1 /*
2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/classLoaderData.hpp"
27 #include "classfile/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/cmsGCStats.hpp"
33 #include "gc/cms/cmsHeap.hpp"
34 #include "gc/cms/cmsOopClosures.inline.hpp"
35 #include "gc/cms/compactibleFreeListSpace.hpp"
36 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp"
37 #include "gc/cms/concurrentMarkSweepThread.hpp"
38 #include "gc/cms/parNewGeneration.hpp"
39 #include "gc/cms/vmCMSOperations.hpp"
40 #include "gc/serial/genMarkSweep.hpp"
41 #include "gc/serial/tenuredGeneration.hpp"
42 #include "gc/shared/adaptiveSizePolicy.hpp"
43 #include "gc/shared/cardGeneration.inline.hpp"
44 #include "gc/shared/cardTableRS.hpp"
45 #include "gc/shared/collectedHeap.inline.hpp"
46 #include "gc/shared/collectorCounters.hpp"
47 #include "gc/shared/collectorPolicy.hpp"
48 #include "gc/shared/gcLocker.hpp"
49 #include "gc/shared/gcPolicyCounters.hpp"
50 #include "gc/shared/gcTimer.hpp"
51 #include "gc/shared/gcTrace.hpp"
52 #include "gc/shared/gcTraceTime.inline.hpp"
53 #include "gc/shared/genCollectedHeap.hpp"
54 #include "gc/shared/genOopClosures.inline.hpp"
55 #include "gc/shared/isGCActiveMark.hpp"
56 #include "gc/shared/referencePolicy.hpp"
57 #include "gc/shared/strongRootsScope.hpp"
58 #include "gc/shared/taskqueue.inline.hpp"
59 #include "gc/shared/weakProcessor.hpp"
60 #include "logging/log.hpp"
61 #include "logging/logStream.hpp"
62 #include "memory/allocation.hpp"
63 #include "memory/binaryTreeDictionary.inline.hpp"
64 #include "memory/iterator.inline.hpp"
65 #include "memory/padded.hpp"
66 #include "memory/resourceArea.hpp"
67 #include "oops/access.inline.hpp"
68 #include "oops/oop.inline.hpp"
69 #include "prims/jvmtiExport.hpp"
70 #include "runtime/atomic.hpp"
71 #include "runtime/globals_extension.hpp"
72 #include "runtime/handles.inline.hpp"
73 #include "runtime/java.hpp"
74 #include "runtime/orderAccess.inline.hpp"
75 #include "runtime/timer.hpp"
76 #include "runtime/vmThread.hpp"
77 #include "services/memoryService.hpp"
78 #include "services/runtimeService.hpp"
79 #include "utilities/align.hpp"
80 #include "utilities/stack.inline.hpp"
81
82 // statics
83 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
84 bool CMSCollector::_full_gc_requested = false;
85 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
86
87 //////////////////////////////////////////////////////////////////
88 // In support of CMS/VM thread synchronization
89 //////////////////////////////////////////////////////////////////
90 // We split use of the CGC_lock into 2 "levels".
91 // The low-level locking is of the usual CGC_lock monitor. We introduce
92 // a higher level "token" (hereafter "CMS token") built on top of the
93 // low level monitor (hereafter "CGC lock").
94 // The token-passing protocol gives priority to the VM thread. The
95 // CMS-lock doesn't provide any fairness guarantees, but clients
96 // should ensure that it is only held for very short, bounded
97 // durations.
98 //
99 // When either of the CMS thread or the VM thread is involved in
100 // collection operations during which it does not want the other
101 // thread to interfere, it obtains the CMS token.
102 //
103 // If either thread tries to get the token while the other has
104 // it, that thread waits. However, if the VM thread and CMS thread
105 // both want the token, then the VM thread gets priority while the
106 // CMS thread waits. This ensures, for instance, that the "concurrent"
107 // phases of the CMS thread's work do not block out the VM thread
108 // for long periods of time as the CMS thread continues to hog
109 // the token. (See bug 4616232).
110 //
111 // The baton-passing functions are, however, controlled by the
112 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
113 // and here the low-level CMS lock, not the high level token,
114 // ensures mutual exclusion.
115 //
116 // Two important conditions that we have to satisfy:
117 // 1. if a thread does a low-level wait on the CMS lock, then it
118 // relinquishes the CMS token if it were holding that token
119 // when it acquired the low-level CMS lock.
120 // 2. any low-level notifications on the low-level lock
121 // should only be sent when a thread has relinquished the token.
122 //
123 // In the absence of either property, we'd have potential deadlock.
124 //
125 // We protect each of the CMS (concurrent and sequential) phases
126 // with the CMS _token_, not the CMS _lock_.
127 //
128 // The only code protected by CMS lock is the token acquisition code
129 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
130 // baton-passing code.
131 //
132 // Unfortunately, i couldn't come up with a good abstraction to factor and
133 // hide the naked CGC_lock manipulation in the baton-passing code
134 // further below. That's something we should try to do. Also, the proof
135 // of correctness of this 2-level locking scheme is far from obvious,
136 // and potentially quite slippery. We have an uneasy suspicion, for instance,
137 // that there may be a theoretical possibility of delay/starvation in the
138 // low-level lock/wait/notify scheme used for the baton-passing because of
139 // potential interference with the priority scheme embodied in the
140 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
141 // invocation further below and marked with "XXX 20011219YSR".
142 // Indeed, as we note elsewhere, this may become yet more slippery
143 // in the presence of multiple CMS and/or multiple VM threads. XXX
144
145 class CMSTokenSync: public StackObj {
146 private:
147 bool _is_cms_thread;
148 public:
149 CMSTokenSync(bool is_cms_thread):
150 _is_cms_thread(is_cms_thread) {
151 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
152 "Incorrect argument to constructor");
153 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
154 }
155
156 ~CMSTokenSync() {
157 assert(_is_cms_thread ?
158 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
159 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
160 "Incorrect state");
161 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
162 }
163 };
164
165 // Convenience class that does a CMSTokenSync, and then acquires
166 // upto three locks.
167 class CMSTokenSyncWithLocks: public CMSTokenSync {
168 private:
169 // Note: locks are acquired in textual declaration order
170 // and released in the opposite order
171 MutexLockerEx _locker1, _locker2, _locker3;
172 public:
173 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
174 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
175 CMSTokenSync(is_cms_thread),
176 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
177 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
178 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
179 { }
180 };
181
182
183 //////////////////////////////////////////////////////////////////
184 // Concurrent Mark-Sweep Generation /////////////////////////////
185 //////////////////////////////////////////////////////////////////
186
187 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
188
189 // This struct contains per-thread things necessary to support parallel
190 // young-gen collection.
191 class CMSParGCThreadState: public CHeapObj<mtGC> {
192 public:
193 CompactibleFreeListSpaceLAB lab;
194 PromotionInfo promo;
195
196 // Constructor.
197 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
198 promo.setSpace(cfls);
199 }
200 };
201
202 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
203 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) :
204 CardGeneration(rs, initial_byte_size, ct),
205 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
206 _did_compact(false)
207 {
208 HeapWord* bottom = (HeapWord*) _virtual_space.low();
209 HeapWord* end = (HeapWord*) _virtual_space.high();
210
211 _direct_allocated_words = 0;
212 NOT_PRODUCT(
213 _numObjectsPromoted = 0;
214 _numWordsPromoted = 0;
215 _numObjectsAllocated = 0;
216 _numWordsAllocated = 0;
217 )
218
219 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end));
220 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
221 _cmsSpace->_old_gen = this;
222
223 _gc_stats = new CMSGCStats();
224
225 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
226 // offsets match. The ability to tell free chunks from objects
227 // depends on this property.
228 debug_only(
229 FreeChunk* junk = NULL;
230 assert(UseCompressedClassPointers ||
231 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
232 "Offset of FreeChunk::_prev within FreeChunk must match"
233 " that of OopDesc::_klass within OopDesc");
234 )
235
236 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC);
237 for (uint i = 0; i < ParallelGCThreads; i++) {
238 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
239 }
240
241 _incremental_collection_failed = false;
242 // The "dilatation_factor" is the expansion that can occur on
243 // account of the fact that the minimum object size in the CMS
244 // generation may be larger than that in, say, a contiguous young
245 // generation.
246 // Ideally, in the calculation below, we'd compute the dilatation
247 // factor as: MinChunkSize/(promoting_gen's min object size)
248 // Since we do not have such a general query interface for the
249 // promoting generation, we'll instead just use the minimum
250 // object size (which today is a header's worth of space);
251 // note that all arithmetic is in units of HeapWords.
252 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
253 assert(_dilatation_factor >= 1.0, "from previous assert");
254 }
255
256
257 // The field "_initiating_occupancy" represents the occupancy percentage
258 // at which we trigger a new collection cycle. Unless explicitly specified
259 // via CMSInitiatingOccupancyFraction (argument "io" below), it
260 // is calculated by:
261 //
262 // Let "f" be MinHeapFreeRatio in
263 //
264 // _initiating_occupancy = 100-f +
265 // f * (CMSTriggerRatio/100)
266 // where CMSTriggerRatio is the argument "tr" below.
267 //
268 // That is, if we assume the heap is at its desired maximum occupancy at the
269 // end of a collection, we let CMSTriggerRatio of the (purported) free
270 // space be allocated before initiating a new collection cycle.
271 //
272 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
273 assert(io <= 100 && tr <= 100, "Check the arguments");
274 if (io >= 0) {
275 _initiating_occupancy = (double)io / 100.0;
276 } else {
277 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
278 (double)(tr * MinHeapFreeRatio) / 100.0)
279 / 100.0;
280 }
281 }
282
283 void ConcurrentMarkSweepGeneration::ref_processor_init() {
284 assert(collector() != NULL, "no collector");
285 collector()->ref_processor_init();
286 }
287
288 void CMSCollector::ref_processor_init() {
289 if (_ref_processor == NULL) {
290 // Allocate and initialize a reference processor
291 _ref_processor =
292 new ReferenceProcessor(_span, // span
293 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
294 ParallelGCThreads, // mt processing degree
295 _cmsGen->refs_discovery_is_mt(), // mt discovery
296 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
297 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
298 &_is_alive_closure); // closure for liveness info
299 // Initialize the _ref_processor field of CMSGen
300 _cmsGen->set_ref_processor(_ref_processor);
301
302 }
303 }
304
305 AdaptiveSizePolicy* CMSCollector::size_policy() {
306 return CMSHeap::heap()->size_policy();
307 }
308
309 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
310
311 const char* gen_name = "old";
312 GenCollectorPolicy* gcp = CMSHeap::heap()->gen_policy();
313 // Generation Counters - generation 1, 1 subspace
314 _gen_counters = new GenerationCounters(gen_name, 1, 1,
315 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
316
317 _space_counters = new GSpaceCounters(gen_name, 0,
318 _virtual_space.reserved_size(),
319 this, _gen_counters);
320 }
321
322 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
323 _cms_gen(cms_gen)
324 {
325 assert(alpha <= 100, "bad value");
326 _saved_alpha = alpha;
327
328 // Initialize the alphas to the bootstrap value of 100.
329 _gc0_alpha = _cms_alpha = 100;
330
331 _cms_begin_time.update();
332 _cms_end_time.update();
333
334 _gc0_duration = 0.0;
335 _gc0_period = 0.0;
336 _gc0_promoted = 0;
337
338 _cms_duration = 0.0;
339 _cms_period = 0.0;
340 _cms_allocated = 0;
341
342 _cms_used_at_gc0_begin = 0;
343 _cms_used_at_gc0_end = 0;
344 _allow_duty_cycle_reduction = false;
345 _valid_bits = 0;
346 }
347
348 double CMSStats::cms_free_adjustment_factor(size_t free) const {
349 // TBD: CR 6909490
350 return 1.0;
351 }
352
353 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
354 }
355
356 // If promotion failure handling is on use
357 // the padded average size of the promotion for each
358 // young generation collection.
359 double CMSStats::time_until_cms_gen_full() const {
360 size_t cms_free = _cms_gen->cmsSpace()->free();
361 CMSHeap* heap = CMSHeap::heap();
362 size_t expected_promotion = MIN2(heap->young_gen()->capacity(),
363 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
364 if (cms_free > expected_promotion) {
365 // Start a cms collection if there isn't enough space to promote
366 // for the next young collection. Use the padded average as
367 // a safety factor.
368 cms_free -= expected_promotion;
369
370 // Adjust by the safety factor.
371 double cms_free_dbl = (double)cms_free;
372 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0;
373 // Apply a further correction factor which tries to adjust
374 // for recent occurance of concurrent mode failures.
375 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
376 cms_free_dbl = cms_free_dbl * cms_adjustment;
377
378 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
379 cms_free, expected_promotion);
380 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0);
381 // Add 1 in case the consumption rate goes to zero.
382 return cms_free_dbl / (cms_consumption_rate() + 1.0);
383 }
384 return 0.0;
385 }
386
387 // Compare the duration of the cms collection to the
388 // time remaining before the cms generation is empty.
389 // Note that the time from the start of the cms collection
390 // to the start of the cms sweep (less than the total
391 // duration of the cms collection) can be used. This
392 // has been tried and some applications experienced
393 // promotion failures early in execution. This was
394 // possibly because the averages were not accurate
395 // enough at the beginning.
396 double CMSStats::time_until_cms_start() const {
397 // We add "gc0_period" to the "work" calculation
398 // below because this query is done (mostly) at the
399 // end of a scavenge, so we need to conservatively
400 // account for that much possible delay
401 // in the query so as to avoid concurrent mode failures
402 // due to starting the collection just a wee bit too
403 // late.
404 double work = cms_duration() + gc0_period();
405 double deadline = time_until_cms_gen_full();
406 // If a concurrent mode failure occurred recently, we want to be
407 // more conservative and halve our expected time_until_cms_gen_full()
408 if (work > deadline) {
409 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ",
410 cms_duration(), gc0_period(), time_until_cms_gen_full());
411 return 0.0;
412 }
413 return work - deadline;
414 }
415
416 #ifndef PRODUCT
417 void CMSStats::print_on(outputStream *st) const {
418 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
419 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
420 gc0_duration(), gc0_period(), gc0_promoted());
421 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
422 cms_duration(), cms_period(), cms_allocated());
423 st->print(",cms_since_beg=%g,cms_since_end=%g",
424 cms_time_since_begin(), cms_time_since_end());
425 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
426 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
427
428 if (valid()) {
429 st->print(",promo_rate=%g,cms_alloc_rate=%g",
430 promotion_rate(), cms_allocation_rate());
431 st->print(",cms_consumption_rate=%g,time_until_full=%g",
432 cms_consumption_rate(), time_until_cms_gen_full());
433 }
434 st->cr();
435 }
436 #endif // #ifndef PRODUCT
437
438 CMSCollector::CollectorState CMSCollector::_collectorState =
439 CMSCollector::Idling;
440 bool CMSCollector::_foregroundGCIsActive = false;
441 bool CMSCollector::_foregroundGCShouldWait = false;
442
443 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
444 CardTableRS* ct,
445 ConcurrentMarkSweepPolicy* cp):
446 _cmsGen(cmsGen),
447 _ct(ct),
448 _ref_processor(NULL), // will be set later
449 _conc_workers(NULL), // may be set later
450 _abort_preclean(false),
451 _start_sampling(false),
452 _between_prologue_and_epilogue(false),
453 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
454 _modUnionTable((CardTable::card_shift - LogHeapWordSize),
455 -1 /* lock-free */, "No_lock" /* dummy */),
456 _modUnionClosurePar(&_modUnionTable),
457 // Adjust my span to cover old (cms) gen
458 _span(cmsGen->reserved()),
459 // Construct the is_alive_closure with _span & markBitMap
460 _is_alive_closure(_span, &_markBitMap),
461 _restart_addr(NULL),
462 _overflow_list(NULL),
463 _stats(cmsGen),
464 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
465 //verify that this lock should be acquired with safepoint check.
466 Monitor::_safepoint_check_sometimes)),
467 _eden_chunk_array(NULL), // may be set in ctor body
468 _eden_chunk_capacity(0), // -- ditto --
469 _eden_chunk_index(0), // -- ditto --
470 _survivor_plab_array(NULL), // -- ditto --
471 _survivor_chunk_array(NULL), // -- ditto --
472 _survivor_chunk_capacity(0), // -- ditto --
473 _survivor_chunk_index(0), // -- ditto --
474 _ser_pmc_preclean_ovflw(0),
475 _ser_kac_preclean_ovflw(0),
476 _ser_pmc_remark_ovflw(0),
477 _par_pmc_remark_ovflw(0),
478 _ser_kac_ovflw(0),
479 _par_kac_ovflw(0),
480 #ifndef PRODUCT
481 _num_par_pushes(0),
482 #endif
483 _collection_count_start(0),
484 _verifying(false),
485 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
486 _completed_initialization(false),
487 _collector_policy(cp),
488 _should_unload_classes(CMSClassUnloadingEnabled),
489 _concurrent_cycles_since_last_unload(0),
490 _roots_scanning_options(GenCollectedHeap::SO_None),
491 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
492 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
493 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
494 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
495 _cms_start_registered(false)
496 {
497 // Now expand the span and allocate the collection support structures
498 // (MUT, marking bit map etc.) to cover both generations subject to
499 // collection.
500
501 // For use by dirty card to oop closures.
502 _cmsGen->cmsSpace()->set_collector(this);
503
504 // Allocate MUT and marking bit map
505 {
506 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
507 if (!_markBitMap.allocate(_span)) {
508 log_warning(gc)("Failed to allocate CMS Bit Map");
509 return;
510 }
511 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
512 }
513 {
514 _modUnionTable.allocate(_span);
515 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
516 }
517
518 if (!_markStack.allocate(MarkStackSize)) {
519 log_warning(gc)("Failed to allocate CMS Marking Stack");
520 return;
521 }
522
523 // Support for multi-threaded concurrent phases
524 if (CMSConcurrentMTEnabled) {
525 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
526 // just for now
527 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4);
528 }
529 if (ConcGCThreads > 1) {
530 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
531 ConcGCThreads, true);
532 if (_conc_workers == NULL) {
533 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled");
534 CMSConcurrentMTEnabled = false;
535 } else {
536 _conc_workers->initialize_workers();
537 }
538 } else {
539 CMSConcurrentMTEnabled = false;
540 }
541 }
542 if (!CMSConcurrentMTEnabled) {
543 ConcGCThreads = 0;
544 } else {
545 // Turn off CMSCleanOnEnter optimization temporarily for
546 // the MT case where it's not fixed yet; see 6178663.
547 CMSCleanOnEnter = false;
548 }
549 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
550 "Inconsistency");
551 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads);
552 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
553
554 // Parallel task queues; these are shared for the
555 // concurrent and stop-world phases of CMS, but
556 // are not shared with parallel scavenge (ParNew).
557 {
558 uint i;
559 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads);
560
561 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
562 || ParallelRefProcEnabled)
563 && num_queues > 0) {
564 _task_queues = new OopTaskQueueSet(num_queues);
565 if (_task_queues == NULL) {
566 log_warning(gc)("task_queues allocation failure.");
567 return;
568 }
569 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
570 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
571 for (i = 0; i < num_queues; i++) {
572 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
573 if (q == NULL) {
574 log_warning(gc)("work_queue allocation failure.");
575 return;
576 }
577 _task_queues->register_queue(i, q);
578 }
579 for (i = 0; i < num_queues; i++) {
580 _task_queues->queue(i)->initialize();
581 _hash_seed[i] = 17; // copied from ParNew
582 }
583 }
584 }
585
586 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
587
588 // Clip CMSBootstrapOccupancy between 0 and 100.
589 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0;
590
591 // Now tell CMS generations the identity of their collector
592 ConcurrentMarkSweepGeneration::set_collector(this);
593
594 // Create & start a CMS thread for this CMS collector
595 _cmsThread = ConcurrentMarkSweepThread::start(this);
596 assert(cmsThread() != NULL, "CMS Thread should have been created");
597 assert(cmsThread()->collector() == this,
598 "CMS Thread should refer to this gen");
599 assert(CGC_lock != NULL, "Where's the CGC_lock?");
600
601 // Support for parallelizing young gen rescan
602 CMSHeap* heap = CMSHeap::heap();
603 assert(heap->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
604 _young_gen = (ParNewGeneration*)heap->young_gen();
605 if (heap->supports_inline_contig_alloc()) {
606 _top_addr = heap->top_addr();
607 _end_addr = heap->end_addr();
608 assert(_young_gen != NULL, "no _young_gen");
609 _eden_chunk_index = 0;
610 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain;
611 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
612 }
613
614 // Support for parallelizing survivor space rescan
615 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
616 const size_t max_plab_samples =
617 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize);
618
619 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
620 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
621 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
622 _survivor_chunk_capacity = max_plab_samples;
623 for (uint i = 0; i < ParallelGCThreads; i++) {
624 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
625 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
626 assert(cur->end() == 0, "Should be 0");
627 assert(cur->array() == vec, "Should be vec");
628 assert(cur->capacity() == max_plab_samples, "Error");
629 }
630 }
631
632 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
633 _gc_counters = new CollectorCounters("CMS", 1);
634 _cgc_counters = new CollectorCounters("CMS stop-the-world phases", 2);
635 _completed_initialization = true;
636 _inter_sweep_timer.start(); // start of time
637 }
638
639 const char* ConcurrentMarkSweepGeneration::name() const {
640 return "concurrent mark-sweep generation";
641 }
642 void ConcurrentMarkSweepGeneration::update_counters() {
643 if (UsePerfData) {
644 _space_counters->update_all();
645 _gen_counters->update_all();
646 }
647 }
648
649 // this is an optimized version of update_counters(). it takes the
650 // used value as a parameter rather than computing it.
651 //
652 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
653 if (UsePerfData) {
654 _space_counters->update_used(used);
655 _space_counters->update_capacity();
656 _gen_counters->update_all();
657 }
658 }
659
660 void ConcurrentMarkSweepGeneration::print() const {
661 Generation::print();
662 cmsSpace()->print();
663 }
664
665 #ifndef PRODUCT
666 void ConcurrentMarkSweepGeneration::print_statistics() {
667 cmsSpace()->printFLCensus(0);
668 }
669 #endif
670
671 size_t
672 ConcurrentMarkSweepGeneration::contiguous_available() const {
673 // dld proposes an improvement in precision here. If the committed
674 // part of the space ends in a free block we should add that to
675 // uncommitted size in the calculation below. Will make this
676 // change later, staying with the approximation below for the
677 // time being. -- ysr.
678 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
679 }
680
681 size_t
682 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
683 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
684 }
685
686 size_t ConcurrentMarkSweepGeneration::max_available() const {
687 return free() + _virtual_space.uncommitted_size();
688 }
689
690 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
691 size_t available = max_available();
692 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
693 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
694 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")",
695 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes);
696 return res;
697 }
698
699 // At a promotion failure dump information on block layout in heap
700 // (cms old generation).
701 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
702 Log(gc, promotion) log;
703 if (log.is_trace()) {
704 LogStream ls(log.trace());
705 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls);
706 }
707 }
708
709 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
710 // Clear the promotion information. These pointers can be adjusted
711 // along with all the other pointers into the heap but
712 // compaction is expected to be a rare event with
713 // a heap using cms so don't do it without seeing the need.
714 for (uint i = 0; i < ParallelGCThreads; i++) {
715 _par_gc_thread_states[i]->promo.reset();
716 }
717 }
718
719 void ConcurrentMarkSweepGeneration::compute_new_size() {
720 assert_locked_or_safepoint(Heap_lock);
721
722 // If incremental collection failed, we just want to expand
723 // to the limit.
724 if (incremental_collection_failed()) {
725 clear_incremental_collection_failed();
726 grow_to_reserved();
727 return;
728 }
729
730 // The heap has been compacted but not reset yet.
731 // Any metric such as free() or used() will be incorrect.
732
733 CardGeneration::compute_new_size();
734
735 // Reset again after a possible resizing
736 if (did_compact()) {
737 cmsSpace()->reset_after_compaction();
738 }
739 }
740
741 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
742 assert_locked_or_safepoint(Heap_lock);
743
744 // If incremental collection failed, we just want to expand
745 // to the limit.
746 if (incremental_collection_failed()) {
747 clear_incremental_collection_failed();
748 grow_to_reserved();
749 return;
750 }
751
752 double free_percentage = ((double) free()) / capacity();
753 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
754 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
755
756 // compute expansion delta needed for reaching desired free percentage
757 if (free_percentage < desired_free_percentage) {
758 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
759 assert(desired_capacity >= capacity(), "invalid expansion size");
760 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
761 Log(gc) log;
762 if (log.is_trace()) {
763 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
764 log.trace("From compute_new_size: ");
765 log.trace(" Free fraction %f", free_percentage);
766 log.trace(" Desired free fraction %f", desired_free_percentage);
767 log.trace(" Maximum free fraction %f", maximum_free_percentage);
768 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000);
769 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000);
770 CMSHeap* heap = CMSHeap::heap();
771 assert(heap->is_old_gen(this), "The CMS generation should always be the old generation");
772 size_t young_size = heap->young_gen()->capacity();
773 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000);
774 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000);
775 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000);
776 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes);
777 }
778 // safe if expansion fails
779 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
780 log.trace(" Expanded free fraction %f", ((double) free()) / capacity());
781 } else {
782 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
783 assert(desired_capacity <= capacity(), "invalid expansion size");
784 size_t shrink_bytes = capacity() - desired_capacity;
785 // Don't shrink unless the delta is greater than the minimum shrink we want
786 if (shrink_bytes >= MinHeapDeltaBytes) {
787 shrink_free_list_by(shrink_bytes);
788 }
789 }
790 }
791
792 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
793 return cmsSpace()->freelistLock();
794 }
795
796 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) {
797 CMSSynchronousYieldRequest yr;
798 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
799 return have_lock_and_allocate(size, tlab);
800 }
801
802 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
803 bool tlab /* ignored */) {
804 assert_lock_strong(freelistLock());
805 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
806 HeapWord* res = cmsSpace()->allocate(adjustedSize);
807 // Allocate the object live (grey) if the background collector has
808 // started marking. This is necessary because the marker may
809 // have passed this address and consequently this object will
810 // not otherwise be greyed and would be incorrectly swept up.
811 // Note that if this object contains references, the writing
812 // of those references will dirty the card containing this object
813 // allowing the object to be blackened (and its references scanned)
814 // either during a preclean phase or at the final checkpoint.
815 if (res != NULL) {
816 // We may block here with an uninitialized object with
817 // its mark-bit or P-bits not yet set. Such objects need
818 // to be safely navigable by block_start().
819 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
820 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
821 collector()->direct_allocated(res, adjustedSize);
822 _direct_allocated_words += adjustedSize;
823 // allocation counters
824 NOT_PRODUCT(
825 _numObjectsAllocated++;
826 _numWordsAllocated += (int)adjustedSize;
827 )
828 }
829 return res;
830 }
831
832 // In the case of direct allocation by mutators in a generation that
833 // is being concurrently collected, the object must be allocated
834 // live (grey) if the background collector has started marking.
835 // This is necessary because the marker may
836 // have passed this address and consequently this object will
837 // not otherwise be greyed and would be incorrectly swept up.
838 // Note that if this object contains references, the writing
839 // of those references will dirty the card containing this object
840 // allowing the object to be blackened (and its references scanned)
841 // either during a preclean phase or at the final checkpoint.
842 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
843 assert(_markBitMap.covers(start, size), "Out of bounds");
844 if (_collectorState >= Marking) {
845 MutexLockerEx y(_markBitMap.lock(),
846 Mutex::_no_safepoint_check_flag);
847 // [see comments preceding SweepClosure::do_blk() below for details]
848 //
849 // Can the P-bits be deleted now? JJJ
850 //
851 // 1. need to mark the object as live so it isn't collected
852 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
853 // 3. need to mark the end of the object so marking, precleaning or sweeping
854 // can skip over uninitialized or unparsable objects. An allocated
855 // object is considered uninitialized for our purposes as long as
856 // its klass word is NULL. All old gen objects are parsable
857 // as soon as they are initialized.)
858 _markBitMap.mark(start); // object is live
859 _markBitMap.mark(start + 1); // object is potentially uninitialized?
860 _markBitMap.mark(start + size - 1);
861 // mark end of object
862 }
863 // check that oop looks uninitialized
864 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
865 }
866
867 void CMSCollector::promoted(bool par, HeapWord* start,
868 bool is_obj_array, size_t obj_size) {
869 assert(_markBitMap.covers(start), "Out of bounds");
870 // See comment in direct_allocated() about when objects should
871 // be allocated live.
872 if (_collectorState >= Marking) {
873 // we already hold the marking bit map lock, taken in
874 // the prologue
875 if (par) {
876 _markBitMap.par_mark(start);
877 } else {
878 _markBitMap.mark(start);
879 }
880 // We don't need to mark the object as uninitialized (as
881 // in direct_allocated above) because this is being done with the
882 // world stopped and the object will be initialized by the
883 // time the marking, precleaning or sweeping get to look at it.
884 // But see the code for copying objects into the CMS generation,
885 // where we need to ensure that concurrent readers of the
886 // block offset table are able to safely navigate a block that
887 // is in flux from being free to being allocated (and in
888 // transition while being copied into) and subsequently
889 // becoming a bona-fide object when the copy/promotion is complete.
890 assert(SafepointSynchronize::is_at_safepoint(),
891 "expect promotion only at safepoints");
892
893 if (_collectorState < Sweeping) {
894 // Mark the appropriate cards in the modUnionTable, so that
895 // this object gets scanned before the sweep. If this is
896 // not done, CMS generation references in the object might
897 // not get marked.
898 // For the case of arrays, which are otherwise precisely
899 // marked, we need to dirty the entire array, not just its head.
900 if (is_obj_array) {
901 // The [par_]mark_range() method expects mr.end() below to
902 // be aligned to the granularity of a bit's representation
903 // in the heap. In the case of the MUT below, that's a
904 // card size.
905 MemRegion mr(start,
906 align_up(start + obj_size,
907 CardTable::card_size /* bytes */));
908 if (par) {
909 _modUnionTable.par_mark_range(mr);
910 } else {
911 _modUnionTable.mark_range(mr);
912 }
913 } else { // not an obj array; we can just mark the head
914 if (par) {
915 _modUnionTable.par_mark(start);
916 } else {
917 _modUnionTable.mark(start);
918 }
919 }
920 }
921 }
922 }
923
924 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
925 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
926 // allocate, copy and if necessary update promoinfo --
927 // delegate to underlying space.
928 assert_lock_strong(freelistLock());
929
930 #ifndef PRODUCT
931 if (CMSHeap::heap()->promotion_should_fail()) {
932 return NULL;
933 }
934 #endif // #ifndef PRODUCT
935
936 oop res = _cmsSpace->promote(obj, obj_size);
937 if (res == NULL) {
938 // expand and retry
939 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
940 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
941 // Since this is the old generation, we don't try to promote
942 // into a more senior generation.
943 res = _cmsSpace->promote(obj, obj_size);
944 }
945 if (res != NULL) {
946 // See comment in allocate() about when objects should
947 // be allocated live.
948 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below");
949 collector()->promoted(false, // Not parallel
950 (HeapWord*)res, obj->is_objArray(), obj_size);
951 // promotion counters
952 NOT_PRODUCT(
953 _numObjectsPromoted++;
954 _numWordsPromoted +=
955 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
956 )
957 }
958 return res;
959 }
960
961
962 // IMPORTANT: Notes on object size recognition in CMS.
963 // ---------------------------------------------------
964 // A block of storage in the CMS generation is always in
965 // one of three states. A free block (FREE), an allocated
966 // object (OBJECT) whose size() method reports the correct size,
967 // and an intermediate state (TRANSIENT) in which its size cannot
968 // be accurately determined.
969 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
970 // -----------------------------------------------------
971 // FREE: klass_word & 1 == 1; mark_word holds block size
972 //
973 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
974 // obj->size() computes correct size
975 //
976 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
977 //
978 // STATE IDENTIFICATION: (64 bit+COOPS)
979 // ------------------------------------
980 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
981 //
982 // OBJECT: klass_word installed; klass_word != 0;
983 // obj->size() computes correct size
984 //
985 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
986 //
987 //
988 // STATE TRANSITION DIAGRAM
989 //
990 // mut / parnew mut / parnew
991 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
992 // ^ |
993 // |------------------------ DEAD <------------------------------------|
994 // sweep mut
995 //
996 // While a block is in TRANSIENT state its size cannot be determined
997 // so readers will either need to come back later or stall until
998 // the size can be determined. Note that for the case of direct
999 // allocation, P-bits, when available, may be used to determine the
1000 // size of an object that may not yet have been initialized.
1001
1002 // Things to support parallel young-gen collection.
1003 oop
1004 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1005 oop old, markOop m,
1006 size_t word_sz) {
1007 #ifndef PRODUCT
1008 if (CMSHeap::heap()->promotion_should_fail()) {
1009 return NULL;
1010 }
1011 #endif // #ifndef PRODUCT
1012
1013 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1014 PromotionInfo* promoInfo = &ps->promo;
1015 // if we are tracking promotions, then first ensure space for
1016 // promotion (including spooling space for saving header if necessary).
1017 // then allocate and copy, then track promoted info if needed.
1018 // When tracking (see PromotionInfo::track()), the mark word may
1019 // be displaced and in this case restoration of the mark word
1020 // occurs in the (oop_since_save_marks_)iterate phase.
1021 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1022 // Out of space for allocating spooling buffers;
1023 // try expanding and allocating spooling buffers.
1024 if (!expand_and_ensure_spooling_space(promoInfo)) {
1025 return NULL;
1026 }
1027 }
1028 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant");
1029 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1030 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1031 if (obj_ptr == NULL) {
1032 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1033 if (obj_ptr == NULL) {
1034 return NULL;
1035 }
1036 }
1037 oop obj = oop(obj_ptr);
1038 OrderAccess::storestore();
1039 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1040 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1041 // IMPORTANT: See note on object initialization for CMS above.
1042 // Otherwise, copy the object. Here we must be careful to insert the
1043 // klass pointer last, since this marks the block as an allocated object.
1044 // Except with compressed oops it's the mark word.
1045 HeapWord* old_ptr = (HeapWord*)old;
1046 // Restore the mark word copied above.
1047 obj->set_mark_raw(m);
1048 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1049 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1050 OrderAccess::storestore();
1051
1052 if (UseCompressedClassPointers) {
1053 // Copy gap missed by (aligned) header size calculation below
1054 obj->set_klass_gap(old->klass_gap());
1055 }
1056 if (word_sz > (size_t)oopDesc::header_size()) {
1057 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1058 obj_ptr + oopDesc::header_size(),
1059 word_sz - oopDesc::header_size());
1060 }
1061
1062 // Now we can track the promoted object, if necessary. We take care
1063 // to delay the transition from uninitialized to full object
1064 // (i.e., insertion of klass pointer) until after, so that it
1065 // atomically becomes a promoted object.
1066 if (promoInfo->tracking()) {
1067 promoInfo->track((PromotedObject*)obj, old->klass());
1068 }
1069 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1070 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1071 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below");
1072
1073 // Finally, install the klass pointer (this should be volatile).
1074 OrderAccess::storestore();
1075 obj->set_klass(old->klass());
1076 // We should now be able to calculate the right size for this object
1077 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1078
1079 collector()->promoted(true, // parallel
1080 obj_ptr, old->is_objArray(), word_sz);
1081
1082 NOT_PRODUCT(
1083 Atomic::inc(&_numObjectsPromoted);
1084 Atomic::add(alloc_sz, &_numWordsPromoted);
1085 )
1086
1087 return obj;
1088 }
1089
1090 void
1091 ConcurrentMarkSweepGeneration::
1092 par_promote_alloc_done(int thread_num) {
1093 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1094 ps->lab.retire(thread_num);
1095 }
1096
1097 void
1098 ConcurrentMarkSweepGeneration::
1099 par_oop_since_save_marks_iterate_done(int thread_num) {
1100 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1101 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1102 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1103
1104 // Because card-scanning has been completed, subsequent phases
1105 // (e.g., reference processing) will not need to recognize which
1106 // objects have been promoted during this GC. So, we can now disable
1107 // promotion tracking.
1108 ps->promo.stopTrackingPromotions();
1109 }
1110
1111 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1112 size_t size,
1113 bool tlab)
1114 {
1115 // We allow a STW collection only if a full
1116 // collection was requested.
1117 return full || should_allocate(size, tlab); // FIX ME !!!
1118 // This and promotion failure handling are connected at the
1119 // hip and should be fixed by untying them.
1120 }
1121
1122 bool CMSCollector::shouldConcurrentCollect() {
1123 LogTarget(Trace, gc) log;
1124
1125 if (_full_gc_requested) {
1126 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)");
1127 return true;
1128 }
1129
1130 FreelistLocker x(this);
1131 // ------------------------------------------------------------------
1132 // Print out lots of information which affects the initiation of
1133 // a collection.
1134 if (log.is_enabled() && stats().valid()) {
1135 log.print("CMSCollector shouldConcurrentCollect: ");
1136
1137 LogStream out(log);
1138 stats().print_on(&out);
1139
1140 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full());
1141 log.print("free=" SIZE_FORMAT, _cmsGen->free());
1142 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available());
1143 log.print("promotion_rate=%g", stats().promotion_rate());
1144 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate());
1145 log.print("occupancy=%3.7f", _cmsGen->occupancy());
1146 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1147 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1148 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1149 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect());
1150 }
1151 // ------------------------------------------------------------------
1152
1153 // If the estimated time to complete a cms collection (cms_duration())
1154 // is less than the estimated time remaining until the cms generation
1155 // is full, start a collection.
1156 if (!UseCMSInitiatingOccupancyOnly) {
1157 if (stats().valid()) {
1158 if (stats().time_until_cms_start() == 0.0) {
1159 return true;
1160 }
1161 } else {
1162 // We want to conservatively collect somewhat early in order
1163 // to try and "bootstrap" our CMS/promotion statistics;
1164 // this branch will not fire after the first successful CMS
1165 // collection because the stats should then be valid.
1166 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1167 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f",
1168 _cmsGen->occupancy(), _bootstrap_occupancy);
1169 return true;
1170 }
1171 }
1172 }
1173
1174 // Otherwise, we start a collection cycle if
1175 // old gen want a collection cycle started. Each may use
1176 // an appropriate criterion for making this decision.
1177 // XXX We need to make sure that the gen expansion
1178 // criterion dovetails well with this. XXX NEED TO FIX THIS
1179 if (_cmsGen->should_concurrent_collect()) {
1180 log.print("CMS old gen initiated");
1181 return true;
1182 }
1183
1184 // We start a collection if we believe an incremental collection may fail;
1185 // this is not likely to be productive in practice because it's probably too
1186 // late anyway.
1187 CMSHeap* heap = CMSHeap::heap();
1188 if (heap->incremental_collection_will_fail(true /* consult_young */)) {
1189 log.print("CMSCollector: collect because incremental collection will fail ");
1190 return true;
1191 }
1192
1193 if (MetaspaceGC::should_concurrent_collect()) {
1194 log.print("CMSCollector: collect for metadata allocation ");
1195 return true;
1196 }
1197
1198 // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1199 if (CMSTriggerInterval >= 0) {
1200 if (CMSTriggerInterval == 0) {
1201 // Trigger always
1202 return true;
1203 }
1204
1205 // Check the CMS time since begin (we do not check the stats validity
1206 // as we want to be able to trigger the first CMS cycle as well)
1207 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1208 if (stats().valid()) {
1209 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1210 stats().cms_time_since_begin());
1211 } else {
1212 log.print("CMSCollector: collect because of trigger interval (first collection)");
1213 }
1214 return true;
1215 }
1216 }
1217
1218 return false;
1219 }
1220
1221 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1222
1223 // Clear _expansion_cause fields of constituent generations
1224 void CMSCollector::clear_expansion_cause() {
1225 _cmsGen->clear_expansion_cause();
1226 }
1227
1228 // We should be conservative in starting a collection cycle. To
1229 // start too eagerly runs the risk of collecting too often in the
1230 // extreme. To collect too rarely falls back on full collections,
1231 // which works, even if not optimum in terms of concurrent work.
1232 // As a work around for too eagerly collecting, use the flag
1233 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1234 // giving the user an easily understandable way of controlling the
1235 // collections.
1236 // We want to start a new collection cycle if any of the following
1237 // conditions hold:
1238 // . our current occupancy exceeds the configured initiating occupancy
1239 // for this generation, or
1240 // . we recently needed to expand this space and have not, since that
1241 // expansion, done a collection of this generation, or
1242 // . the underlying space believes that it may be a good idea to initiate
1243 // a concurrent collection (this may be based on criteria such as the
1244 // following: the space uses linear allocation and linear allocation is
1245 // going to fail, or there is believed to be excessive fragmentation in
1246 // the generation, etc... or ...
1247 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1248 // the case of the old generation; see CR 6543076):
1249 // we may be approaching a point at which allocation requests may fail because
1250 // we will be out of sufficient free space given allocation rate estimates.]
1251 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1252
1253 assert_lock_strong(freelistLock());
1254 if (occupancy() > initiating_occupancy()) {
1255 log_trace(gc)(" %s: collect because of occupancy %f / %f ",
1256 short_name(), occupancy(), initiating_occupancy());
1257 return true;
1258 }
1259 if (UseCMSInitiatingOccupancyOnly) {
1260 return false;
1261 }
1262 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1263 log_trace(gc)(" %s: collect because expanded for allocation ", short_name());
1264 return true;
1265 }
1266 return false;
1267 }
1268
1269 void ConcurrentMarkSweepGeneration::collect(bool full,
1270 bool clear_all_soft_refs,
1271 size_t size,
1272 bool tlab)
1273 {
1274 collector()->collect(full, clear_all_soft_refs, size, tlab);
1275 }
1276
1277 void CMSCollector::collect(bool full,
1278 bool clear_all_soft_refs,
1279 size_t size,
1280 bool tlab)
1281 {
1282 // The following "if" branch is present for defensive reasons.
1283 // In the current uses of this interface, it can be replaced with:
1284 // assert(!GCLocker.is_active(), "Can't be called otherwise");
1285 // But I am not placing that assert here to allow future
1286 // generality in invoking this interface.
1287 if (GCLocker::is_active()) {
1288 // A consistency test for GCLocker
1289 assert(GCLocker::needs_gc(), "Should have been set already");
1290 // Skip this foreground collection, instead
1291 // expanding the heap if necessary.
1292 // Need the free list locks for the call to free() in compute_new_size()
1293 compute_new_size();
1294 return;
1295 }
1296 acquire_control_and_collect(full, clear_all_soft_refs);
1297 }
1298
1299 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1300 CMSHeap* heap = CMSHeap::heap();
1301 unsigned int gc_count = heap->total_full_collections();
1302 if (gc_count == full_gc_count) {
1303 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1304 _full_gc_requested = true;
1305 _full_gc_cause = cause;
1306 CGC_lock->notify(); // nudge CMS thread
1307 } else {
1308 assert(gc_count > full_gc_count, "Error: causal loop");
1309 }
1310 }
1311
1312 bool CMSCollector::is_external_interruption() {
1313 GCCause::Cause cause = CMSHeap::heap()->gc_cause();
1314 return GCCause::is_user_requested_gc(cause) ||
1315 GCCause::is_serviceability_requested_gc(cause);
1316 }
1317
1318 void CMSCollector::report_concurrent_mode_interruption() {
1319 if (is_external_interruption()) {
1320 log_debug(gc)("Concurrent mode interrupted");
1321 } else {
1322 log_debug(gc)("Concurrent mode failure");
1323 _gc_tracer_cm->report_concurrent_mode_failure();
1324 }
1325 }
1326
1327
1328 // The foreground and background collectors need to coordinate in order
1329 // to make sure that they do not mutually interfere with CMS collections.
1330 // When a background collection is active,
1331 // the foreground collector may need to take over (preempt) and
1332 // synchronously complete an ongoing collection. Depending on the
1333 // frequency of the background collections and the heap usage
1334 // of the application, this preemption can be seldom or frequent.
1335 // There are only certain
1336 // points in the background collection that the "collection-baton"
1337 // can be passed to the foreground collector.
1338 //
1339 // The foreground collector will wait for the baton before
1340 // starting any part of the collection. The foreground collector
1341 // will only wait at one location.
1342 //
1343 // The background collector will yield the baton before starting a new
1344 // phase of the collection (e.g., before initial marking, marking from roots,
1345 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1346 // of the loop which switches the phases. The background collector does some
1347 // of the phases (initial mark, final re-mark) with the world stopped.
1348 // Because of locking involved in stopping the world,
1349 // the foreground collector should not block waiting for the background
1350 // collector when it is doing a stop-the-world phase. The background
1351 // collector will yield the baton at an additional point just before
1352 // it enters a stop-the-world phase. Once the world is stopped, the
1353 // background collector checks the phase of the collection. If the
1354 // phase has not changed, it proceeds with the collection. If the
1355 // phase has changed, it skips that phase of the collection. See
1356 // the comments on the use of the Heap_lock in collect_in_background().
1357 //
1358 // Variable used in baton passing.
1359 // _foregroundGCIsActive - Set to true by the foreground collector when
1360 // it wants the baton. The foreground clears it when it has finished
1361 // the collection.
1362 // _foregroundGCShouldWait - Set to true by the background collector
1363 // when it is running. The foreground collector waits while
1364 // _foregroundGCShouldWait is true.
1365 // CGC_lock - monitor used to protect access to the above variables
1366 // and to notify the foreground and background collectors.
1367 // _collectorState - current state of the CMS collection.
1368 //
1369 // The foreground collector
1370 // acquires the CGC_lock
1371 // sets _foregroundGCIsActive
1372 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1373 // various locks acquired in preparation for the collection
1374 // are released so as not to block the background collector
1375 // that is in the midst of a collection
1376 // proceeds with the collection
1377 // clears _foregroundGCIsActive
1378 // returns
1379 //
1380 // The background collector in a loop iterating on the phases of the
1381 // collection
1382 // acquires the CGC_lock
1383 // sets _foregroundGCShouldWait
1384 // if _foregroundGCIsActive is set
1385 // clears _foregroundGCShouldWait, notifies _CGC_lock
1386 // waits on _CGC_lock for _foregroundGCIsActive to become false
1387 // and exits the loop.
1388 // otherwise
1389 // proceed with that phase of the collection
1390 // if the phase is a stop-the-world phase,
1391 // yield the baton once more just before enqueueing
1392 // the stop-world CMS operation (executed by the VM thread).
1393 // returns after all phases of the collection are done
1394 //
1395
1396 void CMSCollector::acquire_control_and_collect(bool full,
1397 bool clear_all_soft_refs) {
1398 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1399 assert(!Thread::current()->is_ConcurrentGC_thread(),
1400 "shouldn't try to acquire control from self!");
1401
1402 // Start the protocol for acquiring control of the
1403 // collection from the background collector (aka CMS thread).
1404 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1405 "VM thread should have CMS token");
1406 // Remember the possibly interrupted state of an ongoing
1407 // concurrent collection
1408 CollectorState first_state = _collectorState;
1409
1410 // Signal to a possibly ongoing concurrent collection that
1411 // we want to do a foreground collection.
1412 _foregroundGCIsActive = true;
1413
1414 // release locks and wait for a notify from the background collector
1415 // releasing the locks in only necessary for phases which
1416 // do yields to improve the granularity of the collection.
1417 assert_lock_strong(bitMapLock());
1418 // We need to lock the Free list lock for the space that we are
1419 // currently collecting.
1420 assert(haveFreelistLocks(), "Must be holding free list locks");
1421 bitMapLock()->unlock();
1422 releaseFreelistLocks();
1423 {
1424 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1425 if (_foregroundGCShouldWait) {
1426 // We are going to be waiting for action for the CMS thread;
1427 // it had better not be gone (for instance at shutdown)!
1428 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(),
1429 "CMS thread must be running");
1430 // Wait here until the background collector gives us the go-ahead
1431 ConcurrentMarkSweepThread::clear_CMS_flag(
1432 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1433 // Get a possibly blocked CMS thread going:
1434 // Note that we set _foregroundGCIsActive true above,
1435 // without protection of the CGC_lock.
1436 CGC_lock->notify();
1437 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1438 "Possible deadlock");
1439 while (_foregroundGCShouldWait) {
1440 // wait for notification
1441 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1442 // Possibility of delay/starvation here, since CMS token does
1443 // not know to give priority to VM thread? Actually, i think
1444 // there wouldn't be any delay/starvation, but the proof of
1445 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1446 }
1447 ConcurrentMarkSweepThread::set_CMS_flag(
1448 ConcurrentMarkSweepThread::CMS_vm_has_token);
1449 }
1450 }
1451 // The CMS_token is already held. Get back the other locks.
1452 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1453 "VM thread should have CMS token");
1454 getFreelistLocks();
1455 bitMapLock()->lock_without_safepoint_check();
1456 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d",
1457 p2i(Thread::current()), first_state);
1458 log_debug(gc, state)(" gets control with state %d", _collectorState);
1459
1460 // Inform cms gen if this was due to partial collection failing.
1461 // The CMS gen may use this fact to determine its expansion policy.
1462 CMSHeap* heap = CMSHeap::heap();
1463 if (heap->incremental_collection_will_fail(false /* don't consult_young */)) {
1464 assert(!_cmsGen->incremental_collection_failed(),
1465 "Should have been noticed, reacted to and cleared");
1466 _cmsGen->set_incremental_collection_failed();
1467 }
1468
1469 if (first_state > Idling) {
1470 report_concurrent_mode_interruption();
1471 }
1472
1473 set_did_compact(true);
1474
1475 // If the collection is being acquired from the background
1476 // collector, there may be references on the discovered
1477 // references lists. Abandon those references, since some
1478 // of them may have become unreachable after concurrent
1479 // discovery; the STW compacting collector will redo discovery
1480 // more precisely, without being subject to floating garbage.
1481 // Leaving otherwise unreachable references in the discovered
1482 // lists would require special handling.
1483 ref_processor()->disable_discovery();
1484 ref_processor()->abandon_partial_discovery();
1485 ref_processor()->verify_no_references_recorded();
1486
1487 if (first_state > Idling) {
1488 save_heap_summary();
1489 }
1490
1491 do_compaction_work(clear_all_soft_refs);
1492
1493 // Has the GC time limit been exceeded?
1494 size_t max_eden_size = _young_gen->max_eden_size();
1495 GCCause::Cause gc_cause = heap->gc_cause();
1496 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1497 _young_gen->eden()->used(),
1498 _cmsGen->max_capacity(),
1499 max_eden_size,
1500 full,
1501 gc_cause,
1502 heap->soft_ref_policy());
1503
1504 // Reset the expansion cause, now that we just completed
1505 // a collection cycle.
1506 clear_expansion_cause();
1507 _foregroundGCIsActive = false;
1508 return;
1509 }
1510
1511 // Resize the tenured generation
1512 // after obtaining the free list locks for the
1513 // two generations.
1514 void CMSCollector::compute_new_size() {
1515 assert_locked_or_safepoint(Heap_lock);
1516 FreelistLocker z(this);
1517 MetaspaceGC::compute_new_size();
1518 _cmsGen->compute_new_size_free_list();
1519 }
1520
1521 // A work method used by the foreground collector to do
1522 // a mark-sweep-compact.
1523 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1524 CMSHeap* heap = CMSHeap::heap();
1525
1526 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1527 gc_timer->register_gc_start();
1528
1529 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1530 gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start());
1531
1532 heap->pre_full_gc_dump(gc_timer);
1533
1534 GCTraceTime(Trace, gc, phases) t("CMS:MSC");
1535
1536 // Temporarily widen the span of the weak reference processing to
1537 // the entire heap.
1538 MemRegion new_span(CMSHeap::heap()->reserved_region());
1539 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1540 // Temporarily, clear the "is_alive_non_header" field of the
1541 // reference processor.
1542 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1543 // Temporarily make reference _processing_ single threaded (non-MT).
1544 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1545 // Temporarily make refs discovery atomic
1546 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1547 // Temporarily make reference _discovery_ single threaded (non-MT)
1548 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1549
1550 ref_processor()->set_enqueuing_is_done(false);
1551 ref_processor()->enable_discovery();
1552 ref_processor()->setup_policy(clear_all_soft_refs);
1553 // If an asynchronous collection finishes, the _modUnionTable is
1554 // all clear. If we are assuming the collection from an asynchronous
1555 // collection, clear the _modUnionTable.
1556 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1557 "_modUnionTable should be clear if the baton was not passed");
1558 _modUnionTable.clear_all();
1559 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(),
1560 "mod union for klasses should be clear if the baton was passed");
1561 _ct->cld_rem_set()->clear_mod_union();
1562
1563
1564 // We must adjust the allocation statistics being maintained
1565 // in the free list space. We do so by reading and clearing
1566 // the sweep timer and updating the block flux rate estimates below.
1567 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1568 if (_inter_sweep_timer.is_active()) {
1569 _inter_sweep_timer.stop();
1570 // Note that we do not use this sample to update the _inter_sweep_estimate.
1571 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1572 _inter_sweep_estimate.padded_average(),
1573 _intra_sweep_estimate.padded_average());
1574 }
1575
1576 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs);
1577 #ifdef ASSERT
1578 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1579 size_t free_size = cms_space->free();
1580 assert(free_size ==
1581 pointer_delta(cms_space->end(), cms_space->compaction_top())
1582 * HeapWordSize,
1583 "All the free space should be compacted into one chunk at top");
1584 assert(cms_space->dictionary()->total_chunk_size(
1585 debug_only(cms_space->freelistLock())) == 0 ||
1586 cms_space->totalSizeInIndexedFreeLists() == 0,
1587 "All the free space should be in a single chunk");
1588 size_t num = cms_space->totalCount();
1589 assert((free_size == 0 && num == 0) ||
1590 (free_size > 0 && (num == 1 || num == 2)),
1591 "There should be at most 2 free chunks after compaction");
1592 #endif // ASSERT
1593 _collectorState = Resetting;
1594 assert(_restart_addr == NULL,
1595 "Should have been NULL'd before baton was passed");
1596 reset_stw();
1597 _cmsGen->reset_after_compaction();
1598 _concurrent_cycles_since_last_unload = 0;
1599
1600 // Clear any data recorded in the PLAB chunk arrays.
1601 if (_survivor_plab_array != NULL) {
1602 reset_survivor_plab_arrays();
1603 }
1604
1605 // Adjust the per-size allocation stats for the next epoch.
1606 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1607 // Restart the "inter sweep timer" for the next epoch.
1608 _inter_sweep_timer.reset();
1609 _inter_sweep_timer.start();
1610
1611 // No longer a need to do a concurrent collection for Metaspace.
1612 MetaspaceGC::set_should_concurrent_collect(false);
1613
1614 heap->post_full_gc_dump(gc_timer);
1615
1616 gc_timer->register_gc_end();
1617
1618 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1619
1620 // For a mark-sweep-compact, compute_new_size() will be called
1621 // in the heap's do_collection() method.
1622 }
1623
1624 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1625 Log(gc, heap) log;
1626 if (!log.is_trace()) {
1627 return;
1628 }
1629
1630 ContiguousSpace* eden_space = _young_gen->eden();
1631 ContiguousSpace* from_space = _young_gen->from();
1632 ContiguousSpace* to_space = _young_gen->to();
1633 // Eden
1634 if (_eden_chunk_array != NULL) {
1635 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1636 p2i(eden_space->bottom()), p2i(eden_space->top()),
1637 p2i(eden_space->end()), eden_space->capacity());
1638 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT,
1639 _eden_chunk_index, _eden_chunk_capacity);
1640 for (size_t i = 0; i < _eden_chunk_index; i++) {
1641 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i]));
1642 }
1643 }
1644 // Survivor
1645 if (_survivor_chunk_array != NULL) {
1646 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1647 p2i(from_space->bottom()), p2i(from_space->top()),
1648 p2i(from_space->end()), from_space->capacity());
1649 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT,
1650 _survivor_chunk_index, _survivor_chunk_capacity);
1651 for (size_t i = 0; i < _survivor_chunk_index; i++) {
1652 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i]));
1653 }
1654 }
1655 }
1656
1657 void CMSCollector::getFreelistLocks() const {
1658 // Get locks for all free lists in all generations that this
1659 // collector is responsible for
1660 _cmsGen->freelistLock()->lock_without_safepoint_check();
1661 }
1662
1663 void CMSCollector::releaseFreelistLocks() const {
1664 // Release locks for all free lists in all generations that this
1665 // collector is responsible for
1666 _cmsGen->freelistLock()->unlock();
1667 }
1668
1669 bool CMSCollector::haveFreelistLocks() const {
1670 // Check locks for all free lists in all generations that this
1671 // collector is responsible for
1672 assert_lock_strong(_cmsGen->freelistLock());
1673 PRODUCT_ONLY(ShouldNotReachHere());
1674 return true;
1675 }
1676
1677 // A utility class that is used by the CMS collector to
1678 // temporarily "release" the foreground collector from its
1679 // usual obligation to wait for the background collector to
1680 // complete an ongoing phase before proceeding.
1681 class ReleaseForegroundGC: public StackObj {
1682 private:
1683 CMSCollector* _c;
1684 public:
1685 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1686 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1687 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1688 // allow a potentially blocked foreground collector to proceed
1689 _c->_foregroundGCShouldWait = false;
1690 if (_c->_foregroundGCIsActive) {
1691 CGC_lock->notify();
1692 }
1693 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1694 "Possible deadlock");
1695 }
1696
1697 ~ReleaseForegroundGC() {
1698 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1699 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1700 _c->_foregroundGCShouldWait = true;
1701 }
1702 };
1703
1704 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1705 assert(Thread::current()->is_ConcurrentGC_thread(),
1706 "A CMS asynchronous collection is only allowed on a CMS thread.");
1707
1708 CMSHeap* heap = CMSHeap::heap();
1709 {
1710 bool safepoint_check = Mutex::_no_safepoint_check_flag;
1711 MutexLockerEx hl(Heap_lock, safepoint_check);
1712 FreelistLocker fll(this);
1713 MutexLockerEx x(CGC_lock, safepoint_check);
1714 if (_foregroundGCIsActive) {
1715 // The foreground collector is. Skip this
1716 // background collection.
1717 assert(!_foregroundGCShouldWait, "Should be clear");
1718 return;
1719 } else {
1720 assert(_collectorState == Idling, "Should be idling before start.");
1721 _collectorState = InitialMarking;
1722 register_gc_start(cause);
1723 // Reset the expansion cause, now that we are about to begin
1724 // a new cycle.
1725 clear_expansion_cause();
1726
1727 // Clear the MetaspaceGC flag since a concurrent collection
1728 // is starting but also clear it after the collection.
1729 MetaspaceGC::set_should_concurrent_collect(false);
1730 }
1731 // Decide if we want to enable class unloading as part of the
1732 // ensuing concurrent GC cycle.
1733 update_should_unload_classes();
1734 _full_gc_requested = false; // acks all outstanding full gc requests
1735 _full_gc_cause = GCCause::_no_gc;
1736 // Signal that we are about to start a collection
1737 heap->increment_total_full_collections(); // ... starting a collection cycle
1738 _collection_count_start = heap->total_full_collections();
1739 }
1740
1741 size_t prev_used = _cmsGen->used();
1742
1743 // The change of the collection state is normally done at this level;
1744 // the exceptions are phases that are executed while the world is
1745 // stopped. For those phases the change of state is done while the
1746 // world is stopped. For baton passing purposes this allows the
1747 // background collector to finish the phase and change state atomically.
1748 // The foreground collector cannot wait on a phase that is done
1749 // while the world is stopped because the foreground collector already
1750 // has the world stopped and would deadlock.
1751 while (_collectorState != Idling) {
1752 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d",
1753 p2i(Thread::current()), _collectorState);
1754 // The foreground collector
1755 // holds the Heap_lock throughout its collection.
1756 // holds the CMS token (but not the lock)
1757 // except while it is waiting for the background collector to yield.
1758 //
1759 // The foreground collector should be blocked (not for long)
1760 // if the background collector is about to start a phase
1761 // executed with world stopped. If the background
1762 // collector has already started such a phase, the
1763 // foreground collector is blocked waiting for the
1764 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
1765 // are executed in the VM thread.
1766 //
1767 // The locking order is
1768 // PendingListLock (PLL) -- if applicable (FinalMarking)
1769 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
1770 // CMS token (claimed in
1771 // stop_world_and_do() -->
1772 // safepoint_synchronize() -->
1773 // CMSThread::synchronize())
1774
1775 {
1776 // Check if the FG collector wants us to yield.
1777 CMSTokenSync x(true); // is cms thread
1778 if (waitForForegroundGC()) {
1779 // We yielded to a foreground GC, nothing more to be
1780 // done this round.
1781 assert(_foregroundGCShouldWait == false, "We set it to false in "
1782 "waitForForegroundGC()");
1783 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1784 p2i(Thread::current()), _collectorState);
1785 return;
1786 } else {
1787 // The background collector can run but check to see if the
1788 // foreground collector has done a collection while the
1789 // background collector was waiting to get the CGC_lock
1790 // above. If yes, break so that _foregroundGCShouldWait
1791 // is cleared before returning.
1792 if (_collectorState == Idling) {
1793 break;
1794 }
1795 }
1796 }
1797
1798 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1799 "should be waiting");
1800
1801 switch (_collectorState) {
1802 case InitialMarking:
1803 {
1804 ReleaseForegroundGC x(this);
1805 stats().record_cms_begin();
1806 VM_CMS_Initial_Mark initial_mark_op(this);
1807 VMThread::execute(&initial_mark_op);
1808 }
1809 // The collector state may be any legal state at this point
1810 // since the background collector may have yielded to the
1811 // foreground collector.
1812 break;
1813 case Marking:
1814 // initial marking in checkpointRootsInitialWork has been completed
1815 if (markFromRoots()) { // we were successful
1816 assert(_collectorState == Precleaning, "Collector state should "
1817 "have changed");
1818 } else {
1819 assert(_foregroundGCIsActive, "Internal state inconsistency");
1820 }
1821 break;
1822 case Precleaning:
1823 // marking from roots in markFromRoots has been completed
1824 preclean();
1825 assert(_collectorState == AbortablePreclean ||
1826 _collectorState == FinalMarking,
1827 "Collector state should have changed");
1828 break;
1829 case AbortablePreclean:
1830 abortable_preclean();
1831 assert(_collectorState == FinalMarking, "Collector state should "
1832 "have changed");
1833 break;
1834 case FinalMarking:
1835 {
1836 ReleaseForegroundGC x(this);
1837
1838 VM_CMS_Final_Remark final_remark_op(this);
1839 VMThread::execute(&final_remark_op);
1840 }
1841 assert(_foregroundGCShouldWait, "block post-condition");
1842 break;
1843 case Sweeping:
1844 // final marking in checkpointRootsFinal has been completed
1845 sweep();
1846 assert(_collectorState == Resizing, "Collector state change "
1847 "to Resizing must be done under the free_list_lock");
1848
1849 case Resizing: {
1850 // Sweeping has been completed...
1851 // At this point the background collection has completed.
1852 // Don't move the call to compute_new_size() down
1853 // into code that might be executed if the background
1854 // collection was preempted.
1855 {
1856 ReleaseForegroundGC x(this); // unblock FG collection
1857 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
1858 CMSTokenSync z(true); // not strictly needed.
1859 if (_collectorState == Resizing) {
1860 compute_new_size();
1861 save_heap_summary();
1862 _collectorState = Resetting;
1863 } else {
1864 assert(_collectorState == Idling, "The state should only change"
1865 " because the foreground collector has finished the collection");
1866 }
1867 }
1868 break;
1869 }
1870 case Resetting:
1871 // CMS heap resizing has been completed
1872 reset_concurrent();
1873 assert(_collectorState == Idling, "Collector state should "
1874 "have changed");
1875
1876 MetaspaceGC::set_should_concurrent_collect(false);
1877
1878 stats().record_cms_end();
1879 // Don't move the concurrent_phases_end() and compute_new_size()
1880 // calls to here because a preempted background collection
1881 // has it's state set to "Resetting".
1882 break;
1883 case Idling:
1884 default:
1885 ShouldNotReachHere();
1886 break;
1887 }
1888 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d",
1889 p2i(Thread::current()), _collectorState);
1890 assert(_foregroundGCShouldWait, "block post-condition");
1891 }
1892
1893 // Should this be in gc_epilogue?
1894 heap->counters()->update_counters();
1895
1896 {
1897 // Clear _foregroundGCShouldWait and, in the event that the
1898 // foreground collector is waiting, notify it, before
1899 // returning.
1900 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1901 _foregroundGCShouldWait = false;
1902 if (_foregroundGCIsActive) {
1903 CGC_lock->notify();
1904 }
1905 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1906 "Possible deadlock");
1907 }
1908 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1909 p2i(Thread::current()), _collectorState);
1910 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)",
1911 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K);
1912 }
1913
1914 void CMSCollector::register_gc_start(GCCause::Cause cause) {
1915 _cms_start_registered = true;
1916 _gc_timer_cm->register_gc_start();
1917 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
1918 }
1919
1920 void CMSCollector::register_gc_end() {
1921 if (_cms_start_registered) {
1922 report_heap_summary(GCWhen::AfterGC);
1923
1924 _gc_timer_cm->register_gc_end();
1925 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1926 _cms_start_registered = false;
1927 }
1928 }
1929
1930 void CMSCollector::save_heap_summary() {
1931 CMSHeap* heap = CMSHeap::heap();
1932 _last_heap_summary = heap->create_heap_summary();
1933 _last_metaspace_summary = heap->create_metaspace_summary();
1934 }
1935
1936 void CMSCollector::report_heap_summary(GCWhen::Type when) {
1937 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
1938 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
1939 }
1940
1941 bool CMSCollector::waitForForegroundGC() {
1942 bool res = false;
1943 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1944 "CMS thread should have CMS token");
1945 // Block the foreground collector until the
1946 // background collectors decides whether to
1947 // yield.
1948 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1949 _foregroundGCShouldWait = true;
1950 if (_foregroundGCIsActive) {
1951 // The background collector yields to the
1952 // foreground collector and returns a value
1953 // indicating that it has yielded. The foreground
1954 // collector can proceed.
1955 res = true;
1956 _foregroundGCShouldWait = false;
1957 ConcurrentMarkSweepThread::clear_CMS_flag(
1958 ConcurrentMarkSweepThread::CMS_cms_has_token);
1959 ConcurrentMarkSweepThread::set_CMS_flag(
1960 ConcurrentMarkSweepThread::CMS_cms_wants_token);
1961 // Get a possibly blocked foreground thread going
1962 CGC_lock->notify();
1963 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
1964 p2i(Thread::current()), _collectorState);
1965 while (_foregroundGCIsActive) {
1966 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1967 }
1968 ConcurrentMarkSweepThread::set_CMS_flag(
1969 ConcurrentMarkSweepThread::CMS_cms_has_token);
1970 ConcurrentMarkSweepThread::clear_CMS_flag(
1971 ConcurrentMarkSweepThread::CMS_cms_wants_token);
1972 }
1973 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
1974 p2i(Thread::current()), _collectorState);
1975 return res;
1976 }
1977
1978 // Because of the need to lock the free lists and other structures in
1979 // the collector, common to all the generations that the collector is
1980 // collecting, we need the gc_prologues of individual CMS generations
1981 // delegate to their collector. It may have been simpler had the
1982 // current infrastructure allowed one to call a prologue on a
1983 // collector. In the absence of that we have the generation's
1984 // prologue delegate to the collector, which delegates back
1985 // some "local" work to a worker method in the individual generations
1986 // that it's responsible for collecting, while itself doing any
1987 // work common to all generations it's responsible for. A similar
1988 // comment applies to the gc_epilogue()'s.
1989 // The role of the variable _between_prologue_and_epilogue is to
1990 // enforce the invocation protocol.
1991 void CMSCollector::gc_prologue(bool full) {
1992 // Call gc_prologue_work() for the CMSGen
1993 // we are responsible for.
1994
1995 // The following locking discipline assumes that we are only called
1996 // when the world is stopped.
1997 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
1998
1999 // The CMSCollector prologue must call the gc_prologues for the
2000 // "generations" that it's responsible
2001 // for.
2002
2003 assert( Thread::current()->is_VM_thread()
2004 || ( CMSScavengeBeforeRemark
2005 && Thread::current()->is_ConcurrentGC_thread()),
2006 "Incorrect thread type for prologue execution");
2007
2008 if (_between_prologue_and_epilogue) {
2009 // We have already been invoked; this is a gc_prologue delegation
2010 // from yet another CMS generation that we are responsible for, just
2011 // ignore it since all relevant work has already been done.
2012 return;
2013 }
2014
2015 // set a bit saying prologue has been called; cleared in epilogue
2016 _between_prologue_and_epilogue = true;
2017 // Claim locks for common data structures, then call gc_prologue_work()
2018 // for each CMSGen.
2019
2020 getFreelistLocks(); // gets free list locks on constituent spaces
2021 bitMapLock()->lock_without_safepoint_check();
2022
2023 // Should call gc_prologue_work() for all cms gens we are responsible for
2024 bool duringMarking = _collectorState >= Marking
2025 && _collectorState < Sweeping;
2026
2027 // The young collections clear the modified oops state, which tells if
2028 // there are any modified oops in the class. The remark phase also needs
2029 // that information. Tell the young collection to save the union of all
2030 // modified klasses.
2031 if (duringMarking) {
2032 _ct->cld_rem_set()->set_accumulate_modified_oops(true);
2033 }
2034
2035 bool registerClosure = duringMarking;
2036
2037 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2038
2039 if (!full) {
2040 stats().record_gc0_begin();
2041 }
2042 }
2043
2044 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2045
2046 _capacity_at_prologue = capacity();
2047 _used_at_prologue = used();
2048
2049 // We enable promotion tracking so that card-scanning can recognize
2050 // which objects have been promoted during this GC and skip them.
2051 for (uint i = 0; i < ParallelGCThreads; i++) {
2052 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2053 }
2054
2055 // Delegate to CMScollector which knows how to coordinate between
2056 // this and any other CMS generations that it is responsible for
2057 // collecting.
2058 collector()->gc_prologue(full);
2059 }
2060
2061 // This is a "private" interface for use by this generation's CMSCollector.
2062 // Not to be called directly by any other entity (for instance,
2063 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2064 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2065 bool registerClosure, ModUnionClosure* modUnionClosure) {
2066 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2067 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2068 "Should be NULL");
2069 if (registerClosure) {
2070 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2071 }
2072 cmsSpace()->gc_prologue();
2073 // Clear stat counters
2074 NOT_PRODUCT(
2075 assert(_numObjectsPromoted == 0, "check");
2076 assert(_numWordsPromoted == 0, "check");
2077 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently",
2078 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2079 _numObjectsAllocated = 0;
2080 _numWordsAllocated = 0;
2081 )
2082 }
2083
2084 void CMSCollector::gc_epilogue(bool full) {
2085 // The following locking discipline assumes that we are only called
2086 // when the world is stopped.
2087 assert(SafepointSynchronize::is_at_safepoint(),
2088 "world is stopped assumption");
2089
2090 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2091 // if linear allocation blocks need to be appropriately marked to allow the
2092 // the blocks to be parsable. We also check here whether we need to nudge the
2093 // CMS collector thread to start a new cycle (if it's not already active).
2094 assert( Thread::current()->is_VM_thread()
2095 || ( CMSScavengeBeforeRemark
2096 && Thread::current()->is_ConcurrentGC_thread()),
2097 "Incorrect thread type for epilogue execution");
2098
2099 if (!_between_prologue_and_epilogue) {
2100 // We have already been invoked; this is a gc_epilogue delegation
2101 // from yet another CMS generation that we are responsible for, just
2102 // ignore it since all relevant work has already been done.
2103 return;
2104 }
2105 assert(haveFreelistLocks(), "must have freelist locks");
2106 assert_lock_strong(bitMapLock());
2107
2108 _ct->cld_rem_set()->set_accumulate_modified_oops(false);
2109
2110 _cmsGen->gc_epilogue_work(full);
2111
2112 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2113 // in case sampling was not already enabled, enable it
2114 _start_sampling = true;
2115 }
2116 // reset _eden_chunk_array so sampling starts afresh
2117 _eden_chunk_index = 0;
2118
2119 size_t cms_used = _cmsGen->cmsSpace()->used();
2120
2121 // update performance counters - this uses a special version of
2122 // update_counters() that allows the utilization to be passed as a
2123 // parameter, avoiding multiple calls to used().
2124 //
2125 _cmsGen->update_counters(cms_used);
2126
2127 bitMapLock()->unlock();
2128 releaseFreelistLocks();
2129
2130 if (!CleanChunkPoolAsync) {
2131 Chunk::clean_chunk_pool();
2132 }
2133
2134 set_did_compact(false);
2135 _between_prologue_and_epilogue = false; // ready for next cycle
2136 }
2137
2138 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2139 collector()->gc_epilogue(full);
2140
2141 // When using ParNew, promotion tracking should have already been
2142 // disabled. However, the prologue (which enables promotion
2143 // tracking) and epilogue are called irrespective of the type of
2144 // GC. So they will also be called before and after Full GCs, during
2145 // which promotion tracking will not be explicitly disabled. So,
2146 // it's safer to also disable it here too (to be symmetric with
2147 // enabling it in the prologue).
2148 for (uint i = 0; i < ParallelGCThreads; i++) {
2149 _par_gc_thread_states[i]->promo.stopTrackingPromotions();
2150 }
2151 }
2152
2153 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2154 assert(!incremental_collection_failed(), "Should have been cleared");
2155 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2156 cmsSpace()->gc_epilogue();
2157 // Print stat counters
2158 NOT_PRODUCT(
2159 assert(_numObjectsAllocated == 0, "check");
2160 assert(_numWordsAllocated == 0, "check");
2161 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
2162 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2163 _numObjectsPromoted = 0;
2164 _numWordsPromoted = 0;
2165 )
2166
2167 // Call down the chain in contiguous_available needs the freelistLock
2168 // so print this out before releasing the freeListLock.
2169 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available());
2170 }
2171
2172 #ifndef PRODUCT
2173 bool CMSCollector::have_cms_token() {
2174 Thread* thr = Thread::current();
2175 if (thr->is_VM_thread()) {
2176 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2177 } else if (thr->is_ConcurrentGC_thread()) {
2178 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2179 } else if (thr->is_GC_task_thread()) {
2180 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2181 ParGCRareEvent_lock->owned_by_self();
2182 }
2183 return false;
2184 }
2185
2186 // Check reachability of the given heap address in CMS generation,
2187 // treating all other generations as roots.
2188 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2189 // We could "guarantee" below, rather than assert, but I'll
2190 // leave these as "asserts" so that an adventurous debugger
2191 // could try this in the product build provided some subset of
2192 // the conditions were met, provided they were interested in the
2193 // results and knew that the computation below wouldn't interfere
2194 // with other concurrent computations mutating the structures
2195 // being read or written.
2196 assert(SafepointSynchronize::is_at_safepoint(),
2197 "Else mutations in object graph will make answer suspect");
2198 assert(have_cms_token(), "Should hold cms token");
2199 assert(haveFreelistLocks(), "must hold free list locks");
2200 assert_lock_strong(bitMapLock());
2201
2202 // Clear the marking bit map array before starting, but, just
2203 // for kicks, first report if the given address is already marked
2204 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr),
2205 _markBitMap.isMarked(addr) ? "" : " not");
2206
2207 if (verify_after_remark()) {
2208 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2209 bool result = verification_mark_bm()->isMarked(addr);
2210 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr),
2211 result ? "IS" : "is NOT");
2212 return result;
2213 } else {
2214 tty->print_cr("Could not compute result");
2215 return false;
2216 }
2217 }
2218 #endif
2219
2220 void
2221 CMSCollector::print_on_error(outputStream* st) {
2222 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2223 if (collector != NULL) {
2224 CMSBitMap* bitmap = &collector->_markBitMap;
2225 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2226 bitmap->print_on_error(st, " Bits: ");
2227
2228 st->cr();
2229
2230 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2231 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2232 mut_bitmap->print_on_error(st, " Bits: ");
2233 }
2234 }
2235
2236 ////////////////////////////////////////////////////////
2237 // CMS Verification Support
2238 ////////////////////////////////////////////////////////
2239 // Following the remark phase, the following invariant
2240 // should hold -- each object in the CMS heap which is
2241 // marked in markBitMap() should be marked in the verification_mark_bm().
2242
2243 class VerifyMarkedClosure: public BitMapClosure {
2244 CMSBitMap* _marks;
2245 bool _failed;
2246
2247 public:
2248 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2249
2250 bool do_bit(size_t offset) {
2251 HeapWord* addr = _marks->offsetToHeapWord(offset);
2252 if (!_marks->isMarked(addr)) {
2253 Log(gc, verify) log;
2254 ResourceMark rm;
2255 LogStream ls(log.error());
2256 oop(addr)->print_on(&ls);
2257 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
2258 _failed = true;
2259 }
2260 return true;
2261 }
2262
2263 bool failed() { return _failed; }
2264 };
2265
2266 bool CMSCollector::verify_after_remark() {
2267 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking.");
2268 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2269 static bool init = false;
2270
2271 assert(SafepointSynchronize::is_at_safepoint(),
2272 "Else mutations in object graph will make answer suspect");
2273 assert(have_cms_token(),
2274 "Else there may be mutual interference in use of "
2275 " verification data structures");
2276 assert(_collectorState > Marking && _collectorState <= Sweeping,
2277 "Else marking info checked here may be obsolete");
2278 assert(haveFreelistLocks(), "must hold free list locks");
2279 assert_lock_strong(bitMapLock());
2280
2281
2282 // Allocate marking bit map if not already allocated
2283 if (!init) { // first time
2284 if (!verification_mark_bm()->allocate(_span)) {
2285 return false;
2286 }
2287 init = true;
2288 }
2289
2290 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2291
2292 // Turn off refs discovery -- so we will be tracing through refs.
2293 // This is as intended, because by this time
2294 // GC must already have cleared any refs that need to be cleared,
2295 // and traced those that need to be marked; moreover,
2296 // the marking done here is not going to interfere in any
2297 // way with the marking information used by GC.
2298 NoRefDiscovery no_discovery(ref_processor());
2299
2300 #if COMPILER2_OR_JVMCI
2301 DerivedPointerTableDeactivate dpt_deact;
2302 #endif
2303
2304 // Clear any marks from a previous round
2305 verification_mark_bm()->clear_all();
2306 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2307 verify_work_stacks_empty();
2308
2309 CMSHeap* heap = CMSHeap::heap();
2310 heap->ensure_parsability(false); // fill TLABs, but no need to retire them
2311 // Update the saved marks which may affect the root scans.
2312 heap->save_marks();
2313
2314 if (CMSRemarkVerifyVariant == 1) {
2315 // In this first variant of verification, we complete
2316 // all marking, then check if the new marks-vector is
2317 // a subset of the CMS marks-vector.
2318 verify_after_remark_work_1();
2319 } else {
2320 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2");
2321 // In this second variant of verification, we flag an error
2322 // (i.e. an object reachable in the new marks-vector not reachable
2323 // in the CMS marks-vector) immediately, also indicating the
2324 // identify of an object (A) that references the unmarked object (B) --
2325 // presumably, a mutation to A failed to be picked up by preclean/remark?
2326 verify_after_remark_work_2();
2327 }
2328
2329 return true;
2330 }
2331
2332 void CMSCollector::verify_after_remark_work_1() {
2333 ResourceMark rm;
2334 HandleMark hm;
2335 CMSHeap* heap = CMSHeap::heap();
2336
2337 // Get a clear set of claim bits for the roots processing to work with.
2338 ClassLoaderDataGraph::clear_claimed_marks();
2339
2340 // Mark from roots one level into CMS
2341 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2342 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2343
2344 {
2345 StrongRootsScope srs(1);
2346
2347 heap->cms_process_roots(&srs,
2348 true, // young gen as roots
2349 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2350 should_unload_classes(),
2351 ¬Older,
2352 NULL);
2353 }
2354
2355 // Now mark from the roots
2356 MarkFromRootsClosure markFromRootsClosure(this, _span,
2357 verification_mark_bm(), verification_mark_stack(),
2358 false /* don't yield */, true /* verifying */);
2359 assert(_restart_addr == NULL, "Expected pre-condition");
2360 verification_mark_bm()->iterate(&markFromRootsClosure);
2361 while (_restart_addr != NULL) {
2362 // Deal with stack overflow: by restarting at the indicated
2363 // address.
2364 HeapWord* ra = _restart_addr;
2365 markFromRootsClosure.reset(ra);
2366 _restart_addr = NULL;
2367 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2368 }
2369 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2370 verify_work_stacks_empty();
2371
2372 // Marking completed -- now verify that each bit marked in
2373 // verification_mark_bm() is also marked in markBitMap(); flag all
2374 // errors by printing corresponding objects.
2375 VerifyMarkedClosure vcl(markBitMap());
2376 verification_mark_bm()->iterate(&vcl);
2377 if (vcl.failed()) {
2378 Log(gc, verify) log;
2379 log.error("Failed marking verification after remark");
2380 ResourceMark rm;
2381 LogStream ls(log.error());
2382 heap->print_on(&ls);
2383 fatal("CMS: failed marking verification after remark");
2384 }
2385 }
2386
2387 class VerifyCLDOopsCLDClosure : public CLDClosure {
2388 class VerifyCLDOopsClosure : public OopClosure {
2389 CMSBitMap* _bitmap;
2390 public:
2391 VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2392 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2393 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2394 } _oop_closure;
2395 public:
2396 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2397 void do_cld(ClassLoaderData* cld) {
2398 cld->oops_do(&_oop_closure, false, false);
2399 }
2400 };
2401
2402 void CMSCollector::verify_after_remark_work_2() {
2403 ResourceMark rm;
2404 HandleMark hm;
2405 CMSHeap* heap = CMSHeap::heap();
2406
2407 // Get a clear set of claim bits for the roots processing to work with.
2408 ClassLoaderDataGraph::clear_claimed_marks();
2409
2410 // Mark from roots one level into CMS
2411 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2412 markBitMap());
2413 CLDToOopClosure cld_closure(¬Older, true);
2414
2415 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2416
2417 {
2418 StrongRootsScope srs(1);
2419
2420 heap->cms_process_roots(&srs,
2421 true, // young gen as roots
2422 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2423 should_unload_classes(),
2424 ¬Older,
2425 &cld_closure);
2426 }
2427
2428 // Now mark from the roots
2429 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2430 verification_mark_bm(), markBitMap(), verification_mark_stack());
2431 assert(_restart_addr == NULL, "Expected pre-condition");
2432 verification_mark_bm()->iterate(&markFromRootsClosure);
2433 while (_restart_addr != NULL) {
2434 // Deal with stack overflow: by restarting at the indicated
2435 // address.
2436 HeapWord* ra = _restart_addr;
2437 markFromRootsClosure.reset(ra);
2438 _restart_addr = NULL;
2439 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2440 }
2441 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2442 verify_work_stacks_empty();
2443
2444 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm());
2445 ClassLoaderDataGraph::cld_do(&verify_cld_oops);
2446
2447 // Marking completed -- now verify that each bit marked in
2448 // verification_mark_bm() is also marked in markBitMap(); flag all
2449 // errors by printing corresponding objects.
2450 VerifyMarkedClosure vcl(markBitMap());
2451 verification_mark_bm()->iterate(&vcl);
2452 assert(!vcl.failed(), "Else verification above should not have succeeded");
2453 }
2454
2455 void ConcurrentMarkSweepGeneration::save_marks() {
2456 // delegate to CMS space
2457 cmsSpace()->save_marks();
2458 }
2459
2460 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2461 return cmsSpace()->no_allocs_since_save_marks();
2462 }
2463
2464 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2465 \
2466 void ConcurrentMarkSweepGeneration:: \
2467 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2468 cl->set_generation(this); \
2469 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2470 cl->reset_generation(); \
2471 save_marks(); \
2472 }
2473
2474 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2475
2476 void
2477 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2478 if (freelistLock()->owned_by_self()) {
2479 Generation::oop_iterate(cl);
2480 } else {
2481 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2482 Generation::oop_iterate(cl);
2483 }
2484 }
2485
2486 void
2487 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2488 if (freelistLock()->owned_by_self()) {
2489 Generation::object_iterate(cl);
2490 } else {
2491 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2492 Generation::object_iterate(cl);
2493 }
2494 }
2495
2496 void
2497 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2498 if (freelistLock()->owned_by_self()) {
2499 Generation::safe_object_iterate(cl);
2500 } else {
2501 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2502 Generation::safe_object_iterate(cl);
2503 }
2504 }
2505
2506 void
2507 ConcurrentMarkSweepGeneration::post_compact() {
2508 }
2509
2510 void
2511 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2512 // Fix the linear allocation blocks to look like free blocks.
2513
2514 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2515 // are not called when the heap is verified during universe initialization and
2516 // at vm shutdown.
2517 if (freelistLock()->owned_by_self()) {
2518 cmsSpace()->prepare_for_verify();
2519 } else {
2520 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2521 cmsSpace()->prepare_for_verify();
2522 }
2523 }
2524
2525 void
2526 ConcurrentMarkSweepGeneration::verify() {
2527 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2528 // are not called when the heap is verified during universe initialization and
2529 // at vm shutdown.
2530 if (freelistLock()->owned_by_self()) {
2531 cmsSpace()->verify();
2532 } else {
2533 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2534 cmsSpace()->verify();
2535 }
2536 }
2537
2538 void CMSCollector::verify() {
2539 _cmsGen->verify();
2540 }
2541
2542 #ifndef PRODUCT
2543 bool CMSCollector::overflow_list_is_empty() const {
2544 assert(_num_par_pushes >= 0, "Inconsistency");
2545 if (_overflow_list == NULL) {
2546 assert(_num_par_pushes == 0, "Inconsistency");
2547 }
2548 return _overflow_list == NULL;
2549 }
2550
2551 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2552 // merely consolidate assertion checks that appear to occur together frequently.
2553 void CMSCollector::verify_work_stacks_empty() const {
2554 assert(_markStack.isEmpty(), "Marking stack should be empty");
2555 assert(overflow_list_is_empty(), "Overflow list should be empty");
2556 }
2557
2558 void CMSCollector::verify_overflow_empty() const {
2559 assert(overflow_list_is_empty(), "Overflow list should be empty");
2560 assert(no_preserved_marks(), "No preserved marks");
2561 }
2562 #endif // PRODUCT
2563
2564 // Decide if we want to enable class unloading as part of the
2565 // ensuing concurrent GC cycle. We will collect and
2566 // unload classes if it's the case that:
2567 // (a) class unloading is enabled at the command line, and
2568 // (b) old gen is getting really full
2569 // NOTE: Provided there is no change in the state of the heap between
2570 // calls to this method, it should have idempotent results. Moreover,
2571 // its results should be monotonically increasing (i.e. going from 0 to 1,
2572 // but not 1 to 0) between successive calls between which the heap was
2573 // not collected. For the implementation below, it must thus rely on
2574 // the property that concurrent_cycles_since_last_unload()
2575 // will not decrease unless a collection cycle happened and that
2576 // _cmsGen->is_too_full() are
2577 // themselves also monotonic in that sense. See check_monotonicity()
2578 // below.
2579 void CMSCollector::update_should_unload_classes() {
2580 _should_unload_classes = false;
2581 if (CMSClassUnloadingEnabled) {
2582 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2583 CMSClassUnloadingMaxInterval)
2584 || _cmsGen->is_too_full();
2585 }
2586 }
2587
2588 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2589 bool res = should_concurrent_collect();
2590 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2591 return res;
2592 }
2593
2594 void CMSCollector::setup_cms_unloading_and_verification_state() {
2595 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2596 || VerifyBeforeExit;
2597 const int rso = GenCollectedHeap::SO_AllCodeCache;
2598
2599 // We set the proper root for this CMS cycle here.
2600 if (should_unload_classes()) { // Should unload classes this cycle
2601 remove_root_scanning_option(rso); // Shrink the root set appropriately
2602 set_verifying(should_verify); // Set verification state for this cycle
2603 return; // Nothing else needs to be done at this time
2604 }
2605
2606 // Not unloading classes this cycle
2607 assert(!should_unload_classes(), "Inconsistency!");
2608
2609 // If we are not unloading classes then add SO_AllCodeCache to root
2610 // scanning options.
2611 add_root_scanning_option(rso);
2612
2613 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2614 set_verifying(true);
2615 } else if (verifying() && !should_verify) {
2616 // We were verifying, but some verification flags got disabled.
2617 set_verifying(false);
2618 // Exclude symbols, strings and code cache elements from root scanning to
2619 // reduce IM and RM pauses.
2620 remove_root_scanning_option(rso);
2621 }
2622 }
2623
2624
2625 #ifndef PRODUCT
2626 HeapWord* CMSCollector::block_start(const void* p) const {
2627 const HeapWord* addr = (HeapWord*)p;
2628 if (_span.contains(p)) {
2629 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2630 return _cmsGen->cmsSpace()->block_start(p);
2631 }
2632 }
2633 return NULL;
2634 }
2635 #endif
2636
2637 HeapWord*
2638 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2639 bool tlab,
2640 bool parallel) {
2641 CMSSynchronousYieldRequest yr;
2642 assert(!tlab, "Can't deal with TLAB allocation");
2643 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2644 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2645 if (GCExpandToAllocateDelayMillis > 0) {
2646 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2647 }
2648 return have_lock_and_allocate(word_size, tlab);
2649 }
2650
2651 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2652 size_t bytes,
2653 size_t expand_bytes,
2654 CMSExpansionCause::Cause cause)
2655 {
2656
2657 bool success = expand(bytes, expand_bytes);
2658
2659 // remember why we expanded; this information is used
2660 // by shouldConcurrentCollect() when making decisions on whether to start
2661 // a new CMS cycle.
2662 if (success) {
2663 set_expansion_cause(cause);
2664 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause));
2665 }
2666 }
2667
2668 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2669 HeapWord* res = NULL;
2670 MutexLocker x(ParGCRareEvent_lock);
2671 while (true) {
2672 // Expansion by some other thread might make alloc OK now:
2673 res = ps->lab.alloc(word_sz);
2674 if (res != NULL) return res;
2675 // If there's not enough expansion space available, give up.
2676 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2677 return NULL;
2678 }
2679 // Otherwise, we try expansion.
2680 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2681 // Now go around the loop and try alloc again;
2682 // A competing par_promote might beat us to the expansion space,
2683 // so we may go around the loop again if promotion fails again.
2684 if (GCExpandToAllocateDelayMillis > 0) {
2685 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2686 }
2687 }
2688 }
2689
2690
2691 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2692 PromotionInfo* promo) {
2693 MutexLocker x(ParGCRareEvent_lock);
2694 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2695 while (true) {
2696 // Expansion by some other thread might make alloc OK now:
2697 if (promo->ensure_spooling_space()) {
2698 assert(promo->has_spooling_space(),
2699 "Post-condition of successful ensure_spooling_space()");
2700 return true;
2701 }
2702 // If there's not enough expansion space available, give up.
2703 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2704 return false;
2705 }
2706 // Otherwise, we try expansion.
2707 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2708 // Now go around the loop and try alloc again;
2709 // A competing allocation might beat us to the expansion space,
2710 // so we may go around the loop again if allocation fails again.
2711 if (GCExpandToAllocateDelayMillis > 0) {
2712 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2713 }
2714 }
2715 }
2716
2717 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2718 // Only shrink if a compaction was done so that all the free space
2719 // in the generation is in a contiguous block at the end.
2720 if (did_compact()) {
2721 CardGeneration::shrink(bytes);
2722 }
2723 }
2724
2725 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2726 assert_locked_or_safepoint(Heap_lock);
2727 }
2728
2729 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2730 assert_locked_or_safepoint(Heap_lock);
2731 assert_lock_strong(freelistLock());
2732 log_trace(gc)("Shrinking of CMS not yet implemented");
2733 return;
2734 }
2735
2736
2737 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2738 // phases.
2739 class CMSPhaseAccounting: public StackObj {
2740 public:
2741 CMSPhaseAccounting(CMSCollector *collector,
2742 const char *title);
2743 ~CMSPhaseAccounting();
2744
2745 private:
2746 CMSCollector *_collector;
2747 const char *_title;
2748 GCTraceConcTime(Info, gc) _trace_time;
2749
2750 public:
2751 // Not MT-safe; so do not pass around these StackObj's
2752 // where they may be accessed by other threads.
2753 double wallclock_millis() {
2754 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time());
2755 }
2756 };
2757
2758 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2759 const char *title) :
2760 _collector(collector), _title(title), _trace_time(title) {
2761
2762 _collector->resetYields();
2763 _collector->resetTimer();
2764 _collector->startTimer();
2765 _collector->gc_timer_cm()->register_gc_concurrent_start(title);
2766 }
2767
2768 CMSPhaseAccounting::~CMSPhaseAccounting() {
2769 _collector->gc_timer_cm()->register_gc_concurrent_end();
2770 _collector->stopTimer();
2771 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks()));
2772 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields());
2773 }
2774
2775 // CMS work
2776
2777 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2778 class CMSParMarkTask : public AbstractGangTask {
2779 protected:
2780 CMSCollector* _collector;
2781 uint _n_workers;
2782 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2783 AbstractGangTask(name),
2784 _collector(collector),
2785 _n_workers(n_workers) {}
2786 // Work method in support of parallel rescan ... of young gen spaces
2787 void do_young_space_rescan(OopsInGenClosure* cl,
2788 ContiguousSpace* space,
2789 HeapWord** chunk_array, size_t chunk_top);
2790 void work_on_young_gen_roots(OopsInGenClosure* cl);
2791 };
2792
2793 // Parallel initial mark task
2794 class CMSParInitialMarkTask: public CMSParMarkTask {
2795 StrongRootsScope* _strong_roots_scope;
2796 public:
2797 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) :
2798 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers),
2799 _strong_roots_scope(strong_roots_scope) {}
2800 void work(uint worker_id);
2801 };
2802
2803 // Checkpoint the roots into this generation from outside
2804 // this generation. [Note this initial checkpoint need only
2805 // be approximate -- we'll do a catch up phase subsequently.]
2806 void CMSCollector::checkpointRootsInitial() {
2807 assert(_collectorState == InitialMarking, "Wrong collector state");
2808 check_correct_thread_executing();
2809 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
2810
2811 save_heap_summary();
2812 report_heap_summary(GCWhen::BeforeGC);
2813
2814 ReferenceProcessor* rp = ref_processor();
2815 assert(_restart_addr == NULL, "Control point invariant");
2816 {
2817 // acquire locks for subsequent manipulations
2818 MutexLockerEx x(bitMapLock(),
2819 Mutex::_no_safepoint_check_flag);
2820 checkpointRootsInitialWork();
2821 // enable ("weak") refs discovery
2822 rp->enable_discovery();
2823 _collectorState = Marking;
2824 }
2825 }
2826
2827 void CMSCollector::checkpointRootsInitialWork() {
2828 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2829 assert(_collectorState == InitialMarking, "just checking");
2830
2831 // Already have locks.
2832 assert_lock_strong(bitMapLock());
2833 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2834
2835 // Setup the verification and class unloading state for this
2836 // CMS collection cycle.
2837 setup_cms_unloading_and_verification_state();
2838
2839 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm);
2840
2841 // Reset all the PLAB chunk arrays if necessary.
2842 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2843 reset_survivor_plab_arrays();
2844 }
2845
2846 ResourceMark rm;
2847 HandleMark hm;
2848
2849 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2850 CMSHeap* heap = CMSHeap::heap();
2851
2852 verify_work_stacks_empty();
2853 verify_overflow_empty();
2854
2855 heap->ensure_parsability(false); // fill TLABs, but no need to retire them
2856 // Update the saved marks which may affect the root scans.
2857 heap->save_marks();
2858
2859 // weak reference processing has not started yet.
2860 ref_processor()->set_enqueuing_is_done(false);
2861
2862 // Need to remember all newly created CLDs,
2863 // so that we can guarantee that the remark finds them.
2864 ClassLoaderDataGraph::remember_new_clds(true);
2865
2866 // Whenever a CLD is found, it will be claimed before proceeding to mark
2867 // the klasses. The claimed marks need to be cleared before marking starts.
2868 ClassLoaderDataGraph::clear_claimed_marks();
2869
2870 print_eden_and_survivor_chunk_arrays();
2871
2872 {
2873 #if COMPILER2_OR_JVMCI
2874 DerivedPointerTableDeactivate dpt_deact;
2875 #endif
2876 if (CMSParallelInitialMarkEnabled) {
2877 // The parallel version.
2878 WorkGang* workers = heap->workers();
2879 assert(workers != NULL, "Need parallel worker threads.");
2880 uint n_workers = workers->active_workers();
2881
2882 StrongRootsScope srs(n_workers);
2883
2884 CMSParInitialMarkTask tsk(this, &srs, n_workers);
2885 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
2886 // If the total workers is greater than 1, then multiple workers
2887 // may be used at some time and the initialization has been set
2888 // such that the single threaded path cannot be used.
2889 if (workers->total_workers() > 1) {
2890 workers->run_task(&tsk);
2891 } else {
2892 tsk.work(0);
2893 }
2894 } else {
2895 // The serial version.
2896 CLDToOopClosure cld_closure(¬Older, true);
2897 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2898
2899 StrongRootsScope srs(1);
2900
2901 heap->cms_process_roots(&srs,
2902 true, // young gen as roots
2903 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2904 should_unload_classes(),
2905 ¬Older,
2906 &cld_closure);
2907 }
2908 }
2909
2910 // Clear mod-union table; it will be dirtied in the prologue of
2911 // CMS generation per each young generation collection.
2912
2913 assert(_modUnionTable.isAllClear(),
2914 "Was cleared in most recent final checkpoint phase"
2915 " or no bits are set in the gc_prologue before the start of the next "
2916 "subsequent marking phase.");
2917
2918 assert(_ct->cld_rem_set()->mod_union_is_clear(), "Must be");
2919
2920 // Save the end of the used_region of the constituent generations
2921 // to be used to limit the extent of sweep in each generation.
2922 save_sweep_limits();
2923 verify_overflow_empty();
2924 }
2925
2926 bool CMSCollector::markFromRoots() {
2927 // we might be tempted to assert that:
2928 // assert(!SafepointSynchronize::is_at_safepoint(),
2929 // "inconsistent argument?");
2930 // However that wouldn't be right, because it's possible that
2931 // a safepoint is indeed in progress as a young generation
2932 // stop-the-world GC happens even as we mark in this generation.
2933 assert(_collectorState == Marking, "inconsistent state?");
2934 check_correct_thread_executing();
2935 verify_overflow_empty();
2936
2937 // Weak ref discovery note: We may be discovering weak
2938 // refs in this generation concurrent (but interleaved) with
2939 // weak ref discovery by the young generation collector.
2940
2941 CMSTokenSyncWithLocks ts(true, bitMapLock());
2942 GCTraceCPUTime tcpu;
2943 CMSPhaseAccounting pa(this, "Concurrent Mark");
2944 bool res = markFromRootsWork();
2945 if (res) {
2946 _collectorState = Precleaning;
2947 } else { // We failed and a foreground collection wants to take over
2948 assert(_foregroundGCIsActive, "internal state inconsistency");
2949 assert(_restart_addr == NULL, "foreground will restart from scratch");
2950 log_debug(gc)("bailing out to foreground collection");
2951 }
2952 verify_overflow_empty();
2953 return res;
2954 }
2955
2956 bool CMSCollector::markFromRootsWork() {
2957 // iterate over marked bits in bit map, doing a full scan and mark
2958 // from these roots using the following algorithm:
2959 // . if oop is to the right of the current scan pointer,
2960 // mark corresponding bit (we'll process it later)
2961 // . else (oop is to left of current scan pointer)
2962 // push oop on marking stack
2963 // . drain the marking stack
2964
2965 // Note that when we do a marking step we need to hold the
2966 // bit map lock -- recall that direct allocation (by mutators)
2967 // and promotion (by the young generation collector) is also
2968 // marking the bit map. [the so-called allocate live policy.]
2969 // Because the implementation of bit map marking is not
2970 // robust wrt simultaneous marking of bits in the same word,
2971 // we need to make sure that there is no such interference
2972 // between concurrent such updates.
2973
2974 // already have locks
2975 assert_lock_strong(bitMapLock());
2976
2977 verify_work_stacks_empty();
2978 verify_overflow_empty();
2979 bool result = false;
2980 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
2981 result = do_marking_mt();
2982 } else {
2983 result = do_marking_st();
2984 }
2985 return result;
2986 }
2987
2988 // Forward decl
2989 class CMSConcMarkingTask;
2990
2991 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
2992 CMSCollector* _collector;
2993 CMSConcMarkingTask* _task;
2994 public:
2995 virtual void yield();
2996
2997 // "n_threads" is the number of threads to be terminated.
2998 // "queue_set" is a set of work queues of other threads.
2999 // "collector" is the CMS collector associated with this task terminator.
3000 // "yield" indicates whether we need the gang as a whole to yield.
3001 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3002 ParallelTaskTerminator(n_threads, queue_set),
3003 _collector(collector) { }
3004
3005 void set_task(CMSConcMarkingTask* task) {
3006 _task = task;
3007 }
3008 };
3009
3010 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3011 CMSConcMarkingTask* _task;
3012 public:
3013 bool should_exit_termination();
3014 void set_task(CMSConcMarkingTask* task) {
3015 _task = task;
3016 }
3017 };
3018
3019 // MT Concurrent Marking Task
3020 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3021 CMSCollector* _collector;
3022 uint _n_workers; // requested/desired # workers
3023 bool _result;
3024 CompactibleFreeListSpace* _cms_space;
3025 char _pad_front[64]; // padding to ...
3026 HeapWord* volatile _global_finger; // ... avoid sharing cache line
3027 char _pad_back[64];
3028 HeapWord* _restart_addr;
3029
3030 // Exposed here for yielding support
3031 Mutex* const _bit_map_lock;
3032
3033 // The per thread work queues, available here for stealing
3034 OopTaskQueueSet* _task_queues;
3035
3036 // Termination (and yielding) support
3037 CMSConcMarkingTerminator _term;
3038 CMSConcMarkingTerminatorTerminator _term_term;
3039
3040 public:
3041 CMSConcMarkingTask(CMSCollector* collector,
3042 CompactibleFreeListSpace* cms_space,
3043 YieldingFlexibleWorkGang* workers,
3044 OopTaskQueueSet* task_queues):
3045 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3046 _collector(collector),
3047 _cms_space(cms_space),
3048 _n_workers(0), _result(true),
3049 _task_queues(task_queues),
3050 _term(_n_workers, task_queues, _collector),
3051 _bit_map_lock(collector->bitMapLock())
3052 {
3053 _requested_size = _n_workers;
3054 _term.set_task(this);
3055 _term_term.set_task(this);
3056 _restart_addr = _global_finger = _cms_space->bottom();
3057 }
3058
3059
3060 OopTaskQueueSet* task_queues() { return _task_queues; }
3061
3062 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3063
3064 HeapWord* volatile* global_finger_addr() { return &_global_finger; }
3065
3066 CMSConcMarkingTerminator* terminator() { return &_term; }
3067
3068 virtual void set_for_termination(uint active_workers) {
3069 terminator()->reset_for_reuse(active_workers);
3070 }
3071
3072 void work(uint worker_id);
3073 bool should_yield() {
3074 return ConcurrentMarkSweepThread::should_yield()
3075 && !_collector->foregroundGCIsActive();
3076 }
3077
3078 virtual void coordinator_yield(); // stuff done by coordinator
3079 bool result() { return _result; }
3080
3081 void reset(HeapWord* ra) {
3082 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3083 _restart_addr = _global_finger = ra;
3084 _term.reset_for_reuse();
3085 }
3086
3087 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3088 OopTaskQueue* work_q);
3089
3090 private:
3091 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3092 void do_work_steal(int i);
3093 void bump_global_finger(HeapWord* f);
3094 };
3095
3096 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3097 assert(_task != NULL, "Error");
3098 return _task->yielding();
3099 // Note that we do not need the disjunct || _task->should_yield() above
3100 // because we want terminating threads to yield only if the task
3101 // is already in the midst of yielding, which happens only after at least one
3102 // thread has yielded.
3103 }
3104
3105 void CMSConcMarkingTerminator::yield() {
3106 if (_task->should_yield()) {
3107 _task->yield();
3108 } else {
3109 ParallelTaskTerminator::yield();
3110 }
3111 }
3112
3113 ////////////////////////////////////////////////////////////////
3114 // Concurrent Marking Algorithm Sketch
3115 ////////////////////////////////////////////////////////////////
3116 // Until all tasks exhausted (both spaces):
3117 // -- claim next available chunk
3118 // -- bump global finger via CAS
3119 // -- find first object that starts in this chunk
3120 // and start scanning bitmap from that position
3121 // -- scan marked objects for oops
3122 // -- CAS-mark target, and if successful:
3123 // . if target oop is above global finger (volatile read)
3124 // nothing to do
3125 // . if target oop is in chunk and above local finger
3126 // then nothing to do
3127 // . else push on work-queue
3128 // -- Deal with possible overflow issues:
3129 // . local work-queue overflow causes stuff to be pushed on
3130 // global (common) overflow queue
3131 // . always first empty local work queue
3132 // . then get a batch of oops from global work queue if any
3133 // . then do work stealing
3134 // -- When all tasks claimed (both spaces)
3135 // and local work queue empty,
3136 // then in a loop do:
3137 // . check global overflow stack; steal a batch of oops and trace
3138 // . try to steal from other threads oif GOS is empty
3139 // . if neither is available, offer termination
3140 // -- Terminate and return result
3141 //
3142 void CMSConcMarkingTask::work(uint worker_id) {
3143 elapsedTimer _timer;
3144 ResourceMark rm;
3145 HandleMark hm;
3146
3147 DEBUG_ONLY(_collector->verify_overflow_empty();)
3148
3149 // Before we begin work, our work queue should be empty
3150 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3151 // Scan the bitmap covering _cms_space, tracing through grey objects.
3152 _timer.start();
3153 do_scan_and_mark(worker_id, _cms_space);
3154 _timer.stop();
3155 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3156
3157 // ... do work stealing
3158 _timer.reset();
3159 _timer.start();
3160 do_work_steal(worker_id);
3161 _timer.stop();
3162 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3163 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3164 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3165 // Note that under the current task protocol, the
3166 // following assertion is true even of the spaces
3167 // expanded since the completion of the concurrent
3168 // marking. XXX This will likely change under a strict
3169 // ABORT semantics.
3170 // After perm removal the comparison was changed to
3171 // greater than or equal to from strictly greater than.
3172 // Before perm removal the highest address sweep would
3173 // have been at the end of perm gen but now is at the
3174 // end of the tenured gen.
3175 assert(_global_finger >= _cms_space->end(),
3176 "All tasks have been completed");
3177 DEBUG_ONLY(_collector->verify_overflow_empty();)
3178 }
3179
3180 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3181 HeapWord* read = _global_finger;
3182 HeapWord* cur = read;
3183 while (f > read) {
3184 cur = read;
3185 read = Atomic::cmpxchg(f, &_global_finger, cur);
3186 if (cur == read) {
3187 // our cas succeeded
3188 assert(_global_finger >= f, "protocol consistency");
3189 break;
3190 }
3191 }
3192 }
3193
3194 // This is really inefficient, and should be redone by
3195 // using (not yet available) block-read and -write interfaces to the
3196 // stack and the work_queue. XXX FIX ME !!!
3197 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3198 OopTaskQueue* work_q) {
3199 // Fast lock-free check
3200 if (ovflw_stk->length() == 0) {
3201 return false;
3202 }
3203 assert(work_q->size() == 0, "Shouldn't steal");
3204 MutexLockerEx ml(ovflw_stk->par_lock(),
3205 Mutex::_no_safepoint_check_flag);
3206 // Grab up to 1/4 the size of the work queue
3207 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3208 (size_t)ParGCDesiredObjsFromOverflowList);
3209 num = MIN2(num, ovflw_stk->length());
3210 for (int i = (int) num; i > 0; i--) {
3211 oop cur = ovflw_stk->pop();
3212 assert(cur != NULL, "Counted wrong?");
3213 work_q->push(cur);
3214 }
3215 return num > 0;
3216 }
3217
3218 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3219 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3220 int n_tasks = pst->n_tasks();
3221 // We allow that there may be no tasks to do here because
3222 // we are restarting after a stack overflow.
3223 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3224 uint nth_task = 0;
3225
3226 HeapWord* aligned_start = sp->bottom();
3227 if (sp->used_region().contains(_restart_addr)) {
3228 // Align down to a card boundary for the start of 0th task
3229 // for this space.
3230 aligned_start = align_down(_restart_addr, CardTable::card_size);
3231 }
3232
3233 size_t chunk_size = sp->marking_task_size();
3234 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3235 // Having claimed the nth task in this space,
3236 // compute the chunk that it corresponds to:
3237 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3238 aligned_start + (nth_task+1)*chunk_size);
3239 // Try and bump the global finger via a CAS;
3240 // note that we need to do the global finger bump
3241 // _before_ taking the intersection below, because
3242 // the task corresponding to that region will be
3243 // deemed done even if the used_region() expands
3244 // because of allocation -- as it almost certainly will
3245 // during start-up while the threads yield in the
3246 // closure below.
3247 HeapWord* finger = span.end();
3248 bump_global_finger(finger); // atomically
3249 // There are null tasks here corresponding to chunks
3250 // beyond the "top" address of the space.
3251 span = span.intersection(sp->used_region());
3252 if (!span.is_empty()) { // Non-null task
3253 HeapWord* prev_obj;
3254 assert(!span.contains(_restart_addr) || nth_task == 0,
3255 "Inconsistency");
3256 if (nth_task == 0) {
3257 // For the 0th task, we'll not need to compute a block_start.
3258 if (span.contains(_restart_addr)) {
3259 // In the case of a restart because of stack overflow,
3260 // we might additionally skip a chunk prefix.
3261 prev_obj = _restart_addr;
3262 } else {
3263 prev_obj = span.start();
3264 }
3265 } else {
3266 // We want to skip the first object because
3267 // the protocol is to scan any object in its entirety
3268 // that _starts_ in this span; a fortiori, any
3269 // object starting in an earlier span is scanned
3270 // as part of an earlier claimed task.
3271 // Below we use the "careful" version of block_start
3272 // so we do not try to navigate uninitialized objects.
3273 prev_obj = sp->block_start_careful(span.start());
3274 // Below we use a variant of block_size that uses the
3275 // Printezis bits to avoid waiting for allocated
3276 // objects to become initialized/parsable.
3277 while (prev_obj < span.start()) {
3278 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3279 if (sz > 0) {
3280 prev_obj += sz;
3281 } else {
3282 // In this case we may end up doing a bit of redundant
3283 // scanning, but that appears unavoidable, short of
3284 // locking the free list locks; see bug 6324141.
3285 break;
3286 }
3287 }
3288 }
3289 if (prev_obj < span.end()) {
3290 MemRegion my_span = MemRegion(prev_obj, span.end());
3291 // Do the marking work within a non-empty span --
3292 // the last argument to the constructor indicates whether the
3293 // iteration should be incremental with periodic yields.
3294 ParMarkFromRootsClosure cl(this, _collector, my_span,
3295 &_collector->_markBitMap,
3296 work_queue(i),
3297 &_collector->_markStack);
3298 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3299 } // else nothing to do for this task
3300 } // else nothing to do for this task
3301 }
3302 // We'd be tempted to assert here that since there are no
3303 // more tasks left to claim in this space, the global_finger
3304 // must exceed space->top() and a fortiori space->end(). However,
3305 // that would not quite be correct because the bumping of
3306 // global_finger occurs strictly after the claiming of a task,
3307 // so by the time we reach here the global finger may not yet
3308 // have been bumped up by the thread that claimed the last
3309 // task.
3310 pst->all_tasks_completed();
3311 }
3312
3313 class ParConcMarkingClosure: public MetadataAwareOopClosure {
3314 private:
3315 CMSCollector* _collector;
3316 CMSConcMarkingTask* _task;
3317 MemRegion _span;
3318 CMSBitMap* _bit_map;
3319 CMSMarkStack* _overflow_stack;
3320 OopTaskQueue* _work_queue;
3321 protected:
3322 DO_OOP_WORK_DEFN
3323 public:
3324 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3325 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3326 MetadataAwareOopClosure(collector->ref_processor()),
3327 _collector(collector),
3328 _task(task),
3329 _span(collector->_span),
3330 _work_queue(work_queue),
3331 _bit_map(bit_map),
3332 _overflow_stack(overflow_stack)
3333 { }
3334 virtual void do_oop(oop* p);
3335 virtual void do_oop(narrowOop* p);
3336
3337 void trim_queue(size_t max);
3338 void handle_stack_overflow(HeapWord* lost);
3339 void do_yield_check() {
3340 if (_task->should_yield()) {
3341 _task->yield();
3342 }
3343 }
3344 };
3345
3346 DO_OOP_WORK_IMPL(ParConcMarkingClosure)
3347
3348 // Grey object scanning during work stealing phase --
3349 // the salient assumption here is that any references
3350 // that are in these stolen objects being scanned must
3351 // already have been initialized (else they would not have
3352 // been published), so we do not need to check for
3353 // uninitialized objects before pushing here.
3354 void ParConcMarkingClosure::do_oop(oop obj) {
3355 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
3356 HeapWord* addr = (HeapWord*)obj;
3357 // Check if oop points into the CMS generation
3358 // and is not marked
3359 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3360 // a white object ...
3361 // If we manage to "claim" the object, by being the
3362 // first thread to mark it, then we push it on our
3363 // marking stack
3364 if (_bit_map->par_mark(addr)) { // ... now grey
3365 // push on work queue (grey set)
3366 bool simulate_overflow = false;
3367 NOT_PRODUCT(
3368 if (CMSMarkStackOverflowALot &&
3369 _collector->simulate_overflow()) {
3370 // simulate a stack overflow
3371 simulate_overflow = true;
3372 }
3373 )
3374 if (simulate_overflow ||
3375 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3376 // stack overflow
3377 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
3378 // We cannot assert that the overflow stack is full because
3379 // it may have been emptied since.
3380 assert(simulate_overflow ||
3381 _work_queue->size() == _work_queue->max_elems(),
3382 "Else push should have succeeded");
3383 handle_stack_overflow(addr);
3384 }
3385 } // Else, some other thread got there first
3386 do_yield_check();
3387 }
3388 }
3389
3390 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); }
3391 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); }
3392
3393 void ParConcMarkingClosure::trim_queue(size_t max) {
3394 while (_work_queue->size() > max) {
3395 oop new_oop;
3396 if (_work_queue->pop_local(new_oop)) {
3397 assert(oopDesc::is_oop(new_oop), "Should be an oop");
3398 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3399 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3400 new_oop->oop_iterate(this); // do_oop() above
3401 do_yield_check();
3402 }
3403 }
3404 }
3405
3406 // Upon stack overflow, we discard (part of) the stack,
3407 // remembering the least address amongst those discarded
3408 // in CMSCollector's _restart_address.
3409 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3410 // We need to do this under a mutex to prevent other
3411 // workers from interfering with the work done below.
3412 MutexLockerEx ml(_overflow_stack->par_lock(),
3413 Mutex::_no_safepoint_check_flag);
3414 // Remember the least grey address discarded
3415 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3416 _collector->lower_restart_addr(ra);
3417 _overflow_stack->reset(); // discard stack contents
3418 _overflow_stack->expand(); // expand the stack if possible
3419 }
3420
3421
3422 void CMSConcMarkingTask::do_work_steal(int i) {
3423 OopTaskQueue* work_q = work_queue(i);
3424 oop obj_to_scan;
3425 CMSBitMap* bm = &(_collector->_markBitMap);
3426 CMSMarkStack* ovflw = &(_collector->_markStack);
3427 int* seed = _collector->hash_seed(i);
3428 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3429 while (true) {
3430 cl.trim_queue(0);
3431 assert(work_q->size() == 0, "Should have been emptied above");
3432 if (get_work_from_overflow_stack(ovflw, work_q)) {
3433 // Can't assert below because the work obtained from the
3434 // overflow stack may already have been stolen from us.
3435 // assert(work_q->size() > 0, "Work from overflow stack");
3436 continue;
3437 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3438 assert(oopDesc::is_oop(obj_to_scan), "Should be an oop");
3439 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3440 obj_to_scan->oop_iterate(&cl);
3441 } else if (terminator()->offer_termination(&_term_term)) {
3442 assert(work_q->size() == 0, "Impossible!");
3443 break;
3444 } else if (yielding() || should_yield()) {
3445 yield();
3446 }
3447 }
3448 }
3449
3450 // This is run by the CMS (coordinator) thread.
3451 void CMSConcMarkingTask::coordinator_yield() {
3452 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3453 "CMS thread should hold CMS token");
3454 // First give up the locks, then yield, then re-lock
3455 // We should probably use a constructor/destructor idiom to
3456 // do this unlock/lock or modify the MutexUnlocker class to
3457 // serve our purpose. XXX
3458 assert_lock_strong(_bit_map_lock);
3459 _bit_map_lock->unlock();
3460 ConcurrentMarkSweepThread::desynchronize(true);
3461 _collector->stopTimer();
3462 _collector->incrementYields();
3463
3464 // It is possible for whichever thread initiated the yield request
3465 // not to get a chance to wake up and take the bitmap lock between
3466 // this thread releasing it and reacquiring it. So, while the
3467 // should_yield() flag is on, let's sleep for a bit to give the
3468 // other thread a chance to wake up. The limit imposed on the number
3469 // of iterations is defensive, to avoid any unforseen circumstances
3470 // putting us into an infinite loop. Since it's always been this
3471 // (coordinator_yield()) method that was observed to cause the
3472 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3473 // which is by default non-zero. For the other seven methods that
3474 // also perform the yield operation, as are using a different
3475 // parameter (CMSYieldSleepCount) which is by default zero. This way we
3476 // can enable the sleeping for those methods too, if necessary.
3477 // See 6442774.
3478 //
3479 // We really need to reconsider the synchronization between the GC
3480 // thread and the yield-requesting threads in the future and we
3481 // should really use wait/notify, which is the recommended
3482 // way of doing this type of interaction. Additionally, we should
3483 // consolidate the eight methods that do the yield operation and they
3484 // are almost identical into one for better maintainability and
3485 // readability. See 6445193.
3486 //
3487 // Tony 2006.06.29
3488 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3489 ConcurrentMarkSweepThread::should_yield() &&
3490 !CMSCollector::foregroundGCIsActive(); ++i) {
3491 os::sleep(Thread::current(), 1, false);
3492 }
3493
3494 ConcurrentMarkSweepThread::synchronize(true);
3495 _bit_map_lock->lock_without_safepoint_check();
3496 _collector->startTimer();
3497 }
3498
3499 bool CMSCollector::do_marking_mt() {
3500 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3501 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3502 conc_workers()->active_workers(),
3503 Threads::number_of_non_daemon_threads());
3504 num_workers = conc_workers()->update_active_workers(num_workers);
3505 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers());
3506
3507 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
3508
3509 CMSConcMarkingTask tsk(this,
3510 cms_space,
3511 conc_workers(),
3512 task_queues());
3513
3514 // Since the actual number of workers we get may be different
3515 // from the number we requested above, do we need to do anything different
3516 // below? In particular, may be we need to subclass the SequantialSubTasksDone
3517 // class?? XXX
3518 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3519
3520 // Refs discovery is already non-atomic.
3521 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3522 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3523 conc_workers()->start_task(&tsk);
3524 while (tsk.yielded()) {
3525 tsk.coordinator_yield();
3526 conc_workers()->continue_task(&tsk);
3527 }
3528 // If the task was aborted, _restart_addr will be non-NULL
3529 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3530 while (_restart_addr != NULL) {
3531 // XXX For now we do not make use of ABORTED state and have not
3532 // yet implemented the right abort semantics (even in the original
3533 // single-threaded CMS case). That needs some more investigation
3534 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3535 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3536 // If _restart_addr is non-NULL, a marking stack overflow
3537 // occurred; we need to do a fresh marking iteration from the
3538 // indicated restart address.
3539 if (_foregroundGCIsActive) {
3540 // We may be running into repeated stack overflows, having
3541 // reached the limit of the stack size, while making very
3542 // slow forward progress. It may be best to bail out and
3543 // let the foreground collector do its job.
3544 // Clear _restart_addr, so that foreground GC
3545 // works from scratch. This avoids the headache of
3546 // a "rescan" which would otherwise be needed because
3547 // of the dirty mod union table & card table.
3548 _restart_addr = NULL;
3549 return false;
3550 }
3551 // Adjust the task to restart from _restart_addr
3552 tsk.reset(_restart_addr);
3553 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3554 _restart_addr);
3555 _restart_addr = NULL;
3556 // Get the workers going again
3557 conc_workers()->start_task(&tsk);
3558 while (tsk.yielded()) {
3559 tsk.coordinator_yield();
3560 conc_workers()->continue_task(&tsk);
3561 }
3562 }
3563 assert(tsk.completed(), "Inconsistency");
3564 assert(tsk.result() == true, "Inconsistency");
3565 return true;
3566 }
3567
3568 bool CMSCollector::do_marking_st() {
3569 ResourceMark rm;
3570 HandleMark hm;
3571
3572 // Temporarily make refs discovery single threaded (non-MT)
3573 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3574 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3575 &_markStack, CMSYield);
3576 // the last argument to iterate indicates whether the iteration
3577 // should be incremental with periodic yields.
3578 _markBitMap.iterate(&markFromRootsClosure);
3579 // If _restart_addr is non-NULL, a marking stack overflow
3580 // occurred; we need to do a fresh iteration from the
3581 // indicated restart address.
3582 while (_restart_addr != NULL) {
3583 if (_foregroundGCIsActive) {
3584 // We may be running into repeated stack overflows, having
3585 // reached the limit of the stack size, while making very
3586 // slow forward progress. It may be best to bail out and
3587 // let the foreground collector do its job.
3588 // Clear _restart_addr, so that foreground GC
3589 // works from scratch. This avoids the headache of
3590 // a "rescan" which would otherwise be needed because
3591 // of the dirty mod union table & card table.
3592 _restart_addr = NULL;
3593 return false; // indicating failure to complete marking
3594 }
3595 // Deal with stack overflow:
3596 // we restart marking from _restart_addr
3597 HeapWord* ra = _restart_addr;
3598 markFromRootsClosure.reset(ra);
3599 _restart_addr = NULL;
3600 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3601 }
3602 return true;
3603 }
3604
3605 void CMSCollector::preclean() {
3606 check_correct_thread_executing();
3607 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3608 verify_work_stacks_empty();
3609 verify_overflow_empty();
3610 _abort_preclean = false;
3611 if (CMSPrecleaningEnabled) {
3612 if (!CMSEdenChunksRecordAlways) {
3613 _eden_chunk_index = 0;
3614 }
3615 size_t used = get_eden_used();
3616 size_t capacity = get_eden_capacity();
3617 // Don't start sampling unless we will get sufficiently
3618 // many samples.
3619 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100)
3620 * CMSScheduleRemarkEdenPenetration)) {
3621 _start_sampling = true;
3622 } else {
3623 _start_sampling = false;
3624 }
3625 GCTraceCPUTime tcpu;
3626 CMSPhaseAccounting pa(this, "Concurrent Preclean");
3627 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3628 }
3629 CMSTokenSync x(true); // is cms thread
3630 if (CMSPrecleaningEnabled) {
3631 sample_eden();
3632 _collectorState = AbortablePreclean;
3633 } else {
3634 _collectorState = FinalMarking;
3635 }
3636 verify_work_stacks_empty();
3637 verify_overflow_empty();
3638 }
3639
3640 // Try and schedule the remark such that young gen
3641 // occupancy is CMSScheduleRemarkEdenPenetration %.
3642 void CMSCollector::abortable_preclean() {
3643 check_correct_thread_executing();
3644 assert(CMSPrecleaningEnabled, "Inconsistent control state");
3645 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3646
3647 // If Eden's current occupancy is below this threshold,
3648 // immediately schedule the remark; else preclean
3649 // past the next scavenge in an effort to
3650 // schedule the pause as described above. By choosing
3651 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3652 // we will never do an actual abortable preclean cycle.
3653 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3654 GCTraceCPUTime tcpu;
3655 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean");
3656 // We need more smarts in the abortable preclean
3657 // loop below to deal with cases where allocation
3658 // in young gen is very very slow, and our precleaning
3659 // is running a losing race against a horde of
3660 // mutators intent on flooding us with CMS updates
3661 // (dirty cards).
3662 // One, admittedly dumb, strategy is to give up
3663 // after a certain number of abortable precleaning loops
3664 // or after a certain maximum time. We want to make
3665 // this smarter in the next iteration.
3666 // XXX FIX ME!!! YSR
3667 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3668 while (!(should_abort_preclean() ||
3669 ConcurrentMarkSweepThread::cmst()->should_terminate())) {
3670 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3671 cumworkdone += workdone;
3672 loops++;
3673 // Voluntarily terminate abortable preclean phase if we have
3674 // been at it for too long.
3675 if ((CMSMaxAbortablePrecleanLoops != 0) &&
3676 loops >= CMSMaxAbortablePrecleanLoops) {
3677 log_debug(gc)(" CMS: abort preclean due to loops ");
3678 break;
3679 }
3680 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3681 log_debug(gc)(" CMS: abort preclean due to time ");
3682 break;
3683 }
3684 // If we are doing little work each iteration, we should
3685 // take a short break.
3686 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3687 // Sleep for some time, waiting for work to accumulate
3688 stopTimer();
3689 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3690 startTimer();
3691 waited++;
3692 }
3693 }
3694 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3695 loops, waited, cumworkdone);
3696 }
3697 CMSTokenSync x(true); // is cms thread
3698 if (_collectorState != Idling) {
3699 assert(_collectorState == AbortablePreclean,
3700 "Spontaneous state transition?");
3701 _collectorState = FinalMarking;
3702 } // Else, a foreground collection completed this CMS cycle.
3703 return;
3704 }
3705
3706 // Respond to an Eden sampling opportunity
3707 void CMSCollector::sample_eden() {
3708 // Make sure a young gc cannot sneak in between our
3709 // reading and recording of a sample.
3710 assert(Thread::current()->is_ConcurrentGC_thread(),
3711 "Only the cms thread may collect Eden samples");
3712 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3713 "Should collect samples while holding CMS token");
3714 if (!_start_sampling) {
3715 return;
3716 }
3717 // When CMSEdenChunksRecordAlways is true, the eden chunk array
3718 // is populated by the young generation.
3719 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3720 if (_eden_chunk_index < _eden_chunk_capacity) {
3721 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
3722 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3723 "Unexpected state of Eden");
3724 // We'd like to check that what we just sampled is an oop-start address;
3725 // however, we cannot do that here since the object may not yet have been
3726 // initialized. So we'll instead do the check when we _use_ this sample
3727 // later.
3728 if (_eden_chunk_index == 0 ||
3729 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3730 _eden_chunk_array[_eden_chunk_index-1])
3731 >= CMSSamplingGrain)) {
3732 _eden_chunk_index++; // commit sample
3733 }
3734 }
3735 }
3736 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3737 size_t used = get_eden_used();
3738 size_t capacity = get_eden_capacity();
3739 assert(used <= capacity, "Unexpected state of Eden");
3740 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3741 _abort_preclean = true;
3742 }
3743 }
3744 }
3745
3746
3747 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3748 assert(_collectorState == Precleaning ||
3749 _collectorState == AbortablePreclean, "incorrect state");
3750 ResourceMark rm;
3751 HandleMark hm;
3752
3753 // Precleaning is currently not MT but the reference processor
3754 // may be set for MT. Disable it temporarily here.
3755 ReferenceProcessor* rp = ref_processor();
3756 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3757
3758 // Do one pass of scrubbing the discovered reference lists
3759 // to remove any reference objects with strongly-reachable
3760 // referents.
3761 if (clean_refs) {
3762 CMSPrecleanRefsYieldClosure yield_cl(this);
3763 assert(rp->span().equals(_span), "Spans should be equal");
3764 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3765 &_markStack, true /* preclean */);
3766 CMSDrainMarkingStackClosure complete_trace(this,
3767 _span, &_markBitMap, &_markStack,
3768 &keep_alive, true /* preclean */);
3769
3770 // We don't want this step to interfere with a young
3771 // collection because we don't want to take CPU
3772 // or memory bandwidth away from the young GC threads
3773 // (which may be as many as there are CPUs).
3774 // Note that we don't need to protect ourselves from
3775 // interference with mutators because they can't
3776 // manipulate the discovered reference lists nor affect
3777 // the computed reachability of the referents, the
3778 // only properties manipulated by the precleaning
3779 // of these reference lists.
3780 stopTimer();
3781 CMSTokenSyncWithLocks x(true /* is cms thread */,
3782 bitMapLock());
3783 startTimer();
3784 sample_eden();
3785
3786 // The following will yield to allow foreground
3787 // collection to proceed promptly. XXX YSR:
3788 // The code in this method may need further
3789 // tweaking for better performance and some restructuring
3790 // for cleaner interfaces.
3791 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3792 rp->preclean_discovered_references(
3793 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3794 gc_timer);
3795 }
3796
3797 if (clean_survivor) { // preclean the active survivor space(s)
3798 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3799 &_markBitMap, &_modUnionTable,
3800 &_markStack, true /* precleaning phase */);
3801 stopTimer();
3802 CMSTokenSyncWithLocks ts(true /* is cms thread */,
3803 bitMapLock());
3804 startTimer();
3805 unsigned int before_count =
3806 CMSHeap::heap()->total_collections();
3807 SurvivorSpacePrecleanClosure
3808 sss_cl(this, _span, &_markBitMap, &_markStack,
3809 &pam_cl, before_count, CMSYield);
3810 _young_gen->from()->object_iterate_careful(&sss_cl);
3811 _young_gen->to()->object_iterate_careful(&sss_cl);
3812 }
3813 MarkRefsIntoAndScanClosure
3814 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3815 &_markStack, this, CMSYield,
3816 true /* precleaning phase */);
3817 // CAUTION: The following closure has persistent state that may need to
3818 // be reset upon a decrease in the sequence of addresses it
3819 // processes.
3820 ScanMarkedObjectsAgainCarefullyClosure
3821 smoac_cl(this, _span,
3822 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3823
3824 // Preclean dirty cards in ModUnionTable and CardTable using
3825 // appropriate convergence criterion;
3826 // repeat CMSPrecleanIter times unless we find that
3827 // we are losing.
3828 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3829 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3830 "Bad convergence multiplier");
3831 assert(CMSPrecleanThreshold >= 100,
3832 "Unreasonably low CMSPrecleanThreshold");
3833
3834 size_t numIter, cumNumCards, lastNumCards, curNumCards;
3835 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3836 numIter < CMSPrecleanIter;
3837 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3838 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
3839 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3840 // Either there are very few dirty cards, so re-mark
3841 // pause will be small anyway, or our pre-cleaning isn't
3842 // that much faster than the rate at which cards are being
3843 // dirtied, so we might as well stop and re-mark since
3844 // precleaning won't improve our re-mark time by much.
3845 if (curNumCards <= CMSPrecleanThreshold ||
3846 (numIter > 0 &&
3847 (curNumCards * CMSPrecleanDenominator >
3848 lastNumCards * CMSPrecleanNumerator))) {
3849 numIter++;
3850 cumNumCards += curNumCards;
3851 break;
3852 }
3853 }
3854
3855 preclean_cld(&mrias_cl, _cmsGen->freelistLock());
3856
3857 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
3858 cumNumCards += curNumCards;
3859 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
3860 curNumCards, cumNumCards, numIter);
3861 return cumNumCards; // as a measure of useful work done
3862 }
3863
3864 // PRECLEANING NOTES:
3865 // Precleaning involves:
3866 // . reading the bits of the modUnionTable and clearing the set bits.
3867 // . For the cards corresponding to the set bits, we scan the
3868 // objects on those cards. This means we need the free_list_lock
3869 // so that we can safely iterate over the CMS space when scanning
3870 // for oops.
3871 // . When we scan the objects, we'll be both reading and setting
3872 // marks in the marking bit map, so we'll need the marking bit map.
3873 // . For protecting _collector_state transitions, we take the CGC_lock.
3874 // Note that any races in the reading of of card table entries by the
3875 // CMS thread on the one hand and the clearing of those entries by the
3876 // VM thread or the setting of those entries by the mutator threads on the
3877 // other are quite benign. However, for efficiency it makes sense to keep
3878 // the VM thread from racing with the CMS thread while the latter is
3879 // dirty card info to the modUnionTable. We therefore also use the
3880 // CGC_lock to protect the reading of the card table and the mod union
3881 // table by the CM thread.
3882 // . We run concurrently with mutator updates, so scanning
3883 // needs to be done carefully -- we should not try to scan
3884 // potentially uninitialized objects.
3885 //
3886 // Locking strategy: While holding the CGC_lock, we scan over and
3887 // reset a maximal dirty range of the mod union / card tables, then lock
3888 // the free_list_lock and bitmap lock to do a full marking, then
3889 // release these locks; and repeat the cycle. This allows for a
3890 // certain amount of fairness in the sharing of these locks between
3891 // the CMS collector on the one hand, and the VM thread and the
3892 // mutators on the other.
3893
3894 // NOTE: preclean_mod_union_table() and preclean_card_table()
3895 // further below are largely identical; if you need to modify
3896 // one of these methods, please check the other method too.
3897
3898 size_t CMSCollector::preclean_mod_union_table(
3899 ConcurrentMarkSweepGeneration* old_gen,
3900 ScanMarkedObjectsAgainCarefullyClosure* cl) {
3901 verify_work_stacks_empty();
3902 verify_overflow_empty();
3903
3904 // strategy: starting with the first card, accumulate contiguous
3905 // ranges of dirty cards; clear these cards, then scan the region
3906 // covered by these cards.
3907
3908 // Since all of the MUT is committed ahead, we can just use
3909 // that, in case the generations expand while we are precleaning.
3910 // It might also be fine to just use the committed part of the
3911 // generation, but we might potentially miss cards when the
3912 // generation is rapidly expanding while we are in the midst
3913 // of precleaning.
3914 HeapWord* startAddr = old_gen->reserved().start();
3915 HeapWord* endAddr = old_gen->reserved().end();
3916
3917 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
3918
3919 size_t numDirtyCards, cumNumDirtyCards;
3920 HeapWord *nextAddr, *lastAddr;
3921 for (cumNumDirtyCards = numDirtyCards = 0,
3922 nextAddr = lastAddr = startAddr;
3923 nextAddr < endAddr;
3924 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
3925
3926 ResourceMark rm;
3927 HandleMark hm;
3928
3929 MemRegion dirtyRegion;
3930 {
3931 stopTimer();
3932 // Potential yield point
3933 CMSTokenSync ts(true);
3934 startTimer();
3935 sample_eden();
3936 // Get dirty region starting at nextOffset (inclusive),
3937 // simultaneously clearing it.
3938 dirtyRegion =
3939 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
3940 assert(dirtyRegion.start() >= nextAddr,
3941 "returned region inconsistent?");
3942 }
3943 // Remember where the next search should begin.
3944 // The returned region (if non-empty) is a right open interval,
3945 // so lastOffset is obtained from the right end of that
3946 // interval.
3947 lastAddr = dirtyRegion.end();
3948 // Should do something more transparent and less hacky XXX
3949 numDirtyCards =
3950 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
3951
3952 // We'll scan the cards in the dirty region (with periodic
3953 // yields for foreground GC as needed).
3954 if (!dirtyRegion.is_empty()) {
3955 assert(numDirtyCards > 0, "consistency check");
3956 HeapWord* stop_point = NULL;
3957 stopTimer();
3958 // Potential yield point
3959 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(),
3960 bitMapLock());
3961 startTimer();
3962 {
3963 verify_work_stacks_empty();
3964 verify_overflow_empty();
3965 sample_eden();
3966 stop_point =
3967 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
3968 }
3969 if (stop_point != NULL) {
3970 // The careful iteration stopped early either because it found an
3971 // uninitialized object, or because we were in the midst of an
3972 // "abortable preclean", which should now be aborted. Redirty
3973 // the bits corresponding to the partially-scanned or unscanned
3974 // cards. We'll either restart at the next block boundary or
3975 // abort the preclean.
3976 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
3977 "Should only be AbortablePreclean.");
3978 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
3979 if (should_abort_preclean()) {
3980 break; // out of preclean loop
3981 } else {
3982 // Compute the next address at which preclean should pick up;
3983 // might need bitMapLock in order to read P-bits.
3984 lastAddr = next_card_start_after_block(stop_point);
3985 }
3986 }
3987 } else {
3988 assert(lastAddr == endAddr, "consistency check");
3989 assert(numDirtyCards == 0, "consistency check");
3990 break;
3991 }
3992 }
3993 verify_work_stacks_empty();
3994 verify_overflow_empty();
3995 return cumNumDirtyCards;
3996 }
3997
3998 // NOTE: preclean_mod_union_table() above and preclean_card_table()
3999 // below are largely identical; if you need to modify
4000 // one of these methods, please check the other method too.
4001
4002 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen,
4003 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4004 // strategy: it's similar to precleamModUnionTable above, in that
4005 // we accumulate contiguous ranges of dirty cards, mark these cards
4006 // precleaned, then scan the region covered by these cards.
4007 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high());
4008 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low());
4009
4010 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
4011
4012 size_t numDirtyCards, cumNumDirtyCards;
4013 HeapWord *lastAddr, *nextAddr;
4014
4015 for (cumNumDirtyCards = numDirtyCards = 0,
4016 nextAddr = lastAddr = startAddr;
4017 nextAddr < endAddr;
4018 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4019
4020 ResourceMark rm;
4021 HandleMark hm;
4022
4023 MemRegion dirtyRegion;
4024 {
4025 // See comments in "Precleaning notes" above on why we
4026 // do this locking. XXX Could the locking overheads be
4027 // too high when dirty cards are sparse? [I don't think so.]
4028 stopTimer();
4029 CMSTokenSync x(true); // is cms thread
4030 startTimer();
4031 sample_eden();
4032 // Get and clear dirty region from card table
4033 dirtyRegion = _ct->dirty_card_range_after_reset(MemRegion(nextAddr, endAddr),
4034 true,
4035 CardTable::precleaned_card_val());
4036
4037 assert(dirtyRegion.start() >= nextAddr,
4038 "returned region inconsistent?");
4039 }
4040 lastAddr = dirtyRegion.end();
4041 numDirtyCards =
4042 dirtyRegion.word_size()/CardTable::card_size_in_words;
4043
4044 if (!dirtyRegion.is_empty()) {
4045 stopTimer();
4046 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock());
4047 startTimer();
4048 sample_eden();
4049 verify_work_stacks_empty();
4050 verify_overflow_empty();
4051 HeapWord* stop_point =
4052 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4053 if (stop_point != NULL) {
4054 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4055 "Should only be AbortablePreclean.");
4056 _ct->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4057 if (should_abort_preclean()) {
4058 break; // out of preclean loop
4059 } else {
4060 // Compute the next address at which preclean should pick up.
4061 lastAddr = next_card_start_after_block(stop_point);
4062 }
4063 }
4064 } else {
4065 break;
4066 }
4067 }
4068 verify_work_stacks_empty();
4069 verify_overflow_empty();
4070 return cumNumDirtyCards;
4071 }
4072
4073 class PrecleanCLDClosure : public CLDClosure {
4074 MetadataAwareOopsInGenClosure* _cm_closure;
4075 public:
4076 PrecleanCLDClosure(MetadataAwareOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {}
4077 void do_cld(ClassLoaderData* cld) {
4078 if (cld->has_accumulated_modified_oops()) {
4079 cld->clear_accumulated_modified_oops();
4080
4081 _cm_closure->do_cld(cld);
4082 }
4083 }
4084 };
4085
4086 // The freelist lock is needed to prevent asserts, is it really needed?
4087 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4088
4089 cl->set_freelistLock(freelistLock);
4090
4091 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4092
4093 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4094 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4095 PrecleanCLDClosure preclean_closure(cl);
4096 ClassLoaderDataGraph::cld_do(&preclean_closure);
4097
4098 verify_work_stacks_empty();
4099 verify_overflow_empty();
4100 }
4101
4102 void CMSCollector::checkpointRootsFinal() {
4103 assert(_collectorState == FinalMarking, "incorrect state transition?");
4104 check_correct_thread_executing();
4105 // world is stopped at this checkpoint
4106 assert(SafepointSynchronize::is_at_safepoint(),
4107 "world should be stopped");
4108 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
4109
4110 verify_work_stacks_empty();
4111 verify_overflow_empty();
4112
4113 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)",
4114 _young_gen->used() / K, _young_gen->capacity() / K);
4115 {
4116 if (CMSScavengeBeforeRemark) {
4117 CMSHeap* heap = CMSHeap::heap();
4118 // Temporarily set flag to false, GCH->do_collection will
4119 // expect it to be false and set to true
4120 FlagSetting fl(heap->_is_gc_active, false);
4121
4122 heap->do_collection(true, // full (i.e. force, see below)
4123 false, // !clear_all_soft_refs
4124 0, // size
4125 false, // is_tlab
4126 GenCollectedHeap::YoungGen // type
4127 );
4128 }
4129 FreelistLocker x(this);
4130 MutexLockerEx y(bitMapLock(),
4131 Mutex::_no_safepoint_check_flag);
4132 checkpointRootsFinalWork();
4133 }
4134 verify_work_stacks_empty();
4135 verify_overflow_empty();
4136 }
4137
4138 void CMSCollector::checkpointRootsFinalWork() {
4139 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm);
4140
4141 assert(haveFreelistLocks(), "must have free list locks");
4142 assert_lock_strong(bitMapLock());
4143
4144 ResourceMark rm;
4145 HandleMark hm;
4146
4147 CMSHeap* heap = CMSHeap::heap();
4148
4149 if (should_unload_classes()) {
4150 CodeCache::gc_prologue();
4151 }
4152 assert(haveFreelistLocks(), "must have free list locks");
4153 assert_lock_strong(bitMapLock());
4154
4155 // We might assume that we need not fill TLAB's when
4156 // CMSScavengeBeforeRemark is set, because we may have just done
4157 // a scavenge which would have filled all TLAB's -- and besides
4158 // Eden would be empty. This however may not always be the case --
4159 // for instance although we asked for a scavenge, it may not have
4160 // happened because of a JNI critical section. We probably need
4161 // a policy for deciding whether we can in that case wait until
4162 // the critical section releases and then do the remark following
4163 // the scavenge, and skip it here. In the absence of that policy,
4164 // or of an indication of whether the scavenge did indeed occur,
4165 // we cannot rely on TLAB's having been filled and must do
4166 // so here just in case a scavenge did not happen.
4167 heap->ensure_parsability(false); // fill TLAB's, but no need to retire them
4168 // Update the saved marks which may affect the root scans.
4169 heap->save_marks();
4170
4171 print_eden_and_survivor_chunk_arrays();
4172
4173 {
4174 #if COMPILER2_OR_JVMCI
4175 DerivedPointerTableDeactivate dpt_deact;
4176 #endif
4177
4178 // Note on the role of the mod union table:
4179 // Since the marker in "markFromRoots" marks concurrently with
4180 // mutators, it is possible for some reachable objects not to have been
4181 // scanned. For instance, an only reference to an object A was
4182 // placed in object B after the marker scanned B. Unless B is rescanned,
4183 // A would be collected. Such updates to references in marked objects
4184 // are detected via the mod union table which is the set of all cards
4185 // dirtied since the first checkpoint in this GC cycle and prior to
4186 // the most recent young generation GC, minus those cleaned up by the
4187 // concurrent precleaning.
4188 if (CMSParallelRemarkEnabled) {
4189 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm);
4190 do_remark_parallel();
4191 } else {
4192 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm);
4193 do_remark_non_parallel();
4194 }
4195 }
4196 verify_work_stacks_empty();
4197 verify_overflow_empty();
4198
4199 {
4200 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm);
4201 refProcessingWork();
4202 }
4203 verify_work_stacks_empty();
4204 verify_overflow_empty();
4205
4206 if (should_unload_classes()) {
4207 CodeCache::gc_epilogue();
4208 }
4209 JvmtiExport::gc_epilogue();
4210
4211 // If we encountered any (marking stack / work queue) overflow
4212 // events during the current CMS cycle, take appropriate
4213 // remedial measures, where possible, so as to try and avoid
4214 // recurrence of that condition.
4215 assert(_markStack.isEmpty(), "No grey objects");
4216 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4217 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4218 if (ser_ovflw > 0) {
4219 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")",
4220 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4221 _markStack.expand();
4222 _ser_pmc_remark_ovflw = 0;
4223 _ser_pmc_preclean_ovflw = 0;
4224 _ser_kac_preclean_ovflw = 0;
4225 _ser_kac_ovflw = 0;
4226 }
4227 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4228 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")",
4229 _par_pmc_remark_ovflw, _par_kac_ovflw);
4230 _par_pmc_remark_ovflw = 0;
4231 _par_kac_ovflw = 0;
4232 }
4233 if (_markStack._hit_limit > 0) {
4234 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")",
4235 _markStack._hit_limit);
4236 }
4237 if (_markStack._failed_double > 0) {
4238 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT,
4239 _markStack._failed_double, _markStack.capacity());
4240 }
4241 _markStack._hit_limit = 0;
4242 _markStack._failed_double = 0;
4243
4244 if ((VerifyAfterGC || VerifyDuringGC) &&
4245 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) {
4246 verify_after_remark();
4247 }
4248
4249 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4250
4251 // Change under the freelistLocks.
4252 _collectorState = Sweeping;
4253 // Call isAllClear() under bitMapLock
4254 assert(_modUnionTable.isAllClear(),
4255 "Should be clear by end of the final marking");
4256 assert(_ct->cld_rem_set()->mod_union_is_clear(),
4257 "Should be clear by end of the final marking");
4258 }
4259
4260 void CMSParInitialMarkTask::work(uint worker_id) {
4261 elapsedTimer _timer;
4262 ResourceMark rm;
4263 HandleMark hm;
4264
4265 // ---------- scan from roots --------------
4266 _timer.start();
4267 CMSHeap* heap = CMSHeap::heap();
4268 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4269
4270 // ---------- young gen roots --------------
4271 {
4272 work_on_young_gen_roots(&par_mri_cl);
4273 _timer.stop();
4274 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4275 }
4276
4277 // ---------- remaining roots --------------
4278 _timer.reset();
4279 _timer.start();
4280
4281 CLDToOopClosure cld_closure(&par_mri_cl, true);
4282
4283 heap->cms_process_roots(_strong_roots_scope,
4284 false, // yg was scanned above
4285 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4286 _collector->should_unload_classes(),
4287 &par_mri_cl,
4288 &cld_closure);
4289 assert(_collector->should_unload_classes()
4290 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4291 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4292 _timer.stop();
4293 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4294 }
4295
4296 // Parallel remark task
4297 class CMSParRemarkTask: public CMSParMarkTask {
4298 CompactibleFreeListSpace* _cms_space;
4299
4300 // The per-thread work queues, available here for stealing.
4301 OopTaskQueueSet* _task_queues;
4302 ParallelTaskTerminator _term;
4303 StrongRootsScope* _strong_roots_scope;
4304
4305 public:
4306 // A value of 0 passed to n_workers will cause the number of
4307 // workers to be taken from the active workers in the work gang.
4308 CMSParRemarkTask(CMSCollector* collector,
4309 CompactibleFreeListSpace* cms_space,
4310 uint n_workers, WorkGang* workers,
4311 OopTaskQueueSet* task_queues,
4312 StrongRootsScope* strong_roots_scope):
4313 CMSParMarkTask("Rescan roots and grey objects in parallel",
4314 collector, n_workers),
4315 _cms_space(cms_space),
4316 _task_queues(task_queues),
4317 _term(n_workers, task_queues),
4318 _strong_roots_scope(strong_roots_scope) { }
4319
4320 OopTaskQueueSet* task_queues() { return _task_queues; }
4321
4322 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4323
4324 ParallelTaskTerminator* terminator() { return &_term; }
4325 uint n_workers() { return _n_workers; }
4326
4327 void work(uint worker_id);
4328
4329 private:
4330 // ... of dirty cards in old space
4331 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4332 ParMarkRefsIntoAndScanClosure* cl);
4333
4334 // ... work stealing for the above
4335 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed);
4336 };
4337
4338 class RemarkCLDClosure : public CLDClosure {
4339 CLDToOopClosure _cm_closure;
4340 public:
4341 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure) {}
4342 void do_cld(ClassLoaderData* cld) {
4343 // Check if we have modified any oops in the CLD during the concurrent marking.
4344 if (cld->has_accumulated_modified_oops()) {
4345 cld->clear_accumulated_modified_oops();
4346
4347 // We could have transfered the current modified marks to the accumulated marks,
4348 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4349 } else if (cld->has_modified_oops()) {
4350 // Don't clear anything, this info is needed by the next young collection.
4351 } else {
4352 // No modified oops in the ClassLoaderData.
4353 return;
4354 }
4355
4356 // The klass has modified fields, need to scan the klass.
4357 _cm_closure.do_cld(cld);
4358 }
4359 };
4360
4361 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) {
4362 ParNewGeneration* young_gen = _collector->_young_gen;
4363 ContiguousSpace* eden_space = young_gen->eden();
4364 ContiguousSpace* from_space = young_gen->from();
4365 ContiguousSpace* to_space = young_gen->to();
4366
4367 HeapWord** eca = _collector->_eden_chunk_array;
4368 size_t ect = _collector->_eden_chunk_index;
4369 HeapWord** sca = _collector->_survivor_chunk_array;
4370 size_t sct = _collector->_survivor_chunk_index;
4371
4372 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4373 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4374
4375 do_young_space_rescan(cl, to_space, NULL, 0);
4376 do_young_space_rescan(cl, from_space, sca, sct);
4377 do_young_space_rescan(cl, eden_space, eca, ect);
4378 }
4379
4380 // work_queue(i) is passed to the closure
4381 // ParMarkRefsIntoAndScanClosure. The "i" parameter
4382 // also is passed to do_dirty_card_rescan_tasks() and to
4383 // do_work_steal() to select the i-th task_queue.
4384
4385 void CMSParRemarkTask::work(uint worker_id) {
4386 elapsedTimer _timer;
4387 ResourceMark rm;
4388 HandleMark hm;
4389
4390 // ---------- rescan from roots --------------
4391 _timer.start();
4392 CMSHeap* heap = CMSHeap::heap();
4393 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4394 _collector->_span, _collector->ref_processor(),
4395 &(_collector->_markBitMap),
4396 work_queue(worker_id));
4397
4398 // Rescan young gen roots first since these are likely
4399 // coarsely partitioned and may, on that account, constitute
4400 // the critical path; thus, it's best to start off that
4401 // work first.
4402 // ---------- young gen roots --------------
4403 {
4404 work_on_young_gen_roots(&par_mrias_cl);
4405 _timer.stop();
4406 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4407 }
4408
4409 // ---------- remaining roots --------------
4410 _timer.reset();
4411 _timer.start();
4412 heap->cms_process_roots(_strong_roots_scope,
4413 false, // yg was scanned above
4414 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4415 _collector->should_unload_classes(),
4416 &par_mrias_cl,
4417 NULL); // The dirty klasses will be handled below
4418
4419 assert(_collector->should_unload_classes()
4420 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4421 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4422 _timer.stop();
4423 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4424
4425 // ---------- unhandled CLD scanning ----------
4426 if (worker_id == 0) { // Single threaded at the moment.
4427 _timer.reset();
4428 _timer.start();
4429
4430 // Scan all new class loader data objects and new dependencies that were
4431 // introduced during concurrent marking.
4432 ResourceMark rm;
4433 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4434 for (int i = 0; i < array->length(); i++) {
4435 par_mrias_cl.do_cld_nv(array->at(i));
4436 }
4437
4438 // We don't need to keep track of new CLDs anymore.
4439 ClassLoaderDataGraph::remember_new_clds(false);
4440
4441 _timer.stop();
4442 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4443 }
4444
4445 // We might have added oops to ClassLoaderData::_handles during the
4446 // concurrent marking phase. These oops do not always point to newly allocated objects
4447 // that are guaranteed to be kept alive. Hence,
4448 // we do have to revisit the _handles block during the remark phase.
4449
4450 // ---------- dirty CLD scanning ----------
4451 if (worker_id == 0) { // Single threaded at the moment.
4452 _timer.reset();
4453 _timer.start();
4454
4455 // Scan all classes that was dirtied during the concurrent marking phase.
4456 RemarkCLDClosure remark_closure(&par_mrias_cl);
4457 ClassLoaderDataGraph::cld_do(&remark_closure);
4458
4459 _timer.stop();
4460 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4461 }
4462
4463
4464 // ---------- rescan dirty cards ------------
4465 _timer.reset();
4466 _timer.start();
4467
4468 // Do the rescan tasks for each of the two spaces
4469 // (cms_space) in turn.
4470 // "worker_id" is passed to select the task_queue for "worker_id"
4471 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4472 _timer.stop();
4473 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4474
4475 // ---------- steal work from other threads ...
4476 // ---------- ... and drain overflow list.
4477 _timer.reset();
4478 _timer.start();
4479 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4480 _timer.stop();
4481 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4482 }
4483
4484 void
4485 CMSParMarkTask::do_young_space_rescan(
4486 OopsInGenClosure* cl, ContiguousSpace* space,
4487 HeapWord** chunk_array, size_t chunk_top) {
4488 // Until all tasks completed:
4489 // . claim an unclaimed task
4490 // . compute region boundaries corresponding to task claimed
4491 // using chunk_array
4492 // . par_oop_iterate(cl) over that region
4493
4494 ResourceMark rm;
4495 HandleMark hm;
4496
4497 SequentialSubTasksDone* pst = space->par_seq_tasks();
4498
4499 uint nth_task = 0;
4500 uint n_tasks = pst->n_tasks();
4501
4502 if (n_tasks > 0) {
4503 assert(pst->valid(), "Uninitialized use?");
4504 HeapWord *start, *end;
4505 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4506 // We claimed task # nth_task; compute its boundaries.
4507 if (chunk_top == 0) { // no samples were taken
4508 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4509 start = space->bottom();
4510 end = space->top();
4511 } else if (nth_task == 0) {
4512 start = space->bottom();
4513 end = chunk_array[nth_task];
4514 } else if (nth_task < (uint)chunk_top) {
4515 assert(nth_task >= 1, "Control point invariant");
4516 start = chunk_array[nth_task - 1];
4517 end = chunk_array[nth_task];
4518 } else {
4519 assert(nth_task == (uint)chunk_top, "Control point invariant");
4520 start = chunk_array[chunk_top - 1];
4521 end = space->top();
4522 }
4523 MemRegion mr(start, end);
4524 // Verify that mr is in space
4525 assert(mr.is_empty() || space->used_region().contains(mr),
4526 "Should be in space");
4527 // Verify that "start" is an object boundary
4528 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())),
4529 "Should be an oop");
4530 space->par_oop_iterate(mr, cl);
4531 }
4532 pst->all_tasks_completed();
4533 }
4534 }
4535
4536 void
4537 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4538 CompactibleFreeListSpace* sp, int i,
4539 ParMarkRefsIntoAndScanClosure* cl) {
4540 // Until all tasks completed:
4541 // . claim an unclaimed task
4542 // . compute region boundaries corresponding to task claimed
4543 // . transfer dirty bits ct->mut for that region
4544 // . apply rescanclosure to dirty mut bits for that region
4545
4546 ResourceMark rm;
4547 HandleMark hm;
4548
4549 OopTaskQueue* work_q = work_queue(i);
4550 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4551 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4552 // CAUTION: This closure has state that persists across calls to
4553 // the work method dirty_range_iterate_clear() in that it has
4554 // embedded in it a (subtype of) UpwardsObjectClosure. The
4555 // use of that state in the embedded UpwardsObjectClosure instance
4556 // assumes that the cards are always iterated (even if in parallel
4557 // by several threads) in monotonically increasing order per each
4558 // thread. This is true of the implementation below which picks
4559 // card ranges (chunks) in monotonically increasing order globally
4560 // and, a-fortiori, in monotonically increasing order per thread
4561 // (the latter order being a subsequence of the former).
4562 // If the work code below is ever reorganized into a more chaotic
4563 // work-partitioning form than the current "sequential tasks"
4564 // paradigm, the use of that persistent state will have to be
4565 // revisited and modified appropriately. See also related
4566 // bug 4756801 work on which should examine this code to make
4567 // sure that the changes there do not run counter to the
4568 // assumptions made here and necessary for correctness and
4569 // efficiency. Note also that this code might yield inefficient
4570 // behavior in the case of very large objects that span one or
4571 // more work chunks. Such objects would potentially be scanned
4572 // several times redundantly. Work on 4756801 should try and
4573 // address that performance anomaly if at all possible. XXX
4574 MemRegion full_span = _collector->_span;
4575 CMSBitMap* bm = &(_collector->_markBitMap); // shared
4576 MarkFromDirtyCardsClosure
4577 greyRescanClosure(_collector, full_span, // entire span of interest
4578 sp, bm, work_q, cl);
4579
4580 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4581 assert(pst->valid(), "Uninitialized use?");
4582 uint nth_task = 0;
4583 const int alignment = CardTable::card_size * BitsPerWord;
4584 MemRegion span = sp->used_region();
4585 HeapWord* start_addr = span.start();
4586 HeapWord* end_addr = align_up(span.end(), alignment);
4587 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4588 assert(is_aligned(start_addr, alignment), "Check alignment");
4589 assert(is_aligned(chunk_size, alignment), "Check alignment");
4590
4591 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4592 // Having claimed the nth_task, compute corresponding mem-region,
4593 // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4594 // The alignment restriction ensures that we do not need any
4595 // synchronization with other gang-workers while setting or
4596 // clearing bits in thus chunk of the MUT.
4597 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4598 start_addr + (nth_task+1)*chunk_size);
4599 // The last chunk's end might be way beyond end of the
4600 // used region. In that case pull back appropriately.
4601 if (this_span.end() > end_addr) {
4602 this_span.set_end(end_addr);
4603 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4604 }
4605 // Iterate over the dirty cards covering this chunk, marking them
4606 // precleaned, and setting the corresponding bits in the mod union
4607 // table. Since we have been careful to partition at Card and MUT-word
4608 // boundaries no synchronization is needed between parallel threads.
4609 _collector->_ct->dirty_card_iterate(this_span,
4610 &modUnionClosure);
4611
4612 // Having transferred these marks into the modUnionTable,
4613 // rescan the marked objects on the dirty cards in the modUnionTable.
4614 // Even if this is at a synchronous collection, the initial marking
4615 // may have been done during an asynchronous collection so there
4616 // may be dirty bits in the mod-union table.
4617 _collector->_modUnionTable.dirty_range_iterate_clear(
4618 this_span, &greyRescanClosure);
4619 _collector->_modUnionTable.verifyNoOneBitsInRange(
4620 this_span.start(),
4621 this_span.end());
4622 }
4623 pst->all_tasks_completed(); // declare that i am done
4624 }
4625
4626 // . see if we can share work_queues with ParNew? XXX
4627 void
4628 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl,
4629 int* seed) {
4630 OopTaskQueue* work_q = work_queue(i);
4631 NOT_PRODUCT(int num_steals = 0;)
4632 oop obj_to_scan;
4633 CMSBitMap* bm = &(_collector->_markBitMap);
4634
4635 while (true) {
4636 // Completely finish any left over work from (an) earlier round(s)
4637 cl->trim_queue(0);
4638 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4639 (size_t)ParGCDesiredObjsFromOverflowList);
4640 // Now check if there's any work in the overflow list
4641 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4642 // only affects the number of attempts made to get work from the
4643 // overflow list and does not affect the number of workers. Just
4644 // pass ParallelGCThreads so this behavior is unchanged.
4645 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4646 work_q,
4647 ParallelGCThreads)) {
4648 // found something in global overflow list;
4649 // not yet ready to go stealing work from others.
4650 // We'd like to assert(work_q->size() != 0, ...)
4651 // because we just took work from the overflow list,
4652 // but of course we can't since all of that could have
4653 // been already stolen from us.
4654 // "He giveth and He taketh away."
4655 continue;
4656 }
4657 // Verify that we have no work before we resort to stealing
4658 assert(work_q->size() == 0, "Have work, shouldn't steal");
4659 // Try to steal from other queues that have work
4660 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4661 NOT_PRODUCT(num_steals++;)
4662 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!");
4663 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4664 // Do scanning work
4665 obj_to_scan->oop_iterate(cl);
4666 // Loop around, finish this work, and try to steal some more
4667 } else if (terminator()->offer_termination()) {
4668 break; // nirvana from the infinite cycle
4669 }
4670 }
4671 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
4672 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4673 "Else our work is not yet done");
4674 }
4675
4676 // Record object boundaries in _eden_chunk_array by sampling the eden
4677 // top in the slow-path eden object allocation code path and record
4678 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4679 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4680 // sampling in sample_eden() that activates during the part of the
4681 // preclean phase.
4682 void CMSCollector::sample_eden_chunk() {
4683 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4684 if (_eden_chunk_lock->try_lock()) {
4685 // Record a sample. This is the critical section. The contents
4686 // of the _eden_chunk_array have to be non-decreasing in the
4687 // address order.
4688 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4689 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4690 "Unexpected state of Eden");
4691 if (_eden_chunk_index == 0 ||
4692 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4693 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4694 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4695 _eden_chunk_index++; // commit sample
4696 }
4697 _eden_chunk_lock->unlock();
4698 }
4699 }
4700 }
4701
4702 // Return a thread-local PLAB recording array, as appropriate.
4703 void* CMSCollector::get_data_recorder(int thr_num) {
4704 if (_survivor_plab_array != NULL &&
4705 (CMSPLABRecordAlways ||
4706 (_collectorState > Marking && _collectorState < FinalMarking))) {
4707 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4708 ChunkArray* ca = &_survivor_plab_array[thr_num];
4709 ca->reset(); // clear it so that fresh data is recorded
4710 return (void*) ca;
4711 } else {
4712 return NULL;
4713 }
4714 }
4715
4716 // Reset all the thread-local PLAB recording arrays
4717 void CMSCollector::reset_survivor_plab_arrays() {
4718 for (uint i = 0; i < ParallelGCThreads; i++) {
4719 _survivor_plab_array[i].reset();
4720 }
4721 }
4722
4723 // Merge the per-thread plab arrays into the global survivor chunk
4724 // array which will provide the partitioning of the survivor space
4725 // for CMS initial scan and rescan.
4726 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4727 int no_of_gc_threads) {
4728 assert(_survivor_plab_array != NULL, "Error");
4729 assert(_survivor_chunk_array != NULL, "Error");
4730 assert(_collectorState == FinalMarking ||
4731 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4732 for (int j = 0; j < no_of_gc_threads; j++) {
4733 _cursor[j] = 0;
4734 }
4735 HeapWord* top = surv->top();
4736 size_t i;
4737 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
4738 HeapWord* min_val = top; // Higher than any PLAB address
4739 uint min_tid = 0; // position of min_val this round
4740 for (int j = 0; j < no_of_gc_threads; j++) {
4741 ChunkArray* cur_sca = &_survivor_plab_array[j];
4742 if (_cursor[j] == cur_sca->end()) {
4743 continue;
4744 }
4745 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4746 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4747 assert(surv->used_region().contains(cur_val), "Out of bounds value");
4748 if (cur_val < min_val) {
4749 min_tid = j;
4750 min_val = cur_val;
4751 } else {
4752 assert(cur_val < top, "All recorded addresses should be less");
4753 }
4754 }
4755 // At this point min_val and min_tid are respectively
4756 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4757 // and the thread (j) that witnesses that address.
4758 // We record this address in the _survivor_chunk_array[i]
4759 // and increment _cursor[min_tid] prior to the next round i.
4760 if (min_val == top) {
4761 break;
4762 }
4763 _survivor_chunk_array[i] = min_val;
4764 _cursor[min_tid]++;
4765 }
4766 // We are all done; record the size of the _survivor_chunk_array
4767 _survivor_chunk_index = i; // exclusive: [0, i)
4768 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4769 // Verify that we used up all the recorded entries
4770 #ifdef ASSERT
4771 size_t total = 0;
4772 for (int j = 0; j < no_of_gc_threads; j++) {
4773 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4774 total += _cursor[j];
4775 }
4776 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
4777 // Check that the merged array is in sorted order
4778 if (total > 0) {
4779 for (size_t i = 0; i < total - 1; i++) {
4780 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
4781 i, p2i(_survivor_chunk_array[i]));
4782 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
4783 "Not sorted");
4784 }
4785 }
4786 #endif // ASSERT
4787 }
4788
4789 // Set up the space's par_seq_tasks structure for work claiming
4790 // for parallel initial scan and rescan of young gen.
4791 // See ParRescanTask where this is currently used.
4792 void
4793 CMSCollector::
4794 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
4795 assert(n_threads > 0, "Unexpected n_threads argument");
4796
4797 // Eden space
4798 if (!_young_gen->eden()->is_empty()) {
4799 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
4800 assert(!pst->valid(), "Clobbering existing data?");
4801 // Each valid entry in [0, _eden_chunk_index) represents a task.
4802 size_t n_tasks = _eden_chunk_index + 1;
4803 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
4804 // Sets the condition for completion of the subtask (how many threads
4805 // need to finish in order to be done).
4806 pst->set_n_threads(n_threads);
4807 pst->set_n_tasks((int)n_tasks);
4808 }
4809
4810 // Merge the survivor plab arrays into _survivor_chunk_array
4811 if (_survivor_plab_array != NULL) {
4812 merge_survivor_plab_arrays(_young_gen->from(), n_threads);
4813 } else {
4814 assert(_survivor_chunk_index == 0, "Error");
4815 }
4816
4817 // To space
4818 {
4819 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
4820 assert(!pst->valid(), "Clobbering existing data?");
4821 // Sets the condition for completion of the subtask (how many threads
4822 // need to finish in order to be done).
4823 pst->set_n_threads(n_threads);
4824 pst->set_n_tasks(1);
4825 assert(pst->valid(), "Error");
4826 }
4827
4828 // From space
4829 {
4830 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
4831 assert(!pst->valid(), "Clobbering existing data?");
4832 size_t n_tasks = _survivor_chunk_index + 1;
4833 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
4834 // Sets the condition for completion of the subtask (how many threads
4835 // need to finish in order to be done).
4836 pst->set_n_threads(n_threads);
4837 pst->set_n_tasks((int)n_tasks);
4838 assert(pst->valid(), "Error");
4839 }
4840 }
4841
4842 // Parallel version of remark
4843 void CMSCollector::do_remark_parallel() {
4844 CMSHeap* heap = CMSHeap::heap();
4845 WorkGang* workers = heap->workers();
4846 assert(workers != NULL, "Need parallel worker threads.");
4847 // Choose to use the number of GC workers most recently set
4848 // into "active_workers".
4849 uint n_workers = workers->active_workers();
4850
4851 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4852
4853 StrongRootsScope srs(n_workers);
4854
4855 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
4856
4857 // We won't be iterating over the cards in the card table updating
4858 // the younger_gen cards, so we shouldn't call the following else
4859 // the verification code as well as subsequent younger_refs_iterate
4860 // code would get confused. XXX
4861 // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
4862
4863 // The young gen rescan work will not be done as part of
4864 // process_roots (which currently doesn't know how to
4865 // parallelize such a scan), but rather will be broken up into
4866 // a set of parallel tasks (via the sampling that the [abortable]
4867 // preclean phase did of eden, plus the [two] tasks of
4868 // scanning the [two] survivor spaces. Further fine-grain
4869 // parallelization of the scanning of the survivor spaces
4870 // themselves, and of precleaning of the young gen itself
4871 // is deferred to the future.
4872 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
4873
4874 // The dirty card rescan work is broken up into a "sequence"
4875 // of parallel tasks (per constituent space) that are dynamically
4876 // claimed by the parallel threads.
4877 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
4878
4879 // It turns out that even when we're using 1 thread, doing the work in a
4880 // separate thread causes wide variance in run times. We can't help this
4881 // in the multi-threaded case, but we special-case n=1 here to get
4882 // repeatable measurements of the 1-thread overhead of the parallel code.
4883 if (n_workers > 1) {
4884 // Make refs discovery MT-safe, if it isn't already: it may not
4885 // necessarily be so, since it's possible that we are doing
4886 // ST marking.
4887 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
4888 workers->run_task(&tsk);
4889 } else {
4890 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4891 tsk.work(0);
4892 }
4893
4894 // restore, single-threaded for now, any preserved marks
4895 // as a result of work_q overflow
4896 restore_preserved_marks_if_any();
4897 }
4898
4899 // Non-parallel version of remark
4900 void CMSCollector::do_remark_non_parallel() {
4901 ResourceMark rm;
4902 HandleMark hm;
4903 CMSHeap* heap = CMSHeap::heap();
4904 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4905
4906 MarkRefsIntoAndScanClosure
4907 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
4908 &_markStack, this,
4909 false /* should_yield */, false /* not precleaning */);
4910 MarkFromDirtyCardsClosure
4911 markFromDirtyCardsClosure(this, _span,
4912 NULL, // space is set further below
4913 &_markBitMap, &_markStack, &mrias_cl);
4914 {
4915 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm);
4916 // Iterate over the dirty cards, setting the corresponding bits in the
4917 // mod union table.
4918 {
4919 ModUnionClosure modUnionClosure(&_modUnionTable);
4920 _ct->dirty_card_iterate(_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 = CardTable::card_size * BitsPerWord;
4928 {
4929 // ... First handle dirty cards in CMS gen
4930 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
4931 MemRegion ur = _cmsGen->used_region();
4932 HeapWord* lb = ur.start();
4933 HeapWord* ub = align_up(ur.end(), alignment);
4934 MemRegion cms_span(lb, ub);
4935 _modUnionTable.dirty_range_iterate_clear(cms_span,
4936 &markFromDirtyCardsClosure);
4937 verify_work_stacks_empty();
4938 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards());
4939 }
4940 }
4941 if (VerifyDuringGC &&
4942 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) {
4943 HandleMark hm; // Discard invalid handles created during verification
4944 Universe::verify();
4945 }
4946 {
4947 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm);
4948
4949 verify_work_stacks_empty();
4950
4951 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
4952 StrongRootsScope srs(1);
4953
4954 heap->cms_process_roots(&srs,
4955 true, // young gen as roots
4956 GenCollectedHeap::ScanningOption(roots_scanning_options()),
4957 should_unload_classes(),
4958 &mrias_cl,
4959 NULL); // The dirty klasses will be handled below
4960
4961 assert(should_unload_classes()
4962 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4963 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4964 }
4965
4966 {
4967 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm);
4968
4969 verify_work_stacks_empty();
4970
4971 // Scan all class loader data objects that might have been introduced
4972 // during concurrent marking.
4973 ResourceMark rm;
4974 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4975 for (int i = 0; i < array->length(); i++) {
4976 mrias_cl.do_cld_nv(array->at(i));
4977 }
4978
4979 // We don't need to keep track of new CLDs anymore.
4980 ClassLoaderDataGraph::remember_new_clds(false);
4981
4982 verify_work_stacks_empty();
4983 }
4984
4985 // We might have added oops to ClassLoaderData::_handles during the
4986 // concurrent marking phase. These oops do not point to newly allocated objects
4987 // that are guaranteed to be kept alive. Hence,
4988 // we do have to revisit the _handles block during the remark phase.
4989 {
4990 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm);
4991
4992 verify_work_stacks_empty();
4993
4994 RemarkCLDClosure remark_closure(&mrias_cl);
4995 ClassLoaderDataGraph::cld_do(&remark_closure);
4996
4997 verify_work_stacks_empty();
4998 }
4999
5000 verify_work_stacks_empty();
5001 // Restore evacuated mark words, if any, used for overflow list links
5002 restore_preserved_marks_if_any();
5003
5004 verify_overflow_empty();
5005 }
5006
5007 ////////////////////////////////////////////////////////
5008 // Parallel Reference Processing Task Proxy Class
5009 ////////////////////////////////////////////////////////
5010 class AbstractGangTaskWOopQueues : public AbstractGangTask {
5011 OopTaskQueueSet* _queues;
5012 ParallelTaskTerminator _terminator;
5013 public:
5014 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) :
5015 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {}
5016 ParallelTaskTerminator* terminator() { return &_terminator; }
5017 OopTaskQueueSet* queues() { return _queues; }
5018 };
5019
5020 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5021 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5022 CMSCollector* _collector;
5023 CMSBitMap* _mark_bit_map;
5024 const MemRegion _span;
5025 ProcessTask& _task;
5026
5027 public:
5028 CMSRefProcTaskProxy(ProcessTask& task,
5029 CMSCollector* collector,
5030 const MemRegion& span,
5031 CMSBitMap* mark_bit_map,
5032 AbstractWorkGang* workers,
5033 OopTaskQueueSet* task_queues):
5034 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5035 task_queues,
5036 workers->active_workers()),
5037 _task(task),
5038 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5039 {
5040 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5041 "Inconsistency in _span");
5042 }
5043
5044 OopTaskQueueSet* task_queues() { return queues(); }
5045
5046 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5047
5048 void do_work_steal(int i,
5049 CMSParDrainMarkingStackClosure* drain,
5050 CMSParKeepAliveClosure* keep_alive,
5051 int* seed);
5052
5053 virtual void work(uint worker_id);
5054 };
5055
5056 void CMSRefProcTaskProxy::work(uint worker_id) {
5057 ResourceMark rm;
5058 HandleMark hm;
5059 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5060 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5061 _mark_bit_map,
5062 work_queue(worker_id));
5063 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5064 _mark_bit_map,
5065 work_queue(worker_id));
5066 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5067 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5068 if (_task.marks_oops_alive()) {
5069 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5070 _collector->hash_seed(worker_id));
5071 }
5072 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5073 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5074 }
5075
5076 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5077 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5078 EnqueueTask& _task;
5079
5080 public:
5081 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5082 : AbstractGangTask("Enqueue reference objects in parallel"),
5083 _task(task)
5084 { }
5085
5086 virtual void work(uint worker_id)
5087 {
5088 _task.work(worker_id);
5089 }
5090 };
5091
5092 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5093 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5094 _span(span),
5095 _bit_map(bit_map),
5096 _work_queue(work_queue),
5097 _mark_and_push(collector, span, bit_map, work_queue),
5098 _low_water_mark(MIN2((work_queue->max_elems()/4),
5099 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5100 { }
5101
5102 // . see if we can share work_queues with ParNew? XXX
5103 void CMSRefProcTaskProxy::do_work_steal(int i,
5104 CMSParDrainMarkingStackClosure* drain,
5105 CMSParKeepAliveClosure* keep_alive,
5106 int* seed) {
5107 OopTaskQueue* work_q = work_queue(i);
5108 NOT_PRODUCT(int num_steals = 0;)
5109 oop obj_to_scan;
5110
5111 while (true) {
5112 // Completely finish any left over work from (an) earlier round(s)
5113 drain->trim_queue(0);
5114 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5115 (size_t)ParGCDesiredObjsFromOverflowList);
5116 // Now check if there's any work in the overflow list
5117 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5118 // only affects the number of attempts made to get work from the
5119 // overflow list and does not affect the number of workers. Just
5120 // pass ParallelGCThreads so this behavior is unchanged.
5121 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5122 work_q,
5123 ParallelGCThreads)) {
5124 // Found something in global overflow list;
5125 // not yet ready to go stealing work from others.
5126 // We'd like to assert(work_q->size() != 0, ...)
5127 // because we just took work from the overflow list,
5128 // but of course we can't, since all of that might have
5129 // been already stolen from us.
5130 continue;
5131 }
5132 // Verify that we have no work before we resort to stealing
5133 assert(work_q->size() == 0, "Have work, shouldn't steal");
5134 // Try to steal from other queues that have work
5135 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5136 NOT_PRODUCT(num_steals++;)
5137 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!");
5138 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5139 // Do scanning work
5140 obj_to_scan->oop_iterate(keep_alive);
5141 // Loop around, finish this work, and try to steal some more
5142 } else if (terminator()->offer_termination()) {
5143 break; // nirvana from the infinite cycle
5144 }
5145 }
5146 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
5147 }
5148
5149 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5150 {
5151 CMSHeap* heap = CMSHeap::heap();
5152 WorkGang* workers = heap->workers();
5153 assert(workers != NULL, "Need parallel worker threads.");
5154 CMSRefProcTaskProxy rp_task(task, &_collector,
5155 _collector.ref_processor()->span(),
5156 _collector.markBitMap(),
5157 workers, _collector.task_queues());
5158 workers->run_task(&rp_task);
5159 }
5160
5161 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5162 {
5163
5164 CMSHeap* heap = CMSHeap::heap();
5165 WorkGang* workers = heap->workers();
5166 assert(workers != NULL, "Need parallel worker threads.");
5167 CMSRefEnqueueTaskProxy enq_task(task);
5168 workers->run_task(&enq_task);
5169 }
5170
5171 void CMSCollector::refProcessingWork() {
5172 ResourceMark rm;
5173 HandleMark hm;
5174
5175 ReferenceProcessor* rp = ref_processor();
5176 assert(rp->span().equals(_span), "Spans should be equal");
5177 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5178 // Process weak references.
5179 rp->setup_policy(false);
5180 verify_work_stacks_empty();
5181
5182 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->num_q());
5183 {
5184 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm);
5185
5186 // Setup keep_alive and complete closures.
5187 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5188 &_markStack, false /* !preclean */);
5189 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5190 _span, &_markBitMap, &_markStack,
5191 &cmsKeepAliveClosure, false /* !preclean */);
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 CMSHeap* heap = CMSHeap::heap();
5201 uint active_workers = ParallelGCThreads;
5202 WorkGang* workers = heap->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 &pt);
5217 } else {
5218 stats = rp->process_discovered_references(&_is_alive_closure,
5219 &cmsKeepAliveClosure,
5220 &cmsDrainMarkingStackClosure,
5221 NULL,
5222 &pt);
5223 }
5224 _gc_tracer_cm->report_gc_reference_stats(stats);
5225 pt.print_all_references();
5226 }
5227
5228 // This is the point where the entire marking should have completed.
5229 verify_work_stacks_empty();
5230
5231 {
5232 GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm);
5233 WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl);
5234 }
5235
5236 if (should_unload_classes()) {
5237 {
5238 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm);
5239
5240 // Unload classes and purge the SystemDictionary.
5241 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure, _gc_timer_cm);
5242
5243 // Unload nmethods.
5244 CodeCache::do_unloading(&_is_alive_closure, purged_class);
5245
5246 // Prune dead klasses from subklass/sibling/implementor lists.
5247 Klass::clean_weak_klass_links(&_is_alive_closure);
5248 }
5249
5250 {
5251 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm);
5252 // Clean up unreferenced symbols in symbol table.
5253 SymbolTable::unlink();
5254 }
5255
5256 {
5257 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm);
5258 // Delete entries for dead interned strings.
5259 StringTable::unlink(&_is_alive_closure);
5260 }
5261 }
5262
5263 // Restore any preserved marks as a result of mark stack or
5264 // work queue overflow
5265 restore_preserved_marks_if_any(); // done single-threaded for now
5266
5267 rp->set_enqueuing_is_done(true);
5268 if (rp->processing_is_mt()) {
5269 rp->balance_all_queues();
5270 CMSRefProcTaskExecutor task_executor(*this);
5271 rp->enqueue_discovered_references(&task_executor, &pt);
5272 } else {
5273 rp->enqueue_discovered_references(NULL, &pt);
5274 }
5275 rp->verify_no_references_recorded();
5276 pt.print_enqueue_phase();
5277 assert(!rp->discovery_enabled(), "should have been disabled");
5278 }
5279
5280 #ifndef PRODUCT
5281 void CMSCollector::check_correct_thread_executing() {
5282 Thread* t = Thread::current();
5283 // Only the VM thread or the CMS thread should be here.
5284 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5285 "Unexpected thread type");
5286 // If this is the vm thread, the foreground process
5287 // should not be waiting. Note that _foregroundGCIsActive is
5288 // true while the foreground collector is waiting.
5289 if (_foregroundGCShouldWait) {
5290 // We cannot be the VM thread
5291 assert(t->is_ConcurrentGC_thread(),
5292 "Should be CMS thread");
5293 } else {
5294 // We can be the CMS thread only if we are in a stop-world
5295 // phase of CMS collection.
5296 if (t->is_ConcurrentGC_thread()) {
5297 assert(_collectorState == InitialMarking ||
5298 _collectorState == FinalMarking,
5299 "Should be a stop-world phase");
5300 // The CMS thread should be holding the CMS_token.
5301 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5302 "Potential interference with concurrently "
5303 "executing VM thread");
5304 }
5305 }
5306 }
5307 #endif
5308
5309 void CMSCollector::sweep() {
5310 assert(_collectorState == Sweeping, "just checking");
5311 check_correct_thread_executing();
5312 verify_work_stacks_empty();
5313 verify_overflow_empty();
5314 increment_sweep_count();
5315 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
5316
5317 _inter_sweep_timer.stop();
5318 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5319
5320 assert(!_intra_sweep_timer.is_active(), "Should not be active");
5321 _intra_sweep_timer.reset();
5322 _intra_sweep_timer.start();
5323 {
5324 GCTraceCPUTime tcpu;
5325 CMSPhaseAccounting pa(this, "Concurrent Sweep");
5326 // First sweep the old gen
5327 {
5328 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5329 bitMapLock());
5330 sweepWork(_cmsGen);
5331 }
5332
5333 // Update Universe::_heap_*_at_gc figures.
5334 // We need all the free list locks to make the abstract state
5335 // transition from Sweeping to Resetting. See detailed note
5336 // further below.
5337 {
5338 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5339 // Update heap occupancy information which is used as
5340 // input to soft ref clearing policy at the next gc.
5341 Universe::update_heap_info_at_gc();
5342 _collectorState = Resizing;
5343 }
5344 }
5345 verify_work_stacks_empty();
5346 verify_overflow_empty();
5347
5348 if (should_unload_classes()) {
5349 // Delay purge to the beginning of the next safepoint. Metaspace::contains
5350 // requires that the virtual spaces are stable and not deleted.
5351 ClassLoaderDataGraph::set_should_purge(true);
5352 }
5353
5354 _intra_sweep_timer.stop();
5355 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5356
5357 _inter_sweep_timer.reset();
5358 _inter_sweep_timer.start();
5359
5360 // We need to use a monotonically non-decreasing time in ms
5361 // or we will see time-warp warnings and os::javaTimeMillis()
5362 // does not guarantee monotonicity.
5363 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5364 update_time_of_last_gc(now);
5365
5366 // NOTE on abstract state transitions:
5367 // Mutators allocate-live and/or mark the mod-union table dirty
5368 // based on the state of the collection. The former is done in
5369 // the interval [Marking, Sweeping] and the latter in the interval
5370 // [Marking, Sweeping). Thus the transitions into the Marking state
5371 // and out of the Sweeping state must be synchronously visible
5372 // globally to the mutators.
5373 // The transition into the Marking state happens with the world
5374 // stopped so the mutators will globally see it. Sweeping is
5375 // done asynchronously by the background collector so the transition
5376 // from the Sweeping state to the Resizing state must be done
5377 // under the freelistLock (as is the check for whether to
5378 // allocate-live and whether to dirty the mod-union table).
5379 assert(_collectorState == Resizing, "Change of collector state to"
5380 " Resizing must be done under the freelistLocks (plural)");
5381
5382 // Now that sweeping has been completed, we clear
5383 // the incremental_collection_failed flag,
5384 // thus inviting a younger gen collection to promote into
5385 // this generation. If such a promotion may still fail,
5386 // the flag will be set again when a young collection is
5387 // attempted.
5388 CMSHeap* heap = CMSHeap::heap();
5389 heap->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
5390 heap->update_full_collections_completed(_collection_count_start);
5391 }
5392
5393 // FIX ME!!! Looks like this belongs in CFLSpace, with
5394 // CMSGen merely delegating to it.
5395 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5396 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5397 HeapWord* minAddr = _cmsSpace->bottom();
5398 HeapWord* largestAddr =
5399 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5400 if (largestAddr == NULL) {
5401 // The dictionary appears to be empty. In this case
5402 // try to coalesce at the end of the heap.
5403 largestAddr = _cmsSpace->end();
5404 }
5405 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5406 size_t nearLargestOffset =
5407 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5408 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5409 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5410 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5411 }
5412
5413 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5414 return addr >= _cmsSpace->nearLargestChunk();
5415 }
5416
5417 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5418 return _cmsSpace->find_chunk_at_end();
5419 }
5420
5421 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation,
5422 bool full) {
5423 // If the young generation has been collected, gather any statistics
5424 // that are of interest at this point.
5425 bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation);
5426 if (!full && current_is_young) {
5427 // Gather statistics on the young generation collection.
5428 collector()->stats().record_gc0_end(used());
5429 }
5430 }
5431
5432 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) {
5433 // We iterate over the space(s) underlying this generation,
5434 // checking the mark bit map to see if the bits corresponding
5435 // to specific blocks are marked or not. Blocks that are
5436 // marked are live and are not swept up. All remaining blocks
5437 // are swept up, with coalescing on-the-fly as we sweep up
5438 // contiguous free and/or garbage blocks:
5439 // We need to ensure that the sweeper synchronizes with allocators
5440 // and stop-the-world collectors. In particular, the following
5441 // locks are used:
5442 // . CMS token: if this is held, a stop the world collection cannot occur
5443 // . freelistLock: if this is held no allocation can occur from this
5444 // generation by another thread
5445 // . bitMapLock: if this is held, no other thread can access or update
5446 //
5447
5448 // Note that we need to hold the freelistLock if we use
5449 // block iterate below; else the iterator might go awry if
5450 // a mutator (or promotion) causes block contents to change
5451 // (for instance if the allocator divvies up a block).
5452 // If we hold the free list lock, for all practical purposes
5453 // young generation GC's can't occur (they'll usually need to
5454 // promote), so we might as well prevent all young generation
5455 // GC's while we do a sweeping step. For the same reason, we might
5456 // as well take the bit map lock for the entire duration
5457
5458 // check that we hold the requisite locks
5459 assert(have_cms_token(), "Should hold cms token");
5460 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5461 assert_lock_strong(old_gen->freelistLock());
5462 assert_lock_strong(bitMapLock());
5463
5464 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5465 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
5466 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5467 _inter_sweep_estimate.padded_average(),
5468 _intra_sweep_estimate.padded_average());
5469 old_gen->setNearLargestChunk();
5470
5471 {
5472 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield);
5473 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5474 // We need to free-up/coalesce garbage/blocks from a
5475 // co-terminal free run. This is done in the SweepClosure
5476 // destructor; so, do not remove this scope, else the
5477 // end-of-sweep-census below will be off by a little bit.
5478 }
5479 old_gen->cmsSpace()->sweep_completed();
5480 old_gen->cmsSpace()->endSweepFLCensus(sweep_count());
5481 if (should_unload_classes()) { // unloaded classes this cycle,
5482 _concurrent_cycles_since_last_unload = 0; // ... reset count
5483 } else { // did not unload classes,
5484 _concurrent_cycles_since_last_unload++; // ... increment count
5485 }
5486 }
5487
5488 // Reset CMS data structures (for now just the marking bit map)
5489 // preparatory for the next cycle.
5490 void CMSCollector::reset_concurrent() {
5491 CMSTokenSyncWithLocks ts(true, bitMapLock());
5492
5493 // If the state is not "Resetting", the foreground thread
5494 // has done a collection and the resetting.
5495 if (_collectorState != Resetting) {
5496 assert(_collectorState == Idling, "The state should only change"
5497 " because the foreground collector has finished the collection");
5498 return;
5499 }
5500
5501 {
5502 // Clear the mark bitmap (no grey objects to start with)
5503 // for the next cycle.
5504 GCTraceCPUTime tcpu;
5505 CMSPhaseAccounting cmspa(this, "Concurrent Reset");
5506
5507 HeapWord* curAddr = _markBitMap.startWord();
5508 while (curAddr < _markBitMap.endWord()) {
5509 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
5510 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5511 _markBitMap.clear_large_range(chunk);
5512 if (ConcurrentMarkSweepThread::should_yield() &&
5513 !foregroundGCIsActive() &&
5514 CMSYield) {
5515 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5516 "CMS thread should hold CMS token");
5517 assert_lock_strong(bitMapLock());
5518 bitMapLock()->unlock();
5519 ConcurrentMarkSweepThread::desynchronize(true);
5520 stopTimer();
5521 incrementYields();
5522
5523 // See the comment in coordinator_yield()
5524 for (unsigned i = 0; i < CMSYieldSleepCount &&
5525 ConcurrentMarkSweepThread::should_yield() &&
5526 !CMSCollector::foregroundGCIsActive(); ++i) {
5527 os::sleep(Thread::current(), 1, false);
5528 }
5529
5530 ConcurrentMarkSweepThread::synchronize(true);
5531 bitMapLock()->lock_without_safepoint_check();
5532 startTimer();
5533 }
5534 curAddr = chunk.end();
5535 }
5536 // A successful mostly concurrent collection has been done.
5537 // Because only the full (i.e., concurrent mode failure) collections
5538 // are being measured for gc overhead limits, clean the "near" flag
5539 // and count.
5540 size_policy()->reset_gc_overhead_limit_count();
5541 _collectorState = Idling;
5542 }
5543
5544 register_gc_end();
5545 }
5546
5547 // Same as above but for STW paths
5548 void CMSCollector::reset_stw() {
5549 // already have the lock
5550 assert(_collectorState == Resetting, "just checking");
5551 assert_lock_strong(bitMapLock());
5552 GCIdMark gc_id_mark(_cmsThread->gc_id());
5553 _markBitMap.clear_all();
5554 _collectorState = Idling;
5555 register_gc_end();
5556 }
5557
5558 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5559 GCTraceCPUTime tcpu;
5560 TraceCollectorStats tcs_cgc(cgc_counters());
5561
5562 switch (op) {
5563 case CMS_op_checkpointRootsInitial: {
5564 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true);
5565 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT);
5566 checkpointRootsInitial();
5567 break;
5568 }
5569 case CMS_op_checkpointRootsFinal: {
5570 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true);
5571 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT);
5572 checkpointRootsFinal();
5573 break;
5574 }
5575 default:
5576 fatal("No such CMS_op");
5577 }
5578 }
5579
5580 #ifndef PRODUCT
5581 size_t const CMSCollector::skip_header_HeapWords() {
5582 return FreeChunk::header_size();
5583 }
5584
5585 // Try and collect here conditions that should hold when
5586 // CMS thread is exiting. The idea is that the foreground GC
5587 // thread should not be blocked if it wants to terminate
5588 // the CMS thread and yet continue to run the VM for a while
5589 // after that.
5590 void CMSCollector::verify_ok_to_terminate() const {
5591 assert(Thread::current()->is_ConcurrentGC_thread(),
5592 "should be called by CMS thread");
5593 assert(!_foregroundGCShouldWait, "should be false");
5594 // We could check here that all the various low-level locks
5595 // are not held by the CMS thread, but that is overkill; see
5596 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5597 // is checked.
5598 }
5599 #endif
5600
5601 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5602 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5603 "missing Printezis mark?");
5604 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5605 size_t size = pointer_delta(nextOneAddr + 1, addr);
5606 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5607 "alignment problem");
5608 assert(size >= 3, "Necessary for Printezis marks to work");
5609 return size;
5610 }
5611
5612 // A variant of the above (block_size_using_printezis_bits()) except
5613 // that we return 0 if the P-bits are not yet set.
5614 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5615 if (_markBitMap.isMarked(addr + 1)) {
5616 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5617 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5618 size_t size = pointer_delta(nextOneAddr + 1, addr);
5619 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5620 "alignment problem");
5621 assert(size >= 3, "Necessary for Printezis marks to work");
5622 return size;
5623 }
5624 return 0;
5625 }
5626
5627 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5628 size_t sz = 0;
5629 oop p = (oop)addr;
5630 if (p->klass_or_null_acquire() != NULL) {
5631 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5632 } else {
5633 sz = block_size_using_printezis_bits(addr);
5634 }
5635 assert(sz > 0, "size must be nonzero");
5636 HeapWord* next_block = addr + sz;
5637 HeapWord* next_card = align_up(next_block, CardTable::card_size);
5638 assert(align_down((uintptr_t)addr, CardTable::card_size) <
5639 align_down((uintptr_t)next_card, CardTable::card_size),
5640 "must be different cards");
5641 return next_card;
5642 }
5643
5644
5645 // CMS Bit Map Wrapper /////////////////////////////////////////
5646
5647 // Construct a CMS bit map infrastructure, but don't create the
5648 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5649 // further below.
5650 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5651 _bm(),
5652 _shifter(shifter),
5653 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5654 Monitor::_safepoint_check_sometimes) : NULL)
5655 {
5656 _bmStartWord = 0;
5657 _bmWordSize = 0;
5658 }
5659
5660 bool CMSBitMap::allocate(MemRegion mr) {
5661 _bmStartWord = mr.start();
5662 _bmWordSize = mr.word_size();
5663 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5664 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5665 if (!brs.is_reserved()) {
5666 log_warning(gc)("CMS bit map allocation failure");
5667 return false;
5668 }
5669 // For now we'll just commit all of the bit map up front.
5670 // Later on we'll try to be more parsimonious with swap.
5671 if (!_virtual_space.initialize(brs, brs.size())) {
5672 log_warning(gc)("CMS bit map backing store failure");
5673 return false;
5674 }
5675 assert(_virtual_space.committed_size() == brs.size(),
5676 "didn't reserve backing store for all of CMS bit map?");
5677 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5678 _bmWordSize, "inconsistency in bit map sizing");
5679 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter);
5680
5681 // bm.clear(); // can we rely on getting zero'd memory? verify below
5682 assert(isAllClear(),
5683 "Expected zero'd memory from ReservedSpace constructor");
5684 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5685 "consistency check");
5686 return true;
5687 }
5688
5689 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5690 HeapWord *next_addr, *end_addr, *last_addr;
5691 assert_locked();
5692 assert(covers(mr), "out-of-range error");
5693 // XXX assert that start and end are appropriately aligned
5694 for (next_addr = mr.start(), end_addr = mr.end();
5695 next_addr < end_addr; next_addr = last_addr) {
5696 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5697 last_addr = dirty_region.end();
5698 if (!dirty_region.is_empty()) {
5699 cl->do_MemRegion(dirty_region);
5700 } else {
5701 assert(last_addr == end_addr, "program logic");
5702 return;
5703 }
5704 }
5705 }
5706
5707 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5708 _bm.print_on_error(st, prefix);
5709 }
5710
5711 #ifndef PRODUCT
5712 void CMSBitMap::assert_locked() const {
5713 CMSLockVerifier::assert_locked(lock());
5714 }
5715
5716 bool CMSBitMap::covers(MemRegion mr) const {
5717 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5718 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5719 "size inconsistency");
5720 return (mr.start() >= _bmStartWord) &&
5721 (mr.end() <= endWord());
5722 }
5723
5724 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5725 return (start >= _bmStartWord && (start + size) <= endWord());
5726 }
5727
5728 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5729 // verify that there are no 1 bits in the interval [left, right)
5730 FalseBitMapClosure falseBitMapClosure;
5731 iterate(&falseBitMapClosure, left, right);
5732 }
5733
5734 void CMSBitMap::region_invariant(MemRegion mr)
5735 {
5736 assert_locked();
5737 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5738 assert(!mr.is_empty(), "unexpected empty region");
5739 assert(covers(mr), "mr should be covered by bit map");
5740 // convert address range into offset range
5741 size_t start_ofs = heapWordToOffset(mr.start());
5742 // Make sure that end() is appropriately aligned
5743 assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))),
5744 "Misaligned mr.end()");
5745 size_t end_ofs = heapWordToOffset(mr.end());
5746 assert(end_ofs > start_ofs, "Should mark at least one bit");
5747 }
5748
5749 #endif
5750
5751 bool CMSMarkStack::allocate(size_t size) {
5752 // allocate a stack of the requisite depth
5753 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5754 size * sizeof(oop)));
5755 if (!rs.is_reserved()) {
5756 log_warning(gc)("CMSMarkStack allocation failure");
5757 return false;
5758 }
5759 if (!_virtual_space.initialize(rs, rs.size())) {
5760 log_warning(gc)("CMSMarkStack backing store failure");
5761 return false;
5762 }
5763 assert(_virtual_space.committed_size() == rs.size(),
5764 "didn't reserve backing store for all of CMS stack?");
5765 _base = (oop*)(_virtual_space.low());
5766 _index = 0;
5767 _capacity = size;
5768 NOT_PRODUCT(_max_depth = 0);
5769 return true;
5770 }
5771
5772 // XXX FIX ME !!! In the MT case we come in here holding a
5773 // leaf lock. For printing we need to take a further lock
5774 // which has lower rank. We need to recalibrate the two
5775 // lock-ranks involved in order to be able to print the
5776 // messages below. (Or defer the printing to the caller.
5777 // For now we take the expedient path of just disabling the
5778 // messages for the problematic case.)
5779 void CMSMarkStack::expand() {
5780 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
5781 if (_capacity == MarkStackSizeMax) {
5782 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) {
5783 // We print a warning message only once per CMS cycle.
5784 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit");
5785 }
5786 return;
5787 }
5788 // Double capacity if possible
5789 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
5790 // Do not give up existing stack until we have managed to
5791 // get the double capacity that we desired.
5792 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5793 new_capacity * sizeof(oop)));
5794 if (rs.is_reserved()) {
5795 // Release the backing store associated with old stack
5796 _virtual_space.release();
5797 // Reinitialize virtual space for new stack
5798 if (!_virtual_space.initialize(rs, rs.size())) {
5799 fatal("Not enough swap for expanded marking stack");
5800 }
5801 _base = (oop*)(_virtual_space.low());
5802 _index = 0;
5803 _capacity = new_capacity;
5804 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) {
5805 // Failed to double capacity, continue;
5806 // we print a detail message only once per CMS cycle.
5807 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
5808 _capacity / K, new_capacity / K);
5809 }
5810 }
5811
5812
5813 // Closures
5814 // XXX: there seems to be a lot of code duplication here;
5815 // should refactor and consolidate common code.
5816
5817 // This closure is used to mark refs into the CMS generation in
5818 // the CMS bit map. Called at the first checkpoint. This closure
5819 // assumes that we do not need to re-mark dirty cards; if the CMS
5820 // generation on which this is used is not an oldest
5821 // generation then this will lose younger_gen cards!
5822
5823 MarkRefsIntoClosure::MarkRefsIntoClosure(
5824 MemRegion span, CMSBitMap* bitMap):
5825 _span(span),
5826 _bitMap(bitMap)
5827 {
5828 assert(ref_processor() == NULL, "deliberately left NULL");
5829 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5830 }
5831
5832 void MarkRefsIntoClosure::do_oop(oop obj) {
5833 // if p points into _span, then mark corresponding bit in _markBitMap
5834 assert(oopDesc::is_oop(obj), "expected an oop");
5835 HeapWord* addr = (HeapWord*)obj;
5836 if (_span.contains(addr)) {
5837 // this should be made more efficient
5838 _bitMap->mark(addr);
5839 }
5840 }
5841
5842 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
5843 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
5844
5845 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure(
5846 MemRegion span, CMSBitMap* bitMap):
5847 _span(span),
5848 _bitMap(bitMap)
5849 {
5850 assert(ref_processor() == NULL, "deliberately left NULL");
5851 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5852 }
5853
5854 void ParMarkRefsIntoClosure::do_oop(oop obj) {
5855 // if p points into _span, then mark corresponding bit in _markBitMap
5856 assert(oopDesc::is_oop(obj), "expected an oop");
5857 HeapWord* addr = (HeapWord*)obj;
5858 if (_span.contains(addr)) {
5859 // this should be made more efficient
5860 _bitMap->par_mark(addr);
5861 }
5862 }
5863
5864 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); }
5865 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); }
5866
5867 // A variant of the above, used for CMS marking verification.
5868 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
5869 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
5870 _span(span),
5871 _verification_bm(verification_bm),
5872 _cms_bm(cms_bm)
5873 {
5874 assert(ref_processor() == NULL, "deliberately left NULL");
5875 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
5876 }
5877
5878 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
5879 // if p points into _span, then mark corresponding bit in _markBitMap
5880 assert(oopDesc::is_oop(obj), "expected an oop");
5881 HeapWord* addr = (HeapWord*)obj;
5882 if (_span.contains(addr)) {
5883 _verification_bm->mark(addr);
5884 if (!_cms_bm->isMarked(addr)) {
5885 Log(gc, verify) log;
5886 ResourceMark rm;
5887 LogStream ls(log.error());
5888 oop(addr)->print_on(&ls);
5889 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
5890 fatal("... aborting");
5891 }
5892 }
5893 }
5894
5895 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5896 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5897
5898 //////////////////////////////////////////////////
5899 // MarkRefsIntoAndScanClosure
5900 //////////////////////////////////////////////////
5901
5902 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
5903 ReferenceProcessor* rp,
5904 CMSBitMap* bit_map,
5905 CMSBitMap* mod_union_table,
5906 CMSMarkStack* mark_stack,
5907 CMSCollector* collector,
5908 bool should_yield,
5909 bool concurrent_precleaning):
5910 _collector(collector),
5911 _span(span),
5912 _bit_map(bit_map),
5913 _mark_stack(mark_stack),
5914 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
5915 mark_stack, concurrent_precleaning),
5916 _yield(should_yield),
5917 _concurrent_precleaning(concurrent_precleaning),
5918 _freelistLock(NULL)
5919 {
5920 // FIXME: Should initialize in base class constructor.
5921 assert(rp != NULL, "ref_processor shouldn't be NULL");
5922 set_ref_processor_internal(rp);
5923 }
5924
5925 // This closure is used to mark refs into the CMS generation at the
5926 // second (final) checkpoint, and to scan and transitively follow
5927 // the unmarked oops. It is also used during the concurrent precleaning
5928 // phase while scanning objects on dirty cards in the CMS generation.
5929 // The marks are made in the marking bit map and the marking stack is
5930 // used for keeping the (newly) grey objects during the scan.
5931 // The parallel version (Par_...) appears further below.
5932 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
5933 if (obj != NULL) {
5934 assert(oopDesc::is_oop(obj), "expected an oop");
5935 HeapWord* addr = (HeapWord*)obj;
5936 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
5937 assert(_collector->overflow_list_is_empty(),
5938 "overflow list should be empty");
5939 if (_span.contains(addr) &&
5940 !_bit_map->isMarked(addr)) {
5941 // mark bit map (object is now grey)
5942 _bit_map->mark(addr);
5943 // push on marking stack (stack should be empty), and drain the
5944 // stack by applying this closure to the oops in the oops popped
5945 // from the stack (i.e. blacken the grey objects)
5946 bool res = _mark_stack->push(obj);
5947 assert(res, "Should have space to push on empty stack");
5948 do {
5949 oop new_oop = _mark_stack->pop();
5950 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
5951 assert(_bit_map->isMarked((HeapWord*)new_oop),
5952 "only grey objects on this stack");
5953 // iterate over the oops in this oop, marking and pushing
5954 // the ones in CMS heap (i.e. in _span).
5955 new_oop->oop_iterate(&_pushAndMarkClosure);
5956 // check if it's time to yield
5957 do_yield_check();
5958 } while (!_mark_stack->isEmpty() ||
5959 (!_concurrent_precleaning && take_from_overflow_list()));
5960 // if marking stack is empty, and we are not doing this
5961 // during precleaning, then check the overflow list
5962 }
5963 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
5964 assert(_collector->overflow_list_is_empty(),
5965 "overflow list was drained above");
5966
5967 assert(_collector->no_preserved_marks(),
5968 "All preserved marks should have been restored above");
5969 }
5970 }
5971
5972 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5973 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5974
5975 void MarkRefsIntoAndScanClosure::do_yield_work() {
5976 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5977 "CMS thread should hold CMS token");
5978 assert_lock_strong(_freelistLock);
5979 assert_lock_strong(_bit_map->lock());
5980 // relinquish the free_list_lock and bitMaplock()
5981 _bit_map->lock()->unlock();
5982 _freelistLock->unlock();
5983 ConcurrentMarkSweepThread::desynchronize(true);
5984 _collector->stopTimer();
5985 _collector->incrementYields();
5986
5987 // See the comment in coordinator_yield()
5988 for (unsigned i = 0;
5989 i < CMSYieldSleepCount &&
5990 ConcurrentMarkSweepThread::should_yield() &&
5991 !CMSCollector::foregroundGCIsActive();
5992 ++i) {
5993 os::sleep(Thread::current(), 1, false);
5994 }
5995
5996 ConcurrentMarkSweepThread::synchronize(true);
5997 _freelistLock->lock_without_safepoint_check();
5998 _bit_map->lock()->lock_without_safepoint_check();
5999 _collector->startTimer();
6000 }
6001
6002 ///////////////////////////////////////////////////////////
6003 // ParMarkRefsIntoAndScanClosure: a parallel version of
6004 // MarkRefsIntoAndScanClosure
6005 ///////////////////////////////////////////////////////////
6006 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure(
6007 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6008 CMSBitMap* bit_map, OopTaskQueue* work_queue):
6009 _span(span),
6010 _bit_map(bit_map),
6011 _work_queue(work_queue),
6012 _low_water_mark(MIN2((work_queue->max_elems()/4),
6013 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6014 _parPushAndMarkClosure(collector, span, rp, bit_map, work_queue)
6015 {
6016 // FIXME: Should initialize in base class constructor.
6017 assert(rp != NULL, "ref_processor shouldn't be NULL");
6018 set_ref_processor_internal(rp);
6019 }
6020
6021 // This closure is used to mark refs into the CMS generation at the
6022 // second (final) checkpoint, and to scan and transitively follow
6023 // the unmarked oops. The marks are made in the marking bit map and
6024 // the work_queue is used for keeping the (newly) grey objects during
6025 // the scan phase whence they are also available for stealing by parallel
6026 // threads. Since the marking bit map is shared, updates are
6027 // synchronized (via CAS).
6028 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) {
6029 if (obj != NULL) {
6030 // Ignore mark word because this could be an already marked oop
6031 // that may be chained at the end of the overflow list.
6032 assert(oopDesc::is_oop(obj, true), "expected an oop");
6033 HeapWord* addr = (HeapWord*)obj;
6034 if (_span.contains(addr) &&
6035 !_bit_map->isMarked(addr)) {
6036 // mark bit map (object will become grey):
6037 // It is possible for several threads to be
6038 // trying to "claim" this object concurrently;
6039 // the unique thread that succeeds in marking the
6040 // object first will do the subsequent push on
6041 // to the work queue (or overflow list).
6042 if (_bit_map->par_mark(addr)) {
6043 // push on work_queue (which may not be empty), and trim the
6044 // queue to an appropriate length by applying this closure to
6045 // the oops in the oops popped from the stack (i.e. blacken the
6046 // grey objects)
6047 bool res = _work_queue->push(obj);
6048 assert(res, "Low water mark should be less than capacity?");
6049 trim_queue(_low_water_mark);
6050 } // Else, another thread claimed the object
6051 }
6052 }
6053 }
6054
6055 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6056 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6057
6058 // This closure is used to rescan the marked objects on the dirty cards
6059 // in the mod union table and the card table proper.
6060 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6061 oop p, MemRegion mr) {
6062
6063 size_t size = 0;
6064 HeapWord* addr = (HeapWord*)p;
6065 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6066 assert(_span.contains(addr), "we are scanning the CMS generation");
6067 // check if it's time to yield
6068 if (do_yield_check()) {
6069 // We yielded for some foreground stop-world work,
6070 // and we have been asked to abort this ongoing preclean cycle.
6071 return 0;
6072 }
6073 if (_bitMap->isMarked(addr)) {
6074 // it's marked; is it potentially uninitialized?
6075 if (p->klass_or_null_acquire() != NULL) {
6076 // an initialized object; ignore mark word in verification below
6077 // since we are running concurrent with mutators
6078 assert(oopDesc::is_oop(p, true), "should be an oop");
6079 if (p->is_objArray()) {
6080 // objArrays are precisely marked; restrict scanning
6081 // to dirty cards only.
6082 size = CompactibleFreeListSpace::adjustObjectSize(
6083 p->oop_iterate_size(_scanningClosure, mr));
6084 } else {
6085 // A non-array may have been imprecisely marked; we need
6086 // to scan object in its entirety.
6087 size = CompactibleFreeListSpace::adjustObjectSize(
6088 p->oop_iterate_size(_scanningClosure));
6089 }
6090 #ifdef ASSERT
6091 size_t direct_size =
6092 CompactibleFreeListSpace::adjustObjectSize(p->size());
6093 assert(size == direct_size, "Inconsistency in size");
6094 assert(size >= 3, "Necessary for Printezis marks to work");
6095 HeapWord* start_pbit = addr + 1;
6096 HeapWord* end_pbit = addr + size - 1;
6097 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit),
6098 "inconsistent Printezis mark");
6099 // Verify inner mark bits (between Printezis bits) are clear,
6100 // but don't repeat if there are multiple dirty regions for
6101 // the same object, to avoid potential O(N^2) performance.
6102 if (addr != _last_scanned_object) {
6103 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit);
6104 _last_scanned_object = addr;
6105 }
6106 #endif // ASSERT
6107 } else {
6108 // An uninitialized object.
6109 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6110 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6111 size = pointer_delta(nextOneAddr + 1, addr);
6112 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6113 "alignment problem");
6114 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6115 // will dirty the card when the klass pointer is installed in the
6116 // object (signaling the completion of initialization).
6117 }
6118 } else {
6119 // Either a not yet marked object or an uninitialized object
6120 if (p->klass_or_null_acquire() == NULL) {
6121 // An uninitialized object, skip to the next card, since
6122 // we may not be able to read its P-bits yet.
6123 assert(size == 0, "Initial value");
6124 } else {
6125 // An object not (yet) reached by marking: we merely need to
6126 // compute its size so as to go look at the next block.
6127 assert(oopDesc::is_oop(p, true), "should be an oop");
6128 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6129 }
6130 }
6131 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6132 return size;
6133 }
6134
6135 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6136 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6137 "CMS thread should hold CMS token");
6138 assert_lock_strong(_freelistLock);
6139 assert_lock_strong(_bitMap->lock());
6140 // relinquish the free_list_lock and bitMaplock()
6141 _bitMap->lock()->unlock();
6142 _freelistLock->unlock();
6143 ConcurrentMarkSweepThread::desynchronize(true);
6144 _collector->stopTimer();
6145 _collector->incrementYields();
6146
6147 // See the comment in coordinator_yield()
6148 for (unsigned i = 0; i < CMSYieldSleepCount &&
6149 ConcurrentMarkSweepThread::should_yield() &&
6150 !CMSCollector::foregroundGCIsActive(); ++i) {
6151 os::sleep(Thread::current(), 1, false);
6152 }
6153
6154 ConcurrentMarkSweepThread::synchronize(true);
6155 _freelistLock->lock_without_safepoint_check();
6156 _bitMap->lock()->lock_without_safepoint_check();
6157 _collector->startTimer();
6158 }
6159
6160
6161 //////////////////////////////////////////////////////////////////
6162 // SurvivorSpacePrecleanClosure
6163 //////////////////////////////////////////////////////////////////
6164 // This (single-threaded) closure is used to preclean the oops in
6165 // the survivor spaces.
6166 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6167
6168 HeapWord* addr = (HeapWord*)p;
6169 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6170 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6171 assert(p->klass_or_null() != NULL, "object should be initialized");
6172 // an initialized object; ignore mark word in verification below
6173 // since we are running concurrent with mutators
6174 assert(oopDesc::is_oop(p, true), "should be an oop");
6175 // Note that we do not yield while we iterate over
6176 // the interior oops of p, pushing the relevant ones
6177 // on our marking stack.
6178 size_t size = p->oop_iterate_size(_scanning_closure);
6179 do_yield_check();
6180 // Observe that below, we do not abandon the preclean
6181 // phase as soon as we should; rather we empty the
6182 // marking stack before returning. This is to satisfy
6183 // some existing assertions. In general, it may be a
6184 // good idea to abort immediately and complete the marking
6185 // from the grey objects at a later time.
6186 while (!_mark_stack->isEmpty()) {
6187 oop new_oop = _mark_stack->pop();
6188 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
6189 assert(_bit_map->isMarked((HeapWord*)new_oop),
6190 "only grey objects on this stack");
6191 // iterate over the oops in this oop, marking and pushing
6192 // the ones in CMS heap (i.e. in _span).
6193 new_oop->oop_iterate(_scanning_closure);
6194 // check if it's time to yield
6195 do_yield_check();
6196 }
6197 unsigned int after_count =
6198 CMSHeap::heap()->total_collections();
6199 bool abort = (_before_count != after_count) ||
6200 _collector->should_abort_preclean();
6201 return abort ? 0 : size;
6202 }
6203
6204 void SurvivorSpacePrecleanClosure::do_yield_work() {
6205 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6206 "CMS thread should hold CMS token");
6207 assert_lock_strong(_bit_map->lock());
6208 // Relinquish the bit map lock
6209 _bit_map->lock()->unlock();
6210 ConcurrentMarkSweepThread::desynchronize(true);
6211 _collector->stopTimer();
6212 _collector->incrementYields();
6213
6214 // See the comment in coordinator_yield()
6215 for (unsigned i = 0; i < CMSYieldSleepCount &&
6216 ConcurrentMarkSweepThread::should_yield() &&
6217 !CMSCollector::foregroundGCIsActive(); ++i) {
6218 os::sleep(Thread::current(), 1, false);
6219 }
6220
6221 ConcurrentMarkSweepThread::synchronize(true);
6222 _bit_map->lock()->lock_without_safepoint_check();
6223 _collector->startTimer();
6224 }
6225
6226 // This closure is used to rescan the marked objects on the dirty cards
6227 // in the mod union table and the card table proper. In the parallel
6228 // case, although the bitMap is shared, we do a single read so the
6229 // isMarked() query is "safe".
6230 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6231 // Ignore mark word because we are running concurrent with mutators
6232 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p));
6233 HeapWord* addr = (HeapWord*)p;
6234 assert(_span.contains(addr), "we are scanning the CMS generation");
6235 bool is_obj_array = false;
6236 #ifdef ASSERT
6237 if (!_parallel) {
6238 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6239 assert(_collector->overflow_list_is_empty(),
6240 "overflow list should be empty");
6241
6242 }
6243 #endif // ASSERT
6244 if (_bit_map->isMarked(addr)) {
6245 // Obj arrays are precisely marked, non-arrays are not;
6246 // so we scan objArrays precisely and non-arrays in their
6247 // entirety.
6248 if (p->is_objArray()) {
6249 is_obj_array = true;
6250 if (_parallel) {
6251 p->oop_iterate(_par_scan_closure, mr);
6252 } else {
6253 p->oop_iterate(_scan_closure, mr);
6254 }
6255 } else {
6256 if (_parallel) {
6257 p->oop_iterate(_par_scan_closure);
6258 } else {
6259 p->oop_iterate(_scan_closure);
6260 }
6261 }
6262 }
6263 #ifdef ASSERT
6264 if (!_parallel) {
6265 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6266 assert(_collector->overflow_list_is_empty(),
6267 "overflow list should be empty");
6268
6269 }
6270 #endif // ASSERT
6271 return is_obj_array;
6272 }
6273
6274 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6275 MemRegion span,
6276 CMSBitMap* bitMap, CMSMarkStack* markStack,
6277 bool should_yield, bool verifying):
6278 _collector(collector),
6279 _span(span),
6280 _bitMap(bitMap),
6281 _mut(&collector->_modUnionTable),
6282 _markStack(markStack),
6283 _yield(should_yield),
6284 _skipBits(0)
6285 {
6286 assert(_markStack->isEmpty(), "stack should be empty");
6287 _finger = _bitMap->startWord();
6288 _threshold = _finger;
6289 assert(_collector->_restart_addr == NULL, "Sanity check");
6290 assert(_span.contains(_finger), "Out of bounds _finger?");
6291 DEBUG_ONLY(_verifying = verifying;)
6292 }
6293
6294 void MarkFromRootsClosure::reset(HeapWord* addr) {
6295 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6296 assert(_span.contains(addr), "Out of bounds _finger?");
6297 _finger = addr;
6298 _threshold = align_up(_finger, CardTable::card_size);
6299 }
6300
6301 // Should revisit to see if this should be restructured for
6302 // greater efficiency.
6303 bool MarkFromRootsClosure::do_bit(size_t offset) {
6304 if (_skipBits > 0) {
6305 _skipBits--;
6306 return true;
6307 }
6308 // convert offset into a HeapWord*
6309 HeapWord* addr = _bitMap->startWord() + offset;
6310 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6311 "address out of range");
6312 assert(_bitMap->isMarked(addr), "tautology");
6313 if (_bitMap->isMarked(addr+1)) {
6314 // this is an allocated but not yet initialized object
6315 assert(_skipBits == 0, "tautology");
6316 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6317 oop p = oop(addr);
6318 if (p->klass_or_null_acquire() == NULL) {
6319 DEBUG_ONLY(if (!_verifying) {)
6320 // We re-dirty the cards on which this object lies and increase
6321 // the _threshold so that we'll come back to scan this object
6322 // during the preclean or remark phase. (CMSCleanOnEnter)
6323 if (CMSCleanOnEnter) {
6324 size_t sz = _collector->block_size_using_printezis_bits(addr);
6325 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size);
6326 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6327 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6328 // Bump _threshold to end_card_addr; note that
6329 // _threshold cannot possibly exceed end_card_addr, anyhow.
6330 // This prevents future clearing of the card as the scan proceeds
6331 // to the right.
6332 assert(_threshold <= end_card_addr,
6333 "Because we are just scanning into this object");
6334 if (_threshold < end_card_addr) {
6335 _threshold = end_card_addr;
6336 }
6337 if (p->klass_or_null_acquire() != NULL) {
6338 // Redirty the range of cards...
6339 _mut->mark_range(redirty_range);
6340 } // ...else the setting of klass will dirty the card anyway.
6341 }
6342 DEBUG_ONLY(})
6343 return true;
6344 }
6345 }
6346 scanOopsInOop(addr);
6347 return true;
6348 }
6349
6350 // We take a break if we've been at this for a while,
6351 // so as to avoid monopolizing the locks involved.
6352 void MarkFromRootsClosure::do_yield_work() {
6353 // First give up the locks, then yield, then re-lock
6354 // We should probably use a constructor/destructor idiom to
6355 // do this unlock/lock or modify the MutexUnlocker class to
6356 // serve our purpose. XXX
6357 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6358 "CMS thread should hold CMS token");
6359 assert_lock_strong(_bitMap->lock());
6360 _bitMap->lock()->unlock();
6361 ConcurrentMarkSweepThread::desynchronize(true);
6362 _collector->stopTimer();
6363 _collector->incrementYields();
6364
6365 // See the comment in coordinator_yield()
6366 for (unsigned i = 0; i < CMSYieldSleepCount &&
6367 ConcurrentMarkSweepThread::should_yield() &&
6368 !CMSCollector::foregroundGCIsActive(); ++i) {
6369 os::sleep(Thread::current(), 1, false);
6370 }
6371
6372 ConcurrentMarkSweepThread::synchronize(true);
6373 _bitMap->lock()->lock_without_safepoint_check();
6374 _collector->startTimer();
6375 }
6376
6377 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6378 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6379 assert(_markStack->isEmpty(),
6380 "should drain stack to limit stack usage");
6381 // convert ptr to an oop preparatory to scanning
6382 oop obj = oop(ptr);
6383 // Ignore mark word in verification below, since we
6384 // may be running concurrent with mutators.
6385 assert(oopDesc::is_oop(obj, true), "should be an oop");
6386 assert(_finger <= ptr, "_finger runneth ahead");
6387 // advance the finger to right end of this object
6388 _finger = ptr + obj->size();
6389 assert(_finger > ptr, "we just incremented it above");
6390 // On large heaps, it may take us some time to get through
6391 // the marking phase. During
6392 // this time it's possible that a lot of mutations have
6393 // accumulated in the card table and the mod union table --
6394 // these mutation records are redundant until we have
6395 // actually traced into the corresponding card.
6396 // Here, we check whether advancing the finger would make
6397 // us cross into a new card, and if so clear corresponding
6398 // cards in the MUT (preclean them in the card-table in the
6399 // future).
6400
6401 DEBUG_ONLY(if (!_verifying) {)
6402 // The clean-on-enter optimization is disabled by default,
6403 // until we fix 6178663.
6404 if (CMSCleanOnEnter && (_finger > _threshold)) {
6405 // [_threshold, _finger) represents the interval
6406 // of cards to be cleared in MUT (or precleaned in card table).
6407 // The set of cards to be cleared is all those that overlap
6408 // with the interval [_threshold, _finger); note that
6409 // _threshold is always kept card-aligned but _finger isn't
6410 // always card-aligned.
6411 HeapWord* old_threshold = _threshold;
6412 assert(is_aligned(old_threshold, CardTable::card_size),
6413 "_threshold should always be card-aligned");
6414 _threshold = align_up(_finger, CardTable::card_size);
6415 MemRegion mr(old_threshold, _threshold);
6416 assert(!mr.is_empty(), "Control point invariant");
6417 assert(_span.contains(mr), "Should clear within span");
6418 _mut->clear_range(mr);
6419 }
6420 DEBUG_ONLY(})
6421 // Note: the finger doesn't advance while we drain
6422 // the stack below.
6423 PushOrMarkClosure pushOrMarkClosure(_collector,
6424 _span, _bitMap, _markStack,
6425 _finger, this);
6426 bool res = _markStack->push(obj);
6427 assert(res, "Empty non-zero size stack should have space for single push");
6428 while (!_markStack->isEmpty()) {
6429 oop new_oop = _markStack->pop();
6430 // Skip verifying header mark word below because we are
6431 // running concurrent with mutators.
6432 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop");
6433 // now scan this oop's oops
6434 new_oop->oop_iterate(&pushOrMarkClosure);
6435 do_yield_check();
6436 }
6437 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6438 }
6439
6440 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task,
6441 CMSCollector* collector, MemRegion span,
6442 CMSBitMap* bit_map,
6443 OopTaskQueue* work_queue,
6444 CMSMarkStack* overflow_stack):
6445 _collector(collector),
6446 _whole_span(collector->_span),
6447 _span(span),
6448 _bit_map(bit_map),
6449 _mut(&collector->_modUnionTable),
6450 _work_queue(work_queue),
6451 _overflow_stack(overflow_stack),
6452 _skip_bits(0),
6453 _task(task)
6454 {
6455 assert(_work_queue->size() == 0, "work_queue should be empty");
6456 _finger = span.start();
6457 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
6458 assert(_span.contains(_finger), "Out of bounds _finger?");
6459 }
6460
6461 // Should revisit to see if this should be restructured for
6462 // greater efficiency.
6463 bool ParMarkFromRootsClosure::do_bit(size_t offset) {
6464 if (_skip_bits > 0) {
6465 _skip_bits--;
6466 return true;
6467 }
6468 // convert offset into a HeapWord*
6469 HeapWord* addr = _bit_map->startWord() + offset;
6470 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6471 "address out of range");
6472 assert(_bit_map->isMarked(addr), "tautology");
6473 if (_bit_map->isMarked(addr+1)) {
6474 // this is an allocated object that might not yet be initialized
6475 assert(_skip_bits == 0, "tautology");
6476 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
6477 oop p = oop(addr);
6478 if (p->klass_or_null_acquire() == NULL) {
6479 // in the case of Clean-on-Enter optimization, redirty card
6480 // and avoid clearing card by increasing the threshold.
6481 return true;
6482 }
6483 }
6484 scan_oops_in_oop(addr);
6485 return true;
6486 }
6487
6488 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6489 assert(_bit_map->isMarked(ptr), "expected bit to be set");
6490 // Should we assert that our work queue is empty or
6491 // below some drain limit?
6492 assert(_work_queue->size() == 0,
6493 "should drain stack to limit stack usage");
6494 // convert ptr to an oop preparatory to scanning
6495 oop obj = oop(ptr);
6496 // Ignore mark word in verification below, since we
6497 // may be running concurrent with mutators.
6498 assert(oopDesc::is_oop(obj, true), "should be an oop");
6499 assert(_finger <= ptr, "_finger runneth ahead");
6500 // advance the finger to right end of this object
6501 _finger = ptr + obj->size();
6502 assert(_finger > ptr, "we just incremented it above");
6503 // On large heaps, it may take us some time to get through
6504 // the marking phase. During
6505 // this time it's possible that a lot of mutations have
6506 // accumulated in the card table and the mod union table --
6507 // these mutation records are redundant until we have
6508 // actually traced into the corresponding card.
6509 // Here, we check whether advancing the finger would make
6510 // us cross into a new card, and if so clear corresponding
6511 // cards in the MUT (preclean them in the card-table in the
6512 // future).
6513
6514 // The clean-on-enter optimization is disabled by default,
6515 // until we fix 6178663.
6516 if (CMSCleanOnEnter && (_finger > _threshold)) {
6517 // [_threshold, _finger) represents the interval
6518 // of cards to be cleared in MUT (or precleaned in card table).
6519 // The set of cards to be cleared is all those that overlap
6520 // with the interval [_threshold, _finger); note that
6521 // _threshold is always kept card-aligned but _finger isn't
6522 // always card-aligned.
6523 HeapWord* old_threshold = _threshold;
6524 assert(is_aligned(old_threshold, CardTable::card_size),
6525 "_threshold should always be card-aligned");
6526 _threshold = align_up(_finger, CardTable::card_size);
6527 MemRegion mr(old_threshold, _threshold);
6528 assert(!mr.is_empty(), "Control point invariant");
6529 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6530 _mut->clear_range(mr);
6531 }
6532
6533 // Note: the local finger doesn't advance while we drain
6534 // the stack below, but the global finger sure can and will.
6535 HeapWord* volatile* gfa = _task->global_finger_addr();
6536 ParPushOrMarkClosure pushOrMarkClosure(_collector,
6537 _span, _bit_map,
6538 _work_queue,
6539 _overflow_stack,
6540 _finger,
6541 gfa, this);
6542 bool res = _work_queue->push(obj); // overflow could occur here
6543 assert(res, "Will hold once we use workqueues");
6544 while (true) {
6545 oop new_oop;
6546 if (!_work_queue->pop_local(new_oop)) {
6547 // We emptied our work_queue; check if there's stuff that can
6548 // be gotten from the overflow stack.
6549 if (CMSConcMarkingTask::get_work_from_overflow_stack(
6550 _overflow_stack, _work_queue)) {
6551 do_yield_check();
6552 continue;
6553 } else { // done
6554 break;
6555 }
6556 }
6557 // Skip verifying header mark word below because we are
6558 // running concurrent with mutators.
6559 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop");
6560 // now scan this oop's oops
6561 new_oop->oop_iterate(&pushOrMarkClosure);
6562 do_yield_check();
6563 }
6564 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6565 }
6566
6567 // Yield in response to a request from VM Thread or
6568 // from mutators.
6569 void ParMarkFromRootsClosure::do_yield_work() {
6570 assert(_task != NULL, "sanity");
6571 _task->yield();
6572 }
6573
6574 // A variant of the above used for verifying CMS marking work.
6575 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6576 MemRegion span,
6577 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6578 CMSMarkStack* mark_stack):
6579 _collector(collector),
6580 _span(span),
6581 _verification_bm(verification_bm),
6582 _cms_bm(cms_bm),
6583 _mark_stack(mark_stack),
6584 _pam_verify_closure(collector, span, verification_bm, cms_bm,
6585 mark_stack)
6586 {
6587 assert(_mark_stack->isEmpty(), "stack should be empty");
6588 _finger = _verification_bm->startWord();
6589 assert(_collector->_restart_addr == NULL, "Sanity check");
6590 assert(_span.contains(_finger), "Out of bounds _finger?");
6591 }
6592
6593 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6594 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6595 assert(_span.contains(addr), "Out of bounds _finger?");
6596 _finger = addr;
6597 }
6598
6599 // Should revisit to see if this should be restructured for
6600 // greater efficiency.
6601 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6602 // convert offset into a HeapWord*
6603 HeapWord* addr = _verification_bm->startWord() + offset;
6604 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6605 "address out of range");
6606 assert(_verification_bm->isMarked(addr), "tautology");
6607 assert(_cms_bm->isMarked(addr), "tautology");
6608
6609 assert(_mark_stack->isEmpty(),
6610 "should drain stack to limit stack usage");
6611 // convert addr to an oop preparatory to scanning
6612 oop obj = oop(addr);
6613 assert(oopDesc::is_oop(obj), "should be an oop");
6614 assert(_finger <= addr, "_finger runneth ahead");
6615 // advance the finger to right end of this object
6616 _finger = addr + obj->size();
6617 assert(_finger > addr, "we just incremented it above");
6618 // Note: the finger doesn't advance while we drain
6619 // the stack below.
6620 bool res = _mark_stack->push(obj);
6621 assert(res, "Empty non-zero size stack should have space for single push");
6622 while (!_mark_stack->isEmpty()) {
6623 oop new_oop = _mark_stack->pop();
6624 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop");
6625 // now scan this oop's oops
6626 new_oop->oop_iterate(&_pam_verify_closure);
6627 }
6628 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6629 return true;
6630 }
6631
6632 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6633 CMSCollector* collector, MemRegion span,
6634 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6635 CMSMarkStack* mark_stack):
6636 MetadataAwareOopClosure(collector->ref_processor()),
6637 _collector(collector),
6638 _span(span),
6639 _verification_bm(verification_bm),
6640 _cms_bm(cms_bm),
6641 _mark_stack(mark_stack)
6642 { }
6643
6644 template <class T> void PushAndMarkVerifyClosure::do_oop_work(T *p) {
6645 oop obj = RawAccess<>::oop_load(p);
6646 do_oop(obj);
6647 }
6648
6649 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6650 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6651
6652 // Upon stack overflow, we discard (part of) the stack,
6653 // remembering the least address amongst those discarded
6654 // in CMSCollector's _restart_address.
6655 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6656 // Remember the least grey address discarded
6657 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6658 _collector->lower_restart_addr(ra);
6659 _mark_stack->reset(); // discard stack contents
6660 _mark_stack->expand(); // expand the stack if possible
6661 }
6662
6663 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6664 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6665 HeapWord* addr = (HeapWord*)obj;
6666 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6667 // Oop lies in _span and isn't yet grey or black
6668 _verification_bm->mark(addr); // now grey
6669 if (!_cms_bm->isMarked(addr)) {
6670 Log(gc, verify) log;
6671 ResourceMark rm;
6672 LogStream ls(log.error());
6673 oop(addr)->print_on(&ls);
6674 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6675 fatal("... aborting");
6676 }
6677
6678 if (!_mark_stack->push(obj)) { // stack overflow
6679 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity());
6680 assert(_mark_stack->isFull(), "Else push should have succeeded");
6681 handle_stack_overflow(addr);
6682 }
6683 // anything including and to the right of _finger
6684 // will be scanned as we iterate over the remainder of the
6685 // bit map
6686 }
6687 }
6688
6689 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6690 MemRegion span,
6691 CMSBitMap* bitMap, CMSMarkStack* markStack,
6692 HeapWord* finger, MarkFromRootsClosure* parent) :
6693 MetadataAwareOopClosure(collector->ref_processor()),
6694 _collector(collector),
6695 _span(span),
6696 _bitMap(bitMap),
6697 _markStack(markStack),
6698 _finger(finger),
6699 _parent(parent)
6700 { }
6701
6702 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector,
6703 MemRegion span,
6704 CMSBitMap* bit_map,
6705 OopTaskQueue* work_queue,
6706 CMSMarkStack* overflow_stack,
6707 HeapWord* finger,
6708 HeapWord* volatile* global_finger_addr,
6709 ParMarkFromRootsClosure* parent) :
6710 MetadataAwareOopClosure(collector->ref_processor()),
6711 _collector(collector),
6712 _whole_span(collector->_span),
6713 _span(span),
6714 _bit_map(bit_map),
6715 _work_queue(work_queue),
6716 _overflow_stack(overflow_stack),
6717 _finger(finger),
6718 _global_finger_addr(global_finger_addr),
6719 _parent(parent)
6720 { }
6721
6722 // Assumes thread-safe access by callers, who are
6723 // responsible for mutual exclusion.
6724 void CMSCollector::lower_restart_addr(HeapWord* low) {
6725 assert(_span.contains(low), "Out of bounds addr");
6726 if (_restart_addr == NULL) {
6727 _restart_addr = low;
6728 } else {
6729 _restart_addr = MIN2(_restart_addr, low);
6730 }
6731 }
6732
6733 // Upon stack overflow, we discard (part of) the stack,
6734 // remembering the least address amongst those discarded
6735 // in CMSCollector's _restart_address.
6736 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6737 // Remember the least grey address discarded
6738 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6739 _collector->lower_restart_addr(ra);
6740 _markStack->reset(); // discard stack contents
6741 _markStack->expand(); // expand the stack if possible
6742 }
6743
6744 // Upon stack overflow, we discard (part of) the stack,
6745 // remembering the least address amongst those discarded
6746 // in CMSCollector's _restart_address.
6747 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6748 // We need to do this under a mutex to prevent other
6749 // workers from interfering with the work done below.
6750 MutexLockerEx ml(_overflow_stack->par_lock(),
6751 Mutex::_no_safepoint_check_flag);
6752 // Remember the least grey address discarded
6753 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
6754 _collector->lower_restart_addr(ra);
6755 _overflow_stack->reset(); // discard stack contents
6756 _overflow_stack->expand(); // expand the stack if possible
6757 }
6758
6759 void PushOrMarkClosure::do_oop(oop obj) {
6760 // Ignore mark word because we are running concurrent with mutators.
6761 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6762 HeapWord* addr = (HeapWord*)obj;
6763 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
6764 // Oop lies in _span and isn't yet grey or black
6765 _bitMap->mark(addr); // now grey
6766 if (addr < _finger) {
6767 // the bit map iteration has already either passed, or
6768 // sampled, this bit in the bit map; we'll need to
6769 // use the marking stack to scan this oop's oops.
6770 bool simulate_overflow = false;
6771 NOT_PRODUCT(
6772 if (CMSMarkStackOverflowALot &&
6773 _collector->simulate_overflow()) {
6774 // simulate a stack overflow
6775 simulate_overflow = true;
6776 }
6777 )
6778 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
6779 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity());
6780 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
6781 handle_stack_overflow(addr);
6782 }
6783 }
6784 // anything including and to the right of _finger
6785 // will be scanned as we iterate over the remainder of the
6786 // bit map
6787 do_yield_check();
6788 }
6789 }
6790
6791 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
6792 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
6793
6794 void ParPushOrMarkClosure::do_oop(oop obj) {
6795 // Ignore mark word because we are running concurrent with mutators.
6796 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6797 HeapWord* addr = (HeapWord*)obj;
6798 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
6799 // Oop lies in _span and isn't yet grey or black
6800 // We read the global_finger (volatile read) strictly after marking oop
6801 bool res = _bit_map->par_mark(addr); // now grey
6802 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
6803 // Should we push this marked oop on our stack?
6804 // -- if someone else marked it, nothing to do
6805 // -- if target oop is above global finger nothing to do
6806 // -- if target oop is in chunk and above local finger
6807 // then nothing to do
6808 // -- else push on work queue
6809 if ( !res // someone else marked it, they will deal with it
6810 || (addr >= *gfa) // will be scanned in a later task
6811 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
6812 return;
6813 }
6814 // the bit map iteration has already either passed, or
6815 // sampled, this bit in the bit map; we'll need to
6816 // use the marking stack to scan this oop's oops.
6817 bool simulate_overflow = false;
6818 NOT_PRODUCT(
6819 if (CMSMarkStackOverflowALot &&
6820 _collector->simulate_overflow()) {
6821 // simulate a stack overflow
6822 simulate_overflow = true;
6823 }
6824 )
6825 if (simulate_overflow ||
6826 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
6827 // stack overflow
6828 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
6829 // We cannot assert that the overflow stack is full because
6830 // it may have been emptied since.
6831 assert(simulate_overflow ||
6832 _work_queue->size() == _work_queue->max_elems(),
6833 "Else push should have succeeded");
6834 handle_stack_overflow(addr);
6835 }
6836 do_yield_check();
6837 }
6838 }
6839
6840 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); }
6841 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); }
6842
6843 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
6844 MemRegion span,
6845 ReferenceProcessor* rp,
6846 CMSBitMap* bit_map,
6847 CMSBitMap* mod_union_table,
6848 CMSMarkStack* mark_stack,
6849 bool concurrent_precleaning):
6850 MetadataAwareOopClosure(rp),
6851 _collector(collector),
6852 _span(span),
6853 _bit_map(bit_map),
6854 _mod_union_table(mod_union_table),
6855 _mark_stack(mark_stack),
6856 _concurrent_precleaning(concurrent_precleaning)
6857 {
6858 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL");
6859 }
6860
6861 // Grey object rescan during pre-cleaning and second checkpoint phases --
6862 // the non-parallel version (the parallel version appears further below.)
6863 void PushAndMarkClosure::do_oop(oop obj) {
6864 // Ignore mark word verification. If during concurrent precleaning,
6865 // the object monitor may be locked. If during the checkpoint
6866 // phases, the object may already have been reached by a different
6867 // path and may be at the end of the global overflow list (so
6868 // the mark word may be NULL).
6869 assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */),
6870 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6871 HeapWord* addr = (HeapWord*)obj;
6872 // Check if oop points into the CMS generation
6873 // and is not marked
6874 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6875 // a white object ...
6876 _bit_map->mark(addr); // ... now grey
6877 // push on the marking stack (grey set)
6878 bool simulate_overflow = false;
6879 NOT_PRODUCT(
6880 if (CMSMarkStackOverflowALot &&
6881 _collector->simulate_overflow()) {
6882 // simulate a stack overflow
6883 simulate_overflow = true;
6884 }
6885 )
6886 if (simulate_overflow || !_mark_stack->push(obj)) {
6887 if (_concurrent_precleaning) {
6888 // During precleaning we can just dirty the appropriate card(s)
6889 // in the mod union table, thus ensuring that the object remains
6890 // in the grey set and continue. In the case of object arrays
6891 // we need to dirty all of the cards that the object spans,
6892 // since the rescan of object arrays will be limited to the
6893 // dirty cards.
6894 // Note that no one can be interfering with us in this action
6895 // of dirtying the mod union table, so no locking or atomics
6896 // are required.
6897 if (obj->is_objArray()) {
6898 size_t sz = obj->size();
6899 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size);
6900 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6901 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6902 _mod_union_table->mark_range(redirty_range);
6903 } else {
6904 _mod_union_table->mark(addr);
6905 }
6906 _collector->_ser_pmc_preclean_ovflw++;
6907 } else {
6908 // During the remark phase, we need to remember this oop
6909 // in the overflow list.
6910 _collector->push_on_overflow_list(obj);
6911 _collector->_ser_pmc_remark_ovflw++;
6912 }
6913 }
6914 }
6915 }
6916
6917 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector,
6918 MemRegion span,
6919 ReferenceProcessor* rp,
6920 CMSBitMap* bit_map,
6921 OopTaskQueue* work_queue):
6922 MetadataAwareOopClosure(rp),
6923 _collector(collector),
6924 _span(span),
6925 _bit_map(bit_map),
6926 _work_queue(work_queue)
6927 {
6928 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL");
6929 }
6930
6931 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
6932 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
6933
6934 // Grey object rescan during second checkpoint phase --
6935 // the parallel version.
6936 void ParPushAndMarkClosure::do_oop(oop obj) {
6937 // In the assert below, we ignore the mark word because
6938 // this oop may point to an already visited object that is
6939 // on the overflow stack (in which case the mark word has
6940 // been hijacked for chaining into the overflow stack --
6941 // if this is the last object in the overflow stack then
6942 // its mark word will be NULL). Because this object may
6943 // have been subsequently popped off the global overflow
6944 // stack, and the mark word possibly restored to the prototypical
6945 // value, by the time we get to examined this failing assert in
6946 // the debugger, is_oop_or_null(false) may subsequently start
6947 // to hold.
6948 assert(oopDesc::is_oop_or_null(obj, true),
6949 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6950 HeapWord* addr = (HeapWord*)obj;
6951 // Check if oop points into the CMS generation
6952 // and is not marked
6953 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6954 // a white object ...
6955 // If we manage to "claim" the object, by being the
6956 // first thread to mark it, then we push it on our
6957 // marking stack
6958 if (_bit_map->par_mark(addr)) { // ... now grey
6959 // push on work queue (grey set)
6960 bool simulate_overflow = false;
6961 NOT_PRODUCT(
6962 if (CMSMarkStackOverflowALot &&
6963 _collector->par_simulate_overflow()) {
6964 // simulate a stack overflow
6965 simulate_overflow = true;
6966 }
6967 )
6968 if (simulate_overflow || !_work_queue->push(obj)) {
6969 _collector->par_push_on_overflow_list(obj);
6970 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
6971 }
6972 } // Else, some other thread got there first
6973 }
6974 }
6975
6976 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); }
6977 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); }
6978
6979 void CMSPrecleanRefsYieldClosure::do_yield_work() {
6980 Mutex* bml = _collector->bitMapLock();
6981 assert_lock_strong(bml);
6982 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6983 "CMS thread should hold CMS token");
6984
6985 bml->unlock();
6986 ConcurrentMarkSweepThread::desynchronize(true);
6987
6988 _collector->stopTimer();
6989 _collector->incrementYields();
6990
6991 // See the comment in coordinator_yield()
6992 for (unsigned i = 0; i < CMSYieldSleepCount &&
6993 ConcurrentMarkSweepThread::should_yield() &&
6994 !CMSCollector::foregroundGCIsActive(); ++i) {
6995 os::sleep(Thread::current(), 1, false);
6996 }
6997
6998 ConcurrentMarkSweepThread::synchronize(true);
6999 bml->lock();
7000
7001 _collector->startTimer();
7002 }
7003
7004 bool CMSPrecleanRefsYieldClosure::should_return() {
7005 if (ConcurrentMarkSweepThread::should_yield()) {
7006 do_yield_work();
7007 }
7008 return _collector->foregroundGCIsActive();
7009 }
7010
7011 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7012 assert(((size_t)mr.start())%CardTable::card_size_in_words == 0,
7013 "mr should be aligned to start at a card boundary");
7014 // We'd like to assert:
7015 // assert(mr.word_size()%CardTable::card_size_in_words == 0,
7016 // "mr should be a range of cards");
7017 // However, that would be too strong in one case -- the last
7018 // partition ends at _unallocated_block which, in general, can be
7019 // an arbitrary boundary, not necessarily card aligned.
7020 _num_dirty_cards += mr.word_size()/CardTable::card_size_in_words;
7021 _space->object_iterate_mem(mr, &_scan_cl);
7022 }
7023
7024 SweepClosure::SweepClosure(CMSCollector* collector,
7025 ConcurrentMarkSweepGeneration* g,
7026 CMSBitMap* bitMap, bool should_yield) :
7027 _collector(collector),
7028 _g(g),
7029 _sp(g->cmsSpace()),
7030 _limit(_sp->sweep_limit()),
7031 _freelistLock(_sp->freelistLock()),
7032 _bitMap(bitMap),
7033 _yield(should_yield),
7034 _inFreeRange(false), // No free range at beginning of sweep
7035 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7036 _lastFreeRangeCoalesced(false),
7037 _freeFinger(g->used_region().start())
7038 {
7039 NOT_PRODUCT(
7040 _numObjectsFreed = 0;
7041 _numWordsFreed = 0;
7042 _numObjectsLive = 0;
7043 _numWordsLive = 0;
7044 _numObjectsAlreadyFree = 0;
7045 _numWordsAlreadyFree = 0;
7046 _last_fc = NULL;
7047
7048 _sp->initializeIndexedFreeListArrayReturnedBytes();
7049 _sp->dictionary()->initialize_dict_returned_bytes();
7050 )
7051 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7052 "sweep _limit out of bounds");
7053 log_develop_trace(gc, sweep)("====================");
7054 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit));
7055 }
7056
7057 void SweepClosure::print_on(outputStream* st) const {
7058 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7059 p2i(_sp->bottom()), p2i(_sp->end()));
7060 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7061 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7062 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7063 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7064 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7065 }
7066
7067 #ifndef PRODUCT
7068 // Assertion checking only: no useful work in product mode --
7069 // however, if any of the flags below become product flags,
7070 // you may need to review this code to see if it needs to be
7071 // enabled in product mode.
7072 SweepClosure::~SweepClosure() {
7073 assert_lock_strong(_freelistLock);
7074 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7075 "sweep _limit out of bounds");
7076 if (inFreeRange()) {
7077 Log(gc, sweep) log;
7078 log.error("inFreeRange() should have been reset; dumping state of SweepClosure");
7079 ResourceMark rm;
7080 LogStream ls(log.error());
7081 print_on(&ls);
7082 ShouldNotReachHere();
7083 }
7084
7085 if (log_is_enabled(Debug, gc, sweep)) {
7086 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7087 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7088 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7089 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7090 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord);
7091 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes);
7092 }
7093
7094 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) {
7095 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7096 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7097 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7098 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes",
7099 returned_bytes, indexListReturnedBytes, dict_returned_bytes);
7100 }
7101 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit));
7102 log_develop_trace(gc, sweep)("================");
7103 }
7104 #endif // PRODUCT
7105
7106 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7107 bool freeRangeInFreeLists) {
7108 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)",
7109 p2i(freeFinger), freeRangeInFreeLists);
7110 assert(!inFreeRange(), "Trampling existing free range");
7111 set_inFreeRange(true);
7112 set_lastFreeRangeCoalesced(false);
7113
7114 set_freeFinger(freeFinger);
7115 set_freeRangeInFreeLists(freeRangeInFreeLists);
7116 if (CMSTestInFreeList) {
7117 if (freeRangeInFreeLists) {
7118 FreeChunk* fc = (FreeChunk*) freeFinger;
7119 assert(fc->is_free(), "A chunk on the free list should be free.");
7120 assert(fc->size() > 0, "Free range should have a size");
7121 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7122 }
7123 }
7124 }
7125
7126 // Note that the sweeper runs concurrently with mutators. Thus,
7127 // it is possible for direct allocation in this generation to happen
7128 // in the middle of the sweep. Note that the sweeper also coalesces
7129 // contiguous free blocks. Thus, unless the sweeper and the allocator
7130 // synchronize appropriately freshly allocated blocks may get swept up.
7131 // This is accomplished by the sweeper locking the free lists while
7132 // it is sweeping. Thus blocks that are determined to be free are
7133 // indeed free. There is however one additional complication:
7134 // blocks that have been allocated since the final checkpoint and
7135 // mark, will not have been marked and so would be treated as
7136 // unreachable and swept up. To prevent this, the allocator marks
7137 // the bit map when allocating during the sweep phase. This leads,
7138 // however, to a further complication -- objects may have been allocated
7139 // but not yet initialized -- in the sense that the header isn't yet
7140 // installed. The sweeper can not then determine the size of the block
7141 // in order to skip over it. To deal with this case, we use a technique
7142 // (due to Printezis) to encode such uninitialized block sizes in the
7143 // bit map. Since the bit map uses a bit per every HeapWord, but the
7144 // CMS generation has a minimum object size of 3 HeapWords, it follows
7145 // that "normal marks" won't be adjacent in the bit map (there will
7146 // always be at least two 0 bits between successive 1 bits). We make use
7147 // of these "unused" bits to represent uninitialized blocks -- the bit
7148 // corresponding to the start of the uninitialized object and the next
7149 // bit are both set. Finally, a 1 bit marks the end of the object that
7150 // started with the two consecutive 1 bits to indicate its potentially
7151 // uninitialized state.
7152
7153 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7154 FreeChunk* fc = (FreeChunk*)addr;
7155 size_t res;
7156
7157 // Check if we are done sweeping. Below we check "addr >= _limit" rather
7158 // than "addr == _limit" because although _limit was a block boundary when
7159 // we started the sweep, it may no longer be one because heap expansion
7160 // may have caused us to coalesce the block ending at the address _limit
7161 // with a newly expanded chunk (this happens when _limit was set to the
7162 // previous _end of the space), so we may have stepped past _limit:
7163 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7164 if (addr >= _limit) { // we have swept up to or past the limit: finish up
7165 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7166 "sweep _limit out of bounds");
7167 assert(addr < _sp->end(), "addr out of bounds");
7168 // Flush any free range we might be holding as a single
7169 // coalesced chunk to the appropriate free list.
7170 if (inFreeRange()) {
7171 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7172 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger()));
7173 flush_cur_free_chunk(freeFinger(),
7174 pointer_delta(addr, freeFinger()));
7175 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]",
7176 p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7177 lastFreeRangeCoalesced() ? 1 : 0);
7178 }
7179
7180 // help the iterator loop finish
7181 return pointer_delta(_sp->end(), addr);
7182 }
7183
7184 assert(addr < _limit, "sweep invariant");
7185 // check if we should yield
7186 do_yield_check(addr);
7187 if (fc->is_free()) {
7188 // Chunk that is already free
7189 res = fc->size();
7190 do_already_free_chunk(fc);
7191 debug_only(_sp->verifyFreeLists());
7192 // If we flush the chunk at hand in lookahead_and_flush()
7193 // and it's coalesced with a preceding chunk, then the
7194 // process of "mangling" the payload of the coalesced block
7195 // will cause erasure of the size information from the
7196 // (erstwhile) header of all the coalesced blocks but the
7197 // first, so the first disjunct in the assert will not hold
7198 // in that specific case (in which case the second disjunct
7199 // will hold).
7200 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7201 "Otherwise the size info doesn't change at this step");
7202 NOT_PRODUCT(
7203 _numObjectsAlreadyFree++;
7204 _numWordsAlreadyFree += res;
7205 )
7206 NOT_PRODUCT(_last_fc = fc;)
7207 } else if (!_bitMap->isMarked(addr)) {
7208 // Chunk is fresh garbage
7209 res = do_garbage_chunk(fc);
7210 debug_only(_sp->verifyFreeLists());
7211 NOT_PRODUCT(
7212 _numObjectsFreed++;
7213 _numWordsFreed += res;
7214 )
7215 } else {
7216 // Chunk that is alive.
7217 res = do_live_chunk(fc);
7218 debug_only(_sp->verifyFreeLists());
7219 NOT_PRODUCT(
7220 _numObjectsLive++;
7221 _numWordsLive += res;
7222 )
7223 }
7224 return res;
7225 }
7226
7227 // For the smart allocation, record following
7228 // split deaths - a free chunk is removed from its free list because
7229 // it is being split into two or more chunks.
7230 // split birth - a free chunk is being added to its free list because
7231 // a larger free chunk has been split and resulted in this free chunk.
7232 // coal death - a free chunk is being removed from its free list because
7233 // it is being coalesced into a large free chunk.
7234 // coal birth - a free chunk is being added to its free list because
7235 // it was created when two or more free chunks where coalesced into
7236 // this free chunk.
7237 //
7238 // These statistics are used to determine the desired number of free
7239 // chunks of a given size. The desired number is chosen to be relative
7240 // to the end of a CMS sweep. The desired number at the end of a sweep
7241 // is the
7242 // count-at-end-of-previous-sweep (an amount that was enough)
7243 // - count-at-beginning-of-current-sweep (the excess)
7244 // + split-births (gains in this size during interval)
7245 // - split-deaths (demands on this size during interval)
7246 // where the interval is from the end of one sweep to the end of the
7247 // next.
7248 //
7249 // When sweeping the sweeper maintains an accumulated chunk which is
7250 // the chunk that is made up of chunks that have been coalesced. That
7251 // will be termed the left-hand chunk. A new chunk of garbage that
7252 // is being considered for coalescing will be referred to as the
7253 // right-hand chunk.
7254 //
7255 // When making a decision on whether to coalesce a right-hand chunk with
7256 // the current left-hand chunk, the current count vs. the desired count
7257 // of the left-hand chunk is considered. Also if the right-hand chunk
7258 // is near the large chunk at the end of the heap (see
7259 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7260 // left-hand chunk is coalesced.
7261 //
7262 // When making a decision about whether to split a chunk, the desired count
7263 // vs. the current count of the candidate to be split is also considered.
7264 // If the candidate is underpopulated (currently fewer chunks than desired)
7265 // a chunk of an overpopulated (currently more chunks than desired) size may
7266 // be chosen. The "hint" associated with a free list, if non-null, points
7267 // to a free list which may be overpopulated.
7268 //
7269
7270 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7271 const size_t size = fc->size();
7272 // Chunks that cannot be coalesced are not in the
7273 // free lists.
7274 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7275 assert(_sp->verify_chunk_in_free_list(fc),
7276 "free chunk should be in free lists");
7277 }
7278 // a chunk that is already free, should not have been
7279 // marked in the bit map
7280 HeapWord* const addr = (HeapWord*) fc;
7281 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7282 // Verify that the bit map has no bits marked between
7283 // addr and purported end of this block.
7284 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7285
7286 // Some chunks cannot be coalesced under any circumstances.
7287 // See the definition of cantCoalesce().
7288 if (!fc->cantCoalesce()) {
7289 // This chunk can potentially be coalesced.
7290 // All the work is done in
7291 do_post_free_or_garbage_chunk(fc, size);
7292 // Note that if the chunk is not coalescable (the else arm
7293 // below), we unconditionally flush, without needing to do
7294 // a "lookahead," as we do below.
7295 if (inFreeRange()) lookahead_and_flush(fc, size);
7296 } else {
7297 // Code path common to both original and adaptive free lists.
7298
7299 // cant coalesce with previous block; this should be treated
7300 // as the end of a free run if any
7301 if (inFreeRange()) {
7302 // we kicked some butt; time to pick up the garbage
7303 assert(freeFinger() < addr, "freeFinger points too high");
7304 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7305 }
7306 // else, nothing to do, just continue
7307 }
7308 }
7309
7310 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7311 // This is a chunk of garbage. It is not in any free list.
7312 // Add it to a free list or let it possibly be coalesced into
7313 // a larger chunk.
7314 HeapWord* const addr = (HeapWord*) fc;
7315 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7316
7317 // Verify that the bit map has no bits marked between
7318 // addr and purported end of just dead object.
7319 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7320 do_post_free_or_garbage_chunk(fc, size);
7321
7322 assert(_limit >= addr + size,
7323 "A freshly garbage chunk can't possibly straddle over _limit");
7324 if (inFreeRange()) lookahead_and_flush(fc, size);
7325 return size;
7326 }
7327
7328 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7329 HeapWord* addr = (HeapWord*) fc;
7330 // The sweeper has just found a live object. Return any accumulated
7331 // left hand chunk to the free lists.
7332 if (inFreeRange()) {
7333 assert(freeFinger() < addr, "freeFinger points too high");
7334 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7335 }
7336
7337 // This object is live: we'd normally expect this to be
7338 // an oop, and like to assert the following:
7339 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop");
7340 // However, as we commented above, this may be an object whose
7341 // header hasn't yet been initialized.
7342 size_t size;
7343 assert(_bitMap->isMarked(addr), "Tautology for this control point");
7344 if (_bitMap->isMarked(addr + 1)) {
7345 // Determine the size from the bit map, rather than trying to
7346 // compute it from the object header.
7347 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7348 size = pointer_delta(nextOneAddr + 1, addr);
7349 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7350 "alignment problem");
7351
7352 #ifdef ASSERT
7353 if (oop(addr)->klass_or_null_acquire() != NULL) {
7354 // Ignore mark word because we are running concurrent with mutators
7355 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop");
7356 assert(size ==
7357 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7358 "P-mark and computed size do not agree");
7359 }
7360 #endif
7361
7362 } else {
7363 // This should be an initialized object that's alive.
7364 assert(oop(addr)->klass_or_null_acquire() != NULL,
7365 "Should be an initialized object");
7366 // Ignore mark word because we are running concurrent with mutators
7367 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop");
7368 // Verify that the bit map has no bits marked between
7369 // addr and purported end of this block.
7370 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7371 assert(size >= 3, "Necessary for Printezis marks to work");
7372 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7373 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7374 }
7375 return size;
7376 }
7377
7378 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7379 size_t chunkSize) {
7380 // do_post_free_or_garbage_chunk() should only be called in the case
7381 // of the adaptive free list allocator.
7382 const bool fcInFreeLists = fc->is_free();
7383 assert((HeapWord*)fc <= _limit, "sweep invariant");
7384 if (CMSTestInFreeList && fcInFreeLists) {
7385 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7386 }
7387
7388 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7389
7390 HeapWord* const fc_addr = (HeapWord*) fc;
7391
7392 bool coalesce = false;
7393 const size_t left = pointer_delta(fc_addr, freeFinger());
7394 const size_t right = chunkSize;
7395 switch (FLSCoalescePolicy) {
7396 // numeric value forms a coalition aggressiveness metric
7397 case 0: { // never coalesce
7398 coalesce = false;
7399 break;
7400 }
7401 case 1: { // coalesce if left & right chunks on overpopulated lists
7402 coalesce = _sp->coalOverPopulated(left) &&
7403 _sp->coalOverPopulated(right);
7404 break;
7405 }
7406 case 2: { // coalesce if left chunk on overpopulated list (default)
7407 coalesce = _sp->coalOverPopulated(left);
7408 break;
7409 }
7410 case 3: { // coalesce if left OR right chunk on overpopulated list
7411 coalesce = _sp->coalOverPopulated(left) ||
7412 _sp->coalOverPopulated(right);
7413 break;
7414 }
7415 case 4: { // always coalesce
7416 coalesce = true;
7417 break;
7418 }
7419 default:
7420 ShouldNotReachHere();
7421 }
7422
7423 // Should the current free range be coalesced?
7424 // If the chunk is in a free range and either we decided to coalesce above
7425 // or the chunk is near the large block at the end of the heap
7426 // (isNearLargestChunk() returns true), then coalesce this chunk.
7427 const bool doCoalesce = inFreeRange()
7428 && (coalesce || _g->isNearLargestChunk(fc_addr));
7429 if (doCoalesce) {
7430 // Coalesce the current free range on the left with the new
7431 // chunk on the right. If either is on a free list,
7432 // it must be removed from the list and stashed in the closure.
7433 if (freeRangeInFreeLists()) {
7434 FreeChunk* const ffc = (FreeChunk*)freeFinger();
7435 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7436 "Size of free range is inconsistent with chunk size.");
7437 if (CMSTestInFreeList) {
7438 assert(_sp->verify_chunk_in_free_list(ffc),
7439 "Chunk is not in free lists");
7440 }
7441 _sp->coalDeath(ffc->size());
7442 _sp->removeFreeChunkFromFreeLists(ffc);
7443 set_freeRangeInFreeLists(false);
7444 }
7445 if (fcInFreeLists) {
7446 _sp->coalDeath(chunkSize);
7447 assert(fc->size() == chunkSize,
7448 "The chunk has the wrong size or is not in the free lists");
7449 _sp->removeFreeChunkFromFreeLists(fc);
7450 }
7451 set_lastFreeRangeCoalesced(true);
7452 print_free_block_coalesced(fc);
7453 } else { // not in a free range and/or should not coalesce
7454 // Return the current free range and start a new one.
7455 if (inFreeRange()) {
7456 // In a free range but cannot coalesce with the right hand chunk.
7457 // Put the current free range into the free lists.
7458 flush_cur_free_chunk(freeFinger(),
7459 pointer_delta(fc_addr, freeFinger()));
7460 }
7461 // Set up for new free range. Pass along whether the right hand
7462 // chunk is in the free lists.
7463 initialize_free_range((HeapWord*)fc, fcInFreeLists);
7464 }
7465 }
7466
7467 // Lookahead flush:
7468 // If we are tracking a free range, and this is the last chunk that
7469 // we'll look at because its end crosses past _limit, we'll preemptively
7470 // flush it along with any free range we may be holding on to. Note that
7471 // this can be the case only for an already free or freshly garbage
7472 // chunk. If this block is an object, it can never straddle
7473 // over _limit. The "straddling" occurs when _limit is set at
7474 // the previous end of the space when this cycle started, and
7475 // a subsequent heap expansion caused the previously co-terminal
7476 // free block to be coalesced with the newly expanded portion,
7477 // thus rendering _limit a non-block-boundary making it dangerous
7478 // for the sweeper to step over and examine.
7479 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7480 assert(inFreeRange(), "Should only be called if currently in a free range.");
7481 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7482 assert(_sp->used_region().contains(eob - 1),
7483 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7484 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7485 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7486 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size);
7487 if (eob >= _limit) {
7488 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7489 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block "
7490 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7491 "[" PTR_FORMAT "," PTR_FORMAT ")",
7492 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7493 // Return the storage we are tracking back into the free lists.
7494 log_develop_trace(gc, sweep)("Flushing ... ");
7495 assert(freeFinger() < eob, "Error");
7496 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7497 }
7498 }
7499
7500 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7501 assert(inFreeRange(), "Should only be called if currently in a free range.");
7502 assert(size > 0,
7503 "A zero sized chunk cannot be added to the free lists.");
7504 if (!freeRangeInFreeLists()) {
7505 if (CMSTestInFreeList) {
7506 FreeChunk* fc = (FreeChunk*) chunk;
7507 fc->set_size(size);
7508 assert(!_sp->verify_chunk_in_free_list(fc),
7509 "chunk should not be in free lists yet");
7510 }
7511 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size);
7512 // A new free range is going to be starting. The current
7513 // free range has not been added to the free lists yet or
7514 // was removed so add it back.
7515 // If the current free range was coalesced, then the death
7516 // of the free range was recorded. Record a birth now.
7517 if (lastFreeRangeCoalesced()) {
7518 _sp->coalBirth(size);
7519 }
7520 _sp->addChunkAndRepairOffsetTable(chunk, size,
7521 lastFreeRangeCoalesced());
7522 } else {
7523 log_develop_trace(gc, sweep)("Already in free list: nothing to flush");
7524 }
7525 set_inFreeRange(false);
7526 set_freeRangeInFreeLists(false);
7527 }
7528
7529 // We take a break if we've been at this for a while,
7530 // so as to avoid monopolizing the locks involved.
7531 void SweepClosure::do_yield_work(HeapWord* addr) {
7532 // Return current free chunk being used for coalescing (if any)
7533 // to the appropriate freelist. After yielding, the next
7534 // free block encountered will start a coalescing range of
7535 // free blocks. If the next free block is adjacent to the
7536 // chunk just flushed, they will need to wait for the next
7537 // sweep to be coalesced.
7538 if (inFreeRange()) {
7539 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7540 }
7541
7542 // First give up the locks, then yield, then re-lock.
7543 // We should probably use a constructor/destructor idiom to
7544 // do this unlock/lock or modify the MutexUnlocker class to
7545 // serve our purpose. XXX
7546 assert_lock_strong(_bitMap->lock());
7547 assert_lock_strong(_freelistLock);
7548 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7549 "CMS thread should hold CMS token");
7550 _bitMap->lock()->unlock();
7551 _freelistLock->unlock();
7552 ConcurrentMarkSweepThread::desynchronize(true);
7553 _collector->stopTimer();
7554 _collector->incrementYields();
7555
7556 // See the comment in coordinator_yield()
7557 for (unsigned i = 0; i < CMSYieldSleepCount &&
7558 ConcurrentMarkSweepThread::should_yield() &&
7559 !CMSCollector::foregroundGCIsActive(); ++i) {
7560 os::sleep(Thread::current(), 1, false);
7561 }
7562
7563 ConcurrentMarkSweepThread::synchronize(true);
7564 _freelistLock->lock();
7565 _bitMap->lock()->lock_without_safepoint_check();
7566 _collector->startTimer();
7567 }
7568
7569 #ifndef PRODUCT
7570 // This is actually very useful in a product build if it can
7571 // be called from the debugger. Compile it into the product
7572 // as needed.
7573 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7574 return debug_cms_space->verify_chunk_in_free_list(fc);
7575 }
7576 #endif
7577
7578 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7579 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7580 p2i(fc), fc->size());
7581 }
7582
7583 // CMSIsAliveClosure
7584 bool CMSIsAliveClosure::do_object_b(oop obj) {
7585 HeapWord* addr = (HeapWord*)obj;
7586 return addr != NULL &&
7587 (!_span.contains(addr) || _bit_map->isMarked(addr));
7588 }
7589
7590
7591 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7592 MemRegion span,
7593 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7594 bool cpc):
7595 _collector(collector),
7596 _span(span),
7597 _bit_map(bit_map),
7598 _mark_stack(mark_stack),
7599 _concurrent_precleaning(cpc) {
7600 assert(!_span.is_empty(), "Empty span could spell trouble");
7601 }
7602
7603
7604 // CMSKeepAliveClosure: the serial version
7605 void CMSKeepAliveClosure::do_oop(oop obj) {
7606 HeapWord* addr = (HeapWord*)obj;
7607 if (_span.contains(addr) &&
7608 !_bit_map->isMarked(addr)) {
7609 _bit_map->mark(addr);
7610 bool simulate_overflow = false;
7611 NOT_PRODUCT(
7612 if (CMSMarkStackOverflowALot &&
7613 _collector->simulate_overflow()) {
7614 // simulate a stack overflow
7615 simulate_overflow = true;
7616 }
7617 )
7618 if (simulate_overflow || !_mark_stack->push(obj)) {
7619 if (_concurrent_precleaning) {
7620 // We dirty the overflown object and let the remark
7621 // phase deal with it.
7622 assert(_collector->overflow_list_is_empty(), "Error");
7623 // In the case of object arrays, we need to dirty all of
7624 // the cards that the object spans. No locking or atomics
7625 // are needed since no one else can be mutating the mod union
7626 // table.
7627 if (obj->is_objArray()) {
7628 size_t sz = obj->size();
7629 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size);
7630 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7631 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7632 _collector->_modUnionTable.mark_range(redirty_range);
7633 } else {
7634 _collector->_modUnionTable.mark(addr);
7635 }
7636 _collector->_ser_kac_preclean_ovflw++;
7637 } else {
7638 _collector->push_on_overflow_list(obj);
7639 _collector->_ser_kac_ovflw++;
7640 }
7641 }
7642 }
7643 }
7644
7645 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7646 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7647
7648 // CMSParKeepAliveClosure: a parallel version of the above.
7649 // The work queues are private to each closure (thread),
7650 // but (may be) available for stealing by other threads.
7651 void CMSParKeepAliveClosure::do_oop(oop obj) {
7652 HeapWord* addr = (HeapWord*)obj;
7653 if (_span.contains(addr) &&
7654 !_bit_map->isMarked(addr)) {
7655 // In general, during recursive tracing, several threads
7656 // may be concurrently getting here; the first one to
7657 // "tag" it, claims it.
7658 if (_bit_map->par_mark(addr)) {
7659 bool res = _work_queue->push(obj);
7660 assert(res, "Low water mark should be much less than capacity");
7661 // Do a recursive trim in the hope that this will keep
7662 // stack usage lower, but leave some oops for potential stealers
7663 trim_queue(_low_water_mark);
7664 } // Else, another thread got there first
7665 }
7666 }
7667
7668 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7669 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7670
7671 void CMSParKeepAliveClosure::trim_queue(uint max) {
7672 while (_work_queue->size() > max) {
7673 oop new_oop;
7674 if (_work_queue->pop_local(new_oop)) {
7675 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
7676 assert(_bit_map->isMarked((HeapWord*)new_oop),
7677 "no white objects on this stack!");
7678 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7679 // iterate over the oops in this oop, marking and pushing
7680 // the ones in CMS heap (i.e. in _span).
7681 new_oop->oop_iterate(&_mark_and_push);
7682 }
7683 }
7684 }
7685
7686 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
7687 CMSCollector* collector,
7688 MemRegion span, CMSBitMap* bit_map,
7689 OopTaskQueue* work_queue):
7690 _collector(collector),
7691 _span(span),
7692 _bit_map(bit_map),
7693 _work_queue(work_queue) { }
7694
7695 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
7696 HeapWord* addr = (HeapWord*)obj;
7697 if (_span.contains(addr) &&
7698 !_bit_map->isMarked(addr)) {
7699 if (_bit_map->par_mark(addr)) {
7700 bool simulate_overflow = false;
7701 NOT_PRODUCT(
7702 if (CMSMarkStackOverflowALot &&
7703 _collector->par_simulate_overflow()) {
7704 // simulate a stack overflow
7705 simulate_overflow = true;
7706 }
7707 )
7708 if (simulate_overflow || !_work_queue->push(obj)) {
7709 _collector->par_push_on_overflow_list(obj);
7710 _collector->_par_kac_ovflw++;
7711 }
7712 } // Else another thread got there already
7713 }
7714 }
7715
7716 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7717 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7718
7719 //////////////////////////////////////////////////////////////////
7720 // CMSExpansionCause /////////////////////////////
7721 //////////////////////////////////////////////////////////////////
7722 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
7723 switch (cause) {
7724 case _no_expansion:
7725 return "No expansion";
7726 case _satisfy_free_ratio:
7727 return "Free ratio";
7728 case _satisfy_promotion:
7729 return "Satisfy promotion";
7730 case _satisfy_allocation:
7731 return "allocation";
7732 case _allocate_par_lab:
7733 return "Par LAB";
7734 case _allocate_par_spooling_space:
7735 return "Par Spooling Space";
7736 case _adaptive_size_policy:
7737 return "Ergonomics";
7738 default:
7739 return "unknown";
7740 }
7741 }
7742
7743 void CMSDrainMarkingStackClosure::do_void() {
7744 // the max number to take from overflow list at a time
7745 const size_t num = _mark_stack->capacity()/4;
7746 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
7747 "Overflow list should be NULL during concurrent phases");
7748 while (!_mark_stack->isEmpty() ||
7749 // if stack is empty, check the overflow list
7750 _collector->take_from_overflow_list(num, _mark_stack)) {
7751 oop obj = _mark_stack->pop();
7752 HeapWord* addr = (HeapWord*)obj;
7753 assert(_span.contains(addr), "Should be within span");
7754 assert(_bit_map->isMarked(addr), "Should be marked");
7755 assert(oopDesc::is_oop(obj), "Should be an oop");
7756 obj->oop_iterate(_keep_alive);
7757 }
7758 }
7759
7760 void CMSParDrainMarkingStackClosure::do_void() {
7761 // drain queue
7762 trim_queue(0);
7763 }
7764
7765 // Trim our work_queue so its length is below max at return
7766 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
7767 while (_work_queue->size() > max) {
7768 oop new_oop;
7769 if (_work_queue->pop_local(new_oop)) {
7770 assert(oopDesc::is_oop(new_oop), "Expected an oop");
7771 assert(_bit_map->isMarked((HeapWord*)new_oop),
7772 "no white objects on this stack!");
7773 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7774 // iterate over the oops in this oop, marking and pushing
7775 // the ones in CMS heap (i.e. in _span).
7776 new_oop->oop_iterate(&_mark_and_push);
7777 }
7778 }
7779 }
7780
7781 ////////////////////////////////////////////////////////////////////
7782 // Support for Marking Stack Overflow list handling and related code
7783 ////////////////////////////////////////////////////////////////////
7784 // Much of the following code is similar in shape and spirit to the
7785 // code used in ParNewGC. We should try and share that code
7786 // as much as possible in the future.
7787
7788 #ifndef PRODUCT
7789 // Debugging support for CMSStackOverflowALot
7790
7791 // It's OK to call this multi-threaded; the worst thing
7792 // that can happen is that we'll get a bunch of closely
7793 // spaced simulated overflows, but that's OK, in fact
7794 // probably good as it would exercise the overflow code
7795 // under contention.
7796 bool CMSCollector::simulate_overflow() {
7797 if (_overflow_counter-- <= 0) { // just being defensive
7798 _overflow_counter = CMSMarkStackOverflowInterval;
7799 return true;
7800 } else {
7801 return false;
7802 }
7803 }
7804
7805 bool CMSCollector::par_simulate_overflow() {
7806 return simulate_overflow();
7807 }
7808 #endif
7809
7810 // Single-threaded
7811 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
7812 assert(stack->isEmpty(), "Expected precondition");
7813 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
7814 size_t i = num;
7815 oop cur = _overflow_list;
7816 const markOop proto = markOopDesc::prototype();
7817 NOT_PRODUCT(ssize_t n = 0;)
7818 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
7819 next = oop(cur->mark_raw());
7820 cur->set_mark_raw(proto); // until proven otherwise
7821 assert(oopDesc::is_oop(cur), "Should be an oop");
7822 bool res = stack->push(cur);
7823 assert(res, "Bit off more than can chew?");
7824 NOT_PRODUCT(n++;)
7825 }
7826 _overflow_list = cur;
7827 #ifndef PRODUCT
7828 assert(_num_par_pushes >= n, "Too many pops?");
7829 _num_par_pushes -=n;
7830 #endif
7831 return !stack->isEmpty();
7832 }
7833
7834 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
7835 // (MT-safe) Get a prefix of at most "num" from the list.
7836 // The overflow list is chained through the mark word of
7837 // each object in the list. We fetch the entire list,
7838 // break off a prefix of the right size and return the
7839 // remainder. If other threads try to take objects from
7840 // the overflow list at that time, they will wait for
7841 // some time to see if data becomes available. If (and
7842 // only if) another thread places one or more object(s)
7843 // on the global list before we have returned the suffix
7844 // to the global list, we will walk down our local list
7845 // to find its end and append the global list to
7846 // our suffix before returning it. This suffix walk can
7847 // prove to be expensive (quadratic in the amount of traffic)
7848 // when there are many objects in the overflow list and
7849 // there is much producer-consumer contention on the list.
7850 // *NOTE*: The overflow list manipulation code here and
7851 // in ParNewGeneration:: are very similar in shape,
7852 // except that in the ParNew case we use the old (from/eden)
7853 // copy of the object to thread the list via its klass word.
7854 // Because of the common code, if you make any changes in
7855 // the code below, please check the ParNew version to see if
7856 // similar changes might be needed.
7857 // CR 6797058 has been filed to consolidate the common code.
7858 bool CMSCollector::par_take_from_overflow_list(size_t num,
7859 OopTaskQueue* work_q,
7860 int no_of_gc_threads) {
7861 assert(work_q->size() == 0, "First empty local work queue");
7862 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
7863 if (_overflow_list == NULL) {
7864 return false;
7865 }
7866 // Grab the entire list; we'll put back a suffix
7867 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
7868 Thread* tid = Thread::current();
7869 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
7870 // set to ParallelGCThreads.
7871 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
7872 size_t sleep_time_millis = MAX2((size_t)1, num/100);
7873 // If the list is busy, we spin for a short while,
7874 // sleeping between attempts to get the list.
7875 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
7876 os::sleep(tid, sleep_time_millis, false);
7877 if (_overflow_list == NULL) {
7878 // Nothing left to take
7879 return false;
7880 } else if (_overflow_list != BUSY) {
7881 // Try and grab the prefix
7882 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
7883 }
7884 }
7885 // If the list was found to be empty, or we spun long
7886 // enough, we give up and return empty-handed. If we leave
7887 // the list in the BUSY state below, it must be the case that
7888 // some other thread holds the overflow list and will set it
7889 // to a non-BUSY state in the future.
7890 if (prefix == NULL || prefix == BUSY) {
7891 // Nothing to take or waited long enough
7892 if (prefix == NULL) {
7893 // Write back the NULL in case we overwrote it with BUSY above
7894 // and it is still the same value.
7895 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
7896 }
7897 return false;
7898 }
7899 assert(prefix != NULL && prefix != BUSY, "Error");
7900 size_t i = num;
7901 oop cur = prefix;
7902 // Walk down the first "num" objects, unless we reach the end.
7903 for (; i > 1 && cur->mark_raw() != NULL; cur = oop(cur->mark_raw()), i--);
7904 if (cur->mark_raw() == NULL) {
7905 // We have "num" or fewer elements in the list, so there
7906 // is nothing to return to the global list.
7907 // Write back the NULL in lieu of the BUSY we wrote
7908 // above, if it is still the same value.
7909 if (_overflow_list == BUSY) {
7910 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
7911 }
7912 } else {
7913 // Chop off the suffix and return it to the global list.
7914 assert(cur->mark_raw() != BUSY, "Error");
7915 oop suffix_head = cur->mark_raw(); // suffix will be put back on global list
7916 cur->set_mark_raw(NULL); // break off suffix
7917 // It's possible that the list is still in the empty(busy) state
7918 // we left it in a short while ago; in that case we may be
7919 // able to place back the suffix without incurring the cost
7920 // of a walk down the list.
7921 oop observed_overflow_list = _overflow_list;
7922 oop cur_overflow_list = observed_overflow_list;
7923 bool attached = false;
7924 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
7925 observed_overflow_list =
7926 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list);
7927 if (cur_overflow_list == observed_overflow_list) {
7928 attached = true;
7929 break;
7930 } else cur_overflow_list = observed_overflow_list;
7931 }
7932 if (!attached) {
7933 // Too bad, someone else sneaked in (at least) an element; we'll need
7934 // to do a splice. Find tail of suffix so we can prepend suffix to global
7935 // list.
7936 for (cur = suffix_head; cur->mark_raw() != NULL; cur = (oop)(cur->mark_raw()));
7937 oop suffix_tail = cur;
7938 assert(suffix_tail != NULL && suffix_tail->mark_raw() == NULL,
7939 "Tautology");
7940 observed_overflow_list = _overflow_list;
7941 do {
7942 cur_overflow_list = observed_overflow_list;
7943 if (cur_overflow_list != BUSY) {
7944 // Do the splice ...
7945 suffix_tail->set_mark_raw(markOop(cur_overflow_list));
7946 } else { // cur_overflow_list == BUSY
7947 suffix_tail->set_mark_raw(NULL);
7948 }
7949 // ... and try to place spliced list back on overflow_list ...
7950 observed_overflow_list =
7951 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list);
7952 } while (cur_overflow_list != observed_overflow_list);
7953 // ... until we have succeeded in doing so.
7954 }
7955 }
7956
7957 // Push the prefix elements on work_q
7958 assert(prefix != NULL, "control point invariant");
7959 const markOop proto = markOopDesc::prototype();
7960 oop next;
7961 NOT_PRODUCT(ssize_t n = 0;)
7962 for (cur = prefix; cur != NULL; cur = next) {
7963 next = oop(cur->mark_raw());
7964 cur->set_mark_raw(proto); // until proven otherwise
7965 assert(oopDesc::is_oop(cur), "Should be an oop");
7966 bool res = work_q->push(cur);
7967 assert(res, "Bit off more than we can chew?");
7968 NOT_PRODUCT(n++;)
7969 }
7970 #ifndef PRODUCT
7971 assert(_num_par_pushes >= n, "Too many pops?");
7972 Atomic::sub(n, &_num_par_pushes);
7973 #endif
7974 return true;
7975 }
7976
7977 // Single-threaded
7978 void CMSCollector::push_on_overflow_list(oop p) {
7979 NOT_PRODUCT(_num_par_pushes++;)
7980 assert(oopDesc::is_oop(p), "Not an oop");
7981 preserve_mark_if_necessary(p);
7982 p->set_mark_raw((markOop)_overflow_list);
7983 _overflow_list = p;
7984 }
7985
7986 // Multi-threaded; use CAS to prepend to overflow list
7987 void CMSCollector::par_push_on_overflow_list(oop p) {
7988 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);)
7989 assert(oopDesc::is_oop(p), "Not an oop");
7990 par_preserve_mark_if_necessary(p);
7991 oop observed_overflow_list = _overflow_list;
7992 oop cur_overflow_list;
7993 do {
7994 cur_overflow_list = observed_overflow_list;
7995 if (cur_overflow_list != BUSY) {
7996 p->set_mark_raw(markOop(cur_overflow_list));
7997 } else {
7998 p->set_mark_raw(NULL);
7999 }
8000 observed_overflow_list =
8001 Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list);
8002 } while (cur_overflow_list != observed_overflow_list);
8003 }
8004 #undef BUSY
8005
8006 // Single threaded
8007 // General Note on GrowableArray: pushes may silently fail
8008 // because we are (temporarily) out of C-heap for expanding
8009 // the stack. The problem is quite ubiquitous and affects
8010 // a lot of code in the JVM. The prudent thing for GrowableArray
8011 // to do (for now) is to exit with an error. However, that may
8012 // be too draconian in some cases because the caller may be
8013 // able to recover without much harm. For such cases, we
8014 // should probably introduce a "soft_push" method which returns
8015 // an indication of success or failure with the assumption that
8016 // the caller may be able to recover from a failure; code in
8017 // the VM can then be changed, incrementally, to deal with such
8018 // failures where possible, thus, incrementally hardening the VM
8019 // in such low resource situations.
8020 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8021 _preserved_oop_stack.push(p);
8022 _preserved_mark_stack.push(m);
8023 assert(m == p->mark_raw(), "Mark word changed");
8024 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8025 "bijection");
8026 }
8027
8028 // Single threaded
8029 void CMSCollector::preserve_mark_if_necessary(oop p) {
8030 markOop m = p->mark_raw();
8031 if (m->must_be_preserved(p)) {
8032 preserve_mark_work(p, m);
8033 }
8034 }
8035
8036 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8037 markOop m = p->mark_raw();
8038 if (m->must_be_preserved(p)) {
8039 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8040 // Even though we read the mark word without holding
8041 // the lock, we are assured that it will not change
8042 // because we "own" this oop, so no other thread can
8043 // be trying to push it on the overflow list; see
8044 // the assertion in preserve_mark_work() that checks
8045 // that m == p->mark_raw().
8046 preserve_mark_work(p, m);
8047 }
8048 }
8049
8050 // We should be able to do this multi-threaded,
8051 // a chunk of stack being a task (this is
8052 // correct because each oop only ever appears
8053 // once in the overflow list. However, it's
8054 // not very easy to completely overlap this with
8055 // other operations, so will generally not be done
8056 // until all work's been completed. Because we
8057 // expect the preserved oop stack (set) to be small,
8058 // it's probably fine to do this single-threaded.
8059 // We can explore cleverer concurrent/overlapped/parallel
8060 // processing of preserved marks if we feel the
8061 // need for this in the future. Stack overflow should
8062 // be so rare in practice and, when it happens, its
8063 // effect on performance so great that this will
8064 // likely just be in the noise anyway.
8065 void CMSCollector::restore_preserved_marks_if_any() {
8066 assert(SafepointSynchronize::is_at_safepoint(),
8067 "world should be stopped");
8068 assert(Thread::current()->is_ConcurrentGC_thread() ||
8069 Thread::current()->is_VM_thread(),
8070 "should be single-threaded");
8071 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8072 "bijection");
8073
8074 while (!_preserved_oop_stack.is_empty()) {
8075 oop p = _preserved_oop_stack.pop();
8076 assert(oopDesc::is_oop(p), "Should be an oop");
8077 assert(_span.contains(p), "oop should be in _span");
8078 assert(p->mark_raw() == markOopDesc::prototype(),
8079 "Set when taken from overflow list");
8080 markOop m = _preserved_mark_stack.pop();
8081 p->set_mark_raw(m);
8082 }
8083 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8084 "stacks were cleared above");
8085 }
8086
8087 #ifndef PRODUCT
8088 bool CMSCollector::no_preserved_marks() const {
8089 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8090 }
8091 #endif
8092
8093 // Transfer some number of overflown objects to usual marking
8094 // stack. Return true if some objects were transferred.
8095 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8096 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8097 (size_t)ParGCDesiredObjsFromOverflowList);
8098
8099 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8100 assert(_collector->overflow_list_is_empty() || res,
8101 "If list is not empty, we should have taken something");
8102 assert(!res || !_mark_stack->isEmpty(),
8103 "If we took something, it should now be on our stack");
8104 return res;
8105 }
8106
8107 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8108 size_t res = _sp->block_size_no_stall(addr, _collector);
8109 if (_sp->block_is_obj(addr)) {
8110 if (_live_bit_map->isMarked(addr)) {
8111 // It can't have been dead in a previous cycle
8112 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8113 } else {
8114 _dead_bit_map->mark(addr); // mark the dead object
8115 }
8116 }
8117 // Could be 0, if the block size could not be computed without stalling.
8118 return res;
8119 }
8120
8121 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8122 GCMemoryManager* manager = CMSHeap::heap()->old_manager();
8123 switch (phase) {
8124 case CMSCollector::InitialMarking:
8125 initialize(manager /* GC manager */ ,
8126 cause /* cause of the GC */,
8127 true /* recordGCBeginTime */,
8128 true /* recordPreGCUsage */,
8129 false /* recordPeakUsage */,
8130 false /* recordPostGCusage */,
8131 true /* recordAccumulatedGCTime */,
8132 false /* recordGCEndTime */,
8133 false /* countCollection */ );
8134 break;
8135
8136 case CMSCollector::FinalMarking:
8137 initialize(manager /* GC manager */ ,
8138 cause /* cause of the GC */,
8139 false /* recordGCBeginTime */,
8140 false /* recordPreGCUsage */,
8141 false /* recordPeakUsage */,
8142 false /* recordPostGCusage */,
8143 true /* recordAccumulatedGCTime */,
8144 false /* recordGCEndTime */,
8145 false /* countCollection */ );
8146 break;
8147
8148 case CMSCollector::Sweeping:
8149 initialize(manager /* GC manager */ ,
8150 cause /* cause of the GC */,
8151 false /* recordGCBeginTime */,
8152 false /* recordPreGCUsage */,
8153 true /* recordPeakUsage */,
8154 true /* recordPostGCusage */,
8155 false /* recordAccumulatedGCTime */,
8156 true /* recordGCEndTime */,
8157 true /* countCollection */ );
8158 break;
8159
8160 default:
8161 ShouldNotReachHere();
8162 }
8163 }