromRootsClosure);
3150 while (_restart_addr != NULL) {
3151 // Deal with stack overflow: by restarting at the indicated
3152 // address.
3153 HeapWord* ra = _restart_addr;
3154 markFromRootsClosure.reset(ra);
3155 _restart_addr = NULL;
3156 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
3157 }
3158 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
3159 verify_work_stacks_empty();
3160
3161 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
3162 ClassLoaderDataGraph::classes_do(&verify_klass_oops);
3163
3164 // Marking completed -- now verify that each bit marked in
3165 // verification_mark_bm() is also marked in markBitMap(); flag all
3166 // errors by printing corresponding objects.
3167 VerifyMarkedClosure vcl(markBitMap());
3168 verification_mark_bm()->iterate(&vcl);
3169 assert(!vcl.failed(), "Else verification above should not have succeeded");
3170 }
3171
3172 void ConcurrentMarkSweepGeneration::save_marks() {
3173 // delegate to CMS space
3174 cmsSpace()->save_marks();
3175 for (uint i = 0; i < ParallelGCThreads; i++) {
3176 _par_gc_thread_states[i]->promo.startTrackingPromotions();
3177 }
3178 }
3179
3180 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
3181 return cmsSpace()->no_allocs_since_save_marks();
3182 }
3183
3184 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
3185 \
3186 void ConcurrentMarkSweepGeneration:: \
3187 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
3188 cl->set_generation(this); \
3189 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
3190 cl->reset_generation(); \
3191 save_marks(); \
3192 }
3193
3194 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
3195
3196 void
3197 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
3198 cl->set_generation(this);
3199 younger_refs_in_space_iterate(_cmsSpace, cl);
3200 cl->reset_generation();
3201 }
3202
3203 void
3204 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
3205 if (freelistLock()->owned_by_self()) {
3206 Generation::oop_iterate(cl);
3207 } else {
3208 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3209 Generation::oop_iterate(cl);
3210 }
3211 }
3212
3213 void
3214 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
3215 if (freelistLock()->owned_by_self()) {
3216 Generation::object_iterate(cl);
3217 } else {
3218 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3219 Generation::object_iterate(cl);
3220 }
3221 }
3222
3223 void
3224 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
3225 if (freelistLock()->owned_by_self()) {
3226 Generation::safe_object_iterate(cl);
3227 } else {
3228 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3229 Generation::safe_object_iterate(cl);
3230 }
3231 }
3232
3233 void
3234 ConcurrentMarkSweepGeneration::post_compact() {
3235 }
3236
3237 void
3238 ConcurrentMarkSweepGeneration::prepare_for_verify() {
3239 // Fix the linear allocation blocks to look like free blocks.
3240
3241 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3242 // are not called when the heap is verified during universe initialization and
3243 // at vm shutdown.
3244 if (freelistLock()->owned_by_self()) {
3245 cmsSpace()->prepare_for_verify();
3246 } else {
3247 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3248 cmsSpace()->prepare_for_verify();
3249 }
3250 }
3251
3252 void
3253 ConcurrentMarkSweepGeneration::verify() {
3254 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3255 // are not called when the heap is verified during universe initialization and
3256 // at vm shutdown.
3257 if (freelistLock()->owned_by_self()) {
3258 cmsSpace()->verify();
3259 } else {
3260 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3261 cmsSpace()->verify();
3262 }
3263 }
3264
3265 void CMSCollector::verify() {
3266 _cmsGen->verify();
3267 }
3268
3269 #ifndef PRODUCT
3270 bool CMSCollector::overflow_list_is_empty() const {
3271 assert(_num_par_pushes >= 0, "Inconsistency");
3272 if (_overflow_list == NULL) {
3273 assert(_num_par_pushes == 0, "Inconsistency");
3274 }
3275 return _overflow_list == NULL;
3276 }
3277
3278 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3279 // merely consolidate assertion checks that appear to occur together frequently.
3280 void CMSCollector::verify_work_stacks_empty() const {
3281 assert(_markStack.isEmpty(), "Marking stack should be empty");
3282 assert(overflow_list_is_empty(), "Overflow list should be empty");
3283 }
3284
3285 void CMSCollector::verify_overflow_empty() const {
3286 assert(overflow_list_is_empty(), "Overflow list should be empty");
3287 assert(no_preserved_marks(), "No preserved marks");
3288 }
3289 #endif // PRODUCT
3290
3291 // Decide if we want to enable class unloading as part of the
3292 // ensuing concurrent GC cycle. We will collect and
3293 // unload classes if it's the case that:
3294 // (1) an explicit gc request has been made and the flag
3295 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3296 // (2) (a) class unloading is enabled at the command line, and
3297 // (b) old gen is getting really full
3298 // NOTE: Provided there is no change in the state of the heap between
3299 // calls to this method, it should have idempotent results. Moreover,
3300 // its results should be monotonically increasing (i.e. going from 0 to 1,
3301 // but not 1 to 0) between successive calls between which the heap was
3302 // not collected. For the implementation below, it must thus rely on
3303 // the property that concurrent_cycles_since_last_unload()
3304 // will not decrease unless a collection cycle happened and that
3305 // _cmsGen->is_too_full() are
3306 // themselves also monotonic in that sense. See check_monotonicity()
3307 // below.
3308 void CMSCollector::update_should_unload_classes() {
3309 _should_unload_classes = false;
3310 // Condition 1 above
3311 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3312 _should_unload_classes = true;
3313 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3314 // Disjuncts 2.b.(i,ii,iii) above
3315 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3316 CMSClassUnloadingMaxInterval)
3317 || _cmsGen->is_too_full();
3318 }
3319 }
3320
3321 bool ConcurrentMarkSweepGeneration::is_too_full() const {
3322 bool res = should_concurrent_collect();
3323 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
3324 return res;
3325 }
3326
3327 void CMSCollector::setup_cms_unloading_and_verification_state() {
3328 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3329 || VerifyBeforeExit;
3330 const int rso = GenCollectedHeap::SO_AllCodeCache;
3331
3332 // We set the proper root for this CMS cycle here.
3333 if (should_unload_classes()) { // Should unload classes this cycle
3334 remove_root_scanning_option(rso); // Shrink the root set appropriately
3335 set_verifying(should_verify); // Set verification state for this cycle
3336 return; // Nothing else needs to be done at this time
3337 }
3338
3339 // Not unloading classes this cycle
3340 assert(!should_unload_classes(), "Inconsitency!");
3341
3342 // If we are not unloading classes then add SO_AllCodeCache to root
3343 // scanning options.
3344 add_root_scanning_option(rso);
3345
3346 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3347 set_verifying(true);
3348 } else if (verifying() && !should_verify) {
3349 // We were verifying, but some verification flags got disabled.
3350 set_verifying(false);
3351 // Exclude symbols, strings and code cache elements from root scanning to
3352 // reduce IM and RM pauses.
3353 remove_root_scanning_option(rso);
3354 }
3355 }
3356
3357
3358 #ifndef PRODUCT
3359 HeapWord* CMSCollector::block_start(const void* p) const {
3360 const HeapWord* addr = (HeapWord*)p;
3361 if (_span.contains(p)) {
3362 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3363 return _cmsGen->cmsSpace()->block_start(p);
3364 }
3365 }
3366 return NULL;
3367 }
3368 #endif
3369
3370 HeapWord*
3371 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3372 bool tlab,
3373 bool parallel) {
3374 CMSSynchronousYieldRequest yr;
3375 assert(!tlab, "Can't deal with TLAB allocation");
3376 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3377 expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3378 CMSExpansionCause::_satisfy_allocation);
3379 if (GCExpandToAllocateDelayMillis > 0) {
3380 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3381 }
3382 return have_lock_and_allocate(word_size, tlab);
3383 }
3384
3385 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3386 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3387 // to CardGeneration and share it...
3388 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) {
3389 return CardGeneration::expand(bytes, expand_bytes);
3390 }
3391
3392 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3393 CMSExpansionCause::Cause cause)
3394 {
3395
3396 bool success = expand(bytes, expand_bytes);
3397
3398 // remember why we expanded; this information is used
3399 // by shouldConcurrentCollect() when making decisions on whether to start
3400 // a new CMS cycle.
3401 if (success) {
3402 set_expansion_cause(cause);
3403 if (PrintGCDetails && Verbose) {
3404 gclog_or_tty->print_cr("Expanded CMS gen for %s",
3405 CMSExpansionCause::to_string(cause));
3406 }
3407 }
3408 }
3409
3410 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3411 HeapWord* res = NULL;
3412 MutexLocker x(ParGCRareEvent_lock);
3413 while (true) {
3414 // Expansion by some other thread might make alloc OK now:
3415 res = ps->lab.alloc(word_sz);
3416 if (res != NULL) return res;
3417 // If there's not enough expansion space available, give up.
3418 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3419 return NULL;
3420 }
3421 // Otherwise, we try expansion.
3422 expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3423 CMSExpansionCause::_allocate_par_lab);
3424 // Now go around the loop and try alloc again;
3425 // A competing par_promote might beat us to the expansion space,
3426 // so we may go around the loop again if promotion fails agaion.
3427 if (GCExpandToAllocateDelayMillis > 0) {
3428 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3429 }
3430 }
3431 }
3432
3433
3434 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3435 PromotionInfo* promo) {
3436 MutexLocker x(ParGCRareEvent_lock);
3437 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3438 while (true) {
3439 // Expansion by some other thread might make alloc OK now:
3440 if (promo->ensure_spooling_space()) {
3441 assert(promo->has_spooling_space(),
3442 "Post-condition of successful ensure_spooling_space()");
3443 return true;
3444 }
3445 // If there's not enough expansion space available, give up.
3446 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3447 return false;
3448 }
3449 // Otherwise, we try expansion.
3450 expand(refill_size_bytes, MinHeapDeltaBytes,
3451 CMSExpansionCause::_allocate_par_spooling_space);
3452 // Now go around the loop and try alloc again;
3453 // A competing allocation might beat us to the expansion space,
3454 // so we may go around the loop again if allocation fails again.
3455 if (GCExpandToAllocateDelayMillis > 0) {
3456 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3457 }
3458 }
3459 }
3460
3461
3462 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3463 assert_locked_or_safepoint(ExpandHeap_lock);
3464 // Shrink committed space
3465 _virtual_space.shrink_by(bytes);
3466 // Shrink space; this also shrinks the space's BOT
3467 _cmsSpace->set_end((HeapWord*) _virtual_space.high());
3468 size_t new_word_size = heap_word_size(_cmsSpace->capacity());
3469 // Shrink the shared block offset array
3470 _bts->resize(new_word_size);
3471 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3472 // Shrink the card table
3473 Universe::heap()->barrier_set()->resize_covered_region(mr);
3474
3475 if (Verbose && PrintGC) {
3476 size_t new_mem_size = _virtual_space.committed_size();
3477 size_t old_mem_size = new_mem_size + bytes;
3478 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
3479 name(), old_mem_size/K, new_mem_size/K);
3480 }
3481 }
3482
3483 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3484 assert_locked_or_safepoint(Heap_lock);
3485 size_t size = ReservedSpace::page_align_size_down(bytes);
3486 // Only shrink if a compaction was done so that all the free space
3487 // in the generation is in a contiguous block at the end.
3488 if (size > 0 && did_compact()) {
3489 shrink_by(size);
3490 }
3491 }
3492
3493 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3494 assert_locked_or_safepoint(Heap_lock);
3495 bool result = _virtual_space.expand_by(bytes);
3496 if (result) {
3497 size_t new_word_size =
3498 heap_word_size(_virtual_space.committed_size());
3499 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3500 _bts->resize(new_word_size); // resize the block offset shared array
3501 Universe::heap()->barrier_set()->resize_covered_region(mr);
3502 // Hmmmm... why doesn't CFLS::set_end verify locking?
3503 // This is quite ugly; FIX ME XXX
3504 _cmsSpace->assert_locked(freelistLock());
3505 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3506
3507 // update the space and generation capacity counters
3508 if (UsePerfData) {
3509 _space_counters->update_capacity();
3510 _gen_counters->update_all();
3511 }
3512
3513 if (Verbose && PrintGC) {
3514 size_t new_mem_size = _virtual_space.committed_size();
3515 size_t old_mem_size = new_mem_size - bytes;
3516 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K",
3517 name(), old_mem_size/K, bytes/K, new_mem_size/K);
3518 }
3519 }
3520 return result;
3521 }
3522
3523 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3524 assert_locked_or_safepoint(Heap_lock);
3525 bool success = true;
3526 const size_t remaining_bytes = _virtual_space.uncommitted_size();
3527 if (remaining_bytes > 0) {
3528 success = grow_by(remaining_bytes);
3529 DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3530 }
3531 return success;
3532 }
3533
3534 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
3535 assert_locked_or_safepoint(Heap_lock);
3536 assert_lock_strong(freelistLock());
3537 if (PrintGCDetails && Verbose) {
3538 warning("Shrinking of CMS not yet implemented");
3539 }
3540 return;
3541 }
3542
3543
3544 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3545 // phases.
3546 class CMSPhaseAccounting: public StackObj {
3547 public:
3548 CMSPhaseAccounting(CMSCollector *collector,
3549 const char *phase,
3550 const GCId gc_id,
3551 bool print_cr = true);
3552 ~CMSPhaseAccounting();
3553
3554 private:
3555 CMSCollector *_collector;
3556 const char *_phase;
3557 elapsedTimer _wallclock;
3558 bool _print_cr;
3559 const GCId _gc_id;
3560
3561 public:
3562 // Not MT-safe; so do not pass around these StackObj's
3563 // where they may be accessed by other threads.
3564 jlong wallclock_millis() {
3565 assert(_wallclock.is_active(), "Wall clock should not stop");
3566 _wallclock.stop(); // to record time
3567 jlong ret = _wallclock.milliseconds();
3568 _wallclock.start(); // restart
3569 return ret;
3570 }
3571 };
3572
3573 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3574 const char *phase,
3575 const GCId gc_id,
3576 bool print_cr) :
3577 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) {
3578
3579 if (PrintCMSStatistics != 0) {
3580 _collector->resetYields();
3581 }
3582 if (PrintGCDetails) {
3583 gclog_or_tty->gclog_stamp(_gc_id);
3584 gclog_or_tty->print_cr("[%s-concurrent-%s-start]",
3585 _collector->cmsGen()->short_name(), _phase);
3586 }
3587 _collector->resetTimer();
3588 _wallclock.start();
3589 _collector->startTimer();
3590 }
3591
3592 CMSPhaseAccounting::~CMSPhaseAccounting() {
3593 assert(_wallclock.is_active(), "Wall clock should not have stopped");
3594 _collector->stopTimer();
3595 _wallclock.stop();
3596 if (PrintGCDetails) {
3597 gclog_or_tty->gclog_stamp(_gc_id);
3598 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3599 _collector->cmsGen()->short_name(),
3600 _phase, _collector->timerValue(), _wallclock.seconds());
3601 if (_print_cr) {
3602 gclog_or_tty->cr();
3603 }
3604 if (PrintCMSStatistics != 0) {
3605 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3606 _collector->yields());
3607 }
3608 }
3609 }
3610
3611 // CMS work
3612
3613 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
3614 class CMSParMarkTask : public AbstractGangTask {
3615 protected:
3616 CMSCollector* _collector;
3617 int _n_workers;
3618 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) :
3619 AbstractGangTask(name),
3620 _collector(collector),
3621 _n_workers(n_workers) {}
3622 // Work method in support of parallel rescan ... of young gen spaces
3623 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
3624 ContiguousSpace* space,
3625 HeapWord** chunk_array, size_t chunk_top);
3626 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
3627 };
3628
3629 // Parallel initial mark task
3630 class CMSParInitialMarkTask: public CMSParMarkTask {
3631 public:
3632 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) :
3633 CMSParMarkTask("Scan roots and young gen for initial mark in parallel",
3634 collector, n_workers) {}
3635 void work(uint worker_id);
3636 };
3637
3638 // Checkpoint the roots into this generation from outside
3639 // this generation. [Note this initial checkpoint need only
3640 // be approximate -- we'll do a catch up phase subsequently.]
3641 void CMSCollector::checkpointRootsInitial(bool asynch) {
3642 assert(_collectorState == InitialMarking, "Wrong collector state");
3643 check_correct_thread_executing();
3644 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
3645
3646 save_heap_summary();
3647 report_heap_summary(GCWhen::BeforeGC);
3648
3649 ReferenceProcessor* rp = ref_processor();
3650 SpecializationStats::clear();
3651 assert(_restart_addr == NULL, "Control point invariant");
3652 if (asynch) {
3653 // acquire locks for subsequent manipulations
3654 MutexLockerEx x(bitMapLock(),
3655 Mutex::_no_safepoint_check_flag);
3656 checkpointRootsInitialWork(asynch);
3657 // enable ("weak") refs discovery
3658 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/);
3659 _collectorState = Marking;
3660 } else {
3661 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3662 // which recognizes if we are a CMS generation, and doesn't try to turn on
3663 // discovery; verify that they aren't meddling.
3664 assert(!rp->discovery_is_atomic(),
3665 "incorrect setting of discovery predicate");
3666 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3667 "ref discovery for this generation kind");
3668 // already have locks
3669 checkpointRootsInitialWork(asynch);
3670 // now enable ("weak") refs discovery
3671 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/);
3672 _collectorState = Marking;
3673 }
3674 SpecializationStats::print();
3675 }
3676
3677 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3678 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3679 assert(_collectorState == InitialMarking, "just checking");
3680
3681 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3682 // precede our marking with a collection of all
3683 // younger generations to keep floating garbage to a minimum.
3684 // XXX: we won't do this for now -- it's an optimization to be done later.
3685
3686 // already have locks
3687 assert_lock_strong(bitMapLock());
3688 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3689
3690 // Setup the verification and class unloading state for this
3691 // CMS collection cycle.
3692 setup_cms_unloading_and_verification_state();
3693
3694 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
3695 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
3696 if (UseAdaptiveSizePolicy) {
3697 size_policy()->checkpoint_roots_initial_begin();
3698 }
3699
3700 // Reset all the PLAB chunk arrays if necessary.
3701 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3702 reset_survivor_plab_arrays();
3703 }
3704
3705 ResourceMark rm;
3706 HandleMark hm;
3707
3708 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
3709 GenCollectedHeap* gch = GenCollectedHeap::heap();
3710
3711 verify_work_stacks_empty();
3712 verify_overflow_empty();
3713
3714 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3715 // Update the saved marks which may affect the root scans.
3716 gch->save_marks();
3717
3718 // weak reference processing has not started yet.
3719 ref_processor()->set_enqueuing_is_done(false);
3720
3721 // Need to remember all newly created CLDs,
3722 // so that we can guarantee that the remark finds them.
3723 ClassLoaderDataGraph::remember_new_clds(true);
3724
3725 // Whenever a CLD is found, it will be claimed before proceeding to mark
3726 // the klasses. The claimed marks need to be cleared before marking starts.
3727 ClassLoaderDataGraph::clear_claimed_marks();
3728
3729 if (CMSPrintEdenSurvivorChunks) {
3730 print_eden_and_survivor_chunk_arrays();
3731 }
3732
3733 {
3734 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3735 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
3736 // The parallel version.
3737 FlexibleWorkGang* workers = gch->workers();
3738 assert(workers != NULL, "Need parallel worker threads.");
3739 int n_workers = workers->active_workers();
3740 CMSParInitialMarkTask tsk(this, n_workers);
3741 gch->set_par_threads(n_workers);
3742 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
3743 if (n_workers > 1) {
3744 GenCollectedHeap::StrongRootsScope srs(gch);
3745 workers->run_task(&tsk);
3746 } else {
3747 GenCollectedHeap::StrongRootsScope srs(gch);
3748 tsk.work(0);
3749 }
3750 gch->set_par_threads(0);
3751 } else {
3752 // The serial version.
3753 CLDToOopClosure cld_closure(¬Older, true);
3754 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3755 gch->gen_process_roots(_cmsGen->level(),
3756 true, // younger gens are roots
3757 true, // activate StrongRootsScope
3758 GenCollectedHeap::ScanningOption(roots_scanning_options()),
3759 should_unload_classes(),
3760 ¬Older,
3761 NULL,
3762 &cld_closure);
3763 }
3764 }
3765
3766 // Clear mod-union table; it will be dirtied in the prologue of
3767 // CMS generation per each younger generation collection.
3768
3769 assert(_modUnionTable.isAllClear(),
3770 "Was cleared in most recent final checkpoint phase"
3771 " or no bits are set in the gc_prologue before the start of the next "
3772 "subsequent marking phase.");
3773
3774 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
3775
3776 // Save the end of the used_region of the constituent generations
3777 // to be used to limit the extent of sweep in each generation.
3778 save_sweep_limits();
3779 if (UseAdaptiveSizePolicy) {
3780 size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3781 }
3782 verify_overflow_empty();
3783 }
3784
3785 bool CMSCollector::markFromRoots(bool asynch) {
3786 // we might be tempted to assert that:
3787 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3788 // "inconsistent argument?");
3789 // However that wouldn't be right, because it's possible that
3790 // a safepoint is indeed in progress as a younger generation
3791 // stop-the-world GC happens even as we mark in this generation.
3792 assert(_collectorState == Marking, "inconsistent state?");
3793 check_correct_thread_executing();
3794 verify_overflow_empty();
3795
3796 bool res;
3797 if (asynch) {
3798
3799 // Start the timers for adaptive size policy for the concurrent phases
3800 // Do it here so that the foreground MS can use the concurrent
3801 // timer since a foreground MS might has the sweep done concurrently
3802 // or STW.
3803 if (UseAdaptiveSizePolicy) {
3804 size_policy()->concurrent_marking_begin();
3805 }
3806
3807 // Weak ref discovery note: We may be discovering weak
3808 // refs in this generation concurrent (but interleaved) with
3809 // weak ref discovery by a younger generation collector.
3810
3811 CMSTokenSyncWithLocks ts(true, bitMapLock());
3812 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3813 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3814 res = markFromRootsWork(asynch);
3815 if (res) {
3816 _collectorState = Precleaning;
3817 } else { // We failed and a foreground collection wants to take over
3818 assert(_foregroundGCIsActive, "internal state inconsistency");
3819 assert(_restart_addr == NULL, "foreground will restart from scratch");
3820 if (PrintGCDetails) {
3821 gclog_or_tty->print_cr("bailing out to foreground collection");
3822 }
3823 }
3824 if (UseAdaptiveSizePolicy) {
3825 size_policy()->concurrent_marking_end();
3826 }
3827 } else {
3828 assert(SafepointSynchronize::is_at_safepoint(),
3829 "inconsistent with asynch == false");
3830 if (UseAdaptiveSizePolicy) {
3831 size_policy()->ms_collection_marking_begin();
3832 }
3833 // already have locks
3834 res = markFromRootsWork(asynch);
3835 _collectorState = FinalMarking;
3836 if (UseAdaptiveSizePolicy) {
3837 GenCollectedHeap* gch = GenCollectedHeap::heap();
3838 size_policy()->ms_collection_marking_end(gch->gc_cause());
3839 }
3840 }
3841 verify_overflow_empty();
3842 return res;
3843 }
3844
3845 bool CMSCollector::markFromRootsWork(bool asynch) {
3846 // iterate over marked bits in bit map, doing a full scan and mark
3847 // from these roots using the following algorithm:
3848 // . if oop is to the right of the current scan pointer,
3849 // mark corresponding bit (we'll process it later)
3850 // . else (oop is to left of current scan pointer)
3851 // push oop on marking stack
3852 // . drain the marking stack
3853
3854 // Note that when we do a marking step we need to hold the
3855 // bit map lock -- recall that direct allocation (by mutators)
3856 // and promotion (by younger generation collectors) is also
3857 // marking the bit map. [the so-called allocate live policy.]
3858 // Because the implementation of bit map marking is not
3859 // robust wrt simultaneous marking of bits in the same word,
3860 // we need to make sure that there is no such interference
3861 // between concurrent such updates.
3862
3863 // already have locks
3864 assert_lock_strong(bitMapLock());
3865
3866 verify_work_stacks_empty();
3867 verify_overflow_empty();
3868 bool result = false;
3869 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3870 result = do_marking_mt(asynch);
3871 } else {
3872 result = do_marking_st(asynch);
3873 }
3874 return result;
3875 }
3876
3877 // Forward decl
3878 class CMSConcMarkingTask;
3879
3880 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3881 CMSCollector* _collector;
3882 CMSConcMarkingTask* _task;
3883 public:
3884 virtual void yield();
3885
3886 // "n_threads" is the number of threads to be terminated.
3887 // "queue_set" is a set of work queues of other threads.
3888 // "collector" is the CMS collector associated with this task terminator.
3889 // "yield" indicates whether we need the gang as a whole to yield.
3890 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3891 ParallelTaskTerminator(n_threads, queue_set),
3892 _collector(collector) { }
3893
3894 void set_task(CMSConcMarkingTask* task) {
3895 _task = task;
3896 }
3897 };
3898
3899 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3900 CMSConcMarkingTask* _task;
3901 public:
3902 bool should_exit_termination();
3903 void set_task(CMSConcMarkingTask* task) {
3904 _task = task;
3905 }
3906 };
3907
3908 // MT Concurrent Marking Task
3909 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3910 CMSCollector* _collector;
3911 int _n_workers; // requested/desired # workers
3912 bool _asynch;
3913 bool _result;
3914 CompactibleFreeListSpace* _cms_space;
3915 char _pad_front[64]; // padding to ...
3916 HeapWord* _global_finger; // ... avoid sharing cache line
3917 char _pad_back[64];
3918 HeapWord* _restart_addr;
3919
3920 // Exposed here for yielding support
3921 Mutex* const _bit_map_lock;
3922
3923 // The per thread work queues, available here for stealing
3924 OopTaskQueueSet* _task_queues;
3925
3926 // Termination (and yielding) support
3927 CMSConcMarkingTerminator _term;
3928 CMSConcMarkingTerminatorTerminator _term_term;
3929
3930 public:
3931 CMSConcMarkingTask(CMSCollector* collector,
3932 CompactibleFreeListSpace* cms_space,
3933 bool asynch,
3934 YieldingFlexibleWorkGang* workers,
3935 OopTaskQueueSet* task_queues):
3936 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3937 _collector(collector),
3938 _cms_space(cms_space),
3939 _asynch(asynch), _n_workers(0), _result(true),
3940 _task_queues(task_queues),
3941 _term(_n_workers, task_queues, _collector),
3942 _bit_map_lock(collector->bitMapLock())
3943 {
3944 _requested_size = _n_workers;
3945 _term.set_task(this);
3946 _term_term.set_task(this);
3947 _restart_addr = _global_finger = _cms_space->bottom();
3948 }
3949
3950
3951 OopTaskQueueSet* task_queues() { return _task_queues; }
3952
3953 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3954
3955 HeapWord** global_finger_addr() { return &_global_finger; }
3956
3957 CMSConcMarkingTerminator* terminator() { return &_term; }
3958
3959 virtual void set_for_termination(int active_workers) {
3960 terminator()->reset_for_reuse(active_workers);
3961 }
3962
3963 void work(uint worker_id);
3964 bool should_yield() {
3965 return ConcurrentMarkSweepThread::should_yield()
3966 && !_collector->foregroundGCIsActive()
3967 && _asynch;
3968 }
3969
3970 virtual void coordinator_yield(); // stuff done by coordinator
3971 bool result() { return _result; }
3972
3973 void reset(HeapWord* ra) {
3974 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3975 _restart_addr = _global_finger = ra;
3976 _term.reset_for_reuse();
3977 }
3978
3979 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3980 OopTaskQueue* work_q);
3981
3982 private:
3983 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3984 void do_work_steal(int i);
3985 void bump_global_finger(HeapWord* f);
3986 };
3987
3988 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3989 assert(_task != NULL, "Error");
3990 return _task->yielding();
3991 // Note that we do not need the disjunct || _task->should_yield() above
3992 // because we want terminating threads to yield only if the task
3993 // is already in the midst of yielding, which happens only after at least one
3994 // thread has yielded.
3995 }
3996
3997 void CMSConcMarkingTerminator::yield() {
3998 if (_task->should_yield()) {
3999 _task->yield();
4000 } else {
4001 ParallelTaskTerminator::yield();
4002 }
4003 }
4004
4005 ////////////////////////////////////////////////////////////////
4006 // Concurrent Marking Algorithm Sketch
4007 ////////////////////////////////////////////////////////////////
4008 // Until all tasks exhausted (both spaces):
4009 // -- claim next available chunk
4010 // -- bump global finger via CAS
4011 // -- find first object that starts in this chunk
4012 // and start scanning bitmap from that position
4013 // -- scan marked objects for oops
4014 // -- CAS-mark target, and if successful:
4015 // . if target oop is above global finger (volatile read)
4016 // nothing to do
4017 // . if target oop is in chunk and above local finger
4018 // then nothing to do
4019 // . else push on work-queue
4020 // -- Deal with possible overflow issues:
4021 // . local work-queue overflow causes stuff to be pushed on
4022 // global (common) overflow queue
4023 // . always first empty local work queue
4024 // . then get a batch of oops from global work queue if any
4025 // . then do work stealing
4026 // -- When all tasks claimed (both spaces)
4027 // and local work queue empty,
4028 // then in a loop do:
4029 // . check global overflow stack; steal a batch of oops and trace
4030 // . try to steal from other threads oif GOS is empty
4031 // . if neither is available, offer termination
4032 // -- Terminate and return result
4033 //
4034 void CMSConcMarkingTask::work(uint worker_id) {
4035 elapsedTimer _timer;
4036 ResourceMark rm;
4037 HandleMark hm;
4038
4039 DEBUG_ONLY(_collector->verify_overflow_empty();)
4040
4041 // Before we begin work, our work queue should be empty
4042 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
4043 // Scan the bitmap covering _cms_space, tracing through grey objects.
4044 _timer.start();
4045 do_scan_and_mark(worker_id, _cms_space);
4046 _timer.stop();
4047 if (PrintCMSStatistics != 0) {
4048 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
4049 worker_id, _timer.seconds());
4050 // XXX: need xxx/xxx type of notation, two timers
4051 }
4052
4053 // ... do work stealing
4054 _timer.reset();
4055 _timer.start();
4056 do_work_steal(worker_id);
4057 _timer.stop();
4058 if (PrintCMSStatistics != 0) {
4059 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
4060 worker_id, _timer.seconds());
4061 // XXX: need xxx/xxx type of notation, two timers
4062 }
4063 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
4064 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
4065 // Note that under the current task protocol, the
4066 // following assertion is true even of the spaces
4067 // expanded since the completion of the concurrent
4068 // marking. XXX This will likely change under a strict
4069 // ABORT semantics.
4070 // After perm removal the comparison was changed to
4071 // greater than or equal to from strictly greater than.
4072 // Before perm removal the highest address sweep would
4073 // have been at the end of perm gen but now is at the
4074 // end of the tenured gen.
4075 assert(_global_finger >= _cms_space->end(),
4076 "All tasks have been completed");
4077 DEBUG_ONLY(_collector->verify_overflow_empty();)
4078 }
4079
4080 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
4081 HeapWord* read = _global_finger;
4082 HeapWord* cur = read;
4083 while (f > read) {
4084 cur = read;
4085 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
4086 if (cur == read) {
4087 // our cas succeeded
4088 assert(_global_finger >= f, "protocol consistency");
4089 break;
4090 }
4091 }
4092 }
4093
4094 // This is really inefficient, and should be redone by
4095 // using (not yet available) block-read and -write interfaces to the
4096 // stack and the work_queue. XXX FIX ME !!!
4097 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
4098 OopTaskQueue* work_q) {
4099 // Fast lock-free check
4100 if (ovflw_stk->length() == 0) {
4101 return false;
4102 }
4103 assert(work_q->size() == 0, "Shouldn't steal");
4104 MutexLockerEx ml(ovflw_stk->par_lock(),
4105 Mutex::_no_safepoint_check_flag);
4106 // Grab up to 1/4 the size of the work queue
4107 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4108 (size_t)ParGCDesiredObjsFromOverflowList);
4109 num = MIN2(num, ovflw_stk->length());
4110 for (int i = (int) num; i > 0; i--) {
4111 oop cur = ovflw_stk->pop();
4112 assert(cur != NULL, "Counted wrong?");
4113 work_q->push(cur);
4114 }
4115 return num > 0;
4116 }
4117
4118 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
4119 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4120 int n_tasks = pst->n_tasks();
4121 // We allow that there may be no tasks to do here because
4122 // we are restarting after a stack overflow.
4123 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
4124 uint nth_task = 0;
4125
4126 HeapWord* aligned_start = sp->bottom();
4127 if (sp->used_region().contains(_restart_addr)) {
4128 // Align down to a card boundary for the start of 0th task
4129 // for this space.
4130 aligned_start =
4131 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
4132 CardTableModRefBS::card_size);
4133 }
4134
4135 size_t chunk_size = sp->marking_task_size();
4136 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4137 // Having claimed the nth task in this space,
4138 // compute the chunk that it corresponds to:
4139 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
4140 aligned_start + (nth_task+1)*chunk_size);
4141 // Try and bump the global finger via a CAS;
4142 // note that we need to do the global finger bump
4143 // _before_ taking the intersection below, because
4144 // the task corresponding to that region will be
4145 // deemed done even if the used_region() expands
4146 // because of allocation -- as it almost certainly will
4147 // during start-up while the threads yield in the
4148 // closure below.
4149 HeapWord* finger = span.end();
4150 bump_global_finger(finger); // atomically
4151 // There are null tasks here corresponding to chunks
4152 // beyond the "top" address of the space.
4153 span = span.intersection(sp->used_region());
4154 if (!span.is_empty()) { // Non-null task
4155 HeapWord* prev_obj;
4156 assert(!span.contains(_restart_addr) || nth_task == 0,
4157 "Inconsistency");
4158 if (nth_task == 0) {
4159 // For the 0th task, we'll not need to compute a block_start.
4160 if (span.contains(_restart_addr)) {
4161 // In the case of a restart because of stack overflow,
4162 // we might additionally skip a chunk prefix.
4163 prev_obj = _restart_addr;
4164 } else {
4165 prev_obj = span.start();
4166 }
4167 } else {
4168 // We want to skip the first object because
4169 // the protocol is to scan any object in its entirety
4170 // that _starts_ in this span; a fortiori, any
4171 // object starting in an earlier span is scanned
4172 // as part of an earlier claimed task.
4173 // Below we use the "careful" version of block_start
4174 // so we do not try to navigate uninitialized objects.
4175 prev_obj = sp->block_start_careful(span.start());
4176 // Below we use a variant of block_size that uses the
4177 // Printezis bits to avoid waiting for allocated
4178 // objects to become initialized/parsable.
4179 while (prev_obj < span.start()) {
4180 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
4181 if (sz > 0) {
4182 prev_obj += sz;
4183 } else {
4184 // In this case we may end up doing a bit of redundant
4185 // scanning, but that appears unavoidable, short of
4186 // locking the free list locks; see bug 6324141.
4187 break;
4188 }
4189 }
4190 }
4191 if (prev_obj < span.end()) {
4192 MemRegion my_span = MemRegion(prev_obj, span.end());
4193 // Do the marking work within a non-empty span --
4194 // the last argument to the constructor indicates whether the
4195 // iteration should be incremental with periodic yields.
4196 Par_MarkFromRootsClosure cl(this, _collector, my_span,
4197 &_collector->_markBitMap,
4198 work_queue(i),
4199 &_collector->_markStack,
4200 _asynch);
4201 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
4202 } // else nothing to do for this task
4203 } // else nothing to do for this task
4204 }
4205 // We'd be tempted to assert here that since there are no
4206 // more tasks left to claim in this space, the global_finger
4207 // must exceed space->top() and a fortiori space->end(). However,
4208 // that would not quite be correct because the bumping of
4209 // global_finger occurs strictly after the claiming of a task,
4210 // so by the time we reach here the global finger may not yet
4211 // have been bumped up by the thread that claimed the last
4212 // task.
4213 pst->all_tasks_completed();
4214 }
4215
4216 class Par_ConcMarkingClosure: public MetadataAwareOopClosure {
4217 private:
4218 CMSCollector* _collector;
4219 CMSConcMarkingTask* _task;
4220 MemRegion _span;
4221 CMSBitMap* _bit_map;
4222 CMSMarkStack* _overflow_stack;
4223 OopTaskQueue* _work_queue;
4224 protected:
4225 DO_OOP_WORK_DEFN
4226 public:
4227 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
4228 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
4229 MetadataAwareOopClosure(collector->ref_processor()),
4230 _collector(collector),
4231 _task(task),
4232 _span(collector->_span),
4233 _work_queue(work_queue),
4234 _bit_map(bit_map),
4235 _overflow_stack(overflow_stack)
4236 { }
4237 virtual void do_oop(oop* p);
4238 virtual void do_oop(narrowOop* p);
4239
4240 void trim_queue(size_t max);
4241 void handle_stack_overflow(HeapWord* lost);
4242 void do_yield_check() {
4243 if (_task->should_yield()) {
4244 _task->yield();
4245 }
4246 }
4247 };
4248
4249 // Grey object scanning during work stealing phase --
4250 // the salient assumption here is that any references
4251 // that are in these stolen objects being scanned must
4252 // already have been initialized (else they would not have
4253 // been published), so we do not need to check for
4254 // uninitialized objects before pushing here.
4255 void Par_ConcMarkingClosure::do_oop(oop obj) {
4256 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
4257 HeapWord* addr = (HeapWord*)obj;
4258 // Check if oop points into the CMS generation
4259 // and is not marked
4260 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
4261 // a white object ...
4262 // If we manage to "claim" the object, by being the
4263 // first thread to mark it, then we push it on our
4264 // marking stack
4265 if (_bit_map->par_mark(addr)) { // ... now grey
4266 // push on work queue (grey set)
4267 bool simulate_overflow = false;
4268 NOT_PRODUCT(
4269 if (CMSMarkStackOverflowALot &&
4270 _collector->simulate_overflow()) {
4271 // simulate a stack overflow
4272 simulate_overflow = true;
4273 }
4274 )
4275 if (simulate_overflow ||
4276 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
4277 // stack overflow
4278 if (PrintCMSStatistics != 0) {
4279 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
4280 SIZE_FORMAT, _overflow_stack->capacity());
4281 }
4282 // We cannot assert that the overflow stack is full because
4283 // it may have been emptied since.
4284 assert(simulate_overflow ||
4285 _work_queue->size() == _work_queue->max_elems(),
4286 "Else push should have succeeded");
4287 handle_stack_overflow(addr);
4288 }
4289 } // Else, some other thread got there first
4290 do_yield_check();
4291 }
4292 }
4293
4294 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4295 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4296
4297 void Par_ConcMarkingClosure::trim_queue(size_t max) {
4298 while (_work_queue->size() > max) {
4299 oop new_oop;
4300 if (_work_queue->pop_local(new_oop)) {
4301 assert(new_oop->is_oop(), "Should be an oop");
4302 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
4303 assert(_span.contains((HeapWord*)new_oop), "Not in span");
4304 new_oop->oop_iterate(this); // do_oop() above
4305 do_yield_check();
4306 }
4307 }
4308 }
4309
4310 // Upon stack overflow, we discard (part of) the stack,
4311 // remembering the least address amongst those discarded
4312 // in CMSCollector's _restart_address.
4313 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
4314 // We need to do this under a mutex to prevent other
4315 // workers from interfering with the work done below.
4316 MutexLockerEx ml(_overflow_stack->par_lock(),
4317 Mutex::_no_safepoint_check_flag);
4318 // Remember the least grey address discarded
4319 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
4320 _collector->lower_restart_addr(ra);
4321 _overflow_stack->reset(); // discard stack contents
4322 _overflow_stack->expand(); // expand the stack if possible
4323 }
4324
4325
4326 void CMSConcMarkingTask::do_work_steal(int i) {
4327 OopTaskQueue* work_q = work_queue(i);
4328 oop obj_to_scan;
4329 CMSBitMap* bm = &(_collector->_markBitMap);
4330 CMSMarkStack* ovflw = &(_collector->_markStack);
4331 int* seed = _collector->hash_seed(i);
4332 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
4333 while (true) {
4334 cl.trim_queue(0);
4335 assert(work_q->size() == 0, "Should have been emptied above");
4336 if (get_work_from_overflow_stack(ovflw, work_q)) {
4337 // Can't assert below because the work obtained from the
4338 // overflow stack may already have been stolen from us.
4339 // assert(work_q->size() > 0, "Work from overflow stack");
4340 continue;
4341 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4342 assert(obj_to_scan->is_oop(), "Should be an oop");
4343 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
4344 obj_to_scan->oop_iterate(&cl);
4345 } else if (terminator()->offer_termination(&_term_term)) {
4346 assert(work_q->size() == 0, "Impossible!");
4347 break;
4348 } else if (yielding() || should_yield()) {
4349 yield();
4350 }
4351 }
4352 }
4353
4354 // This is run by the CMS (coordinator) thread.
4355 void CMSConcMarkingTask::coordinator_yield() {
4356 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4357 "CMS thread should hold CMS token");
4358 // First give up the locks, then yield, then re-lock
4359 // We should probably use a constructor/destructor idiom to
4360 // do this unlock/lock or modify the MutexUnlocker class to
4361 // serve our purpose. XXX
4362 assert_lock_strong(_bit_map_lock);
4363 _bit_map_lock->unlock();
4364 ConcurrentMarkSweepThread::desynchronize(true);
4365 ConcurrentMarkSweepThread::acknowledge_yield_request();
4366 _collector->stopTimer();
4367 if (PrintCMSStatistics != 0) {
4368 _collector->incrementYields();
4369 }
4370 _collector->icms_wait();
4371
4372 // It is possible for whichever thread initiated the yield request
4373 // not to get a chance to wake up and take the bitmap lock between
4374 // this thread releasing it and reacquiring it. So, while the
4375 // should_yield() flag is on, let's sleep for a bit to give the
4376 // other thread a chance to wake up. The limit imposed on the number
4377 // of iterations is defensive, to avoid any unforseen circumstances
4378 // putting us into an infinite loop. Since it's always been this
4379 // (coordinator_yield()) method that was observed to cause the
4380 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4381 // which is by default non-zero. For the other seven methods that
4382 // also perform the yield operation, as are using a different
4383 // parameter (CMSYieldSleepCount) which is by default zero. This way we
4384 // can enable the sleeping for those methods too, if necessary.
4385 // See 6442774.
4386 //
4387 // We really need to reconsider the synchronization between the GC
4388 // thread and the yield-requesting threads in the future and we
4389 // should really use wait/notify, which is the recommended
4390 // way of doing this type of interaction. Additionally, we should
4391 // consolidate the eight methods that do the yield operation and they
4392 // are almost identical into one for better maintenability and
4393 // readability. See 6445193.
4394 //
4395 // Tony 2006.06.29
4396 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4397 ConcurrentMarkSweepThread::should_yield() &&
4398 !CMSCollector::foregroundGCIsActive(); ++i) {
4399 os::sleep(Thread::current(), 1, false);
4400 ConcurrentMarkSweepThread::acknowledge_yield_request();
4401 }
4402
4403 ConcurrentMarkSweepThread::synchronize(true);
4404 _bit_map_lock->lock_without_safepoint_check();
4405 _collector->startTimer();
4406 }
4407
4408 bool CMSCollector::do_marking_mt(bool asynch) {
4409 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
4410 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers(
4411 conc_workers()->total_workers(),
4412 conc_workers()->active_workers(),
4413 Threads::number_of_non_daemon_threads());
4414 conc_workers()->set_active_workers(num_workers);
4415
4416 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4417
4418 CMSConcMarkingTask tsk(this,
4419 cms_space,
4420 asynch,
4421 conc_workers(),
4422 task_queues());
4423
4424 // Since the actual number of workers we get may be different
4425 // from the number we requested above, do we need to do anything different
4426 // below? In particular, may be we need to subclass the SequantialSubTasksDone
4427 // class?? XXX
4428 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4429
4430 // Refs discovery is already non-atomic.
4431 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4432 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
4433 conc_workers()->start_task(&tsk);
4434 while (tsk.yielded()) {
4435 tsk.coordinator_yield();
4436 conc_workers()->continue_task(&tsk);
4437 }
4438 // If the task was aborted, _restart_addr will be non-NULL
4439 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4440 while (_restart_addr != NULL) {
4441 // XXX For now we do not make use of ABORTED state and have not
4442 // yet implemented the right abort semantics (even in the original
4443 // single-threaded CMS case). That needs some more investigation
4444 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4445 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4446 // If _restart_addr is non-NULL, a marking stack overflow
4447 // occurred; we need to do a fresh marking iteration from the
4448 // indicated restart address.
4449 if (_foregroundGCIsActive && asynch) {
4450 // We may be running into repeated stack overflows, having
4451 // reached the limit of the stack size, while making very
4452 // slow forward progress. It may be best to bail out and
4453 // let the foreground collector do its job.
4454 // Clear _restart_addr, so that foreground GC
4455 // works from scratch. This avoids the headache of
4456 // a "rescan" which would otherwise be needed because
4457 // of the dirty mod union table & card table.
4458 _restart_addr = NULL;
4459 return false;
4460 }
4461 // Adjust the task to restart from _restart_addr
4462 tsk.reset(_restart_addr);
4463 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4464 _restart_addr);
4465 _restart_addr = NULL;
4466 // Get the workers going again
4467 conc_workers()->start_task(&tsk);
4468 while (tsk.yielded()) {
4469 tsk.coordinator_yield();
4470 conc_workers()->continue_task(&tsk);
4471 }
4472 }
4473 assert(tsk.completed(), "Inconsistency");
4474 assert(tsk.result() == true, "Inconsistency");
4475 return true;
4476 }
4477
4478 bool CMSCollector::do_marking_st(bool asynch) {
4479 ResourceMark rm;
4480 HandleMark hm;
4481
4482 // Temporarily make refs discovery single threaded (non-MT)
4483 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
4484 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4485 &_markStack, CMSYield && asynch);
4486 // the last argument to iterate indicates whether the iteration
4487 // should be incremental with periodic yields.
4488 _markBitMap.iterate(&markFromRootsClosure);
4489 // If _restart_addr is non-NULL, a marking stack overflow
4490 // occurred; we need to do a fresh iteration from the
4491 // indicated restart address.
4492 while (_restart_addr != NULL) {
4493 if (_foregroundGCIsActive && asynch) {
4494 // We may be running into repeated stack overflows, having
4495 // reached the limit of the stack size, while making very
4496 // slow forward progress. It may be best to bail out and
4497 // let the foreground collector do its job.
4498 // Clear _restart_addr, so that foreground GC
4499 // works from scratch. This avoids the headache of
4500 // a "rescan" which would otherwise be needed because
4501 // of the dirty mod union table & card table.
4502 _restart_addr = NULL;
4503 return false; // indicating failure to complete marking
4504 }
4505 // Deal with stack overflow:
4506 // we restart marking from _restart_addr
4507 HeapWord* ra = _restart_addr;
4508 markFromRootsClosure.reset(ra);
4509 _restart_addr = NULL;
4510 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4511 }
4512 return true;
4513 }
4514
4515 void CMSCollector::preclean() {
4516 check_correct_thread_executing();
4517 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4518 verify_work_stacks_empty();
4519 verify_overflow_empty();
4520 _abort_preclean = false;
4521 if (CMSPrecleaningEnabled) {
4522 if (!CMSEdenChunksRecordAlways) {
4523 _eden_chunk_index = 0;
4524 }
4525 size_t used = get_eden_used();
4526 size_t capacity = get_eden_capacity();
4527 // Don't start sampling unless we will get sufficiently
4528 // many samples.
4529 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4530 * CMSScheduleRemarkEdenPenetration)) {
4531 _start_sampling = true;
4532 } else {
4533 _start_sampling = false;
4534 }
4535 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4536 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
4537 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4538 }
4539 CMSTokenSync x(true); // is cms thread
4540 if (CMSPrecleaningEnabled) {
4541 sample_eden();
4542 _collectorState = AbortablePreclean;
4543 } else {
4544 _collectorState = FinalMarking;
4545 }
4546 verify_work_stacks_empty();
4547 verify_overflow_empty();
4548 }
4549
4550 // Try and schedule the remark such that young gen
4551 // occupancy is CMSScheduleRemarkEdenPenetration %.
4552 void CMSCollector::abortable_preclean() {
4553 check_correct_thread_executing();
4554 assert(CMSPrecleaningEnabled, "Inconsistent control state");
4555 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4556
4557 // If Eden's current occupancy is below this threshold,
4558 // immediately schedule the remark; else preclean
4559 // past the next scavenge in an effort to
4560 // schedule the pause as described avove. By choosing
4561 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4562 // we will never do an actual abortable preclean cycle.
4563 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4564 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4565 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
4566 // We need more smarts in the abortable preclean
4567 // loop below to deal with cases where allocation
4568 // in young gen is very very slow, and our precleaning
4569 // is running a losing race against a horde of
4570 // mutators intent on flooding us with CMS updates
4571 // (dirty cards).
4572 // One, admittedly dumb, strategy is to give up
4573 // after a certain number of abortable precleaning loops
4574 // or after a certain maximum time. We want to make
4575 // this smarter in the next iteration.
4576 // XXX FIX ME!!! YSR
4577 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4578 while (!(should_abort_preclean() ||
4579 ConcurrentMarkSweepThread::should_terminate())) {
4580 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4581 cumworkdone += workdone;
4582 loops++;
4583 // Voluntarily terminate abortable preclean phase if we have
4584 // been at it for too long.
4585 if ((CMSMaxAbortablePrecleanLoops != 0) &&
4586 loops >= CMSMaxAbortablePrecleanLoops) {
4587 if (PrintGCDetails) {
4588 gclog_or_tty->print(" CMS: abort preclean due to loops ");
4589 }
4590 break;
4591 }
4592 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4593 if (PrintGCDetails) {
4594 gclog_or_tty->print(" CMS: abort preclean due to time ");
4595 }
4596 break;
4597 }
4598 // If we are doing little work each iteration, we should
4599 // take a short break.
4600 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4601 // Sleep for some time, waiting for work to accumulate
4602 stopTimer();
4603 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4604 startTimer();
4605 waited++;
4606 }
4607 }
4608 if (PrintCMSStatistics > 0) {
4609 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4610 loops, waited, cumworkdone);
4611 }
4612 }
4613 CMSTokenSync x(true); // is cms thread
4614 if (_collectorState != Idling) {
4615 assert(_collectorState == AbortablePreclean,
4616 "Spontaneous state transition?");
4617 _collectorState = FinalMarking;
4618 } // Else, a foreground collection completed this CMS cycle.
4619 return;
4620 }
4621
4622 // Respond to an Eden sampling opportunity
4623 void CMSCollector::sample_eden() {
4624 // Make sure a young gc cannot sneak in between our
4625 // reading and recording of a sample.
4626 assert(Thread::current()->is_ConcurrentGC_thread(),
4627 "Only the cms thread may collect Eden samples");
4628 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4629 "Should collect samples while holding CMS token");
4630 if (!_start_sampling) {
4631 return;
4632 }
4633 // When CMSEdenChunksRecordAlways is true, the eden chunk array
4634 // is populated by the young generation.
4635 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
4636 if (_eden_chunk_index < _eden_chunk_capacity) {
4637 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
4638 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4639 "Unexpected state of Eden");
4640 // We'd like to check that what we just sampled is an oop-start address;
4641 // however, we cannot do that here since the object may not yet have been
4642 // initialized. So we'll instead do the check when we _use_ this sample
4643 // later.
4644 if (_eden_chunk_index == 0 ||
4645 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4646 _eden_chunk_array[_eden_chunk_index-1])
4647 >= CMSSamplingGrain)) {
4648 _eden_chunk_index++; // commit sample
4649 }
4650 }
4651 }
4652 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4653 size_t used = get_eden_used();
4654 size_t capacity = get_eden_capacity();
4655 assert(used <= capacity, "Unexpected state of Eden");
4656 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4657 _abort_preclean = true;
4658 }
4659 }
4660 }
4661
4662
4663 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4664 assert(_collectorState == Precleaning ||
4665 _collectorState == AbortablePreclean, "incorrect state");
4666 ResourceMark rm;
4667 HandleMark hm;
4668
4669 // Precleaning is currently not MT but the reference processor
4670 // may be set for MT. Disable it temporarily here.
4671 ReferenceProcessor* rp = ref_processor();
4672 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
4673
4674 // Do one pass of scrubbing the discovered reference lists
4675 // to remove any reference objects with strongly-reachable
4676 // referents.
4677 if (clean_refs) {
4678 CMSPrecleanRefsYieldClosure yield_cl(this);
4679 assert(rp->span().equals(_span), "Spans should be equal");
4680 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4681 &_markStack, true /* preclean */);
4682 CMSDrainMarkingStackClosure complete_trace(this,
4683 _span, &_markBitMap, &_markStack,
4684 &keep_alive, true /* preclean */);
4685
4686 // We don't want this step to interfere with a young
4687 // collection because we don't want to take CPU
4688 // or memory bandwidth away from the young GC threads
4689 // (which may be as many as there are CPUs).
4690 // Note that we don't need to protect ourselves from
4691 // interference with mutators because they can't
4692 // manipulate the discovered reference lists nor affect
4693 // the computed reachability of the referents, the
4694 // only properties manipulated by the precleaning
4695 // of these reference lists.
4696 stopTimer();
4697 CMSTokenSyncWithLocks x(true /* is cms thread */,
4698 bitMapLock());
4699 startTimer();
4700 sample_eden();
4701
4702 // The following will yield to allow foreground
4703 // collection to proceed promptly. XXX YSR:
4704 // The code in this method may need further
4705 // tweaking for better performance and some restructuring
4706 // for cleaner interfaces.
4707 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
4708 rp->preclean_discovered_references(
4709 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
4710 gc_timer, _gc_tracer_cm->gc_id());
4711 }
4712
4713 if (clean_survivor) { // preclean the active survivor space(s)
4714 assert(_young_gen->kind() == Generation::DefNew ||
4715 _young_gen->kind() == Generation::ParNew ||
4716 _young_gen->kind() == Generation::ASParNew,
4717 "incorrect type for cast");
4718 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4719 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4720 &_markBitMap, &_modUnionTable,
4721 &_markStack, true /* precleaning phase */);
4722 stopTimer();
4723 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4724 bitMapLock());
4725 startTimer();
4726 unsigned int before_count =
4727 GenCollectedHeap::heap()->total_collections();
4728 SurvivorSpacePrecleanClosure
4729 sss_cl(this, _span, &_markBitMap, &_markStack,
4730 &pam_cl, before_count, CMSYield);
4731 dng->from()->object_iterate_careful(&sss_cl);
4732 dng->to()->object_iterate_careful(&sss_cl);
4733 }
4734 MarkRefsIntoAndScanClosure
4735 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4736 &_markStack, this, CMSYield,
4737 true /* precleaning phase */);
4738 // CAUTION: The following closure has persistent state that may need to
4739 // be reset upon a decrease in the sequence of addresses it
4740 // processes.
4741 ScanMarkedObjectsAgainCarefullyClosure
4742 smoac_cl(this, _span,
4743 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
4744
4745 // Preclean dirty cards in ModUnionTable and CardTable using
4746 // appropriate convergence criterion;
4747 // repeat CMSPrecleanIter times unless we find that
4748 // we are losing.
4749 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4750 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4751 "Bad convergence multiplier");
4752 assert(CMSPrecleanThreshold >= 100,
4753 "Unreasonably low CMSPrecleanThreshold");
4754
4755 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4756 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4757 numIter < CMSPrecleanIter;
4758 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4759 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4760 if (Verbose && PrintGCDetails) {
4761 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4762 }
4763 // Either there are very few dirty cards, so re-mark
4764 // pause will be small anyway, or our pre-cleaning isn't
4765 // that much faster than the rate at which cards are being
4766 // dirtied, so we might as well stop and re-mark since
4767 // precleaning won't improve our re-mark time by much.
4768 if (curNumCards <= CMSPrecleanThreshold ||
4769 (numIter > 0 &&
4770 (curNumCards * CMSPrecleanDenominator >
4771 lastNumCards * CMSPrecleanNumerator))) {
4772 numIter++;
4773 cumNumCards += curNumCards;
4774 break;
4775 }
4776 }
4777
4778 preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
4779
4780 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4781 cumNumCards += curNumCards;
4782 if (PrintGCDetails && PrintCMSStatistics != 0) {
4783 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4784 curNumCards, cumNumCards, numIter);
4785 }
4786 return cumNumCards; // as a measure of useful work done
4787 }
4788
4789 // PRECLEANING NOTES:
4790 // Precleaning involves:
4791 // . reading the bits of the modUnionTable and clearing the set bits.
4792 // . For the cards corresponding to the set bits, we scan the
4793 // objects on those cards. This means we need the free_list_lock
4794 // so that we can safely iterate over the CMS space when scanning
4795 // for oops.
4796 // . When we scan the objects, we'll be both reading and setting
4797 // marks in the marking bit map, so we'll need the marking bit map.
4798 // . For protecting _collector_state transitions, we take the CGC_lock.
4799 // Note that any races in the reading of of card table entries by the
4800 // CMS thread on the one hand and the clearing of those entries by the
4801 // VM thread or the setting of those entries by the mutator threads on the
4802 // other are quite benign. However, for efficiency it makes sense to keep
4803 // the VM thread from racing with the CMS thread while the latter is
4804 // dirty card info to the modUnionTable. We therefore also use the
4805 // CGC_lock to protect the reading of the card table and the mod union
4806 // table by the CM thread.
4807 // . We run concurrently with mutator updates, so scanning
4808 // needs to be done carefully -- we should not try to scan
4809 // potentially uninitialized objects.
4810 //
4811 // Locking strategy: While holding the CGC_lock, we scan over and
4812 // reset a maximal dirty range of the mod union / card tables, then lock
4813 // the free_list_lock and bitmap lock to do a full marking, then
4814 // release these locks; and repeat the cycle. This allows for a
4815 // certain amount of fairness in the sharing of these locks between
4816 // the CMS collector on the one hand, and the VM thread and the
4817 // mutators on the other.
4818
4819 // NOTE: preclean_mod_union_table() and preclean_card_table()
4820 // further below are largely identical; if you need to modify
4821 // one of these methods, please check the other method too.
4822
4823 size_t CMSCollector::preclean_mod_union_table(
4824 ConcurrentMarkSweepGeneration* gen,
4825 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4826 verify_work_stacks_empty();
4827 verify_overflow_empty();
4828
4829 // strategy: starting with the first card, accumulate contiguous
4830 // ranges of dirty cards; clear these cards, then scan the region
4831 // covered by these cards.
4832
4833 // Since all of the MUT is committed ahead, we can just use
4834 // that, in case the generations expand while we are precleaning.
4835 // It might also be fine to just use the committed part of the
4836 // generation, but we might potentially miss cards when the
4837 // generation is rapidly expanding while we are in the midst
4838 // of precleaning.
4839 HeapWord* startAddr = gen->reserved().start();
4840 HeapWord* endAddr = gen->reserved().end();
4841
4842 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4843
4844 size_t numDirtyCards, cumNumDirtyCards;
4845 HeapWord *nextAddr, *lastAddr;
4846 for (cumNumDirtyCards = numDirtyCards = 0,
4847 nextAddr = lastAddr = startAddr;
4848 nextAddr < endAddr;
4849 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4850
4851 ResourceMark rm;
4852 HandleMark hm;
4853
4854 MemRegion dirtyRegion;
4855 {
4856 stopTimer();
4857 // Potential yield point
4858 CMSTokenSync ts(true);
4859 startTimer();
4860 sample_eden();
4861 // Get dirty region starting at nextOffset (inclusive),
4862 // simultaneously clearing it.
4863 dirtyRegion =
4864 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4865 assert(dirtyRegion.start() >= nextAddr,
4866 "returned region inconsistent?");
4867 }
4868 // Remember where the next search should begin.
4869 // The returned region (if non-empty) is a right open interval,
4870 // so lastOffset is obtained from the right end of that
4871 // interval.
4872 lastAddr = dirtyRegion.end();
4873 // Should do something more transparent and less hacky XXX
4874 numDirtyCards =
4875 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4876
4877 // We'll scan the cards in the dirty region (with periodic
4878 // yields for foreground GC as needed).
4879 if (!dirtyRegion.is_empty()) {
4880 assert(numDirtyCards > 0, "consistency check");
4881 HeapWord* stop_point = NULL;
4882 stopTimer();
4883 // Potential yield point
4884 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4885 bitMapLock());
4886 startTimer();
4887 {
4888 verify_work_stacks_empty();
4889 verify_overflow_empty();
4890 sample_eden();
4891 stop_point =
4892 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4893 }
4894 if (stop_point != NULL) {
4895 // The careful iteration stopped early either because it found an
4896 // uninitialized object, or because we were in the midst of an
4897 // "abortable preclean", which should now be aborted. Redirty
4898 // the bits corresponding to the partially-scanned or unscanned
4899 // cards. We'll either restart at the next block boundary or
4900 // abort the preclean.
4901 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4902 "Should only be AbortablePreclean.");
4903 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4904 if (should_abort_preclean()) {
4905 break; // out of preclean loop
4906 } else {
4907 // Compute the next address at which preclean should pick up;
4908 // might need bitMapLock in order to read P-bits.
4909 lastAddr = next_card_start_after_block(stop_point);
4910 }
4911 }
4912 } else {
4913 assert(lastAddr == endAddr, "consistency check");
4914 assert(numDirtyCards == 0, "consistency check");
4915 break;
4916 }
4917 }
4918 verify_work_stacks_empty();
4919 verify_overflow_empty();
4920 return cumNumDirtyCards;
4921 }
4922
4923 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4924 // below are largely identical; if you need to modify
4925 // one of these methods, please check the other method too.
4926
4927 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4928 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4929 // strategy: it's similar to precleamModUnionTable above, in that
4930 // we accumulate contiguous ranges of dirty cards, mark these cards
4931 // precleaned, then scan the region covered by these cards.
4932 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4933 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4934
4935 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4936
4937 size_t numDirtyCards, cumNumDirtyCards;
4938 HeapWord *lastAddr, *nextAddr;
4939
4940 for (cumNumDirtyCards = numDirtyCards = 0,
4941 nextAddr = lastAddr = startAddr;
4942 nextAddr < endAddr;
4943 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4944
4945 ResourceMark rm;
4946 HandleMark hm;
4947
4948 MemRegion dirtyRegion;
4949 {
4950 // See comments in "Precleaning notes" above on why we
4951 // do this locking. XXX Could the locking overheads be
4952 // too high when dirty cards are sparse? [I don't think so.]
4953 stopTimer();
4954 CMSTokenSync x(true); // is cms thread
4955 startTimer();
4956 sample_eden();
4957 // Get and clear dirty region from card table
4958 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4959 MemRegion(nextAddr, endAddr),
4960 true,
4961 CardTableModRefBS::precleaned_card_val());
4962
4963 assert(dirtyRegion.start() >= nextAddr,
4964 "returned region inconsistent?");
4965 }
4966 lastAddr = dirtyRegion.end();
4967 numDirtyCards =
4968 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4969
4970 if (!dirtyRegion.is_empty()) {
4971 stopTimer();
4972 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4973 startTimer();
4974 sample_eden();
4975 verify_work_stacks_empty();
4976 verify_overflow_empty();
4977 HeapWord* stop_point =
4978 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4979 if (stop_point != NULL) {
4980 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4981 "Should only be AbortablePreclean.");
4982 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4983 if (should_abort_preclean()) {
4984 break; // out of preclean loop
4985 } else {
4986 // Compute the next address at which preclean should pick up.
4987 lastAddr = next_card_start_after_block(stop_point);
4988 }
4989 }
4990 } else {
4991 break;
4992 }
4993 }
4994 verify_work_stacks_empty();
4995 verify_overflow_empty();
4996 return cumNumDirtyCards;
4997 }
4998
4999 class PrecleanKlassClosure : public KlassClosure {
5000 KlassToOopClosure _cm_klass_closure;
5001 public:
5002 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
5003 void do_klass(Klass* k) {
5004 if (k->has_accumulated_modified_oops()) {
5005 k->clear_accumulated_modified_oops();
5006
5007 _cm_klass_closure.do_klass(k);
5008 }
5009 }
5010 };
5011
5012 // The freelist lock is needed to prevent asserts, is it really needed?
5013 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
5014
5015 cl->set_freelistLock(freelistLock);
5016
5017 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
5018
5019 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
5020 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
5021 PrecleanKlassClosure preclean_klass_closure(cl);
5022 ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
5023
5024 verify_work_stacks_empty();
5025 verify_overflow_empty();
5026 }
5027
5028 void CMSCollector::checkpointRootsFinal(bool asynch,
5029 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
5030 assert(_collectorState == FinalMarking, "incorrect state transition?");
5031 check_correct_thread_executing();
5032 // world is stopped at this checkpoint
5033 assert(SafepointSynchronize::is_at_safepoint(),
5034 "world should be stopped");
5035 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5036
5037 verify_work_stacks_empty();
5038 verify_overflow_empty();
5039
5040 SpecializationStats::clear();
5041 if (PrintGCDetails) {
5042 gclog_or_tty->print("[YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)]",
5043 _young_gen->used() / K,
5044 _young_gen->capacity() / K);
5045 }
5046 if (asynch) {
5047 if (CMSScavengeBeforeRemark) {
5048 GenCollectedHeap* gch = GenCollectedHeap::heap();
5049 // Temporarily set flag to false, GCH->do_collection will
5050 // expect it to be false and set to true
5051 FlagSetting fl(gch->_is_gc_active, false);
5052 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
5053 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
5054 int level = _cmsGen->level() - 1;
5055 if (level >= 0) {
5056 gch->do_collection(true, // full (i.e. force, see below)
5057 false, // !clear_all_soft_refs
5058 0, // size
5059 false, // is_tlab
5060 level // max_level
5061 );
5062 }
5063 }
5064 FreelistLocker x(this);
5065 MutexLockerEx y(bitMapLock(),
5066 Mutex::_no_safepoint_check_flag);
5067 assert(!init_mark_was_synchronous, "but that's impossible!");
5068 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
5069 } else {
5070 // already have all the locks
5071 checkpointRootsFinalWork(asynch, clear_all_soft_refs,
5072 init_mark_was_synchronous);
5073 }
5074 verify_work_stacks_empty();
5075 verify_overflow_empty();
5076 SpecializationStats::print();
5077 }
5078
5079 void CMSCollector::checkpointRootsFinalWork(bool asynch,
5080 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
5081
5082 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
5083
5084 assert(haveFreelistLocks(), "must have free list locks");
5085 assert_lock_strong(bitMapLock());
5086
5087 if (UseAdaptiveSizePolicy) {
5088 size_policy()->checkpoint_roots_final_begin();
5089 }
5090
5091 ResourceMark rm;
5092 HandleMark hm;
5093
5094 GenCollectedHeap* gch = GenCollectedHeap::heap();
5095
5096 if (should_unload_classes()) {
5097 CodeCache::gc_prologue();
5098 }
5099 assert(haveFreelistLocks(), "must have free list locks");
5100 assert_lock_strong(bitMapLock());
5101
5102 if (!init_mark_was_synchronous) {
5103 // We might assume that we need not fill TLAB's when
5104 // CMSScavengeBeforeRemark is set, because we may have just done
5105 // a scavenge which would have filled all TLAB's -- and besides
5106 // Eden would be empty. This however may not always be the case --
5107 // for instance although we asked for a scavenge, it may not have
5108 // happened because of a JNI critical section. We probably need
5109 // a policy for deciding whether we can in that case wait until
5110 // the critical section releases and then do the remark following
5111 // the scavenge, and skip it here. In the absence of that policy,
5112 // or of an indication of whether the scavenge did indeed occur,
5113 // we cannot rely on TLAB's having been filled and must do
5114 // so here just in case a scavenge did not happen.
5115 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
5116 // Update the saved marks which may affect the root scans.
5117 gch->save_marks();
5118
5119 if (CMSPrintEdenSurvivorChunks) {
5120 print_eden_and_survivor_chunk_arrays();
5121 }
5122
5123 {
5124 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
5125
5126 // Note on the role of the mod union table:
5127 // Since the marker in "markFromRoots" marks concurrently with
5128 // mutators, it is possible for some reachable objects not to have been
5129 // scanned. For instance, an only reference to an object A was
5130 // placed in object B after the marker scanned B. Unless B is rescanned,
5131 // A would be collected. Such updates to references in marked objects
5132 // are detected via the mod union table which is the set of all cards
5133 // dirtied since the first checkpoint in this GC cycle and prior to
5134 // the most recent young generation GC, minus those cleaned up by the
5135 // concurrent precleaning.
5136 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
5137 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5138 do_remark_parallel();
5139 } else {
5140 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
5141 _gc_timer_cm, _gc_tracer_cm->gc_id());
5142 do_remark_non_parallel();
5143 }
5144 }
5145 } else {
5146 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
5147 // The initial mark was stop-world, so there's no rescanning to
5148 // do; go straight on to the next step below.
5149 }
5150 verify_work_stacks_empty();
5151 verify_overflow_empty();
5152
5153 {
5154 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
5155 refProcessingWork(asynch, clear_all_soft_refs);
5156 }
5157 verify_work_stacks_empty();
5158 verify_overflow_empty();
5159
5160 if (should_unload_classes()) {
5161 CodeCache::gc_epilogue();
5162 }
5163 JvmtiExport::gc_epilogue();
5164
5165 // If we encountered any (marking stack / work queue) overflow
5166 // events during the current CMS cycle, take appropriate
5167 // remedial measures, where possible, so as to try and avoid
5168 // recurrence of that condition.
5169 assert(_markStack.isEmpty(), "No grey objects");
5170 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
5171 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
5172 if (ser_ovflw > 0) {
5173 if (PrintCMSStatistics != 0) {
5174 gclog_or_tty->print_cr("Marking stack overflow (benign) "
5175 "(pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT
5176 ", kac_preclean=" SIZE_FORMAT ")",
5177 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
5178 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
5179 }
5180 _markStack.expand();
5181 _ser_pmc_remark_ovflw = 0;
5182 _ser_pmc_preclean_ovflw = 0;
5183 _ser_kac_preclean_ovflw = 0;
5184 _ser_kac_ovflw = 0;
5185 }
5186 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
5187 if (PrintCMSStatistics != 0) {
5188 gclog_or_tty->print_cr("Work queue overflow (benign) "
5189 "(pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")",
5190 _par_pmc_remark_ovflw, _par_kac_ovflw);
5191 }
5192 _par_pmc_remark_ovflw = 0;
5193 _par_kac_ovflw = 0;
5194 }
5195 if (PrintCMSStatistics != 0) {
5196 if (_markStack._hit_limit > 0) {
5197 gclog_or_tty->print_cr(" (benign) Hit max stack size limit (" SIZE_FORMAT ")",
5198 _markStack._hit_limit);
5199 }
5200 if (_markStack._failed_double > 0) {
5201 gclog_or_tty->print_cr(" (benign) Failed stack doubling (" SIZE_FORMAT "),"
5202 " current capacity " SIZE_FORMAT,
5203 _markStack._failed_double,
5204 _markStack.capacity());
5205 }
5206 }
5207 _markStack._hit_limit = 0;
5208 _markStack._failed_double = 0;
5209
5210 if ((VerifyAfterGC || VerifyDuringGC) &&
5211 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5212 verify_after_remark();
5213 }
5214
5215 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
5216
5217 // Change under the freelistLocks.
5218 _collectorState = Sweeping;
5219 // Call isAllClear() under bitMapLock
5220 assert(_modUnionTable.isAllClear(),
5221 "Should be clear by end of the final marking");
5222 assert(_ct->klass_rem_set()->mod_union_is_clear(),
5223 "Should be clear by end of the final marking");
5224 if (UseAdaptiveSizePolicy) {
5225 size_policy()->checkpoint_roots_final_end(gch->gc_cause());
5226 }
5227 }
5228
5229 void CMSParInitialMarkTask::work(uint worker_id) {
5230 elapsedTimer _timer;
5231 ResourceMark rm;
5232 HandleMark hm;
5233
5234 // ---------- scan from roots --------------
5235 _timer.start();
5236 GenCollectedHeap* gch = GenCollectedHeap::heap();
5237 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
5238
5239 // ---------- young gen roots --------------
5240 {
5241 work_on_young_gen_roots(worker_id, &par_mri_cl);
5242 _timer.stop();
5243 if (PrintCMSStatistics != 0) {
5244 gclog_or_tty->print_cr(
5245 "Finished young gen initial mark scan work in %dth thread: %3.3f sec",
5246 worker_id, _timer.seconds());
5247 }
5248 }
5249
5250 // ---------- remaining roots --------------
5251 _timer.reset();
5252 _timer.start();
5253
5254 CLDToOopClosure cld_closure(&par_mri_cl, true);
5255
5256 gch->gen_process_roots(_collector->_cmsGen->level(),
5257 false, // yg was scanned above
5258 false, // this is parallel code
5259 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5260 _collector->should_unload_classes(),
5261 &par_mri_cl,
5262 NULL,
5263 &cld_closure);
5264 assert(_collector->should_unload_classes()
5265 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
5266 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5267 _timer.stop();
5268 if (PrintCMSStatistics != 0) {
5269 gclog_or_tty->print_cr(
5270 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec",
5271 worker_id, _timer.seconds());
5272 }
5273 }
5274
5275 // Parallel remark task
5276 class CMSParRemarkTask: public CMSParMarkTask {
5277 CompactibleFreeListSpace* _cms_space;
5278
5279 // The per-thread work queues, available here for stealing.
5280 OopTaskQueueSet* _task_queues;
5281 ParallelTaskTerminator _term;
5282
5283 public:
5284 // A value of 0 passed to n_workers will cause the number of
5285 // workers to be taken from the active workers in the work gang.
5286 CMSParRemarkTask(CMSCollector* collector,
5287 CompactibleFreeListSpace* cms_space,
5288 int n_workers, FlexibleWorkGang* workers,
5289 OopTaskQueueSet* task_queues):
5290 CMSParMarkTask("Rescan roots and grey objects in parallel",
5291 collector, n_workers),
5292 _cms_space(cms_space),
5293 _task_queues(task_queues),
5294 _term(n_workers, task_queues) { }
5295
5296 OopTaskQueueSet* task_queues() { return _task_queues; }
5297
5298 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5299
5300 ParallelTaskTerminator* terminator() { return &_term; }
5301 int n_workers() { return _n_workers; }
5302
5303 void work(uint worker_id);
5304
5305 private:
5306 // ... of dirty cards in old space
5307 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
5308 Par_MarkRefsIntoAndScanClosure* cl);
5309
5310 // ... work stealing for the above
5311 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
5312 };
5313
5314 class RemarkKlassClosure : public KlassClosure {
5315 KlassToOopClosure _cm_klass_closure;
5316 public:
5317 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
5318 void do_klass(Klass* k) {
5319 // Check if we have modified any oops in the Klass during the concurrent marking.
5320 if (k->has_accumulated_modified_oops()) {
5321 k->clear_accumulated_modified_oops();
5322
5323 // We could have transfered the current modified marks to the accumulated marks,
5324 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
5325 } else if (k->has_modified_oops()) {
5326 // Don't clear anything, this info is needed by the next young collection.
5327 } else {
5328 // No modified oops in the Klass.
5329 return;
5330 }
5331
5332 // The klass has modified fields, need to scan the klass.
5333 _cm_klass_closure.do_klass(k);
5334 }
5335 };
5336
5337 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
5338 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
5339 EdenSpace* eden_space = dng->eden();
5340 ContiguousSpace* from_space = dng->from();
5341 ContiguousSpace* to_space = dng->to();
5342
5343 HeapWord** eca = _collector->_eden_chunk_array;
5344 size_t ect = _collector->_eden_chunk_index;
5345 HeapWord** sca = _collector->_survivor_chunk_array;
5346 size_t sct = _collector->_survivor_chunk_index;
5347
5348 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
5349 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
5350
5351 do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
5352 do_young_space_rescan(worker_id, cl, from_space, sca, sct);
5353 do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
5354 }
5355
5356 // work_queue(i) is passed to the closure
5357 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
5358 // also is passed to do_dirty_card_rescan_tasks() and to
5359 // do_work_steal() to select the i-th task_queue.
5360
5361 void CMSParRemarkTask::work(uint worker_id) {
5362 elapsedTimer _timer;
5363 ResourceMark rm;
5364 HandleMark hm;
5365
5366 // ---------- rescan from roots --------------
5367 _timer.start();
5368 GenCollectedHeap* gch = GenCollectedHeap::heap();
5369 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
5370 _collector->_span, _collector->ref_processor(),
5371 &(_collector->_markBitMap),
5372 work_queue(worker_id));
5373
5374 // Rescan young gen roots first since these are likely
5375 // coarsely partitioned and may, on that account, constitute
5376 // the critical path; thus, it's best to start off that
5377 // work first.
5378 // ---------- young gen roots --------------
5379 {
5380 work_on_young_gen_roots(worker_id, &par_mrias_cl);
5381 _timer.stop();
5382 if (PrintCMSStatistics != 0) {
5383 gclog_or_tty->print_cr(
5384 "Finished young gen rescan work in %dth thread: %3.3f sec",
5385 worker_id, _timer.seconds());
5386 }
5387 }
5388
5389 // ---------- remaining roots --------------
5390 _timer.reset();
5391 _timer.start();
5392 gch->gen_process_roots(_collector->_cmsGen->level(),
5393 false, // yg was scanned above
5394 false, // this is parallel code
5395 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5396 _collector->should_unload_classes(),
5397 &par_mrias_cl,
5398 NULL,
5399 NULL); // The dirty klasses will be handled below
5400
5401 assert(_collector->should_unload_classes()
5402 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
5403 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5404 _timer.stop();
5405 if (PrintCMSStatistics != 0) {
5406 gclog_or_tty->print_cr(
5407 "Finished remaining root rescan work in %dth thread: %3.3f sec",
5408 worker_id, _timer.seconds());
5409 }
5410
5411 // ---------- unhandled CLD scanning ----------
5412 if (worker_id == 0) { // Single threaded at the moment.
5413 _timer.reset();
5414 _timer.start();
5415
5416 // Scan all new class loader data objects and new dependencies that were
5417 // introduced during concurrent marking.
5418 ResourceMark rm;
5419 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5420 for (int i = 0; i < array->length(); i++) {
5421 par_mrias_cl.do_class_loader_data(array->at(i));
5422 }
5423
5424 // We don't need to keep track of new CLDs anymore.
5425 ClassLoaderDataGraph::remember_new_clds(false);
5426
5427 _timer.stop();
5428 if (PrintCMSStatistics != 0) {
5429 gclog_or_tty->print_cr(
5430 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec",
5431 worker_id, _timer.seconds());
5432 }
5433 }
5434
5435 // ---------- dirty klass scanning ----------
5436 if (worker_id == 0) { // Single threaded at the moment.
5437 _timer.reset();
5438 _timer.start();
5439
5440 // Scan all classes that was dirtied during the concurrent marking phase.
5441 RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
5442 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5443
5444 _timer.stop();
5445 if (PrintCMSStatistics != 0) {
5446 gclog_or_tty->print_cr(
5447 "Finished dirty klass scanning work in %dth thread: %3.3f sec",
5448 worker_id, _timer.seconds());
5449 }
5450 }
5451
5452 // We might have added oops to ClassLoaderData::_handles during the
5453 // concurrent marking phase. These oops point to newly allocated objects
5454 // that are guaranteed to be kept alive. Either by the direct allocation
5455 // code, or when the young collector processes the roots. Hence,
5456 // we don't have to revisit the _handles block during the remark phase.
5457
5458 // ---------- rescan dirty cards ------------
5459 _timer.reset();
5460 _timer.start();
5461
5462 // Do the rescan tasks for each of the two spaces
5463 // (cms_space) in turn.
5464 // "worker_id" is passed to select the task_queue for "worker_id"
5465 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
5466 _timer.stop();
5467 if (PrintCMSStatistics != 0) {
5468 gclog_or_tty->print_cr(
5469 "Finished dirty card rescan work in %dth thread: %3.3f sec",
5470 worker_id, _timer.seconds());
5471 }
5472
5473 // ---------- steal work from other threads ...
5474 // ---------- ... and drain overflow list.
5475 _timer.reset();
5476 _timer.start();
5477 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
5478 _timer.stop();
5479 if (PrintCMSStatistics != 0) {
5480 gclog_or_tty->print_cr(
5481 "Finished work stealing in %dth thread: %3.3f sec",
5482 worker_id, _timer.seconds());
5483 }
5484 }
5485
5486 // Note that parameter "i" is not used.
5487 void
5488 CMSParMarkTask::do_young_space_rescan(uint worker_id,
5489 OopsInGenClosure* cl, ContiguousSpace* space,
5490 HeapWord** chunk_array, size_t chunk_top) {
5491 // Until all tasks completed:
5492 // . claim an unclaimed task
5493 // . compute region boundaries corresponding to task claimed
5494 // using chunk_array
5495 // . par_oop_iterate(cl) over that region
5496
5497 ResourceMark rm;
5498 HandleMark hm;
5499
5500 SequentialSubTasksDone* pst = space->par_seq_tasks();
5501
5502 uint nth_task = 0;
5503 uint n_tasks = pst->n_tasks();
5504
5505 if (n_tasks > 0) {
5506 assert(pst->valid(), "Uninitialized use?");
5507 HeapWord *start, *end;
5508 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5509 // We claimed task # nth_task; compute its boundaries.
5510 if (chunk_top == 0) { // no samples were taken
5511 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
5512 start = space->bottom();
5513 end = space->top();
5514 } else if (nth_task == 0) {
5515 start = space->bottom();
5516 end = chunk_array[nth_task];
5517 } else if (nth_task < (uint)chunk_top) {
5518 assert(nth_task >= 1, "Control point invariant");
5519 start = chunk_array[nth_task - 1];
5520 end = chunk_array[nth_task];
5521 } else {
5522 assert(nth_task == (uint)chunk_top, "Control point invariant");
5523 start = chunk_array[chunk_top - 1];
5524 end = space->top();
5525 }
5526 MemRegion mr(start, end);
5527 // Verify that mr is in space
5528 assert(mr.is_empty() || space->used_region().contains(mr),
5529 "Should be in space");
5530 // Verify that "start" is an object boundary
5531 assert(mr.is_empty() || oop(mr.start())->is_oop(),
5532 "Should be an oop");
5533 space->par_oop_iterate(mr, cl);
5534 }
5535 pst->all_tasks_completed();
5536 }
5537 }
5538
5539 void
5540 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5541 CompactibleFreeListSpace* sp, int i,
5542 Par_MarkRefsIntoAndScanClosure* cl) {
5543 // Until all tasks completed:
5544 // . claim an unclaimed task
5545 // . compute region boundaries corresponding to task claimed
5546 // . transfer dirty bits ct->mut for that region
5547 // . apply rescanclosure to dirty mut bits for that region
5548
5549 ResourceMark rm;
5550 HandleMark hm;
5551
5552 OopTaskQueue* work_q = work_queue(i);
5553 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5554 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5555 // CAUTION: This closure has state that persists across calls to
5556 // the work method dirty_range_iterate_clear() in that it has
5557 // imbedded in it a (subtype of) UpwardsObjectClosure. The
5558 // use of that state in the imbedded UpwardsObjectClosure instance
5559 // assumes that the cards are always iterated (even if in parallel
5560 // by several threads) in monotonically increasing order per each
5561 // thread. This is true of the implementation below which picks
5562 // card ranges (chunks) in monotonically increasing order globally
5563 // and, a-fortiori, in monotonically increasing order per thread
5564 // (the latter order being a subsequence of the former).
5565 // If the work code below is ever reorganized into a more chaotic
5566 // work-partitioning form than the current "sequential tasks"
5567 // paradigm, the use of that persistent state will have to be
5568 // revisited and modified appropriately. See also related
5569 // bug 4756801 work on which should examine this code to make
5570 // sure that the changes there do not run counter to the
5571 // assumptions made here and necessary for correctness and
5572 // efficiency. Note also that this code might yield inefficient
5573 // behaviour in the case of very large objects that span one or
5574 // more work chunks. Such objects would potentially be scanned
5575 // several times redundantly. Work on 4756801 should try and
5576 // address that performance anomaly if at all possible. XXX
5577 MemRegion full_span = _collector->_span;
5578 CMSBitMap* bm = &(_collector->_markBitMap); // shared
5579 MarkFromDirtyCardsClosure
5580 greyRescanClosure(_collector, full_span, // entire span of interest
5581 sp, bm, work_q, cl);
5582
5583 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5584 assert(pst->valid(), "Uninitialized use?");
5585 uint nth_task = 0;
5586 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5587 MemRegion span = sp->used_region();
5588 HeapWord* start_addr = span.start();
5589 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5590 alignment);
5591 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5592 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5593 start_addr, "Check alignment");
5594 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5595 chunk_size, "Check alignment");
5596
5597 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5598 // Having claimed the nth_task, compute corresponding mem-region,
5599 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
5600 // The alignment restriction ensures that we do not need any
5601 // synchronization with other gang-workers while setting or
5602 // clearing bits in thus chunk of the MUT.
5603 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5604 start_addr + (nth_task+1)*chunk_size);
5605 // The last chunk's end might be way beyond end of the
5606 // used region. In that case pull back appropriately.
5607 if (this_span.end() > end_addr) {
5608 this_span.set_end(end_addr);
5609 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5610 }
5611 // Iterate over the dirty cards covering this chunk, marking them
5612 // precleaned, and setting the corresponding bits in the mod union
5613 // table. Since we have been careful to partition at Card and MUT-word
5614 // boundaries no synchronization is needed between parallel threads.
5615 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5616 &modUnionClosure);
5617
5618 // Having transferred these marks into the modUnionTable,
5619 // rescan the marked objects on the dirty cards in the modUnionTable.
5620 // Even if this is at a synchronous collection, the initial marking
5621 // may have been done during an asynchronous collection so there
5622 // may be dirty bits in the mod-union table.
5623 _collector->_modUnionTable.dirty_range_iterate_clear(
5624 this_span, &greyRescanClosure);
5625 _collector->_modUnionTable.verifyNoOneBitsInRange(
5626 this_span.start(),
5627 this_span.end());
5628 }
5629 pst->all_tasks_completed(); // declare that i am done
5630 }
5631
5632 // . see if we can share work_queues with ParNew? XXX
5633 void
5634 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5635 int* seed) {
5636 OopTaskQueue* work_q = work_queue(i);
5637 NOT_PRODUCT(int num_steals = 0;)
5638 oop obj_to_scan;
5639 CMSBitMap* bm = &(_collector->_markBitMap);
5640
5641 while (true) {
5642 // Completely finish any left over work from (an) earlier round(s)
5643 cl->trim_queue(0);
5644 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5645 (size_t)ParGCDesiredObjsFromOverflowList);
5646 // Now check if there's any work in the overflow list
5647 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5648 // only affects the number of attempts made to get work from the
5649 // overflow list and does not affect the number of workers. Just
5650 // pass ParallelGCThreads so this behavior is unchanged.
5651 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5652 work_q,
5653 ParallelGCThreads)) {
5654 // found something in global overflow list;
5655 // not yet ready to go stealing work from others.
5656 // We'd like to assert(work_q->size() != 0, ...)
5657 // because we just took work from the overflow list,
5658 // but of course we can't since all of that could have
5659 // been already stolen from us.
5660 // "He giveth and He taketh away."
5661 continue;
5662 }
5663 // Verify that we have no work before we resort to stealing
5664 assert(work_q->size() == 0, "Have work, shouldn't steal");
5665 // Try to steal from other queues that have work
5666 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5667 NOT_PRODUCT(num_steals++;)
5668 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5669 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5670 // Do scanning work
5671 obj_to_scan->oop_iterate(cl);
5672 // Loop around, finish this work, and try to steal some more
5673 } else if (terminator()->offer_termination()) {
5674 break; // nirvana from the infinite cycle
5675 }
5676 }
5677 NOT_PRODUCT(
5678 if (PrintCMSStatistics != 0) {
5679 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5680 }
5681 )
5682 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5683 "Else our work is not yet done");
5684 }
5685
5686 // Record object boundaries in _eden_chunk_array by sampling the eden
5687 // top in the slow-path eden object allocation code path and record
5688 // the boundaries, if CMSEdenChunksRecordAlways is true. If
5689 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
5690 // sampling in sample_eden() that activates during the part of the
5691 // preclean phase.
5692 void CMSCollector::sample_eden_chunk() {
5693 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
5694 if (_eden_chunk_lock->try_lock()) {
5695 // Record a sample. This is the critical section. The contents
5696 // of the _eden_chunk_array have to be non-decreasing in the
5697 // address order.
5698 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
5699 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
5700 "Unexpected state of Eden");
5701 if (_eden_chunk_index == 0 ||
5702 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
5703 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
5704 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
5705 _eden_chunk_index++; // commit sample
5706 }
5707 _eden_chunk_lock->unlock();
5708 }
5709 }
5710 }
5711
5712 // Return a thread-local PLAB recording array, as appropriate.
5713 void* CMSCollector::get_data_recorder(int thr_num) {
5714 if (_survivor_plab_array != NULL &&
5715 (CMSPLABRecordAlways ||
5716 (_collectorState > Marking && _collectorState < FinalMarking))) {
5717 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5718 ChunkArray* ca = &_survivor_plab_array[thr_num];
5719 ca->reset(); // clear it so that fresh data is recorded
5720 return (void*) ca;
5721 } else {
5722 return NULL;
5723 }
5724 }
5725
5726 // Reset all the thread-local PLAB recording arrays
5727 void CMSCollector::reset_survivor_plab_arrays() {
5728 for (uint i = 0; i < ParallelGCThreads; i++) {
5729 _survivor_plab_array[i].reset();
5730 }
5731 }
5732
5733 // Merge the per-thread plab arrays into the global survivor chunk
5734 // array which will provide the partitioning of the survivor space
5735 // for CMS initial scan and rescan.
5736 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
5737 int no_of_gc_threads) {
5738 assert(_survivor_plab_array != NULL, "Error");
5739 assert(_survivor_chunk_array != NULL, "Error");
5740 assert(_collectorState == FinalMarking ||
5741 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
5742 for (int j = 0; j < no_of_gc_threads; j++) {
5743 _cursor[j] = 0;
5744 }
5745 HeapWord* top = surv->top();
5746 size_t i;
5747 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
5748 HeapWord* min_val = top; // Higher than any PLAB address
5749 uint min_tid = 0; // position of min_val this round
5750 for (int j = 0; j < no_of_gc_threads; j++) {
5751 ChunkArray* cur_sca = &_survivor_plab_array[j];
5752 if (_cursor[j] == cur_sca->end()) {
5753 continue;
5754 }
5755 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5756 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5757 assert(surv->used_region().contains(cur_val), "Out of bounds value");
5758 if (cur_val < min_val) {
5759 min_tid = j;
5760 min_val = cur_val;
5761 } else {
5762 assert(cur_val < top, "All recorded addresses should be less");
5763 }
5764 }
5765 // At this point min_val and min_tid are respectively
5766 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5767 // and the thread (j) that witnesses that address.
5768 // We record this address in the _survivor_chunk_array[i]
5769 // and increment _cursor[min_tid] prior to the next round i.
5770 if (min_val == top) {
5771 break;
5772 }
5773 _survivor_chunk_array[i] = min_val;
5774 _cursor[min_tid]++;
5775 }
5776 // We are all done; record the size of the _survivor_chunk_array
5777 _survivor_chunk_index = i; // exclusive: [0, i)
5778 if (PrintCMSStatistics > 0) {
5779 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5780 }
5781 // Verify that we used up all the recorded entries
5782 #ifdef ASSERT
5783 size_t total = 0;
5784 for (int j = 0; j < no_of_gc_threads; j++) {
5785 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5786 total += _cursor[j];
5787 }
5788 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5789 // Check that the merged array is in sorted order
5790 if (total > 0) {
5791 for (size_t i = 0; i < total - 1; i++) {
5792 if (PrintCMSStatistics > 0) {
5793 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5794 i, _survivor_chunk_array[i]);
5795 }
5796 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5797 "Not sorted");
5798 }
5799 }
5800 #endif // ASSERT
5801 }
5802
5803 // Set up the space's par_seq_tasks structure for work claiming
5804 // for parallel initial scan and rescan of young gen.
5805 // See ParRescanTask where this is currently used.
5806 void
5807 CMSCollector::
5808 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5809 assert(n_threads > 0, "Unexpected n_threads argument");
5810 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5811
5812 // Eden space
5813 if (!dng->eden()->is_empty()) {
5814 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5815 assert(!pst->valid(), "Clobbering existing data?");
5816 // Each valid entry in [0, _eden_chunk_index) represents a task.
5817 size_t n_tasks = _eden_chunk_index + 1;
5818 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5819 // Sets the condition for completion of the subtask (how many threads
5820 // need to finish in order to be done).
5821 pst->set_n_threads(n_threads);
5822 pst->set_n_tasks((int)n_tasks);
5823 }
5824
5825 // Merge the survivor plab arrays into _survivor_chunk_array
5826 if (_survivor_plab_array != NULL) {
5827 merge_survivor_plab_arrays(dng->from(), n_threads);
5828 } else {
5829 assert(_survivor_chunk_index == 0, "Error");
5830 }
5831
5832 // To space
5833 {
5834 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5835 assert(!pst->valid(), "Clobbering existing data?");
5836 // Sets the condition for completion of the subtask (how many threads
5837 // need to finish in order to be done).
5838 pst->set_n_threads(n_threads);
5839 pst->set_n_tasks(1);
5840 assert(pst->valid(), "Error");
5841 }
5842
5843 // From space
5844 {
5845 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5846 assert(!pst->valid(), "Clobbering existing data?");
5847 size_t n_tasks = _survivor_chunk_index + 1;
5848 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5849 // Sets the condition for completion of the subtask (how many threads
5850 // need to finish in order to be done).
5851 pst->set_n_threads(n_threads);
5852 pst->set_n_tasks((int)n_tasks);
5853 assert(pst->valid(), "Error");
5854 }
5855 }
5856
5857 // Parallel version of remark
5858 void CMSCollector::do_remark_parallel() {
5859 GenCollectedHeap* gch = GenCollectedHeap::heap();
5860 FlexibleWorkGang* workers = gch->workers();
5861 assert(workers != NULL, "Need parallel worker threads.");
5862 // Choose to use the number of GC workers most recently set
5863 // into "active_workers". If active_workers is not set, set it
5864 // to ParallelGCThreads.
5865 int n_workers = workers->active_workers();
5866 if (n_workers == 0) {
5867 assert(n_workers > 0, "Should have been set during scavenge");
5868 n_workers = ParallelGCThreads;
5869 workers->set_active_workers(n_workers);
5870 }
5871 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5872
5873 CMSParRemarkTask tsk(this,
5874 cms_space,
5875 n_workers, workers, task_queues());
5876
5877 // Set up for parallel process_roots work.
5878 gch->set_par_threads(n_workers);
5879 // We won't be iterating over the cards in the card table updating
5880 // the younger_gen cards, so we shouldn't call the following else
5881 // the verification code as well as subsequent younger_refs_iterate
5882 // code would get confused. XXX
5883 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5884
5885 // The young gen rescan work will not be done as part of
5886 // process_roots (which currently doesn't know how to
5887 // parallelize such a scan), but rather will be broken up into
5888 // a set of parallel tasks (via the sampling that the [abortable]
5889 // preclean phase did of EdenSpace, plus the [two] tasks of
5890 // scanning the [two] survivor spaces. Further fine-grain
5891 // parallelization of the scanning of the survivor spaces
5892 // themselves, and of precleaning of the younger gen itself
5893 // is deferred to the future.
5894 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5895
5896 // The dirty card rescan work is broken up into a "sequence"
5897 // of parallel tasks (per constituent space) that are dynamically
5898 // claimed by the parallel threads.
5899 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5900
5901 // It turns out that even when we're using 1 thread, doing the work in a
5902 // separate thread causes wide variance in run times. We can't help this
5903 // in the multi-threaded case, but we special-case n=1 here to get
5904 // repeatable measurements of the 1-thread overhead of the parallel code.
5905 if (n_workers > 1) {
5906 // Make refs discovery MT-safe, if it isn't already: it may not
5907 // necessarily be so, since it's possible that we are doing
5908 // ST marking.
5909 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5910 GenCollectedHeap::StrongRootsScope srs(gch);
5911 workers->run_task(&tsk);
5912 } else {
5913 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5914 GenCollectedHeap::StrongRootsScope srs(gch);
5915 tsk.work(0);
5916 }
5917
5918 gch->set_par_threads(0); // 0 ==> non-parallel.
5919 // restore, single-threaded for now, any preserved marks
5920 // as a result of work_q overflow
5921 restore_preserved_marks_if_any();
5922 }
5923
5924 // Non-parallel version of remark
5925 void CMSCollector::do_remark_non_parallel() {
5926 ResourceMark rm;
5927 HandleMark hm;
5928 GenCollectedHeap* gch = GenCollectedHeap::heap();
5929 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5930
5931 MarkRefsIntoAndScanClosure
5932 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
5933 &_markStack, this,
5934 false /* should_yield */, false /* not precleaning */);
5935 MarkFromDirtyCardsClosure
5936 markFromDirtyCardsClosure(this, _span,
5937 NULL, // space is set further below
5938 &_markBitMap, &_markStack, &mrias_cl);
5939 {
5940 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5941 // Iterate over the dirty cards, setting the corresponding bits in the
5942 // mod union table.
5943 {
5944 ModUnionClosure modUnionClosure(&_modUnionTable);
5945 _ct->ct_bs()->dirty_card_iterate(
5946 _cmsGen->used_region(),
5947 &modUnionClosure);
5948 }
5949 // Having transferred these marks into the modUnionTable, we just need
5950 // to rescan the marked objects on the dirty cards in the modUnionTable.
5951 // The initial marking may have been done during an asynchronous
5952 // collection so there may be dirty bits in the mod-union table.
5953 const int alignment =
5954 CardTableModRefBS::card_size * BitsPerWord;
5955 {
5956 // ... First handle dirty cards in CMS gen
5957 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5958 MemRegion ur = _cmsGen->used_region();
5959 HeapWord* lb = ur.start();
5960 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5961 MemRegion cms_span(lb, ub);
5962 _modUnionTable.dirty_range_iterate_clear(cms_span,
5963 &markFromDirtyCardsClosure);
5964 verify_work_stacks_empty();
5965 if (PrintCMSStatistics != 0) {
5966 gclog_or_tty->print(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ",
5967 markFromDirtyCardsClosure.num_dirty_cards());
5968 }
5969 }
5970 }
5971 if (VerifyDuringGC &&
5972 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5973 HandleMark hm; // Discard invalid handles created during verification
5974 Universe::verify();
5975 }
5976 {
5977 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5978
5979 verify_work_stacks_empty();
5980
5981 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5982 GenCollectedHeap::StrongRootsScope srs(gch);
5983
5984 gch->gen_process_roots(_cmsGen->level(),
5985 true, // younger gens as roots
5986 false, // use the local StrongRootsScope
5987 GenCollectedHeap::ScanningOption(roots_scanning_options()),
5988 should_unload_classes(),
5989 &mrias_cl,
5990 NULL,
5991 NULL); // The dirty klasses will be handled below
5992
5993 assert(should_unload_classes()
5994 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
5995 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5996 }
5997
5998 {
5999 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
6000
6001 verify_work_stacks_empty();
6002
6003 // Scan all class loader data objects that might have been introduced
6004 // during concurrent marking.
6005 ResourceMark rm;
6006 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
6007 for (int i = 0; i < array->length(); i++) {
6008 mrias_cl.do_class_loader_data(array->at(i));
6009 }
6010
6011 // We don't need to keep track of new CLDs anymore.
6012 ClassLoaderDataGraph::remember_new_clds(false);
6013
6014 verify_work_stacks_empty();
6015 }
6016
6017 {
6018 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
6019
6020 verify_work_stacks_empty();
6021
6022 RemarkKlassClosure remark_klass_closure(&mrias_cl);
6023 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
6024
6025 verify_work_stacks_empty();
6026 }
6027
6028 // We might have added oops to ClassLoaderData::_handles during the
6029 // concurrent marking phase. These oops point to newly allocated objects
6030 // that are guaranteed to be kept alive. Either by the direct allocation
6031 // code, or when the young collector processes the roots. Hence,
6032 // we don't have to revisit the _handles block during the remark phase.
6033
6034 verify_work_stacks_empty();
6035 // Restore evacuated mark words, if any, used for overflow list links
6036 if (!CMSOverflowEarlyRestoration) {
6037 restore_preserved_marks_if_any();
6038 }
6039 verify_overflow_empty();
6040 }
6041
6042 ////////////////////////////////////////////////////////
6043 // Parallel Reference Processing Task Proxy Class
6044 ////////////////////////////////////////////////////////
6045 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
6046 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
6047 CMSCollector* _collector;
6048 CMSBitMap* _mark_bit_map;
6049 const MemRegion _span;
6050 ProcessTask& _task;
6051
6052 public:
6053 CMSRefProcTaskProxy(ProcessTask& task,
6054 CMSCollector* collector,
6055 const MemRegion& span,
6056 CMSBitMap* mark_bit_map,
6057 AbstractWorkGang* workers,
6058 OopTaskQueueSet* task_queues):
6059 // XXX Should superclass AGTWOQ also know about AWG since it knows
6060 // about the task_queues used by the AWG? Then it could initialize
6061 // the terminator() object. See 6984287. The set_for_termination()
6062 // below is a temporary band-aid for the regression in 6984287.
6063 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
6064 task_queues),
6065 _task(task),
6066 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
6067 {
6068 assert(_collector->_span.equals(_span) && !_span.is_empty(),
6069 "Inconsistency in _span");
6070 set_for_termination(workers->active_workers());
6071 }
6072
6073 OopTaskQueueSet* task_queues() { return queues(); }
6074
6075 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
6076
6077 void do_work_steal(int i,
6078 CMSParDrainMarkingStackClosure* drain,
6079 CMSParKeepAliveClosure* keep_alive,
6080 int* seed);
6081
6082 virtual void work(uint worker_id);
6083 };
6084
6085 void CMSRefProcTaskProxy::work(uint worker_id) {
6086 ResourceMark rm;
6087 HandleMark hm;
6088 assert(_collector->_span.equals(_span), "Inconsistency in _span");
6089 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
6090 _mark_bit_map,
6091 work_queue(worker_id));
6092 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
6093 _mark_bit_map,
6094 work_queue(worker_id));
6095 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
6096 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
6097 if (_task.marks_oops_alive()) {
6098 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
6099 _collector->hash_seed(worker_id));
6100 }
6101 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
6102 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
6103 }
6104
6105 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
6106 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
6107 EnqueueTask& _task;
6108
6109 public:
6110 CMSRefEnqueueTaskProxy(EnqueueTask& task)
6111 : AbstractGangTask("Enqueue reference objects in parallel"),
6112 _task(task)
6113 { }
6114
6115 virtual void work(uint worker_id)
6116 {
6117 _task.work(worker_id);
6118 }
6119 };
6120
6121 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
6122 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
6123 _span(span),
6124 _bit_map(bit_map),
6125 _work_queue(work_queue),
6126 _mark_and_push(collector, span, bit_map, work_queue),
6127 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6128 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
6129 { }
6130
6131 // . see if we can share work_queues with ParNew? XXX
6132 void CMSRefProcTaskProxy::do_work_steal(int i,
6133 CMSParDrainMarkingStackClosure* drain,
6134 CMSParKeepAliveClosure* keep_alive,
6135 int* seed) {
6136 OopTaskQueue* work_q = work_queue(i);
6137 NOT_PRODUCT(int num_steals = 0;)
6138 oop obj_to_scan;
6139
6140 while (true) {
6141 // Completely finish any left over work from (an) earlier round(s)
6142 drain->trim_queue(0);
6143 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
6144 (size_t)ParGCDesiredObjsFromOverflowList);
6145 // Now check if there's any work in the overflow list
6146 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
6147 // only affects the number of attempts made to get work from the
6148 // overflow list and does not affect the number of workers. Just
6149 // pass ParallelGCThreads so this behavior is unchanged.
6150 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
6151 work_q,
6152 ParallelGCThreads)) {
6153 // Found something in global overflow list;
6154 // not yet ready to go stealing work from others.
6155 // We'd like to assert(work_q->size() != 0, ...)
6156 // because we just took work from the overflow list,
6157 // but of course we can't, since all of that might have
6158 // been already stolen from us.
6159 continue;
6160 }
6161 // Verify that we have no work before we resort to stealing
6162 assert(work_q->size() == 0, "Have work, shouldn't steal");
6163 // Try to steal from other queues that have work
6164 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
6165 NOT_PRODUCT(num_steals++;)
6166 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
6167 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
6168 // Do scanning work
6169 obj_to_scan->oop_iterate(keep_alive);
6170 // Loop around, finish this work, and try to steal some more
6171 } else if (terminator()->offer_termination()) {
6172 break; // nirvana from the infinite cycle
6173 }
6174 }
6175 NOT_PRODUCT(
6176 if (PrintCMSStatistics != 0) {
6177 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
6178 }
6179 )
6180 }
6181
6182 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
6183 {
6184 GenCollectedHeap* gch = GenCollectedHeap::heap();
6185 FlexibleWorkGang* workers = gch->workers();
6186 assert(workers != NULL, "Need parallel worker threads.");
6187 CMSRefProcTaskProxy rp_task(task, &_collector,
6188 _collector.ref_processor()->span(),
6189 _collector.markBitMap(),
6190 workers, _collector.task_queues());
6191 workers->run_task(&rp_task);
6192 }
6193
6194 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
6195 {
6196
6197 GenCollectedHeap* gch = GenCollectedHeap::heap();
6198 FlexibleWorkGang* workers = gch->workers();
6199 assert(workers != NULL, "Need parallel worker threads.");
6200 CMSRefEnqueueTaskProxy enq_task(task);
6201 workers->run_task(&enq_task);
6202 }
6203
6204 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
6205
6206 ResourceMark rm;
6207 HandleMark hm;
6208
6209 ReferenceProcessor* rp = ref_processor();
6210 assert(rp->span().equals(_span), "Spans should be equal");
6211 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
6212 // Process weak references.
6213 rp->setup_policy(clear_all_soft_refs);
6214 verify_work_stacks_empty();
6215
6216 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
6217 &_markStack, false /* !preclean */);
6218 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
6219 _span, &_markBitMap, &_markStack,
6220 &cmsKeepAliveClosure, false /* !preclean */);
6221 {
6222 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
6223
6224 ReferenceProcessorStats stats;
6225 if (rp->processing_is_mt()) {
6226 // Set the degree of MT here. If the discovery is done MT, there
6227 // may have been a different number of threads doing the discovery
6228 // and a different number of discovered lists may have Ref objects.
6229 // That is OK as long as the Reference lists are balanced (see
6230 // balance_all_queues() and balance_queues()).
6231 GenCollectedHeap* gch = GenCollectedHeap::heap();
6232 int active_workers = ParallelGCThreads;
6233 FlexibleWorkGang* workers = gch->workers();
6234 if (workers != NULL) {
6235 active_workers = workers->active_workers();
6236 // The expectation is that active_workers will have already
6237 // been set to a reasonable value. If it has not been set,
6238 // investigate.
6239 assert(active_workers > 0, "Should have been set during scavenge");
6240 }
6241 rp->set_active_mt_degree(active_workers);
6242 CMSRefProcTaskExecutor task_executor(*this);
6243 stats = rp->process_discovered_references(&_is_alive_closure,
6244 &cmsKeepAliveClosure,
6245 &cmsDrainMarkingStackClosure,
6246 &task_executor,
6247 _gc_timer_cm,
6248 _gc_tracer_cm->gc_id());
6249 } else {
6250 stats = rp->process_discovered_references(&_is_alive_closure,
6251 &cmsKeepAliveClosure,
6252 &cmsDrainMarkingStackClosure,
6253 NULL,
6254 _gc_timer_cm,
6255 _gc_tracer_cm->gc_id());
6256 }
6257 _gc_tracer_cm->report_gc_reference_stats(stats);
6258
6259 }
6260
6261 // This is the point where the entire marking should have completed.
6262 verify_work_stacks_empty();
6263
6264 if (should_unload_classes()) {
6265 {
6266 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
6267
6268 // Unload classes and purge the SystemDictionary.
6269 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
6270
6271 // Unload nmethods.
6272 CodeCache::do_unloading(&_is_alive_closure, purged_class);
6273
6274 // Prune dead klasses from subklass/sibling/implementor lists.
6275 Klass::clean_weak_klass_links(&_is_alive_closure);
6276 }
6277
6278 {
6279 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
6280 // Clean up unreferenced symbols in symbol table.
6281 SymbolTable::unlink();
6282 }
6283
6284 {
6285 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
6286 // Delete entries for dead interned strings.
6287 StringTable::unlink(&_is_alive_closure);
6288 }
6289 }
6290
6291
6292 // Restore any preserved marks as a result of mark stack or
6293 // work queue overflow
6294 restore_preserved_marks_if_any(); // done single-threaded for now
6295
6296 rp->set_enqueuing_is_done(true);
6297 if (rp->processing_is_mt()) {
6298 rp->balance_all_queues();
6299 CMSRefProcTaskExecutor task_executor(*this);
6300 rp->enqueue_discovered_references(&task_executor);
6301 } else {
6302 rp->enqueue_discovered_references(NULL);
6303 }
6304 rp->verify_no_references_recorded();
6305 assert(!rp->discovery_enabled(), "should have been disabled");
6306 }
6307
6308 #ifndef PRODUCT
6309 void CMSCollector::check_correct_thread_executing() {
6310 Thread* t = Thread::current();
6311 // Only the VM thread or the CMS thread should be here.
6312 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
6313 "Unexpected thread type");
6314 // If this is the vm thread, the foreground process
6315 // should not be waiting. Note that _foregroundGCIsActive is
6316 // true while the foreground collector is waiting.
6317 if (_foregroundGCShouldWait) {
6318 // We cannot be the VM thread
6319 assert(t->is_ConcurrentGC_thread(),
6320 "Should be CMS thread");
6321 } else {
6322 // We can be the CMS thread only if we are in a stop-world
6323 // phase of CMS collection.
6324 if (t->is_ConcurrentGC_thread()) {
6325 assert(_collectorState == InitialMarking ||
6326 _collectorState == FinalMarking,
6327 "Should be a stop-world phase");
6328 // The CMS thread should be holding the CMS_token.
6329 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6330 "Potential interference with concurrently "
6331 "executing VM thread");
6332 }
6333 }
6334 }
6335 #endif
6336
6337 void CMSCollector::sweep(bool asynch) {
6338 assert(_collectorState == Sweeping, "just checking");
6339 check_correct_thread_executing();
6340 verify_work_stacks_empty();
6341 verify_overflow_empty();
6342 increment_sweep_count();
6343 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
6344
6345 _inter_sweep_timer.stop();
6346 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
6347 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
6348
6349 assert(!_intra_sweep_timer.is_active(), "Should not be active");
6350 _intra_sweep_timer.reset();
6351 _intra_sweep_timer.start();
6352 if (asynch) {
6353 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6354 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails);
6355 // First sweep the old gen
6356 {
6357 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6358 bitMapLock());
6359 sweepWork(_cmsGen, asynch);
6360 }
6361
6362 // Update Universe::_heap_*_at_gc figures.
6363 // We need all the free list locks to make the abstract state
6364 // transition from Sweeping to Resetting. See detailed note
6365 // further below.
6366 {
6367 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
6368 // Update heap occupancy information which is used as
6369 // input to soft ref clearing policy at the next gc.
6370 Universe::update_heap_info_at_gc();
6371 _collectorState = Resizing;
6372 }
6373 } else {
6374 // already have needed locks
6375 sweepWork(_cmsGen, asynch);
6376 // Update heap occupancy information which is used as
6377 // input to soft ref clearing policy at the next gc.
6378 Universe::update_heap_info_at_gc();
6379 _collectorState = Resizing;
6380 }
6381 verify_work_stacks_empty();
6382 verify_overflow_empty();
6383
6384 if (should_unload_classes()) {
6385 // Delay purge to the beginning of the next safepoint. Metaspace::contains
6386 // requires that the virtual spaces are stable and not deleted.
6387 ClassLoaderDataGraph::set_should_purge(true);
6388 }
6389
6390 _intra_sweep_timer.stop();
6391 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
6392
6393 _inter_sweep_timer.reset();
6394 _inter_sweep_timer.start();
6395
6396 // We need to use a monotonically non-deccreasing time in ms
6397 // or we will see time-warp warnings and os::javaTimeMillis()
6398 // does not guarantee monotonicity.
6399 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
6400 update_time_of_last_gc(now);
6401
6402 // NOTE on abstract state transitions:
6403 // Mutators allocate-live and/or mark the mod-union table dirty
6404 // based on the state of the collection. The former is done in
6405 // the interval [Marking, Sweeping] and the latter in the interval
6406 // [Marking, Sweeping). Thus the transitions into the Marking state
6407 // and out of the Sweeping state must be synchronously visible
6408 // globally to the mutators.
6409 // The transition into the Marking state happens with the world
6410 // stopped so the mutators will globally see it. Sweeping is
6411 // done asynchronously by the background collector so the transition
6412 // from the Sweeping state to the Resizing state must be done
6413 // under the freelistLock (as is the check for whether to
6414 // allocate-live and whether to dirty the mod-union table).
6415 assert(_collectorState == Resizing, "Change of collector state to"
6416 " Resizing must be done under the freelistLocks (plural)");
6417
6418 // Now that sweeping has been completed, we clear
6419 // the incremental_collection_failed flag,
6420 // thus inviting a younger gen collection to promote into
6421 // this generation. If such a promotion may still fail,
6422 // the flag will be set again when a young collection is
6423 // attempted.
6424 GenCollectedHeap* gch = GenCollectedHeap::heap();
6425 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
6426 gch->update_full_collections_completed(_collection_count_start);
6427 }
6428
6429 // FIX ME!!! Looks like this belongs in CFLSpace, with
6430 // CMSGen merely delegating to it.
6431 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
6432 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
6433 HeapWord* minAddr = _cmsSpace->bottom();
6434 HeapWord* largestAddr =
6435 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
6436 if (largestAddr == NULL) {
6437 // The dictionary appears to be empty. In this case
6438 // try to coalesce at the end of the heap.
6439 largestAddr = _cmsSpace->end();
6440 }
6441 size_t largestOffset = pointer_delta(largestAddr, minAddr);
6442 size_t nearLargestOffset =
6443 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
6444 if (PrintFLSStatistics != 0) {
6445 gclog_or_tty->print_cr(
6446 "CMS: Large Block: " PTR_FORMAT ";"
6447 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
6448 largestAddr,
6449 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset);
6450 }
6451 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
6452 }
6453
6454 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
6455 return addr >= _cmsSpace->nearLargestChunk();
6456 }
6457
6458 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
6459 return _cmsSpace->find_chunk_at_end();
6460 }
6461
6462 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
6463 bool full) {
6464 // The next lower level has been collected. Gather any statistics
6465 // that are of interest at this point.
6466 if (!full && (current_level + 1) == level()) {
6467 // Gather statistics on the young generation collection.
6468 collector()->stats().record_gc0_end(used());
6469 }
6470 }
6471
6472 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
6473 GenCollectedHeap* gch = GenCollectedHeap::heap();
6474 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
6475 "Wrong type of heap");
6476 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
6477 gch->gen_policy()->size_policy();
6478 assert(sp->is_gc_cms_adaptive_size_policy(),
6479 "Wrong type of size policy");
6480 return sp;
6481 }
6482
6483 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
6484 if (PrintGCDetails && Verbose) {
6485 gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
6486 }
6487 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
6488 _debug_collection_type =
6489 (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
6490 if (PrintGCDetails && Verbose) {
6491 gclog_or_tty->print_cr("to %d ", _debug_collection_type);
6492 }
6493 }
6494
6495 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
6496 bool asynch) {
6497 // We iterate over the space(s) underlying this generation,
6498 // checking the mark bit map to see if the bits corresponding
6499 // to specific blocks are marked or not. Blocks that are
6500 // marked are live and are not swept up. All remaining blocks
6501 // are swept up, with coalescing on-the-fly as we sweep up
6502 // contiguous free and/or garbage blocks:
6503 // We need to ensure that the sweeper synchronizes with allocators
6504 // and stop-the-world collectors. In particular, the following
6505 // locks are used:
6506 // . CMS token: if this is held, a stop the world collection cannot occur
6507 // . freelistLock: if this is held no allocation can occur from this
6508 // generation by another thread
6509 // . bitMapLock: if this is held, no other thread can access or update
6510 //
6511
6512 // Note that we need to hold the freelistLock if we use
6513 // block iterate below; else the iterator might go awry if
6514 // a mutator (or promotion) causes block contents to change
6515 // (for instance if the allocator divvies up a block).
6516 // If we hold the free list lock, for all practical purposes
6517 // young generation GC's can't occur (they'll usually need to
6518 // promote), so we might as well prevent all young generation
6519 // GC's while we do a sweeping step. For the same reason, we might
6520 // as well take the bit map lock for the entire duration
6521
6522 // check that we hold the requisite locks
6523 assert(have_cms_token(), "Should hold cms token");
6524 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
6525 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
6526 "Should possess CMS token to sweep");
6527 assert_lock_strong(gen->freelistLock());
6528 assert_lock_strong(bitMapLock());
6529
6530 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
6531 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
6532 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
6533 _inter_sweep_estimate.padded_average(),
6534 _intra_sweep_estimate.padded_average());
6535 gen->setNearLargestChunk();
6536
6537 {
6538 SweepClosure sweepClosure(this, gen, &_markBitMap,
6539 CMSYield && asynch);
6540 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
6541 // We need to free-up/coalesce garbage/blocks from a
6542 // co-terminal free run. This is done in the SweepClosure
6543 // destructor; so, do not remove this scope, else the
6544 // end-of-sweep-census below will be off by a little bit.
6545 }
6546 gen->cmsSpace()->sweep_completed();
6547 gen->cmsSpace()->endSweepFLCensus(sweep_count());
6548 if (should_unload_classes()) { // unloaded classes this cycle,
6549 _concurrent_cycles_since_last_unload = 0; // ... reset count
6550 } else { // did not unload classes,
6551 _concurrent_cycles_since_last_unload++; // ... increment count
6552 }
6553 }
6554
6555 // Reset CMS data structures (for now just the marking bit map)
6556 // preparatory for the next cycle.
6557 void CMSCollector::reset(bool asynch) {
6558 GenCollectedHeap* gch = GenCollectedHeap::heap();
6559 CMSAdaptiveSizePolicy* sp = size_policy();
6560 AdaptiveSizePolicyOutput(sp, gch->total_collections());
6561 if (asynch) {
6562 CMSTokenSyncWithLocks ts(true, bitMapLock());
6563
6564 // If the state is not "Resetting", the foreground thread
6565 // has done a collection and the resetting.
6566 if (_collectorState != Resetting) {
6567 assert(_collectorState == Idling, "The state should only change"
6568 " because the foreground collector has finished the collection");
6569 return;
6570 }
6571
6572 // Clear the mark bitmap (no grey objects to start with)
6573 // for the next cycle.
6574 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6575 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails);
6576
6577 HeapWord* curAddr = _markBitMap.startWord();
6578 while (curAddr < _markBitMap.endWord()) {
6579 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
6580 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
6581 _markBitMap.clear_large_range(chunk);
6582 if (ConcurrentMarkSweepThread::should_yield() &&
6583 !foregroundGCIsActive() &&
6584 CMSYield) {
6585 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6586 "CMS thread should hold CMS token");
6587 assert_lock_strong(bitMapLock());
6588 bitMapLock()->unlock();
6589 ConcurrentMarkSweepThread::desynchronize(true);
6590 ConcurrentMarkSweepThread::acknowledge_yield_request();
6591 stopTimer();
6592 if (PrintCMSStatistics != 0) {
6593 incrementYields();
6594 }
6595 icms_wait();
6596
6597 // See the comment in coordinator_yield()
6598 for (unsigned i = 0; i < CMSYieldSleepCount &&
6599 ConcurrentMarkSweepThread::should_yield() &&
6600 !CMSCollector::foregroundGCIsActive(); ++i) {
6601 os::sleep(Thread::current(), 1, false);
6602 ConcurrentMarkSweepThread::acknowledge_yield_request();
6603 }
6604
6605 ConcurrentMarkSweepThread::synchronize(true);
6606 bitMapLock()->lock_without_safepoint_check();
6607 startTimer();
6608 }
6609 curAddr = chunk.end();
6610 }
6611 // A successful mostly concurrent collection has been done.
6612 // Because only the full (i.e., concurrent mode failure) collections
6613 // are being measured for gc overhead limits, clean the "near" flag
6614 // and count.
6615 sp->reset_gc_overhead_limit_count();
6616 _collectorState = Idling;
6617 } else {
6618 // already have the lock
6619 assert(_collectorState == Resetting, "just checking");
6620 assert_lock_strong(bitMapLock());
6621 _markBitMap.clear_all();
6622 _collectorState = Idling;
6623 }
6624
6625 // Stop incremental mode after a cycle completes, so that any future cycles
6626 // are triggered by allocation.
6627 stop_icms();
6628
6629 NOT_PRODUCT(
6630 if (RotateCMSCollectionTypes) {
6631 _cmsGen->rotate_debug_collection_type();
6632 }
6633 )
6634
6635 register_gc_end();
6636 }
6637
6638 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
6639 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6640 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id());
6641 TraceCollectorStats tcs(counters());
6642
6643 switch (op) {
6644 case CMS_op_checkpointRootsInitial: {
6645 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6646 checkpointRootsInitial(true); // asynch
6647 if (PrintGC) {
6648 _cmsGen->printOccupancy("initial-mark");
6649 }
6650 break;
6651 }
6652 case CMS_op_checkpointRootsFinal: {
6653 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6654 checkpointRootsFinal(true, // asynch
6655 false, // !clear_all_soft_refs
6656 false); // !init_mark_was_synchronous
6657 if (PrintGC) {
6658 _cmsGen->printOccupancy("remark");
6659 }
6660 break;
6661 }
6662 default:
6663 fatal("No such CMS_op");
6664 }
6665 }
6666
6667 #ifndef PRODUCT
6668 size_t const CMSCollector::skip_header_HeapWords() {
6669 return FreeChunk::header_size();
6670 }
6671
6672 // Try and collect here conditions that should hold when
6673 // CMS thread is exiting. The idea is that the foreground GC
6674 // thread should not be blocked if it wants to terminate
6675 // the CMS thread and yet continue to run the VM for a while
6676 // after that.
6677 void CMSCollector::verify_ok_to_terminate() const {
6678 assert(Thread::current()->is_ConcurrentGC_thread(),
6679 "should be called by CMS thread");
6680 assert(!_foregroundGCShouldWait, "should be false");
6681 // We could check here that all the various low-level locks
6682 // are not held by the CMS thread, but that is overkill; see
6683 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6684 // is checked.
6685 }
6686 #endif
6687
6688 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6689 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6690 "missing Printezis mark?");
6691 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6692 size_t size = pointer_delta(nextOneAddr + 1, addr);
6693 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6694 "alignment problem");
6695 assert(size >= 3, "Necessary for Printezis marks to work");
6696 return size;
6697 }
6698
6699 // A variant of the above (block_size_using_printezis_bits()) except
6700 // that we return 0 if the P-bits are not yet set.
6701 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6702 if (_markBitMap.isMarked(addr + 1)) {
6703 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
6704 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6705 size_t size = pointer_delta(nextOneAddr + 1, addr);
6706 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6707 "alignment problem");
6708 assert(size >= 3, "Necessary for Printezis marks to work");
6709 return size;
6710 }
6711 return 0;
6712 }
6713
6714 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6715 size_t sz = 0;
6716 oop p = (oop)addr;
6717 if (p->klass_or_null() != NULL) {
6718 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6719 } else {
6720 sz = block_size_using_printezis_bits(addr);
6721 }
6722 assert(sz > 0, "size must be nonzero");
6723 HeapWord* next_block = addr + sz;
6724 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
6725 CardTableModRefBS::card_size);
6726 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
6727 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6728 "must be different cards");
6729 return next_card;
6730 }
6731
6732
6733 // CMS Bit Map Wrapper /////////////////////////////////////////
6734
6735 // Construct a CMS bit map infrastructure, but don't create the
6736 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6737 // further below.
6738 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6739 _bm(),
6740 _shifter(shifter),
6741 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6742 {
6743 _bmStartWord = 0;
6744 _bmWordSize = 0;
6745 }
6746
6747 bool CMSBitMap::allocate(MemRegion mr) {
6748 _bmStartWord = mr.start();
6749 _bmWordSize = mr.word_size();
6750 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6751 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6752 if (!brs.is_reserved()) {
6753 warning("CMS bit map allocation failure");
6754 return false;
6755 }
6756 // For now we'll just commit all of the bit map up fromt.
6757 // Later on we'll try to be more parsimonious with swap.
6758 if (!_virtual_space.initialize(brs, brs.size())) {
6759 warning("CMS bit map backing store failure");
6760 return false;
6761 }
6762 assert(_virtual_space.committed_size() == brs.size(),
6763 "didn't reserve backing store for all of CMS bit map?");
6764 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
6765 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6766 _bmWordSize, "inconsistency in bit map sizing");
6767 _bm.set_size(_bmWordSize >> _shifter);
6768
6769 // bm.clear(); // can we rely on getting zero'd memory? verify below
6770 assert(isAllClear(),
6771 "Expected zero'd memory from ReservedSpace constructor");
6772 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6773 "consistency check");
6774 return true;
6775 }
6776
6777 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6778 HeapWord *next_addr, *end_addr, *last_addr;
6779 assert_locked();
6780 assert(covers(mr), "out-of-range error");
6781 // XXX assert that start and end are appropriately aligned
6782 for (next_addr = mr.start(), end_addr = mr.end();
6783 next_addr < end_addr; next_addr = last_addr) {
6784 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6785 last_addr = dirty_region.end();
6786 if (!dirty_region.is_empty()) {
6787 cl->do_MemRegion(dirty_region);
6788 } else {
6789 assert(last_addr == end_addr, "program logic");
6790 return;
6791 }
6792 }
6793 }
6794
6795 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
6796 _bm.print_on_error(st, prefix);
6797 }
6798
6799 #ifndef PRODUCT
6800 void CMSBitMap::assert_locked() const {
6801 CMSLockVerifier::assert_locked(lock());
6802 }
6803
6804 bool CMSBitMap::covers(MemRegion mr) const {
6805 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6806 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6807 "size inconsistency");
6808 return (mr.start() >= _bmStartWord) &&
6809 (mr.end() <= endWord());
6810 }
6811
6812 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6813 return (start >= _bmStartWord && (start + size) <= endWord());
6814 }
6815
6816 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6817 // verify that there are no 1 bits in the interval [left, right)
6818 FalseBitMapClosure falseBitMapClosure;
6819 iterate(&falseBitMapClosure, left, right);
6820 }
6821
6822 void CMSBitMap::region_invariant(MemRegion mr)
6823 {
6824 assert_locked();
6825 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6826 assert(!mr.is_empty(), "unexpected empty region");
6827 assert(covers(mr), "mr should be covered by bit map");
6828 // convert address range into offset range
6829 size_t start_ofs = heapWordToOffset(mr.start());
6830 // Make sure that end() is appropriately aligned
6831 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6832 ((intptr_t) 1 << (_shifter+LogHeapWordSize))),
6833 "Misaligned mr.end()");
6834 size_t end_ofs = heapWordToOffset(mr.end());
6835 assert(end_ofs > start_ofs, "Should mark at least one bit");
6836 }
6837
6838 #endif
6839
6840 bool CMSMarkStack::allocate(size_t size) {
6841 // allocate a stack of the requisite depth
6842 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6843 size * sizeof(oop)));
6844 if (!rs.is_reserved()) {
6845 warning("CMSMarkStack allocation failure");
6846 return false;
6847 }
6848 if (!_virtual_space.initialize(rs, rs.size())) {
6849 warning("CMSMarkStack backing store failure");
6850 return false;
6851 }
6852 assert(_virtual_space.committed_size() == rs.size(),
6853 "didn't reserve backing store for all of CMS stack?");
6854 _base = (oop*)(_virtual_space.low());
6855 _index = 0;
6856 _capacity = size;
6857 NOT_PRODUCT(_max_depth = 0);
6858 return true;
6859 }
6860
6861 // XXX FIX ME !!! In the MT case we come in here holding a
6862 // leaf lock. For printing we need to take a further lock
6863 // which has lower rank. We need to recallibrate the two
6864 // lock-ranks involved in order to be able to rpint the
6865 // messages below. (Or defer the printing to the caller.
6866 // For now we take the expedient path of just disabling the
6867 // messages for the problematic case.)
6868 void CMSMarkStack::expand() {
6869 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6870 if (_capacity == MarkStackSizeMax) {
6871 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6872 // We print a warning message only once per CMS cycle.
6873 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6874 }
6875 return;
6876 }
6877 // Double capacity if possible
6878 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6879 // Do not give up existing stack until we have managed to
6880 // get the double capacity that we desired.
6881 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6882 new_capacity * sizeof(oop)));
6883 if (rs.is_reserved()) {
6884 // Release the backing store associated with old stack
6885 _virtual_space.release();
6886 // Reinitialize virtual space for new stack
6887 if (!_virtual_space.initialize(rs, rs.size())) {
6888 fatal("Not enough swap for expanded marking stack");
6889 }
6890 _base = (oop*)(_virtual_space.low());
6891 _index = 0;
6892 _capacity = new_capacity;
6893 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6894 // Failed to double capacity, continue;
6895 // we print a detail message only once per CMS cycle.
6896 gclog_or_tty->print(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to "
6897 SIZE_FORMAT "K",
6898 _capacity / K, new_capacity / K);
6899 }
6900 }
6901
6902
6903 // Closures
6904 // XXX: there seems to be a lot of code duplication here;
6905 // should refactor and consolidate common code.
6906
6907 // This closure is used to mark refs into the CMS generation in
6908 // the CMS bit map. Called at the first checkpoint. This closure
6909 // assumes that we do not need to re-mark dirty cards; if the CMS
6910 // generation on which this is used is not an oldest
6911 // generation then this will lose younger_gen cards!
6912
6913 MarkRefsIntoClosure::MarkRefsIntoClosure(
6914 MemRegion span, CMSBitMap* bitMap):
6915 _span(span),
6916 _bitMap(bitMap)
6917 {
6918 assert(_ref_processor == NULL, "deliberately left NULL");
6919 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6920 }
6921
6922 void MarkRefsIntoClosure::do_oop(oop obj) {
6923 // if p points into _span, then mark corresponding bit in _markBitMap
6924 assert(obj->is_oop(), "expected an oop");
6925 HeapWord* addr = (HeapWord*)obj;
6926 if (_span.contains(addr)) {
6927 // this should be made more efficient
6928 _bitMap->mark(addr);
6929 }
6930 }
6931
6932 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6933 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6934
6935 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
6936 MemRegion span, CMSBitMap* bitMap):
6937 _span(span),
6938 _bitMap(bitMap)
6939 {
6940 assert(_ref_processor == NULL, "deliberately left NULL");
6941 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6942 }
6943
6944 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
6945 // if p points into _span, then mark corresponding bit in _markBitMap
6946 assert(obj->is_oop(), "expected an oop");
6947 HeapWord* addr = (HeapWord*)obj;
6948 if (_span.contains(addr)) {
6949 // this should be made more efficient
6950 _bitMap->par_mark(addr);
6951 }
6952 }
6953
6954 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6955 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6956
6957 // A variant of the above, used for CMS marking verification.
6958 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6959 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6960 _span(span),
6961 _verification_bm(verification_bm),
6962 _cms_bm(cms_bm)
6963 {
6964 assert(_ref_processor == NULL, "deliberately left NULL");
6965 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6966 }
6967
6968 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6969 // if p points into _span, then mark corresponding bit in _markBitMap
6970 assert(obj->is_oop(), "expected an oop");
6971 HeapWord* addr = (HeapWord*)obj;
6972 if (_span.contains(addr)) {
6973 _verification_bm->mark(addr);
6974 if (!_cms_bm->isMarked(addr)) {
6975 oop(addr)->print();
6976 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6977 fatal("... aborting");
6978 }
6979 }
6980 }
6981
6982 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6983 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6984
6985 //////////////////////////////////////////////////
6986 // MarkRefsIntoAndScanClosure
6987 //////////////////////////////////////////////////
6988
6989 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6990 ReferenceProcessor* rp,
6991 CMSBitMap* bit_map,
6992 CMSBitMap* mod_union_table,
6993 CMSMarkStack* mark_stack,
6994 CMSCollector* collector,
6995 bool should_yield,
6996 bool concurrent_precleaning):
6997 _collector(collector),
6998 _span(span),
6999 _bit_map(bit_map),
7000 _mark_stack(mark_stack),
7001 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
7002 mark_stack, concurrent_precleaning),
7003 _yield(should_yield),
7004 _concurrent_precleaning(concurrent_precleaning),
7005 _freelistLock(NULL)
7006 {
7007 _ref_processor = rp;
7008 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7009 }
7010
7011 // This closure is used to mark refs into the CMS generation at the
7012 // second (final) checkpoint, and to scan and transitively follow
7013 // the unmarked oops. It is also used during the concurrent precleaning
7014 // phase while scanning objects on dirty cards in the CMS generation.
7015 // The marks are made in the marking bit map and the marking stack is
7016 // used for keeping the (newly) grey objects during the scan.
7017 // The parallel version (Par_...) appears further below.
7018 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
7019 if (obj != NULL) {
7020 assert(obj->is_oop(), "expected an oop");
7021 HeapWord* addr = (HeapWord*)obj;
7022 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
7023 assert(_collector->overflow_list_is_empty(),
7024 "overflow list should be empty");
7025 if (_span.contains(addr) &&
7026 !_bit_map->isMarked(addr)) {
7027 // mark bit map (object is now grey)
7028 _bit_map->mark(addr);
7029 // push on marking stack (stack should be empty), and drain the
7030 // stack by applying this closure to the oops in the oops popped
7031 // from the stack (i.e. blacken the grey objects)
7032 bool res = _mark_stack->push(obj);
7033 assert(res, "Should have space to push on empty stack");
7034 do {
7035 oop new_oop = _mark_stack->pop();
7036 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
7037 assert(_bit_map->isMarked((HeapWord*)new_oop),
7038 "only grey objects on this stack");
7039 // iterate over the oops in this oop, marking and pushing
7040 // the ones in CMS heap (i.e. in _span).
7041 new_oop->oop_iterate(&_pushAndMarkClosure);
7042 // check if it's time to yield
7043 do_yield_check();
7044 } while (!_mark_stack->isEmpty() ||
7045 (!_concurrent_precleaning && take_from_overflow_list()));
7046 // if marking stack is empty, and we are not doing this
7047 // during precleaning, then check the overflow list
7048 }
7049 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
7050 assert(_collector->overflow_list_is_empty(),
7051 "overflow list was drained above");
7052 // We could restore evacuated mark words, if any, used for
7053 // overflow list links here because the overflow list is
7054 // provably empty here. That would reduce the maximum
7055 // size requirements for preserved_{oop,mark}_stack.
7056 // But we'll just postpone it until we are all done
7057 // so we can just stream through.
7058 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
7059 _collector->restore_preserved_marks_if_any();
7060 assert(_collector->no_preserved_marks(), "No preserved marks");
7061 }
7062 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
7063 "All preserved marks should have been restored above");
7064 }
7065 }
7066
7067 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
7068 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
7069
7070 void MarkRefsIntoAndScanClosure::do_yield_work() {
7071 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7072 "CMS thread should hold CMS token");
7073 assert_lock_strong(_freelistLock);
7074 assert_lock_strong(_bit_map->lock());
7075 // relinquish the free_list_lock and bitMaplock()
7076 _bit_map->lock()->unlock();
7077 _freelistLock->unlock();
7078 ConcurrentMarkSweepThread::desynchronize(true);
7079 ConcurrentMarkSweepThread::acknowledge_yield_request();
7080 _collector->stopTimer();
7081 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7082 if (PrintCMSStatistics != 0) {
7083 _collector->incrementYields();
7084 }
7085 _collector->icms_wait();
7086
7087 // See the comment in coordinator_yield()
7088 for (unsigned i = 0;
7089 i < CMSYieldSleepCount &&
7090 ConcurrentMarkSweepThread::should_yield() &&
7091 !CMSCollector::foregroundGCIsActive();
7092 ++i) {
7093 os::sleep(Thread::current(), 1, false);
7094 ConcurrentMarkSweepThread::acknowledge_yield_request();
7095 }
7096
7097 ConcurrentMarkSweepThread::synchronize(true);
7098 _freelistLock->lock_without_safepoint_check();
7099 _bit_map->lock()->lock_without_safepoint_check();
7100 _collector->startTimer();
7101 }
7102
7103 ///////////////////////////////////////////////////////////
7104 // Par_MarkRefsIntoAndScanClosure: a parallel version of
7105 // MarkRefsIntoAndScanClosure
7106 ///////////////////////////////////////////////////////////
7107 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
7108 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
7109 CMSBitMap* bit_map, OopTaskQueue* work_queue):
7110 _span(span),
7111 _bit_map(bit_map),
7112 _work_queue(work_queue),
7113 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
7114 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
7115 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
7116 {
7117 _ref_processor = rp;
7118 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7119 }
7120
7121 // This closure is used to mark refs into the CMS generation at the
7122 // second (final) checkpoint, and to scan and transitively follow
7123 // the unmarked oops. The marks are made in the marking bit map and
7124 // the work_queue is used for keeping the (newly) grey objects during
7125 // the scan phase whence they are also available for stealing by parallel
7126 // threads. Since the marking bit map is shared, updates are
7127 // synchronized (via CAS).
7128 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
7129 if (obj != NULL) {
7130 // Ignore mark word because this could be an already marked oop
7131 // that may be chained at the end of the overflow list.
7132 assert(obj->is_oop(true), "expected an oop");
7133 HeapWord* addr = (HeapWord*)obj;
7134 if (_span.contains(addr) &&
7135 !_bit_map->isMarked(addr)) {
7136 // mark bit map (object will become grey):
7137 // It is possible for several threads to be
7138 // trying to "claim" this object concurrently;
7139 // the unique thread that succeeds in marking the
7140 // object first will do the subsequent push on
7141 // to the work queue (or overflow list).
7142 if (_bit_map->par_mark(addr)) {
7143 // push on work_queue (which may not be empty), and trim the
7144 // queue to an appropriate length by applying this closure to
7145 // the oops in the oops popped from the stack (i.e. blacken the
7146 // grey objects)
7147 bool res = _work_queue->push(obj);
7148 assert(res, "Low water mark should be less than capacity?");
7149 trim_queue(_low_water_mark);
7150 } // Else, another thread claimed the object
7151 }
7152 }
7153 }
7154
7155 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
7156 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
7157
7158 // This closure is used to rescan the marked objects on the dirty cards
7159 // in the mod union table and the card table proper.
7160 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
7161 oop p, MemRegion mr) {
7162
7163 size_t size = 0;
7164 HeapWord* addr = (HeapWord*)p;
7165 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7166 assert(_span.contains(addr), "we are scanning the CMS generation");
7167 // check if it's time to yield
7168 if (do_yield_check()) {
7169 // We yielded for some foreground stop-world work,
7170 // and we have been asked to abort this ongoing preclean cycle.
7171 return 0;
7172 }
7173 if (_bitMap->isMarked(addr)) {
7174 // it's marked; is it potentially uninitialized?
7175 if (p->klass_or_null() != NULL) {
7176 // an initialized object; ignore mark word in verification below
7177 // since we are running concurrent with mutators
7178 assert(p->is_oop(true), "should be an oop");
7179 if (p->is_objArray()) {
7180 // objArrays are precisely marked; restrict scanning
7181 // to dirty cards only.
7182 size = CompactibleFreeListSpace::adjustObjectSize(
7183 p->oop_iterate(_scanningClosure, mr));
7184 } else {
7185 // A non-array may have been imprecisely marked; we need
7186 // to scan object in its entirety.
7187 size = CompactibleFreeListSpace::adjustObjectSize(
7188 p->oop_iterate(_scanningClosure));
7189 }
7190 #ifdef ASSERT
7191 size_t direct_size =
7192 CompactibleFreeListSpace::adjustObjectSize(p->size());
7193 assert(size == direct_size, "Inconsistency in size");
7194 assert(size >= 3, "Necessary for Printezis marks to work");
7195 if (!_bitMap->isMarked(addr+1)) {
7196 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
7197 } else {
7198 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
7199 assert(_bitMap->isMarked(addr+size-1),
7200 "inconsistent Printezis mark");
7201 }
7202 #endif // ASSERT
7203 } else {
7204 // an unitialized object
7205 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
7206 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7207 size = pointer_delta(nextOneAddr + 1, addr);
7208 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7209 "alignment problem");
7210 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
7211 // will dirty the card when the klass pointer is installed in the
7212 // object (signalling the completion of initialization).
7213 }
7214 } else {
7215 // Either a not yet marked object or an uninitialized object
7216 if (p->klass_or_null() == NULL) {
7217 // An uninitialized object, skip to the next card, since
7218 // we may not be able to read its P-bits yet.
7219 assert(size == 0, "Initial value");
7220 } else {
7221 // An object not (yet) reached by marking: we merely need to
7222 // compute its size so as to go look at the next block.
7223 assert(p->is_oop(true), "should be an oop");
7224 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
7225 }
7226 }
7227 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7228 return size;
7229 }
7230
7231 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
7232 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7233 "CMS thread should hold CMS token");
7234 assert_lock_strong(_freelistLock);
7235 assert_lock_strong(_bitMap->lock());
7236 // relinquish the free_list_lock and bitMaplock()
7237 _bitMap->lock()->unlock();
7238 _freelistLock->unlock();
7239 ConcurrentMarkSweepThread::desynchronize(true);
7240 ConcurrentMarkSweepThread::acknowledge_yield_request();
7241 _collector->stopTimer();
7242 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7243 if (PrintCMSStatistics != 0) {
7244 _collector->incrementYields();
7245 }
7246 _collector->icms_wait();
7247
7248 // See the comment in coordinator_yield()
7249 for (unsigned i = 0; i < CMSYieldSleepCount &&
7250 ConcurrentMarkSweepThread::should_yield() &&
7251 !CMSCollector::foregroundGCIsActive(); ++i) {
7252 os::sleep(Thread::current(), 1, false);
7253 ConcurrentMarkSweepThread::acknowledge_yield_request();
7254 }
7255
7256 ConcurrentMarkSweepThread::synchronize(true);
7257 _freelistLock->lock_without_safepoint_check();
7258 _bitMap->lock()->lock_without_safepoint_check();
7259 _collector->startTimer();
7260 }
7261
7262
7263 //////////////////////////////////////////////////////////////////
7264 // SurvivorSpacePrecleanClosure
7265 //////////////////////////////////////////////////////////////////
7266 // This (single-threaded) closure is used to preclean the oops in
7267 // the survivor spaces.
7268 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
7269
7270 HeapWord* addr = (HeapWord*)p;
7271 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7272 assert(!_span.contains(addr), "we are scanning the survivor spaces");
7273 assert(p->klass_or_null() != NULL, "object should be initializd");
7274 // an initialized object; ignore mark word in verification below
7275 // since we are running concurrent with mutators
7276 assert(p->is_oop(true), "should be an oop");
7277 // Note that we do not yield while we iterate over
7278 // the interior oops of p, pushing the relevant ones
7279 // on our marking stack.
7280 size_t size = p->oop_iterate(_scanning_closure);
7281 do_yield_check();
7282 // Observe that below, we do not abandon the preclean
7283 // phase as soon as we should; rather we empty the
7284 // marking stack before returning. This is to satisfy
7285 // some existing assertions. In general, it may be a
7286 // good idea to abort immediately and complete the marking
7287 // from the grey objects at a later time.
7288 while (!_mark_stack->isEmpty()) {
7289 oop new_oop = _mark_stack->pop();
7290 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
7291 assert(_bit_map->isMarked((HeapWord*)new_oop),
7292 "only grey objects on this stack");
7293 // iterate over the oops in this oop, marking and pushing
7294 // the ones in CMS heap (i.e. in _span).
7295 new_oop->oop_iterate(_scanning_closure);
7296 // check if it's time to yield
7297 do_yield_check();
7298 }
7299 unsigned int after_count =
7300 GenCollectedHeap::heap()->total_collections();
7301 bool abort = (_before_count != after_count) ||
7302 _collector->should_abort_preclean();
7303 return abort ? 0 : size;
7304 }
7305
7306 void SurvivorSpacePrecleanClosure::do_yield_work() {
7307 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7308 "CMS thread should hold CMS token");
7309 assert_lock_strong(_bit_map->lock());
7310 // Relinquish the bit map lock
7311 _bit_map->lock()->unlock();
7312 ConcurrentMarkSweepThread::desynchronize(true);
7313 ConcurrentMarkSweepThread::acknowledge_yield_request();
7314 _collector->stopTimer();
7315 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7316 if (PrintCMSStatistics != 0) {
7317 _collector->incrementYields();
7318 }
7319 _collector->icms_wait();
7320
7321 // See the comment in coordinator_yield()
7322 for (unsigned i = 0; i < CMSYieldSleepCount &&
7323 ConcurrentMarkSweepThread::should_yield() &&
7324 !CMSCollector::foregroundGCIsActive(); ++i) {
7325 os::sleep(Thread::current(), 1, false);
7326 ConcurrentMarkSweepThread::acknowledge_yield_request();
7327 }
7328
7329 ConcurrentMarkSweepThread::synchronize(true);
7330 _bit_map->lock()->lock_without_safepoint_check();
7331 _collector->startTimer();
7332 }
7333
7334 // This closure is used to rescan the marked objects on the dirty cards
7335 // in the mod union table and the card table proper. In the parallel
7336 // case, although the bitMap is shared, we do a single read so the
7337 // isMarked() query is "safe".
7338 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
7339 // Ignore mark word because we are running concurrent with mutators
7340 assert(p->is_oop_or_null(true), "expected an oop or null");
7341 HeapWord* addr = (HeapWord*)p;
7342 assert(_span.contains(addr), "we are scanning the CMS generation");
7343 bool is_obj_array = false;
7344 #ifdef ASSERT
7345 if (!_parallel) {
7346 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
7347 assert(_collector->overflow_list_is_empty(),
7348 "overflow list should be empty");
7349
7350 }
7351 #endif // ASSERT
7352 if (_bit_map->isMarked(addr)) {
7353 // Obj arrays are precisely marked, non-arrays are not;
7354 // so we scan objArrays precisely and non-arrays in their
7355 // entirety.
7356 if (p->is_objArray()) {
7357 is_obj_array = true;
7358 if (_parallel) {
7359 p->oop_iterate(_par_scan_closure, mr);
7360 } else {
7361 p->oop_iterate(_scan_closure, mr);
7362 }
7363 } else {
7364 if (_parallel) {
7365 p->oop_iterate(_par_scan_closure);
7366 } else {
7367 p->oop_iterate(_scan_closure);
7368 }
7369 }
7370 }
7371 #ifdef ASSERT
7372 if (!_parallel) {
7373 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
7374 assert(_collector->overflow_list_is_empty(),
7375 "overflow list should be empty");
7376
7377 }
7378 #endif // ASSERT
7379 return is_obj_array;
7380 }
7381
7382 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
7383 MemRegion span,
7384 CMSBitMap* bitMap, CMSMarkStack* markStack,
7385 bool should_yield, bool verifying):
7386 _collector(collector),
7387 _span(span),
7388 _bitMap(bitMap),
7389 _mut(&collector->_modUnionTable),
7390 _markStack(markStack),
7391 _yield(should_yield),
7392 _skipBits(0)
7393 {
7394 assert(_markStack->isEmpty(), "stack should be empty");
7395 _finger = _bitMap->startWord();
7396 _threshold = _finger;
7397 assert(_collector->_restart_addr == NULL, "Sanity check");
7398 assert(_span.contains(_finger), "Out of bounds _finger?");
7399 DEBUG_ONLY(_verifying = verifying;)
7400 }
7401
7402 void MarkFromRootsClosure::reset(HeapWord* addr) {
7403 assert(_markStack->isEmpty(), "would cause duplicates on stack");
7404 assert(_span.contains(addr), "Out of bounds _finger?");
7405 _finger = addr;
7406 _threshold = (HeapWord*)round_to(
7407 (intptr_t)_finger, CardTableModRefBS::card_size);
7408 }
7409
7410 // Should revisit to see if this should be restructured for
7411 // greater efficiency.
7412 bool MarkFromRootsClosure::do_bit(size_t offset) {
7413 if (_skipBits > 0) {
7414 _skipBits--;
7415 return true;
7416 }
7417 // convert offset into a HeapWord*
7418 HeapWord* addr = _bitMap->startWord() + offset;
7419 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
7420 "address out of range");
7421 assert(_bitMap->isMarked(addr), "tautology");
7422 if (_bitMap->isMarked(addr+1)) {
7423 // this is an allocated but not yet initialized object
7424 assert(_skipBits == 0, "tautology");
7425 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
7426 oop p = oop(addr);
7427 if (p->klass_or_null() == NULL) {
7428 DEBUG_ONLY(if (!_verifying) {)
7429 // We re-dirty the cards on which this object lies and increase
7430 // the _threshold so that we'll come back to scan this object
7431 // during the preclean or remark phase. (CMSCleanOnEnter)
7432 if (CMSCleanOnEnter) {
7433 size_t sz = _collector->block_size_using_printezis_bits(addr);
7434 HeapWord* end_card_addr = (HeapWord*)round_to(
7435 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7436 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7437 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7438 // Bump _threshold to end_card_addr; note that
7439 // _threshold cannot possibly exceed end_card_addr, anyhow.
7440 // This prevents future clearing of the card as the scan proceeds
7441 // to the right.
7442 assert(_threshold <= end_card_addr,
7443 "Because we are just scanning into this object");
7444 if (_threshold < end_card_addr) {
7445 _threshold = end_card_addr;
7446 }
7447 if (p->klass_or_null() != NULL) {
7448 // Redirty the range of cards...
7449 _mut->mark_range(redirty_range);
7450 } // ...else the setting of klass will dirty the card anyway.
7451 }
7452 DEBUG_ONLY(})
7453 return true;
7454 }
7455 }
7456 scanOopsInOop(addr);
7457 return true;
7458 }
7459
7460 // We take a break if we've been at this for a while,
7461 // so as to avoid monopolizing the locks involved.
7462 void MarkFromRootsClosure::do_yield_work() {
7463 // First give up the locks, then yield, then re-lock
7464 // We should probably use a constructor/destructor idiom to
7465 // do this unlock/lock or modify the MutexUnlocker class to
7466 // serve our purpose. XXX
7467 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7468 "CMS thread should hold CMS token");
7469 assert_lock_strong(_bitMap->lock());
7470 _bitMap->lock()->unlock();
7471 ConcurrentMarkSweepThread::desynchronize(true);
7472 ConcurrentMarkSweepThread::acknowledge_yield_request();
7473 _collector->stopTimer();
7474 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7475 if (PrintCMSStatistics != 0) {
7476 _collector->incrementYields();
7477 }
7478 _collector->icms_wait();
7479
7480 // See the comment in coordinator_yield()
7481 for (unsigned i = 0; i < CMSYieldSleepCount &&
7482 ConcurrentMarkSweepThread::should_yield() &&
7483 !CMSCollector::foregroundGCIsActive(); ++i) {
7484 os::sleep(Thread::current(), 1, false);
7485 ConcurrentMarkSweepThread::acknowledge_yield_request();
7486 }
7487
7488 ConcurrentMarkSweepThread::synchronize(true);
7489 _bitMap->lock()->lock_without_safepoint_check();
7490 _collector->startTimer();
7491 }
7492
7493 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
7494 assert(_bitMap->isMarked(ptr), "expected bit to be set");
7495 assert(_markStack->isEmpty(),
7496 "should drain stack to limit stack usage");
7497 // convert ptr to an oop preparatory to scanning
7498 oop obj = oop(ptr);
7499 // Ignore mark word in verification below, since we
7500 // may be running concurrent with mutators.
7501 assert(obj->is_oop(true), "should be an oop");
7502 assert(_finger <= ptr, "_finger runneth ahead");
7503 // advance the finger to right end of this object
7504 _finger = ptr + obj->size();
7505 assert(_finger > ptr, "we just incremented it above");
7506 // On large heaps, it may take us some time to get through
7507 // the marking phase (especially if running iCMS). During
7508 // this time it's possible that a lot of mutations have
7509 // accumulated in the card table and the mod union table --
7510 // these mutation records are redundant until we have
7511 // actually traced into the corresponding card.
7512 // Here, we check whether advancing the finger would make
7513 // us cross into a new card, and if so clear corresponding
7514 // cards in the MUT (preclean them in the card-table in the
7515 // future).
7516
7517 DEBUG_ONLY(if (!_verifying) {)
7518 // The clean-on-enter optimization is disabled by default,
7519 // until we fix 6178663.
7520 if (CMSCleanOnEnter && (_finger > _threshold)) {
7521 // [_threshold, _finger) represents the interval
7522 // of cards to be cleared in MUT (or precleaned in card table).
7523 // The set of cards to be cleared is all those that overlap
7524 // with the interval [_threshold, _finger); note that
7525 // _threshold is always kept card-aligned but _finger isn't
7526 // always card-aligned.
7527 HeapWord* old_threshold = _threshold;
7528 assert(old_threshold == (HeapWord*)round_to(
7529 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7530 "_threshold should always be card-aligned");
7531 _threshold = (HeapWord*)round_to(
7532 (intptr_t)_finger, CardTableModRefBS::card_size);
7533 MemRegion mr(old_threshold, _threshold);
7534 assert(!mr.is_empty(), "Control point invariant");
7535 assert(_span.contains(mr), "Should clear within span");
7536 _mut->clear_range(mr);
7537 }
7538 DEBUG_ONLY(})
7539 // Note: the finger doesn't advance while we drain
7540 // the stack below.
7541 PushOrMarkClosure pushOrMarkClosure(_collector,
7542 _span, _bitMap, _markStack,
7543 _finger, this);
7544 bool res = _markStack->push(obj);
7545 assert(res, "Empty non-zero size stack should have space for single push");
7546 while (!_markStack->isEmpty()) {
7547 oop new_oop = _markStack->pop();
7548 // Skip verifying header mark word below because we are
7549 // running concurrent with mutators.
7550 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7551 // now scan this oop's oops
7552 new_oop->oop_iterate(&pushOrMarkClosure);
7553 do_yield_check();
7554 }
7555 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
7556 }
7557
7558 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
7559 CMSCollector* collector, MemRegion span,
7560 CMSBitMap* bit_map,
7561 OopTaskQueue* work_queue,
7562 CMSMarkStack* overflow_stack,
7563 bool should_yield):
7564 _collector(collector),
7565 _whole_span(collector->_span),
7566 _span(span),
7567 _bit_map(bit_map),
7568 _mut(&collector->_modUnionTable),
7569 _work_queue(work_queue),
7570 _overflow_stack(overflow_stack),
7571 _yield(should_yield),
7572 _skip_bits(0),
7573 _task(task)
7574 {
7575 assert(_work_queue->size() == 0, "work_queue should be empty");
7576 _finger = span.start();
7577 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
7578 assert(_span.contains(_finger), "Out of bounds _finger?");
7579 }
7580
7581 // Should revisit to see if this should be restructured for
7582 // greater efficiency.
7583 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
7584 if (_skip_bits > 0) {
7585 _skip_bits--;
7586 return true;
7587 }
7588 // convert offset into a HeapWord*
7589 HeapWord* addr = _bit_map->startWord() + offset;
7590 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
7591 "address out of range");
7592 assert(_bit_map->isMarked(addr), "tautology");
7593 if (_bit_map->isMarked(addr+1)) {
7594 // this is an allocated object that might not yet be initialized
7595 assert(_skip_bits == 0, "tautology");
7596 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
7597 oop p = oop(addr);
7598 if (p->klass_or_null() == NULL) {
7599 // in the case of Clean-on-Enter optimization, redirty card
7600 // and avoid clearing card by increasing the threshold.
7601 return true;
7602 }
7603 }
7604 scan_oops_in_oop(addr);
7605 return true;
7606 }
7607
7608 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
7609 assert(_bit_map->isMarked(ptr), "expected bit to be set");
7610 // Should we assert that our work queue is empty or
7611 // below some drain limit?
7612 assert(_work_queue->size() == 0,
7613 "should drain stack to limit stack usage");
7614 // convert ptr to an oop preparatory to scanning
7615 oop obj = oop(ptr);
7616 // Ignore mark word in verification below, since we
7617 // may be running concurrent with mutators.
7618 assert(obj->is_oop(true), "should be an oop");
7619 assert(_finger <= ptr, "_finger runneth ahead");
7620 // advance the finger to right end of this object
7621 _finger = ptr + obj->size();
7622 assert(_finger > ptr, "we just incremented it above");
7623 // On large heaps, it may take us some time to get through
7624 // the marking phase (especially if running iCMS). During
7625 // this time it's possible that a lot of mutations have
7626 // accumulated in the card table and the mod union table --
7627 // these mutation records are redundant until we have
7628 // actually traced into the corresponding card.
7629 // Here, we check whether advancing the finger would make
7630 // us cross into a new card, and if so clear corresponding
7631 // cards in the MUT (preclean them in the card-table in the
7632 // future).
7633
7634 // The clean-on-enter optimization is disabled by default,
7635 // until we fix 6178663.
7636 if (CMSCleanOnEnter && (_finger > _threshold)) {
7637 // [_threshold, _finger) represents the interval
7638 // of cards to be cleared in MUT (or precleaned in card table).
7639 // The set of cards to be cleared is all those that overlap
7640 // with the interval [_threshold, _finger); note that
7641 // _threshold is always kept card-aligned but _finger isn't
7642 // always card-aligned.
7643 HeapWord* old_threshold = _threshold;
7644 assert(old_threshold == (HeapWord*)round_to(
7645 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7646 "_threshold should always be card-aligned");
7647 _threshold = (HeapWord*)round_to(
7648 (intptr_t)_finger, CardTableModRefBS::card_size);
7649 MemRegion mr(old_threshold, _threshold);
7650 assert(!mr.is_empty(), "Control point invariant");
7651 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7652 _mut->clear_range(mr);
7653 }
7654
7655 // Note: the local finger doesn't advance while we drain
7656 // the stack below, but the global finger sure can and will.
7657 HeapWord** gfa = _task->global_finger_addr();
7658 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7659 _span, _bit_map,
7660 _work_queue,
7661 _overflow_stack,
7662 _finger,
7663 gfa, this);
7664 bool res = _work_queue->push(obj); // overflow could occur here
7665 assert(res, "Will hold once we use workqueues");
7666 while (true) {
7667 oop new_oop;
7668 if (!_work_queue->pop_local(new_oop)) {
7669 // We emptied our work_queue; check if there's stuff that can
7670 // be gotten from the overflow stack.
7671 if (CMSConcMarkingTask::get_work_from_overflow_stack(
7672 _overflow_stack, _work_queue)) {
7673 do_yield_check();
7674 continue;
7675 } else { // done
7676 break;
7677 }
7678 }
7679 // Skip verifying header mark word below because we are
7680 // running concurrent with mutators.
7681 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7682 // now scan this oop's oops
7683 new_oop->oop_iterate(&pushOrMarkClosure);
7684 do_yield_check();
7685 }
7686 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7687 }
7688
7689 // Yield in response to a request from VM Thread or
7690 // from mutators.
7691 void Par_MarkFromRootsClosure::do_yield_work() {
7692 assert(_task != NULL, "sanity");
7693 _task->yield();
7694 }
7695
7696 // A variant of the above used for verifying CMS marking work.
7697 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7698 MemRegion span,
7699 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7700 CMSMarkStack* mark_stack):
7701 _collector(collector),
7702 _span(span),
7703 _verification_bm(verification_bm),
7704 _cms_bm(cms_bm),
7705 _mark_stack(mark_stack),
7706 _pam_verify_closure(collector, span, verification_bm, cms_bm,
7707 mark_stack)
7708 {
7709 assert(_mark_stack->isEmpty(), "stack should be empty");
7710 _finger = _verification_bm->startWord();
7711 assert(_collector->_restart_addr == NULL, "Sanity check");
7712 assert(_span.contains(_finger), "Out of bounds _finger?");
7713 }
7714
7715 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7716 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7717 assert(_span.contains(addr), "Out of bounds _finger?");
7718 _finger = addr;
7719 }
7720
7721 // Should revisit to see if this should be restructured for
7722 // greater efficiency.
7723 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7724 // convert offset into a HeapWord*
7725 HeapWord* addr = _verification_bm->startWord() + offset;
7726 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7727 "address out of range");
7728 assert(_verification_bm->isMarked(addr), "tautology");
7729 assert(_cms_bm->isMarked(addr), "tautology");
7730
7731 assert(_mark_stack->isEmpty(),
7732 "should drain stack to limit stack usage");
7733 // convert addr to an oop preparatory to scanning
7734 oop obj = oop(addr);
7735 assert(obj->is_oop(), "should be an oop");
7736 assert(_finger <= addr, "_finger runneth ahead");
7737 // advance the finger to right end of this object
7738 _finger = addr + obj->size();
7739 assert(_finger > addr, "we just incremented it above");
7740 // Note: the finger doesn't advance while we drain
7741 // the stack below.
7742 bool res = _mark_stack->push(obj);
7743 assert(res, "Empty non-zero size stack should have space for single push");
7744 while (!_mark_stack->isEmpty()) {
7745 oop new_oop = _mark_stack->pop();
7746 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7747 // now scan this oop's oops
7748 new_oop->oop_iterate(&_pam_verify_closure);
7749 }
7750 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7751 return true;
7752 }
7753
7754 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7755 CMSCollector* collector, MemRegion span,
7756 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7757 CMSMarkStack* mark_stack):
7758 MetadataAwareOopClosure(collector->ref_processor()),
7759 _collector(collector),
7760 _span(span),
7761 _verification_bm(verification_bm),
7762 _cms_bm(cms_bm),
7763 _mark_stack(mark_stack)
7764 { }
7765
7766 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7767 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7768
7769 // Upon stack overflow, we discard (part of) the stack,
7770 // remembering the least address amongst those discarded
7771 // in CMSCollector's _restart_address.
7772 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7773 // Remember the least grey address discarded
7774 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7775 _collector->lower_restart_addr(ra);
7776 _mark_stack->reset(); // discard stack contents
7777 _mark_stack->expand(); // expand the stack if possible
7778 }
7779
7780 void PushAndMarkVerifyClosure::do_oop(oop obj) {
7781 assert(obj->is_oop_or_null(), "expected an oop or NULL");
7782 HeapWord* addr = (HeapWord*)obj;
7783 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7784 // Oop lies in _span and isn't yet grey or black
7785 _verification_bm->mark(addr); // now grey
7786 if (!_cms_bm->isMarked(addr)) {
7787 oop(addr)->print();
7788 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
7789 addr);
7790 fatal("... aborting");
7791 }
7792
7793 if (!_mark_stack->push(obj)) { // stack overflow
7794 if (PrintCMSStatistics != 0) {
7795 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7796 SIZE_FORMAT, _mark_stack->capacity());
7797 }
7798 assert(_mark_stack->isFull(), "Else push should have succeeded");
7799 handle_stack_overflow(addr);
7800 }
7801 // anything including and to the right of _finger
7802 // will be scanned as we iterate over the remainder of the
7803 // bit map
7804 }
7805 }
7806
7807 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7808 MemRegion span,
7809 CMSBitMap* bitMap, CMSMarkStack* markStack,
7810 HeapWord* finger, MarkFromRootsClosure* parent) :
7811 MetadataAwareOopClosure(collector->ref_processor()),
7812 _collector(collector),
7813 _span(span),
7814 _bitMap(bitMap),
7815 _markStack(markStack),
7816 _finger(finger),
7817 _parent(parent)
7818 { }
7819
7820 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7821 MemRegion span,
7822 CMSBitMap* bit_map,
7823 OopTaskQueue* work_queue,
7824 CMSMarkStack* overflow_stack,
7825 HeapWord* finger,
7826 HeapWord** global_finger_addr,
7827 Par_MarkFromRootsClosure* parent) :
7828 MetadataAwareOopClosure(collector->ref_processor()),
7829 _collector(collector),
7830 _whole_span(collector->_span),
7831 _span(span),
7832 _bit_map(bit_map),
7833 _work_queue(work_queue),
7834 _overflow_stack(overflow_stack),
7835 _finger(finger),
7836 _global_finger_addr(global_finger_addr),
7837 _parent(parent)
7838 { }
7839
7840 // Assumes thread-safe access by callers, who are
7841 // responsible for mutual exclusion.
7842 void CMSCollector::lower_restart_addr(HeapWord* low) {
7843 assert(_span.contains(low), "Out of bounds addr");
7844 if (_restart_addr == NULL) {
7845 _restart_addr = low;
7846 } else {
7847 _restart_addr = MIN2(_restart_addr, low);
7848 }
7849 }
7850
7851 // Upon stack overflow, we discard (part of) the stack,
7852 // remembering the least address amongst those discarded
7853 // in CMSCollector's _restart_address.
7854 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7855 // Remember the least grey address discarded
7856 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7857 _collector->lower_restart_addr(ra);
7858 _markStack->reset(); // discard stack contents
7859 _markStack->expand(); // expand the stack if possible
7860 }
7861
7862 // Upon stack overflow, we discard (part of) the stack,
7863 // remembering the least address amongst those discarded
7864 // in CMSCollector's _restart_address.
7865 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7866 // We need to do this under a mutex to prevent other
7867 // workers from interfering with the work done below.
7868 MutexLockerEx ml(_overflow_stack->par_lock(),
7869 Mutex::_no_safepoint_check_flag);
7870 // Remember the least grey address discarded
7871 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7872 _collector->lower_restart_addr(ra);
7873 _overflow_stack->reset(); // discard stack contents
7874 _overflow_stack->expand(); // expand the stack if possible
7875 }
7876
7877 void PushOrMarkClosure::do_oop(oop obj) {
7878 // Ignore mark word because we are running concurrent with mutators.
7879 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7880 HeapWord* addr = (HeapWord*)obj;
7881 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7882 // Oop lies in _span and isn't yet grey or black
7883 _bitMap->mark(addr); // now grey
7884 if (addr < _finger) {
7885 // the bit map iteration has already either passed, or
7886 // sampled, this bit in the bit map; we'll need to
7887 // use the marking stack to scan this oop's oops.
7888 bool simulate_overflow = false;
7889 NOT_PRODUCT(
7890 if (CMSMarkStackOverflowALot &&
7891 _collector->simulate_overflow()) {
7892 // simulate a stack overflow
7893 simulate_overflow = true;
7894 }
7895 )
7896 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7897 if (PrintCMSStatistics != 0) {
7898 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7899 SIZE_FORMAT, _markStack->capacity());
7900 }
7901 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7902 handle_stack_overflow(addr);
7903 }
7904 }
7905 // anything including and to the right of _finger
7906 // will be scanned as we iterate over the remainder of the
7907 // bit map
7908 do_yield_check();
7909 }
7910 }
7911
7912 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7913 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7914
7915 void Par_PushOrMarkClosure::do_oop(oop obj) {
7916 // Ignore mark word because we are running concurrent with mutators.
7917 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7918 HeapWord* addr = (HeapWord*)obj;
7919 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7920 // Oop lies in _span and isn't yet grey or black
7921 // We read the global_finger (volatile read) strictly after marking oop
7922 bool res = _bit_map->par_mark(addr); // now grey
7923 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7924 // Should we push this marked oop on our stack?
7925 // -- if someone else marked it, nothing to do
7926 // -- if target oop is above global finger nothing to do
7927 // -- if target oop is in chunk and above local finger
7928 // then nothing to do
7929 // -- else push on work queue
7930 if ( !res // someone else marked it, they will deal with it
7931 || (addr >= *gfa) // will be scanned in a later task
7932 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7933 return;
7934 }
7935 // the bit map iteration has already either passed, or
7936 // sampled, this bit in the bit map; we'll need to
7937 // use the marking stack to scan this oop's oops.
7938 bool simulate_overflow = false;
7939 NOT_PRODUCT(
7940 if (CMSMarkStackOverflowALot &&
7941 _collector->simulate_overflow()) {
7942 // simulate a stack overflow
7943 simulate_overflow = true;
7944 }
7945 )
7946 if (simulate_overflow ||
7947 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7948 // stack overflow
7949 if (PrintCMSStatistics != 0) {
7950 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7951 SIZE_FORMAT, _overflow_stack->capacity());
7952 }
7953 // We cannot assert that the overflow stack is full because
7954 // it may have been emptied since.
7955 assert(simulate_overflow ||
7956 _work_queue->size() == _work_queue->max_elems(),
7957 "Else push should have succeeded");
7958 handle_stack_overflow(addr);
7959 }
7960 do_yield_check();
7961 }
7962 }
7963
7964 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7965 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7966
7967 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7968 MemRegion span,
7969 ReferenceProcessor* rp,
7970 CMSBitMap* bit_map,
7971 CMSBitMap* mod_union_table,
7972 CMSMarkStack* mark_stack,
7973 bool concurrent_precleaning):
7974 MetadataAwareOopClosure(rp),
7975 _collector(collector),
7976 _span(span),
7977 _bit_map(bit_map),
7978 _mod_union_table(mod_union_table),
7979 _mark_stack(mark_stack),
7980 _concurrent_precleaning(concurrent_precleaning)
7981 {
7982 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7983 }
7984
7985 // Grey object rescan during pre-cleaning and second checkpoint phases --
7986 // the non-parallel version (the parallel version appears further below.)
7987 void PushAndMarkClosure::do_oop(oop obj) {
7988 // Ignore mark word verification. If during concurrent precleaning,
7989 // the object monitor may be locked. If during the checkpoint
7990 // phases, the object may already have been reached by a different
7991 // path and may be at the end of the global overflow list (so
7992 // the mark word may be NULL).
7993 assert(obj->is_oop_or_null(true /* ignore mark word */),
7994 "expected an oop or NULL");
7995 HeapWord* addr = (HeapWord*)obj;
7996 // Check if oop points into the CMS generation
7997 // and is not marked
7998 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7999 // a white object ...
8000 _bit_map->mark(addr); // ... now grey
8001 // push on the marking stack (grey set)
8002 bool simulate_overflow = false;
8003 NOT_PRODUCT(
8004 if (CMSMarkStackOverflowALot &&
8005 _collector->simulate_overflow()) {
8006 // simulate a stack overflow
8007 simulate_overflow = true;
8008 }
8009 )
8010 if (simulate_overflow || !_mark_stack->push(obj)) {
8011 if (_concurrent_precleaning) {
8012 // During precleaning we can just dirty the appropriate card(s)
8013 // in the mod union table, thus ensuring that the object remains
8014 // in the grey set and continue. In the case of object arrays
8015 // we need to dirty all of the cards that the object spans,
8016 // since the rescan of object arrays will be limited to the
8017 // dirty cards.
8018 // Note that no one can be intefering with us in this action
8019 // of dirtying the mod union table, so no locking or atomics
8020 // are required.
8021 if (obj->is_objArray()) {
8022 size_t sz = obj->size();
8023 HeapWord* end_card_addr = (HeapWord*)round_to(
8024 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
8025 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8026 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8027 _mod_union_table->mark_range(redirty_range);
8028 } else {
8029 _mod_union_table->mark(addr);
8030 }
8031 _collector->_ser_pmc_preclean_ovflw++;
8032 } else {
8033 // During the remark phase, we need to remember this oop
8034 // in the overflow list.
8035 _collector->push_on_overflow_list(obj);
8036 _collector->_ser_pmc_remark_ovflw++;
8037 }
8038 }
8039 }
8040 }
8041
8042 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
8043 MemRegion span,
8044 ReferenceProcessor* rp,
8045 CMSBitMap* bit_map,
8046 OopTaskQueue* work_queue):
8047 MetadataAwareOopClosure(rp),
8048 _collector(collector),
8049 _span(span),
8050 _bit_map(bit_map),
8051 _work_queue(work_queue)
8052 {
8053 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
8054 }
8055
8056 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
8057 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
8058
8059 // Grey object rescan during second checkpoint phase --
8060 // the parallel version.
8061 void Par_PushAndMarkClosure::do_oop(oop obj) {
8062 // In the assert below, we ignore the mark word because
8063 // this oop may point to an already visited object that is
8064 // on the overflow stack (in which case the mark word has
8065 // been hijacked for chaining into the overflow stack --
8066 // if this is the last object in the overflow stack then
8067 // its mark word will be NULL). Because this object may
8068 // have been subsequently popped off the global overflow
8069 // stack, and the mark word possibly restored to the prototypical
8070 // value, by the time we get to examined this failing assert in
8071 // the debugger, is_oop_or_null(false) may subsequently start
8072 // to hold.
8073 assert(obj->is_oop_or_null(true),
8074 "expected an oop or NULL");
8075 HeapWord* addr = (HeapWord*)obj;
8076 // Check if oop points into the CMS generation
8077 // and is not marked
8078 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
8079 // a white object ...
8080 // If we manage to "claim" the object, by being the
8081 // first thread to mark it, then we push it on our
8082 // marking stack
8083 if (_bit_map->par_mark(addr)) { // ... now grey
8084 // push on work queue (grey set)
8085 bool simulate_overflow = false;
8086 NOT_PRODUCT(
8087 if (CMSMarkStackOverflowALot &&
8088 _collector->par_simulate_overflow()) {
8089 // simulate a stack overflow
8090 simulate_overflow = true;
8091 }
8092 )
8093 if (simulate_overflow || !_work_queue->push(obj)) {
8094 _collector->par_push_on_overflow_list(obj);
8095 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
8096 }
8097 } // Else, some other thread got there first
8098 }
8099 }
8100
8101 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
8102 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
8103
8104 void CMSPrecleanRefsYieldClosure::do_yield_work() {
8105 Mutex* bml = _collector->bitMapLock();
8106 assert_lock_strong(bml);
8107 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8108 "CMS thread should hold CMS token");
8109
8110 bml->unlock();
8111 ConcurrentMarkSweepThread::desynchronize(true);
8112
8113 ConcurrentMarkSweepThread::acknowledge_yield_request();
8114
8115 _collector->stopTimer();
8116 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8117 if (PrintCMSStatistics != 0) {
8118 _collector->incrementYields();
8119 }
8120 _collector->icms_wait();
8121
8122 // See the comment in coordinator_yield()
8123 for (unsigned i = 0; i < CMSYieldSleepCount &&
8124 ConcurrentMarkSweepThread::should_yield() &&
8125 !CMSCollector::foregroundGCIsActive(); ++i) {
8126 os::sleep(Thread::current(), 1, false);
8127 ConcurrentMarkSweepThread::acknowledge_yield_request();
8128 }
8129
8130 ConcurrentMarkSweepThread::synchronize(true);
8131 bml->lock();
8132
8133 _collector->startTimer();
8134 }
8135
8136 bool CMSPrecleanRefsYieldClosure::should_return() {
8137 if (ConcurrentMarkSweepThread::should_yield()) {
8138 do_yield_work();
8139 }
8140 return _collector->foregroundGCIsActive();
8141 }
8142
8143 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
8144 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
8145 "mr should be aligned to start at a card boundary");
8146 // We'd like to assert:
8147 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
8148 // "mr should be a range of cards");
8149 // However, that would be too strong in one case -- the last
8150 // partition ends at _unallocated_block which, in general, can be
8151 // an arbitrary boundary, not necessarily card aligned.
8152 if (PrintCMSStatistics != 0) {
8153 _num_dirty_cards +=
8154 mr.word_size()/CardTableModRefBS::card_size_in_words;
8155 }
8156 _space->object_iterate_mem(mr, &_scan_cl);
8157 }
8158
8159 SweepClosure::SweepClosure(CMSCollector* collector,
8160 ConcurrentMarkSweepGeneration* g,
8161 CMSBitMap* bitMap, bool should_yield) :
8162 _collector(collector),
8163 _g(g),
8164 _sp(g->cmsSpace()),
8165 _limit(_sp->sweep_limit()),
8166 _freelistLock(_sp->freelistLock()),
8167 _bitMap(bitMap),
8168 _yield(should_yield),
8169 _inFreeRange(false), // No free range at beginning of sweep
8170 _freeRangeInFreeLists(false), // No free range at beginning of sweep
8171 _lastFreeRangeCoalesced(false),
8172 _freeFinger(g->used_region().start())
8173 {
8174 NOT_PRODUCT(
8175 _numObjectsFreed = 0;
8176 _numWordsFreed = 0;
8177 _numObjectsLive = 0;
8178 _numWordsLive = 0;
8179 _numObjectsAlreadyFree = 0;
8180 _numWordsAlreadyFree = 0;
8181 _last_fc = NULL;
8182
8183 _sp->initializeIndexedFreeListArrayReturnedBytes();
8184 _sp->dictionary()->initialize_dict_returned_bytes();
8185 )
8186 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8187 "sweep _limit out of bounds");
8188 if (CMSTraceSweeper) {
8189 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
8190 _limit);
8191 }
8192 }
8193
8194 void SweepClosure::print_on(outputStream* st) const {
8195 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
8196 _sp->bottom(), _sp->end());
8197 tty->print_cr("_limit = " PTR_FORMAT, _limit);
8198 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger);
8199 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);)
8200 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
8201 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
8202 }
8203
8204 #ifndef PRODUCT
8205 // Assertion checking only: no useful work in product mode --
8206 // however, if any of the flags below become product flags,
8207 // you may need to review this code to see if it needs to be
8208 // enabled in product mode.
8209 SweepClosure::~SweepClosure() {
8210 assert_lock_strong(_freelistLock);
8211 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8212 "sweep _limit out of bounds");
8213 if (inFreeRange()) {
8214 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
8215 print();
8216 ShouldNotReachHere();
8217 }
8218 if (Verbose && PrintGC) {
8219 gclog_or_tty->print("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
8220 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
8221 gclog_or_tty->print_cr("\nLive " SIZE_FORMAT " objects, "
8222 SIZE_FORMAT " bytes "
8223 "Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
8224 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
8225 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
8226 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
8227 * sizeof(HeapWord);
8228 gclog_or_tty->print_cr("Total sweep: " SIZE_FORMAT " bytes", totalBytes);
8229
8230 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
8231 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
8232 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
8233 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
8234 gclog_or_tty->print("Returned " SIZE_FORMAT " bytes", returned_bytes);
8235 gclog_or_tty->print(" Indexed List Returned " SIZE_FORMAT " bytes",
8236 indexListReturnedBytes);
8237 gclog_or_tty->print_cr(" Dictionary Returned " SIZE_FORMAT " bytes",
8238 dict_returned_bytes);
8239 }
8240 }
8241 if (CMSTraceSweeper) {
8242 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
8243 _limit);
8244 }
8245 }
8246 #endif // PRODUCT
8247
8248 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
8249 bool freeRangeInFreeLists) {
8250 if (CMSTraceSweeper) {
8251 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n",
8252 freeFinger, freeRangeInFreeLists);
8253 }
8254 assert(!inFreeRange(), "Trampling existing free range");
8255 set_inFreeRange(true);
8256 set_lastFreeRangeCoalesced(false);
8257
8258 set_freeFinger(freeFinger);
8259 set_freeRangeInFreeLists(freeRangeInFreeLists);
8260 if (CMSTestInFreeList) {
8261 if (freeRangeInFreeLists) {
8262 FreeChunk* fc = (FreeChunk*) freeFinger;
8263 assert(fc->is_free(), "A chunk on the free list should be free.");
8264 assert(fc->size() > 0, "Free range should have a size");
8265 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
8266 }
8267 }
8268 }
8269
8270 // Note that the sweeper runs concurrently with mutators. Thus,
8271 // it is possible for direct allocation in this generation to happen
8272 // in the middle of the sweep. Note that the sweeper also coalesces
8273 // contiguous free blocks. Thus, unless the sweeper and the allocator
8274 // synchronize appropriately freshly allocated blocks may get swept up.
8275 // This is accomplished by the sweeper locking the free lists while
8276 // it is sweeping. Thus blocks that are determined to be free are
8277 // indeed free. There is however one additional complication:
8278 // blocks that have been allocated since the final checkpoint and
8279 // mark, will not have been marked and so would be treated as
8280 // unreachable and swept up. To prevent this, the allocator marks
8281 // the bit map when allocating during the sweep phase. This leads,
8282 // however, to a further complication -- objects may have been allocated
8283 // but not yet initialized -- in the sense that the header isn't yet
8284 // installed. The sweeper can not then determine the size of the block
8285 // in order to skip over it. To deal with this case, we use a technique
8286 // (due to Printezis) to encode such uninitialized block sizes in the
8287 // bit map. Since the bit map uses a bit per every HeapWord, but the
8288 // CMS generation has a minimum object size of 3 HeapWords, it follows
8289 // that "normal marks" won't be adjacent in the bit map (there will
8290 // always be at least two 0 bits between successive 1 bits). We make use
8291 // of these "unused" bits to represent uninitialized blocks -- the bit
8292 // corresponding to the start of the uninitialized object and the next
8293 // bit are both set. Finally, a 1 bit marks the end of the object that
8294 // started with the two consecutive 1 bits to indicate its potentially
8295 // uninitialized state.
8296
8297 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
8298 FreeChunk* fc = (FreeChunk*)addr;
8299 size_t res;
8300
8301 // Check if we are done sweeping. Below we check "addr >= _limit" rather
8302 // than "addr == _limit" because although _limit was a block boundary when
8303 // we started the sweep, it may no longer be one because heap expansion
8304 // may have caused us to coalesce the block ending at the address _limit
8305 // with a newly expanded chunk (this happens when _limit was set to the
8306 // previous _end of the space), so we may have stepped past _limit:
8307 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
8308 if (addr >= _limit) { // we have swept up to or past the limit: finish up
8309 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8310 "sweep _limit out of bounds");
8311 assert(addr < _sp->end(), "addr out of bounds");
8312 // Flush any free range we might be holding as a single
8313 // coalesced chunk to the appropriate free list.
8314 if (inFreeRange()) {
8315 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
8316 err_msg("freeFinger() " PTR_FORMAT " is out-of-bounds", freeFinger()));
8317 flush_cur_free_chunk(freeFinger(),
8318 pointer_delta(addr, freeFinger()));
8319 if (CMSTraceSweeper) {
8320 gclog_or_tty->print("Sweep: last chunk: ");
8321 gclog_or_tty->print("put_free_blk 0x%x (" SIZE_FORMAT ") "
8322 "[coalesced:" SIZE_FORMAT "]\n",
8323 freeFinger(), pointer_delta(addr, freeFinger()),
8324 lastFreeRangeCoalesced());
8325 }
8326 }
8327
8328 // help the iterator loop finish
8329 return pointer_delta(_sp->end(), addr);
8330 }
8331
8332 assert(addr < _limit, "sweep invariant");
8333 // check if we should yield
8334 do_yield_check(addr);
8335 if (fc->is_free()) {
8336 // Chunk that is already free
8337 res = fc->size();
8338 do_already_free_chunk(fc);
8339 debug_only(_sp->verifyFreeLists());
8340 // If we flush the chunk at hand in lookahead_and_flush()
8341 // and it's coalesced with a preceding chunk, then the
8342 // process of "mangling" the payload of the coalesced block
8343 // will cause erasure of the size information from the
8344 // (erstwhile) header of all the coalesced blocks but the
8345 // first, so the first disjunct in the assert will not hold
8346 // in that specific case (in which case the second disjunct
8347 // will hold).
8348 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
8349 "Otherwise the size info doesn't change at this step");
8350 NOT_PRODUCT(
8351 _numObjectsAlreadyFree++;
8352 _numWordsAlreadyFree += res;
8353 )
8354 NOT_PRODUCT(_last_fc = fc;)
8355 } else if (!_bitMap->isMarked(addr)) {
8356 // Chunk is fresh garbage
8357 res = do_garbage_chunk(fc);
8358 debug_only(_sp->verifyFreeLists());
8359 NOT_PRODUCT(
8360 _numObjectsFreed++;
8361 _numWordsFreed += res;
8362 )
8363 } else {
8364 // Chunk that is alive.
8365 res = do_live_chunk(fc);
8366 debug_only(_sp->verifyFreeLists());
8367 NOT_PRODUCT(
8368 _numObjectsLive++;
8369 _numWordsLive += res;
8370 )
8371 }
8372 return res;
8373 }
8374
8375 // For the smart allocation, record following
8376 // split deaths - a free chunk is removed from its free list because
8377 // it is being split into two or more chunks.
8378 // split birth - a free chunk is being added to its free list because
8379 // a larger free chunk has been split and resulted in this free chunk.
8380 // coal death - a free chunk is being removed from its free list because
8381 // it is being coalesced into a large free chunk.
8382 // coal birth - a free chunk is being added to its free list because
8383 // it was created when two or more free chunks where coalesced into
8384 // this free chunk.
8385 //
8386 // These statistics are used to determine the desired number of free
8387 // chunks of a given size. The desired number is chosen to be relative
8388 // to the end of a CMS sweep. The desired number at the end of a sweep
8389 // is the
8390 // count-at-end-of-previous-sweep (an amount that was enough)
8391 // - count-at-beginning-of-current-sweep (the excess)
8392 // + split-births (gains in this size during interval)
8393 // - split-deaths (demands on this size during interval)
8394 // where the interval is from the end of one sweep to the end of the
8395 // next.
8396 //
8397 // When sweeping the sweeper maintains an accumulated chunk which is
8398 // the chunk that is made up of chunks that have been coalesced. That
8399 // will be termed the left-hand chunk. A new chunk of garbage that
8400 // is being considered for coalescing will be referred to as the
8401 // right-hand chunk.
8402 //
8403 // When making a decision on whether to coalesce a right-hand chunk with
8404 // the current left-hand chunk, the current count vs. the desired count
8405 // of the left-hand chunk is considered. Also if the right-hand chunk
8406 // is near the large chunk at the end of the heap (see
8407 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
8408 // left-hand chunk is coalesced.
8409 //
8410 // When making a decision about whether to split a chunk, the desired count
8411 // vs. the current count of the candidate to be split is also considered.
8412 // If the candidate is underpopulated (currently fewer chunks than desired)
8413 // a chunk of an overpopulated (currently more chunks than desired) size may
8414 // be chosen. The "hint" associated with a free list, if non-null, points
8415 // to a free list which may be overpopulated.
8416 //
8417
8418 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
8419 const size_t size = fc->size();
8420 // Chunks that cannot be coalesced are not in the
8421 // free lists.
8422 if (CMSTestInFreeList && !fc->cantCoalesce()) {
8423 assert(_sp->verify_chunk_in_free_list(fc),
8424 "free chunk should be in free lists");
8425 }
8426 // a chunk that is already free, should not have been
8427 // marked in the bit map
8428 HeapWord* const addr = (HeapWord*) fc;
8429 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
8430 // Verify that the bit map has no bits marked between
8431 // addr and purported end of this block.
8432 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8433
8434 // Some chunks cannot be coalesced under any circumstances.
8435 // See the definition of cantCoalesce().
8436 if (!fc->cantCoalesce()) {
8437 // This chunk can potentially be coalesced.
8438 if (_sp->adaptive_freelists()) {
8439 // All the work is done in
8440 do_post_free_or_garbage_chunk(fc, size);
8441 } else { // Not adaptive free lists
8442 // this is a free chunk that can potentially be coalesced by the sweeper;
8443 if (!inFreeRange()) {
8444 // if the next chunk is a free block that can't be coalesced
8445 // it doesn't make sense to remove this chunk from the free lists
8446 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
8447 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
8448 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
8449 nextChunk->is_free() && // ... which is free...
8450 nextChunk->cantCoalesce()) { // ... but can't be coalesced
8451 // nothing to do
8452 } else {
8453 // Potentially the start of a new free range:
8454 // Don't eagerly remove it from the free lists.
8455 // No need to remove it if it will just be put
8456 // back again. (Also from a pragmatic point of view
8457 // if it is a free block in a region that is beyond
8458 // any allocated blocks, an assertion will fail)
8459 // Remember the start of a free run.
8460 initialize_free_range(addr, true);
8461 // end - can coalesce with next chunk
8462 }
8463 } else {
8464 // the midst of a free range, we are coalescing
8465 print_free_block_coalesced(fc);
8466 if (CMSTraceSweeper) {
8467 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size);
8468 }
8469 // remove it from the free lists
8470 _sp->removeFreeChunkFromFreeLists(fc);
8471 set_lastFreeRangeCoalesced(true);
8472 // If the chunk is being coalesced and the current free range is
8473 // in the free lists, remove the current free range so that it
8474 // will be returned to the free lists in its entirety - all
8475 // the coalesced pieces included.
8476 if (freeRangeInFreeLists()) {
8477 FreeChunk* ffc = (FreeChunk*) freeFinger();
8478 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8479 "Size of free range is inconsistent with chunk size.");
8480 if (CMSTestInFreeList) {
8481 assert(_sp->verify_chunk_in_free_list(ffc),
8482 "free range is not in free lists");
8483 }
8484 _sp->removeFreeChunkFromFreeLists(ffc);
8485 set_freeRangeInFreeLists(false);
8486 }
8487 }
8488 }
8489 // Note that if the chunk is not coalescable (the else arm
8490 // below), we unconditionally flush, without needing to do
8491 // a "lookahead," as we do below.
8492 if (inFreeRange()) lookahead_and_flush(fc, size);
8493 } else {
8494 // Code path common to both original and adaptive free lists.
8495
8496 // cant coalesce with previous block; this should be treated
8497 // as the end of a free run if any
8498 if (inFreeRange()) {
8499 // we kicked some butt; time to pick up the garbage
8500 assert(freeFinger() < addr, "freeFinger points too high");
8501 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8502 }
8503 // else, nothing to do, just continue
8504 }
8505 }
8506
8507 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
8508 // This is a chunk of garbage. It is not in any free list.
8509 // Add it to a free list or let it possibly be coalesced into
8510 // a larger chunk.
8511 HeapWord* const addr = (HeapWord*) fc;
8512 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8513
8514 if (_sp->adaptive_freelists()) {
8515 // Verify that the bit map has no bits marked between
8516 // addr and purported end of just dead object.
8517 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8518
8519 do_post_free_or_garbage_chunk(fc, size);
8520 } else {
8521 if (!inFreeRange()) {
8522 // start of a new free range
8523 assert(size > 0, "A free range should have a size");
8524 initialize_free_range(addr, false);
8525 } else {
8526 // this will be swept up when we hit the end of the
8527 // free range
8528 if (CMSTraceSweeper) {
8529 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size);
8530 }
8531 // If the chunk is being coalesced and the current free range is
8532 // in the free lists, remove the current free range so that it
8533 // will be returned to the free lists in its entirety - all
8534 // the coalesced pieces included.
8535 if (freeRangeInFreeLists()) {
8536 FreeChunk* ffc = (FreeChunk*)freeFinger();
8537 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8538 "Size of free range is inconsistent with chunk size.");
8539 if (CMSTestInFreeList) {
8540 assert(_sp->verify_chunk_in_free_list(ffc),
8541 "free range is not in free lists");
8542 }
8543 _sp->removeFreeChunkFromFreeLists(ffc);
8544 set_freeRangeInFreeLists(false);
8545 }
8546 set_lastFreeRangeCoalesced(true);
8547 }
8548 // this will be swept up when we hit the end of the free range
8549
8550 // Verify that the bit map has no bits marked between
8551 // addr and purported end of just dead object.
8552 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8553 }
8554 assert(_limit >= addr + size,
8555 "A freshly garbage chunk can't possibly straddle over _limit");
8556 if (inFreeRange()) lookahead_and_flush(fc, size);
8557 return size;
8558 }
8559
8560 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
8561 HeapWord* addr = (HeapWord*) fc;
8562 // The sweeper has just found a live object. Return any accumulated
8563 // left hand chunk to the free lists.
8564 if (inFreeRange()) {
8565 assert(freeFinger() < addr, "freeFinger points too high");
8566 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8567 }
8568
8569 // This object is live: we'd normally expect this to be
8570 // an oop, and like to assert the following:
8571 // assert(oop(addr)->is_oop(), "live block should be an oop");
8572 // However, as we commented above, this may be an object whose
8573 // header hasn't yet been initialized.
8574 size_t size;
8575 assert(_bitMap->isMarked(addr), "Tautology for this control point");
8576 if (_bitMap->isMarked(addr + 1)) {
8577 // Determine the size from the bit map, rather than trying to
8578 // compute it from the object header.
8579 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
8580 size = pointer_delta(nextOneAddr + 1, addr);
8581 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
8582 "alignment problem");
8583
8584 #ifdef ASSERT
8585 if (oop(addr)->klass_or_null() != NULL) {
8586 // Ignore mark word because we are running concurrent with mutators
8587 assert(oop(addr)->is_oop(true), "live block should be an oop");
8588 assert(size ==
8589 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
8590 "P-mark and computed size do not agree");
8591 }
8592 #endif
8593
8594 } else {
8595 // This should be an initialized object that's alive.
8596 assert(oop(addr)->klass_or_null() != NULL,
8597 "Should be an initialized object");
8598 // Ignore mark word because we are running concurrent with mutators
8599 assert(oop(addr)->is_oop(true), "live block should be an oop");
8600 // Verify that the bit map has no bits marked between
8601 // addr and purported end of this block.
8602 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8603 assert(size >= 3, "Necessary for Printezis marks to work");
8604 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
8605 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
8606 }
8607 return size;
8608 }
8609
8610 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
8611 size_t chunkSize) {
8612 // do_post_free_or_garbage_chunk() should only be called in the case
8613 // of the adaptive free list allocator.
8614 const bool fcInFreeLists = fc->is_free();
8615 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
8616 assert((HeapWord*)fc <= _limit, "sweep invariant");
8617 if (CMSTestInFreeList && fcInFreeLists) {
8618 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
8619 }
8620
8621 if (CMSTraceSweeper) {
8622 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
8623 }
8624
8625 HeapWord* const fc_addr = (HeapWord*) fc;
8626
8627 bool coalesce = false;
8628 const size_t left = pointer_delta(fc_addr, freeFinger());
8629 const size_t right = chunkSize;
8630 switch (FLSCoalescePolicy) {
8631 // numeric value forms a coalition aggressiveness metric
8632 case 0: { // never coalesce
8633 coalesce = false;
8634 break;
8635 }
8636 case 1: { // coalesce if left & right chunks on overpopulated lists
8637 coalesce = _sp->coalOverPopulated(left) &&
8638 _sp->coalOverPopulated(right);
8639 break;
8640 }
8641 case 2: { // coalesce if left chunk on overpopulated list (default)
8642 coalesce = _sp->coalOverPopulated(left);
8643 break;
8644 }
8645 case 3: { // coalesce if left OR right chunk on overpopulated list
8646 coalesce = _sp->coalOverPopulated(left) ||
8647 _sp->coalOverPopulated(right);
8648 break;
8649 }
8650 case 4: { // always coalesce
8651 coalesce = true;
8652 break;
8653 }
8654 default:
8655 ShouldNotReachHere();
8656 }
8657
8658 // Should the current free range be coalesced?
8659 // If the chunk is in a free range and either we decided to coalesce above
8660 // or the chunk is near the large block at the end of the heap
8661 // (isNearLargestChunk() returns true), then coalesce this chunk.
8662 const bool doCoalesce = inFreeRange()
8663 && (coalesce || _g->isNearLargestChunk(fc_addr));
8664 if (doCoalesce) {
8665 // Coalesce the current free range on the left with the new
8666 // chunk on the right. If either is on a free list,
8667 // it must be removed from the list and stashed in the closure.
8668 if (freeRangeInFreeLists()) {
8669 FreeChunk* const ffc = (FreeChunk*)freeFinger();
8670 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
8671 "Size of free range is inconsistent with chunk size.");
8672 if (CMSTestInFreeList) {
8673 assert(_sp->verify_chunk_in_free_list(ffc),
8674 "Chunk is not in free lists");
8675 }
8676 _sp->coalDeath(ffc->size());
8677 _sp->removeFreeChunkFromFreeLists(ffc);
8678 set_freeRangeInFreeLists(false);
8679 }
8680 if (fcInFreeLists) {
8681 _sp->coalDeath(chunkSize);
8682 assert(fc->size() == chunkSize,
8683 "The chunk has the wrong size or is not in the free lists");
8684 _sp->removeFreeChunkFromFreeLists(fc);
8685 }
8686 set_lastFreeRangeCoalesced(true);
8687 print_free_block_coalesced(fc);
8688 } else { // not in a free range and/or should not coalesce
8689 // Return the current free range and start a new one.
8690 if (inFreeRange()) {
8691 // In a free range but cannot coalesce with the right hand chunk.
8692 // Put the current free range into the free lists.
8693 flush_cur_free_chunk(freeFinger(),
8694 pointer_delta(fc_addr, freeFinger()));
8695 }
8696 // Set up for new free range. Pass along whether the right hand
8697 // chunk is in the free lists.
8698 initialize_free_range((HeapWord*)fc, fcInFreeLists);
8699 }
8700 }
8701
8702 // Lookahead flush:
8703 // If we are tracking a free range, and this is the last chunk that
8704 // we'll look at because its end crosses past _limit, we'll preemptively
8705 // flush it along with any free range we may be holding on to. Note that
8706 // this can be the case only for an already free or freshly garbage
8707 // chunk. If this block is an object, it can never straddle
8708 // over _limit. The "straddling" occurs when _limit is set at
8709 // the previous end of the space when this cycle started, and
8710 // a subsequent heap expansion caused the previously co-terminal
8711 // free block to be coalesced with the newly expanded portion,
8712 // thus rendering _limit a non-block-boundary making it dangerous
8713 // for the sweeper to step over and examine.
8714 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
8715 assert(inFreeRange(), "Should only be called if currently in a free range.");
8716 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
8717 assert(_sp->used_region().contains(eob - 1),
8718 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
8719 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
8720 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
8721 eob, eob-1, _limit, _sp->bottom(), _sp->end(), fc, chunk_size));
8722 if (eob >= _limit) {
8723 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
8724 if (CMSTraceSweeper) {
8725 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
8726 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
8727 "[" PTR_FORMAT "," PTR_FORMAT ")",
8728 _limit, fc, eob, _sp->bottom(), _sp->end());
8729 }
8730 // Return the storage we are tracking back into the free lists.
8731 if (CMSTraceSweeper) {
8732 gclog_or_tty->print_cr("Flushing ... ");
8733 }
8734 assert(freeFinger() < eob, "Error");
8735 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
8736 }
8737 }
8738
8739 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
8740 assert(inFreeRange(), "Should only be called if currently in a free range.");
8741 assert(size > 0,
8742 "A zero sized chunk cannot be added to the free lists.");
8743 if (!freeRangeInFreeLists()) {
8744 if (CMSTestInFreeList) {
8745 FreeChunk* fc = (FreeChunk*) chunk;
8746 fc->set_size(size);
8747 assert(!_sp->verify_chunk_in_free_list(fc),
8748 "chunk should not be in free lists yet");
8749 }
8750 if (CMSTraceSweeper) {
8751 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8752 chunk, size);
8753 }
8754 // A new free range is going to be starting. The current
8755 // free range has not been added to the free lists yet or
8756 // was removed so add it back.
8757 // If the current free range was coalesced, then the death
8758 // of the free range was recorded. Record a birth now.
8759 if (lastFreeRangeCoalesced()) {
8760 _sp->coalBirth(size);
8761 }
8762 _sp->addChunkAndRepairOffsetTable(chunk, size,
8763 lastFreeRangeCoalesced());
8764 } else if (CMSTraceSweeper) {
8765 gclog_or_tty->print_cr("Already in free list: nothing to flush");
8766 }
8767 set_inFreeRange(false);
8768 set_freeRangeInFreeLists(false);
8769 }
8770
8771 // We take a break if we've been at this for a while,
8772 // so as to avoid monopolizing the locks involved.
8773 void SweepClosure::do_yield_work(HeapWord* addr) {
8774 // Return current free chunk being used for coalescing (if any)
8775 // to the appropriate freelist. After yielding, the next
8776 // free block encountered will start a coalescing range of
8777 // free blocks. If the next free block is adjacent to the
8778 // chunk just flushed, they will need to wait for the next
8779 // sweep to be coalesced.
8780 if (inFreeRange()) {
8781 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8782 }
8783
8784 // First give up the locks, then yield, then re-lock.
8785 // We should probably use a constructor/destructor idiom to
8786 // do this unlock/lock or modify the MutexUnlocker class to
8787 // serve our purpose. XXX
8788 assert_lock_strong(_bitMap->lock());
8789 assert_lock_strong(_freelistLock);
8790 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8791 "CMS thread should hold CMS token");
8792 _bitMap->lock()->unlock();
8793 _freelistLock->unlock();
8794 ConcurrentMarkSweepThread::desynchronize(true);
8795 ConcurrentMarkSweepThread::acknowledge_yield_request();
8796 _collector->stopTimer();
8797 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8798 if (PrintCMSStatistics != 0) {
8799 _collector->incrementYields();
8800 }
8801 _collector->icms_wait();
8802
8803 // See the comment in coordinator_yield()
8804 for (unsigned i = 0; i < CMSYieldSleepCount &&
8805 ConcurrentMarkSweepThread::should_yield() &&
8806 !CMSCollector::foregroundGCIsActive(); ++i) {
8807 os::sleep(Thread::current(), 1, false);
8808 ConcurrentMarkSweepThread::acknowledge_yield_request();
8809 }
8810
8811 ConcurrentMarkSweepThread::synchronize(true);
8812 _freelistLock->lock();
8813 _bitMap->lock()->lock_without_safepoint_check();
8814 _collector->startTimer();
8815 }
8816
8817 #ifndef PRODUCT
8818 // This is actually very useful in a product build if it can
8819 // be called from the debugger. Compile it into the product
8820 // as needed.
8821 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
8822 return debug_cms_space->verify_chunk_in_free_list(fc);
8823 }
8824 #endif
8825
8826 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
8827 if (CMSTraceSweeper) {
8828 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
8829 fc, fc->size());
8830 }
8831 }
8832
8833 // CMSIsAliveClosure
8834 bool CMSIsAliveClosure::do_object_b(oop obj) {
8835 HeapWord* addr = (HeapWord*)obj;
8836 return addr != NULL &&
8837 (!_span.contains(addr) || _bit_map->isMarked(addr));
8838 }
8839
8840
8841 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
8842 MemRegion span,
8843 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
8844 bool cpc):
8845 _collector(collector),
8846 _span(span),
8847 _bit_map(bit_map),
8848 _mark_stack(mark_stack),
8849 _concurrent_precleaning(cpc) {
8850 assert(!_span.is_empty(), "Empty span could spell trouble");
8851 }
8852
8853
8854 // CMSKeepAliveClosure: the serial version
8855 void CMSKeepAliveClosure::do_oop(oop obj) {
8856 HeapWord* addr = (HeapWord*)obj;
8857 if (_span.contains(addr) &&
8858 !_bit_map->isMarked(addr)) {
8859 _bit_map->mark(addr);
8860 bool simulate_overflow = false;
8861 NOT_PRODUCT(
8862 if (CMSMarkStackOverflowALot &&
8863 _collector->simulate_overflow()) {
8864 // simulate a stack overflow
8865 simulate_overflow = true;
8866 }
8867 )
8868 if (simulate_overflow || !_mark_stack->push(obj)) {
8869 if (_concurrent_precleaning) {
8870 // We dirty the overflown object and let the remark
8871 // phase deal with it.
8872 assert(_collector->overflow_list_is_empty(), "Error");
8873 // In the case of object arrays, we need to dirty all of
8874 // the cards that the object spans. No locking or atomics
8875 // are needed since no one else can be mutating the mod union
8876 // table.
8877 if (obj->is_objArray()) {
8878 size_t sz = obj->size();
8879 HeapWord* end_card_addr =
8880 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
8881 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8882 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8883 _collector->_modUnionTable.mark_range(redirty_range);
8884 } else {
8885 _collector->_modUnionTable.mark(addr);
8886 }
8887 _collector->_ser_kac_preclean_ovflw++;
8888 } else {
8889 _collector->push_on_overflow_list(obj);
8890 _collector->_ser_kac_ovflw++;
8891 }
8892 }
8893 }
8894 }
8895
8896 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8897 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8898
8899 // CMSParKeepAliveClosure: a parallel version of the above.
8900 // The work queues are private to each closure (thread),
8901 // but (may be) available for stealing by other threads.
8902 void CMSParKeepAliveClosure::do_oop(oop obj) {
8903 HeapWord* addr = (HeapWord*)obj;
8904 if (_span.contains(addr) &&
8905 !_bit_map->isMarked(addr)) {
8906 // In general, during recursive tracing, several threads
8907 // may be concurrently getting here; the first one to
8908 // "tag" it, claims it.
8909 if (_bit_map->par_mark(addr)) {
8910 bool res = _work_queue->push(obj);
8911 assert(res, "Low water mark should be much less than capacity");
8912 // Do a recursive trim in the hope that this will keep
8913 // stack usage lower, but leave some oops for potential stealers
8914 trim_queue(_low_water_mark);
8915 } // Else, another thread got there first
8916 }
8917 }
8918
8919 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8920 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8921
8922 void CMSParKeepAliveClosure::trim_queue(uint max) {
8923 while (_work_queue->size() > max) {
8924 oop new_oop;
8925 if (_work_queue->pop_local(new_oop)) {
8926 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8927 assert(_bit_map->isMarked((HeapWord*)new_oop),
8928 "no white objects on this stack!");
8929 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8930 // iterate over the oops in this oop, marking and pushing
8931 // the ones in CMS heap (i.e. in _span).
8932 new_oop->oop_iterate(&_mark_and_push);
8933 }
8934 }
8935 }
8936
8937 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8938 CMSCollector* collector,
8939 MemRegion span, CMSBitMap* bit_map,
8940 OopTaskQueue* work_queue):
8941 _collector(collector),
8942 _span(span),
8943 _bit_map(bit_map),
8944 _work_queue(work_queue) { }
8945
8946 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8947 HeapWord* addr = (HeapWord*)obj;
8948 if (_span.contains(addr) &&
8949 !_bit_map->isMarked(addr)) {
8950 if (_bit_map->par_mark(addr)) {
8951 bool simulate_overflow = false;
8952 NOT_PRODUCT(
8953 if (CMSMarkStackOverflowALot &&
8954 _collector->par_simulate_overflow()) {
8955 // simulate a stack overflow
8956 simulate_overflow = true;
8957 }
8958 )
8959 if (simulate_overflow || !_work_queue->push(obj)) {
8960 _collector->par_push_on_overflow_list(obj);
8961 _collector->_par_kac_ovflw++;
8962 }
8963 } // Else another thread got there already
8964 }
8965 }
8966
8967 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8968 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8969
8970 //////////////////////////////////////////////////////////////////
8971 // CMSExpansionCause /////////////////////////////
8972 //////////////////////////////////////////////////////////////////
8973 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8974 switch (cause) {
8975 case _no_expansion:
8976 return "No expansion";
8977 case _satisfy_free_ratio:
8978 return "Free ratio";
8979 case _satisfy_promotion:
8980 return "Satisfy promotion";
8981 case _satisfy_allocation:
8982 return "allocation";
8983 case _allocate_par_lab:
8984 return "Par LAB";
8985 case _allocate_par_spooling_space:
8986 return "Par Spooling Space";
8987 case _adaptive_size_policy:
8988 return "Ergonomics";
8989 default:
8990 return "unknown";
8991 }
8992 }
8993
8994 void CMSDrainMarkingStackClosure::do_void() {
8995 // the max number to take from overflow list at a time
8996 const size_t num = _mark_stack->capacity()/4;
8997 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8998 "Overflow list should be NULL during concurrent phases");
8999 while (!_mark_stack->isEmpty() ||
9000 // if stack is empty, check the overflow list
9001 _collector->take_from_overflow_list(num, _mark_stack)) {
9002 oop obj = _mark_stack->pop();
9003 HeapWord* addr = (HeapWord*)obj;
9004 assert(_span.contains(addr), "Should be within span");
9005 assert(_bit_map->isMarked(addr), "Should be marked");
9006 assert(obj->is_oop(), "Should be an oop");
9007 obj->oop_iterate(_keep_alive);
9008 }
9009 }
9010
9011 void CMSParDrainMarkingStackClosure::do_void() {
9012 // drain queue
9013 trim_queue(0);
9014 }
9015
9016 // Trim our work_queue so its length is below max at return
9017 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
9018 while (_work_queue->size() > max) {
9019 oop new_oop;
9020 if (_work_queue->pop_local(new_oop)) {
9021 assert(new_oop->is_oop(), "Expected an oop");
9022 assert(_bit_map->isMarked((HeapWord*)new_oop),
9023 "no white objects on this stack!");
9024 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
9025 // iterate over the oops in this oop, marking and pushing
9026 // the ones in CMS heap (i.e. in _span).
9027 new_oop->oop_iterate(&_mark_and_push);
9028 }
9029 }
9030 }
9031
9032 ////////////////////////////////////////////////////////////////////
9033 // Support for Marking Stack Overflow list handling and related code
9034 ////////////////////////////////////////////////////////////////////
9035 // Much of the following code is similar in shape and spirit to the
9036 // code used in ParNewGC. We should try and share that code
9037 // as much as possible in the future.
9038
9039 #ifndef PRODUCT
9040 // Debugging support for CMSStackOverflowALot
9041
9042 // It's OK to call this multi-threaded; the worst thing
9043 // that can happen is that we'll get a bunch of closely
9044 // spaced simulated oveflows, but that's OK, in fact
9045 // probably good as it would exercise the overflow code
9046 // under contention.
9047 bool CMSCollector::simulate_overflow() {
9048 if (_overflow_counter-- <= 0) { // just being defensive
9049 _overflow_counter = CMSMarkStackOverflowInterval;
9050 return true;
9051 } else {
9052 return false;
9053 }
9054 }
9055
9056 bool CMSCollector::par_simulate_overflow() {
9057 return simulate_overflow();
9058 }
9059 #endif
9060
9061 // Single-threaded
9062 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
9063 assert(stack->isEmpty(), "Expected precondition");
9064 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
9065 size_t i = num;
9066 oop cur = _overflow_list;
9067 const markOop proto = markOopDesc::prototype();
9068 NOT_PRODUCT(ssize_t n = 0;)
9069 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
9070 next = oop(cur->mark());
9071 cur->set_mark(proto); // until proven otherwise
9072 assert(cur->is_oop(), "Should be an oop");
9073 bool res = stack->push(cur);
9074 assert(res, "Bit off more than can chew?");
9075 NOT_PRODUCT(n++;)
9076 }
9077 _overflow_list = cur;
9078 #ifndef PRODUCT
9079 assert(_num_par_pushes >= n, "Too many pops?");
9080 _num_par_pushes -=n;
9081 #endif
9082 return !stack->isEmpty();
9083 }
9084
9085 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
9086 // (MT-safe) Get a prefix of at most "num" from the list.
9087 // The overflow list is chained through the mark word of
9088 // each object in the list. We fetch the entire list,
9089 // break off a prefix of the right size and return the
9090 // remainder. If other threads try to take objects from
9091 // the overflow list at that time, they will wait for
9092 // some time to see if data becomes available. If (and
9093 // only if) another thread places one or more object(s)
9094 // on the global list before we have returned the suffix
9095 // to the global list, we will walk down our local list
9096 // to find its end and append the global list to
9097 // our suffix before returning it. This suffix walk can
9098 // prove to be expensive (quadratic in the amount of traffic)
9099 // when there are many objects in the overflow list and
9100 // there is much producer-consumer contention on the list.
9101 // *NOTE*: The overflow list manipulation code here and
9102 // in ParNewGeneration:: are very similar in shape,
9103 // except that in the ParNew case we use the old (from/eden)
9104 // copy of the object to thread the list via its klass word.
9105 // Because of the common code, if you make any changes in
9106 // the code below, please check the ParNew version to see if
9107 // similar changes might be needed.
9108 // CR 6797058 has been filed to consolidate the common code.
9109 bool CMSCollector::par_take_from_overflow_list(size_t num,
9110 OopTaskQueue* work_q,
9111 int no_of_gc_threads) {
9112 assert(work_q->size() == 0, "First empty local work queue");
9113 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
9114 if (_overflow_list == NULL) {
9115 return false;
9116 }
9117 // Grab the entire list; we'll put back a suffix
9118 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
9119 Thread* tid = Thread::current();
9120 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
9121 // set to ParallelGCThreads.
9122 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
9123 size_t sleep_time_millis = MAX2((size_t)1, num/100);
9124 // If the list is busy, we spin for a short while,
9125 // sleeping between attempts to get the list.
9126 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
9127 os::sleep(tid, sleep_time_millis, false);
9128 if (_overflow_list == NULL) {
9129 // Nothing left to take
9130 return false;
9131 } else if (_overflow_list != BUSY) {
9132 // Try and grab the prefix
9133 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
9134 }
9135 }
9136 // If the list was found to be empty, or we spun long
9137 // enough, we give up and return empty-handed. If we leave
9138 // the list in the BUSY state below, it must be the case that
9139 // some other thread holds the overflow list and will set it
9140 // to a non-BUSY state in the future.
9141 if (prefix == NULL || prefix == BUSY) {
9142 // Nothing to take or waited long enough
9143 if (prefix == NULL) {
9144 // Write back the NULL in case we overwrote it with BUSY above
9145 // and it is still the same value.
9146 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
9147 }
9148 return false;
9149 }
9150 assert(prefix != NULL && prefix != BUSY, "Error");
9151 size_t i = num;
9152 oop cur = prefix;
9153 // Walk down the first "num" objects, unless we reach the end.
9154 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
9155 if (cur->mark() == NULL) {
9156 // We have "num" or fewer elements in the list, so there
9157 // is nothing to return to the global list.
9158 // Write back the NULL in lieu of the BUSY we wrote
9159 // above, if it is still the same value.
9160 if (_overflow_list == BUSY) {
9161 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
9162 }
9163 } else {
9164 // Chop off the suffix and rerturn it to the global list.
9165 assert(cur->mark() != BUSY, "Error");
9166 oop suffix_head = cur->mark(); // suffix will be put back on global list
9167 cur->set_mark(NULL); // break off suffix
9168 // It's possible that the list is still in the empty(busy) state
9169 // we left it in a short while ago; in that case we may be
9170 // able to place back the suffix without incurring the cost
9171 // of a walk down the list.
9172 oop observed_overflow_list = _overflow_list;
9173 oop cur_overflow_list = observed_overflow_list;
9174 bool attached = false;
9175 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
9176 observed_overflow_list =
9177 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
9178 if (cur_overflow_list == observed_overflow_list) {
9179 attached = true;
9180 break;
9181 } else cur_overflow_list = observed_overflow_list;
9182 }
9183 if (!attached) {
9184 // Too bad, someone else sneaked in (at least) an element; we'll need
9185 // to do a splice. Find tail of suffix so we can prepend suffix to global
9186 // list.
9187 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
9188 oop suffix_tail = cur;
9189 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
9190 "Tautology");
9191 observed_overflow_list = _overflow_list;
9192 do {
9193 cur_overflow_list = observed_overflow_list;
9194 if (cur_overflow_list != BUSY) {
9195 // Do the splice ...
9196 suffix_tail->set_mark(markOop(cur_overflow_list));
9197 } else { // cur_overflow_list == BUSY
9198 suffix_tail->set_mark(NULL);
9199 }
9200 // ... and try to place spliced list back on overflow_list ...
9201 observed_overflow_list =
9202 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
9203 } while (cur_overflow_list != observed_overflow_list);
9204 // ... until we have succeeded in doing so.
9205 }
9206 }
9207
9208 // Push the prefix elements on work_q
9209 assert(prefix != NULL, "control point invariant");
9210 const markOop proto = markOopDesc::prototype();
9211 oop next;
9212 NOT_PRODUCT(ssize_t n = 0;)
9213 for (cur = prefix; cur != NULL; cur = next) {
9214 next = oop(cur->mark());
9215 cur->set_mark(proto); // until proven otherwise
9216 assert(cur->is_oop(), "Should be an oop");
9217 bool res = work_q->push(cur);
9218 assert(res, "Bit off more than we can chew?");
9219 NOT_PRODUCT(n++;)
9220 }
9221 #ifndef PRODUCT
9222 assert(_num_par_pushes >= n, "Too many pops?");
9223 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
9224 #endif
9225 return true;
9226 }
9227
9228 // Single-threaded
9229 void CMSCollector::push_on_overflow_list(oop p) {
9230 NOT_PRODUCT(_num_par_pushes++;)
9231 assert(p->is_oop(), "Not an oop");
9232 preserve_mark_if_necessary(p);
9233 p->set_mark((markOop)_overflow_list);
9234 _overflow_list = p;
9235 }
9236
9237 // Multi-threaded; use CAS to prepend to overflow list
9238 void CMSCollector::par_push_on_overflow_list(oop p) {
9239 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
9240 assert(p->is_oop(), "Not an oop");
9241 par_preserve_mark_if_necessary(p);
9242 oop observed_overflow_list = _overflow_list;
9243 oop cur_overflow_list;
9244 do {
9245 cur_overflow_list = observed_overflow_list;
9246 if (cur_overflow_list != BUSY) {
9247 p->set_mark(markOop(cur_overflow_list));
9248 } else {
9249 p->set_mark(NULL);
9250 }
9251 observed_overflow_list =
9252 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
9253 } while (cur_overflow_list != observed_overflow_list);
9254 }
9255 #undef BUSY
9256
9257 // Single threaded
9258 // General Note on GrowableArray: pushes may silently fail
9259 // because we are (temporarily) out of C-heap for expanding
9260 // the stack. The problem is quite ubiquitous and affects
9261 // a lot of code in the JVM. The prudent thing for GrowableArray
9262 // to do (for now) is to exit with an error. However, that may
9263 // be too draconian in some cases because the caller may be
9264 // able to recover without much harm. For such cases, we
9265 // should probably introduce a "soft_push" method which returns
9266 // an indication of success or failure with the assumption that
9267 // the caller may be able to recover from a failure; code in
9268 // the VM can then be changed, incrementally, to deal with such
9269 // failures where possible, thus, incrementally hardening the VM
9270 // in such low resource situations.
9271 void CMSCollector::preserve_mark_work(oop p, markOop m) {
9272 _preserved_oop_stack.push(p);
9273 _preserved_mark_stack.push(m);
9274 assert(m == p->mark(), "Mark word changed");
9275 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9276 "bijection");
9277 }
9278
9279 // Single threaded
9280 void CMSCollector::preserve_mark_if_necessary(oop p) {
9281 markOop m = p->mark();
9282 if (m->must_be_preserved(p)) {
9283 preserve_mark_work(p, m);
9284 }
9285 }
9286
9287 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
9288 markOop m = p->mark();
9289 if (m->must_be_preserved(p)) {
9290 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
9291 // Even though we read the mark word without holding
9292 // the lock, we are assured that it will not change
9293 // because we "own" this oop, so no other thread can
9294 // be trying to push it on the overflow list; see
9295 // the assertion in preserve_mark_work() that checks
9296 // that m == p->mark().
9297 preserve_mark_work(p, m);
9298 }
9299 }
9300
9301 // We should be able to do this multi-threaded,
9302 // a chunk of stack being a task (this is
9303 // correct because each oop only ever appears
9304 // once in the overflow list. However, it's
9305 // not very easy to completely overlap this with
9306 // other operations, so will generally not be done
9307 // until all work's been completed. Because we
9308 // expect the preserved oop stack (set) to be small,
9309 // it's probably fine to do this single-threaded.
9310 // We can explore cleverer concurrent/overlapped/parallel
9311 // processing of preserved marks if we feel the
9312 // need for this in the future. Stack overflow should
9313 // be so rare in practice and, when it happens, its
9314 // effect on performance so great that this will
9315 // likely just be in the noise anyway.
9316 void CMSCollector::restore_preserved_marks_if_any() {
9317 assert(SafepointSynchronize::is_at_safepoint(),
9318 "world should be stopped");
9319 assert(Thread::current()->is_ConcurrentGC_thread() ||
9320 Thread::current()->is_VM_thread(),
9321 "should be single-threaded");
9322 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9323 "bijection");
9324
9325 while (!_preserved_oop_stack.is_empty()) {
9326 oop p = _preserved_oop_stack.pop();
9327 assert(p->is_oop(), "Should be an oop");
9328 assert(_span.contains(p), "oop should be in _span");
9329 assert(p->mark() == markOopDesc::prototype(),
9330 "Set when taken from overflow list");
9331 markOop m = _preserved_mark_stack.pop();
9332 p->set_mark(m);
9333 }
9334 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
9335 "stacks were cleared above");
9336 }
9337
9338 #ifndef PRODUCT
9339 bool CMSCollector::no_preserved_marks() const {
9340 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
9341 }
9342 #endif
9343
9344 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
9345 {
9346 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
9347 CMSAdaptiveSizePolicy* size_policy =
9348 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
9349 assert(size_policy->is_gc_cms_adaptive_size_policy(),
9350 "Wrong type for size policy");
9351 return size_policy;
9352 }
9353
9354 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
9355 size_t desired_promo_size) {
9356 if (cur_promo_size < desired_promo_size) {
9357 size_t expand_bytes = desired_promo_size - cur_promo_size;
9358 if (PrintAdaptiveSizePolicy && Verbose) {
9359 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9360 "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
9361 expand_bytes);
9362 }
9363 expand(expand_bytes,
9364 MinHeapDeltaBytes,
9365 CMSExpansionCause::_adaptive_size_policy);
9366 } else if (desired_promo_size < cur_promo_size) {
9367 size_t shrink_bytes = cur_promo_size - desired_promo_size;
9368 if (PrintAdaptiveSizePolicy && Verbose) {
9369 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9370 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
9371 shrink_bytes);
9372 }
9373 shrink(shrink_bytes);
9374 }
9375 }
9376
9377 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
9378 GenCollectedHeap* gch = GenCollectedHeap::heap();
9379 CMSGCAdaptivePolicyCounters* counters =
9380 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
9381 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
9382 "Wrong kind of counters");
9383 return counters;
9384 }
9385
9386
9387 void ASConcurrentMarkSweepGeneration::update_counters() {
9388 if (UsePerfData) {
9389 _space_counters->update_all();
9390 _gen_counters->update_all();
9391 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9392 GenCollectedHeap* gch = GenCollectedHeap::heap();
9393 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9394 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9395 "Wrong gc statistics type");
9396 counters->update_counters(gc_stats_l);
9397 }
9398 }
9399
9400 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
9401 if (UsePerfData) {
9402 _space_counters->update_used(used);
9403 _space_counters->update_capacity();
9404 _gen_counters->update_all();
9405
9406 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9407 GenCollectedHeap* gch = GenCollectedHeap::heap();
9408 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9409 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9410 "Wrong gc statistics type");
9411 counters->update_counters(gc_stats_l);
9412 }
9413 }
9414
9415 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
9416 assert_locked_or_safepoint(Heap_lock);
9417 assert_lock_strong(freelistLock());
9418 HeapWord* old_end = _cmsSpace->end();
9419 HeapWord* unallocated_start = _cmsSpace->unallocated_block();
9420 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
9421 FreeChunk* chunk_at_end = find_chunk_at_end();
9422 if (chunk_at_end == NULL) {
9423 // No room to shrink
9424 if (PrintGCDetails && Verbose) {
9425 gclog_or_tty->print_cr("No room to shrink: old_end "
9426 PTR_FORMAT " unallocated_start " PTR_FORMAT
9427 " chunk_at_end " PTR_FORMAT,
9428 old_end, unallocated_start, chunk_at_end);
9429 }
9430 return;
9431 } else {
9432
9433 // Find the chunk at the end of the space and determine
9434 // how much it can be shrunk.
9435 size_t shrinkable_size_in_bytes = chunk_at_end->size();
9436 size_t aligned_shrinkable_size_in_bytes =
9437 align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
9438 assert(unallocated_start <= (HeapWord*) chunk_at_end->end(),
9439 "Inconsistent chunk at end of space");
9440 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
9441 size_t word_size_before = heap_word_size(_virtual_space.committed_size());
9442
9443 // Shrink the underlying space
9444 _virtual_space.shrink_by(bytes);
9445 if (PrintGCDetails && Verbose) {
9446 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
9447 " desired_bytes " SIZE_FORMAT
9448 " shrinkable_size_in_bytes " SIZE_FORMAT
9449 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
9450 " bytes " SIZE_FORMAT,
9451 desired_bytes, shrinkable_size_in_bytes,
9452 aligned_shrinkable_size_in_bytes, bytes);
9453 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT
9454 " unallocated_start " SIZE_FORMAT,
9455 old_end, unallocated_start);
9456 }
9457
9458 // If the space did shrink (shrinking is not guaranteed),
9459 // shrink the chunk at the end by the appropriate amount.
9460 if (((HeapWord*)_virtual_space.high()) < old_end) {
9461 size_t new_word_size =
9462 heap_word_size(_virtual_space.committed_size());
9463
9464 // Have to remove the chunk from the dictionary because it is changing
9465 // size and might be someplace elsewhere in the dictionary.
9466
9467 // Get the chunk at end, shrink it, and put it
9468 // back.
9469 _cmsSpace->removeChunkFromDictionary(chunk_at_end);
9470 size_t word_size_change = word_size_before - new_word_size;
9471 size_t chunk_at_end_old_size = chunk_at_end->size();
9472 assert(chunk_at_end_old_size >= word_size_change,
9473 "Shrink is too large");
9474 chunk_at_end->set_size(chunk_at_end_old_size -
9475 word_size_change);
9476 _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
9477 word_size_change);
9478
9479 _cmsSpace->returnChunkToDictionary(chunk_at_end);
9480
9481 MemRegion mr(_cmsSpace->bottom(), new_word_size);
9482 _bts->resize(new_word_size); // resize the block offset shared array
9483 Universe::heap()->barrier_set()->resize_covered_region(mr);
9484 _cmsSpace->assert_locked();
9485 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
9486
9487 NOT_PRODUCT(_cmsSpace->dictionary()->verify());
9488
9489 // update the space and generation capacity counters
9490 if (UsePerfData) {
9491 _space_counters->update_capacity();
9492 _gen_counters->update_all();
9493 }
9494
9495 if (Verbose && PrintGCDetails) {
9496 size_t new_mem_size = _virtual_space.committed_size();
9497 size_t old_mem_size = new_mem_size + bytes;
9498 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K",
9499 name(), old_mem_size/K, bytes/K, new_mem_size/K);
9500 }
9501 }
9502
9503 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
9504 "Inconsistency at end of space");
9505 assert(chunk_at_end->end() == (uintptr_t*) _cmsSpace->end(),
9506 "Shrinking is inconsistent");
9507 return;
9508 }
9509 }
9510 // Transfer some number of overflown objects to usual marking
9511 // stack. Return true if some objects were transferred.
9512 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
9513 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
9514 (size_t)ParGCDesiredObjsFromOverflowList);
9515
9516 bool res = _collector->take_from_overflow_list(num, _mark_stack);
9517 assert(_collector->overflow_list_is_empty() || res,
9518 "If list is not empty, we should have taken something");
9519 assert(!res || !_mark_stack->isEmpty(),
9520 "If we took something, it should now be on our stack");
9521 return res;
9522 }
9523
9524 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
9525 size_t res = _sp->block_size_no_stall(addr, _collector);
9526 if (_sp->block_is_obj(addr)) {
9527 if (_live_bit_map->isMarked(addr)) {
9528 // It can't have been dead in a previous cycle
9529 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
9530 } else {
9531 _dead_bit_map->mark(addr); // mark the dead object
9532 }
9533 }
9534 // Could be 0, if the block size could not be computed without stalling.
9535 return res;
9536 }
9537
9538 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
9539
9540 switch (phase) {
9541 case CMSCollector::InitialMarking:
9542 initialize(true /* fullGC */ ,
9543 cause /* cause of the GC */,
9544 true /* allMemoryPoolsAffected */,
9545 true /* recordGCBeginTime */,
9546 true /* recordPreGCUsage */,
9547 false /* recordPeakUsage */,
9548 false /* recordPostGCusage */,
9549 true /* recordAccumulatedGCTime */,
9550 false /* recordGCEndTime */,
9551 false /* countCollection */ );
9552 break;
9553
9554 case CMSCollector::FinalMarking:
9555 initialize(true /* fullGC */ ,
9556 cause /* cause of the GC */,
9557 true /* allMemoryPoolsAffected */,
9558 false /* recordGCBeginTime */,
9559 false /* recordPreGCUsage */,
9560 false /* recordPeakUsage */,
9561 false /* recordPostGCusage */,
9562 true /* recordAccumulatedGCTime */,
9563 false /* recordGCEndTime */,
9564 false /* countCollection */ );
9565 break;
9566
9567 case CMSCollector::Sweeping:
9568 initialize(true /* fullGC */ ,
9569 cause /* cause of the GC */,
9570 true /* allMemoryPoolsAffected */,
9571 false /* recordGCBeginTime */,
9572 false /* recordPreGCUsage */,
9573 true /* recordPeakUsage */,
9574 true /* recordPostGCusage */,
9575 false /* recordAccumulatedGCTime */,
9576 true /* recordGCEndTime */,
9577 true /* countCollection */ );
9578 break;
9579
9580 default:
9581 ShouldNotReachHere();
9582 }
9583 }