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