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