1 /*
2 * Copyright (c) 2001, 2015, 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/stringTable.hpp"
28 #include "classfile/symbolTable.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/codeCache.hpp"
31 #include "gc/cms/cmsCollectorPolicy.hpp"
32 #include "gc/cms/cmsOopClosures.inline.hpp"
33 #include "gc/cms/compactibleFreeListSpace.hpp"
34 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp"
35 #include "gc/cms/concurrentMarkSweepThread.hpp"
36 #include "gc/cms/parNewGeneration.hpp"
37 #include "gc/cms/vmCMSOperations.hpp"
38 #include "gc/serial/genMarkSweep.hpp"
39 #include "gc/serial/tenuredGeneration.hpp"
40 #include "gc/shared/adaptiveSizePolicy.hpp"
41 #include "gc/shared/cardGeneration.inline.hpp"
42 #include "gc/shared/cardTableRS.hpp"
43 #include "gc/shared/collectedHeap.inline.hpp"
44 #include "gc/shared/collectorCounters.hpp"
45 #include "gc/shared/collectorPolicy.hpp"
46 #include "gc/shared/gcLocker.inline.hpp"
47 #include "gc/shared/gcPolicyCounters.hpp"
48 #include "gc/shared/gcTimer.hpp"
49 #include "gc/shared/gcTrace.hpp"
50 #include "gc/shared/gcTraceTime.hpp"
51 #include "gc/shared/genCollectedHeap.hpp"
52 #include "gc/shared/genOopClosures.inline.hpp"
53 #include "gc/shared/isGCActiveMark.hpp"
54 #include "gc/shared/referencePolicy.hpp"
55 #include "gc/shared/strongRootsScope.hpp"
56 #include "gc/shared/taskqueue.inline.hpp"
57 #include "memory/allocation.hpp"
58 #include "memory/iterator.inline.hpp"
59 #include "memory/padded.hpp"
60 #include "memory/resourceArea.hpp"
61 #include "oops/oop.inline.hpp"
62 #include "prims/jvmtiExport.hpp"
63 #include "runtime/atomic.inline.hpp"
64 #include "runtime/globals_extension.hpp"
65 #include "runtime/handles.inline.hpp"
66 #include "runtime/java.hpp"
67 #include "runtime/orderAccess.inline.hpp"
68 #include "runtime/vmThread.hpp"
69 #include "services/memoryService.hpp"
70 #include "services/runtimeService.hpp"
71 #include "utilities/stack.inline.hpp"
72
73 // statics
74 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
75 bool CMSCollector::_full_gc_requested = false;
76 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
77
78 //////////////////////////////////////////////////////////////////
79 // In support of CMS/VM thread synchronization
80 //////////////////////////////////////////////////////////////////
81 // We split use of the CGC_lock into 2 "levels".
82 // The low-level locking is of the usual CGC_lock monitor. We introduce
83 // a higher level "token" (hereafter "CMS token") built on top of the
84 // low level monitor (hereafter "CGC lock").
85 // The token-passing protocol gives priority to the VM thread. The
86 // CMS-lock doesn't provide any fairness guarantees, but clients
87 // should ensure that it is only held for very short, bounded
88 // durations.
89 //
90 // When either of the CMS thread or the VM thread is involved in
91 // collection operations during which it does not want the other
92 // thread to interfere, it obtains the CMS token.
93 //
94 // If either thread tries to get the token while the other has
95 // it, that thread waits. However, if the VM thread and CMS thread
96 // both want the token, then the VM thread gets priority while the
97 // CMS thread waits. This ensures, for instance, that the "concurrent"
98 // phases of the CMS thread's work do not block out the VM thread
99 // for long periods of time as the CMS thread continues to hog
100 // the token. (See bug 4616232).
101 //
102 // The baton-passing functions are, however, controlled by the
103 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
104 // and here the low-level CMS lock, not the high level token,
105 // ensures mutual exclusion.
106 //
107 // Two important conditions that we have to satisfy:
108 // 1. if a thread does a low-level wait on the CMS lock, then it
109 // relinquishes the CMS token if it were holding that token
110 // when it acquired the low-level CMS lock.
111 // 2. any low-level notifications on the low-level lock
112 // should only be sent when a thread has relinquished the token.
113 //
114 // In the absence of either property, we'd have potential deadlock.
115 //
116 // We protect each of the CMS (concurrent and sequential) phases
117 // with the CMS _token_, not the CMS _lock_.
118 //
119 // The only code protected by CMS lock is the token acquisition code
120 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
121 // baton-passing code.
122 //
123 // Unfortunately, i couldn't come up with a good abstraction to factor and
124 // hide the naked CGC_lock manipulation in the baton-passing code
125 // further below. That's something we should try to do. Also, the proof
126 // of correctness of this 2-level locking scheme is far from obvious,
127 // and potentially quite slippery. We have an uneasy suspicion, for instance,
128 // that there may be a theoretical possibility of delay/starvation in the
129 // low-level lock/wait/notify scheme used for the baton-passing because of
130 // potential interference with the priority scheme embodied in the
131 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
132 // invocation further below and marked with "XXX 20011219YSR".
133 // Indeed, as we note elsewhere, this may become yet more slippery
134 // in the presence of multiple CMS and/or multiple VM threads. XXX
135
136 class CMSTokenSync: public StackObj {
137 private:
138 bool _is_cms_thread;
139 public:
140 CMSTokenSync(bool is_cms_thread):
141 _is_cms_thread(is_cms_thread) {
142 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
143 "Incorrect argument to constructor");
144 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
145 }
146
147 ~CMSTokenSync() {
148 assert(_is_cms_thread ?
149 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
150 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
151 "Incorrect state");
152 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
153 }
154 };
155
156 // Convenience class that does a CMSTokenSync, and then acquires
157 // upto three locks.
158 class CMSTokenSyncWithLocks: public CMSTokenSync {
159 private:
160 // Note: locks are acquired in textual declaration order
161 // and released in the opposite order
162 MutexLockerEx _locker1, _locker2, _locker3;
163 public:
164 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
165 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
166 CMSTokenSync(is_cms_thread),
167 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
168 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
169 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
170 { }
171 };
172
173
174 //////////////////////////////////////////////////////////////////
175 // Concurrent Mark-Sweep Generation /////////////////////////////
176 //////////////////////////////////////////////////////////////////
177
178 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
179
180 // This struct contains per-thread things necessary to support parallel
181 // young-gen collection.
182 class CMSParGCThreadState: public CHeapObj<mtGC> {
183 public:
184 CFLS_LAB lab;
185 PromotionInfo promo;
186
187 // Constructor.
188 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
189 promo.setSpace(cfls);
190 }
191 };
192
193 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
194 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) :
195 CardGeneration(rs, initial_byte_size, ct),
196 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
197 _did_compact(false)
198 {
199 HeapWord* bottom = (HeapWord*) _virtual_space.low();
200 HeapWord* end = (HeapWord*) _virtual_space.high();
201
202 _direct_allocated_words = 0;
203 NOT_PRODUCT(
204 _numObjectsPromoted = 0;
205 _numWordsPromoted = 0;
206 _numObjectsAllocated = 0;
207 _numWordsAllocated = 0;
208 )
209
210 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end));
211 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
212 _cmsSpace->_old_gen = this;
213
214 _gc_stats = new CMSGCStats();
215
216 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
217 // offsets match. The ability to tell free chunks from objects
218 // depends on this property.
219 debug_only(
220 FreeChunk* junk = NULL;
221 assert(UseCompressedClassPointers ||
222 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
223 "Offset of FreeChunk::_prev within FreeChunk must match"
224 " that of OopDesc::_klass within OopDesc");
225 )
226
227 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC);
228 for (uint i = 0; i < ParallelGCThreads; i++) {
229 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
230 }
231
232 _incremental_collection_failed = false;
233 // The "dilatation_factor" is the expansion that can occur on
234 // account of the fact that the minimum object size in the CMS
235 // generation may be larger than that in, say, a contiguous young
236 // generation.
237 // Ideally, in the calculation below, we'd compute the dilatation
238 // factor as: MinChunkSize/(promoting_gen's min object size)
239 // Since we do not have such a general query interface for the
240 // promoting generation, we'll instead just use the minimum
241 // object size (which today is a header's worth of space);
242 // note that all arithmetic is in units of HeapWords.
243 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
244 assert(_dilatation_factor >= 1.0, "from previous assert");
245 }
246
247
248 // The field "_initiating_occupancy" represents the occupancy percentage
249 // at which we trigger a new collection cycle. Unless explicitly specified
250 // via CMSInitiatingOccupancyFraction (argument "io" below), it
251 // is calculated by:
252 //
253 // Let "f" be MinHeapFreeRatio in
254 //
255 // _initiating_occupancy = 100-f +
256 // f * (CMSTriggerRatio/100)
257 // where CMSTriggerRatio is the argument "tr" below.
258 //
259 // That is, if we assume the heap is at its desired maximum occupancy at the
260 // end of a collection, we let CMSTriggerRatio of the (purported) free
261 // space be allocated before initiating a new collection cycle.
262 //
263 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
264 assert(io <= 100 && tr <= 100, "Check the arguments");
265 if (io >= 0) {
266 _initiating_occupancy = (double)io / 100.0;
267 } else {
268 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
269 (double)(tr * MinHeapFreeRatio) / 100.0)
270 / 100.0;
271 }
272 }
273
274 void ConcurrentMarkSweepGeneration::ref_processor_init() {
275 assert(collector() != NULL, "no collector");
276 collector()->ref_processor_init();
277 }
278
279 void CMSCollector::ref_processor_init() {
280 if (_ref_processor == NULL) {
281 // Allocate and initialize a reference processor
282 _ref_processor =
283 new ReferenceProcessor(_span, // span
284 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
285 ParallelGCThreads, // mt processing degree
286 _cmsGen->refs_discovery_is_mt(), // mt discovery
287 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
288 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
289 &_is_alive_closure); // closure for liveness info
290 // Initialize the _ref_processor field of CMSGen
291 _cmsGen->set_ref_processor(_ref_processor);
292
293 }
294 }
295
296 AdaptiveSizePolicy* CMSCollector::size_policy() {
297 GenCollectedHeap* gch = GenCollectedHeap::heap();
298 return gch->gen_policy()->size_policy();
299 }
300
301 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
302
303 const char* gen_name = "old";
304 GenCollectorPolicy* gcp = GenCollectedHeap::heap()->gen_policy();
305 // Generation Counters - generation 1, 1 subspace
306 _gen_counters = new GenerationCounters(gen_name, 1, 1,
307 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
308
309 _space_counters = new GSpaceCounters(gen_name, 0,
310 _virtual_space.reserved_size(),
311 this, _gen_counters);
312 }
313
314 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
315 _cms_gen(cms_gen)
316 {
317 assert(alpha <= 100, "bad value");
318 _saved_alpha = alpha;
319
320 // Initialize the alphas to the bootstrap value of 100.
321 _gc0_alpha = _cms_alpha = 100;
322
323 _cms_begin_time.update();
324 _cms_end_time.update();
325
326 _gc0_duration = 0.0;
327 _gc0_period = 0.0;
328 _gc0_promoted = 0;
329
330 _cms_duration = 0.0;
331 _cms_period = 0.0;
332 _cms_allocated = 0;
333
334 _cms_used_at_gc0_begin = 0;
335 _cms_used_at_gc0_end = 0;
336 _allow_duty_cycle_reduction = false;
337 _valid_bits = 0;
338 }
339
340 double CMSStats::cms_free_adjustment_factor(size_t free) const {
341 // TBD: CR 6909490
342 return 1.0;
343 }
344
345 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
346 }
347
348 // If promotion failure handling is on use
349 // the padded average size of the promotion for each
350 // young generation collection.
351 double CMSStats::time_until_cms_gen_full() const {
352 size_t cms_free = _cms_gen->cmsSpace()->free();
353 GenCollectedHeap* gch = GenCollectedHeap::heap();
354 size_t expected_promotion = MIN2(gch->young_gen()->capacity(),
355 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
356 if (cms_free > expected_promotion) {
357 // Start a cms collection if there isn't enough space to promote
358 // for the next young collection. Use the padded average as
359 // a safety factor.
360 cms_free -= expected_promotion;
361
362 // Adjust by the safety factor.
363 double cms_free_dbl = (double)cms_free;
364 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0;
365 // Apply a further correction factor which tries to adjust
366 // for recent occurance of concurrent mode failures.
367 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
368 cms_free_dbl = cms_free_dbl * cms_adjustment;
369
370 if (PrintGCDetails && Verbose) {
371 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
372 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
373 cms_free, expected_promotion);
374 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
375 cms_free_dbl, cms_consumption_rate() + 1.0);
376 }
377 // Add 1 in case the consumption rate goes to zero.
378 return cms_free_dbl / (cms_consumption_rate() + 1.0);
379 }
380 return 0.0;
381 }
382
383 // Compare the duration of the cms collection to the
384 // time remaining before the cms generation is empty.
385 // Note that the time from the start of the cms collection
386 // to the start of the cms sweep (less than the total
387 // duration of the cms collection) can be used. This
388 // has been tried and some applications experienced
389 // promotion failures early in execution. This was
390 // possibly because the averages were not accurate
391 // enough at the beginning.
392 double CMSStats::time_until_cms_start() const {
393 // We add "gc0_period" to the "work" calculation
394 // below because this query is done (mostly) at the
395 // end of a scavenge, so we need to conservatively
396 // account for that much possible delay
397 // in the query so as to avoid concurrent mode failures
398 // due to starting the collection just a wee bit too
399 // late.
400 double work = cms_duration() + gc0_period();
401 double deadline = time_until_cms_gen_full();
402 // If a concurrent mode failure occurred recently, we want to be
403 // more conservative and halve our expected time_until_cms_gen_full()
404 if (work > deadline) {
405 if (Verbose && PrintGCDetails) {
406 gclog_or_tty->print(
407 " CMSCollector: collect because of anticipated promotion "
408 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
409 gc0_period(), time_until_cms_gen_full());
410 }
411 return 0.0;
412 }
413 return work - deadline;
414 }
415
416 #ifndef PRODUCT
417 void CMSStats::print_on(outputStream *st) const {
418 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
419 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
420 gc0_duration(), gc0_period(), gc0_promoted());
421 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
422 cms_duration(), cms_period(), cms_allocated());
423 st->print(",cms_since_beg=%g,cms_since_end=%g",
424 cms_time_since_begin(), cms_time_since_end());
425 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
426 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
427
428 if (valid()) {
429 st->print(",promo_rate=%g,cms_alloc_rate=%g",
430 promotion_rate(), cms_allocation_rate());
431 st->print(",cms_consumption_rate=%g,time_until_full=%g",
432 cms_consumption_rate(), time_until_cms_gen_full());
433 }
434 st->print(" ");
435 }
436 #endif // #ifndef PRODUCT
437
438 CMSCollector::CollectorState CMSCollector::_collectorState =
439 CMSCollector::Idling;
440 bool CMSCollector::_foregroundGCIsActive = false;
441 bool CMSCollector::_foregroundGCShouldWait = false;
442
443 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
444 CardTableRS* ct,
445 ConcurrentMarkSweepPolicy* cp):
446 _cmsGen(cmsGen),
447 _ct(ct),
448 _ref_processor(NULL), // will be set later
449 _conc_workers(NULL), // may be set later
450 _abort_preclean(false),
451 _start_sampling(false),
452 _between_prologue_and_epilogue(false),
453 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
454 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
455 -1 /* lock-free */, "No_lock" /* dummy */),
456 _modUnionClosurePar(&_modUnionTable),
457 // Adjust my span to cover old (cms) gen
458 _span(cmsGen->reserved()),
459 // Construct the is_alive_closure with _span & markBitMap
460 _is_alive_closure(_span, &_markBitMap),
461 _restart_addr(NULL),
462 _overflow_list(NULL),
463 _stats(cmsGen),
464 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
465 //verify that this lock should be acquired with safepoint check.
466 Monitor::_safepoint_check_sometimes)),
467 _eden_chunk_array(NULL), // may be set in ctor body
468 _eden_chunk_capacity(0), // -- ditto --
469 _eden_chunk_index(0), // -- ditto --
470 _survivor_plab_array(NULL), // -- ditto --
471 _survivor_chunk_array(NULL), // -- ditto --
472 _survivor_chunk_capacity(0), // -- ditto --
473 _survivor_chunk_index(0), // -- ditto --
474 _ser_pmc_preclean_ovflw(0),
475 _ser_kac_preclean_ovflw(0),
476 _ser_pmc_remark_ovflw(0),
477 _par_pmc_remark_ovflw(0),
478 _ser_kac_ovflw(0),
479 _par_kac_ovflw(0),
480 #ifndef PRODUCT
481 _num_par_pushes(0),
482 #endif
483 _collection_count_start(0),
484 _verifying(false),
485 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
486 _completed_initialization(false),
487 _collector_policy(cp),
488 _should_unload_classes(CMSClassUnloadingEnabled),
489 _concurrent_cycles_since_last_unload(0),
490 _roots_scanning_options(GenCollectedHeap::SO_None),
491 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
492 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
493 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
494 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
495 _cms_start_registered(false)
496 {
497 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
498 ExplicitGCInvokesConcurrent = true;
499 }
500 // Now expand the span and allocate the collection support structures
501 // (MUT, marking bit map etc.) to cover both generations subject to
502 // collection.
503
504 // For use by dirty card to oop closures.
505 _cmsGen->cmsSpace()->set_collector(this);
506
507 // Allocate MUT and marking bit map
508 {
509 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
510 if (!_markBitMap.allocate(_span)) {
511 warning("Failed to allocate CMS Bit Map");
512 return;
513 }
514 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
515 }
516 {
517 _modUnionTable.allocate(_span);
518 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
519 }
520
521 if (!_markStack.allocate(MarkStackSize)) {
522 warning("Failed to allocate CMS Marking Stack");
523 return;
524 }
525
526 // Support for multi-threaded concurrent phases
527 if (CMSConcurrentMTEnabled) {
528 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
529 // just for now
530 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4);
531 }
532 if (ConcGCThreads > 1) {
533 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
534 ConcGCThreads, true);
535 if (_conc_workers == NULL) {
536 warning("GC/CMS: _conc_workers allocation failure: "
537 "forcing -CMSConcurrentMTEnabled");
538 CMSConcurrentMTEnabled = false;
539 } else {
540 _conc_workers->initialize_workers();
541 }
542 } else {
543 CMSConcurrentMTEnabled = false;
544 }
545 }
546 if (!CMSConcurrentMTEnabled) {
547 ConcGCThreads = 0;
548 } else {
549 // Turn off CMSCleanOnEnter optimization temporarily for
550 // the MT case where it's not fixed yet; see 6178663.
551 CMSCleanOnEnter = false;
552 }
553 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
554 "Inconsistency");
555
556 // Parallel task queues; these are shared for the
557 // concurrent and stop-world phases of CMS, but
558 // are not shared with parallel scavenge (ParNew).
559 {
560 uint i;
561 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads);
562
563 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
564 || ParallelRefProcEnabled)
565 && num_queues > 0) {
566 _task_queues = new OopTaskQueueSet(num_queues);
567 if (_task_queues == NULL) {
568 warning("task_queues allocation failure.");
569 return;
570 }
571 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
572 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
573 for (i = 0; i < num_queues; i++) {
574 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
575 if (q == NULL) {
576 warning("work_queue allocation failure.");
577 return;
578 }
579 _task_queues->register_queue(i, q);
580 }
581 for (i = 0; i < num_queues; i++) {
582 _task_queues->queue(i)->initialize();
583 _hash_seed[i] = 17; // copied from ParNew
584 }
585 }
586 }
587
588 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
589
590 // Clip CMSBootstrapOccupancy between 0 and 100.
591 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0;
592
593 // Now tell CMS generations the identity of their collector
594 ConcurrentMarkSweepGeneration::set_collector(this);
595
596 // Create & start a CMS thread for this CMS collector
597 _cmsThread = ConcurrentMarkSweepThread::start(this);
598 assert(cmsThread() != NULL, "CMS Thread should have been created");
599 assert(cmsThread()->collector() == this,
600 "CMS Thread should refer to this gen");
601 assert(CGC_lock != NULL, "Where's the CGC_lock?");
602
603 // Support for parallelizing young gen rescan
604 GenCollectedHeap* gch = GenCollectedHeap::heap();
605 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
606 _young_gen = (ParNewGeneration*)gch->young_gen();
607 if (gch->supports_inline_contig_alloc()) {
608 _top_addr = gch->top_addr();
609 _end_addr = gch->end_addr();
610 assert(_young_gen != NULL, "no _young_gen");
611 _eden_chunk_index = 0;
612 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain;
613 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
614 }
615
616 // Support for parallelizing survivor space rescan
617 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
618 const size_t max_plab_samples =
619 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize);
620
621 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
622 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
623 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
624 _survivor_chunk_capacity = max_plab_samples;
625 for (uint i = 0; i < ParallelGCThreads; i++) {
626 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
627 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
628 assert(cur->end() == 0, "Should be 0");
629 assert(cur->array() == vec, "Should be vec");
630 assert(cur->capacity() == max_plab_samples, "Error");
631 }
632 }
633
634 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
635 _gc_counters = new CollectorCounters("CMS", 1);
636 _completed_initialization = true;
637 _inter_sweep_timer.start(); // start of time
638 }
639
640 const char* ConcurrentMarkSweepGeneration::name() const {
641 return "concurrent mark-sweep generation";
642 }
643 void ConcurrentMarkSweepGeneration::update_counters() {
644 if (UsePerfData) {
645 _space_counters->update_all();
646 _gen_counters->update_all();
647 }
648 }
649
650 // this is an optimized version of update_counters(). it takes the
651 // used value as a parameter rather than computing it.
652 //
653 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
654 if (UsePerfData) {
655 _space_counters->update_used(used);
656 _space_counters->update_capacity();
657 _gen_counters->update_all();
658 }
659 }
660
661 void ConcurrentMarkSweepGeneration::print() const {
662 Generation::print();
663 cmsSpace()->print();
664 }
665
666 #ifndef PRODUCT
667 void ConcurrentMarkSweepGeneration::print_statistics() {
668 cmsSpace()->printFLCensus(0);
669 }
670 #endif
671
672 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
673 GenCollectedHeap* gch = GenCollectedHeap::heap();
674 if (PrintGCDetails) {
675 // I didn't want to change the logging when removing the level concept,
676 // but I guess this logging could say "old" or something instead of "1".
677 assert(gch->is_old_gen(this),
678 "The CMS generation should be the old generation");
679 uint level = 1;
680 if (Verbose) {
681 gclog_or_tty->print("[%u %s-%s: " SIZE_FORMAT "(" SIZE_FORMAT ")]",
682 level, short_name(), s, used(), capacity());
683 } else {
684 gclog_or_tty->print("[%u %s-%s: " SIZE_FORMAT "K(" SIZE_FORMAT "K)]",
685 level, short_name(), s, used() / K, capacity() / K);
686 }
687 }
688 if (Verbose) {
689 gclog_or_tty->print(" " SIZE_FORMAT "(" SIZE_FORMAT ")",
690 gch->used(), gch->capacity());
691 } else {
692 gclog_or_tty->print(" " SIZE_FORMAT "K(" SIZE_FORMAT "K)",
693 gch->used() / K, gch->capacity() / K);
694 }
695 }
696
697 size_t
698 ConcurrentMarkSweepGeneration::contiguous_available() const {
699 // dld proposes an improvement in precision here. If the committed
700 // part of the space ends in a free block we should add that to
701 // uncommitted size in the calculation below. Will make this
702 // change later, staying with the approximation below for the
703 // time being. -- ysr.
704 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
705 }
706
707 size_t
708 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
709 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
710 }
711
712 size_t ConcurrentMarkSweepGeneration::max_available() const {
713 return free() + _virtual_space.uncommitted_size();
714 }
715
716 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
717 size_t available = max_available();
718 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
719 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
720 if (Verbose && PrintGCDetails) {
721 gclog_or_tty->print_cr(
722 "CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "),"
723 "max_promo(" SIZE_FORMAT ")",
724 res? "":" not", available, res? ">=":"<",
725 av_promo, max_promotion_in_bytes);
726 }
727 return res;
728 }
729
730 // At a promotion failure dump information on block layout in heap
731 // (cms old generation).
732 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
733 if (CMSDumpAtPromotionFailure) {
734 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
735 }
736 }
737
738 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
739 // Clear the promotion information. These pointers can be adjusted
740 // along with all the other pointers into the heap but
741 // compaction is expected to be a rare event with
742 // a heap using cms so don't do it without seeing the need.
743 for (uint i = 0; i < ParallelGCThreads; i++) {
744 _par_gc_thread_states[i]->promo.reset();
745 }
746 }
747
748 void ConcurrentMarkSweepGeneration::compute_new_size() {
749 assert_locked_or_safepoint(Heap_lock);
750
751 // If incremental collection failed, we just want to expand
752 // to the limit.
753 if (incremental_collection_failed()) {
754 clear_incremental_collection_failed();
755 grow_to_reserved();
756 return;
757 }
758
759 // The heap has been compacted but not reset yet.
760 // Any metric such as free() or used() will be incorrect.
761
762 CardGeneration::compute_new_size();
763
764 // Reset again after a possible resizing
765 if (did_compact()) {
766 cmsSpace()->reset_after_compaction();
767 }
768 }
769
770 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
771 assert_locked_or_safepoint(Heap_lock);
772
773 // If incremental collection failed, we just want to expand
774 // to the limit.
775 if (incremental_collection_failed()) {
776 clear_incremental_collection_failed();
777 grow_to_reserved();
778 return;
779 }
780
781 double free_percentage = ((double) free()) / capacity();
782 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
783 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
784
785 // compute expansion delta needed for reaching desired free percentage
786 if (free_percentage < desired_free_percentage) {
787 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
788 assert(desired_capacity >= capacity(), "invalid expansion size");
789 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
790 if (PrintGCDetails && Verbose) {
791 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
792 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
793 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
794 gclog_or_tty->print_cr(" Desired free fraction %f", desired_free_percentage);
795 gclog_or_tty->print_cr(" Maximum free fraction %f", maximum_free_percentage);
796 gclog_or_tty->print_cr(" Capacity " SIZE_FORMAT, capacity() / 1000);
797 gclog_or_tty->print_cr(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000);
798 GenCollectedHeap* gch = GenCollectedHeap::heap();
799 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation");
800 size_t young_size = gch->young_gen()->capacity();
801 gclog_or_tty->print_cr(" Young gen size " SIZE_FORMAT, young_size / 1000);
802 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000);
803 gclog_or_tty->print_cr(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000);
804 gclog_or_tty->print_cr(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes);
805 }
806 // safe if expansion fails
807 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
808 if (PrintGCDetails && Verbose) {
809 gclog_or_tty->print_cr(" Expanded free fraction %f", ((double) free()) / capacity());
810 }
811 } else {
812 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
813 assert(desired_capacity <= capacity(), "invalid expansion size");
814 size_t shrink_bytes = capacity() - desired_capacity;
815 // Don't shrink unless the delta is greater than the minimum shrink we want
816 if (shrink_bytes >= MinHeapDeltaBytes) {
817 shrink_free_list_by(shrink_bytes);
818 }
819 }
820 }
821
822 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
823 return cmsSpace()->freelistLock();
824 }
825
826 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) {
827 CMSSynchronousYieldRequest yr;
828 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
829 return have_lock_and_allocate(size, tlab);
830 }
831
832 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
833 bool tlab /* ignored */) {
834 assert_lock_strong(freelistLock());
835 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
836 HeapWord* res = cmsSpace()->allocate(adjustedSize);
837 // Allocate the object live (grey) if the background collector has
838 // started marking. This is necessary because the marker may
839 // have passed this address and consequently this object will
840 // not otherwise be greyed and would be incorrectly swept up.
841 // Note that if this object contains references, the writing
842 // of those references will dirty the card containing this object
843 // allowing the object to be blackened (and its references scanned)
844 // either during a preclean phase or at the final checkpoint.
845 if (res != NULL) {
846 // We may block here with an uninitialized object with
847 // its mark-bit or P-bits not yet set. Such objects need
848 // to be safely navigable by block_start().
849 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
850 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
851 collector()->direct_allocated(res, adjustedSize);
852 _direct_allocated_words += adjustedSize;
853 // allocation counters
854 NOT_PRODUCT(
855 _numObjectsAllocated++;
856 _numWordsAllocated += (int)adjustedSize;
857 )
858 }
859 return res;
860 }
861
862 // In the case of direct allocation by mutators in a generation that
863 // is being concurrently collected, the object must be allocated
864 // live (grey) if the background collector has started marking.
865 // This is necessary because the marker may
866 // have passed this address and consequently this object will
867 // not otherwise be greyed and would be incorrectly swept up.
868 // Note that if this object contains references, the writing
869 // of those references will dirty the card containing this object
870 // allowing the object to be blackened (and its references scanned)
871 // either during a preclean phase or at the final checkpoint.
872 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
873 assert(_markBitMap.covers(start, size), "Out of bounds");
874 if (_collectorState >= Marking) {
875 MutexLockerEx y(_markBitMap.lock(),
876 Mutex::_no_safepoint_check_flag);
877 // [see comments preceding SweepClosure::do_blk() below for details]
878 //
879 // Can the P-bits be deleted now? JJJ
880 //
881 // 1. need to mark the object as live so it isn't collected
882 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
883 // 3. need to mark the end of the object so marking, precleaning or sweeping
884 // can skip over uninitialized or unparsable objects. An allocated
885 // object is considered uninitialized for our purposes as long as
886 // its klass word is NULL. All old gen objects are parsable
887 // as soon as they are initialized.)
888 _markBitMap.mark(start); // object is live
889 _markBitMap.mark(start + 1); // object is potentially uninitialized?
890 _markBitMap.mark(start + size - 1);
891 // mark end of object
892 }
893 // check that oop looks uninitialized
894 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
895 }
896
897 void CMSCollector::promoted(bool par, HeapWord* start,
898 bool is_obj_array, size_t obj_size) {
899 assert(_markBitMap.covers(start), "Out of bounds");
900 // See comment in direct_allocated() about when objects should
901 // be allocated live.
902 if (_collectorState >= Marking) {
903 // we already hold the marking bit map lock, taken in
904 // the prologue
905 if (par) {
906 _markBitMap.par_mark(start);
907 } else {
908 _markBitMap.mark(start);
909 }
910 // We don't need to mark the object as uninitialized (as
911 // in direct_allocated above) because this is being done with the
912 // world stopped and the object will be initialized by the
913 // time the marking, precleaning or sweeping get to look at it.
914 // But see the code for copying objects into the CMS generation,
915 // where we need to ensure that concurrent readers of the
916 // block offset table are able to safely navigate a block that
917 // is in flux from being free to being allocated (and in
918 // transition while being copied into) and subsequently
919 // becoming a bona-fide object when the copy/promotion is complete.
920 assert(SafepointSynchronize::is_at_safepoint(),
921 "expect promotion only at safepoints");
922
923 if (_collectorState < Sweeping) {
924 // Mark the appropriate cards in the modUnionTable, so that
925 // this object gets scanned before the sweep. If this is
926 // not done, CMS generation references in the object might
927 // not get marked.
928 // For the case of arrays, which are otherwise precisely
929 // marked, we need to dirty the entire array, not just its head.
930 if (is_obj_array) {
931 // The [par_]mark_range() method expects mr.end() below to
932 // be aligned to the granularity of a bit's representation
933 // in the heap. In the case of the MUT below, that's a
934 // card size.
935 MemRegion mr(start,
936 (HeapWord*)round_to((intptr_t)(start + obj_size),
937 CardTableModRefBS::card_size /* bytes */));
938 if (par) {
939 _modUnionTable.par_mark_range(mr);
940 } else {
941 _modUnionTable.mark_range(mr);
942 }
943 } else { // not an obj array; we can just mark the head
944 if (par) {
945 _modUnionTable.par_mark(start);
946 } else {
947 _modUnionTable.mark(start);
948 }
949 }
950 }
951 }
952 }
953
954 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
955 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
956 // allocate, copy and if necessary update promoinfo --
957 // delegate to underlying space.
958 assert_lock_strong(freelistLock());
959
960 #ifndef PRODUCT
961 if (GenCollectedHeap::heap()->promotion_should_fail()) {
962 return NULL;
963 }
964 #endif // #ifndef PRODUCT
965
966 oop res = _cmsSpace->promote(obj, obj_size);
967 if (res == NULL) {
968 // expand and retry
969 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
970 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
971 // Since this is the old generation, we don't try to promote
972 // into a more senior generation.
973 res = _cmsSpace->promote(obj, obj_size);
974 }
975 if (res != NULL) {
976 // See comment in allocate() about when objects should
977 // be allocated live.
978 assert(obj->is_oop(), "Will dereference klass pointer below");
979 collector()->promoted(false, // Not parallel
980 (HeapWord*)res, obj->is_objArray(), obj_size);
981 // promotion counters
982 NOT_PRODUCT(
983 _numObjectsPromoted++;
984 _numWordsPromoted +=
985 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
986 )
987 }
988 return res;
989 }
990
991
992 // IMPORTANT: Notes on object size recognition in CMS.
993 // ---------------------------------------------------
994 // A block of storage in the CMS generation is always in
995 // one of three states. A free block (FREE), an allocated
996 // object (OBJECT) whose size() method reports the correct size,
997 // and an intermediate state (TRANSIENT) in which its size cannot
998 // be accurately determined.
999 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1000 // -----------------------------------------------------
1001 // FREE: klass_word & 1 == 1; mark_word holds block size
1002 //
1003 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1004 // obj->size() computes correct size
1005 //
1006 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1007 //
1008 // STATE IDENTIFICATION: (64 bit+COOPS)
1009 // ------------------------------------
1010 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1011 //
1012 // OBJECT: klass_word installed; klass_word != 0;
1013 // obj->size() computes correct size
1014 //
1015 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1016 //
1017 //
1018 // STATE TRANSITION DIAGRAM
1019 //
1020 // mut / parnew mut / parnew
1021 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1022 // ^ |
1023 // |------------------------ DEAD <------------------------------------|
1024 // sweep mut
1025 //
1026 // While a block is in TRANSIENT state its size cannot be determined
1027 // so readers will either need to come back later or stall until
1028 // the size can be determined. Note that for the case of direct
1029 // allocation, P-bits, when available, may be used to determine the
1030 // size of an object that may not yet have been initialized.
1031
1032 // Things to support parallel young-gen collection.
1033 oop
1034 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1035 oop old, markOop m,
1036 size_t word_sz) {
1037 #ifndef PRODUCT
1038 if (GenCollectedHeap::heap()->promotion_should_fail()) {
1039 return NULL;
1040 }
1041 #endif // #ifndef PRODUCT
1042
1043 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1044 PromotionInfo* promoInfo = &ps->promo;
1045 // if we are tracking promotions, then first ensure space for
1046 // promotion (including spooling space for saving header if necessary).
1047 // then allocate and copy, then track promoted info if needed.
1048 // When tracking (see PromotionInfo::track()), the mark word may
1049 // be displaced and in this case restoration of the mark word
1050 // occurs in the (oop_since_save_marks_)iterate phase.
1051 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1052 // Out of space for allocating spooling buffers;
1053 // try expanding and allocating spooling buffers.
1054 if (!expand_and_ensure_spooling_space(promoInfo)) {
1055 return NULL;
1056 }
1057 }
1058 assert(promoInfo->has_spooling_space(), "Control point invariant");
1059 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1060 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1061 if (obj_ptr == NULL) {
1062 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1063 if (obj_ptr == NULL) {
1064 return NULL;
1065 }
1066 }
1067 oop obj = oop(obj_ptr);
1068 OrderAccess::storestore();
1069 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1070 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1071 // IMPORTANT: See note on object initialization for CMS above.
1072 // Otherwise, copy the object. Here we must be careful to insert the
1073 // klass pointer last, since this marks the block as an allocated object.
1074 // Except with compressed oops it's the mark word.
1075 HeapWord* old_ptr = (HeapWord*)old;
1076 // Restore the mark word copied above.
1077 obj->set_mark(m);
1078 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1079 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1080 OrderAccess::storestore();
1081
1082 if (UseCompressedClassPointers) {
1083 // Copy gap missed by (aligned) header size calculation below
1084 obj->set_klass_gap(old->klass_gap());
1085 }
1086 if (word_sz > (size_t)oopDesc::header_size()) {
1087 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1088 obj_ptr + oopDesc::header_size(),
1089 word_sz - oopDesc::header_size());
1090 }
1091
1092 // Now we can track the promoted object, if necessary. We take care
1093 // to delay the transition from uninitialized to full object
1094 // (i.e., insertion of klass pointer) until after, so that it
1095 // atomically becomes a promoted object.
1096 if (promoInfo->tracking()) {
1097 promoInfo->track((PromotedObject*)obj, old->klass());
1098 }
1099 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1100 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1101 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1102
1103 // Finally, install the klass pointer (this should be volatile).
1104 OrderAccess::storestore();
1105 obj->set_klass(old->klass());
1106 // We should now be able to calculate the right size for this object
1107 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1108
1109 collector()->promoted(true, // parallel
1110 obj_ptr, old->is_objArray(), word_sz);
1111
1112 NOT_PRODUCT(
1113 Atomic::inc_ptr(&_numObjectsPromoted);
1114 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1115 )
1116
1117 return obj;
1118 }
1119
1120 void
1121 ConcurrentMarkSweepGeneration::
1122 par_promote_alloc_done(int thread_num) {
1123 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1124 ps->lab.retire(thread_num);
1125 }
1126
1127 void
1128 ConcurrentMarkSweepGeneration::
1129 par_oop_since_save_marks_iterate_done(int thread_num) {
1130 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1131 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1132 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1133 }
1134
1135 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1136 size_t size,
1137 bool tlab)
1138 {
1139 // We allow a STW collection only if a full
1140 // collection was requested.
1141 return full || should_allocate(size, tlab); // FIX ME !!!
1142 // This and promotion failure handling are connected at the
1143 // hip and should be fixed by untying them.
1144 }
1145
1146 bool CMSCollector::shouldConcurrentCollect() {
1147 if (_full_gc_requested) {
1148 if (Verbose && PrintGCDetails) {
1149 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1150 " gc request (or gc_locker)");
1151 }
1152 return true;
1153 }
1154
1155 FreelistLocker x(this);
1156 // ------------------------------------------------------------------
1157 // Print out lots of information which affects the initiation of
1158 // a collection.
1159 if (PrintCMSInitiationStatistics && stats().valid()) {
1160 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1161 gclog_or_tty->stamp();
1162 gclog_or_tty->cr();
1163 stats().print_on(gclog_or_tty);
1164 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1165 stats().time_until_cms_gen_full());
1166 gclog_or_tty->print_cr("free=" SIZE_FORMAT, _cmsGen->free());
1167 gclog_or_tty->print_cr("contiguous_available=" SIZE_FORMAT,
1168 _cmsGen->contiguous_available());
1169 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1170 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1171 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1172 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1173 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1174 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1175 gclog_or_tty->print_cr("metadata initialized %d",
1176 MetaspaceGC::should_concurrent_collect());
1177 }
1178 // ------------------------------------------------------------------
1179
1180 // If the estimated time to complete a cms collection (cms_duration())
1181 // is less than the estimated time remaining until the cms generation
1182 // is full, start a collection.
1183 if (!UseCMSInitiatingOccupancyOnly) {
1184 if (stats().valid()) {
1185 if (stats().time_until_cms_start() == 0.0) {
1186 return true;
1187 }
1188 } else {
1189 // We want to conservatively collect somewhat early in order
1190 // to try and "bootstrap" our CMS/promotion statistics;
1191 // this branch will not fire after the first successful CMS
1192 // collection because the stats should then be valid.
1193 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1194 if (Verbose && PrintGCDetails) {
1195 gclog_or_tty->print_cr(
1196 " CMSCollector: collect for bootstrapping statistics:"
1197 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1198 _bootstrap_occupancy);
1199 }
1200 return true;
1201 }
1202 }
1203 }
1204
1205 // Otherwise, we start a collection cycle if
1206 // old gen want a collection cycle started. Each may use
1207 // an appropriate criterion for making this decision.
1208 // XXX We need to make sure that the gen expansion
1209 // criterion dovetails well with this. XXX NEED TO FIX THIS
1210 if (_cmsGen->should_concurrent_collect()) {
1211 if (Verbose && PrintGCDetails) {
1212 gclog_or_tty->print_cr("CMS old gen initiated");
1213 }
1214 return true;
1215 }
1216
1217 // We start a collection if we believe an incremental collection may fail;
1218 // this is not likely to be productive in practice because it's probably too
1219 // late anyway.
1220 GenCollectedHeap* gch = GenCollectedHeap::heap();
1221 assert(gch->collector_policy()->is_generation_policy(),
1222 "You may want to check the correctness of the following");
1223 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1224 if (Verbose && PrintGCDetails) {
1225 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1226 }
1227 return true;
1228 }
1229
1230 if (MetaspaceGC::should_concurrent_collect()) {
1231 if (Verbose && PrintGCDetails) {
1232 gclog_or_tty->print("CMSCollector: collect for metadata allocation ");
1233 }
1234 return true;
1235 }
1236
1237 // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1238 if (CMSTriggerInterval >= 0) {
1239 if (CMSTriggerInterval == 0) {
1240 // Trigger always
1241 return true;
1242 }
1243
1244 // Check the CMS time since begin (we do not check the stats validity
1245 // as we want to be able to trigger the first CMS cycle as well)
1246 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1247 if (Verbose && PrintGCDetails) {
1248 if (stats().valid()) {
1249 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1250 stats().cms_time_since_begin());
1251 } else {
1252 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)");
1253 }
1254 }
1255 return true;
1256 }
1257 }
1258
1259 return false;
1260 }
1261
1262 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1263
1264 // Clear _expansion_cause fields of constituent generations
1265 void CMSCollector::clear_expansion_cause() {
1266 _cmsGen->clear_expansion_cause();
1267 }
1268
1269 // We should be conservative in starting a collection cycle. To
1270 // start too eagerly runs the risk of collecting too often in the
1271 // extreme. To collect too rarely falls back on full collections,
1272 // which works, even if not optimum in terms of concurrent work.
1273 // As a work around for too eagerly collecting, use the flag
1274 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1275 // giving the user an easily understandable way of controlling the
1276 // collections.
1277 // We want to start a new collection cycle if any of the following
1278 // conditions hold:
1279 // . our current occupancy exceeds the configured initiating occupancy
1280 // for this generation, or
1281 // . we recently needed to expand this space and have not, since that
1282 // expansion, done a collection of this generation, or
1283 // . the underlying space believes that it may be a good idea to initiate
1284 // a concurrent collection (this may be based on criteria such as the
1285 // following: the space uses linear allocation and linear allocation is
1286 // going to fail, or there is believed to be excessive fragmentation in
1287 // the generation, etc... or ...
1288 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1289 // the case of the old generation; see CR 6543076):
1290 // we may be approaching a point at which allocation requests may fail because
1291 // we will be out of sufficient free space given allocation rate estimates.]
1292 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1293
1294 assert_lock_strong(freelistLock());
1295 if (occupancy() > initiating_occupancy()) {
1296 if (PrintGCDetails && Verbose) {
1297 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1298 short_name(), occupancy(), initiating_occupancy());
1299 }
1300 return true;
1301 }
1302 if (UseCMSInitiatingOccupancyOnly) {
1303 return false;
1304 }
1305 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1306 if (PrintGCDetails && Verbose) {
1307 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1308 short_name());
1309 }
1310 return true;
1311 }
1312 return false;
1313 }
1314
1315 void ConcurrentMarkSweepGeneration::collect(bool full,
1316 bool clear_all_soft_refs,
1317 size_t size,
1318 bool tlab)
1319 {
1320 collector()->collect(full, clear_all_soft_refs, size, tlab);
1321 }
1322
1323 void CMSCollector::collect(bool full,
1324 bool clear_all_soft_refs,
1325 size_t size,
1326 bool tlab)
1327 {
1328 // The following "if" branch is present for defensive reasons.
1329 // In the current uses of this interface, it can be replaced with:
1330 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1331 // But I am not placing that assert here to allow future
1332 // generality in invoking this interface.
1333 if (GC_locker::is_active()) {
1334 // A consistency test for GC_locker
1335 assert(GC_locker::needs_gc(), "Should have been set already");
1336 // Skip this foreground collection, instead
1337 // expanding the heap if necessary.
1338 // Need the free list locks for the call to free() in compute_new_size()
1339 compute_new_size();
1340 return;
1341 }
1342 acquire_control_and_collect(full, clear_all_soft_refs);
1343 }
1344
1345 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1346 GenCollectedHeap* gch = GenCollectedHeap::heap();
1347 unsigned int gc_count = gch->total_full_collections();
1348 if (gc_count == full_gc_count) {
1349 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1350 _full_gc_requested = true;
1351 _full_gc_cause = cause;
1352 CGC_lock->notify(); // nudge CMS thread
1353 } else {
1354 assert(gc_count > full_gc_count, "Error: causal loop");
1355 }
1356 }
1357
1358 bool CMSCollector::is_external_interruption() {
1359 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1360 return GCCause::is_user_requested_gc(cause) ||
1361 GCCause::is_serviceability_requested_gc(cause);
1362 }
1363
1364 void CMSCollector::report_concurrent_mode_interruption() {
1365 if (is_external_interruption()) {
1366 if (PrintGCDetails) {
1367 gclog_or_tty->print(" (concurrent mode interrupted)");
1368 }
1369 } else {
1370 if (PrintGCDetails) {
1371 gclog_or_tty->print(" (concurrent mode failure)");
1372 }
1373 _gc_tracer_cm->report_concurrent_mode_failure();
1374 }
1375 }
1376
1377
1378 // The foreground and background collectors need to coordinate in order
1379 // to make sure that they do not mutually interfere with CMS collections.
1380 // When a background collection is active,
1381 // the foreground collector may need to take over (preempt) and
1382 // synchronously complete an ongoing collection. Depending on the
1383 // frequency of the background collections and the heap usage
1384 // of the application, this preemption can be seldom or frequent.
1385 // There are only certain
1386 // points in the background collection that the "collection-baton"
1387 // can be passed to the foreground collector.
1388 //
1389 // The foreground collector will wait for the baton before
1390 // starting any part of the collection. The foreground collector
1391 // will only wait at one location.
1392 //
1393 // The background collector will yield the baton before starting a new
1394 // phase of the collection (e.g., before initial marking, marking from roots,
1395 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1396 // of the loop which switches the phases. The background collector does some
1397 // of the phases (initial mark, final re-mark) with the world stopped.
1398 // Because of locking involved in stopping the world,
1399 // the foreground collector should not block waiting for the background
1400 // collector when it is doing a stop-the-world phase. The background
1401 // collector will yield the baton at an additional point just before
1402 // it enters a stop-the-world phase. Once the world is stopped, the
1403 // background collector checks the phase of the collection. If the
1404 // phase has not changed, it proceeds with the collection. If the
1405 // phase has changed, it skips that phase of the collection. See
1406 // the comments on the use of the Heap_lock in collect_in_background().
1407 //
1408 // Variable used in baton passing.
1409 // _foregroundGCIsActive - Set to true by the foreground collector when
1410 // it wants the baton. The foreground clears it when it has finished
1411 // the collection.
1412 // _foregroundGCShouldWait - Set to true by the background collector
1413 // when it is running. The foreground collector waits while
1414 // _foregroundGCShouldWait is true.
1415 // CGC_lock - monitor used to protect access to the above variables
1416 // and to notify the foreground and background collectors.
1417 // _collectorState - current state of the CMS collection.
1418 //
1419 // The foreground collector
1420 // acquires the CGC_lock
1421 // sets _foregroundGCIsActive
1422 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1423 // various locks acquired in preparation for the collection
1424 // are released so as not to block the background collector
1425 // that is in the midst of a collection
1426 // proceeds with the collection
1427 // clears _foregroundGCIsActive
1428 // returns
1429 //
1430 // The background collector in a loop iterating on the phases of the
1431 // collection
1432 // acquires the CGC_lock
1433 // sets _foregroundGCShouldWait
1434 // if _foregroundGCIsActive is set
1435 // clears _foregroundGCShouldWait, notifies _CGC_lock
1436 // waits on _CGC_lock for _foregroundGCIsActive to become false
1437 // and exits the loop.
1438 // otherwise
1439 // proceed with that phase of the collection
1440 // if the phase is a stop-the-world phase,
1441 // yield the baton once more just before enqueueing
1442 // the stop-world CMS operation (executed by the VM thread).
1443 // returns after all phases of the collection are done
1444 //
1445
1446 void CMSCollector::acquire_control_and_collect(bool full,
1447 bool clear_all_soft_refs) {
1448 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1449 assert(!Thread::current()->is_ConcurrentGC_thread(),
1450 "shouldn't try to acquire control from self!");
1451
1452 // Start the protocol for acquiring control of the
1453 // collection from the background collector (aka CMS thread).
1454 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1455 "VM thread should have CMS token");
1456 // Remember the possibly interrupted state of an ongoing
1457 // concurrent collection
1458 CollectorState first_state = _collectorState;
1459
1460 // Signal to a possibly ongoing concurrent collection that
1461 // we want to do a foreground collection.
1462 _foregroundGCIsActive = true;
1463
1464 // release locks and wait for a notify from the background collector
1465 // releasing the locks in only necessary for phases which
1466 // do yields to improve the granularity of the collection.
1467 assert_lock_strong(bitMapLock());
1468 // We need to lock the Free list lock for the space that we are
1469 // currently collecting.
1470 assert(haveFreelistLocks(), "Must be holding free list locks");
1471 bitMapLock()->unlock();
1472 releaseFreelistLocks();
1473 {
1474 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1475 if (_foregroundGCShouldWait) {
1476 // We are going to be waiting for action for the CMS thread;
1477 // it had better not be gone (for instance at shutdown)!
1478 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1479 "CMS thread must be running");
1480 // Wait here until the background collector gives us the go-ahead
1481 ConcurrentMarkSweepThread::clear_CMS_flag(
1482 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1483 // Get a possibly blocked CMS thread going:
1484 // Note that we set _foregroundGCIsActive true above,
1485 // without protection of the CGC_lock.
1486 CGC_lock->notify();
1487 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1488 "Possible deadlock");
1489 while (_foregroundGCShouldWait) {
1490 // wait for notification
1491 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1492 // Possibility of delay/starvation here, since CMS token does
1493 // not know to give priority to VM thread? Actually, i think
1494 // there wouldn't be any delay/starvation, but the proof of
1495 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1496 }
1497 ConcurrentMarkSweepThread::set_CMS_flag(
1498 ConcurrentMarkSweepThread::CMS_vm_has_token);
1499 }
1500 }
1501 // The CMS_token is already held. Get back the other locks.
1502 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1503 "VM thread should have CMS token");
1504 getFreelistLocks();
1505 bitMapLock()->lock_without_safepoint_check();
1506 if (TraceCMSState) {
1507 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1508 INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state);
1509 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1510 }
1511
1512 // Inform cms gen if this was due to partial collection failing.
1513 // The CMS gen may use this fact to determine its expansion policy.
1514 GenCollectedHeap* gch = GenCollectedHeap::heap();
1515 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1516 assert(!_cmsGen->incremental_collection_failed(),
1517 "Should have been noticed, reacted to and cleared");
1518 _cmsGen->set_incremental_collection_failed();
1519 }
1520
1521 if (first_state > Idling) {
1522 report_concurrent_mode_interruption();
1523 }
1524
1525 set_did_compact(true);
1526
1527 // If the collection is being acquired from the background
1528 // collector, there may be references on the discovered
1529 // references lists. Abandon those references, since some
1530 // of them may have become unreachable after concurrent
1531 // discovery; the STW compacting collector will redo discovery
1532 // more precisely, without being subject to floating garbage.
1533 // Leaving otherwise unreachable references in the discovered
1534 // lists would require special handling.
1535 ref_processor()->disable_discovery();
1536 ref_processor()->abandon_partial_discovery();
1537 ref_processor()->verify_no_references_recorded();
1538
1539 if (first_state > Idling) {
1540 save_heap_summary();
1541 }
1542
1543 do_compaction_work(clear_all_soft_refs);
1544
1545 // Has the GC time limit been exceeded?
1546 size_t max_eden_size = _young_gen->max_eden_size();
1547 GCCause::Cause gc_cause = gch->gc_cause();
1548 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1549 _young_gen->eden()->used(),
1550 _cmsGen->max_capacity(),
1551 max_eden_size,
1552 full,
1553 gc_cause,
1554 gch->collector_policy());
1555
1556 // Reset the expansion cause, now that we just completed
1557 // a collection cycle.
1558 clear_expansion_cause();
1559 _foregroundGCIsActive = false;
1560 return;
1561 }
1562
1563 // Resize the tenured generation
1564 // after obtaining the free list locks for the
1565 // two generations.
1566 void CMSCollector::compute_new_size() {
1567 assert_locked_or_safepoint(Heap_lock);
1568 FreelistLocker z(this);
1569 MetaspaceGC::compute_new_size();
1570 _cmsGen->compute_new_size_free_list();
1571 }
1572
1573 // A work method used by the foreground collector to do
1574 // a mark-sweep-compact.
1575 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1576 GenCollectedHeap* gch = GenCollectedHeap::heap();
1577
1578 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1579 gc_timer->register_gc_start();
1580
1581 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1582 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
1583
1584 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL);
1585
1586 // Temporarily widen the span of the weak reference processing to
1587 // the entire heap.
1588 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1589 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1590 // Temporarily, clear the "is_alive_non_header" field of the
1591 // reference processor.
1592 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1593 // Temporarily make reference _processing_ single threaded (non-MT).
1594 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1595 // Temporarily make refs discovery atomic
1596 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1597 // Temporarily make reference _discovery_ single threaded (non-MT)
1598 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1599
1600 ref_processor()->set_enqueuing_is_done(false);
1601 ref_processor()->enable_discovery();
1602 ref_processor()->setup_policy(clear_all_soft_refs);
1603 // If an asynchronous collection finishes, the _modUnionTable is
1604 // all clear. If we are assuming the collection from an asynchronous
1605 // collection, clear the _modUnionTable.
1606 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1607 "_modUnionTable should be clear if the baton was not passed");
1608 _modUnionTable.clear_all();
1609 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
1610 "mod union for klasses should be clear if the baton was passed");
1611 _ct->klass_rem_set()->clear_mod_union();
1612
1613 // We must adjust the allocation statistics being maintained
1614 // in the free list space. We do so by reading and clearing
1615 // the sweep timer and updating the block flux rate estimates below.
1616 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1617 if (_inter_sweep_timer.is_active()) {
1618 _inter_sweep_timer.stop();
1619 // Note that we do not use this sample to update the _inter_sweep_estimate.
1620 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1621 _inter_sweep_estimate.padded_average(),
1622 _intra_sweep_estimate.padded_average());
1623 }
1624
1625 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs);
1626 #ifdef ASSERT
1627 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1628 size_t free_size = cms_space->free();
1629 assert(free_size ==
1630 pointer_delta(cms_space->end(), cms_space->compaction_top())
1631 * HeapWordSize,
1632 "All the free space should be compacted into one chunk at top");
1633 assert(cms_space->dictionary()->total_chunk_size(
1634 debug_only(cms_space->freelistLock())) == 0 ||
1635 cms_space->totalSizeInIndexedFreeLists() == 0,
1636 "All the free space should be in a single chunk");
1637 size_t num = cms_space->totalCount();
1638 assert((free_size == 0 && num == 0) ||
1639 (free_size > 0 && (num == 1 || num == 2)),
1640 "There should be at most 2 free chunks after compaction");
1641 #endif // ASSERT
1642 _collectorState = Resetting;
1643 assert(_restart_addr == NULL,
1644 "Should have been NULL'd before baton was passed");
1645 reset_stw();
1646 _cmsGen->reset_after_compaction();
1647 _concurrent_cycles_since_last_unload = 0;
1648
1649 // Clear any data recorded in the PLAB chunk arrays.
1650 if (_survivor_plab_array != NULL) {
1651 reset_survivor_plab_arrays();
1652 }
1653
1654 // Adjust the per-size allocation stats for the next epoch.
1655 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1656 // Restart the "inter sweep timer" for the next epoch.
1657 _inter_sweep_timer.reset();
1658 _inter_sweep_timer.start();
1659
1660 gc_timer->register_gc_end();
1661
1662 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1663
1664 // For a mark-sweep-compact, compute_new_size() will be called
1665 // in the heap's do_collection() method.
1666 }
1667
1668 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1669 ContiguousSpace* eden_space = _young_gen->eden();
1670 ContiguousSpace* from_space = _young_gen->from();
1671 ContiguousSpace* to_space = _young_gen->to();
1672 // Eden
1673 if (_eden_chunk_array != NULL) {
1674 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1675 p2i(eden_space->bottom()), p2i(eden_space->top()),
1676 p2i(eden_space->end()), eden_space->capacity());
1677 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", "
1678 "_eden_chunk_capacity=" SIZE_FORMAT,
1679 _eden_chunk_index, _eden_chunk_capacity);
1680 for (size_t i = 0; i < _eden_chunk_index; i++) {
1681 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1682 i, p2i(_eden_chunk_array[i]));
1683 }
1684 }
1685 // Survivor
1686 if (_survivor_chunk_array != NULL) {
1687 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1688 p2i(from_space->bottom()), p2i(from_space->top()),
1689 p2i(from_space->end()), from_space->capacity());
1690 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", "
1691 "_survivor_chunk_capacity=" SIZE_FORMAT,
1692 _survivor_chunk_index, _survivor_chunk_capacity);
1693 for (size_t i = 0; i < _survivor_chunk_index; i++) {
1694 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1695 i, p2i(_survivor_chunk_array[i]));
1696 }
1697 }
1698 }
1699
1700 void CMSCollector::getFreelistLocks() const {
1701 // Get locks for all free lists in all generations that this
1702 // collector is responsible for
1703 _cmsGen->freelistLock()->lock_without_safepoint_check();
1704 }
1705
1706 void CMSCollector::releaseFreelistLocks() const {
1707 // Release locks for all free lists in all generations that this
1708 // collector is responsible for
1709 _cmsGen->freelistLock()->unlock();
1710 }
1711
1712 bool CMSCollector::haveFreelistLocks() const {
1713 // Check locks for all free lists in all generations that this
1714 // collector is responsible for
1715 assert_lock_strong(_cmsGen->freelistLock());
1716 PRODUCT_ONLY(ShouldNotReachHere());
1717 return true;
1718 }
1719
1720 // A utility class that is used by the CMS collector to
1721 // temporarily "release" the foreground collector from its
1722 // usual obligation to wait for the background collector to
1723 // complete an ongoing phase before proceeding.
1724 class ReleaseForegroundGC: public StackObj {
1725 private:
1726 CMSCollector* _c;
1727 public:
1728 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1729 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1730 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1731 // allow a potentially blocked foreground collector to proceed
1732 _c->_foregroundGCShouldWait = false;
1733 if (_c->_foregroundGCIsActive) {
1734 CGC_lock->notify();
1735 }
1736 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1737 "Possible deadlock");
1738 }
1739
1740 ~ReleaseForegroundGC() {
1741 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1742 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1743 _c->_foregroundGCShouldWait = true;
1744 }
1745 };
1746
1747 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1748 assert(Thread::current()->is_ConcurrentGC_thread(),
1749 "A CMS asynchronous collection is only allowed on a CMS thread.");
1750
1751 GenCollectedHeap* gch = GenCollectedHeap::heap();
1752 {
1753 bool safepoint_check = Mutex::_no_safepoint_check_flag;
1754 MutexLockerEx hl(Heap_lock, safepoint_check);
1755 FreelistLocker fll(this);
1756 MutexLockerEx x(CGC_lock, safepoint_check);
1757 if (_foregroundGCIsActive) {
1758 // The foreground collector is. Skip this
1759 // background collection.
1760 assert(!_foregroundGCShouldWait, "Should be clear");
1761 return;
1762 } else {
1763 assert(_collectorState == Idling, "Should be idling before start.");
1764 _collectorState = InitialMarking;
1765 register_gc_start(cause);
1766 // Reset the expansion cause, now that we are about to begin
1767 // a new cycle.
1768 clear_expansion_cause();
1769
1770 // Clear the MetaspaceGC flag since a concurrent collection
1771 // is starting but also clear it after the collection.
1772 MetaspaceGC::set_should_concurrent_collect(false);
1773 }
1774 // Decide if we want to enable class unloading as part of the
1775 // ensuing concurrent GC cycle.
1776 update_should_unload_classes();
1777 _full_gc_requested = false; // acks all outstanding full gc requests
1778 _full_gc_cause = GCCause::_no_gc;
1779 // Signal that we are about to start a collection
1780 gch->increment_total_full_collections(); // ... starting a collection cycle
1781 _collection_count_start = gch->total_full_collections();
1782 }
1783
1784 // Used for PrintGC
1785 size_t prev_used = 0;
1786 if (PrintGC && Verbose) {
1787 prev_used = _cmsGen->used();
1788 }
1789
1790 // The change of the collection state is normally done at this level;
1791 // the exceptions are phases that are executed while the world is
1792 // stopped. For those phases the change of state is done while the
1793 // world is stopped. For baton passing purposes this allows the
1794 // background collector to finish the phase and change state atomically.
1795 // The foreground collector cannot wait on a phase that is done
1796 // while the world is stopped because the foreground collector already
1797 // has the world stopped and would deadlock.
1798 while (_collectorState != Idling) {
1799 if (TraceCMSState) {
1800 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
1801 p2i(Thread::current()), _collectorState);
1802 }
1803 // The foreground collector
1804 // holds the Heap_lock throughout its collection.
1805 // holds the CMS token (but not the lock)
1806 // except while it is waiting for the background collector to yield.
1807 //
1808 // The foreground collector should be blocked (not for long)
1809 // if the background collector is about to start a phase
1810 // executed with world stopped. If the background
1811 // collector has already started such a phase, the
1812 // foreground collector is blocked waiting for the
1813 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
1814 // are executed in the VM thread.
1815 //
1816 // The locking order is
1817 // PendingListLock (PLL) -- if applicable (FinalMarking)
1818 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
1819 // CMS token (claimed in
1820 // stop_world_and_do() -->
1821 // safepoint_synchronize() -->
1822 // CMSThread::synchronize())
1823
1824 {
1825 // Check if the FG collector wants us to yield.
1826 CMSTokenSync x(true); // is cms thread
1827 if (waitForForegroundGC()) {
1828 // We yielded to a foreground GC, nothing more to be
1829 // done this round.
1830 assert(_foregroundGCShouldWait == false, "We set it to false in "
1831 "waitForForegroundGC()");
1832 if (TraceCMSState) {
1833 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1834 " exiting collection CMS state %d",
1835 p2i(Thread::current()), _collectorState);
1836 }
1837 return;
1838 } else {
1839 // The background collector can run but check to see if the
1840 // foreground collector has done a collection while the
1841 // background collector was waiting to get the CGC_lock
1842 // above. If yes, break so that _foregroundGCShouldWait
1843 // is cleared before returning.
1844 if (_collectorState == Idling) {
1845 break;
1846 }
1847 }
1848 }
1849
1850 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1851 "should be waiting");
1852
1853 switch (_collectorState) {
1854 case InitialMarking:
1855 {
1856 ReleaseForegroundGC x(this);
1857 stats().record_cms_begin();
1858 VM_CMS_Initial_Mark initial_mark_op(this);
1859 VMThread::execute(&initial_mark_op);
1860 }
1861 // The collector state may be any legal state at this point
1862 // since the background collector may have yielded to the
1863 // foreground collector.
1864 break;
1865 case Marking:
1866 // initial marking in checkpointRootsInitialWork has been completed
1867 if (markFromRoots()) { // we were successful
1868 assert(_collectorState == Precleaning, "Collector state should "
1869 "have changed");
1870 } else {
1871 assert(_foregroundGCIsActive, "Internal state inconsistency");
1872 }
1873 break;
1874 case Precleaning:
1875 // marking from roots in markFromRoots has been completed
1876 preclean();
1877 assert(_collectorState == AbortablePreclean ||
1878 _collectorState == FinalMarking,
1879 "Collector state should have changed");
1880 break;
1881 case AbortablePreclean:
1882 abortable_preclean();
1883 assert(_collectorState == FinalMarking, "Collector state should "
1884 "have changed");
1885 break;
1886 case FinalMarking:
1887 {
1888 ReleaseForegroundGC x(this);
1889
1890 VM_CMS_Final_Remark final_remark_op(this);
1891 VMThread::execute(&final_remark_op);
1892 }
1893 assert(_foregroundGCShouldWait, "block post-condition");
1894 break;
1895 case Sweeping:
1896 // final marking in checkpointRootsFinal has been completed
1897 sweep();
1898 assert(_collectorState == Resizing, "Collector state change "
1899 "to Resizing must be done under the free_list_lock");
1900
1901 case Resizing: {
1902 // Sweeping has been completed...
1903 // At this point the background collection has completed.
1904 // Don't move the call to compute_new_size() down
1905 // into code that might be executed if the background
1906 // collection was preempted.
1907 {
1908 ReleaseForegroundGC x(this); // unblock FG collection
1909 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
1910 CMSTokenSync z(true); // not strictly needed.
1911 if (_collectorState == Resizing) {
1912 compute_new_size();
1913 save_heap_summary();
1914 _collectorState = Resetting;
1915 } else {
1916 assert(_collectorState == Idling, "The state should only change"
1917 " because the foreground collector has finished the collection");
1918 }
1919 }
1920 break;
1921 }
1922 case Resetting:
1923 // CMS heap resizing has been completed
1924 reset_concurrent();
1925 assert(_collectorState == Idling, "Collector state should "
1926 "have changed");
1927
1928 MetaspaceGC::set_should_concurrent_collect(false);
1929
1930 stats().record_cms_end();
1931 // Don't move the concurrent_phases_end() and compute_new_size()
1932 // calls to here because a preempted background collection
1933 // has it's state set to "Resetting".
1934 break;
1935 case Idling:
1936 default:
1937 ShouldNotReachHere();
1938 break;
1939 }
1940 if (TraceCMSState) {
1941 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
1942 p2i(Thread::current()), _collectorState);
1943 }
1944 assert(_foregroundGCShouldWait, "block post-condition");
1945 }
1946
1947 // Should this be in gc_epilogue?
1948 collector_policy()->counters()->update_counters();
1949
1950 {
1951 // Clear _foregroundGCShouldWait and, in the event that the
1952 // foreground collector is waiting, notify it, before
1953 // returning.
1954 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1955 _foregroundGCShouldWait = false;
1956 if (_foregroundGCIsActive) {
1957 CGC_lock->notify();
1958 }
1959 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1960 "Possible deadlock");
1961 }
1962 if (TraceCMSState) {
1963 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1964 " exiting collection CMS state %d",
1965 p2i(Thread::current()), _collectorState);
1966 }
1967 if (PrintGC && Verbose) {
1968 _cmsGen->print_heap_change(prev_used);
1969 }
1970 }
1971
1972 void CMSCollector::register_gc_start(GCCause::Cause cause) {
1973 _cms_start_registered = true;
1974 _gc_timer_cm->register_gc_start();
1975 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
1976 }
1977
1978 void CMSCollector::register_gc_end() {
1979 if (_cms_start_registered) {
1980 report_heap_summary(GCWhen::AfterGC);
1981
1982 _gc_timer_cm->register_gc_end();
1983 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1984 _cms_start_registered = false;
1985 }
1986 }
1987
1988 void CMSCollector::save_heap_summary() {
1989 GenCollectedHeap* gch = GenCollectedHeap::heap();
1990 _last_heap_summary = gch->create_heap_summary();
1991 _last_metaspace_summary = gch->create_metaspace_summary();
1992 }
1993
1994 void CMSCollector::report_heap_summary(GCWhen::Type when) {
1995 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
1996 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
1997 }
1998
1999 bool CMSCollector::waitForForegroundGC() {
2000 bool res = false;
2001 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2002 "CMS thread should have CMS token");
2003 // Block the foreground collector until the
2004 // background collectors decides whether to
2005 // yield.
2006 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2007 _foregroundGCShouldWait = true;
2008 if (_foregroundGCIsActive) {
2009 // The background collector yields to the
2010 // foreground collector and returns a value
2011 // indicating that it has yielded. The foreground
2012 // collector can proceed.
2013 res = true;
2014 _foregroundGCShouldWait = false;
2015 ConcurrentMarkSweepThread::clear_CMS_flag(
2016 ConcurrentMarkSweepThread::CMS_cms_has_token);
2017 ConcurrentMarkSweepThread::set_CMS_flag(
2018 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2019 // Get a possibly blocked foreground thread going
2020 CGC_lock->notify();
2021 if (TraceCMSState) {
2022 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2023 p2i(Thread::current()), _collectorState);
2024 }
2025 while (_foregroundGCIsActive) {
2026 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2027 }
2028 ConcurrentMarkSweepThread::set_CMS_flag(
2029 ConcurrentMarkSweepThread::CMS_cms_has_token);
2030 ConcurrentMarkSweepThread::clear_CMS_flag(
2031 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2032 }
2033 if (TraceCMSState) {
2034 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2035 p2i(Thread::current()), _collectorState);
2036 }
2037 return res;
2038 }
2039
2040 // Because of the need to lock the free lists and other structures in
2041 // the collector, common to all the generations that the collector is
2042 // collecting, we need the gc_prologues of individual CMS generations
2043 // delegate to their collector. It may have been simpler had the
2044 // current infrastructure allowed one to call a prologue on a
2045 // collector. In the absence of that we have the generation's
2046 // prologue delegate to the collector, which delegates back
2047 // some "local" work to a worker method in the individual generations
2048 // that it's responsible for collecting, while itself doing any
2049 // work common to all generations it's responsible for. A similar
2050 // comment applies to the gc_epilogue()'s.
2051 // The role of the variable _between_prologue_and_epilogue is to
2052 // enforce the invocation protocol.
2053 void CMSCollector::gc_prologue(bool full) {
2054 // Call gc_prologue_work() for the CMSGen
2055 // we are responsible for.
2056
2057 // The following locking discipline assumes that we are only called
2058 // when the world is stopped.
2059 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2060
2061 // The CMSCollector prologue must call the gc_prologues for the
2062 // "generations" that it's responsible
2063 // for.
2064
2065 assert( Thread::current()->is_VM_thread()
2066 || ( CMSScavengeBeforeRemark
2067 && Thread::current()->is_ConcurrentGC_thread()),
2068 "Incorrect thread type for prologue execution");
2069
2070 if (_between_prologue_and_epilogue) {
2071 // We have already been invoked; this is a gc_prologue delegation
2072 // from yet another CMS generation that we are responsible for, just
2073 // ignore it since all relevant work has already been done.
2074 return;
2075 }
2076
2077 // set a bit saying prologue has been called; cleared in epilogue
2078 _between_prologue_and_epilogue = true;
2079 // Claim locks for common data structures, then call gc_prologue_work()
2080 // for each CMSGen.
2081
2082 getFreelistLocks(); // gets free list locks on constituent spaces
2083 bitMapLock()->lock_without_safepoint_check();
2084
2085 // Should call gc_prologue_work() for all cms gens we are responsible for
2086 bool duringMarking = _collectorState >= Marking
2087 && _collectorState < Sweeping;
2088
2089 // The young collections clear the modified oops state, which tells if
2090 // there are any modified oops in the class. The remark phase also needs
2091 // that information. Tell the young collection to save the union of all
2092 // modified klasses.
2093 if (duringMarking) {
2094 _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2095 }
2096
2097 bool registerClosure = duringMarking;
2098
2099 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2100
2101 if (!full) {
2102 stats().record_gc0_begin();
2103 }
2104 }
2105
2106 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2107
2108 _capacity_at_prologue = capacity();
2109 _used_at_prologue = used();
2110
2111 // Delegate to CMScollector which knows how to coordinate between
2112 // this and any other CMS generations that it is responsible for
2113 // collecting.
2114 collector()->gc_prologue(full);
2115 }
2116
2117 // This is a "private" interface for use by this generation's CMSCollector.
2118 // Not to be called directly by any other entity (for instance,
2119 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2120 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2121 bool registerClosure, ModUnionClosure* modUnionClosure) {
2122 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2123 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2124 "Should be NULL");
2125 if (registerClosure) {
2126 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2127 }
2128 cmsSpace()->gc_prologue();
2129 // Clear stat counters
2130 NOT_PRODUCT(
2131 assert(_numObjectsPromoted == 0, "check");
2132 assert(_numWordsPromoted == 0, "check");
2133 if (Verbose && PrintGC) {
2134 gclog_or_tty->print("Allocated " SIZE_FORMAT " objects, "
2135 SIZE_FORMAT " bytes concurrently",
2136 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2137 }
2138 _numObjectsAllocated = 0;
2139 _numWordsAllocated = 0;
2140 )
2141 }
2142
2143 void CMSCollector::gc_epilogue(bool full) {
2144 // The following locking discipline assumes that we are only called
2145 // when the world is stopped.
2146 assert(SafepointSynchronize::is_at_safepoint(),
2147 "world is stopped assumption");
2148
2149 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2150 // if linear allocation blocks need to be appropriately marked to allow the
2151 // the blocks to be parsable. We also check here whether we need to nudge the
2152 // CMS collector thread to start a new cycle (if it's not already active).
2153 assert( Thread::current()->is_VM_thread()
2154 || ( CMSScavengeBeforeRemark
2155 && Thread::current()->is_ConcurrentGC_thread()),
2156 "Incorrect thread type for epilogue execution");
2157
2158 if (!_between_prologue_and_epilogue) {
2159 // We have already been invoked; this is a gc_epilogue delegation
2160 // from yet another CMS generation that we are responsible for, just
2161 // ignore it since all relevant work has already been done.
2162 return;
2163 }
2164 assert(haveFreelistLocks(), "must have freelist locks");
2165 assert_lock_strong(bitMapLock());
2166
2167 _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2168
2169 _cmsGen->gc_epilogue_work(full);
2170
2171 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2172 // in case sampling was not already enabled, enable it
2173 _start_sampling = true;
2174 }
2175 // reset _eden_chunk_array so sampling starts afresh
2176 _eden_chunk_index = 0;
2177
2178 size_t cms_used = _cmsGen->cmsSpace()->used();
2179
2180 // update performance counters - this uses a special version of
2181 // update_counters() that allows the utilization to be passed as a
2182 // parameter, avoiding multiple calls to used().
2183 //
2184 _cmsGen->update_counters(cms_used);
2185
2186 bitMapLock()->unlock();
2187 releaseFreelistLocks();
2188
2189 if (!CleanChunkPoolAsync) {
2190 Chunk::clean_chunk_pool();
2191 }
2192
2193 set_did_compact(false);
2194 _between_prologue_and_epilogue = false; // ready for next cycle
2195 }
2196
2197 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2198 collector()->gc_epilogue(full);
2199
2200 // Also reset promotion tracking in par gc thread states.
2201 for (uint i = 0; i < ParallelGCThreads; i++) {
2202 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2203 }
2204 }
2205
2206 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2207 assert(!incremental_collection_failed(), "Should have been cleared");
2208 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2209 cmsSpace()->gc_epilogue();
2210 // Print stat counters
2211 NOT_PRODUCT(
2212 assert(_numObjectsAllocated == 0, "check");
2213 assert(_numWordsAllocated == 0, "check");
2214 if (Verbose && PrintGC) {
2215 gclog_or_tty->print("Promoted " SIZE_FORMAT " objects, "
2216 SIZE_FORMAT " bytes",
2217 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2218 }
2219 _numObjectsPromoted = 0;
2220 _numWordsPromoted = 0;
2221 )
2222
2223 if (PrintGC && Verbose) {
2224 // Call down the chain in contiguous_available needs the freelistLock
2225 // so print this out before releasing the freeListLock.
2226 gclog_or_tty->print(" Contiguous available " SIZE_FORMAT " bytes ",
2227 contiguous_available());
2228 }
2229 }
2230
2231 #ifndef PRODUCT
2232 bool CMSCollector::have_cms_token() {
2233 Thread* thr = Thread::current();
2234 if (thr->is_VM_thread()) {
2235 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2236 } else if (thr->is_ConcurrentGC_thread()) {
2237 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2238 } else if (thr->is_GC_task_thread()) {
2239 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2240 ParGCRareEvent_lock->owned_by_self();
2241 }
2242 return false;
2243 }
2244 #endif
2245
2246 // Check reachability of the given heap address in CMS generation,
2247 // treating all other generations as roots.
2248 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2249 // We could "guarantee" below, rather than assert, but I'll
2250 // leave these as "asserts" so that an adventurous debugger
2251 // could try this in the product build provided some subset of
2252 // the conditions were met, provided they were interested in the
2253 // results and knew that the computation below wouldn't interfere
2254 // with other concurrent computations mutating the structures
2255 // being read or written.
2256 assert(SafepointSynchronize::is_at_safepoint(),
2257 "Else mutations in object graph will make answer suspect");
2258 assert(have_cms_token(), "Should hold cms token");
2259 assert(haveFreelistLocks(), "must hold free list locks");
2260 assert_lock_strong(bitMapLock());
2261
2262 // Clear the marking bit map array before starting, but, just
2263 // for kicks, first report if the given address is already marked
2264 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr),
2265 _markBitMap.isMarked(addr) ? "" : " not");
2266
2267 if (verify_after_remark()) {
2268 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2269 bool result = verification_mark_bm()->isMarked(addr);
2270 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr),
2271 result ? "IS" : "is NOT");
2272 return result;
2273 } else {
2274 gclog_or_tty->print_cr("Could not compute result");
2275 return false;
2276 }
2277 }
2278
2279
2280 void
2281 CMSCollector::print_on_error(outputStream* st) {
2282 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2283 if (collector != NULL) {
2284 CMSBitMap* bitmap = &collector->_markBitMap;
2285 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2286 bitmap->print_on_error(st, " Bits: ");
2287
2288 st->cr();
2289
2290 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2291 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2292 mut_bitmap->print_on_error(st, " Bits: ");
2293 }
2294 }
2295
2296 ////////////////////////////////////////////////////////
2297 // CMS Verification Support
2298 ////////////////////////////////////////////////////////
2299 // Following the remark phase, the following invariant
2300 // should hold -- each object in the CMS heap which is
2301 // marked in markBitMap() should be marked in the verification_mark_bm().
2302
2303 class VerifyMarkedClosure: public BitMapClosure {
2304 CMSBitMap* _marks;
2305 bool _failed;
2306
2307 public:
2308 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2309
2310 bool do_bit(size_t offset) {
2311 HeapWord* addr = _marks->offsetToHeapWord(offset);
2312 if (!_marks->isMarked(addr)) {
2313 oop(addr)->print_on(gclog_or_tty);
2314 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
2315 _failed = true;
2316 }
2317 return true;
2318 }
2319
2320 bool failed() { return _failed; }
2321 };
2322
2323 bool CMSCollector::verify_after_remark(bool silent) {
2324 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... ");
2325 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2326 static bool init = false;
2327
2328 assert(SafepointSynchronize::is_at_safepoint(),
2329 "Else mutations in object graph will make answer suspect");
2330 assert(have_cms_token(),
2331 "Else there may be mutual interference in use of "
2332 " verification data structures");
2333 assert(_collectorState > Marking && _collectorState <= Sweeping,
2334 "Else marking info checked here may be obsolete");
2335 assert(haveFreelistLocks(), "must hold free list locks");
2336 assert_lock_strong(bitMapLock());
2337
2338
2339 // Allocate marking bit map if not already allocated
2340 if (!init) { // first time
2341 if (!verification_mark_bm()->allocate(_span)) {
2342 return false;
2343 }
2344 init = true;
2345 }
2346
2347 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2348
2349 // Turn off refs discovery -- so we will be tracing through refs.
2350 // This is as intended, because by this time
2351 // GC must already have cleared any refs that need to be cleared,
2352 // and traced those that need to be marked; moreover,
2353 // the marking done here is not going to interfere in any
2354 // way with the marking information used by GC.
2355 NoRefDiscovery no_discovery(ref_processor());
2356
2357 #if defined(COMPILER2) || INCLUDE_JVMCI
2358 DerivedPointerTableDeactivate dpt_deact;
2359 #endif
2360
2361 // Clear any marks from a previous round
2362 verification_mark_bm()->clear_all();
2363 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2364 verify_work_stacks_empty();
2365
2366 GenCollectedHeap* gch = GenCollectedHeap::heap();
2367 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2368 // Update the saved marks which may affect the root scans.
2369 gch->save_marks();
2370
2371 if (CMSRemarkVerifyVariant == 1) {
2372 // In this first variant of verification, we complete
2373 // all marking, then check if the new marks-vector is
2374 // a subset of the CMS marks-vector.
2375 verify_after_remark_work_1();
2376 } else if (CMSRemarkVerifyVariant == 2) {
2377 // In this second variant of verification, we flag an error
2378 // (i.e. an object reachable in the new marks-vector not reachable
2379 // in the CMS marks-vector) immediately, also indicating the
2380 // identify of an object (A) that references the unmarked object (B) --
2381 // presumably, a mutation to A failed to be picked up by preclean/remark?
2382 verify_after_remark_work_2();
2383 } else {
2384 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant",
2385 CMSRemarkVerifyVariant);
2386 }
2387 if (!silent) gclog_or_tty->print(" done] ");
2388 return true;
2389 }
2390
2391 void CMSCollector::verify_after_remark_work_1() {
2392 ResourceMark rm;
2393 HandleMark hm;
2394 GenCollectedHeap* gch = GenCollectedHeap::heap();
2395
2396 // Get a clear set of claim bits for the roots processing to work with.
2397 ClassLoaderDataGraph::clear_claimed_marks();
2398
2399 // Mark from roots one level into CMS
2400 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2401 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2402
2403 {
2404 StrongRootsScope srs(1);
2405
2406 gch->gen_process_roots(&srs,
2407 GenCollectedHeap::OldGen,
2408 true, // young gen as roots
2409 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2410 should_unload_classes(),
2411 ¬Older,
2412 NULL,
2413 NULL);
2414 }
2415
2416 // Now mark from the roots
2417 MarkFromRootsClosure markFromRootsClosure(this, _span,
2418 verification_mark_bm(), verification_mark_stack(),
2419 false /* don't yield */, true /* verifying */);
2420 assert(_restart_addr == NULL, "Expected pre-condition");
2421 verification_mark_bm()->iterate(&markFromRootsClosure);
2422 while (_restart_addr != NULL) {
2423 // Deal with stack overflow: by restarting at the indicated
2424 // address.
2425 HeapWord* ra = _restart_addr;
2426 markFromRootsClosure.reset(ra);
2427 _restart_addr = NULL;
2428 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2429 }
2430 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2431 verify_work_stacks_empty();
2432
2433 // Marking completed -- now verify that each bit marked in
2434 // verification_mark_bm() is also marked in markBitMap(); flag all
2435 // errors by printing corresponding objects.
2436 VerifyMarkedClosure vcl(markBitMap());
2437 verification_mark_bm()->iterate(&vcl);
2438 if (vcl.failed()) {
2439 gclog_or_tty->print("Verification failed");
2440 gch->print_on(gclog_or_tty);
2441 fatal("CMS: failed marking verification after remark");
2442 }
2443 }
2444
2445 class VerifyKlassOopsKlassClosure : public KlassClosure {
2446 class VerifyKlassOopsClosure : public OopClosure {
2447 CMSBitMap* _bitmap;
2448 public:
2449 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2450 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2451 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2452 } _oop_closure;
2453 public:
2454 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2455 void do_klass(Klass* k) {
2456 k->oops_do(&_oop_closure);
2457 }
2458 };
2459
2460 void CMSCollector::verify_after_remark_work_2() {
2461 ResourceMark rm;
2462 HandleMark hm;
2463 GenCollectedHeap* gch = GenCollectedHeap::heap();
2464
2465 // Get a clear set of claim bits for the roots processing to work with.
2466 ClassLoaderDataGraph::clear_claimed_marks();
2467
2468 // Mark from roots one level into CMS
2469 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2470 markBitMap());
2471 CLDToOopClosure cld_closure(¬Older, true);
2472
2473 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2474
2475 {
2476 StrongRootsScope srs(1);
2477
2478 gch->gen_process_roots(&srs,
2479 GenCollectedHeap::OldGen,
2480 true, // young gen as roots
2481 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2482 should_unload_classes(),
2483 ¬Older,
2484 NULL,
2485 &cld_closure);
2486 }
2487
2488 // Now mark from the roots
2489 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2490 verification_mark_bm(), markBitMap(), verification_mark_stack());
2491 assert(_restart_addr == NULL, "Expected pre-condition");
2492 verification_mark_bm()->iterate(&markFromRootsClosure);
2493 while (_restart_addr != NULL) {
2494 // Deal with stack overflow: by restarting at the indicated
2495 // address.
2496 HeapWord* ra = _restart_addr;
2497 markFromRootsClosure.reset(ra);
2498 _restart_addr = NULL;
2499 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2500 }
2501 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2502 verify_work_stacks_empty();
2503
2504 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
2505 ClassLoaderDataGraph::classes_do(&verify_klass_oops);
2506
2507 // Marking completed -- now verify that each bit marked in
2508 // verification_mark_bm() is also marked in markBitMap(); flag all
2509 // errors by printing corresponding objects.
2510 VerifyMarkedClosure vcl(markBitMap());
2511 verification_mark_bm()->iterate(&vcl);
2512 assert(!vcl.failed(), "Else verification above should not have succeeded");
2513 }
2514
2515 void ConcurrentMarkSweepGeneration::save_marks() {
2516 // delegate to CMS space
2517 cmsSpace()->save_marks();
2518 for (uint i = 0; i < ParallelGCThreads; i++) {
2519 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2520 }
2521 }
2522
2523 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2524 return cmsSpace()->no_allocs_since_save_marks();
2525 }
2526
2527 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2528 \
2529 void ConcurrentMarkSweepGeneration:: \
2530 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2531 cl->set_generation(this); \
2532 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2533 cl->reset_generation(); \
2534 save_marks(); \
2535 }
2536
2537 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2538
2539 void
2540 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2541 if (freelistLock()->owned_by_self()) {
2542 Generation::oop_iterate(cl);
2543 } else {
2544 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2545 Generation::oop_iterate(cl);
2546 }
2547 }
2548
2549 void
2550 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2551 if (freelistLock()->owned_by_self()) {
2552 Generation::object_iterate(cl);
2553 } else {
2554 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2555 Generation::object_iterate(cl);
2556 }
2557 }
2558
2559 void
2560 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2561 if (freelistLock()->owned_by_self()) {
2562 Generation::safe_object_iterate(cl);
2563 } else {
2564 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2565 Generation::safe_object_iterate(cl);
2566 }
2567 }
2568
2569 void
2570 ConcurrentMarkSweepGeneration::post_compact() {
2571 }
2572
2573 void
2574 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2575 // Fix the linear allocation blocks to look like free blocks.
2576
2577 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2578 // are not called when the heap is verified during universe initialization and
2579 // at vm shutdown.
2580 if (freelistLock()->owned_by_self()) {
2581 cmsSpace()->prepare_for_verify();
2582 } else {
2583 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2584 cmsSpace()->prepare_for_verify();
2585 }
2586 }
2587
2588 void
2589 ConcurrentMarkSweepGeneration::verify() {
2590 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2591 // are not called when the heap is verified during universe initialization and
2592 // at vm shutdown.
2593 if (freelistLock()->owned_by_self()) {
2594 cmsSpace()->verify();
2595 } else {
2596 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2597 cmsSpace()->verify();
2598 }
2599 }
2600
2601 void CMSCollector::verify() {
2602 _cmsGen->verify();
2603 }
2604
2605 #ifndef PRODUCT
2606 bool CMSCollector::overflow_list_is_empty() const {
2607 assert(_num_par_pushes >= 0, "Inconsistency");
2608 if (_overflow_list == NULL) {
2609 assert(_num_par_pushes == 0, "Inconsistency");
2610 }
2611 return _overflow_list == NULL;
2612 }
2613
2614 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2615 // merely consolidate assertion checks that appear to occur together frequently.
2616 void CMSCollector::verify_work_stacks_empty() const {
2617 assert(_markStack.isEmpty(), "Marking stack should be empty");
2618 assert(overflow_list_is_empty(), "Overflow list should be empty");
2619 }
2620
2621 void CMSCollector::verify_overflow_empty() const {
2622 assert(overflow_list_is_empty(), "Overflow list should be empty");
2623 assert(no_preserved_marks(), "No preserved marks");
2624 }
2625 #endif // PRODUCT
2626
2627 // Decide if we want to enable class unloading as part of the
2628 // ensuing concurrent GC cycle. We will collect and
2629 // unload classes if it's the case that:
2630 // (1) an explicit gc request has been made and the flag
2631 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
2632 // (2) (a) class unloading is enabled at the command line, and
2633 // (b) old gen is getting really full
2634 // NOTE: Provided there is no change in the state of the heap between
2635 // calls to this method, it should have idempotent results. Moreover,
2636 // its results should be monotonically increasing (i.e. going from 0 to 1,
2637 // but not 1 to 0) between successive calls between which the heap was
2638 // not collected. For the implementation below, it must thus rely on
2639 // the property that concurrent_cycles_since_last_unload()
2640 // will not decrease unless a collection cycle happened and that
2641 // _cmsGen->is_too_full() are
2642 // themselves also monotonic in that sense. See check_monotonicity()
2643 // below.
2644 void CMSCollector::update_should_unload_classes() {
2645 _should_unload_classes = false;
2646 // Condition 1 above
2647 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
2648 _should_unload_classes = true;
2649 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
2650 // Disjuncts 2.b.(i,ii,iii) above
2651 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2652 CMSClassUnloadingMaxInterval)
2653 || _cmsGen->is_too_full();
2654 }
2655 }
2656
2657 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2658 bool res = should_concurrent_collect();
2659 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2660 return res;
2661 }
2662
2663 void CMSCollector::setup_cms_unloading_and_verification_state() {
2664 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2665 || VerifyBeforeExit;
2666 const int rso = GenCollectedHeap::SO_AllCodeCache;
2667
2668 // We set the proper root for this CMS cycle here.
2669 if (should_unload_classes()) { // Should unload classes this cycle
2670 remove_root_scanning_option(rso); // Shrink the root set appropriately
2671 set_verifying(should_verify); // Set verification state for this cycle
2672 return; // Nothing else needs to be done at this time
2673 }
2674
2675 // Not unloading classes this cycle
2676 assert(!should_unload_classes(), "Inconsistency!");
2677
2678 // If we are not unloading classes then add SO_AllCodeCache to root
2679 // scanning options.
2680 add_root_scanning_option(rso);
2681
2682 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2683 set_verifying(true);
2684 } else if (verifying() && !should_verify) {
2685 // We were verifying, but some verification flags got disabled.
2686 set_verifying(false);
2687 // Exclude symbols, strings and code cache elements from root scanning to
2688 // reduce IM and RM pauses.
2689 remove_root_scanning_option(rso);
2690 }
2691 }
2692
2693
2694 #ifndef PRODUCT
2695 HeapWord* CMSCollector::block_start(const void* p) const {
2696 const HeapWord* addr = (HeapWord*)p;
2697 if (_span.contains(p)) {
2698 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2699 return _cmsGen->cmsSpace()->block_start(p);
2700 }
2701 }
2702 return NULL;
2703 }
2704 #endif
2705
2706 HeapWord*
2707 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2708 bool tlab,
2709 bool parallel) {
2710 CMSSynchronousYieldRequest yr;
2711 assert(!tlab, "Can't deal with TLAB allocation");
2712 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2713 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2714 if (GCExpandToAllocateDelayMillis > 0) {
2715 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2716 }
2717 return have_lock_and_allocate(word_size, tlab);
2718 }
2719
2720 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2721 size_t bytes,
2722 size_t expand_bytes,
2723 CMSExpansionCause::Cause cause)
2724 {
2725
2726 bool success = expand(bytes, expand_bytes);
2727
2728 // remember why we expanded; this information is used
2729 // by shouldConcurrentCollect() when making decisions on whether to start
2730 // a new CMS cycle.
2731 if (success) {
2732 set_expansion_cause(cause);
2733 if (PrintGCDetails && Verbose) {
2734 gclog_or_tty->print_cr("Expanded CMS gen for %s",
2735 CMSExpansionCause::to_string(cause));
2736 }
2737 }
2738 }
2739
2740 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2741 HeapWord* res = NULL;
2742 MutexLocker x(ParGCRareEvent_lock);
2743 while (true) {
2744 // Expansion by some other thread might make alloc OK now:
2745 res = ps->lab.alloc(word_sz);
2746 if (res != NULL) return res;
2747 // If there's not enough expansion space available, give up.
2748 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2749 return NULL;
2750 }
2751 // Otherwise, we try expansion.
2752 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2753 // Now go around the loop and try alloc again;
2754 // A competing par_promote might beat us to the expansion space,
2755 // so we may go around the loop again if promotion fails again.
2756 if (GCExpandToAllocateDelayMillis > 0) {
2757 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2758 }
2759 }
2760 }
2761
2762
2763 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2764 PromotionInfo* promo) {
2765 MutexLocker x(ParGCRareEvent_lock);
2766 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2767 while (true) {
2768 // Expansion by some other thread might make alloc OK now:
2769 if (promo->ensure_spooling_space()) {
2770 assert(promo->has_spooling_space(),
2771 "Post-condition of successful ensure_spooling_space()");
2772 return true;
2773 }
2774 // If there's not enough expansion space available, give up.
2775 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2776 return false;
2777 }
2778 // Otherwise, we try expansion.
2779 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2780 // Now go around the loop and try alloc again;
2781 // A competing allocation might beat us to the expansion space,
2782 // so we may go around the loop again if allocation fails again.
2783 if (GCExpandToAllocateDelayMillis > 0) {
2784 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2785 }
2786 }
2787 }
2788
2789 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2790 // Only shrink if a compaction was done so that all the free space
2791 // in the generation is in a contiguous block at the end.
2792 if (did_compact()) {
2793 CardGeneration::shrink(bytes);
2794 }
2795 }
2796
2797 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2798 assert_locked_or_safepoint(Heap_lock);
2799 }
2800
2801 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2802 assert_locked_or_safepoint(Heap_lock);
2803 assert_lock_strong(freelistLock());
2804 if (PrintGCDetails && Verbose) {
2805 warning("Shrinking of CMS not yet implemented");
2806 }
2807 return;
2808 }
2809
2810
2811 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2812 // phases.
2813 class CMSPhaseAccounting: public StackObj {
2814 public:
2815 CMSPhaseAccounting(CMSCollector *collector,
2816 const char *phase,
2817 bool print_cr = true);
2818 ~CMSPhaseAccounting();
2819
2820 private:
2821 CMSCollector *_collector;
2822 const char *_phase;
2823 elapsedTimer _wallclock;
2824 bool _print_cr;
2825
2826 public:
2827 // Not MT-safe; so do not pass around these StackObj's
2828 // where they may be accessed by other threads.
2829 jlong wallclock_millis() {
2830 assert(_wallclock.is_active(), "Wall clock should not stop");
2831 _wallclock.stop(); // to record time
2832 jlong ret = _wallclock.milliseconds();
2833 _wallclock.start(); // restart
2834 return ret;
2835 }
2836 };
2837
2838 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2839 const char *phase,
2840 bool print_cr) :
2841 _collector(collector), _phase(phase), _print_cr(print_cr) {
2842
2843 if (PrintCMSStatistics != 0) {
2844 _collector->resetYields();
2845 }
2846 if (PrintGCDetails) {
2847 gclog_or_tty->gclog_stamp();
2848 gclog_or_tty->print_cr("[%s-concurrent-%s-start]",
2849 _collector->cmsGen()->short_name(), _phase);
2850 }
2851 _collector->resetTimer();
2852 _wallclock.start();
2853 _collector->startTimer();
2854 }
2855
2856 CMSPhaseAccounting::~CMSPhaseAccounting() {
2857 assert(_wallclock.is_active(), "Wall clock should not have stopped");
2858 _collector->stopTimer();
2859 _wallclock.stop();
2860 if (PrintGCDetails) {
2861 gclog_or_tty->gclog_stamp();
2862 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
2863 _collector->cmsGen()->short_name(),
2864 _phase, _collector->timerValue(), _wallclock.seconds());
2865 if (_print_cr) {
2866 gclog_or_tty->cr();
2867 }
2868 if (PrintCMSStatistics != 0) {
2869 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
2870 _collector->yields());
2871 }
2872 }
2873 }
2874
2875 // CMS work
2876
2877 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2878 class CMSParMarkTask : public AbstractGangTask {
2879 protected:
2880 CMSCollector* _collector;
2881 uint _n_workers;
2882 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2883 AbstractGangTask(name),
2884 _collector(collector),
2885 _n_workers(n_workers) {}
2886 // Work method in support of parallel rescan ... of young gen spaces
2887 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
2888 ContiguousSpace* space,
2889 HeapWord** chunk_array, size_t chunk_top);
2890 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
2891 };
2892
2893 // Parallel initial mark task
2894 class CMSParInitialMarkTask: public CMSParMarkTask {
2895 StrongRootsScope* _strong_roots_scope;
2896 public:
2897 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) :
2898 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers),
2899 _strong_roots_scope(strong_roots_scope) {}
2900 void work(uint worker_id);
2901 };
2902
2903 // Checkpoint the roots into this generation from outside
2904 // this generation. [Note this initial checkpoint need only
2905 // be approximate -- we'll do a catch up phase subsequently.]
2906 void CMSCollector::checkpointRootsInitial() {
2907 assert(_collectorState == InitialMarking, "Wrong collector state");
2908 check_correct_thread_executing();
2909 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
2910
2911 save_heap_summary();
2912 report_heap_summary(GCWhen::BeforeGC);
2913
2914 ReferenceProcessor* rp = ref_processor();
2915 assert(_restart_addr == NULL, "Control point invariant");
2916 {
2917 // acquire locks for subsequent manipulations
2918 MutexLockerEx x(bitMapLock(),
2919 Mutex::_no_safepoint_check_flag);
2920 checkpointRootsInitialWork();
2921 // enable ("weak") refs discovery
2922 rp->enable_discovery();
2923 _collectorState = Marking;
2924 }
2925 }
2926
2927 void CMSCollector::checkpointRootsInitialWork() {
2928 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2929 assert(_collectorState == InitialMarking, "just checking");
2930
2931 // Already have locks.
2932 assert_lock_strong(bitMapLock());
2933 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2934
2935 // Setup the verification and class unloading state for this
2936 // CMS collection cycle.
2937 setup_cms_unloading_and_verification_state();
2938
2939 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
2940 PrintGCDetails && Verbose, true, _gc_timer_cm);)
2941
2942 // Reset all the PLAB chunk arrays if necessary.
2943 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2944 reset_survivor_plab_arrays();
2945 }
2946
2947 ResourceMark rm;
2948 HandleMark hm;
2949
2950 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2951 GenCollectedHeap* gch = GenCollectedHeap::heap();
2952
2953 verify_work_stacks_empty();
2954 verify_overflow_empty();
2955
2956 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2957 // Update the saved marks which may affect the root scans.
2958 gch->save_marks();
2959
2960 // weak reference processing has not started yet.
2961 ref_processor()->set_enqueuing_is_done(false);
2962
2963 // Need to remember all newly created CLDs,
2964 // so that we can guarantee that the remark finds them.
2965 ClassLoaderDataGraph::remember_new_clds(true);
2966
2967 // Whenever a CLD is found, it will be claimed before proceeding to mark
2968 // the klasses. The claimed marks need to be cleared before marking starts.
2969 ClassLoaderDataGraph::clear_claimed_marks();
2970
2971 if (CMSPrintEdenSurvivorChunks) {
2972 print_eden_and_survivor_chunk_arrays();
2973 }
2974
2975 {
2976 #if defined(COMPILER2) || INCLUDE_JVMCI
2977 DerivedPointerTableDeactivate dpt_deact;
2978 #endif
2979 if (CMSParallelInitialMarkEnabled) {
2980 // The parallel version.
2981 WorkGang* workers = gch->workers();
2982 assert(workers != NULL, "Need parallel worker threads.");
2983 uint n_workers = workers->active_workers();
2984
2985 StrongRootsScope srs(n_workers);
2986
2987 CMSParInitialMarkTask tsk(this, &srs, n_workers);
2988 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
2989 if (n_workers > 1) {
2990 workers->run_task(&tsk);
2991 } else {
2992 tsk.work(0);
2993 }
2994 } else {
2995 // The serial version.
2996 CLDToOopClosure cld_closure(¬Older, true);
2997 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2998
2999 StrongRootsScope srs(1);
3000
3001 gch->gen_process_roots(&srs,
3002 GenCollectedHeap::OldGen,
3003 true, // young gen as roots
3004 GenCollectedHeap::ScanningOption(roots_scanning_options()),
3005 should_unload_classes(),
3006 ¬Older,
3007 NULL,
3008 &cld_closure);
3009 }
3010 }
3011
3012 // Clear mod-union table; it will be dirtied in the prologue of
3013 // CMS generation per each young generation collection.
3014
3015 assert(_modUnionTable.isAllClear(),
3016 "Was cleared in most recent final checkpoint phase"
3017 " or no bits are set in the gc_prologue before the start of the next "
3018 "subsequent marking phase.");
3019
3020 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
3021
3022 // Save the end of the used_region of the constituent generations
3023 // to be used to limit the extent of sweep in each generation.
3024 save_sweep_limits();
3025 verify_overflow_empty();
3026 }
3027
3028 bool CMSCollector::markFromRoots() {
3029 // we might be tempted to assert that:
3030 // assert(!SafepointSynchronize::is_at_safepoint(),
3031 // "inconsistent argument?");
3032 // However that wouldn't be right, because it's possible that
3033 // a safepoint is indeed in progress as a young generation
3034 // stop-the-world GC happens even as we mark in this generation.
3035 assert(_collectorState == Marking, "inconsistent state?");
3036 check_correct_thread_executing();
3037 verify_overflow_empty();
3038
3039 // Weak ref discovery note: We may be discovering weak
3040 // refs in this generation concurrent (but interleaved) with
3041 // weak ref discovery by the young generation collector.
3042
3043 CMSTokenSyncWithLocks ts(true, bitMapLock());
3044 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3045 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3046 bool res = markFromRootsWork();
3047 if (res) {
3048 _collectorState = Precleaning;
3049 } else { // We failed and a foreground collection wants to take over
3050 assert(_foregroundGCIsActive, "internal state inconsistency");
3051 assert(_restart_addr == NULL, "foreground will restart from scratch");
3052 if (PrintGCDetails) {
3053 gclog_or_tty->print_cr("bailing out to foreground collection");
3054 }
3055 }
3056 verify_overflow_empty();
3057 return res;
3058 }
3059
3060 bool CMSCollector::markFromRootsWork() {
3061 // iterate over marked bits in bit map, doing a full scan and mark
3062 // from these roots using the following algorithm:
3063 // . if oop is to the right of the current scan pointer,
3064 // mark corresponding bit (we'll process it later)
3065 // . else (oop is to left of current scan pointer)
3066 // push oop on marking stack
3067 // . drain the marking stack
3068
3069 // Note that when we do a marking step we need to hold the
3070 // bit map lock -- recall that direct allocation (by mutators)
3071 // and promotion (by the young generation collector) is also
3072 // marking the bit map. [the so-called allocate live policy.]
3073 // Because the implementation of bit map marking is not
3074 // robust wrt simultaneous marking of bits in the same word,
3075 // we need to make sure that there is no such interference
3076 // between concurrent such updates.
3077
3078 // already have locks
3079 assert_lock_strong(bitMapLock());
3080
3081 verify_work_stacks_empty();
3082 verify_overflow_empty();
3083 bool result = false;
3084 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3085 result = do_marking_mt();
3086 } else {
3087 result = do_marking_st();
3088 }
3089 return result;
3090 }
3091
3092 // Forward decl
3093 class CMSConcMarkingTask;
3094
3095 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3096 CMSCollector* _collector;
3097 CMSConcMarkingTask* _task;
3098 public:
3099 virtual void yield();
3100
3101 // "n_threads" is the number of threads to be terminated.
3102 // "queue_set" is a set of work queues of other threads.
3103 // "collector" is the CMS collector associated with this task terminator.
3104 // "yield" indicates whether we need the gang as a whole to yield.
3105 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3106 ParallelTaskTerminator(n_threads, queue_set),
3107 _collector(collector) { }
3108
3109 void set_task(CMSConcMarkingTask* task) {
3110 _task = task;
3111 }
3112 };
3113
3114 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3115 CMSConcMarkingTask* _task;
3116 public:
3117 bool should_exit_termination();
3118 void set_task(CMSConcMarkingTask* task) {
3119 _task = task;
3120 }
3121 };
3122
3123 // MT Concurrent Marking Task
3124 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3125 CMSCollector* _collector;
3126 uint _n_workers; // requested/desired # workers
3127 bool _result;
3128 CompactibleFreeListSpace* _cms_space;
3129 char _pad_front[64]; // padding to ...
3130 HeapWord* _global_finger; // ... avoid sharing cache line
3131 char _pad_back[64];
3132 HeapWord* _restart_addr;
3133
3134 // Exposed here for yielding support
3135 Mutex* const _bit_map_lock;
3136
3137 // The per thread work queues, available here for stealing
3138 OopTaskQueueSet* _task_queues;
3139
3140 // Termination (and yielding) support
3141 CMSConcMarkingTerminator _term;
3142 CMSConcMarkingTerminatorTerminator _term_term;
3143
3144 public:
3145 CMSConcMarkingTask(CMSCollector* collector,
3146 CompactibleFreeListSpace* cms_space,
3147 YieldingFlexibleWorkGang* workers,
3148 OopTaskQueueSet* task_queues):
3149 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3150 _collector(collector),
3151 _cms_space(cms_space),
3152 _n_workers(0), _result(true),
3153 _task_queues(task_queues),
3154 _term(_n_workers, task_queues, _collector),
3155 _bit_map_lock(collector->bitMapLock())
3156 {
3157 _requested_size = _n_workers;
3158 _term.set_task(this);
3159 _term_term.set_task(this);
3160 _restart_addr = _global_finger = _cms_space->bottom();
3161 }
3162
3163
3164 OopTaskQueueSet* task_queues() { return _task_queues; }
3165
3166 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3167
3168 HeapWord** global_finger_addr() { return &_global_finger; }
3169
3170 CMSConcMarkingTerminator* terminator() { return &_term; }
3171
3172 virtual void set_for_termination(uint active_workers) {
3173 terminator()->reset_for_reuse(active_workers);
3174 }
3175
3176 void work(uint worker_id);
3177 bool should_yield() {
3178 return ConcurrentMarkSweepThread::should_yield()
3179 && !_collector->foregroundGCIsActive();
3180 }
3181
3182 virtual void coordinator_yield(); // stuff done by coordinator
3183 bool result() { return _result; }
3184
3185 void reset(HeapWord* ra) {
3186 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3187 _restart_addr = _global_finger = ra;
3188 _term.reset_for_reuse();
3189 }
3190
3191 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3192 OopTaskQueue* work_q);
3193
3194 private:
3195 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3196 void do_work_steal(int i);
3197 void bump_global_finger(HeapWord* f);
3198 };
3199
3200 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3201 assert(_task != NULL, "Error");
3202 return _task->yielding();
3203 // Note that we do not need the disjunct || _task->should_yield() above
3204 // because we want terminating threads to yield only if the task
3205 // is already in the midst of yielding, which happens only after at least one
3206 // thread has yielded.
3207 }
3208
3209 void CMSConcMarkingTerminator::yield() {
3210 if (_task->should_yield()) {
3211 _task->yield();
3212 } else {
3213 ParallelTaskTerminator::yield();
3214 }
3215 }
3216
3217 ////////////////////////////////////////////////////////////////
3218 // Concurrent Marking Algorithm Sketch
3219 ////////////////////////////////////////////////////////////////
3220 // Until all tasks exhausted (both spaces):
3221 // -- claim next available chunk
3222 // -- bump global finger via CAS
3223 // -- find first object that starts in this chunk
3224 // and start scanning bitmap from that position
3225 // -- scan marked objects for oops
3226 // -- CAS-mark target, and if successful:
3227 // . if target oop is above global finger (volatile read)
3228 // nothing to do
3229 // . if target oop is in chunk and above local finger
3230 // then nothing to do
3231 // . else push on work-queue
3232 // -- Deal with possible overflow issues:
3233 // . local work-queue overflow causes stuff to be pushed on
3234 // global (common) overflow queue
3235 // . always first empty local work queue
3236 // . then get a batch of oops from global work queue if any
3237 // . then do work stealing
3238 // -- When all tasks claimed (both spaces)
3239 // and local work queue empty,
3240 // then in a loop do:
3241 // . check global overflow stack; steal a batch of oops and trace
3242 // . try to steal from other threads oif GOS is empty
3243 // . if neither is available, offer termination
3244 // -- Terminate and return result
3245 //
3246 void CMSConcMarkingTask::work(uint worker_id) {
3247 elapsedTimer _timer;
3248 ResourceMark rm;
3249 HandleMark hm;
3250
3251 DEBUG_ONLY(_collector->verify_overflow_empty();)
3252
3253 // Before we begin work, our work queue should be empty
3254 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3255 // Scan the bitmap covering _cms_space, tracing through grey objects.
3256 _timer.start();
3257 do_scan_and_mark(worker_id, _cms_space);
3258 _timer.stop();
3259 if (PrintCMSStatistics != 0) {
3260 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3261 worker_id, _timer.seconds());
3262 // XXX: need xxx/xxx type of notation, two timers
3263 }
3264
3265 // ... do work stealing
3266 _timer.reset();
3267 _timer.start();
3268 do_work_steal(worker_id);
3269 _timer.stop();
3270 if (PrintCMSStatistics != 0) {
3271 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3272 worker_id, _timer.seconds());
3273 // XXX: need xxx/xxx type of notation, two timers
3274 }
3275 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3276 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3277 // Note that under the current task protocol, the
3278 // following assertion is true even of the spaces
3279 // expanded since the completion of the concurrent
3280 // marking. XXX This will likely change under a strict
3281 // ABORT semantics.
3282 // After perm removal the comparison was changed to
3283 // greater than or equal to from strictly greater than.
3284 // Before perm removal the highest address sweep would
3285 // have been at the end of perm gen but now is at the
3286 // end of the tenured gen.
3287 assert(_global_finger >= _cms_space->end(),
3288 "All tasks have been completed");
3289 DEBUG_ONLY(_collector->verify_overflow_empty();)
3290 }
3291
3292 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3293 HeapWord* read = _global_finger;
3294 HeapWord* cur = read;
3295 while (f > read) {
3296 cur = read;
3297 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3298 if (cur == read) {
3299 // our cas succeeded
3300 assert(_global_finger >= f, "protocol consistency");
3301 break;
3302 }
3303 }
3304 }
3305
3306 // This is really inefficient, and should be redone by
3307 // using (not yet available) block-read and -write interfaces to the
3308 // stack and the work_queue. XXX FIX ME !!!
3309 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3310 OopTaskQueue* work_q) {
3311 // Fast lock-free check
3312 if (ovflw_stk->length() == 0) {
3313 return false;
3314 }
3315 assert(work_q->size() == 0, "Shouldn't steal");
3316 MutexLockerEx ml(ovflw_stk->par_lock(),
3317 Mutex::_no_safepoint_check_flag);
3318 // Grab up to 1/4 the size of the work queue
3319 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3320 (size_t)ParGCDesiredObjsFromOverflowList);
3321 num = MIN2(num, ovflw_stk->length());
3322 for (int i = (int) num; i > 0; i--) {
3323 oop cur = ovflw_stk->pop();
3324 assert(cur != NULL, "Counted wrong?");
3325 work_q->push(cur);
3326 }
3327 return num > 0;
3328 }
3329
3330 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3331 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3332 int n_tasks = pst->n_tasks();
3333 // We allow that there may be no tasks to do here because
3334 // we are restarting after a stack overflow.
3335 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3336 uint nth_task = 0;
3337
3338 HeapWord* aligned_start = sp->bottom();
3339 if (sp->used_region().contains(_restart_addr)) {
3340 // Align down to a card boundary for the start of 0th task
3341 // for this space.
3342 aligned_start =
3343 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3344 CardTableModRefBS::card_size);
3345 }
3346
3347 size_t chunk_size = sp->marking_task_size();
3348 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3349 // Having claimed the nth task in this space,
3350 // compute the chunk that it corresponds to:
3351 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3352 aligned_start + (nth_task+1)*chunk_size);
3353 // Try and bump the global finger via a CAS;
3354 // note that we need to do the global finger bump
3355 // _before_ taking the intersection below, because
3356 // the task corresponding to that region will be
3357 // deemed done even if the used_region() expands
3358 // because of allocation -- as it almost certainly will
3359 // during start-up while the threads yield in the
3360 // closure below.
3361 HeapWord* finger = span.end();
3362 bump_global_finger(finger); // atomically
3363 // There are null tasks here corresponding to chunks
3364 // beyond the "top" address of the space.
3365 span = span.intersection(sp->used_region());
3366 if (!span.is_empty()) { // Non-null task
3367 HeapWord* prev_obj;
3368 assert(!span.contains(_restart_addr) || nth_task == 0,
3369 "Inconsistency");
3370 if (nth_task == 0) {
3371 // For the 0th task, we'll not need to compute a block_start.
3372 if (span.contains(_restart_addr)) {
3373 // In the case of a restart because of stack overflow,
3374 // we might additionally skip a chunk prefix.
3375 prev_obj = _restart_addr;
3376 } else {
3377 prev_obj = span.start();
3378 }
3379 } else {
3380 // We want to skip the first object because
3381 // the protocol is to scan any object in its entirety
3382 // that _starts_ in this span; a fortiori, any
3383 // object starting in an earlier span is scanned
3384 // as part of an earlier claimed task.
3385 // Below we use the "careful" version of block_start
3386 // so we do not try to navigate uninitialized objects.
3387 prev_obj = sp->block_start_careful(span.start());
3388 // Below we use a variant of block_size that uses the
3389 // Printezis bits to avoid waiting for allocated
3390 // objects to become initialized/parsable.
3391 while (prev_obj < span.start()) {
3392 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3393 if (sz > 0) {
3394 prev_obj += sz;
3395 } else {
3396 // In this case we may end up doing a bit of redundant
3397 // scanning, but that appears unavoidable, short of
3398 // locking the free list locks; see bug 6324141.
3399 break;
3400 }
3401 }
3402 }
3403 if (prev_obj < span.end()) {
3404 MemRegion my_span = MemRegion(prev_obj, span.end());
3405 // Do the marking work within a non-empty span --
3406 // the last argument to the constructor indicates whether the
3407 // iteration should be incremental with periodic yields.
3408 Par_MarkFromRootsClosure cl(this, _collector, my_span,
3409 &_collector->_markBitMap,
3410 work_queue(i),
3411 &_collector->_markStack);
3412 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3413 } // else nothing to do for this task
3414 } // else nothing to do for this task
3415 }
3416 // We'd be tempted to assert here that since there are no
3417 // more tasks left to claim in this space, the global_finger
3418 // must exceed space->top() and a fortiori space->end(). However,
3419 // that would not quite be correct because the bumping of
3420 // global_finger occurs strictly after the claiming of a task,
3421 // so by the time we reach here the global finger may not yet
3422 // have been bumped up by the thread that claimed the last
3423 // task.
3424 pst->all_tasks_completed();
3425 }
3426
3427 class Par_ConcMarkingClosure: public MetadataAwareOopClosure {
3428 private:
3429 CMSCollector* _collector;
3430 CMSConcMarkingTask* _task;
3431 MemRegion _span;
3432 CMSBitMap* _bit_map;
3433 CMSMarkStack* _overflow_stack;
3434 OopTaskQueue* _work_queue;
3435 protected:
3436 DO_OOP_WORK_DEFN
3437 public:
3438 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3439 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3440 MetadataAwareOopClosure(collector->ref_processor()),
3441 _collector(collector),
3442 _task(task),
3443 _span(collector->_span),
3444 _work_queue(work_queue),
3445 _bit_map(bit_map),
3446 _overflow_stack(overflow_stack)
3447 { }
3448 virtual void do_oop(oop* p);
3449 virtual void do_oop(narrowOop* p);
3450
3451 void trim_queue(size_t max);
3452 void handle_stack_overflow(HeapWord* lost);
3453 void do_yield_check() {
3454 if (_task->should_yield()) {
3455 _task->yield();
3456 }
3457 }
3458 };
3459
3460 // Grey object scanning during work stealing phase --
3461 // the salient assumption here is that any references
3462 // that are in these stolen objects being scanned must
3463 // already have been initialized (else they would not have
3464 // been published), so we do not need to check for
3465 // uninitialized objects before pushing here.
3466 void Par_ConcMarkingClosure::do_oop(oop obj) {
3467 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
3468 HeapWord* addr = (HeapWord*)obj;
3469 // Check if oop points into the CMS generation
3470 // and is not marked
3471 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3472 // a white object ...
3473 // If we manage to "claim" the object, by being the
3474 // first thread to mark it, then we push it on our
3475 // marking stack
3476 if (_bit_map->par_mark(addr)) { // ... now grey
3477 // push on work queue (grey set)
3478 bool simulate_overflow = false;
3479 NOT_PRODUCT(
3480 if (CMSMarkStackOverflowALot &&
3481 _collector->simulate_overflow()) {
3482 // simulate a stack overflow
3483 simulate_overflow = true;
3484 }
3485 )
3486 if (simulate_overflow ||
3487 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3488 // stack overflow
3489 if (PrintCMSStatistics != 0) {
3490 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
3491 SIZE_FORMAT, _overflow_stack->capacity());
3492 }
3493 // We cannot assert that the overflow stack is full because
3494 // it may have been emptied since.
3495 assert(simulate_overflow ||
3496 _work_queue->size() == _work_queue->max_elems(),
3497 "Else push should have succeeded");
3498 handle_stack_overflow(addr);
3499 }
3500 } // Else, some other thread got there first
3501 do_yield_check();
3502 }
3503 }
3504
3505 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3506 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3507
3508 void Par_ConcMarkingClosure::trim_queue(size_t max) {
3509 while (_work_queue->size() > max) {
3510 oop new_oop;
3511 if (_work_queue->pop_local(new_oop)) {
3512 assert(new_oop->is_oop(), "Should be an oop");
3513 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3514 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3515 new_oop->oop_iterate(this); // do_oop() above
3516 do_yield_check();
3517 }
3518 }
3519 }
3520
3521 // Upon stack overflow, we discard (part of) the stack,
3522 // remembering the least address amongst those discarded
3523 // in CMSCollector's _restart_address.
3524 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3525 // We need to do this under a mutex to prevent other
3526 // workers from interfering with the work done below.
3527 MutexLockerEx ml(_overflow_stack->par_lock(),
3528 Mutex::_no_safepoint_check_flag);
3529 // Remember the least grey address discarded
3530 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3531 _collector->lower_restart_addr(ra);
3532 _overflow_stack->reset(); // discard stack contents
3533 _overflow_stack->expand(); // expand the stack if possible
3534 }
3535
3536
3537 void CMSConcMarkingTask::do_work_steal(int i) {
3538 OopTaskQueue* work_q = work_queue(i);
3539 oop obj_to_scan;
3540 CMSBitMap* bm = &(_collector->_markBitMap);
3541 CMSMarkStack* ovflw = &(_collector->_markStack);
3542 int* seed = _collector->hash_seed(i);
3543 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3544 while (true) {
3545 cl.trim_queue(0);
3546 assert(work_q->size() == 0, "Should have been emptied above");
3547 if (get_work_from_overflow_stack(ovflw, work_q)) {
3548 // Can't assert below because the work obtained from the
3549 // overflow stack may already have been stolen from us.
3550 // assert(work_q->size() > 0, "Work from overflow stack");
3551 continue;
3552 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3553 assert(obj_to_scan->is_oop(), "Should be an oop");
3554 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3555 obj_to_scan->oop_iterate(&cl);
3556 } else if (terminator()->offer_termination(&_term_term)) {
3557 assert(work_q->size() == 0, "Impossible!");
3558 break;
3559 } else if (yielding() || should_yield()) {
3560 yield();
3561 }
3562 }
3563 }
3564
3565 // This is run by the CMS (coordinator) thread.
3566 void CMSConcMarkingTask::coordinator_yield() {
3567 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3568 "CMS thread should hold CMS token");
3569 // First give up the locks, then yield, then re-lock
3570 // We should probably use a constructor/destructor idiom to
3571 // do this unlock/lock or modify the MutexUnlocker class to
3572 // serve our purpose. XXX
3573 assert_lock_strong(_bit_map_lock);
3574 _bit_map_lock->unlock();
3575 ConcurrentMarkSweepThread::desynchronize(true);
3576 _collector->stopTimer();
3577 if (PrintCMSStatistics != 0) {
3578 _collector->incrementYields();
3579 }
3580
3581 // It is possible for whichever thread initiated the yield request
3582 // not to get a chance to wake up and take the bitmap lock between
3583 // this thread releasing it and reacquiring it. So, while the
3584 // should_yield() flag is on, let's sleep for a bit to give the
3585 // other thread a chance to wake up. The limit imposed on the number
3586 // of iterations is defensive, to avoid any unforseen circumstances
3587 // putting us into an infinite loop. Since it's always been this
3588 // (coordinator_yield()) method that was observed to cause the
3589 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3590 // which is by default non-zero. For the other seven methods that
3591 // also perform the yield operation, as are using a different
3592 // parameter (CMSYieldSleepCount) which is by default zero. This way we
3593 // can enable the sleeping for those methods too, if necessary.
3594 // See 6442774.
3595 //
3596 // We really need to reconsider the synchronization between the GC
3597 // thread and the yield-requesting threads in the future and we
3598 // should really use wait/notify, which is the recommended
3599 // way of doing this type of interaction. Additionally, we should
3600 // consolidate the eight methods that do the yield operation and they
3601 // are almost identical into one for better maintainability and
3602 // readability. See 6445193.
3603 //
3604 // Tony 2006.06.29
3605 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3606 ConcurrentMarkSweepThread::should_yield() &&
3607 !CMSCollector::foregroundGCIsActive(); ++i) {
3608 os::sleep(Thread::current(), 1, false);
3609 }
3610
3611 ConcurrentMarkSweepThread::synchronize(true);
3612 _bit_map_lock->lock_without_safepoint_check();
3613 _collector->startTimer();
3614 }
3615
3616 bool CMSCollector::do_marking_mt() {
3617 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3618 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3619 conc_workers()->active_workers(),
3620 Threads::number_of_non_daemon_threads());
3621 conc_workers()->set_active_workers(num_workers);
3622
3623 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
3624
3625 CMSConcMarkingTask tsk(this,
3626 cms_space,
3627 conc_workers(),
3628 task_queues());
3629
3630 // Since the actual number of workers we get may be different
3631 // from the number we requested above, do we need to do anything different
3632 // below? In particular, may be we need to subclass the SequantialSubTasksDone
3633 // class?? XXX
3634 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3635
3636 // Refs discovery is already non-atomic.
3637 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3638 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3639 conc_workers()->start_task(&tsk);
3640 while (tsk.yielded()) {
3641 tsk.coordinator_yield();
3642 conc_workers()->continue_task(&tsk);
3643 }
3644 // If the task was aborted, _restart_addr will be non-NULL
3645 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3646 while (_restart_addr != NULL) {
3647 // XXX For now we do not make use of ABORTED state and have not
3648 // yet implemented the right abort semantics (even in the original
3649 // single-threaded CMS case). That needs some more investigation
3650 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3651 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3652 // If _restart_addr is non-NULL, a marking stack overflow
3653 // occurred; we need to do a fresh marking iteration from the
3654 // indicated restart address.
3655 if (_foregroundGCIsActive) {
3656 // We may be running into repeated stack overflows, having
3657 // reached the limit of the stack size, while making very
3658 // slow forward progress. It may be best to bail out and
3659 // let the foreground collector do its job.
3660 // Clear _restart_addr, so that foreground GC
3661 // works from scratch. This avoids the headache of
3662 // a "rescan" which would otherwise be needed because
3663 // of the dirty mod union table & card table.
3664 _restart_addr = NULL;
3665 return false;
3666 }
3667 // Adjust the task to restart from _restart_addr
3668 tsk.reset(_restart_addr);
3669 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3670 _restart_addr);
3671 _restart_addr = NULL;
3672 // Get the workers going again
3673 conc_workers()->start_task(&tsk);
3674 while (tsk.yielded()) {
3675 tsk.coordinator_yield();
3676 conc_workers()->continue_task(&tsk);
3677 }
3678 }
3679 assert(tsk.completed(), "Inconsistency");
3680 assert(tsk.result() == true, "Inconsistency");
3681 return true;
3682 }
3683
3684 bool CMSCollector::do_marking_st() {
3685 ResourceMark rm;
3686 HandleMark hm;
3687
3688 // Temporarily make refs discovery single threaded (non-MT)
3689 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3690 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3691 &_markStack, CMSYield);
3692 // the last argument to iterate indicates whether the iteration
3693 // should be incremental with periodic yields.
3694 _markBitMap.iterate(&markFromRootsClosure);
3695 // If _restart_addr is non-NULL, a marking stack overflow
3696 // occurred; we need to do a fresh iteration from the
3697 // indicated restart address.
3698 while (_restart_addr != NULL) {
3699 if (_foregroundGCIsActive) {
3700 // We may be running into repeated stack overflows, having
3701 // reached the limit of the stack size, while making very
3702 // slow forward progress. It may be best to bail out and
3703 // let the foreground collector do its job.
3704 // Clear _restart_addr, so that foreground GC
3705 // works from scratch. This avoids the headache of
3706 // a "rescan" which would otherwise be needed because
3707 // of the dirty mod union table & card table.
3708 _restart_addr = NULL;
3709 return false; // indicating failure to complete marking
3710 }
3711 // Deal with stack overflow:
3712 // we restart marking from _restart_addr
3713 HeapWord* ra = _restart_addr;
3714 markFromRootsClosure.reset(ra);
3715 _restart_addr = NULL;
3716 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3717 }
3718 return true;
3719 }
3720
3721 void CMSCollector::preclean() {
3722 check_correct_thread_executing();
3723 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3724 verify_work_stacks_empty();
3725 verify_overflow_empty();
3726 _abort_preclean = false;
3727 if (CMSPrecleaningEnabled) {
3728 if (!CMSEdenChunksRecordAlways) {
3729 _eden_chunk_index = 0;
3730 }
3731 size_t used = get_eden_used();
3732 size_t capacity = get_eden_capacity();
3733 // Don't start sampling unless we will get sufficiently
3734 // many samples.
3735 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
3736 * CMSScheduleRemarkEdenPenetration)) {
3737 _start_sampling = true;
3738 } else {
3739 _start_sampling = false;
3740 }
3741 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3742 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
3743 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3744 }
3745 CMSTokenSync x(true); // is cms thread
3746 if (CMSPrecleaningEnabled) {
3747 sample_eden();
3748 _collectorState = AbortablePreclean;
3749 } else {
3750 _collectorState = FinalMarking;
3751 }
3752 verify_work_stacks_empty();
3753 verify_overflow_empty();
3754 }
3755
3756 // Try and schedule the remark such that young gen
3757 // occupancy is CMSScheduleRemarkEdenPenetration %.
3758 void CMSCollector::abortable_preclean() {
3759 check_correct_thread_executing();
3760 assert(CMSPrecleaningEnabled, "Inconsistent control state");
3761 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3762
3763 // If Eden's current occupancy is below this threshold,
3764 // immediately schedule the remark; else preclean
3765 // past the next scavenge in an effort to
3766 // schedule the pause as described above. By choosing
3767 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3768 // we will never do an actual abortable preclean cycle.
3769 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3770 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3771 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
3772 // We need more smarts in the abortable preclean
3773 // loop below to deal with cases where allocation
3774 // in young gen is very very slow, and our precleaning
3775 // is running a losing race against a horde of
3776 // mutators intent on flooding us with CMS updates
3777 // (dirty cards).
3778 // One, admittedly dumb, strategy is to give up
3779 // after a certain number of abortable precleaning loops
3780 // or after a certain maximum time. We want to make
3781 // this smarter in the next iteration.
3782 // XXX FIX ME!!! YSR
3783 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3784 while (!(should_abort_preclean() ||
3785 ConcurrentMarkSweepThread::should_terminate())) {
3786 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3787 cumworkdone += workdone;
3788 loops++;
3789 // Voluntarily terminate abortable preclean phase if we have
3790 // been at it for too long.
3791 if ((CMSMaxAbortablePrecleanLoops != 0) &&
3792 loops >= CMSMaxAbortablePrecleanLoops) {
3793 if (PrintGCDetails) {
3794 gclog_or_tty->print(" CMS: abort preclean due to loops ");
3795 }
3796 break;
3797 }
3798 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3799 if (PrintGCDetails) {
3800 gclog_or_tty->print(" CMS: abort preclean due to time ");
3801 }
3802 break;
3803 }
3804 // If we are doing little work each iteration, we should
3805 // take a short break.
3806 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3807 // Sleep for some time, waiting for work to accumulate
3808 stopTimer();
3809 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3810 startTimer();
3811 waited++;
3812 }
3813 }
3814 if (PrintCMSStatistics > 0) {
3815 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3816 loops, waited, cumworkdone);
3817 }
3818 }
3819 CMSTokenSync x(true); // is cms thread
3820 if (_collectorState != Idling) {
3821 assert(_collectorState == AbortablePreclean,
3822 "Spontaneous state transition?");
3823 _collectorState = FinalMarking;
3824 } // Else, a foreground collection completed this CMS cycle.
3825 return;
3826 }
3827
3828 // Respond to an Eden sampling opportunity
3829 void CMSCollector::sample_eden() {
3830 // Make sure a young gc cannot sneak in between our
3831 // reading and recording of a sample.
3832 assert(Thread::current()->is_ConcurrentGC_thread(),
3833 "Only the cms thread may collect Eden samples");
3834 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3835 "Should collect samples while holding CMS token");
3836 if (!_start_sampling) {
3837 return;
3838 }
3839 // When CMSEdenChunksRecordAlways is true, the eden chunk array
3840 // is populated by the young generation.
3841 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3842 if (_eden_chunk_index < _eden_chunk_capacity) {
3843 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
3844 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3845 "Unexpected state of Eden");
3846 // We'd like to check that what we just sampled is an oop-start address;
3847 // however, we cannot do that here since the object may not yet have been
3848 // initialized. So we'll instead do the check when we _use_ this sample
3849 // later.
3850 if (_eden_chunk_index == 0 ||
3851 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3852 _eden_chunk_array[_eden_chunk_index-1])
3853 >= CMSSamplingGrain)) {
3854 _eden_chunk_index++; // commit sample
3855 }
3856 }
3857 }
3858 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3859 size_t used = get_eden_used();
3860 size_t capacity = get_eden_capacity();
3861 assert(used <= capacity, "Unexpected state of Eden");
3862 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3863 _abort_preclean = true;
3864 }
3865 }
3866 }
3867
3868
3869 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3870 assert(_collectorState == Precleaning ||
3871 _collectorState == AbortablePreclean, "incorrect state");
3872 ResourceMark rm;
3873 HandleMark hm;
3874
3875 // Precleaning is currently not MT but the reference processor
3876 // may be set for MT. Disable it temporarily here.
3877 ReferenceProcessor* rp = ref_processor();
3878 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3879
3880 // Do one pass of scrubbing the discovered reference lists
3881 // to remove any reference objects with strongly-reachable
3882 // referents.
3883 if (clean_refs) {
3884 CMSPrecleanRefsYieldClosure yield_cl(this);
3885 assert(rp->span().equals(_span), "Spans should be equal");
3886 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3887 &_markStack, true /* preclean */);
3888 CMSDrainMarkingStackClosure complete_trace(this,
3889 _span, &_markBitMap, &_markStack,
3890 &keep_alive, true /* preclean */);
3891
3892 // We don't want this step to interfere with a young
3893 // collection because we don't want to take CPU
3894 // or memory bandwidth away from the young GC threads
3895 // (which may be as many as there are CPUs).
3896 // Note that we don't need to protect ourselves from
3897 // interference with mutators because they can't
3898 // manipulate the discovered reference lists nor affect
3899 // the computed reachability of the referents, the
3900 // only properties manipulated by the precleaning
3901 // of these reference lists.
3902 stopTimer();
3903 CMSTokenSyncWithLocks x(true /* is cms thread */,
3904 bitMapLock());
3905 startTimer();
3906 sample_eden();
3907
3908 // The following will yield to allow foreground
3909 // collection to proceed promptly. XXX YSR:
3910 // The code in this method may need further
3911 // tweaking for better performance and some restructuring
3912 // for cleaner interfaces.
3913 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3914 rp->preclean_discovered_references(
3915 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3916 gc_timer);
3917 }
3918
3919 if (clean_survivor) { // preclean the active survivor space(s)
3920 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3921 &_markBitMap, &_modUnionTable,
3922 &_markStack, true /* precleaning phase */);
3923 stopTimer();
3924 CMSTokenSyncWithLocks ts(true /* is cms thread */,
3925 bitMapLock());
3926 startTimer();
3927 unsigned int before_count =
3928 GenCollectedHeap::heap()->total_collections();
3929 SurvivorSpacePrecleanClosure
3930 sss_cl(this, _span, &_markBitMap, &_markStack,
3931 &pam_cl, before_count, CMSYield);
3932 _young_gen->from()->object_iterate_careful(&sss_cl);
3933 _young_gen->to()->object_iterate_careful(&sss_cl);
3934 }
3935 MarkRefsIntoAndScanClosure
3936 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3937 &_markStack, this, CMSYield,
3938 true /* precleaning phase */);
3939 // CAUTION: The following closure has persistent state that may need to
3940 // be reset upon a decrease in the sequence of addresses it
3941 // processes.
3942 ScanMarkedObjectsAgainCarefullyClosure
3943 smoac_cl(this, _span,
3944 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3945
3946 // Preclean dirty cards in ModUnionTable and CardTable using
3947 // appropriate convergence criterion;
3948 // repeat CMSPrecleanIter times unless we find that
3949 // we are losing.
3950 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3951 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3952 "Bad convergence multiplier");
3953 assert(CMSPrecleanThreshold >= 100,
3954 "Unreasonably low CMSPrecleanThreshold");
3955
3956 size_t numIter, cumNumCards, lastNumCards, curNumCards;
3957 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3958 numIter < CMSPrecleanIter;
3959 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3960 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
3961 if (Verbose && PrintGCDetails) {
3962 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3963 }
3964 // Either there are very few dirty cards, so re-mark
3965 // pause will be small anyway, or our pre-cleaning isn't
3966 // that much faster than the rate at which cards are being
3967 // dirtied, so we might as well stop and re-mark since
3968 // precleaning won't improve our re-mark time by much.
3969 if (curNumCards <= CMSPrecleanThreshold ||
3970 (numIter > 0 &&
3971 (curNumCards * CMSPrecleanDenominator >
3972 lastNumCards * CMSPrecleanNumerator))) {
3973 numIter++;
3974 cumNumCards += curNumCards;
3975 break;
3976 }
3977 }
3978
3979 preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
3980
3981 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
3982 cumNumCards += curNumCards;
3983 if (PrintGCDetails && PrintCMSStatistics != 0) {
3984 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
3985 curNumCards, cumNumCards, numIter);
3986 }
3987 return cumNumCards; // as a measure of useful work done
3988 }
3989
3990 // PRECLEANING NOTES:
3991 // Precleaning involves:
3992 // . reading the bits of the modUnionTable and clearing the set bits.
3993 // . For the cards corresponding to the set bits, we scan the
3994 // objects on those cards. This means we need the free_list_lock
3995 // so that we can safely iterate over the CMS space when scanning
3996 // for oops.
3997 // . When we scan the objects, we'll be both reading and setting
3998 // marks in the marking bit map, so we'll need the marking bit map.
3999 // . For protecting _collector_state transitions, we take the CGC_lock.
4000 // Note that any races in the reading of of card table entries by the
4001 // CMS thread on the one hand and the clearing of those entries by the
4002 // VM thread or the setting of those entries by the mutator threads on the
4003 // other are quite benign. However, for efficiency it makes sense to keep
4004 // the VM thread from racing with the CMS thread while the latter is
4005 // dirty card info to the modUnionTable. We therefore also use the
4006 // CGC_lock to protect the reading of the card table and the mod union
4007 // table by the CM thread.
4008 // . We run concurrently with mutator updates, so scanning
4009 // needs to be done carefully -- we should not try to scan
4010 // potentially uninitialized objects.
4011 //
4012 // Locking strategy: While holding the CGC_lock, we scan over and
4013 // reset a maximal dirty range of the mod union / card tables, then lock
4014 // the free_list_lock and bitmap lock to do a full marking, then
4015 // release these locks; and repeat the cycle. This allows for a
4016 // certain amount of fairness in the sharing of these locks between
4017 // the CMS collector on the one hand, and the VM thread and the
4018 // mutators on the other.
4019
4020 // NOTE: preclean_mod_union_table() and preclean_card_table()
4021 // further below are largely identical; if you need to modify
4022 // one of these methods, please check the other method too.
4023
4024 size_t CMSCollector::preclean_mod_union_table(
4025 ConcurrentMarkSweepGeneration* old_gen,
4026 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4027 verify_work_stacks_empty();
4028 verify_overflow_empty();
4029
4030 // strategy: starting with the first card, accumulate contiguous
4031 // ranges of dirty cards; clear these cards, then scan the region
4032 // covered by these cards.
4033
4034 // Since all of the MUT is committed ahead, we can just use
4035 // that, in case the generations expand while we are precleaning.
4036 // It might also be fine to just use the committed part of the
4037 // generation, but we might potentially miss cards when the
4038 // generation is rapidly expanding while we are in the midst
4039 // of precleaning.
4040 HeapWord* startAddr = old_gen->reserved().start();
4041 HeapWord* endAddr = old_gen->reserved().end();
4042
4043 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
4044
4045 size_t numDirtyCards, cumNumDirtyCards;
4046 HeapWord *nextAddr, *lastAddr;
4047 for (cumNumDirtyCards = numDirtyCards = 0,
4048 nextAddr = lastAddr = startAddr;
4049 nextAddr < endAddr;
4050 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4051
4052 ResourceMark rm;
4053 HandleMark hm;
4054
4055 MemRegion dirtyRegion;
4056 {
4057 stopTimer();
4058 // Potential yield point
4059 CMSTokenSync ts(true);
4060 startTimer();
4061 sample_eden();
4062 // Get dirty region starting at nextOffset (inclusive),
4063 // simultaneously clearing it.
4064 dirtyRegion =
4065 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4066 assert(dirtyRegion.start() >= nextAddr,
4067 "returned region inconsistent?");
4068 }
4069 // Remember where the next search should begin.
4070 // The returned region (if non-empty) is a right open interval,
4071 // so lastOffset is obtained from the right end of that
4072 // interval.
4073 lastAddr = dirtyRegion.end();
4074 // Should do something more transparent and less hacky XXX
4075 numDirtyCards =
4076 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4077
4078 // We'll scan the cards in the dirty region (with periodic
4079 // yields for foreground GC as needed).
4080 if (!dirtyRegion.is_empty()) {
4081 assert(numDirtyCards > 0, "consistency check");
4082 HeapWord* stop_point = NULL;
4083 stopTimer();
4084 // Potential yield point
4085 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(),
4086 bitMapLock());
4087 startTimer();
4088 {
4089 verify_work_stacks_empty();
4090 verify_overflow_empty();
4091 sample_eden();
4092 stop_point =
4093 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4094 }
4095 if (stop_point != NULL) {
4096 // The careful iteration stopped early either because it found an
4097 // uninitialized object, or because we were in the midst of an
4098 // "abortable preclean", which should now be aborted. Redirty
4099 // the bits corresponding to the partially-scanned or unscanned
4100 // cards. We'll either restart at the next block boundary or
4101 // abort the preclean.
4102 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4103 "Should only be AbortablePreclean.");
4104 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4105 if (should_abort_preclean()) {
4106 break; // out of preclean loop
4107 } else {
4108 // Compute the next address at which preclean should pick up;
4109 // might need bitMapLock in order to read P-bits.
4110 lastAddr = next_card_start_after_block(stop_point);
4111 }
4112 }
4113 } else {
4114 assert(lastAddr == endAddr, "consistency check");
4115 assert(numDirtyCards == 0, "consistency check");
4116 break;
4117 }
4118 }
4119 verify_work_stacks_empty();
4120 verify_overflow_empty();
4121 return cumNumDirtyCards;
4122 }
4123
4124 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4125 // below are largely identical; if you need to modify
4126 // one of these methods, please check the other method too.
4127
4128 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen,
4129 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4130 // strategy: it's similar to precleamModUnionTable above, in that
4131 // we accumulate contiguous ranges of dirty cards, mark these cards
4132 // precleaned, then scan the region covered by these cards.
4133 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high());
4134 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low());
4135
4136 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
4137
4138 size_t numDirtyCards, cumNumDirtyCards;
4139 HeapWord *lastAddr, *nextAddr;
4140
4141 for (cumNumDirtyCards = numDirtyCards = 0,
4142 nextAddr = lastAddr = startAddr;
4143 nextAddr < endAddr;
4144 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4145
4146 ResourceMark rm;
4147 HandleMark hm;
4148
4149 MemRegion dirtyRegion;
4150 {
4151 // See comments in "Precleaning notes" above on why we
4152 // do this locking. XXX Could the locking overheads be
4153 // too high when dirty cards are sparse? [I don't think so.]
4154 stopTimer();
4155 CMSTokenSync x(true); // is cms thread
4156 startTimer();
4157 sample_eden();
4158 // Get and clear dirty region from card table
4159 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4160 MemRegion(nextAddr, endAddr),
4161 true,
4162 CardTableModRefBS::precleaned_card_val());
4163
4164 assert(dirtyRegion.start() >= nextAddr,
4165 "returned region inconsistent?");
4166 }
4167 lastAddr = dirtyRegion.end();
4168 numDirtyCards =
4169 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4170
4171 if (!dirtyRegion.is_empty()) {
4172 stopTimer();
4173 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock());
4174 startTimer();
4175 sample_eden();
4176 verify_work_stacks_empty();
4177 verify_overflow_empty();
4178 HeapWord* stop_point =
4179 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4180 if (stop_point != NULL) {
4181 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4182 "Should only be AbortablePreclean.");
4183 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4184 if (should_abort_preclean()) {
4185 break; // out of preclean loop
4186 } else {
4187 // Compute the next address at which preclean should pick up.
4188 lastAddr = next_card_start_after_block(stop_point);
4189 }
4190 }
4191 } else {
4192 break;
4193 }
4194 }
4195 verify_work_stacks_empty();
4196 verify_overflow_empty();
4197 return cumNumDirtyCards;
4198 }
4199
4200 class PrecleanKlassClosure : public KlassClosure {
4201 KlassToOopClosure _cm_klass_closure;
4202 public:
4203 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4204 void do_klass(Klass* k) {
4205 if (k->has_accumulated_modified_oops()) {
4206 k->clear_accumulated_modified_oops();
4207
4208 _cm_klass_closure.do_klass(k);
4209 }
4210 }
4211 };
4212
4213 // The freelist lock is needed to prevent asserts, is it really needed?
4214 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4215
4216 cl->set_freelistLock(freelistLock);
4217
4218 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4219
4220 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4221 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4222 PrecleanKlassClosure preclean_klass_closure(cl);
4223 ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
4224
4225 verify_work_stacks_empty();
4226 verify_overflow_empty();
4227 }
4228
4229 void CMSCollector::checkpointRootsFinal() {
4230 assert(_collectorState == FinalMarking, "incorrect state transition?");
4231 check_correct_thread_executing();
4232 // world is stopped at this checkpoint
4233 assert(SafepointSynchronize::is_at_safepoint(),
4234 "world should be stopped");
4235 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4236
4237 verify_work_stacks_empty();
4238 verify_overflow_empty();
4239
4240 if (PrintGCDetails) {
4241 gclog_or_tty->print("[YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)]",
4242 _young_gen->used() / K,
4243 _young_gen->capacity() / K);
4244 }
4245 {
4246 if (CMSScavengeBeforeRemark) {
4247 GenCollectedHeap* gch = GenCollectedHeap::heap();
4248 // Temporarily set flag to false, GCH->do_collection will
4249 // expect it to be false and set to true
4250 FlagSetting fl(gch->_is_gc_active, false);
4251 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
4252 PrintGCDetails && Verbose, true, _gc_timer_cm);)
4253 gch->do_collection(true, // full (i.e. force, see below)
4254 false, // !clear_all_soft_refs
4255 0, // size
4256 false, // is_tlab
4257 GenCollectedHeap::YoungGen // type
4258 );
4259 }
4260 FreelistLocker x(this);
4261 MutexLockerEx y(bitMapLock(),
4262 Mutex::_no_safepoint_check_flag);
4263 checkpointRootsFinalWork();
4264 }
4265 verify_work_stacks_empty();
4266 verify_overflow_empty();
4267 }
4268
4269 void CMSCollector::checkpointRootsFinalWork() {
4270 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);)
4271
4272 assert(haveFreelistLocks(), "must have free list locks");
4273 assert_lock_strong(bitMapLock());
4274
4275 ResourceMark rm;
4276 HandleMark hm;
4277
4278 GenCollectedHeap* gch = GenCollectedHeap::heap();
4279
4280 if (should_unload_classes()) {
4281 CodeCache::gc_prologue();
4282 }
4283 assert(haveFreelistLocks(), "must have free list locks");
4284 assert_lock_strong(bitMapLock());
4285
4286 // We might assume that we need not fill TLAB's when
4287 // CMSScavengeBeforeRemark is set, because we may have just done
4288 // a scavenge which would have filled all TLAB's -- and besides
4289 // Eden would be empty. This however may not always be the case --
4290 // for instance although we asked for a scavenge, it may not have
4291 // happened because of a JNI critical section. We probably need
4292 // a policy for deciding whether we can in that case wait until
4293 // the critical section releases and then do the remark following
4294 // the scavenge, and skip it here. In the absence of that policy,
4295 // or of an indication of whether the scavenge did indeed occur,
4296 // we cannot rely on TLAB's having been filled and must do
4297 // so here just in case a scavenge did not happen.
4298 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4299 // Update the saved marks which may affect the root scans.
4300 gch->save_marks();
4301
4302 if (CMSPrintEdenSurvivorChunks) {
4303 print_eden_and_survivor_chunk_arrays();
4304 }
4305
4306 {
4307 #if defined(COMPILER2) || INCLUDE_JVMCI
4308 DerivedPointerTableDeactivate dpt_deact;
4309 #endif
4310
4311 // Note on the role of the mod union table:
4312 // Since the marker in "markFromRoots" marks concurrently with
4313 // mutators, it is possible for some reachable objects not to have been
4314 // scanned. For instance, an only reference to an object A was
4315 // placed in object B after the marker scanned B. Unless B is rescanned,
4316 // A would be collected. Such updates to references in marked objects
4317 // are detected via the mod union table which is the set of all cards
4318 // dirtied since the first checkpoint in this GC cycle and prior to
4319 // the most recent young generation GC, minus those cleaned up by the
4320 // concurrent precleaning.
4321 if (CMSParallelRemarkEnabled) {
4322 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm);
4323 do_remark_parallel();
4324 } else {
4325 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, _gc_timer_cm);
4326 do_remark_non_parallel();
4327 }
4328 }
4329 verify_work_stacks_empty();
4330 verify_overflow_empty();
4331
4332 {
4333 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);)
4334 refProcessingWork();
4335 }
4336 verify_work_stacks_empty();
4337 verify_overflow_empty();
4338
4339 if (should_unload_classes()) {
4340 CodeCache::gc_epilogue();
4341 }
4342 JvmtiExport::gc_epilogue();
4343
4344 // If we encountered any (marking stack / work queue) overflow
4345 // events during the current CMS cycle, take appropriate
4346 // remedial measures, where possible, so as to try and avoid
4347 // recurrence of that condition.
4348 assert(_markStack.isEmpty(), "No grey objects");
4349 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4350 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4351 if (ser_ovflw > 0) {
4352 if (PrintCMSStatistics != 0) {
4353 gclog_or_tty->print_cr("Marking stack overflow (benign) "
4354 "(pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT
4355 ", kac_preclean=" SIZE_FORMAT ")",
4356 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4357 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4358 }
4359 _markStack.expand();
4360 _ser_pmc_remark_ovflw = 0;
4361 _ser_pmc_preclean_ovflw = 0;
4362 _ser_kac_preclean_ovflw = 0;
4363 _ser_kac_ovflw = 0;
4364 }
4365 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4366 if (PrintCMSStatistics != 0) {
4367 gclog_or_tty->print_cr("Work queue overflow (benign) "
4368 "(pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")",
4369 _par_pmc_remark_ovflw, _par_kac_ovflw);
4370 }
4371 _par_pmc_remark_ovflw = 0;
4372 _par_kac_ovflw = 0;
4373 }
4374 if (PrintCMSStatistics != 0) {
4375 if (_markStack._hit_limit > 0) {
4376 gclog_or_tty->print_cr(" (benign) Hit max stack size limit (" SIZE_FORMAT ")",
4377 _markStack._hit_limit);
4378 }
4379 if (_markStack._failed_double > 0) {
4380 gclog_or_tty->print_cr(" (benign) Failed stack doubling (" SIZE_FORMAT "),"
4381 " current capacity " SIZE_FORMAT,
4382 _markStack._failed_double,
4383 _markStack.capacity());
4384 }
4385 }
4386 _markStack._hit_limit = 0;
4387 _markStack._failed_double = 0;
4388
4389 if ((VerifyAfterGC || VerifyDuringGC) &&
4390 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4391 verify_after_remark();
4392 }
4393
4394 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4395
4396 // Change under the freelistLocks.
4397 _collectorState = Sweeping;
4398 // Call isAllClear() under bitMapLock
4399 assert(_modUnionTable.isAllClear(),
4400 "Should be clear by end of the final marking");
4401 assert(_ct->klass_rem_set()->mod_union_is_clear(),
4402 "Should be clear by end of the final marking");
4403 }
4404
4405 void CMSParInitialMarkTask::work(uint worker_id) {
4406 elapsedTimer _timer;
4407 ResourceMark rm;
4408 HandleMark hm;
4409
4410 // ---------- scan from roots --------------
4411 _timer.start();
4412 GenCollectedHeap* gch = GenCollectedHeap::heap();
4413 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4414
4415 // ---------- young gen roots --------------
4416 {
4417 work_on_young_gen_roots(worker_id, &par_mri_cl);
4418 _timer.stop();
4419 if (PrintCMSStatistics != 0) {
4420 gclog_or_tty->print_cr(
4421 "Finished young gen initial mark scan work in %dth thread: %3.3f sec",
4422 worker_id, _timer.seconds());
4423 }
4424 }
4425
4426 // ---------- remaining roots --------------
4427 _timer.reset();
4428 _timer.start();
4429
4430 CLDToOopClosure cld_closure(&par_mri_cl, true);
4431
4432 gch->gen_process_roots(_strong_roots_scope,
4433 GenCollectedHeap::OldGen,
4434 false, // yg was scanned above
4435 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4436 _collector->should_unload_classes(),
4437 &par_mri_cl,
4438 NULL,
4439 &cld_closure);
4440 assert(_collector->should_unload_classes()
4441 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4442 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4443 _timer.stop();
4444 if (PrintCMSStatistics != 0) {
4445 gclog_or_tty->print_cr(
4446 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec",
4447 worker_id, _timer.seconds());
4448 }
4449 }
4450
4451 // Parallel remark task
4452 class CMSParRemarkTask: public CMSParMarkTask {
4453 CompactibleFreeListSpace* _cms_space;
4454
4455 // The per-thread work queues, available here for stealing.
4456 OopTaskQueueSet* _task_queues;
4457 ParallelTaskTerminator _term;
4458 StrongRootsScope* _strong_roots_scope;
4459
4460 public:
4461 // A value of 0 passed to n_workers will cause the number of
4462 // workers to be taken from the active workers in the work gang.
4463 CMSParRemarkTask(CMSCollector* collector,
4464 CompactibleFreeListSpace* cms_space,
4465 uint n_workers, WorkGang* workers,
4466 OopTaskQueueSet* task_queues,
4467 StrongRootsScope* strong_roots_scope):
4468 CMSParMarkTask("Rescan roots and grey objects in parallel",
4469 collector, n_workers),
4470 _cms_space(cms_space),
4471 _task_queues(task_queues),
4472 _term(n_workers, task_queues),
4473 _strong_roots_scope(strong_roots_scope) { }
4474
4475 OopTaskQueueSet* task_queues() { return _task_queues; }
4476
4477 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4478
4479 ParallelTaskTerminator* terminator() { return &_term; }
4480 uint n_workers() { return _n_workers; }
4481
4482 void work(uint worker_id);
4483
4484 private:
4485 // ... of dirty cards in old space
4486 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4487 Par_MarkRefsIntoAndScanClosure* cl);
4488
4489 // ... work stealing for the above
4490 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4491 };
4492
4493 class RemarkKlassClosure : public KlassClosure {
4494 KlassToOopClosure _cm_klass_closure;
4495 public:
4496 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4497 void do_klass(Klass* k) {
4498 // Check if we have modified any oops in the Klass during the concurrent marking.
4499 if (k->has_accumulated_modified_oops()) {
4500 k->clear_accumulated_modified_oops();
4501
4502 // We could have transfered the current modified marks to the accumulated marks,
4503 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4504 } else if (k->has_modified_oops()) {
4505 // Don't clear anything, this info is needed by the next young collection.
4506 } else {
4507 // No modified oops in the Klass.
4508 return;
4509 }
4510
4511 // The klass has modified fields, need to scan the klass.
4512 _cm_klass_closure.do_klass(k);
4513 }
4514 };
4515
4516 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
4517 ParNewGeneration* young_gen = _collector->_young_gen;
4518 ContiguousSpace* eden_space = young_gen->eden();
4519 ContiguousSpace* from_space = young_gen->from();
4520 ContiguousSpace* to_space = young_gen->to();
4521
4522 HeapWord** eca = _collector->_eden_chunk_array;
4523 size_t ect = _collector->_eden_chunk_index;
4524 HeapWord** sca = _collector->_survivor_chunk_array;
4525 size_t sct = _collector->_survivor_chunk_index;
4526
4527 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4528 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4529
4530 do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
4531 do_young_space_rescan(worker_id, cl, from_space, sca, sct);
4532 do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
4533 }
4534
4535 // work_queue(i) is passed to the closure
4536 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
4537 // also is passed to do_dirty_card_rescan_tasks() and to
4538 // do_work_steal() to select the i-th task_queue.
4539
4540 void CMSParRemarkTask::work(uint worker_id) {
4541 elapsedTimer _timer;
4542 ResourceMark rm;
4543 HandleMark hm;
4544
4545 // ---------- rescan from roots --------------
4546 _timer.start();
4547 GenCollectedHeap* gch = GenCollectedHeap::heap();
4548 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4549 _collector->_span, _collector->ref_processor(),
4550 &(_collector->_markBitMap),
4551 work_queue(worker_id));
4552
4553 // Rescan young gen roots first since these are likely
4554 // coarsely partitioned and may, on that account, constitute
4555 // the critical path; thus, it's best to start off that
4556 // work first.
4557 // ---------- young gen roots --------------
4558 {
4559 work_on_young_gen_roots(worker_id, &par_mrias_cl);
4560 _timer.stop();
4561 if (PrintCMSStatistics != 0) {
4562 gclog_or_tty->print_cr(
4563 "Finished young gen rescan work in %dth thread: %3.3f sec",
4564 worker_id, _timer.seconds());
4565 }
4566 }
4567
4568 // ---------- remaining roots --------------
4569 _timer.reset();
4570 _timer.start();
4571 gch->gen_process_roots(_strong_roots_scope,
4572 GenCollectedHeap::OldGen,
4573 false, // yg was scanned above
4574 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4575 _collector->should_unload_classes(),
4576 &par_mrias_cl,
4577 NULL,
4578 NULL); // The dirty klasses will be handled below
4579
4580 assert(_collector->should_unload_classes()
4581 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4582 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4583 _timer.stop();
4584 if (PrintCMSStatistics != 0) {
4585 gclog_or_tty->print_cr(
4586 "Finished remaining root rescan work in %dth thread: %3.3f sec",
4587 worker_id, _timer.seconds());
4588 }
4589
4590 // ---------- unhandled CLD scanning ----------
4591 if (worker_id == 0) { // Single threaded at the moment.
4592 _timer.reset();
4593 _timer.start();
4594
4595 // Scan all new class loader data objects and new dependencies that were
4596 // introduced during concurrent marking.
4597 ResourceMark rm;
4598 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4599 for (int i = 0; i < array->length(); i++) {
4600 par_mrias_cl.do_cld_nv(array->at(i));
4601 }
4602
4603 // We don't need to keep track of new CLDs anymore.
4604 ClassLoaderDataGraph::remember_new_clds(false);
4605
4606 _timer.stop();
4607 if (PrintCMSStatistics != 0) {
4608 gclog_or_tty->print_cr(
4609 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec",
4610 worker_id, _timer.seconds());
4611 }
4612 }
4613
4614 // ---------- dirty klass scanning ----------
4615 if (worker_id == 0) { // Single threaded at the moment.
4616 _timer.reset();
4617 _timer.start();
4618
4619 // Scan all classes that was dirtied during the concurrent marking phase.
4620 RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
4621 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
4622
4623 _timer.stop();
4624 if (PrintCMSStatistics != 0) {
4625 gclog_or_tty->print_cr(
4626 "Finished dirty klass scanning work in %dth thread: %3.3f sec",
4627 worker_id, _timer.seconds());
4628 }
4629 }
4630
4631 // We might have added oops to ClassLoaderData::_handles during the
4632 // concurrent marking phase. These oops point to newly allocated objects
4633 // that are guaranteed to be kept alive. Either by the direct allocation
4634 // code, or when the young collector processes the roots. Hence,
4635 // we don't have to revisit the _handles block during the remark phase.
4636
4637 // ---------- rescan dirty cards ------------
4638 _timer.reset();
4639 _timer.start();
4640
4641 // Do the rescan tasks for each of the two spaces
4642 // (cms_space) in turn.
4643 // "worker_id" is passed to select the task_queue for "worker_id"
4644 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4645 _timer.stop();
4646 if (PrintCMSStatistics != 0) {
4647 gclog_or_tty->print_cr(
4648 "Finished dirty card rescan work in %dth thread: %3.3f sec",
4649 worker_id, _timer.seconds());
4650 }
4651
4652 // ---------- steal work from other threads ...
4653 // ---------- ... and drain overflow list.
4654 _timer.reset();
4655 _timer.start();
4656 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4657 _timer.stop();
4658 if (PrintCMSStatistics != 0) {
4659 gclog_or_tty->print_cr(
4660 "Finished work stealing in %dth thread: %3.3f sec",
4661 worker_id, _timer.seconds());
4662 }
4663 }
4664
4665 // Note that parameter "i" is not used.
4666 void
4667 CMSParMarkTask::do_young_space_rescan(uint worker_id,
4668 OopsInGenClosure* cl, ContiguousSpace* space,
4669 HeapWord** chunk_array, size_t chunk_top) {
4670 // Until all tasks completed:
4671 // . claim an unclaimed task
4672 // . compute region boundaries corresponding to task claimed
4673 // using chunk_array
4674 // . par_oop_iterate(cl) over that region
4675
4676 ResourceMark rm;
4677 HandleMark hm;
4678
4679 SequentialSubTasksDone* pst = space->par_seq_tasks();
4680
4681 uint nth_task = 0;
4682 uint n_tasks = pst->n_tasks();
4683
4684 if (n_tasks > 0) {
4685 assert(pst->valid(), "Uninitialized use?");
4686 HeapWord *start, *end;
4687 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4688 // We claimed task # nth_task; compute its boundaries.
4689 if (chunk_top == 0) { // no samples were taken
4690 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4691 start = space->bottom();
4692 end = space->top();
4693 } else if (nth_task == 0) {
4694 start = space->bottom();
4695 end = chunk_array[nth_task];
4696 } else if (nth_task < (uint)chunk_top) {
4697 assert(nth_task >= 1, "Control point invariant");
4698 start = chunk_array[nth_task - 1];
4699 end = chunk_array[nth_task];
4700 } else {
4701 assert(nth_task == (uint)chunk_top, "Control point invariant");
4702 start = chunk_array[chunk_top - 1];
4703 end = space->top();
4704 }
4705 MemRegion mr(start, end);
4706 // Verify that mr is in space
4707 assert(mr.is_empty() || space->used_region().contains(mr),
4708 "Should be in space");
4709 // Verify that "start" is an object boundary
4710 assert(mr.is_empty() || oop(mr.start())->is_oop(),
4711 "Should be an oop");
4712 space->par_oop_iterate(mr, cl);
4713 }
4714 pst->all_tasks_completed();
4715 }
4716 }
4717
4718 void
4719 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4720 CompactibleFreeListSpace* sp, int i,
4721 Par_MarkRefsIntoAndScanClosure* cl) {
4722 // Until all tasks completed:
4723 // . claim an unclaimed task
4724 // . compute region boundaries corresponding to task claimed
4725 // . transfer dirty bits ct->mut for that region
4726 // . apply rescanclosure to dirty mut bits for that region
4727
4728 ResourceMark rm;
4729 HandleMark hm;
4730
4731 OopTaskQueue* work_q = work_queue(i);
4732 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4733 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4734 // CAUTION: This closure has state that persists across calls to
4735 // the work method dirty_range_iterate_clear() in that it has
4736 // embedded in it a (subtype of) UpwardsObjectClosure. The
4737 // use of that state in the embedded UpwardsObjectClosure instance
4738 // assumes that the cards are always iterated (even if in parallel
4739 // by several threads) in monotonically increasing order per each
4740 // thread. This is true of the implementation below which picks
4741 // card ranges (chunks) in monotonically increasing order globally
4742 // and, a-fortiori, in monotonically increasing order per thread
4743 // (the latter order being a subsequence of the former).
4744 // If the work code below is ever reorganized into a more chaotic
4745 // work-partitioning form than the current "sequential tasks"
4746 // paradigm, the use of that persistent state will have to be
4747 // revisited and modified appropriately. See also related
4748 // bug 4756801 work on which should examine this code to make
4749 // sure that the changes there do not run counter to the
4750 // assumptions made here and necessary for correctness and
4751 // efficiency. Note also that this code might yield inefficient
4752 // behavior in the case of very large objects that span one or
4753 // more work chunks. Such objects would potentially be scanned
4754 // several times redundantly. Work on 4756801 should try and
4755 // address that performance anomaly if at all possible. XXX
4756 MemRegion full_span = _collector->_span;
4757 CMSBitMap* bm = &(_collector->_markBitMap); // shared
4758 MarkFromDirtyCardsClosure
4759 greyRescanClosure(_collector, full_span, // entire span of interest
4760 sp, bm, work_q, cl);
4761
4762 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4763 assert(pst->valid(), "Uninitialized use?");
4764 uint nth_task = 0;
4765 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
4766 MemRegion span = sp->used_region();
4767 HeapWord* start_addr = span.start();
4768 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
4769 alignment);
4770 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4771 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
4772 start_addr, "Check alignment");
4773 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
4774 chunk_size, "Check alignment");
4775
4776 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4777 // Having claimed the nth_task, compute corresponding mem-region,
4778 // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4779 // The alignment restriction ensures that we do not need any
4780 // synchronization with other gang-workers while setting or
4781 // clearing bits in thus chunk of the MUT.
4782 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4783 start_addr + (nth_task+1)*chunk_size);
4784 // The last chunk's end might be way beyond end of the
4785 // used region. In that case pull back appropriately.
4786 if (this_span.end() > end_addr) {
4787 this_span.set_end(end_addr);
4788 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4789 }
4790 // Iterate over the dirty cards covering this chunk, marking them
4791 // precleaned, and setting the corresponding bits in the mod union
4792 // table. Since we have been careful to partition at Card and MUT-word
4793 // boundaries no synchronization is needed between parallel threads.
4794 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
4795 &modUnionClosure);
4796
4797 // Having transferred these marks into the modUnionTable,
4798 // rescan the marked objects on the dirty cards in the modUnionTable.
4799 // Even if this is at a synchronous collection, the initial marking
4800 // may have been done during an asynchronous collection so there
4801 // may be dirty bits in the mod-union table.
4802 _collector->_modUnionTable.dirty_range_iterate_clear(
4803 this_span, &greyRescanClosure);
4804 _collector->_modUnionTable.verifyNoOneBitsInRange(
4805 this_span.start(),
4806 this_span.end());
4807 }
4808 pst->all_tasks_completed(); // declare that i am done
4809 }
4810
4811 // . see if we can share work_queues with ParNew? XXX
4812 void
4813 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
4814 int* seed) {
4815 OopTaskQueue* work_q = work_queue(i);
4816 NOT_PRODUCT(int num_steals = 0;)
4817 oop obj_to_scan;
4818 CMSBitMap* bm = &(_collector->_markBitMap);
4819
4820 while (true) {
4821 // Completely finish any left over work from (an) earlier round(s)
4822 cl->trim_queue(0);
4823 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4824 (size_t)ParGCDesiredObjsFromOverflowList);
4825 // Now check if there's any work in the overflow list
4826 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4827 // only affects the number of attempts made to get work from the
4828 // overflow list and does not affect the number of workers. Just
4829 // pass ParallelGCThreads so this behavior is unchanged.
4830 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4831 work_q,
4832 ParallelGCThreads)) {
4833 // found something in global overflow list;
4834 // not yet ready to go stealing work from others.
4835 // We'd like to assert(work_q->size() != 0, ...)
4836 // because we just took work from the overflow list,
4837 // but of course we can't since all of that could have
4838 // been already stolen from us.
4839 // "He giveth and He taketh away."
4840 continue;
4841 }
4842 // Verify that we have no work before we resort to stealing
4843 assert(work_q->size() == 0, "Have work, shouldn't steal");
4844 // Try to steal from other queues that have work
4845 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4846 NOT_PRODUCT(num_steals++;)
4847 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
4848 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4849 // Do scanning work
4850 obj_to_scan->oop_iterate(cl);
4851 // Loop around, finish this work, and try to steal some more
4852 } else if (terminator()->offer_termination()) {
4853 break; // nirvana from the infinite cycle
4854 }
4855 }
4856 NOT_PRODUCT(
4857 if (PrintCMSStatistics != 0) {
4858 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
4859 }
4860 )
4861 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4862 "Else our work is not yet done");
4863 }
4864
4865 // Record object boundaries in _eden_chunk_array by sampling the eden
4866 // top in the slow-path eden object allocation code path and record
4867 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4868 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4869 // sampling in sample_eden() that activates during the part of the
4870 // preclean phase.
4871 void CMSCollector::sample_eden_chunk() {
4872 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4873 if (_eden_chunk_lock->try_lock()) {
4874 // Record a sample. This is the critical section. The contents
4875 // of the _eden_chunk_array have to be non-decreasing in the
4876 // address order.
4877 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4878 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4879 "Unexpected state of Eden");
4880 if (_eden_chunk_index == 0 ||
4881 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4882 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4883 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4884 _eden_chunk_index++; // commit sample
4885 }
4886 _eden_chunk_lock->unlock();
4887 }
4888 }
4889 }
4890
4891 // Return a thread-local PLAB recording array, as appropriate.
4892 void* CMSCollector::get_data_recorder(int thr_num) {
4893 if (_survivor_plab_array != NULL &&
4894 (CMSPLABRecordAlways ||
4895 (_collectorState > Marking && _collectorState < FinalMarking))) {
4896 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4897 ChunkArray* ca = &_survivor_plab_array[thr_num];
4898 ca->reset(); // clear it so that fresh data is recorded
4899 return (void*) ca;
4900 } else {
4901 return NULL;
4902 }
4903 }
4904
4905 // Reset all the thread-local PLAB recording arrays
4906 void CMSCollector::reset_survivor_plab_arrays() {
4907 for (uint i = 0; i < ParallelGCThreads; i++) {
4908 _survivor_plab_array[i].reset();
4909 }
4910 }
4911
4912 // Merge the per-thread plab arrays into the global survivor chunk
4913 // array which will provide the partitioning of the survivor space
4914 // for CMS initial scan and rescan.
4915 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4916 int no_of_gc_threads) {
4917 assert(_survivor_plab_array != NULL, "Error");
4918 assert(_survivor_chunk_array != NULL, "Error");
4919 assert(_collectorState == FinalMarking ||
4920 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4921 for (int j = 0; j < no_of_gc_threads; j++) {
4922 _cursor[j] = 0;
4923 }
4924 HeapWord* top = surv->top();
4925 size_t i;
4926 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
4927 HeapWord* min_val = top; // Higher than any PLAB address
4928 uint min_tid = 0; // position of min_val this round
4929 for (int j = 0; j < no_of_gc_threads; j++) {
4930 ChunkArray* cur_sca = &_survivor_plab_array[j];
4931 if (_cursor[j] == cur_sca->end()) {
4932 continue;
4933 }
4934 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4935 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4936 assert(surv->used_region().contains(cur_val), "Out of bounds value");
4937 if (cur_val < min_val) {
4938 min_tid = j;
4939 min_val = cur_val;
4940 } else {
4941 assert(cur_val < top, "All recorded addresses should be less");
4942 }
4943 }
4944 // At this point min_val and min_tid are respectively
4945 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4946 // and the thread (j) that witnesses that address.
4947 // We record this address in the _survivor_chunk_array[i]
4948 // and increment _cursor[min_tid] prior to the next round i.
4949 if (min_val == top) {
4950 break;
4951 }
4952 _survivor_chunk_array[i] = min_val;
4953 _cursor[min_tid]++;
4954 }
4955 // We are all done; record the size of the _survivor_chunk_array
4956 _survivor_chunk_index = i; // exclusive: [0, i)
4957 if (PrintCMSStatistics > 0) {
4958 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4959 }
4960 // Verify that we used up all the recorded entries
4961 #ifdef ASSERT
4962 size_t total = 0;
4963 for (int j = 0; j < no_of_gc_threads; j++) {
4964 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4965 total += _cursor[j];
4966 }
4967 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
4968 // Check that the merged array is in sorted order
4969 if (total > 0) {
4970 for (size_t i = 0; i < total - 1; i++) {
4971 if (PrintCMSStatistics > 0) {
4972 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
4973 i, p2i(_survivor_chunk_array[i]));
4974 }
4975 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
4976 "Not sorted");
4977 }
4978 }
4979 #endif // ASSERT
4980 }
4981
4982 // Set up the space's par_seq_tasks structure for work claiming
4983 // for parallel initial scan and rescan of young gen.
4984 // See ParRescanTask where this is currently used.
4985 void
4986 CMSCollector::
4987 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
4988 assert(n_threads > 0, "Unexpected n_threads argument");
4989
4990 // Eden space
4991 if (!_young_gen->eden()->is_empty()) {
4992 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
4993 assert(!pst->valid(), "Clobbering existing data?");
4994 // Each valid entry in [0, _eden_chunk_index) represents a task.
4995 size_t n_tasks = _eden_chunk_index + 1;
4996 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
4997 // Sets the condition for completion of the subtask (how many threads
4998 // need to finish in order to be done).
4999 pst->set_n_threads(n_threads);
5000 pst->set_n_tasks((int)n_tasks);
5001 }
5002
5003 // Merge the survivor plab arrays into _survivor_chunk_array
5004 if (_survivor_plab_array != NULL) {
5005 merge_survivor_plab_arrays(_young_gen->from(), n_threads);
5006 } else {
5007 assert(_survivor_chunk_index == 0, "Error");
5008 }
5009
5010 // To space
5011 {
5012 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
5013 assert(!pst->valid(), "Clobbering existing data?");
5014 // Sets the condition for completion of the subtask (how many threads
5015 // need to finish in order to be done).
5016 pst->set_n_threads(n_threads);
5017 pst->set_n_tasks(1);
5018 assert(pst->valid(), "Error");
5019 }
5020
5021 // From space
5022 {
5023 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
5024 assert(!pst->valid(), "Clobbering existing data?");
5025 size_t n_tasks = _survivor_chunk_index + 1;
5026 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5027 // Sets the condition for completion of the subtask (how many threads
5028 // need to finish in order to be done).
5029 pst->set_n_threads(n_threads);
5030 pst->set_n_tasks((int)n_tasks);
5031 assert(pst->valid(), "Error");
5032 }
5033 }
5034
5035 // Parallel version of remark
5036 void CMSCollector::do_remark_parallel() {
5037 GenCollectedHeap* gch = GenCollectedHeap::heap();
5038 WorkGang* workers = gch->workers();
5039 assert(workers != NULL, "Need parallel worker threads.");
5040 // Choose to use the number of GC workers most recently set
5041 // into "active_workers".
5042 uint n_workers = workers->active_workers();
5043
5044 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5045
5046 StrongRootsScope srs(n_workers);
5047
5048 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
5049
5050 // We won't be iterating over the cards in the card table updating
5051 // the younger_gen cards, so we shouldn't call the following else
5052 // the verification code as well as subsequent younger_refs_iterate
5053 // code would get confused. XXX
5054 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5055
5056 // The young gen rescan work will not be done as part of
5057 // process_roots (which currently doesn't know how to
5058 // parallelize such a scan), but rather will be broken up into
5059 // a set of parallel tasks (via the sampling that the [abortable]
5060 // preclean phase did of eden, plus the [two] tasks of
5061 // scanning the [two] survivor spaces. Further fine-grain
5062 // parallelization of the scanning of the survivor spaces
5063 // themselves, and of precleaning of the young gen itself
5064 // is deferred to the future.
5065 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5066
5067 // The dirty card rescan work is broken up into a "sequence"
5068 // of parallel tasks (per constituent space) that are dynamically
5069 // claimed by the parallel threads.
5070 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5071
5072 // It turns out that even when we're using 1 thread, doing the work in a
5073 // separate thread causes wide variance in run times. We can't help this
5074 // in the multi-threaded case, but we special-case n=1 here to get
5075 // repeatable measurements of the 1-thread overhead of the parallel code.
5076 if (n_workers > 1) {
5077 // Make refs discovery MT-safe, if it isn't already: it may not
5078 // necessarily be so, since it's possible that we are doing
5079 // ST marking.
5080 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5081 workers->run_task(&tsk);
5082 } else {
5083 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5084 tsk.work(0);
5085 }
5086
5087 // restore, single-threaded for now, any preserved marks
5088 // as a result of work_q overflow
5089 restore_preserved_marks_if_any();
5090 }
5091
5092 // Non-parallel version of remark
5093 void CMSCollector::do_remark_non_parallel() {
5094 ResourceMark rm;
5095 HandleMark hm;
5096 GenCollectedHeap* gch = GenCollectedHeap::heap();
5097 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5098
5099 MarkRefsIntoAndScanClosure
5100 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
5101 &_markStack, this,
5102 false /* should_yield */, false /* not precleaning */);
5103 MarkFromDirtyCardsClosure
5104 markFromDirtyCardsClosure(this, _span,
5105 NULL, // space is set further below
5106 &_markBitMap, &_markStack, &mrias_cl);
5107 {
5108 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm);
5109 // Iterate over the dirty cards, setting the corresponding bits in the
5110 // mod union table.
5111 {
5112 ModUnionClosure modUnionClosure(&_modUnionTable);
5113 _ct->ct_bs()->dirty_card_iterate(
5114 _cmsGen->used_region(),
5115 &modUnionClosure);
5116 }
5117 // Having transferred these marks into the modUnionTable, we just need
5118 // to rescan the marked objects on the dirty cards in the modUnionTable.
5119 // The initial marking may have been done during an asynchronous
5120 // collection so there may be dirty bits in the mod-union table.
5121 const int alignment =
5122 CardTableModRefBS::card_size * BitsPerWord;
5123 {
5124 // ... First handle dirty cards in CMS gen
5125 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5126 MemRegion ur = _cmsGen->used_region();
5127 HeapWord* lb = ur.start();
5128 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5129 MemRegion cms_span(lb, ub);
5130 _modUnionTable.dirty_range_iterate_clear(cms_span,
5131 &markFromDirtyCardsClosure);
5132 verify_work_stacks_empty();
5133 if (PrintCMSStatistics != 0) {
5134 gclog_or_tty->print(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ",
5135 markFromDirtyCardsClosure.num_dirty_cards());
5136 }
5137 }
5138 }
5139 if (VerifyDuringGC &&
5140 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5141 HandleMark hm; // Discard invalid handles created during verification
5142 Universe::verify();
5143 }
5144 {
5145 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm);
5146
5147 verify_work_stacks_empty();
5148
5149 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5150 StrongRootsScope srs(1);
5151
5152 gch->gen_process_roots(&srs,
5153 GenCollectedHeap::OldGen,
5154 true, // young gen as roots
5155 GenCollectedHeap::ScanningOption(roots_scanning_options()),
5156 should_unload_classes(),
5157 &mrias_cl,
5158 NULL,
5159 NULL); // The dirty klasses will be handled below
5160
5161 assert(should_unload_classes()
5162 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
5163 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5164 }
5165
5166 {
5167 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm);
5168
5169 verify_work_stacks_empty();
5170
5171 // Scan all class loader data objects that might have been introduced
5172 // during concurrent marking.
5173 ResourceMark rm;
5174 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5175 for (int i = 0; i < array->length(); i++) {
5176 mrias_cl.do_cld_nv(array->at(i));
5177 }
5178
5179 // We don't need to keep track of new CLDs anymore.
5180 ClassLoaderDataGraph::remember_new_clds(false);
5181
5182 verify_work_stacks_empty();
5183 }
5184
5185 {
5186 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm);
5187
5188 verify_work_stacks_empty();
5189
5190 RemarkKlassClosure remark_klass_closure(&mrias_cl);
5191 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5192
5193 verify_work_stacks_empty();
5194 }
5195
5196 // We might have added oops to ClassLoaderData::_handles during the
5197 // concurrent marking phase. These oops point to newly allocated objects
5198 // that are guaranteed to be kept alive. Either by the direct allocation
5199 // code, or when the young collector processes the roots. Hence,
5200 // we don't have to revisit the _handles block during the remark phase.
5201
5202 verify_work_stacks_empty();
5203 // Restore evacuated mark words, if any, used for overflow list links
5204 restore_preserved_marks_if_any();
5205
5206 verify_overflow_empty();
5207 }
5208
5209 ////////////////////////////////////////////////////////
5210 // Parallel Reference Processing Task Proxy Class
5211 ////////////////////////////////////////////////////////
5212 class AbstractGangTaskWOopQueues : public AbstractGangTask {
5213 OopTaskQueueSet* _queues;
5214 ParallelTaskTerminator _terminator;
5215 public:
5216 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) :
5217 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {}
5218 ParallelTaskTerminator* terminator() { return &_terminator; }
5219 OopTaskQueueSet* queues() { return _queues; }
5220 };
5221
5222 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5223 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5224 CMSCollector* _collector;
5225 CMSBitMap* _mark_bit_map;
5226 const MemRegion _span;
5227 ProcessTask& _task;
5228
5229 public:
5230 CMSRefProcTaskProxy(ProcessTask& task,
5231 CMSCollector* collector,
5232 const MemRegion& span,
5233 CMSBitMap* mark_bit_map,
5234 AbstractWorkGang* workers,
5235 OopTaskQueueSet* task_queues):
5236 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5237 task_queues,
5238 workers->active_workers()),
5239 _task(task),
5240 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5241 {
5242 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5243 "Inconsistency in _span");
5244 }
5245
5246 OopTaskQueueSet* task_queues() { return queues(); }
5247
5248 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5249
5250 void do_work_steal(int i,
5251 CMSParDrainMarkingStackClosure* drain,
5252 CMSParKeepAliveClosure* keep_alive,
5253 int* seed);
5254
5255 virtual void work(uint worker_id);
5256 };
5257
5258 void CMSRefProcTaskProxy::work(uint worker_id) {
5259 ResourceMark rm;
5260 HandleMark hm;
5261 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5262 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5263 _mark_bit_map,
5264 work_queue(worker_id));
5265 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5266 _mark_bit_map,
5267 work_queue(worker_id));
5268 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5269 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5270 if (_task.marks_oops_alive()) {
5271 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5272 _collector->hash_seed(worker_id));
5273 }
5274 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5275 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5276 }
5277
5278 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5279 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5280 EnqueueTask& _task;
5281
5282 public:
5283 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5284 : AbstractGangTask("Enqueue reference objects in parallel"),
5285 _task(task)
5286 { }
5287
5288 virtual void work(uint worker_id)
5289 {
5290 _task.work(worker_id);
5291 }
5292 };
5293
5294 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5295 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5296 _span(span),
5297 _bit_map(bit_map),
5298 _work_queue(work_queue),
5299 _mark_and_push(collector, span, bit_map, work_queue),
5300 _low_water_mark(MIN2((work_queue->max_elems()/4),
5301 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5302 { }
5303
5304 // . see if we can share work_queues with ParNew? XXX
5305 void CMSRefProcTaskProxy::do_work_steal(int i,
5306 CMSParDrainMarkingStackClosure* drain,
5307 CMSParKeepAliveClosure* keep_alive,
5308 int* seed) {
5309 OopTaskQueue* work_q = work_queue(i);
5310 NOT_PRODUCT(int num_steals = 0;)
5311 oop obj_to_scan;
5312
5313 while (true) {
5314 // Completely finish any left over work from (an) earlier round(s)
5315 drain->trim_queue(0);
5316 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5317 (size_t)ParGCDesiredObjsFromOverflowList);
5318 // Now check if there's any work in the overflow list
5319 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5320 // only affects the number of attempts made to get work from the
5321 // overflow list and does not affect the number of workers. Just
5322 // pass ParallelGCThreads so this behavior is unchanged.
5323 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5324 work_q,
5325 ParallelGCThreads)) {
5326 // Found something in global overflow list;
5327 // not yet ready to go stealing work from others.
5328 // We'd like to assert(work_q->size() != 0, ...)
5329 // because we just took work from the overflow list,
5330 // but of course we can't, since all of that might have
5331 // been already stolen from us.
5332 continue;
5333 }
5334 // Verify that we have no work before we resort to stealing
5335 assert(work_q->size() == 0, "Have work, shouldn't steal");
5336 // Try to steal from other queues that have work
5337 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5338 NOT_PRODUCT(num_steals++;)
5339 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5340 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5341 // Do scanning work
5342 obj_to_scan->oop_iterate(keep_alive);
5343 // Loop around, finish this work, and try to steal some more
5344 } else if (terminator()->offer_termination()) {
5345 break; // nirvana from the infinite cycle
5346 }
5347 }
5348 NOT_PRODUCT(
5349 if (PrintCMSStatistics != 0) {
5350 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5351 }
5352 )
5353 }
5354
5355 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5356 {
5357 GenCollectedHeap* gch = GenCollectedHeap::heap();
5358 WorkGang* workers = gch->workers();
5359 assert(workers != NULL, "Need parallel worker threads.");
5360 CMSRefProcTaskProxy rp_task(task, &_collector,
5361 _collector.ref_processor()->span(),
5362 _collector.markBitMap(),
5363 workers, _collector.task_queues());
5364 workers->run_task(&rp_task);
5365 }
5366
5367 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5368 {
5369
5370 GenCollectedHeap* gch = GenCollectedHeap::heap();
5371 WorkGang* workers = gch->workers();
5372 assert(workers != NULL, "Need parallel worker threads.");
5373 CMSRefEnqueueTaskProxy enq_task(task);
5374 workers->run_task(&enq_task);
5375 }
5376
5377 void CMSCollector::refProcessingWork() {
5378 ResourceMark rm;
5379 HandleMark hm;
5380
5381 ReferenceProcessor* rp = ref_processor();
5382 assert(rp->span().equals(_span), "Spans should be equal");
5383 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5384 // Process weak references.
5385 rp->setup_policy(false);
5386 verify_work_stacks_empty();
5387
5388 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5389 &_markStack, false /* !preclean */);
5390 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5391 _span, &_markBitMap, &_markStack,
5392 &cmsKeepAliveClosure, false /* !preclean */);
5393 {
5394 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm);
5395
5396 ReferenceProcessorStats stats;
5397 if (rp->processing_is_mt()) {
5398 // Set the degree of MT here. If the discovery is done MT, there
5399 // may have been a different number of threads doing the discovery
5400 // and a different number of discovered lists may have Ref objects.
5401 // That is OK as long as the Reference lists are balanced (see
5402 // balance_all_queues() and balance_queues()).
5403 GenCollectedHeap* gch = GenCollectedHeap::heap();
5404 uint active_workers = ParallelGCThreads;
5405 WorkGang* workers = gch->workers();
5406 if (workers != NULL) {
5407 active_workers = workers->active_workers();
5408 // The expectation is that active_workers will have already
5409 // been set to a reasonable value. If it has not been set,
5410 // investigate.
5411 assert(active_workers > 0, "Should have been set during scavenge");
5412 }
5413 rp->set_active_mt_degree(active_workers);
5414 CMSRefProcTaskExecutor task_executor(*this);
5415 stats = rp->process_discovered_references(&_is_alive_closure,
5416 &cmsKeepAliveClosure,
5417 &cmsDrainMarkingStackClosure,
5418 &task_executor,
5419 _gc_timer_cm);
5420 } else {
5421 stats = rp->process_discovered_references(&_is_alive_closure,
5422 &cmsKeepAliveClosure,
5423 &cmsDrainMarkingStackClosure,
5424 NULL,
5425 _gc_timer_cm);
5426 }
5427 _gc_tracer_cm->report_gc_reference_stats(stats);
5428
5429 }
5430
5431 // This is the point where the entire marking should have completed.
5432 verify_work_stacks_empty();
5433
5434 if (should_unload_classes()) {
5435 {
5436 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm);
5437
5438 // Unload classes and purge the SystemDictionary.
5439 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5440
5441 // Unload nmethods.
5442 CodeCache::do_unloading(&_is_alive_closure, purged_class);
5443
5444 // Prune dead klasses from subklass/sibling/implementor lists.
5445 Klass::clean_weak_klass_links(&_is_alive_closure);
5446 }
5447
5448 {
5449 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm);
5450 // Clean up unreferenced symbols in symbol table.
5451 SymbolTable::unlink();
5452 }
5453
5454 {
5455 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm);
5456 // Delete entries for dead interned strings.
5457 StringTable::unlink(&_is_alive_closure);
5458 }
5459 }
5460
5461
5462 // Restore any preserved marks as a result of mark stack or
5463 // work queue overflow
5464 restore_preserved_marks_if_any(); // done single-threaded for now
5465
5466 rp->set_enqueuing_is_done(true);
5467 if (rp->processing_is_mt()) {
5468 rp->balance_all_queues();
5469 CMSRefProcTaskExecutor task_executor(*this);
5470 rp->enqueue_discovered_references(&task_executor);
5471 } else {
5472 rp->enqueue_discovered_references(NULL);
5473 }
5474 rp->verify_no_references_recorded();
5475 assert(!rp->discovery_enabled(), "should have been disabled");
5476 }
5477
5478 #ifndef PRODUCT
5479 void CMSCollector::check_correct_thread_executing() {
5480 Thread* t = Thread::current();
5481 // Only the VM thread or the CMS thread should be here.
5482 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5483 "Unexpected thread type");
5484 // If this is the vm thread, the foreground process
5485 // should not be waiting. Note that _foregroundGCIsActive is
5486 // true while the foreground collector is waiting.
5487 if (_foregroundGCShouldWait) {
5488 // We cannot be the VM thread
5489 assert(t->is_ConcurrentGC_thread(),
5490 "Should be CMS thread");
5491 } else {
5492 // We can be the CMS thread only if we are in a stop-world
5493 // phase of CMS collection.
5494 if (t->is_ConcurrentGC_thread()) {
5495 assert(_collectorState == InitialMarking ||
5496 _collectorState == FinalMarking,
5497 "Should be a stop-world phase");
5498 // The CMS thread should be holding the CMS_token.
5499 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5500 "Potential interference with concurrently "
5501 "executing VM thread");
5502 }
5503 }
5504 }
5505 #endif
5506
5507 void CMSCollector::sweep() {
5508 assert(_collectorState == Sweeping, "just checking");
5509 check_correct_thread_executing();
5510 verify_work_stacks_empty();
5511 verify_overflow_empty();
5512 increment_sweep_count();
5513 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5514
5515 _inter_sweep_timer.stop();
5516 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5517
5518 assert(!_intra_sweep_timer.is_active(), "Should not be active");
5519 _intra_sweep_timer.reset();
5520 _intra_sweep_timer.start();
5521 {
5522 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5523 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
5524 // First sweep the old gen
5525 {
5526 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5527 bitMapLock());
5528 sweepWork(_cmsGen);
5529 }
5530
5531 // Update Universe::_heap_*_at_gc figures.
5532 // We need all the free list locks to make the abstract state
5533 // transition from Sweeping to Resetting. See detailed note
5534 // further below.
5535 {
5536 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5537 // Update heap occupancy information which is used as
5538 // input to soft ref clearing policy at the next gc.
5539 Universe::update_heap_info_at_gc();
5540 _collectorState = Resizing;
5541 }
5542 }
5543 verify_work_stacks_empty();
5544 verify_overflow_empty();
5545
5546 if (should_unload_classes()) {
5547 // Delay purge to the beginning of the next safepoint. Metaspace::contains
5548 // requires that the virtual spaces are stable and not deleted.
5549 ClassLoaderDataGraph::set_should_purge(true);
5550 }
5551
5552 _intra_sweep_timer.stop();
5553 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5554
5555 _inter_sweep_timer.reset();
5556 _inter_sweep_timer.start();
5557
5558 // We need to use a monotonically non-decreasing time in ms
5559 // or we will see time-warp warnings and os::javaTimeMillis()
5560 // does not guarantee monotonicity.
5561 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5562 update_time_of_last_gc(now);
5563
5564 // NOTE on abstract state transitions:
5565 // Mutators allocate-live and/or mark the mod-union table dirty
5566 // based on the state of the collection. The former is done in
5567 // the interval [Marking, Sweeping] and the latter in the interval
5568 // [Marking, Sweeping). Thus the transitions into the Marking state
5569 // and out of the Sweeping state must be synchronously visible
5570 // globally to the mutators.
5571 // The transition into the Marking state happens with the world
5572 // stopped so the mutators will globally see it. Sweeping is
5573 // done asynchronously by the background collector so the transition
5574 // from the Sweeping state to the Resizing state must be done
5575 // under the freelistLock (as is the check for whether to
5576 // allocate-live and whether to dirty the mod-union table).
5577 assert(_collectorState == Resizing, "Change of collector state to"
5578 " Resizing must be done under the freelistLocks (plural)");
5579
5580 // Now that sweeping has been completed, we clear
5581 // the incremental_collection_failed flag,
5582 // thus inviting a younger gen collection to promote into
5583 // this generation. If such a promotion may still fail,
5584 // the flag will be set again when a young collection is
5585 // attempted.
5586 GenCollectedHeap* gch = GenCollectedHeap::heap();
5587 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
5588 gch->update_full_collections_completed(_collection_count_start);
5589 }
5590
5591 // FIX ME!!! Looks like this belongs in CFLSpace, with
5592 // CMSGen merely delegating to it.
5593 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5594 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5595 HeapWord* minAddr = _cmsSpace->bottom();
5596 HeapWord* largestAddr =
5597 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5598 if (largestAddr == NULL) {
5599 // The dictionary appears to be empty. In this case
5600 // try to coalesce at the end of the heap.
5601 largestAddr = _cmsSpace->end();
5602 }
5603 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5604 size_t nearLargestOffset =
5605 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5606 if (PrintFLSStatistics != 0) {
5607 gclog_or_tty->print_cr(
5608 "CMS: Large Block: " PTR_FORMAT ";"
5609 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5610 p2i(largestAddr),
5611 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5612 }
5613 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5614 }
5615
5616 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5617 return addr >= _cmsSpace->nearLargestChunk();
5618 }
5619
5620 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5621 return _cmsSpace->find_chunk_at_end();
5622 }
5623
5624 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation,
5625 bool full) {
5626 // If the young generation has been collected, gather any statistics
5627 // that are of interest at this point.
5628 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation);
5629 if (!full && current_is_young) {
5630 // Gather statistics on the young generation collection.
5631 collector()->stats().record_gc0_end(used());
5632 }
5633 }
5634
5635 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) {
5636 // We iterate over the space(s) underlying this generation,
5637 // checking the mark bit map to see if the bits corresponding
5638 // to specific blocks are marked or not. Blocks that are
5639 // marked are live and are not swept up. All remaining blocks
5640 // are swept up, with coalescing on-the-fly as we sweep up
5641 // contiguous free and/or garbage blocks:
5642 // We need to ensure that the sweeper synchronizes with allocators
5643 // and stop-the-world collectors. In particular, the following
5644 // locks are used:
5645 // . CMS token: if this is held, a stop the world collection cannot occur
5646 // . freelistLock: if this is held no allocation can occur from this
5647 // generation by another thread
5648 // . bitMapLock: if this is held, no other thread can access or update
5649 //
5650
5651 // Note that we need to hold the freelistLock if we use
5652 // block iterate below; else the iterator might go awry if
5653 // a mutator (or promotion) causes block contents to change
5654 // (for instance if the allocator divvies up a block).
5655 // If we hold the free list lock, for all practical purposes
5656 // young generation GC's can't occur (they'll usually need to
5657 // promote), so we might as well prevent all young generation
5658 // GC's while we do a sweeping step. For the same reason, we might
5659 // as well take the bit map lock for the entire duration
5660
5661 // check that we hold the requisite locks
5662 assert(have_cms_token(), "Should hold cms token");
5663 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5664 assert_lock_strong(old_gen->freelistLock());
5665 assert_lock_strong(bitMapLock());
5666
5667 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5668 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
5669 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5670 _inter_sweep_estimate.padded_average(),
5671 _intra_sweep_estimate.padded_average());
5672 old_gen->setNearLargestChunk();
5673
5674 {
5675 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield);
5676 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5677 // We need to free-up/coalesce garbage/blocks from a
5678 // co-terminal free run. This is done in the SweepClosure
5679 // destructor; so, do not remove this scope, else the
5680 // end-of-sweep-census below will be off by a little bit.
5681 }
5682 old_gen->cmsSpace()->sweep_completed();
5683 old_gen->cmsSpace()->endSweepFLCensus(sweep_count());
5684 if (should_unload_classes()) { // unloaded classes this cycle,
5685 _concurrent_cycles_since_last_unload = 0; // ... reset count
5686 } else { // did not unload classes,
5687 _concurrent_cycles_since_last_unload++; // ... increment count
5688 }
5689 }
5690
5691 // Reset CMS data structures (for now just the marking bit map)
5692 // preparatory for the next cycle.
5693 void CMSCollector::reset_concurrent() {
5694 CMSTokenSyncWithLocks ts(true, bitMapLock());
5695
5696 // If the state is not "Resetting", the foreground thread
5697 // has done a collection and the resetting.
5698 if (_collectorState != Resetting) {
5699 assert(_collectorState == Idling, "The state should only change"
5700 " because the foreground collector has finished the collection");
5701 return;
5702 }
5703
5704 // Clear the mark bitmap (no grey objects to start with)
5705 // for the next cycle.
5706 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5707 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
5708
5709 HeapWord* curAddr = _markBitMap.startWord();
5710 while (curAddr < _markBitMap.endWord()) {
5711 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
5712 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5713 _markBitMap.clear_large_range(chunk);
5714 if (ConcurrentMarkSweepThread::should_yield() &&
5715 !foregroundGCIsActive() &&
5716 CMSYield) {
5717 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5718 "CMS thread should hold CMS token");
5719 assert_lock_strong(bitMapLock());
5720 bitMapLock()->unlock();
5721 ConcurrentMarkSweepThread::desynchronize(true);
5722 stopTimer();
5723 if (PrintCMSStatistics != 0) {
5724 incrementYields();
5725 }
5726
5727 // See the comment in coordinator_yield()
5728 for (unsigned i = 0; i < CMSYieldSleepCount &&
5729 ConcurrentMarkSweepThread::should_yield() &&
5730 !CMSCollector::foregroundGCIsActive(); ++i) {
5731 os::sleep(Thread::current(), 1, false);
5732 }
5733
5734 ConcurrentMarkSweepThread::synchronize(true);
5735 bitMapLock()->lock_without_safepoint_check();
5736 startTimer();
5737 }
5738 curAddr = chunk.end();
5739 }
5740 // A successful mostly concurrent collection has been done.
5741 // Because only the full (i.e., concurrent mode failure) collections
5742 // are being measured for gc overhead limits, clean the "near" flag
5743 // and count.
5744 size_policy()->reset_gc_overhead_limit_count();
5745 _collectorState = Idling;
5746
5747 register_gc_end();
5748 }
5749
5750 // Same as above but for STW paths
5751 void CMSCollector::reset_stw() {
5752 // already have the lock
5753 assert(_collectorState == Resetting, "just checking");
5754 assert_lock_strong(bitMapLock());
5755 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id());
5756 _markBitMap.clear_all();
5757 _collectorState = Idling;
5758 register_gc_end();
5759 }
5760
5761 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5762 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5763 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL);
5764 TraceCollectorStats tcs(counters());
5765
5766 switch (op) {
5767 case CMS_op_checkpointRootsInitial: {
5768 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5769 checkpointRootsInitial();
5770 if (PrintGC) {
5771 _cmsGen->printOccupancy("initial-mark");
5772 }
5773 break;
5774 }
5775 case CMS_op_checkpointRootsFinal: {
5776 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5777 checkpointRootsFinal();
5778 if (PrintGC) {
5779 _cmsGen->printOccupancy("remark");
5780 }
5781 break;
5782 }
5783 default:
5784 fatal("No such CMS_op");
5785 }
5786 }
5787
5788 #ifndef PRODUCT
5789 size_t const CMSCollector::skip_header_HeapWords() {
5790 return FreeChunk::header_size();
5791 }
5792
5793 // Try and collect here conditions that should hold when
5794 // CMS thread is exiting. The idea is that the foreground GC
5795 // thread should not be blocked if it wants to terminate
5796 // the CMS thread and yet continue to run the VM for a while
5797 // after that.
5798 void CMSCollector::verify_ok_to_terminate() const {
5799 assert(Thread::current()->is_ConcurrentGC_thread(),
5800 "should be called by CMS thread");
5801 assert(!_foregroundGCShouldWait, "should be false");
5802 // We could check here that all the various low-level locks
5803 // are not held by the CMS thread, but that is overkill; see
5804 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5805 // is checked.
5806 }
5807 #endif
5808
5809 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5810 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5811 "missing Printezis mark?");
5812 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5813 size_t size = pointer_delta(nextOneAddr + 1, addr);
5814 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5815 "alignment problem");
5816 assert(size >= 3, "Necessary for Printezis marks to work");
5817 return size;
5818 }
5819
5820 // A variant of the above (block_size_using_printezis_bits()) except
5821 // that we return 0 if the P-bits are not yet set.
5822 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5823 if (_markBitMap.isMarked(addr + 1)) {
5824 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5825 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5826 size_t size = pointer_delta(nextOneAddr + 1, addr);
5827 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5828 "alignment problem");
5829 assert(size >= 3, "Necessary for Printezis marks to work");
5830 return size;
5831 }
5832 return 0;
5833 }
5834
5835 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5836 size_t sz = 0;
5837 oop p = (oop)addr;
5838 if (p->klass_or_null() != NULL) {
5839 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5840 } else {
5841 sz = block_size_using_printezis_bits(addr);
5842 }
5843 assert(sz > 0, "size must be nonzero");
5844 HeapWord* next_block = addr + sz;
5845 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
5846 CardTableModRefBS::card_size);
5847 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
5848 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
5849 "must be different cards");
5850 return next_card;
5851 }
5852
5853
5854 // CMS Bit Map Wrapper /////////////////////////////////////////
5855
5856 // Construct a CMS bit map infrastructure, but don't create the
5857 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5858 // further below.
5859 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5860 _bm(),
5861 _shifter(shifter),
5862 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5863 Monitor::_safepoint_check_sometimes) : NULL)
5864 {
5865 _bmStartWord = 0;
5866 _bmWordSize = 0;
5867 }
5868
5869 bool CMSBitMap::allocate(MemRegion mr) {
5870 _bmStartWord = mr.start();
5871 _bmWordSize = mr.word_size();
5872 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5873 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5874 if (!brs.is_reserved()) {
5875 warning("CMS bit map allocation failure");
5876 return false;
5877 }
5878 // For now we'll just commit all of the bit map up front.
5879 // Later on we'll try to be more parsimonious with swap.
5880 if (!_virtual_space.initialize(brs, brs.size())) {
5881 warning("CMS bit map backing store failure");
5882 return false;
5883 }
5884 assert(_virtual_space.committed_size() == brs.size(),
5885 "didn't reserve backing store for all of CMS bit map?");
5886 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
5887 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5888 _bmWordSize, "inconsistency in bit map sizing");
5889 _bm.set_size(_bmWordSize >> _shifter);
5890
5891 // bm.clear(); // can we rely on getting zero'd memory? verify below
5892 assert(isAllClear(),
5893 "Expected zero'd memory from ReservedSpace constructor");
5894 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5895 "consistency check");
5896 return true;
5897 }
5898
5899 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5900 HeapWord *next_addr, *end_addr, *last_addr;
5901 assert_locked();
5902 assert(covers(mr), "out-of-range error");
5903 // XXX assert that start and end are appropriately aligned
5904 for (next_addr = mr.start(), end_addr = mr.end();
5905 next_addr < end_addr; next_addr = last_addr) {
5906 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5907 last_addr = dirty_region.end();
5908 if (!dirty_region.is_empty()) {
5909 cl->do_MemRegion(dirty_region);
5910 } else {
5911 assert(last_addr == end_addr, "program logic");
5912 return;
5913 }
5914 }
5915 }
5916
5917 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5918 _bm.print_on_error(st, prefix);
5919 }
5920
5921 #ifndef PRODUCT
5922 void CMSBitMap::assert_locked() const {
5923 CMSLockVerifier::assert_locked(lock());
5924 }
5925
5926 bool CMSBitMap::covers(MemRegion mr) const {
5927 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5928 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5929 "size inconsistency");
5930 return (mr.start() >= _bmStartWord) &&
5931 (mr.end() <= endWord());
5932 }
5933
5934 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5935 return (start >= _bmStartWord && (start + size) <= endWord());
5936 }
5937
5938 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5939 // verify that there are no 1 bits in the interval [left, right)
5940 FalseBitMapClosure falseBitMapClosure;
5941 iterate(&falseBitMapClosure, left, right);
5942 }
5943
5944 void CMSBitMap::region_invariant(MemRegion mr)
5945 {
5946 assert_locked();
5947 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5948 assert(!mr.is_empty(), "unexpected empty region");
5949 assert(covers(mr), "mr should be covered by bit map");
5950 // convert address range into offset range
5951 size_t start_ofs = heapWordToOffset(mr.start());
5952 // Make sure that end() is appropriately aligned
5953 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
5954 (1 << (_shifter+LogHeapWordSize))),
5955 "Misaligned mr.end()");
5956 size_t end_ofs = heapWordToOffset(mr.end());
5957 assert(end_ofs > start_ofs, "Should mark at least one bit");
5958 }
5959
5960 #endif
5961
5962 bool CMSMarkStack::allocate(size_t size) {
5963 // allocate a stack of the requisite depth
5964 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5965 size * sizeof(oop)));
5966 if (!rs.is_reserved()) {
5967 warning("CMSMarkStack allocation failure");
5968 return false;
5969 }
5970 if (!_virtual_space.initialize(rs, rs.size())) {
5971 warning("CMSMarkStack backing store failure");
5972 return false;
5973 }
5974 assert(_virtual_space.committed_size() == rs.size(),
5975 "didn't reserve backing store for all of CMS stack?");
5976 _base = (oop*)(_virtual_space.low());
5977 _index = 0;
5978 _capacity = size;
5979 NOT_PRODUCT(_max_depth = 0);
5980 return true;
5981 }
5982
5983 // XXX FIX ME !!! In the MT case we come in here holding a
5984 // leaf lock. For printing we need to take a further lock
5985 // which has lower rank. We need to recalibrate the two
5986 // lock-ranks involved in order to be able to print the
5987 // messages below. (Or defer the printing to the caller.
5988 // For now we take the expedient path of just disabling the
5989 // messages for the problematic case.)
5990 void CMSMarkStack::expand() {
5991 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
5992 if (_capacity == MarkStackSizeMax) {
5993 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
5994 // We print a warning message only once per CMS cycle.
5995 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
5996 }
5997 return;
5998 }
5999 // Double capacity if possible
6000 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6001 // Do not give up existing stack until we have managed to
6002 // get the double capacity that we desired.
6003 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6004 new_capacity * sizeof(oop)));
6005 if (rs.is_reserved()) {
6006 // Release the backing store associated with old stack
6007 _virtual_space.release();
6008 // Reinitialize virtual space for new stack
6009 if (!_virtual_space.initialize(rs, rs.size())) {
6010 fatal("Not enough swap for expanded marking stack");
6011 }
6012 _base = (oop*)(_virtual_space.low());
6013 _index = 0;
6014 _capacity = new_capacity;
6015 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6016 // Failed to double capacity, continue;
6017 // we print a detail message only once per CMS cycle.
6018 gclog_or_tty->print(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to "
6019 SIZE_FORMAT "K",
6020 _capacity / K, new_capacity / K);
6021 }
6022 }
6023
6024
6025 // Closures
6026 // XXX: there seems to be a lot of code duplication here;
6027 // should refactor and consolidate common code.
6028
6029 // This closure is used to mark refs into the CMS generation in
6030 // the CMS bit map. Called at the first checkpoint. This closure
6031 // assumes that we do not need to re-mark dirty cards; if the CMS
6032 // generation on which this is used is not an oldest
6033 // generation then this will lose younger_gen cards!
6034
6035 MarkRefsIntoClosure::MarkRefsIntoClosure(
6036 MemRegion span, CMSBitMap* bitMap):
6037 _span(span),
6038 _bitMap(bitMap)
6039 {
6040 assert(ref_processor() == NULL, "deliberately left NULL");
6041 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6042 }
6043
6044 void MarkRefsIntoClosure::do_oop(oop obj) {
6045 // if p points into _span, then mark corresponding bit in _markBitMap
6046 assert(obj->is_oop(), "expected an oop");
6047 HeapWord* addr = (HeapWord*)obj;
6048 if (_span.contains(addr)) {
6049 // this should be made more efficient
6050 _bitMap->mark(addr);
6051 }
6052 }
6053
6054 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6055 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6056
6057 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
6058 MemRegion span, CMSBitMap* bitMap):
6059 _span(span),
6060 _bitMap(bitMap)
6061 {
6062 assert(ref_processor() == NULL, "deliberately left NULL");
6063 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6064 }
6065
6066 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
6067 // if p points into _span, then mark corresponding bit in _markBitMap
6068 assert(obj->is_oop(), "expected an oop");
6069 HeapWord* addr = (HeapWord*)obj;
6070 if (_span.contains(addr)) {
6071 // this should be made more efficient
6072 _bitMap->par_mark(addr);
6073 }
6074 }
6075
6076 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6077 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6078
6079 // A variant of the above, used for CMS marking verification.
6080 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6081 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6082 _span(span),
6083 _verification_bm(verification_bm),
6084 _cms_bm(cms_bm)
6085 {
6086 assert(ref_processor() == NULL, "deliberately left NULL");
6087 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6088 }
6089
6090 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6091 // if p points into _span, then mark corresponding bit in _markBitMap
6092 assert(obj->is_oop(), "expected an oop");
6093 HeapWord* addr = (HeapWord*)obj;
6094 if (_span.contains(addr)) {
6095 _verification_bm->mark(addr);
6096 if (!_cms_bm->isMarked(addr)) {
6097 oop(addr)->print();
6098 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6099 fatal("... aborting");
6100 }
6101 }
6102 }
6103
6104 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6105 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6106
6107 //////////////////////////////////////////////////
6108 // MarkRefsIntoAndScanClosure
6109 //////////////////////////////////////////////////
6110
6111 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6112 ReferenceProcessor* rp,
6113 CMSBitMap* bit_map,
6114 CMSBitMap* mod_union_table,
6115 CMSMarkStack* mark_stack,
6116 CMSCollector* collector,
6117 bool should_yield,
6118 bool concurrent_precleaning):
6119 _collector(collector),
6120 _span(span),
6121 _bit_map(bit_map),
6122 _mark_stack(mark_stack),
6123 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6124 mark_stack, concurrent_precleaning),
6125 _yield(should_yield),
6126 _concurrent_precleaning(concurrent_precleaning),
6127 _freelistLock(NULL)
6128 {
6129 // FIXME: Should initialize in base class constructor.
6130 assert(rp != NULL, "ref_processor shouldn't be NULL");
6131 set_ref_processor_internal(rp);
6132 }
6133
6134 // This closure is used to mark refs into the CMS generation at the
6135 // second (final) checkpoint, and to scan and transitively follow
6136 // the unmarked oops. It is also used during the concurrent precleaning
6137 // phase while scanning objects on dirty cards in the CMS generation.
6138 // The marks are made in the marking bit map and the marking stack is
6139 // used for keeping the (newly) grey objects during the scan.
6140 // The parallel version (Par_...) appears further below.
6141 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6142 if (obj != NULL) {
6143 assert(obj->is_oop(), "expected an oop");
6144 HeapWord* addr = (HeapWord*)obj;
6145 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6146 assert(_collector->overflow_list_is_empty(),
6147 "overflow list should be empty");
6148 if (_span.contains(addr) &&
6149 !_bit_map->isMarked(addr)) {
6150 // mark bit map (object is now grey)
6151 _bit_map->mark(addr);
6152 // push on marking stack (stack should be empty), and drain the
6153 // stack by applying this closure to the oops in the oops popped
6154 // from the stack (i.e. blacken the grey objects)
6155 bool res = _mark_stack->push(obj);
6156 assert(res, "Should have space to push on empty stack");
6157 do {
6158 oop new_oop = _mark_stack->pop();
6159 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6160 assert(_bit_map->isMarked((HeapWord*)new_oop),
6161 "only grey objects on this stack");
6162 // iterate over the oops in this oop, marking and pushing
6163 // the ones in CMS heap (i.e. in _span).
6164 new_oop->oop_iterate(&_pushAndMarkClosure);
6165 // check if it's time to yield
6166 do_yield_check();
6167 } while (!_mark_stack->isEmpty() ||
6168 (!_concurrent_precleaning && take_from_overflow_list()));
6169 // if marking stack is empty, and we are not doing this
6170 // during precleaning, then check the overflow list
6171 }
6172 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6173 assert(_collector->overflow_list_is_empty(),
6174 "overflow list was drained above");
6175
6176 assert(_collector->no_preserved_marks(),
6177 "All preserved marks should have been restored above");
6178 }
6179 }
6180
6181 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6182 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6183
6184 void MarkRefsIntoAndScanClosure::do_yield_work() {
6185 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6186 "CMS thread should hold CMS token");
6187 assert_lock_strong(_freelistLock);
6188 assert_lock_strong(_bit_map->lock());
6189 // relinquish the free_list_lock and bitMaplock()
6190 _bit_map->lock()->unlock();
6191 _freelistLock->unlock();
6192 ConcurrentMarkSweepThread::desynchronize(true);
6193 _collector->stopTimer();
6194 if (PrintCMSStatistics != 0) {
6195 _collector->incrementYields();
6196 }
6197
6198 // See the comment in coordinator_yield()
6199 for (unsigned i = 0;
6200 i < CMSYieldSleepCount &&
6201 ConcurrentMarkSweepThread::should_yield() &&
6202 !CMSCollector::foregroundGCIsActive();
6203 ++i) {
6204 os::sleep(Thread::current(), 1, false);
6205 }
6206
6207 ConcurrentMarkSweepThread::synchronize(true);
6208 _freelistLock->lock_without_safepoint_check();
6209 _bit_map->lock()->lock_without_safepoint_check();
6210 _collector->startTimer();
6211 }
6212
6213 ///////////////////////////////////////////////////////////
6214 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6215 // MarkRefsIntoAndScanClosure
6216 ///////////////////////////////////////////////////////////
6217 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6218 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6219 CMSBitMap* bit_map, OopTaskQueue* work_queue):
6220 _span(span),
6221 _bit_map(bit_map),
6222 _work_queue(work_queue),
6223 _low_water_mark(MIN2((work_queue->max_elems()/4),
6224 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6225 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
6226 {
6227 // FIXME: Should initialize in base class constructor.
6228 assert(rp != NULL, "ref_processor shouldn't be NULL");
6229 set_ref_processor_internal(rp);
6230 }
6231
6232 // This closure is used to mark refs into the CMS generation at the
6233 // second (final) checkpoint, and to scan and transitively follow
6234 // the unmarked oops. The marks are made in the marking bit map and
6235 // the work_queue is used for keeping the (newly) grey objects during
6236 // the scan phase whence they are also available for stealing by parallel
6237 // threads. Since the marking bit map is shared, updates are
6238 // synchronized (via CAS).
6239 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6240 if (obj != NULL) {
6241 // Ignore mark word because this could be an already marked oop
6242 // that may be chained at the end of the overflow list.
6243 assert(obj->is_oop(true), "expected an oop");
6244 HeapWord* addr = (HeapWord*)obj;
6245 if (_span.contains(addr) &&
6246 !_bit_map->isMarked(addr)) {
6247 // mark bit map (object will become grey):
6248 // It is possible for several threads to be
6249 // trying to "claim" this object concurrently;
6250 // the unique thread that succeeds in marking the
6251 // object first will do the subsequent push on
6252 // to the work queue (or overflow list).
6253 if (_bit_map->par_mark(addr)) {
6254 // push on work_queue (which may not be empty), and trim the
6255 // queue to an appropriate length by applying this closure to
6256 // the oops in the oops popped from the stack (i.e. blacken the
6257 // grey objects)
6258 bool res = _work_queue->push(obj);
6259 assert(res, "Low water mark should be less than capacity?");
6260 trim_queue(_low_water_mark);
6261 } // Else, another thread claimed the object
6262 }
6263 }
6264 }
6265
6266 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6267 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6268
6269 // This closure is used to rescan the marked objects on the dirty cards
6270 // in the mod union table and the card table proper.
6271 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6272 oop p, MemRegion mr) {
6273
6274 size_t size = 0;
6275 HeapWord* addr = (HeapWord*)p;
6276 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6277 assert(_span.contains(addr), "we are scanning the CMS generation");
6278 // check if it's time to yield
6279 if (do_yield_check()) {
6280 // We yielded for some foreground stop-world work,
6281 // and we have been asked to abort this ongoing preclean cycle.
6282 return 0;
6283 }
6284 if (_bitMap->isMarked(addr)) {
6285 // it's marked; is it potentially uninitialized?
6286 if (p->klass_or_null() != NULL) {
6287 // an initialized object; ignore mark word in verification below
6288 // since we are running concurrent with mutators
6289 assert(p->is_oop(true), "should be an oop");
6290 if (p->is_objArray()) {
6291 // objArrays are precisely marked; restrict scanning
6292 // to dirty cards only.
6293 size = CompactibleFreeListSpace::adjustObjectSize(
6294 p->oop_iterate_size(_scanningClosure, mr));
6295 } else {
6296 // A non-array may have been imprecisely marked; we need
6297 // to scan object in its entirety.
6298 size = CompactibleFreeListSpace::adjustObjectSize(
6299 p->oop_iterate_size(_scanningClosure));
6300 }
6301 #ifdef ASSERT
6302 size_t direct_size =
6303 CompactibleFreeListSpace::adjustObjectSize(p->size());
6304 assert(size == direct_size, "Inconsistency in size");
6305 assert(size >= 3, "Necessary for Printezis marks to work");
6306 if (!_bitMap->isMarked(addr+1)) {
6307 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6308 } else {
6309 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6310 assert(_bitMap->isMarked(addr+size-1),
6311 "inconsistent Printezis mark");
6312 }
6313 #endif // ASSERT
6314 } else {
6315 // An uninitialized object.
6316 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6317 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6318 size = pointer_delta(nextOneAddr + 1, addr);
6319 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6320 "alignment problem");
6321 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6322 // will dirty the card when the klass pointer is installed in the
6323 // object (signaling the completion of initialization).
6324 }
6325 } else {
6326 // Either a not yet marked object or an uninitialized object
6327 if (p->klass_or_null() == NULL) {
6328 // An uninitialized object, skip to the next card, since
6329 // we may not be able to read its P-bits yet.
6330 assert(size == 0, "Initial value");
6331 } else {
6332 // An object not (yet) reached by marking: we merely need to
6333 // compute its size so as to go look at the next block.
6334 assert(p->is_oop(true), "should be an oop");
6335 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6336 }
6337 }
6338 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6339 return size;
6340 }
6341
6342 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6343 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6344 "CMS thread should hold CMS token");
6345 assert_lock_strong(_freelistLock);
6346 assert_lock_strong(_bitMap->lock());
6347 // relinquish the free_list_lock and bitMaplock()
6348 _bitMap->lock()->unlock();
6349 _freelistLock->unlock();
6350 ConcurrentMarkSweepThread::desynchronize(true);
6351 _collector->stopTimer();
6352 if (PrintCMSStatistics != 0) {
6353 _collector->incrementYields();
6354 }
6355
6356 // See the comment in coordinator_yield()
6357 for (unsigned i = 0; i < CMSYieldSleepCount &&
6358 ConcurrentMarkSweepThread::should_yield() &&
6359 !CMSCollector::foregroundGCIsActive(); ++i) {
6360 os::sleep(Thread::current(), 1, false);
6361 }
6362
6363 ConcurrentMarkSweepThread::synchronize(true);
6364 _freelistLock->lock_without_safepoint_check();
6365 _bitMap->lock()->lock_without_safepoint_check();
6366 _collector->startTimer();
6367 }
6368
6369
6370 //////////////////////////////////////////////////////////////////
6371 // SurvivorSpacePrecleanClosure
6372 //////////////////////////////////////////////////////////////////
6373 // This (single-threaded) closure is used to preclean the oops in
6374 // the survivor spaces.
6375 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6376
6377 HeapWord* addr = (HeapWord*)p;
6378 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6379 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6380 assert(p->klass_or_null() != NULL, "object should be initialized");
6381 // an initialized object; ignore mark word in verification below
6382 // since we are running concurrent with mutators
6383 assert(p->is_oop(true), "should be an oop");
6384 // Note that we do not yield while we iterate over
6385 // the interior oops of p, pushing the relevant ones
6386 // on our marking stack.
6387 size_t size = p->oop_iterate_size(_scanning_closure);
6388 do_yield_check();
6389 // Observe that below, we do not abandon the preclean
6390 // phase as soon as we should; rather we empty the
6391 // marking stack before returning. This is to satisfy
6392 // some existing assertions. In general, it may be a
6393 // good idea to abort immediately and complete the marking
6394 // from the grey objects at a later time.
6395 while (!_mark_stack->isEmpty()) {
6396 oop new_oop = _mark_stack->pop();
6397 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6398 assert(_bit_map->isMarked((HeapWord*)new_oop),
6399 "only grey objects on this stack");
6400 // iterate over the oops in this oop, marking and pushing
6401 // the ones in CMS heap (i.e. in _span).
6402 new_oop->oop_iterate(_scanning_closure);
6403 // check if it's time to yield
6404 do_yield_check();
6405 }
6406 unsigned int after_count =
6407 GenCollectedHeap::heap()->total_collections();
6408 bool abort = (_before_count != after_count) ||
6409 _collector->should_abort_preclean();
6410 return abort ? 0 : size;
6411 }
6412
6413 void SurvivorSpacePrecleanClosure::do_yield_work() {
6414 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6415 "CMS thread should hold CMS token");
6416 assert_lock_strong(_bit_map->lock());
6417 // Relinquish the bit map lock
6418 _bit_map->lock()->unlock();
6419 ConcurrentMarkSweepThread::desynchronize(true);
6420 _collector->stopTimer();
6421 if (PrintCMSStatistics != 0) {
6422 _collector->incrementYields();
6423 }
6424
6425 // See the comment in coordinator_yield()
6426 for (unsigned i = 0; i < CMSYieldSleepCount &&
6427 ConcurrentMarkSweepThread::should_yield() &&
6428 !CMSCollector::foregroundGCIsActive(); ++i) {
6429 os::sleep(Thread::current(), 1, false);
6430 }
6431
6432 ConcurrentMarkSweepThread::synchronize(true);
6433 _bit_map->lock()->lock_without_safepoint_check();
6434 _collector->startTimer();
6435 }
6436
6437 // This closure is used to rescan the marked objects on the dirty cards
6438 // in the mod union table and the card table proper. In the parallel
6439 // case, although the bitMap is shared, we do a single read so the
6440 // isMarked() query is "safe".
6441 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6442 // Ignore mark word because we are running concurrent with mutators
6443 assert(p->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p));
6444 HeapWord* addr = (HeapWord*)p;
6445 assert(_span.contains(addr), "we are scanning the CMS generation");
6446 bool is_obj_array = false;
6447 #ifdef ASSERT
6448 if (!_parallel) {
6449 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6450 assert(_collector->overflow_list_is_empty(),
6451 "overflow list should be empty");
6452
6453 }
6454 #endif // ASSERT
6455 if (_bit_map->isMarked(addr)) {
6456 // Obj arrays are precisely marked, non-arrays are not;
6457 // so we scan objArrays precisely and non-arrays in their
6458 // entirety.
6459 if (p->is_objArray()) {
6460 is_obj_array = true;
6461 if (_parallel) {
6462 p->oop_iterate(_par_scan_closure, mr);
6463 } else {
6464 p->oop_iterate(_scan_closure, mr);
6465 }
6466 } else {
6467 if (_parallel) {
6468 p->oop_iterate(_par_scan_closure);
6469 } else {
6470 p->oop_iterate(_scan_closure);
6471 }
6472 }
6473 }
6474 #ifdef ASSERT
6475 if (!_parallel) {
6476 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6477 assert(_collector->overflow_list_is_empty(),
6478 "overflow list should be empty");
6479
6480 }
6481 #endif // ASSERT
6482 return is_obj_array;
6483 }
6484
6485 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6486 MemRegion span,
6487 CMSBitMap* bitMap, CMSMarkStack* markStack,
6488 bool should_yield, bool verifying):
6489 _collector(collector),
6490 _span(span),
6491 _bitMap(bitMap),
6492 _mut(&collector->_modUnionTable),
6493 _markStack(markStack),
6494 _yield(should_yield),
6495 _skipBits(0)
6496 {
6497 assert(_markStack->isEmpty(), "stack should be empty");
6498 _finger = _bitMap->startWord();
6499 _threshold = _finger;
6500 assert(_collector->_restart_addr == NULL, "Sanity check");
6501 assert(_span.contains(_finger), "Out of bounds _finger?");
6502 DEBUG_ONLY(_verifying = verifying;)
6503 }
6504
6505 void MarkFromRootsClosure::reset(HeapWord* addr) {
6506 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6507 assert(_span.contains(addr), "Out of bounds _finger?");
6508 _finger = addr;
6509 _threshold = (HeapWord*)round_to(
6510 (intptr_t)_finger, CardTableModRefBS::card_size);
6511 }
6512
6513 // Should revisit to see if this should be restructured for
6514 // greater efficiency.
6515 bool MarkFromRootsClosure::do_bit(size_t offset) {
6516 if (_skipBits > 0) {
6517 _skipBits--;
6518 return true;
6519 }
6520 // convert offset into a HeapWord*
6521 HeapWord* addr = _bitMap->startWord() + offset;
6522 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6523 "address out of range");
6524 assert(_bitMap->isMarked(addr), "tautology");
6525 if (_bitMap->isMarked(addr+1)) {
6526 // this is an allocated but not yet initialized object
6527 assert(_skipBits == 0, "tautology");
6528 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6529 oop p = oop(addr);
6530 if (p->klass_or_null() == NULL) {
6531 DEBUG_ONLY(if (!_verifying) {)
6532 // We re-dirty the cards on which this object lies and increase
6533 // the _threshold so that we'll come back to scan this object
6534 // during the preclean or remark phase. (CMSCleanOnEnter)
6535 if (CMSCleanOnEnter) {
6536 size_t sz = _collector->block_size_using_printezis_bits(addr);
6537 HeapWord* end_card_addr = (HeapWord*)round_to(
6538 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6539 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6540 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6541 // Bump _threshold to end_card_addr; note that
6542 // _threshold cannot possibly exceed end_card_addr, anyhow.
6543 // This prevents future clearing of the card as the scan proceeds
6544 // to the right.
6545 assert(_threshold <= end_card_addr,
6546 "Because we are just scanning into this object");
6547 if (_threshold < end_card_addr) {
6548 _threshold = end_card_addr;
6549 }
6550 if (p->klass_or_null() != NULL) {
6551 // Redirty the range of cards...
6552 _mut->mark_range(redirty_range);
6553 } // ...else the setting of klass will dirty the card anyway.
6554 }
6555 DEBUG_ONLY(})
6556 return true;
6557 }
6558 }
6559 scanOopsInOop(addr);
6560 return true;
6561 }
6562
6563 // We take a break if we've been at this for a while,
6564 // so as to avoid monopolizing the locks involved.
6565 void MarkFromRootsClosure::do_yield_work() {
6566 // First give up the locks, then yield, then re-lock
6567 // We should probably use a constructor/destructor idiom to
6568 // do this unlock/lock or modify the MutexUnlocker class to
6569 // serve our purpose. XXX
6570 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6571 "CMS thread should hold CMS token");
6572 assert_lock_strong(_bitMap->lock());
6573 _bitMap->lock()->unlock();
6574 ConcurrentMarkSweepThread::desynchronize(true);
6575 _collector->stopTimer();
6576 if (PrintCMSStatistics != 0) {
6577 _collector->incrementYields();
6578 }
6579
6580 // See the comment in coordinator_yield()
6581 for (unsigned i = 0; i < CMSYieldSleepCount &&
6582 ConcurrentMarkSweepThread::should_yield() &&
6583 !CMSCollector::foregroundGCIsActive(); ++i) {
6584 os::sleep(Thread::current(), 1, false);
6585 }
6586
6587 ConcurrentMarkSweepThread::synchronize(true);
6588 _bitMap->lock()->lock_without_safepoint_check();
6589 _collector->startTimer();
6590 }
6591
6592 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6593 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6594 assert(_markStack->isEmpty(),
6595 "should drain stack to limit stack usage");
6596 // convert ptr to an oop preparatory to scanning
6597 oop obj = oop(ptr);
6598 // Ignore mark word in verification below, since we
6599 // may be running concurrent with mutators.
6600 assert(obj->is_oop(true), "should be an oop");
6601 assert(_finger <= ptr, "_finger runneth ahead");
6602 // advance the finger to right end of this object
6603 _finger = ptr + obj->size();
6604 assert(_finger > ptr, "we just incremented it above");
6605 // On large heaps, it may take us some time to get through
6606 // the marking phase. During
6607 // this time it's possible that a lot of mutations have
6608 // accumulated in the card table and the mod union table --
6609 // these mutation records are redundant until we have
6610 // actually traced into the corresponding card.
6611 // Here, we check whether advancing the finger would make
6612 // us cross into a new card, and if so clear corresponding
6613 // cards in the MUT (preclean them in the card-table in the
6614 // future).
6615
6616 DEBUG_ONLY(if (!_verifying) {)
6617 // The clean-on-enter optimization is disabled by default,
6618 // until we fix 6178663.
6619 if (CMSCleanOnEnter && (_finger > _threshold)) {
6620 // [_threshold, _finger) represents the interval
6621 // of cards to be cleared in MUT (or precleaned in card table).
6622 // The set of cards to be cleared is all those that overlap
6623 // with the interval [_threshold, _finger); note that
6624 // _threshold is always kept card-aligned but _finger isn't
6625 // always card-aligned.
6626 HeapWord* old_threshold = _threshold;
6627 assert(old_threshold == (HeapWord*)round_to(
6628 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6629 "_threshold should always be card-aligned");
6630 _threshold = (HeapWord*)round_to(
6631 (intptr_t)_finger, CardTableModRefBS::card_size);
6632 MemRegion mr(old_threshold, _threshold);
6633 assert(!mr.is_empty(), "Control point invariant");
6634 assert(_span.contains(mr), "Should clear within span");
6635 _mut->clear_range(mr);
6636 }
6637 DEBUG_ONLY(})
6638 // Note: the finger doesn't advance while we drain
6639 // the stack below.
6640 PushOrMarkClosure pushOrMarkClosure(_collector,
6641 _span, _bitMap, _markStack,
6642 _finger, this);
6643 bool res = _markStack->push(obj);
6644 assert(res, "Empty non-zero size stack should have space for single push");
6645 while (!_markStack->isEmpty()) {
6646 oop new_oop = _markStack->pop();
6647 // Skip verifying header mark word below because we are
6648 // running concurrent with mutators.
6649 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6650 // now scan this oop's oops
6651 new_oop->oop_iterate(&pushOrMarkClosure);
6652 do_yield_check();
6653 }
6654 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6655 }
6656
6657 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
6658 CMSCollector* collector, MemRegion span,
6659 CMSBitMap* bit_map,
6660 OopTaskQueue* work_queue,
6661 CMSMarkStack* overflow_stack):
6662 _collector(collector),
6663 _whole_span(collector->_span),
6664 _span(span),
6665 _bit_map(bit_map),
6666 _mut(&collector->_modUnionTable),
6667 _work_queue(work_queue),
6668 _overflow_stack(overflow_stack),
6669 _skip_bits(0),
6670 _task(task)
6671 {
6672 assert(_work_queue->size() == 0, "work_queue should be empty");
6673 _finger = span.start();
6674 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
6675 assert(_span.contains(_finger), "Out of bounds _finger?");
6676 }
6677
6678 // Should revisit to see if this should be restructured for
6679 // greater efficiency.
6680 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
6681 if (_skip_bits > 0) {
6682 _skip_bits--;
6683 return true;
6684 }
6685 // convert offset into a HeapWord*
6686 HeapWord* addr = _bit_map->startWord() + offset;
6687 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6688 "address out of range");
6689 assert(_bit_map->isMarked(addr), "tautology");
6690 if (_bit_map->isMarked(addr+1)) {
6691 // this is an allocated object that might not yet be initialized
6692 assert(_skip_bits == 0, "tautology");
6693 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
6694 oop p = oop(addr);
6695 if (p->klass_or_null() == NULL) {
6696 // in the case of Clean-on-Enter optimization, redirty card
6697 // and avoid clearing card by increasing the threshold.
6698 return true;
6699 }
6700 }
6701 scan_oops_in_oop(addr);
6702 return true;
6703 }
6704
6705 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6706 assert(_bit_map->isMarked(ptr), "expected bit to be set");
6707 // Should we assert that our work queue is empty or
6708 // below some drain limit?
6709 assert(_work_queue->size() == 0,
6710 "should drain stack to limit stack usage");
6711 // convert ptr to an oop preparatory to scanning
6712 oop obj = oop(ptr);
6713 // Ignore mark word in verification below, since we
6714 // may be running concurrent with mutators.
6715 assert(obj->is_oop(true), "should be an oop");
6716 assert(_finger <= ptr, "_finger runneth ahead");
6717 // advance the finger to right end of this object
6718 _finger = ptr + obj->size();
6719 assert(_finger > ptr, "we just incremented it above");
6720 // On large heaps, it may take us some time to get through
6721 // the marking phase. During
6722 // this time it's possible that a lot of mutations have
6723 // accumulated in the card table and the mod union table --
6724 // these mutation records are redundant until we have
6725 // actually traced into the corresponding card.
6726 // Here, we check whether advancing the finger would make
6727 // us cross into a new card, and if so clear corresponding
6728 // cards in the MUT (preclean them in the card-table in the
6729 // future).
6730
6731 // The clean-on-enter optimization is disabled by default,
6732 // until we fix 6178663.
6733 if (CMSCleanOnEnter && (_finger > _threshold)) {
6734 // [_threshold, _finger) represents the interval
6735 // of cards to be cleared in MUT (or precleaned in card table).
6736 // The set of cards to be cleared is all those that overlap
6737 // with the interval [_threshold, _finger); note that
6738 // _threshold is always kept card-aligned but _finger isn't
6739 // always card-aligned.
6740 HeapWord* old_threshold = _threshold;
6741 assert(old_threshold == (HeapWord*)round_to(
6742 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6743 "_threshold should always be card-aligned");
6744 _threshold = (HeapWord*)round_to(
6745 (intptr_t)_finger, CardTableModRefBS::card_size);
6746 MemRegion mr(old_threshold, _threshold);
6747 assert(!mr.is_empty(), "Control point invariant");
6748 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6749 _mut->clear_range(mr);
6750 }
6751
6752 // Note: the local finger doesn't advance while we drain
6753 // the stack below, but the global finger sure can and will.
6754 HeapWord** gfa = _task->global_finger_addr();
6755 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
6756 _span, _bit_map,
6757 _work_queue,
6758 _overflow_stack,
6759 _finger,
6760 gfa, this);
6761 bool res = _work_queue->push(obj); // overflow could occur here
6762 assert(res, "Will hold once we use workqueues");
6763 while (true) {
6764 oop new_oop;
6765 if (!_work_queue->pop_local(new_oop)) {
6766 // We emptied our work_queue; check if there's stuff that can
6767 // be gotten from the overflow stack.
6768 if (CMSConcMarkingTask::get_work_from_overflow_stack(
6769 _overflow_stack, _work_queue)) {
6770 do_yield_check();
6771 continue;
6772 } else { // done
6773 break;
6774 }
6775 }
6776 // Skip verifying header mark word below because we are
6777 // running concurrent with mutators.
6778 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6779 // now scan this oop's oops
6780 new_oop->oop_iterate(&pushOrMarkClosure);
6781 do_yield_check();
6782 }
6783 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6784 }
6785
6786 // Yield in response to a request from VM Thread or
6787 // from mutators.
6788 void Par_MarkFromRootsClosure::do_yield_work() {
6789 assert(_task != NULL, "sanity");
6790 _task->yield();
6791 }
6792
6793 // A variant of the above used for verifying CMS marking work.
6794 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6795 MemRegion span,
6796 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6797 CMSMarkStack* mark_stack):
6798 _collector(collector),
6799 _span(span),
6800 _verification_bm(verification_bm),
6801 _cms_bm(cms_bm),
6802 _mark_stack(mark_stack),
6803 _pam_verify_closure(collector, span, verification_bm, cms_bm,
6804 mark_stack)
6805 {
6806 assert(_mark_stack->isEmpty(), "stack should be empty");
6807 _finger = _verification_bm->startWord();
6808 assert(_collector->_restart_addr == NULL, "Sanity check");
6809 assert(_span.contains(_finger), "Out of bounds _finger?");
6810 }
6811
6812 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6813 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6814 assert(_span.contains(addr), "Out of bounds _finger?");
6815 _finger = addr;
6816 }
6817
6818 // Should revisit to see if this should be restructured for
6819 // greater efficiency.
6820 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6821 // convert offset into a HeapWord*
6822 HeapWord* addr = _verification_bm->startWord() + offset;
6823 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6824 "address out of range");
6825 assert(_verification_bm->isMarked(addr), "tautology");
6826 assert(_cms_bm->isMarked(addr), "tautology");
6827
6828 assert(_mark_stack->isEmpty(),
6829 "should drain stack to limit stack usage");
6830 // convert addr to an oop preparatory to scanning
6831 oop obj = oop(addr);
6832 assert(obj->is_oop(), "should be an oop");
6833 assert(_finger <= addr, "_finger runneth ahead");
6834 // advance the finger to right end of this object
6835 _finger = addr + obj->size();
6836 assert(_finger > addr, "we just incremented it above");
6837 // Note: the finger doesn't advance while we drain
6838 // the stack below.
6839 bool res = _mark_stack->push(obj);
6840 assert(res, "Empty non-zero size stack should have space for single push");
6841 while (!_mark_stack->isEmpty()) {
6842 oop new_oop = _mark_stack->pop();
6843 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
6844 // now scan this oop's oops
6845 new_oop->oop_iterate(&_pam_verify_closure);
6846 }
6847 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6848 return true;
6849 }
6850
6851 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6852 CMSCollector* collector, MemRegion span,
6853 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6854 CMSMarkStack* mark_stack):
6855 MetadataAwareOopClosure(collector->ref_processor()),
6856 _collector(collector),
6857 _span(span),
6858 _verification_bm(verification_bm),
6859 _cms_bm(cms_bm),
6860 _mark_stack(mark_stack)
6861 { }
6862
6863 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6864 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6865
6866 // Upon stack overflow, we discard (part of) the stack,
6867 // remembering the least address amongst those discarded
6868 // in CMSCollector's _restart_address.
6869 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6870 // Remember the least grey address discarded
6871 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6872 _collector->lower_restart_addr(ra);
6873 _mark_stack->reset(); // discard stack contents
6874 _mark_stack->expand(); // expand the stack if possible
6875 }
6876
6877 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6878 assert(obj->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6879 HeapWord* addr = (HeapWord*)obj;
6880 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6881 // Oop lies in _span and isn't yet grey or black
6882 _verification_bm->mark(addr); // now grey
6883 if (!_cms_bm->isMarked(addr)) {
6884 oop(addr)->print();
6885 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
6886 p2i(addr));
6887 fatal("... aborting");
6888 }
6889
6890 if (!_mark_stack->push(obj)) { // stack overflow
6891 if (PrintCMSStatistics != 0) {
6892 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
6893 SIZE_FORMAT, _mark_stack->capacity());
6894 }
6895 assert(_mark_stack->isFull(), "Else push should have succeeded");
6896 handle_stack_overflow(addr);
6897 }
6898 // anything including and to the right of _finger
6899 // will be scanned as we iterate over the remainder of the
6900 // bit map
6901 }
6902 }
6903
6904 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6905 MemRegion span,
6906 CMSBitMap* bitMap, CMSMarkStack* markStack,
6907 HeapWord* finger, MarkFromRootsClosure* parent) :
6908 MetadataAwareOopClosure(collector->ref_processor()),
6909 _collector(collector),
6910 _span(span),
6911 _bitMap(bitMap),
6912 _markStack(markStack),
6913 _finger(finger),
6914 _parent(parent)
6915 { }
6916
6917 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
6918 MemRegion span,
6919 CMSBitMap* bit_map,
6920 OopTaskQueue* work_queue,
6921 CMSMarkStack* overflow_stack,
6922 HeapWord* finger,
6923 HeapWord** global_finger_addr,
6924 Par_MarkFromRootsClosure* parent) :
6925 MetadataAwareOopClosure(collector->ref_processor()),
6926 _collector(collector),
6927 _whole_span(collector->_span),
6928 _span(span),
6929 _bit_map(bit_map),
6930 _work_queue(work_queue),
6931 _overflow_stack(overflow_stack),
6932 _finger(finger),
6933 _global_finger_addr(global_finger_addr),
6934 _parent(parent)
6935 { }
6936
6937 // Assumes thread-safe access by callers, who are
6938 // responsible for mutual exclusion.
6939 void CMSCollector::lower_restart_addr(HeapWord* low) {
6940 assert(_span.contains(low), "Out of bounds addr");
6941 if (_restart_addr == NULL) {
6942 _restart_addr = low;
6943 } else {
6944 _restart_addr = MIN2(_restart_addr, low);
6945 }
6946 }
6947
6948 // Upon stack overflow, we discard (part of) the stack,
6949 // remembering the least address amongst those discarded
6950 // in CMSCollector's _restart_address.
6951 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6952 // Remember the least grey address discarded
6953 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6954 _collector->lower_restart_addr(ra);
6955 _markStack->reset(); // discard stack contents
6956 _markStack->expand(); // expand the stack if possible
6957 }
6958
6959 // Upon stack overflow, we discard (part of) the stack,
6960 // remembering the least address amongst those discarded
6961 // in CMSCollector's _restart_address.
6962 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6963 // We need to do this under a mutex to prevent other
6964 // workers from interfering with the work done below.
6965 MutexLockerEx ml(_overflow_stack->par_lock(),
6966 Mutex::_no_safepoint_check_flag);
6967 // Remember the least grey address discarded
6968 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
6969 _collector->lower_restart_addr(ra);
6970 _overflow_stack->reset(); // discard stack contents
6971 _overflow_stack->expand(); // expand the stack if possible
6972 }
6973
6974 void PushOrMarkClosure::do_oop(oop obj) {
6975 // Ignore mark word because we are running concurrent with mutators.
6976 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6977 HeapWord* addr = (HeapWord*)obj;
6978 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
6979 // Oop lies in _span and isn't yet grey or black
6980 _bitMap->mark(addr); // now grey
6981 if (addr < _finger) {
6982 // the bit map iteration has already either passed, or
6983 // sampled, this bit in the bit map; we'll need to
6984 // use the marking stack to scan this oop's oops.
6985 bool simulate_overflow = false;
6986 NOT_PRODUCT(
6987 if (CMSMarkStackOverflowALot &&
6988 _collector->simulate_overflow()) {
6989 // simulate a stack overflow
6990 simulate_overflow = true;
6991 }
6992 )
6993 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
6994 if (PrintCMSStatistics != 0) {
6995 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
6996 SIZE_FORMAT, _markStack->capacity());
6997 }
6998 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
6999 handle_stack_overflow(addr);
7000 }
7001 }
7002 // anything including and to the right of _finger
7003 // will be scanned as we iterate over the remainder of the
7004 // bit map
7005 do_yield_check();
7006 }
7007 }
7008
7009 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7010 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7011
7012 void Par_PushOrMarkClosure::do_oop(oop obj) {
7013 // Ignore mark word because we are running concurrent with mutators.
7014 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
7015 HeapWord* addr = (HeapWord*)obj;
7016 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7017 // Oop lies in _span and isn't yet grey or black
7018 // We read the global_finger (volatile read) strictly after marking oop
7019 bool res = _bit_map->par_mark(addr); // now grey
7020 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7021 // Should we push this marked oop on our stack?
7022 // -- if someone else marked it, nothing to do
7023 // -- if target oop is above global finger nothing to do
7024 // -- if target oop is in chunk and above local finger
7025 // then nothing to do
7026 // -- else push on work queue
7027 if ( !res // someone else marked it, they will deal with it
7028 || (addr >= *gfa) // will be scanned in a later task
7029 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7030 return;
7031 }
7032 // the bit map iteration has already either passed, or
7033 // sampled, this bit in the bit map; we'll need to
7034 // use the marking stack to scan this oop's oops.
7035 bool simulate_overflow = false;
7036 NOT_PRODUCT(
7037 if (CMSMarkStackOverflowALot &&
7038 _collector->simulate_overflow()) {
7039 // simulate a stack overflow
7040 simulate_overflow = true;
7041 }
7042 )
7043 if (simulate_overflow ||
7044 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7045 // stack overflow
7046 if (PrintCMSStatistics != 0) {
7047 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7048 SIZE_FORMAT, _overflow_stack->capacity());
7049 }
7050 // We cannot assert that the overflow stack is full because
7051 // it may have been emptied since.
7052 assert(simulate_overflow ||
7053 _work_queue->size() == _work_queue->max_elems(),
7054 "Else push should have succeeded");
7055 handle_stack_overflow(addr);
7056 }
7057 do_yield_check();
7058 }
7059 }
7060
7061 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7062 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7063
7064 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7065 MemRegion span,
7066 ReferenceProcessor* rp,
7067 CMSBitMap* bit_map,
7068 CMSBitMap* mod_union_table,
7069 CMSMarkStack* mark_stack,
7070 bool concurrent_precleaning):
7071 MetadataAwareOopClosure(rp),
7072 _collector(collector),
7073 _span(span),
7074 _bit_map(bit_map),
7075 _mod_union_table(mod_union_table),
7076 _mark_stack(mark_stack),
7077 _concurrent_precleaning(concurrent_precleaning)
7078 {
7079 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL");
7080 }
7081
7082 // Grey object rescan during pre-cleaning and second checkpoint phases --
7083 // the non-parallel version (the parallel version appears further below.)
7084 void PushAndMarkClosure::do_oop(oop obj) {
7085 // Ignore mark word verification. If during concurrent precleaning,
7086 // the object monitor may be locked. If during the checkpoint
7087 // phases, the object may already have been reached by a different
7088 // path and may be at the end of the global overflow list (so
7089 // the mark word may be NULL).
7090 assert(obj->is_oop_or_null(true /* ignore mark word */),
7091 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
7092 HeapWord* addr = (HeapWord*)obj;
7093 // Check if oop points into the CMS generation
7094 // and is not marked
7095 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7096 // a white object ...
7097 _bit_map->mark(addr); // ... now grey
7098 // push on the marking stack (grey set)
7099 bool simulate_overflow = false;
7100 NOT_PRODUCT(
7101 if (CMSMarkStackOverflowALot &&
7102 _collector->simulate_overflow()) {
7103 // simulate a stack overflow
7104 simulate_overflow = true;
7105 }
7106 )
7107 if (simulate_overflow || !_mark_stack->push(obj)) {
7108 if (_concurrent_precleaning) {
7109 // During precleaning we can just dirty the appropriate card(s)
7110 // in the mod union table, thus ensuring that the object remains
7111 // in the grey set and continue. In the case of object arrays
7112 // we need to dirty all of the cards that the object spans,
7113 // since the rescan of object arrays will be limited to the
7114 // dirty cards.
7115 // Note that no one can be interfering with us in this action
7116 // of dirtying the mod union table, so no locking or atomics
7117 // are required.
7118 if (obj->is_objArray()) {
7119 size_t sz = obj->size();
7120 HeapWord* end_card_addr = (HeapWord*)round_to(
7121 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7122 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7123 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7124 _mod_union_table->mark_range(redirty_range);
7125 } else {
7126 _mod_union_table->mark(addr);
7127 }
7128 _collector->_ser_pmc_preclean_ovflw++;
7129 } else {
7130 // During the remark phase, we need to remember this oop
7131 // in the overflow list.
7132 _collector->push_on_overflow_list(obj);
7133 _collector->_ser_pmc_remark_ovflw++;
7134 }
7135 }
7136 }
7137 }
7138
7139 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7140 MemRegion span,
7141 ReferenceProcessor* rp,
7142 CMSBitMap* bit_map,
7143 OopTaskQueue* work_queue):
7144 MetadataAwareOopClosure(rp),
7145 _collector(collector),
7146 _span(span),
7147 _bit_map(bit_map),
7148 _work_queue(work_queue)
7149 {
7150 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL");
7151 }
7152
7153 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7154 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7155
7156 // Grey object rescan during second checkpoint phase --
7157 // the parallel version.
7158 void Par_PushAndMarkClosure::do_oop(oop obj) {
7159 // In the assert below, we ignore the mark word because
7160 // this oop may point to an already visited object that is
7161 // on the overflow stack (in which case the mark word has
7162 // been hijacked for chaining into the overflow stack --
7163 // if this is the last object in the overflow stack then
7164 // its mark word will be NULL). Because this object may
7165 // have been subsequently popped off the global overflow
7166 // stack, and the mark word possibly restored to the prototypical
7167 // value, by the time we get to examined this failing assert in
7168 // the debugger, is_oop_or_null(false) may subsequently start
7169 // to hold.
7170 assert(obj->is_oop_or_null(true),
7171 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
7172 HeapWord* addr = (HeapWord*)obj;
7173 // Check if oop points into the CMS generation
7174 // and is not marked
7175 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7176 // a white object ...
7177 // If we manage to "claim" the object, by being the
7178 // first thread to mark it, then we push it on our
7179 // marking stack
7180 if (_bit_map->par_mark(addr)) { // ... now grey
7181 // push on work queue (grey set)
7182 bool simulate_overflow = false;
7183 NOT_PRODUCT(
7184 if (CMSMarkStackOverflowALot &&
7185 _collector->par_simulate_overflow()) {
7186 // simulate a stack overflow
7187 simulate_overflow = true;
7188 }
7189 )
7190 if (simulate_overflow || !_work_queue->push(obj)) {
7191 _collector->par_push_on_overflow_list(obj);
7192 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7193 }
7194 } // Else, some other thread got there first
7195 }
7196 }
7197
7198 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7199 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7200
7201 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7202 Mutex* bml = _collector->bitMapLock();
7203 assert_lock_strong(bml);
7204 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7205 "CMS thread should hold CMS token");
7206
7207 bml->unlock();
7208 ConcurrentMarkSweepThread::desynchronize(true);
7209
7210 _collector->stopTimer();
7211 if (PrintCMSStatistics != 0) {
7212 _collector->incrementYields();
7213 }
7214
7215 // See the comment in coordinator_yield()
7216 for (unsigned i = 0; i < CMSYieldSleepCount &&
7217 ConcurrentMarkSweepThread::should_yield() &&
7218 !CMSCollector::foregroundGCIsActive(); ++i) {
7219 os::sleep(Thread::current(), 1, false);
7220 }
7221
7222 ConcurrentMarkSweepThread::synchronize(true);
7223 bml->lock();
7224
7225 _collector->startTimer();
7226 }
7227
7228 bool CMSPrecleanRefsYieldClosure::should_return() {
7229 if (ConcurrentMarkSweepThread::should_yield()) {
7230 do_yield_work();
7231 }
7232 return _collector->foregroundGCIsActive();
7233 }
7234
7235 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7236 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7237 "mr should be aligned to start at a card boundary");
7238 // We'd like to assert:
7239 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7240 // "mr should be a range of cards");
7241 // However, that would be too strong in one case -- the last
7242 // partition ends at _unallocated_block which, in general, can be
7243 // an arbitrary boundary, not necessarily card aligned.
7244 if (PrintCMSStatistics != 0) {
7245 _num_dirty_cards +=
7246 mr.word_size()/CardTableModRefBS::card_size_in_words;
7247 }
7248 _space->object_iterate_mem(mr, &_scan_cl);
7249 }
7250
7251 SweepClosure::SweepClosure(CMSCollector* collector,
7252 ConcurrentMarkSweepGeneration* g,
7253 CMSBitMap* bitMap, bool should_yield) :
7254 _collector(collector),
7255 _g(g),
7256 _sp(g->cmsSpace()),
7257 _limit(_sp->sweep_limit()),
7258 _freelistLock(_sp->freelistLock()),
7259 _bitMap(bitMap),
7260 _yield(should_yield),
7261 _inFreeRange(false), // No free range at beginning of sweep
7262 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7263 _lastFreeRangeCoalesced(false),
7264 _freeFinger(g->used_region().start())
7265 {
7266 NOT_PRODUCT(
7267 _numObjectsFreed = 0;
7268 _numWordsFreed = 0;
7269 _numObjectsLive = 0;
7270 _numWordsLive = 0;
7271 _numObjectsAlreadyFree = 0;
7272 _numWordsAlreadyFree = 0;
7273 _last_fc = NULL;
7274
7275 _sp->initializeIndexedFreeListArrayReturnedBytes();
7276 _sp->dictionary()->initialize_dict_returned_bytes();
7277 )
7278 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7279 "sweep _limit out of bounds");
7280 if (CMSTraceSweeper) {
7281 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
7282 p2i(_limit));
7283 }
7284 }
7285
7286 void SweepClosure::print_on(outputStream* st) const {
7287 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7288 p2i(_sp->bottom()), p2i(_sp->end()));
7289 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7290 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7291 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7292 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7293 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7294 }
7295
7296 #ifndef PRODUCT
7297 // Assertion checking only: no useful work in product mode --
7298 // however, if any of the flags below become product flags,
7299 // you may need to review this code to see if it needs to be
7300 // enabled in product mode.
7301 SweepClosure::~SweepClosure() {
7302 assert_lock_strong(_freelistLock);
7303 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7304 "sweep _limit out of bounds");
7305 if (inFreeRange()) {
7306 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
7307 print();
7308 ShouldNotReachHere();
7309 }
7310 if (Verbose && PrintGC) {
7311 gclog_or_tty->print("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7312 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7313 gclog_or_tty->print_cr("\nLive " SIZE_FORMAT " objects, "
7314 SIZE_FORMAT " bytes "
7315 "Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7316 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7317 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7318 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
7319 * sizeof(HeapWord);
7320 gclog_or_tty->print_cr("Total sweep: " SIZE_FORMAT " bytes", totalBytes);
7321
7322 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7323 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7324 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7325 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7326 gclog_or_tty->print("Returned " SIZE_FORMAT " bytes", returned_bytes);
7327 gclog_or_tty->print(" Indexed List Returned " SIZE_FORMAT " bytes",
7328 indexListReturnedBytes);
7329 gclog_or_tty->print_cr(" Dictionary Returned " SIZE_FORMAT " bytes",
7330 dict_returned_bytes);
7331 }
7332 }
7333 if (CMSTraceSweeper) {
7334 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
7335 p2i(_limit));
7336 }
7337 }
7338 #endif // PRODUCT
7339
7340 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7341 bool freeRangeInFreeLists) {
7342 if (CMSTraceSweeper) {
7343 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n",
7344 p2i(freeFinger), freeRangeInFreeLists);
7345 }
7346 assert(!inFreeRange(), "Trampling existing free range");
7347 set_inFreeRange(true);
7348 set_lastFreeRangeCoalesced(false);
7349
7350 set_freeFinger(freeFinger);
7351 set_freeRangeInFreeLists(freeRangeInFreeLists);
7352 if (CMSTestInFreeList) {
7353 if (freeRangeInFreeLists) {
7354 FreeChunk* fc = (FreeChunk*) freeFinger;
7355 assert(fc->is_free(), "A chunk on the free list should be free.");
7356 assert(fc->size() > 0, "Free range should have a size");
7357 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7358 }
7359 }
7360 }
7361
7362 // Note that the sweeper runs concurrently with mutators. Thus,
7363 // it is possible for direct allocation in this generation to happen
7364 // in the middle of the sweep. Note that the sweeper also coalesces
7365 // contiguous free blocks. Thus, unless the sweeper and the allocator
7366 // synchronize appropriately freshly allocated blocks may get swept up.
7367 // This is accomplished by the sweeper locking the free lists while
7368 // it is sweeping. Thus blocks that are determined to be free are
7369 // indeed free. There is however one additional complication:
7370 // blocks that have been allocated since the final checkpoint and
7371 // mark, will not have been marked and so would be treated as
7372 // unreachable and swept up. To prevent this, the allocator marks
7373 // the bit map when allocating during the sweep phase. This leads,
7374 // however, to a further complication -- objects may have been allocated
7375 // but not yet initialized -- in the sense that the header isn't yet
7376 // installed. The sweeper can not then determine the size of the block
7377 // in order to skip over it. To deal with this case, we use a technique
7378 // (due to Printezis) to encode such uninitialized block sizes in the
7379 // bit map. Since the bit map uses a bit per every HeapWord, but the
7380 // CMS generation has a minimum object size of 3 HeapWords, it follows
7381 // that "normal marks" won't be adjacent in the bit map (there will
7382 // always be at least two 0 bits between successive 1 bits). We make use
7383 // of these "unused" bits to represent uninitialized blocks -- the bit
7384 // corresponding to the start of the uninitialized object and the next
7385 // bit are both set. Finally, a 1 bit marks the end of the object that
7386 // started with the two consecutive 1 bits to indicate its potentially
7387 // uninitialized state.
7388
7389 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7390 FreeChunk* fc = (FreeChunk*)addr;
7391 size_t res;
7392
7393 // Check if we are done sweeping. Below we check "addr >= _limit" rather
7394 // than "addr == _limit" because although _limit was a block boundary when
7395 // we started the sweep, it may no longer be one because heap expansion
7396 // may have caused us to coalesce the block ending at the address _limit
7397 // with a newly expanded chunk (this happens when _limit was set to the
7398 // previous _end of the space), so we may have stepped past _limit:
7399 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7400 if (addr >= _limit) { // we have swept up to or past the limit: finish up
7401 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7402 "sweep _limit out of bounds");
7403 assert(addr < _sp->end(), "addr out of bounds");
7404 // Flush any free range we might be holding as a single
7405 // coalesced chunk to the appropriate free list.
7406 if (inFreeRange()) {
7407 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7408 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger()));
7409 flush_cur_free_chunk(freeFinger(),
7410 pointer_delta(addr, freeFinger()));
7411 if (CMSTraceSweeper) {
7412 gclog_or_tty->print("Sweep: last chunk: ");
7413 gclog_or_tty->print("put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") "
7414 "[coalesced:%d]\n",
7415 p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7416 lastFreeRangeCoalesced() ? 1 : 0);
7417 }
7418 }
7419
7420 // help the iterator loop finish
7421 return pointer_delta(_sp->end(), addr);
7422 }
7423
7424 assert(addr < _limit, "sweep invariant");
7425 // check if we should yield
7426 do_yield_check(addr);
7427 if (fc->is_free()) {
7428 // Chunk that is already free
7429 res = fc->size();
7430 do_already_free_chunk(fc);
7431 debug_only(_sp->verifyFreeLists());
7432 // If we flush the chunk at hand in lookahead_and_flush()
7433 // and it's coalesced with a preceding chunk, then the
7434 // process of "mangling" the payload of the coalesced block
7435 // will cause erasure of the size information from the
7436 // (erstwhile) header of all the coalesced blocks but the
7437 // first, so the first disjunct in the assert will not hold
7438 // in that specific case (in which case the second disjunct
7439 // will hold).
7440 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7441 "Otherwise the size info doesn't change at this step");
7442 NOT_PRODUCT(
7443 _numObjectsAlreadyFree++;
7444 _numWordsAlreadyFree += res;
7445 )
7446 NOT_PRODUCT(_last_fc = fc;)
7447 } else if (!_bitMap->isMarked(addr)) {
7448 // Chunk is fresh garbage
7449 res = do_garbage_chunk(fc);
7450 debug_only(_sp->verifyFreeLists());
7451 NOT_PRODUCT(
7452 _numObjectsFreed++;
7453 _numWordsFreed += res;
7454 )
7455 } else {
7456 // Chunk that is alive.
7457 res = do_live_chunk(fc);
7458 debug_only(_sp->verifyFreeLists());
7459 NOT_PRODUCT(
7460 _numObjectsLive++;
7461 _numWordsLive += res;
7462 )
7463 }
7464 return res;
7465 }
7466
7467 // For the smart allocation, record following
7468 // split deaths - a free chunk is removed from its free list because
7469 // it is being split into two or more chunks.
7470 // split birth - a free chunk is being added to its free list because
7471 // a larger free chunk has been split and resulted in this free chunk.
7472 // coal death - a free chunk is being removed from its free list because
7473 // it is being coalesced into a large free chunk.
7474 // coal birth - a free chunk is being added to its free list because
7475 // it was created when two or more free chunks where coalesced into
7476 // this free chunk.
7477 //
7478 // These statistics are used to determine the desired number of free
7479 // chunks of a given size. The desired number is chosen to be relative
7480 // to the end of a CMS sweep. The desired number at the end of a sweep
7481 // is the
7482 // count-at-end-of-previous-sweep (an amount that was enough)
7483 // - count-at-beginning-of-current-sweep (the excess)
7484 // + split-births (gains in this size during interval)
7485 // - split-deaths (demands on this size during interval)
7486 // where the interval is from the end of one sweep to the end of the
7487 // next.
7488 //
7489 // When sweeping the sweeper maintains an accumulated chunk which is
7490 // the chunk that is made up of chunks that have been coalesced. That
7491 // will be termed the left-hand chunk. A new chunk of garbage that
7492 // is being considered for coalescing will be referred to as the
7493 // right-hand chunk.
7494 //
7495 // When making a decision on whether to coalesce a right-hand chunk with
7496 // the current left-hand chunk, the current count vs. the desired count
7497 // of the left-hand chunk is considered. Also if the right-hand chunk
7498 // is near the large chunk at the end of the heap (see
7499 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7500 // left-hand chunk is coalesced.
7501 //
7502 // When making a decision about whether to split a chunk, the desired count
7503 // vs. the current count of the candidate to be split is also considered.
7504 // If the candidate is underpopulated (currently fewer chunks than desired)
7505 // a chunk of an overpopulated (currently more chunks than desired) size may
7506 // be chosen. The "hint" associated with a free list, if non-null, points
7507 // to a free list which may be overpopulated.
7508 //
7509
7510 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7511 const size_t size = fc->size();
7512 // Chunks that cannot be coalesced are not in the
7513 // free lists.
7514 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7515 assert(_sp->verify_chunk_in_free_list(fc),
7516 "free chunk should be in free lists");
7517 }
7518 // a chunk that is already free, should not have been
7519 // marked in the bit map
7520 HeapWord* const addr = (HeapWord*) fc;
7521 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7522 // Verify that the bit map has no bits marked between
7523 // addr and purported end of this block.
7524 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7525
7526 // Some chunks cannot be coalesced under any circumstances.
7527 // See the definition of cantCoalesce().
7528 if (!fc->cantCoalesce()) {
7529 // This chunk can potentially be coalesced.
7530 // All the work is done in
7531 do_post_free_or_garbage_chunk(fc, size);
7532 // Note that if the chunk is not coalescable (the else arm
7533 // below), we unconditionally flush, without needing to do
7534 // a "lookahead," as we do below.
7535 if (inFreeRange()) lookahead_and_flush(fc, size);
7536 } else {
7537 // Code path common to both original and adaptive free lists.
7538
7539 // cant coalesce with previous block; this should be treated
7540 // as the end of a free run if any
7541 if (inFreeRange()) {
7542 // we kicked some butt; time to pick up the garbage
7543 assert(freeFinger() < addr, "freeFinger points too high");
7544 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7545 }
7546 // else, nothing to do, just continue
7547 }
7548 }
7549
7550 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7551 // This is a chunk of garbage. It is not in any free list.
7552 // Add it to a free list or let it possibly be coalesced into
7553 // a larger chunk.
7554 HeapWord* const addr = (HeapWord*) fc;
7555 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7556
7557 // Verify that the bit map has no bits marked between
7558 // addr and purported end of just dead object.
7559 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7560 do_post_free_or_garbage_chunk(fc, size);
7561
7562 assert(_limit >= addr + size,
7563 "A freshly garbage chunk can't possibly straddle over _limit");
7564 if (inFreeRange()) lookahead_and_flush(fc, size);
7565 return size;
7566 }
7567
7568 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7569 HeapWord* addr = (HeapWord*) fc;
7570 // The sweeper has just found a live object. Return any accumulated
7571 // left hand chunk to the free lists.
7572 if (inFreeRange()) {
7573 assert(freeFinger() < addr, "freeFinger points too high");
7574 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7575 }
7576
7577 // This object is live: we'd normally expect this to be
7578 // an oop, and like to assert the following:
7579 // assert(oop(addr)->is_oop(), "live block should be an oop");
7580 // However, as we commented above, this may be an object whose
7581 // header hasn't yet been initialized.
7582 size_t size;
7583 assert(_bitMap->isMarked(addr), "Tautology for this control point");
7584 if (_bitMap->isMarked(addr + 1)) {
7585 // Determine the size from the bit map, rather than trying to
7586 // compute it from the object header.
7587 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7588 size = pointer_delta(nextOneAddr + 1, addr);
7589 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7590 "alignment problem");
7591
7592 #ifdef ASSERT
7593 if (oop(addr)->klass_or_null() != NULL) {
7594 // Ignore mark word because we are running concurrent with mutators
7595 assert(oop(addr)->is_oop(true), "live block should be an oop");
7596 assert(size ==
7597 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7598 "P-mark and computed size do not agree");
7599 }
7600 #endif
7601
7602 } else {
7603 // This should be an initialized object that's alive.
7604 assert(oop(addr)->klass_or_null() != NULL,
7605 "Should be an initialized object");
7606 // Ignore mark word because we are running concurrent with mutators
7607 assert(oop(addr)->is_oop(true), "live block should be an oop");
7608 // Verify that the bit map has no bits marked between
7609 // addr and purported end of this block.
7610 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7611 assert(size >= 3, "Necessary for Printezis marks to work");
7612 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7613 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7614 }
7615 return size;
7616 }
7617
7618 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7619 size_t chunkSize) {
7620 // do_post_free_or_garbage_chunk() should only be called in the case
7621 // of the adaptive free list allocator.
7622 const bool fcInFreeLists = fc->is_free();
7623 assert((HeapWord*)fc <= _limit, "sweep invariant");
7624 if (CMSTestInFreeList && fcInFreeLists) {
7625 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7626 }
7627
7628 if (CMSTraceSweeper) {
7629 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7630 }
7631
7632 HeapWord* const fc_addr = (HeapWord*) fc;
7633
7634 bool coalesce = false;
7635 const size_t left = pointer_delta(fc_addr, freeFinger());
7636 const size_t right = chunkSize;
7637 switch (FLSCoalescePolicy) {
7638 // numeric value forms a coalition aggressiveness metric
7639 case 0: { // never coalesce
7640 coalesce = false;
7641 break;
7642 }
7643 case 1: { // coalesce if left & right chunks on overpopulated lists
7644 coalesce = _sp->coalOverPopulated(left) &&
7645 _sp->coalOverPopulated(right);
7646 break;
7647 }
7648 case 2: { // coalesce if left chunk on overpopulated list (default)
7649 coalesce = _sp->coalOverPopulated(left);
7650 break;
7651 }
7652 case 3: { // coalesce if left OR right chunk on overpopulated list
7653 coalesce = _sp->coalOverPopulated(left) ||
7654 _sp->coalOverPopulated(right);
7655 break;
7656 }
7657 case 4: { // always coalesce
7658 coalesce = true;
7659 break;
7660 }
7661 default:
7662 ShouldNotReachHere();
7663 }
7664
7665 // Should the current free range be coalesced?
7666 // If the chunk is in a free range and either we decided to coalesce above
7667 // or the chunk is near the large block at the end of the heap
7668 // (isNearLargestChunk() returns true), then coalesce this chunk.
7669 const bool doCoalesce = inFreeRange()
7670 && (coalesce || _g->isNearLargestChunk(fc_addr));
7671 if (doCoalesce) {
7672 // Coalesce the current free range on the left with the new
7673 // chunk on the right. If either is on a free list,
7674 // it must be removed from the list and stashed in the closure.
7675 if (freeRangeInFreeLists()) {
7676 FreeChunk* const ffc = (FreeChunk*)freeFinger();
7677 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7678 "Size of free range is inconsistent with chunk size.");
7679 if (CMSTestInFreeList) {
7680 assert(_sp->verify_chunk_in_free_list(ffc),
7681 "Chunk is not in free lists");
7682 }
7683 _sp->coalDeath(ffc->size());
7684 _sp->removeFreeChunkFromFreeLists(ffc);
7685 set_freeRangeInFreeLists(false);
7686 }
7687 if (fcInFreeLists) {
7688 _sp->coalDeath(chunkSize);
7689 assert(fc->size() == chunkSize,
7690 "The chunk has the wrong size or is not in the free lists");
7691 _sp->removeFreeChunkFromFreeLists(fc);
7692 }
7693 set_lastFreeRangeCoalesced(true);
7694 print_free_block_coalesced(fc);
7695 } else { // not in a free range and/or should not coalesce
7696 // Return the current free range and start a new one.
7697 if (inFreeRange()) {
7698 // In a free range but cannot coalesce with the right hand chunk.
7699 // Put the current free range into the free lists.
7700 flush_cur_free_chunk(freeFinger(),
7701 pointer_delta(fc_addr, freeFinger()));
7702 }
7703 // Set up for new free range. Pass along whether the right hand
7704 // chunk is in the free lists.
7705 initialize_free_range((HeapWord*)fc, fcInFreeLists);
7706 }
7707 }
7708
7709 // Lookahead flush:
7710 // If we are tracking a free range, and this is the last chunk that
7711 // we'll look at because its end crosses past _limit, we'll preemptively
7712 // flush it along with any free range we may be holding on to. Note that
7713 // this can be the case only for an already free or freshly garbage
7714 // chunk. If this block is an object, it can never straddle
7715 // over _limit. The "straddling" occurs when _limit is set at
7716 // the previous end of the space when this cycle started, and
7717 // a subsequent heap expansion caused the previously co-terminal
7718 // free block to be coalesced with the newly expanded portion,
7719 // thus rendering _limit a non-block-boundary making it dangerous
7720 // for the sweeper to step over and examine.
7721 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7722 assert(inFreeRange(), "Should only be called if currently in a free range.");
7723 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7724 assert(_sp->used_region().contains(eob - 1),
7725 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7726 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7727 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7728 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size);
7729 if (eob >= _limit) {
7730 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7731 if (CMSTraceSweeper) {
7732 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
7733 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7734 "[" PTR_FORMAT "," PTR_FORMAT ")",
7735 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7736 }
7737 // Return the storage we are tracking back into the free lists.
7738 if (CMSTraceSweeper) {
7739 gclog_or_tty->print_cr("Flushing ... ");
7740 }
7741 assert(freeFinger() < eob, "Error");
7742 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7743 }
7744 }
7745
7746 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7747 assert(inFreeRange(), "Should only be called if currently in a free range.");
7748 assert(size > 0,
7749 "A zero sized chunk cannot be added to the free lists.");
7750 if (!freeRangeInFreeLists()) {
7751 if (CMSTestInFreeList) {
7752 FreeChunk* fc = (FreeChunk*) chunk;
7753 fc->set_size(size);
7754 assert(!_sp->verify_chunk_in_free_list(fc),
7755 "chunk should not be in free lists yet");
7756 }
7757 if (CMSTraceSweeper) {
7758 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists",
7759 p2i(chunk), size);
7760 }
7761 // A new free range is going to be starting. The current
7762 // free range has not been added to the free lists yet or
7763 // was removed so add it back.
7764 // If the current free range was coalesced, then the death
7765 // of the free range was recorded. Record a birth now.
7766 if (lastFreeRangeCoalesced()) {
7767 _sp->coalBirth(size);
7768 }
7769 _sp->addChunkAndRepairOffsetTable(chunk, size,
7770 lastFreeRangeCoalesced());
7771 } else if (CMSTraceSweeper) {
7772 gclog_or_tty->print_cr("Already in free list: nothing to flush");
7773 }
7774 set_inFreeRange(false);
7775 set_freeRangeInFreeLists(false);
7776 }
7777
7778 // We take a break if we've been at this for a while,
7779 // so as to avoid monopolizing the locks involved.
7780 void SweepClosure::do_yield_work(HeapWord* addr) {
7781 // Return current free chunk being used for coalescing (if any)
7782 // to the appropriate freelist. After yielding, the next
7783 // free block encountered will start a coalescing range of
7784 // free blocks. If the next free block is adjacent to the
7785 // chunk just flushed, they will need to wait for the next
7786 // sweep to be coalesced.
7787 if (inFreeRange()) {
7788 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7789 }
7790
7791 // First give up the locks, then yield, then re-lock.
7792 // We should probably use a constructor/destructor idiom to
7793 // do this unlock/lock or modify the MutexUnlocker class to
7794 // serve our purpose. XXX
7795 assert_lock_strong(_bitMap->lock());
7796 assert_lock_strong(_freelistLock);
7797 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7798 "CMS thread should hold CMS token");
7799 _bitMap->lock()->unlock();
7800 _freelistLock->unlock();
7801 ConcurrentMarkSweepThread::desynchronize(true);
7802 _collector->stopTimer();
7803 if (PrintCMSStatistics != 0) {
7804 _collector->incrementYields();
7805 }
7806
7807 // See the comment in coordinator_yield()
7808 for (unsigned i = 0; i < CMSYieldSleepCount &&
7809 ConcurrentMarkSweepThread::should_yield() &&
7810 !CMSCollector::foregroundGCIsActive(); ++i) {
7811 os::sleep(Thread::current(), 1, false);
7812 }
7813
7814 ConcurrentMarkSweepThread::synchronize(true);
7815 _freelistLock->lock();
7816 _bitMap->lock()->lock_without_safepoint_check();
7817 _collector->startTimer();
7818 }
7819
7820 #ifndef PRODUCT
7821 // This is actually very useful in a product build if it can
7822 // be called from the debugger. Compile it into the product
7823 // as needed.
7824 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7825 return debug_cms_space->verify_chunk_in_free_list(fc);
7826 }
7827 #endif
7828
7829 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7830 if (CMSTraceSweeper) {
7831 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7832 p2i(fc), fc->size());
7833 }
7834 }
7835
7836 // CMSIsAliveClosure
7837 bool CMSIsAliveClosure::do_object_b(oop obj) {
7838 HeapWord* addr = (HeapWord*)obj;
7839 return addr != NULL &&
7840 (!_span.contains(addr) || _bit_map->isMarked(addr));
7841 }
7842
7843
7844 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7845 MemRegion span,
7846 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7847 bool cpc):
7848 _collector(collector),
7849 _span(span),
7850 _bit_map(bit_map),
7851 _mark_stack(mark_stack),
7852 _concurrent_precleaning(cpc) {
7853 assert(!_span.is_empty(), "Empty span could spell trouble");
7854 }
7855
7856
7857 // CMSKeepAliveClosure: the serial version
7858 void CMSKeepAliveClosure::do_oop(oop obj) {
7859 HeapWord* addr = (HeapWord*)obj;
7860 if (_span.contains(addr) &&
7861 !_bit_map->isMarked(addr)) {
7862 _bit_map->mark(addr);
7863 bool simulate_overflow = false;
7864 NOT_PRODUCT(
7865 if (CMSMarkStackOverflowALot &&
7866 _collector->simulate_overflow()) {
7867 // simulate a stack overflow
7868 simulate_overflow = true;
7869 }
7870 )
7871 if (simulate_overflow || !_mark_stack->push(obj)) {
7872 if (_concurrent_precleaning) {
7873 // We dirty the overflown object and let the remark
7874 // phase deal with it.
7875 assert(_collector->overflow_list_is_empty(), "Error");
7876 // In the case of object arrays, we need to dirty all of
7877 // the cards that the object spans. No locking or atomics
7878 // are needed since no one else can be mutating the mod union
7879 // table.
7880 if (obj->is_objArray()) {
7881 size_t sz = obj->size();
7882 HeapWord* end_card_addr =
7883 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
7884 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7885 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7886 _collector->_modUnionTable.mark_range(redirty_range);
7887 } else {
7888 _collector->_modUnionTable.mark(addr);
7889 }
7890 _collector->_ser_kac_preclean_ovflw++;
7891 } else {
7892 _collector->push_on_overflow_list(obj);
7893 _collector->_ser_kac_ovflw++;
7894 }
7895 }
7896 }
7897 }
7898
7899 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7900 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7901
7902 // CMSParKeepAliveClosure: a parallel version of the above.
7903 // The work queues are private to each closure (thread),
7904 // but (may be) available for stealing by other threads.
7905 void CMSParKeepAliveClosure::do_oop(oop obj) {
7906 HeapWord* addr = (HeapWord*)obj;
7907 if (_span.contains(addr) &&
7908 !_bit_map->isMarked(addr)) {
7909 // In general, during recursive tracing, several threads
7910 // may be concurrently getting here; the first one to
7911 // "tag" it, claims it.
7912 if (_bit_map->par_mark(addr)) {
7913 bool res = _work_queue->push(obj);
7914 assert(res, "Low water mark should be much less than capacity");
7915 // Do a recursive trim in the hope that this will keep
7916 // stack usage lower, but leave some oops for potential stealers
7917 trim_queue(_low_water_mark);
7918 } // Else, another thread got there first
7919 }
7920 }
7921
7922 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7923 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7924
7925 void CMSParKeepAliveClosure::trim_queue(uint max) {
7926 while (_work_queue->size() > max) {
7927 oop new_oop;
7928 if (_work_queue->pop_local(new_oop)) {
7929 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
7930 assert(_bit_map->isMarked((HeapWord*)new_oop),
7931 "no white objects on this stack!");
7932 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7933 // iterate over the oops in this oop, marking and pushing
7934 // the ones in CMS heap (i.e. in _span).
7935 new_oop->oop_iterate(&_mark_and_push);
7936 }
7937 }
7938 }
7939
7940 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
7941 CMSCollector* collector,
7942 MemRegion span, CMSBitMap* bit_map,
7943 OopTaskQueue* work_queue):
7944 _collector(collector),
7945 _span(span),
7946 _bit_map(bit_map),
7947 _work_queue(work_queue) { }
7948
7949 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
7950 HeapWord* addr = (HeapWord*)obj;
7951 if (_span.contains(addr) &&
7952 !_bit_map->isMarked(addr)) {
7953 if (_bit_map->par_mark(addr)) {
7954 bool simulate_overflow = false;
7955 NOT_PRODUCT(
7956 if (CMSMarkStackOverflowALot &&
7957 _collector->par_simulate_overflow()) {
7958 // simulate a stack overflow
7959 simulate_overflow = true;
7960 }
7961 )
7962 if (simulate_overflow || !_work_queue->push(obj)) {
7963 _collector->par_push_on_overflow_list(obj);
7964 _collector->_par_kac_ovflw++;
7965 }
7966 } // Else another thread got there already
7967 }
7968 }
7969
7970 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7971 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7972
7973 //////////////////////////////////////////////////////////////////
7974 // CMSExpansionCause /////////////////////////////
7975 //////////////////////////////////////////////////////////////////
7976 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
7977 switch (cause) {
7978 case _no_expansion:
7979 return "No expansion";
7980 case _satisfy_free_ratio:
7981 return "Free ratio";
7982 case _satisfy_promotion:
7983 return "Satisfy promotion";
7984 case _satisfy_allocation:
7985 return "allocation";
7986 case _allocate_par_lab:
7987 return "Par LAB";
7988 case _allocate_par_spooling_space:
7989 return "Par Spooling Space";
7990 case _adaptive_size_policy:
7991 return "Ergonomics";
7992 default:
7993 return "unknown";
7994 }
7995 }
7996
7997 void CMSDrainMarkingStackClosure::do_void() {
7998 // the max number to take from overflow list at a time
7999 const size_t num = _mark_stack->capacity()/4;
8000 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8001 "Overflow list should be NULL during concurrent phases");
8002 while (!_mark_stack->isEmpty() ||
8003 // if stack is empty, check the overflow list
8004 _collector->take_from_overflow_list(num, _mark_stack)) {
8005 oop obj = _mark_stack->pop();
8006 HeapWord* addr = (HeapWord*)obj;
8007 assert(_span.contains(addr), "Should be within span");
8008 assert(_bit_map->isMarked(addr), "Should be marked");
8009 assert(obj->is_oop(), "Should be an oop");
8010 obj->oop_iterate(_keep_alive);
8011 }
8012 }
8013
8014 void CMSParDrainMarkingStackClosure::do_void() {
8015 // drain queue
8016 trim_queue(0);
8017 }
8018
8019 // Trim our work_queue so its length is below max at return
8020 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8021 while (_work_queue->size() > max) {
8022 oop new_oop;
8023 if (_work_queue->pop_local(new_oop)) {
8024 assert(new_oop->is_oop(), "Expected an oop");
8025 assert(_bit_map->isMarked((HeapWord*)new_oop),
8026 "no white objects on this stack!");
8027 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8028 // iterate over the oops in this oop, marking and pushing
8029 // the ones in CMS heap (i.e. in _span).
8030 new_oop->oop_iterate(&_mark_and_push);
8031 }
8032 }
8033 }
8034
8035 ////////////////////////////////////////////////////////////////////
8036 // Support for Marking Stack Overflow list handling and related code
8037 ////////////////////////////////////////////////////////////////////
8038 // Much of the following code is similar in shape and spirit to the
8039 // code used in ParNewGC. We should try and share that code
8040 // as much as possible in the future.
8041
8042 #ifndef PRODUCT
8043 // Debugging support for CMSStackOverflowALot
8044
8045 // It's OK to call this multi-threaded; the worst thing
8046 // that can happen is that we'll get a bunch of closely
8047 // spaced simulated overflows, but that's OK, in fact
8048 // probably good as it would exercise the overflow code
8049 // under contention.
8050 bool CMSCollector::simulate_overflow() {
8051 if (_overflow_counter-- <= 0) { // just being defensive
8052 _overflow_counter = CMSMarkStackOverflowInterval;
8053 return true;
8054 } else {
8055 return false;
8056 }
8057 }
8058
8059 bool CMSCollector::par_simulate_overflow() {
8060 return simulate_overflow();
8061 }
8062 #endif
8063
8064 // Single-threaded
8065 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8066 assert(stack->isEmpty(), "Expected precondition");
8067 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8068 size_t i = num;
8069 oop cur = _overflow_list;
8070 const markOop proto = markOopDesc::prototype();
8071 NOT_PRODUCT(ssize_t n = 0;)
8072 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8073 next = oop(cur->mark());
8074 cur->set_mark(proto); // until proven otherwise
8075 assert(cur->is_oop(), "Should be an oop");
8076 bool res = stack->push(cur);
8077 assert(res, "Bit off more than can chew?");
8078 NOT_PRODUCT(n++;)
8079 }
8080 _overflow_list = cur;
8081 #ifndef PRODUCT
8082 assert(_num_par_pushes >= n, "Too many pops?");
8083 _num_par_pushes -=n;
8084 #endif
8085 return !stack->isEmpty();
8086 }
8087
8088 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
8089 // (MT-safe) Get a prefix of at most "num" from the list.
8090 // The overflow list is chained through the mark word of
8091 // each object in the list. We fetch the entire list,
8092 // break off a prefix of the right size and return the
8093 // remainder. If other threads try to take objects from
8094 // the overflow list at that time, they will wait for
8095 // some time to see if data becomes available. If (and
8096 // only if) another thread places one or more object(s)
8097 // on the global list before we have returned the suffix
8098 // to the global list, we will walk down our local list
8099 // to find its end and append the global list to
8100 // our suffix before returning it. This suffix walk can
8101 // prove to be expensive (quadratic in the amount of traffic)
8102 // when there are many objects in the overflow list and
8103 // there is much producer-consumer contention on the list.
8104 // *NOTE*: The overflow list manipulation code here and
8105 // in ParNewGeneration:: are very similar in shape,
8106 // except that in the ParNew case we use the old (from/eden)
8107 // copy of the object to thread the list via its klass word.
8108 // Because of the common code, if you make any changes in
8109 // the code below, please check the ParNew version to see if
8110 // similar changes might be needed.
8111 // CR 6797058 has been filed to consolidate the common code.
8112 bool CMSCollector::par_take_from_overflow_list(size_t num,
8113 OopTaskQueue* work_q,
8114 int no_of_gc_threads) {
8115 assert(work_q->size() == 0, "First empty local work queue");
8116 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8117 if (_overflow_list == NULL) {
8118 return false;
8119 }
8120 // Grab the entire list; we'll put back a suffix
8121 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8122 Thread* tid = Thread::current();
8123 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
8124 // set to ParallelGCThreads.
8125 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
8126 size_t sleep_time_millis = MAX2((size_t)1, num/100);
8127 // If the list is busy, we spin for a short while,
8128 // sleeping between attempts to get the list.
8129 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
8130 os::sleep(tid, sleep_time_millis, false);
8131 if (_overflow_list == NULL) {
8132 // Nothing left to take
8133 return false;
8134 } else if (_overflow_list != BUSY) {
8135 // Try and grab the prefix
8136 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8137 }
8138 }
8139 // If the list was found to be empty, or we spun long
8140 // enough, we give up and return empty-handed. If we leave
8141 // the list in the BUSY state below, it must be the case that
8142 // some other thread holds the overflow list and will set it
8143 // to a non-BUSY state in the future.
8144 if (prefix == NULL || prefix == BUSY) {
8145 // Nothing to take or waited long enough
8146 if (prefix == NULL) {
8147 // Write back the NULL in case we overwrote it with BUSY above
8148 // and it is still the same value.
8149 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8150 }
8151 return false;
8152 }
8153 assert(prefix != NULL && prefix != BUSY, "Error");
8154 size_t i = num;
8155 oop cur = prefix;
8156 // Walk down the first "num" objects, unless we reach the end.
8157 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8158 if (cur->mark() == NULL) {
8159 // We have "num" or fewer elements in the list, so there
8160 // is nothing to return to the global list.
8161 // Write back the NULL in lieu of the BUSY we wrote
8162 // above, if it is still the same value.
8163 if (_overflow_list == BUSY) {
8164 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8165 }
8166 } else {
8167 // Chop off the suffix and return it to the global list.
8168 assert(cur->mark() != BUSY, "Error");
8169 oop suffix_head = cur->mark(); // suffix will be put back on global list
8170 cur->set_mark(NULL); // break off suffix
8171 // It's possible that the list is still in the empty(busy) state
8172 // we left it in a short while ago; in that case we may be
8173 // able to place back the suffix without incurring the cost
8174 // of a walk down the list.
8175 oop observed_overflow_list = _overflow_list;
8176 oop cur_overflow_list = observed_overflow_list;
8177 bool attached = false;
8178 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
8179 observed_overflow_list =
8180 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8181 if (cur_overflow_list == observed_overflow_list) {
8182 attached = true;
8183 break;
8184 } else cur_overflow_list = observed_overflow_list;
8185 }
8186 if (!attached) {
8187 // Too bad, someone else sneaked in (at least) an element; we'll need
8188 // to do a splice. Find tail of suffix so we can prepend suffix to global
8189 // list.
8190 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8191 oop suffix_tail = cur;
8192 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8193 "Tautology");
8194 observed_overflow_list = _overflow_list;
8195 do {
8196 cur_overflow_list = observed_overflow_list;
8197 if (cur_overflow_list != BUSY) {
8198 // Do the splice ...
8199 suffix_tail->set_mark(markOop(cur_overflow_list));
8200 } else { // cur_overflow_list == BUSY
8201 suffix_tail->set_mark(NULL);
8202 }
8203 // ... and try to place spliced list back on overflow_list ...
8204 observed_overflow_list =
8205 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8206 } while (cur_overflow_list != observed_overflow_list);
8207 // ... until we have succeeded in doing so.
8208 }
8209 }
8210
8211 // Push the prefix elements on work_q
8212 assert(prefix != NULL, "control point invariant");
8213 const markOop proto = markOopDesc::prototype();
8214 oop next;
8215 NOT_PRODUCT(ssize_t n = 0;)
8216 for (cur = prefix; cur != NULL; cur = next) {
8217 next = oop(cur->mark());
8218 cur->set_mark(proto); // until proven otherwise
8219 assert(cur->is_oop(), "Should be an oop");
8220 bool res = work_q->push(cur);
8221 assert(res, "Bit off more than we can chew?");
8222 NOT_PRODUCT(n++;)
8223 }
8224 #ifndef PRODUCT
8225 assert(_num_par_pushes >= n, "Too many pops?");
8226 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8227 #endif
8228 return true;
8229 }
8230
8231 // Single-threaded
8232 void CMSCollector::push_on_overflow_list(oop p) {
8233 NOT_PRODUCT(_num_par_pushes++;)
8234 assert(p->is_oop(), "Not an oop");
8235 preserve_mark_if_necessary(p);
8236 p->set_mark((markOop)_overflow_list);
8237 _overflow_list = p;
8238 }
8239
8240 // Multi-threaded; use CAS to prepend to overflow list
8241 void CMSCollector::par_push_on_overflow_list(oop p) {
8242 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8243 assert(p->is_oop(), "Not an oop");
8244 par_preserve_mark_if_necessary(p);
8245 oop observed_overflow_list = _overflow_list;
8246 oop cur_overflow_list;
8247 do {
8248 cur_overflow_list = observed_overflow_list;
8249 if (cur_overflow_list != BUSY) {
8250 p->set_mark(markOop(cur_overflow_list));
8251 } else {
8252 p->set_mark(NULL);
8253 }
8254 observed_overflow_list =
8255 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8256 } while (cur_overflow_list != observed_overflow_list);
8257 }
8258 #undef BUSY
8259
8260 // Single threaded
8261 // General Note on GrowableArray: pushes may silently fail
8262 // because we are (temporarily) out of C-heap for expanding
8263 // the stack. The problem is quite ubiquitous and affects
8264 // a lot of code in the JVM. The prudent thing for GrowableArray
8265 // to do (for now) is to exit with an error. However, that may
8266 // be too draconian in some cases because the caller may be
8267 // able to recover without much harm. For such cases, we
8268 // should probably introduce a "soft_push" method which returns
8269 // an indication of success or failure with the assumption that
8270 // the caller may be able to recover from a failure; code in
8271 // the VM can then be changed, incrementally, to deal with such
8272 // failures where possible, thus, incrementally hardening the VM
8273 // in such low resource situations.
8274 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8275 _preserved_oop_stack.push(p);
8276 _preserved_mark_stack.push(m);
8277 assert(m == p->mark(), "Mark word changed");
8278 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8279 "bijection");
8280 }
8281
8282 // Single threaded
8283 void CMSCollector::preserve_mark_if_necessary(oop p) {
8284 markOop m = p->mark();
8285 if (m->must_be_preserved(p)) {
8286 preserve_mark_work(p, m);
8287 }
8288 }
8289
8290 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8291 markOop m = p->mark();
8292 if (m->must_be_preserved(p)) {
8293 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8294 // Even though we read the mark word without holding
8295 // the lock, we are assured that it will not change
8296 // because we "own" this oop, so no other thread can
8297 // be trying to push it on the overflow list; see
8298 // the assertion in preserve_mark_work() that checks
8299 // that m == p->mark().
8300 preserve_mark_work(p, m);
8301 }
8302 }
8303
8304 // We should be able to do this multi-threaded,
8305 // a chunk of stack being a task (this is
8306 // correct because each oop only ever appears
8307 // once in the overflow list. However, it's
8308 // not very easy to completely overlap this with
8309 // other operations, so will generally not be done
8310 // until all work's been completed. Because we
8311 // expect the preserved oop stack (set) to be small,
8312 // it's probably fine to do this single-threaded.
8313 // We can explore cleverer concurrent/overlapped/parallel
8314 // processing of preserved marks if we feel the
8315 // need for this in the future. Stack overflow should
8316 // be so rare in practice and, when it happens, its
8317 // effect on performance so great that this will
8318 // likely just be in the noise anyway.
8319 void CMSCollector::restore_preserved_marks_if_any() {
8320 assert(SafepointSynchronize::is_at_safepoint(),
8321 "world should be stopped");
8322 assert(Thread::current()->is_ConcurrentGC_thread() ||
8323 Thread::current()->is_VM_thread(),
8324 "should be single-threaded");
8325 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8326 "bijection");
8327
8328 while (!_preserved_oop_stack.is_empty()) {
8329 oop p = _preserved_oop_stack.pop();
8330 assert(p->is_oop(), "Should be an oop");
8331 assert(_span.contains(p), "oop should be in _span");
8332 assert(p->mark() == markOopDesc::prototype(),
8333 "Set when taken from overflow list");
8334 markOop m = _preserved_mark_stack.pop();
8335 p->set_mark(m);
8336 }
8337 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8338 "stacks were cleared above");
8339 }
8340
8341 #ifndef PRODUCT
8342 bool CMSCollector::no_preserved_marks() const {
8343 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8344 }
8345 #endif
8346
8347 // Transfer some number of overflown objects to usual marking
8348 // stack. Return true if some objects were transferred.
8349 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8350 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8351 (size_t)ParGCDesiredObjsFromOverflowList);
8352
8353 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8354 assert(_collector->overflow_list_is_empty() || res,
8355 "If list is not empty, we should have taken something");
8356 assert(!res || !_mark_stack->isEmpty(),
8357 "If we took something, it should now be on our stack");
8358 return res;
8359 }
8360
8361 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8362 size_t res = _sp->block_size_no_stall(addr, _collector);
8363 if (_sp->block_is_obj(addr)) {
8364 if (_live_bit_map->isMarked(addr)) {
8365 // It can't have been dead in a previous cycle
8366 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8367 } else {
8368 _dead_bit_map->mark(addr); // mark the dead object
8369 }
8370 }
8371 // Could be 0, if the block size could not be computed without stalling.
8372 return res;
8373 }
8374
8375 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8376
8377 switch (phase) {
8378 case CMSCollector::InitialMarking:
8379 initialize(true /* fullGC */ ,
8380 cause /* cause of the GC */,
8381 true /* recordGCBeginTime */,
8382 true /* recordPreGCUsage */,
8383 false /* recordPeakUsage */,
8384 false /* recordPostGCusage */,
8385 true /* recordAccumulatedGCTime */,
8386 false /* recordGCEndTime */,
8387 false /* countCollection */ );
8388 break;
8389
8390 case CMSCollector::FinalMarking:
8391 initialize(true /* fullGC */ ,
8392 cause /* cause of the GC */,
8393 false /* recordGCBeginTime */,
8394 false /* recordPreGCUsage */,
8395 false /* recordPeakUsage */,
8396 false /* recordPostGCusage */,
8397 true /* recordAccumulatedGCTime */,
8398 false /* recordGCEndTime */,
8399 false /* countCollection */ );
8400 break;
8401
8402 case CMSCollector::Sweeping:
8403 initialize(true /* fullGC */ ,
8404 cause /* cause of the GC */,
8405 false /* recordGCBeginTime */,
8406 false /* recordPreGCUsage */,
8407 true /* recordPeakUsage */,
8408 true /* recordPostGCusage */,
8409 false /* recordAccumulatedGCTime */,
8410 true /* recordGCEndTime */,
8411 true /* countCollection */ );
8412 break;
8413
8414 default:
8415 ShouldNotReachHere();
8416 }
8417 }