1 /*
2 * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/classLoaderData.hpp"
27 #include "classfile/stringTable.hpp"
28 #include "classfile/symbolTable.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/codeCache.hpp"
31 #include "gc/cms/cmsCollectorPolicy.hpp"
32 #include "gc/cms/cmsOopClosures.inline.hpp"
33 #include "gc/cms/compactibleFreeListSpace.hpp"
34 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp"
35 #include "gc/cms/concurrentMarkSweepThread.hpp"
36 #include "gc/cms/parNewGeneration.hpp"
37 #include "gc/cms/vmCMSOperations.hpp"
38 #include "gc/serial/genMarkSweep.hpp"
39 #include "gc/serial/tenuredGeneration.hpp"
40 #include "gc/shared/adaptiveSizePolicy.hpp"
41 #include "gc/shared/cardGeneration.inline.hpp"
42 #include "gc/shared/cardTableRS.hpp"
43 #include "gc/shared/collectedHeap.inline.hpp"
44 #include "gc/shared/collectorCounters.hpp"
45 #include "gc/shared/collectorPolicy.hpp"
46 #include "gc/shared/gcLocker.inline.hpp"
47 #include "gc/shared/gcPolicyCounters.hpp"
48 #include "gc/shared/gcTimer.hpp"
49 #include "gc/shared/gcTrace.hpp"
50 #include "gc/shared/gcTraceTime.inline.hpp"
51 #include "gc/shared/genCollectedHeap.hpp"
52 #include "gc/shared/genOopClosures.inline.hpp"
53 #include "gc/shared/isGCActiveMark.hpp"
54 #include "gc/shared/referencePolicy.hpp"
55 #include "gc/shared/strongRootsScope.hpp"
56 #include "gc/shared/taskqueue.inline.hpp"
57 #include "logging/log.hpp"
58 #include "memory/allocation.hpp"
59 #include "memory/iterator.inline.hpp"
60 #include "memory/padded.hpp"
61 #include "memory/resourceArea.hpp"
62 #include "oops/oop.inline.hpp"
63 #include "prims/jvmtiExport.hpp"
64 #include "runtime/atomic.inline.hpp"
65 #include "runtime/globals_extension.hpp"
66 #include "runtime/handles.inline.hpp"
67 #include "runtime/java.hpp"
68 #include "runtime/orderAccess.inline.hpp"
69 #include "runtime/timer.hpp"
70 #include "runtime/vmThread.hpp"
71 #include "services/memoryService.hpp"
72 #include "services/runtimeService.hpp"
73 #include "utilities/stack.inline.hpp"
74
75 // statics
76 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
77 bool CMSCollector::_full_gc_requested = false;
78 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
79
80 //////////////////////////////////////////////////////////////////
81 // In support of CMS/VM thread synchronization
82 //////////////////////////////////////////////////////////////////
83 // We split use of the CGC_lock into 2 "levels".
84 // The low-level locking is of the usual CGC_lock monitor. We introduce
85 // a higher level "token" (hereafter "CMS token") built on top of the
86 // low level monitor (hereafter "CGC lock").
87 // The token-passing protocol gives priority to the VM thread. The
88 // CMS-lock doesn't provide any fairness guarantees, but clients
89 // should ensure that it is only held for very short, bounded
90 // durations.
91 //
92 // When either of the CMS thread or the VM thread is involved in
93 // collection operations during which it does not want the other
94 // thread to interfere, it obtains the CMS token.
95 //
96 // If either thread tries to get the token while the other has
97 // it, that thread waits. However, if the VM thread and CMS thread
98 // both want the token, then the VM thread gets priority while the
99 // CMS thread waits. This ensures, for instance, that the "concurrent"
100 // phases of the CMS thread's work do not block out the VM thread
101 // for long periods of time as the CMS thread continues to hog
102 // the token. (See bug 4616232).
103 //
104 // The baton-passing functions are, however, controlled by the
105 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
106 // and here the low-level CMS lock, not the high level token,
107 // ensures mutual exclusion.
108 //
109 // Two important conditions that we have to satisfy:
110 // 1. if a thread does a low-level wait on the CMS lock, then it
111 // relinquishes the CMS token if it were holding that token
112 // when it acquired the low-level CMS lock.
113 // 2. any low-level notifications on the low-level lock
114 // should only be sent when a thread has relinquished the token.
115 //
116 // In the absence of either property, we'd have potential deadlock.
117 //
118 // We protect each of the CMS (concurrent and sequential) phases
119 // with the CMS _token_, not the CMS _lock_.
120 //
121 // The only code protected by CMS lock is the token acquisition code
122 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
123 // baton-passing code.
124 //
125 // Unfortunately, i couldn't come up with a good abstraction to factor and
126 // hide the naked CGC_lock manipulation in the baton-passing code
127 // further below. That's something we should try to do. Also, the proof
128 // of correctness of this 2-level locking scheme is far from obvious,
129 // and potentially quite slippery. We have an uneasy suspicion, for instance,
130 // that there may be a theoretical possibility of delay/starvation in the
131 // low-level lock/wait/notify scheme used for the baton-passing because of
132 // potential interference with the priority scheme embodied in the
133 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
134 // invocation further below and marked with "XXX 20011219YSR".
135 // Indeed, as we note elsewhere, this may become yet more slippery
136 // in the presence of multiple CMS and/or multiple VM threads. XXX
137
138 class CMSTokenSync: public StackObj {
139 private:
140 bool _is_cms_thread;
141 public:
142 CMSTokenSync(bool is_cms_thread):
143 _is_cms_thread(is_cms_thread) {
144 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
145 "Incorrect argument to constructor");
146 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
147 }
148
149 ~CMSTokenSync() {
150 assert(_is_cms_thread ?
151 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
152 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
153 "Incorrect state");
154 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
155 }
156 };
157
158 // Convenience class that does a CMSTokenSync, and then acquires
159 // upto three locks.
160 class CMSTokenSyncWithLocks: public CMSTokenSync {
161 private:
162 // Note: locks are acquired in textual declaration order
163 // and released in the opposite order
164 MutexLockerEx _locker1, _locker2, _locker3;
165 public:
166 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
167 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
168 CMSTokenSync(is_cms_thread),
169 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
170 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
171 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
172 { }
173 };
174
175
176 //////////////////////////////////////////////////////////////////
177 // Concurrent Mark-Sweep Generation /////////////////////////////
178 //////////////////////////////////////////////////////////////////
179
180 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
181
182 // This struct contains per-thread things necessary to support parallel
183 // young-gen collection.
184 class CMSParGCThreadState: public CHeapObj<mtGC> {
185 public:
186 CompactibleFreeListSpaceLAB lab;
187 PromotionInfo promo;
188
189 // Constructor.
190 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
191 promo.setSpace(cfls);
192 }
193 };
194
195 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
196 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) :
197 CardGeneration(rs, initial_byte_size, ct),
198 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
199 _did_compact(false)
200 {
201 HeapWord* bottom = (HeapWord*) _virtual_space.low();
202 HeapWord* end = (HeapWord*) _virtual_space.high();
203
204 _direct_allocated_words = 0;
205 NOT_PRODUCT(
206 _numObjectsPromoted = 0;
207 _numWordsPromoted = 0;
208 _numObjectsAllocated = 0;
209 _numWordsAllocated = 0;
210 )
211
212 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end));
213 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
214 _cmsSpace->_old_gen = this;
215
216 _gc_stats = new CMSGCStats();
217
218 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
219 // offsets match. The ability to tell free chunks from objects
220 // depends on this property.
221 debug_only(
222 FreeChunk* junk = NULL;
223 assert(UseCompressedClassPointers ||
224 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
225 "Offset of FreeChunk::_prev within FreeChunk must match"
226 " that of OopDesc::_klass within OopDesc");
227 )
228
229 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC);
230 for (uint i = 0; i < ParallelGCThreads; i++) {
231 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
232 }
233
234 _incremental_collection_failed = false;
235 // The "dilatation_factor" is the expansion that can occur on
236 // account of the fact that the minimum object size in the CMS
237 // generation may be larger than that in, say, a contiguous young
238 // generation.
239 // Ideally, in the calculation below, we'd compute the dilatation
240 // factor as: MinChunkSize/(promoting_gen's min object size)
241 // Since we do not have such a general query interface for the
242 // promoting generation, we'll instead just use the minimum
243 // object size (which today is a header's worth of space);
244 // note that all arithmetic is in units of HeapWords.
245 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
246 assert(_dilatation_factor >= 1.0, "from previous assert");
247 }
248
249
250 // The field "_initiating_occupancy" represents the occupancy percentage
251 // at which we trigger a new collection cycle. Unless explicitly specified
252 // via CMSInitiatingOccupancyFraction (argument "io" below), it
253 // is calculated by:
254 //
255 // Let "f" be MinHeapFreeRatio in
256 //
257 // _initiating_occupancy = 100-f +
258 // f * (CMSTriggerRatio/100)
259 // where CMSTriggerRatio is the argument "tr" below.
260 //
261 // That is, if we assume the heap is at its desired maximum occupancy at the
262 // end of a collection, we let CMSTriggerRatio of the (purported) free
263 // space be allocated before initiating a new collection cycle.
264 //
265 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
266 assert(io <= 100 && tr <= 100, "Check the arguments");
267 if (io >= 0) {
268 _initiating_occupancy = (double)io / 100.0;
269 } else {
270 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
271 (double)(tr * MinHeapFreeRatio) / 100.0)
272 / 100.0;
273 }
274 }
275
276 void ConcurrentMarkSweepGeneration::ref_processor_init() {
277 assert(collector() != NULL, "no collector");
278 collector()->ref_processor_init();
279 }
280
281 void CMSCollector::ref_processor_init() {
282 if (_ref_processor == NULL) {
283 // Allocate and initialize a reference processor
284 _ref_processor =
285 new ReferenceProcessor(_span, // span
286 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
287 ParallelGCThreads, // mt processing degree
288 _cmsGen->refs_discovery_is_mt(), // mt discovery
289 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
290 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
291 &_is_alive_closure); // closure for liveness info
292 // Initialize the _ref_processor field of CMSGen
293 _cmsGen->set_ref_processor(_ref_processor);
294
295 }
296 }
297
298 AdaptiveSizePolicy* CMSCollector::size_policy() {
299 GenCollectedHeap* gch = GenCollectedHeap::heap();
300 return gch->gen_policy()->size_policy();
301 }
302
303 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
304
305 const char* gen_name = "old";
306 GenCollectorPolicy* gcp = GenCollectedHeap::heap()->gen_policy();
307 // Generation Counters - generation 1, 1 subspace
308 _gen_counters = new GenerationCounters(gen_name, 1, 1,
309 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
310
311 _space_counters = new GSpaceCounters(gen_name, 0,
312 _virtual_space.reserved_size(),
313 this, _gen_counters);
314 }
315
316 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
317 _cms_gen(cms_gen)
318 {
319 assert(alpha <= 100, "bad value");
320 _saved_alpha = alpha;
321
322 // Initialize the alphas to the bootstrap value of 100.
323 _gc0_alpha = _cms_alpha = 100;
324
325 _cms_begin_time.update();
326 _cms_end_time.update();
327
328 _gc0_duration = 0.0;
329 _gc0_period = 0.0;
330 _gc0_promoted = 0;
331
332 _cms_duration = 0.0;
333 _cms_period = 0.0;
334 _cms_allocated = 0;
335
336 _cms_used_at_gc0_begin = 0;
337 _cms_used_at_gc0_end = 0;
338 _allow_duty_cycle_reduction = false;
339 _valid_bits = 0;
340 }
341
342 double CMSStats::cms_free_adjustment_factor(size_t free) const {
343 // TBD: CR 6909490
344 return 1.0;
345 }
346
347 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
348 }
349
350 // If promotion failure handling is on use
351 // the padded average size of the promotion for each
352 // young generation collection.
353 double CMSStats::time_until_cms_gen_full() const {
354 size_t cms_free = _cms_gen->cmsSpace()->free();
355 GenCollectedHeap* gch = GenCollectedHeap::heap();
356 size_t expected_promotion = MIN2(gch->young_gen()->capacity(),
357 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
358 if (cms_free > expected_promotion) {
359 // Start a cms collection if there isn't enough space to promote
360 // for the next young collection. Use the padded average as
361 // a safety factor.
362 cms_free -= expected_promotion;
363
364 // Adjust by the safety factor.
365 double cms_free_dbl = (double)cms_free;
366 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0;
367 // Apply a further correction factor which tries to adjust
368 // for recent occurance of concurrent mode failures.
369 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
370 cms_free_dbl = cms_free_dbl * cms_adjustment;
371
372 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
373 cms_free, expected_promotion);
374 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0);
375 // Add 1 in case the consumption rate goes to zero.
376 return cms_free_dbl / (cms_consumption_rate() + 1.0);
377 }
378 return 0.0;
379 }
380
381 // Compare the duration of the cms collection to the
382 // time remaining before the cms generation is empty.
383 // Note that the time from the start of the cms collection
384 // to the start of the cms sweep (less than the total
385 // duration of the cms collection) can be used. This
386 // has been tried and some applications experienced
387 // promotion failures early in execution. This was
388 // possibly because the averages were not accurate
389 // enough at the beginning.
390 double CMSStats::time_until_cms_start() const {
391 // We add "gc0_period" to the "work" calculation
392 // below because this query is done (mostly) at the
393 // end of a scavenge, so we need to conservatively
394 // account for that much possible delay
395 // in the query so as to avoid concurrent mode failures
396 // due to starting the collection just a wee bit too
397 // late.
398 double work = cms_duration() + gc0_period();
399 double deadline = time_until_cms_gen_full();
400 // If a concurrent mode failure occurred recently, we want to be
401 // more conservative and halve our expected time_until_cms_gen_full()
402 if (work > deadline) {
403 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ",
404 cms_duration(), gc0_period(), time_until_cms_gen_full());
405 return 0.0;
406 }
407 return work - deadline;
408 }
409
410 #ifndef PRODUCT
411 void CMSStats::print_on(outputStream *st) const {
412 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
413 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
414 gc0_duration(), gc0_period(), gc0_promoted());
415 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
416 cms_duration(), cms_period(), cms_allocated());
417 st->print(",cms_since_beg=%g,cms_since_end=%g",
418 cms_time_since_begin(), cms_time_since_end());
419 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
420 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
421
422 if (valid()) {
423 st->print(",promo_rate=%g,cms_alloc_rate=%g",
424 promotion_rate(), cms_allocation_rate());
425 st->print(",cms_consumption_rate=%g,time_until_full=%g",
426 cms_consumption_rate(), time_until_cms_gen_full());
427 }
428 st->cr();
429 }
430 #endif // #ifndef PRODUCT
431
432 CMSCollector::CollectorState CMSCollector::_collectorState =
433 CMSCollector::Idling;
434 bool CMSCollector::_foregroundGCIsActive = false;
435 bool CMSCollector::_foregroundGCShouldWait = false;
436
437 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
438 CardTableRS* ct,
439 ConcurrentMarkSweepPolicy* cp):
440 _cmsGen(cmsGen),
441 _ct(ct),
442 _ref_processor(NULL), // will be set later
443 _conc_workers(NULL), // may be set later
444 _abort_preclean(false),
445 _start_sampling(false),
446 _between_prologue_and_epilogue(false),
447 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
448 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
449 -1 /* lock-free */, "No_lock" /* dummy */),
450 _modUnionClosurePar(&_modUnionTable),
451 // Adjust my span to cover old (cms) gen
452 _span(cmsGen->reserved()),
453 // Construct the is_alive_closure with _span & markBitMap
454 _is_alive_closure(_span, &_markBitMap),
455 _restart_addr(NULL),
456 _overflow_list(NULL),
457 _stats(cmsGen),
458 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
459 //verify that this lock should be acquired with safepoint check.
460 Monitor::_safepoint_check_sometimes)),
461 _eden_chunk_array(NULL), // may be set in ctor body
462 _eden_chunk_capacity(0), // -- ditto --
463 _eden_chunk_index(0), // -- ditto --
464 _survivor_plab_array(NULL), // -- ditto --
465 _survivor_chunk_array(NULL), // -- ditto --
466 _survivor_chunk_capacity(0), // -- ditto --
467 _survivor_chunk_index(0), // -- ditto --
468 _ser_pmc_preclean_ovflw(0),
469 _ser_kac_preclean_ovflw(0),
470 _ser_pmc_remark_ovflw(0),
471 _par_pmc_remark_ovflw(0),
472 _ser_kac_ovflw(0),
473 _par_kac_ovflw(0),
474 #ifndef PRODUCT
475 _num_par_pushes(0),
476 #endif
477 _collection_count_start(0),
478 _verifying(false),
479 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
480 _completed_initialization(false),
481 _collector_policy(cp),
482 _should_unload_classes(CMSClassUnloadingEnabled),
483 _concurrent_cycles_since_last_unload(0),
484 _roots_scanning_options(GenCollectedHeap::SO_None),
485 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
486 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
487 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
488 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
489 _cms_start_registered(false)
490 {
491 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
492 ExplicitGCInvokesConcurrent = true;
493 }
494 // Now expand the span and allocate the collection support structures
495 // (MUT, marking bit map etc.) to cover both generations subject to
496 // collection.
497
498 // For use by dirty card to oop closures.
499 _cmsGen->cmsSpace()->set_collector(this);
500
501 // Allocate MUT and marking bit map
502 {
503 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
504 if (!_markBitMap.allocate(_span)) {
505 log_warning(gc)("Failed to allocate CMS Bit Map");
506 return;
507 }
508 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
509 }
510 {
511 _modUnionTable.allocate(_span);
512 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
513 }
514
515 if (!_markStack.allocate(MarkStackSize)) {
516 log_warning(gc)("Failed to allocate CMS Marking Stack");
517 return;
518 }
519
520 // Support for multi-threaded concurrent phases
521 if (CMSConcurrentMTEnabled) {
522 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
523 // just for now
524 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4);
525 }
526 if (ConcGCThreads > 1) {
527 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
528 ConcGCThreads, true);
529 if (_conc_workers == NULL) {
530 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled");
531 CMSConcurrentMTEnabled = false;
532 } else {
533 _conc_workers->initialize_workers();
534 }
535 } else {
536 CMSConcurrentMTEnabled = false;
537 }
538 }
539 if (!CMSConcurrentMTEnabled) {
540 ConcGCThreads = 0;
541 } else {
542 // Turn off CMSCleanOnEnter optimization temporarily for
543 // the MT case where it's not fixed yet; see 6178663.
544 CMSCleanOnEnter = false;
545 }
546 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
547 "Inconsistency");
548 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads);
549 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
550
551 // Parallel task queues; these are shared for the
552 // concurrent and stop-world phases of CMS, but
553 // are not shared with parallel scavenge (ParNew).
554 {
555 uint i;
556 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads);
557
558 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
559 || ParallelRefProcEnabled)
560 && num_queues > 0) {
561 _task_queues = new OopTaskQueueSet(num_queues);
562 if (_task_queues == NULL) {
563 log_warning(gc)("task_queues allocation failure.");
564 return;
565 }
566 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
567 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
568 for (i = 0; i < num_queues; i++) {
569 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
570 if (q == NULL) {
571 log_warning(gc)("work_queue allocation failure.");
572 return;
573 }
574 _task_queues->register_queue(i, q);
575 }
576 for (i = 0; i < num_queues; i++) {
577 _task_queues->queue(i)->initialize();
578 _hash_seed[i] = 17; // copied from ParNew
579 }
580 }
581 }
582
583 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
584
585 // Clip CMSBootstrapOccupancy between 0 and 100.
586 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0;
587
588 // Now tell CMS generations the identity of their collector
589 ConcurrentMarkSweepGeneration::set_collector(this);
590
591 // Create & start a CMS thread for this CMS collector
592 _cmsThread = ConcurrentMarkSweepThread::start(this);
593 assert(cmsThread() != NULL, "CMS Thread should have been created");
594 assert(cmsThread()->collector() == this,
595 "CMS Thread should refer to this gen");
596 assert(CGC_lock != NULL, "Where's the CGC_lock?");
597
598 // Support for parallelizing young gen rescan
599 GenCollectedHeap* gch = GenCollectedHeap::heap();
600 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
601 _young_gen = (ParNewGeneration*)gch->young_gen();
602 if (gch->supports_inline_contig_alloc()) {
603 _top_addr = gch->top_addr();
604 _end_addr = gch->end_addr();
605 assert(_young_gen != NULL, "no _young_gen");
606 _eden_chunk_index = 0;
607 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain;
608 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
609 }
610
611 // Support for parallelizing survivor space rescan
612 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
613 const size_t max_plab_samples =
614 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize);
615
616 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
617 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
618 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
619 _survivor_chunk_capacity = max_plab_samples;
620 for (uint i = 0; i < ParallelGCThreads; i++) {
621 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
622 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
623 assert(cur->end() == 0, "Should be 0");
624 assert(cur->array() == vec, "Should be vec");
625 assert(cur->capacity() == max_plab_samples, "Error");
626 }
627 }
628
629 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
630 _gc_counters = new CollectorCounters("CMS", 1);
631 _completed_initialization = true;
632 _inter_sweep_timer.start(); // start of time
633 }
634
635 const char* ConcurrentMarkSweepGeneration::name() const {
636 return "concurrent mark-sweep generation";
637 }
638 void ConcurrentMarkSweepGeneration::update_counters() {
639 if (UsePerfData) {
640 _space_counters->update_all();
641 _gen_counters->update_all();
642 }
643 }
644
645 // this is an optimized version of update_counters(). it takes the
646 // used value as a parameter rather than computing it.
647 //
648 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
649 if (UsePerfData) {
650 _space_counters->update_used(used);
651 _space_counters->update_capacity();
652 _gen_counters->update_all();
653 }
654 }
655
656 void ConcurrentMarkSweepGeneration::print() const {
657 Generation::print();
658 cmsSpace()->print();
659 }
660
661 #ifndef PRODUCT
662 void ConcurrentMarkSweepGeneration::print_statistics() {
663 cmsSpace()->printFLCensus(0);
664 }
665 #endif
666
667 size_t
668 ConcurrentMarkSweepGeneration::contiguous_available() const {
669 // dld proposes an improvement in precision here. If the committed
670 // part of the space ends in a free block we should add that to
671 // uncommitted size in the calculation below. Will make this
672 // change later, staying with the approximation below for the
673 // time being. -- ysr.
674 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
675 }
676
677 size_t
678 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
679 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
680 }
681
682 size_t ConcurrentMarkSweepGeneration::max_available() const {
683 return free() + _virtual_space.uncommitted_size();
684 }
685
686 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
687 size_t available = max_available();
688 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
689 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
690 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")",
691 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes);
692 return res;
693 }
694
695 // At a promotion failure dump information on block layout in heap
696 // (cms old generation).
697 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
698 Log(gc, promotion) log;
699 if (log.is_trace()) {
700 ResourceMark rm;
701 cmsSpace()->dump_at_safepoint_with_locks(collector(), log.trace_stream());
702 }
703 }
704
705 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
706 // Clear the promotion information. These pointers can be adjusted
707 // along with all the other pointers into the heap but
708 // compaction is expected to be a rare event with
709 // a heap using cms so don't do it without seeing the need.
710 for (uint i = 0; i < ParallelGCThreads; i++) {
711 _par_gc_thread_states[i]->promo.reset();
712 }
713 }
714
715 void ConcurrentMarkSweepGeneration::compute_new_size() {
716 assert_locked_or_safepoint(Heap_lock);
717
718 // If incremental collection failed, we just want to expand
719 // to the limit.
720 if (incremental_collection_failed()) {
721 clear_incremental_collection_failed();
722 grow_to_reserved();
723 return;
724 }
725
726 // The heap has been compacted but not reset yet.
727 // Any metric such as free() or used() will be incorrect.
728
729 CardGeneration::compute_new_size();
730
731 // Reset again after a possible resizing
732 if (did_compact()) {
733 cmsSpace()->reset_after_compaction();
734 }
735 }
736
737 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
738 assert_locked_or_safepoint(Heap_lock);
739
740 // If incremental collection failed, we just want to expand
741 // to the limit.
742 if (incremental_collection_failed()) {
743 clear_incremental_collection_failed();
744 grow_to_reserved();
745 return;
746 }
747
748 double free_percentage = ((double) free()) / capacity();
749 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
750 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
751
752 // compute expansion delta needed for reaching desired free percentage
753 if (free_percentage < desired_free_percentage) {
754 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
755 assert(desired_capacity >= capacity(), "invalid expansion size");
756 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
757 Log(gc) log;
758 if (log.is_trace()) {
759 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
760 log.trace("From compute_new_size: ");
761 log.trace(" Free fraction %f", free_percentage);
762 log.trace(" Desired free fraction %f", desired_free_percentage);
763 log.trace(" Maximum free fraction %f", maximum_free_percentage);
764 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000);
765 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000);
766 GenCollectedHeap* gch = GenCollectedHeap::heap();
767 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation");
768 size_t young_size = gch->young_gen()->capacity();
769 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000);
770 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000);
771 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000);
772 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes);
773 }
774 // safe if expansion fails
775 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
776 log.trace(" Expanded free fraction %f", ((double) free()) / capacity());
777 } else {
778 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
779 assert(desired_capacity <= capacity(), "invalid expansion size");
780 size_t shrink_bytes = capacity() - desired_capacity;
781 // Don't shrink unless the delta is greater than the minimum shrink we want
782 if (shrink_bytes >= MinHeapDeltaBytes) {
783 shrink_free_list_by(shrink_bytes);
784 }
785 }
786 }
787
788 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
789 return cmsSpace()->freelistLock();
790 }
791
792 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) {
793 CMSSynchronousYieldRequest yr;
794 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
795 return have_lock_and_allocate(size, tlab);
796 }
797
798 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
799 bool tlab /* ignored */) {
800 assert_lock_strong(freelistLock());
801 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
802 HeapWord* res = cmsSpace()->allocate(adjustedSize);
803 // Allocate the object live (grey) if the background collector has
804 // started marking. This is necessary because the marker may
805 // have passed this address and consequently this object will
806 // not otherwise be greyed and would be incorrectly swept up.
807 // Note that if this object contains references, the writing
808 // of those references will dirty the card containing this object
809 // allowing the object to be blackened (and its references scanned)
810 // either during a preclean phase or at the final checkpoint.
811 if (res != NULL) {
812 // We may block here with an uninitialized object with
813 // its mark-bit or P-bits not yet set. Such objects need
814 // to be safely navigable by block_start().
815 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
816 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
817 collector()->direct_allocated(res, adjustedSize);
818 _direct_allocated_words += adjustedSize;
819 // allocation counters
820 NOT_PRODUCT(
821 _numObjectsAllocated++;
822 _numWordsAllocated += (int)adjustedSize;
823 )
824 }
825 return res;
826 }
827
828 // In the case of direct allocation by mutators in a generation that
829 // is being concurrently collected, the object must be allocated
830 // live (grey) if the background collector has started marking.
831 // This is necessary because the marker may
832 // have passed this address and consequently this object will
833 // not otherwise be greyed and would be incorrectly swept up.
834 // Note that if this object contains references, the writing
835 // of those references will dirty the card containing this object
836 // allowing the object to be blackened (and its references scanned)
837 // either during a preclean phase or at the final checkpoint.
838 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
839 assert(_markBitMap.covers(start, size), "Out of bounds");
840 if (_collectorState >= Marking) {
841 MutexLockerEx y(_markBitMap.lock(),
842 Mutex::_no_safepoint_check_flag);
843 // [see comments preceding SweepClosure::do_blk() below for details]
844 //
845 // Can the P-bits be deleted now? JJJ
846 //
847 // 1. need to mark the object as live so it isn't collected
848 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
849 // 3. need to mark the end of the object so marking, precleaning or sweeping
850 // can skip over uninitialized or unparsable objects. An allocated
851 // object is considered uninitialized for our purposes as long as
852 // its klass word is NULL. All old gen objects are parsable
853 // as soon as they are initialized.)
854 _markBitMap.mark(start); // object is live
855 _markBitMap.mark(start + 1); // object is potentially uninitialized?
856 _markBitMap.mark(start + size - 1);
857 // mark end of object
858 }
859 // check that oop looks uninitialized
860 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
861 }
862
863 void CMSCollector::promoted(bool par, HeapWord* start,
864 bool is_obj_array, size_t obj_size) {
865 assert(_markBitMap.covers(start), "Out of bounds");
866 // See comment in direct_allocated() about when objects should
867 // be allocated live.
868 if (_collectorState >= Marking) {
869 // we already hold the marking bit map lock, taken in
870 // the prologue
871 if (par) {
872 _markBitMap.par_mark(start);
873 } else {
874 _markBitMap.mark(start);
875 }
876 // We don't need to mark the object as uninitialized (as
877 // in direct_allocated above) because this is being done with the
878 // world stopped and the object will be initialized by the
879 // time the marking, precleaning or sweeping get to look at it.
880 // But see the code for copying objects into the CMS generation,
881 // where we need to ensure that concurrent readers of the
882 // block offset table are able to safely navigate a block that
883 // is in flux from being free to being allocated (and in
884 // transition while being copied into) and subsequently
885 // becoming a bona-fide object when the copy/promotion is complete.
886 assert(SafepointSynchronize::is_at_safepoint(),
887 "expect promotion only at safepoints");
888
889 if (_collectorState < Sweeping) {
890 // Mark the appropriate cards in the modUnionTable, so that
891 // this object gets scanned before the sweep. If this is
892 // not done, CMS generation references in the object might
893 // not get marked.
894 // For the case of arrays, which are otherwise precisely
895 // marked, we need to dirty the entire array, not just its head.
896 if (is_obj_array) {
897 // The [par_]mark_range() method expects mr.end() below to
898 // be aligned to the granularity of a bit's representation
899 // in the heap. In the case of the MUT below, that's a
900 // card size.
901 MemRegion mr(start,
902 (HeapWord*)round_to((intptr_t)(start + obj_size),
903 CardTableModRefBS::card_size /* bytes */));
904 if (par) {
905 _modUnionTable.par_mark_range(mr);
906 } else {
907 _modUnionTable.mark_range(mr);
908 }
909 } else { // not an obj array; we can just mark the head
910 if (par) {
911 _modUnionTable.par_mark(start);
912 } else {
913 _modUnionTable.mark(start);
914 }
915 }
916 }
917 }
918 }
919
920 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
921 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
922 // allocate, copy and if necessary update promoinfo --
923 // delegate to underlying space.
924 assert_lock_strong(freelistLock());
925
926 #ifndef PRODUCT
927 if (GenCollectedHeap::heap()->promotion_should_fail()) {
928 return NULL;
929 }
930 #endif // #ifndef PRODUCT
931
932 oop res = _cmsSpace->promote(obj, obj_size);
933 if (res == NULL) {
934 // expand and retry
935 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
936 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
937 // Since this is the old generation, we don't try to promote
938 // into a more senior generation.
939 res = _cmsSpace->promote(obj, obj_size);
940 }
941 if (res != NULL) {
942 // See comment in allocate() about when objects should
943 // be allocated live.
944 assert(obj->is_oop(), "Will dereference klass pointer below");
945 collector()->promoted(false, // Not parallel
946 (HeapWord*)res, obj->is_objArray(), obj_size);
947 // promotion counters
948 NOT_PRODUCT(
949 _numObjectsPromoted++;
950 _numWordsPromoted +=
951 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
952 )
953 }
954 return res;
955 }
956
957
958 // IMPORTANT: Notes on object size recognition in CMS.
959 // ---------------------------------------------------
960 // A block of storage in the CMS generation is always in
961 // one of three states. A free block (FREE), an allocated
962 // object (OBJECT) whose size() method reports the correct size,
963 // and an intermediate state (TRANSIENT) in which its size cannot
964 // be accurately determined.
965 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
966 // -----------------------------------------------------
967 // FREE: klass_word & 1 == 1; mark_word holds block size
968 //
969 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
970 // obj->size() computes correct size
971 //
972 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
973 //
974 // STATE IDENTIFICATION: (64 bit+COOPS)
975 // ------------------------------------
976 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
977 //
978 // OBJECT: klass_word installed; klass_word != 0;
979 // obj->size() computes correct size
980 //
981 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
982 //
983 //
984 // STATE TRANSITION DIAGRAM
985 //
986 // mut / parnew mut / parnew
987 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
988 // ^ |
989 // |------------------------ DEAD <------------------------------------|
990 // sweep mut
991 //
992 // While a block is in TRANSIENT state its size cannot be determined
993 // so readers will either need to come back later or stall until
994 // the size can be determined. Note that for the case of direct
995 // allocation, P-bits, when available, may be used to determine the
996 // size of an object that may not yet have been initialized.
997
998 // Things to support parallel young-gen collection.
999 oop
1000 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1001 oop old, markOop m,
1002 size_t word_sz) {
1003 #ifndef PRODUCT
1004 if (GenCollectedHeap::heap()->promotion_should_fail()) {
1005 return NULL;
1006 }
1007 #endif // #ifndef PRODUCT
1008
1009 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1010 PromotionInfo* promoInfo = &ps->promo;
1011 // if we are tracking promotions, then first ensure space for
1012 // promotion (including spooling space for saving header if necessary).
1013 // then allocate and copy, then track promoted info if needed.
1014 // When tracking (see PromotionInfo::track()), the mark word may
1015 // be displaced and in this case restoration of the mark word
1016 // occurs in the (oop_since_save_marks_)iterate phase.
1017 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1018 // Out of space for allocating spooling buffers;
1019 // try expanding and allocating spooling buffers.
1020 if (!expand_and_ensure_spooling_space(promoInfo)) {
1021 return NULL;
1022 }
1023 }
1024 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant");
1025 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1026 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1027 if (obj_ptr == NULL) {
1028 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1029 if (obj_ptr == NULL) {
1030 return NULL;
1031 }
1032 }
1033 oop obj = oop(obj_ptr);
1034 OrderAccess::storestore();
1035 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1036 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1037 // IMPORTANT: See note on object initialization for CMS above.
1038 // Otherwise, copy the object. Here we must be careful to insert the
1039 // klass pointer last, since this marks the block as an allocated object.
1040 // Except with compressed oops it's the mark word.
1041 HeapWord* old_ptr = (HeapWord*)old;
1042 // Restore the mark word copied above.
1043 obj->set_mark(m);
1044 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1045 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1046 OrderAccess::storestore();
1047
1048 if (UseCompressedClassPointers) {
1049 // Copy gap missed by (aligned) header size calculation below
1050 obj->set_klass_gap(old->klass_gap());
1051 }
1052 if (word_sz > (size_t)oopDesc::header_size()) {
1053 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1054 obj_ptr + oopDesc::header_size(),
1055 word_sz - oopDesc::header_size());
1056 }
1057
1058 // Now we can track the promoted object, if necessary. We take care
1059 // to delay the transition from uninitialized to full object
1060 // (i.e., insertion of klass pointer) until after, so that it
1061 // atomically becomes a promoted object.
1062 if (promoInfo->tracking()) {
1063 promoInfo->track((PromotedObject*)obj, old->klass());
1064 }
1065 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1066 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1067 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1068
1069 // Finally, install the klass pointer (this should be volatile).
1070 OrderAccess::storestore();
1071 obj->set_klass(old->klass());
1072 // We should now be able to calculate the right size for this object
1073 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1074
1075 collector()->promoted(true, // parallel
1076 obj_ptr, old->is_objArray(), word_sz);
1077
1078 NOT_PRODUCT(
1079 Atomic::inc_ptr(&_numObjectsPromoted);
1080 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1081 )
1082
1083 return obj;
1084 }
1085
1086 void
1087 ConcurrentMarkSweepGeneration::
1088 par_promote_alloc_done(int thread_num) {
1089 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1090 ps->lab.retire(thread_num);
1091 }
1092
1093 void
1094 ConcurrentMarkSweepGeneration::
1095 par_oop_since_save_marks_iterate_done(int thread_num) {
1096 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1097 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1098 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1099
1100 // Because card-scanning has been completed, subsequent phases
1101 // (e.g., reference processing) will not need to recognize which
1102 // objects have been promoted during this GC. So, we can now disable
1103 // promotion tracking.
1104 ps->promo.stopTrackingPromotions();
1105 }
1106
1107 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1108 size_t size,
1109 bool tlab)
1110 {
1111 // We allow a STW collection only if a full
1112 // collection was requested.
1113 return full || should_allocate(size, tlab); // FIX ME !!!
1114 // This and promotion failure handling are connected at the
1115 // hip and should be fixed by untying them.
1116 }
1117
1118 bool CMSCollector::shouldConcurrentCollect() {
1119 LogTarget(Trace, gc) log;
1120
1121 if (_full_gc_requested) {
1122 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)");
1123 return true;
1124 }
1125
1126 FreelistLocker x(this);
1127 // ------------------------------------------------------------------
1128 // Print out lots of information which affects the initiation of
1129 // a collection.
1130 if (log.is_enabled() && stats().valid()) {
1131 log.print("CMSCollector shouldConcurrentCollect: ");
1132
1133 LogStream out(log);
1134 stats().print_on(&out);
1135
1136 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full());
1137 log.print("free=" SIZE_FORMAT, _cmsGen->free());
1138 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available());
1139 log.print("promotion_rate=%g", stats().promotion_rate());
1140 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate());
1141 log.print("occupancy=%3.7f", _cmsGen->occupancy());
1142 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1143 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1144 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1145 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect());
1146 }
1147 // ------------------------------------------------------------------
1148
1149 // If the estimated time to complete a cms collection (cms_duration())
1150 // is less than the estimated time remaining until the cms generation
1151 // is full, start a collection.
1152 if (!UseCMSInitiatingOccupancyOnly) {
1153 if (stats().valid()) {
1154 if (stats().time_until_cms_start() == 0.0) {
1155 return true;
1156 }
1157 } else {
1158 // We want to conservatively collect somewhat early in order
1159 // to try and "bootstrap" our CMS/promotion statistics;
1160 // this branch will not fire after the first successful CMS
1161 // collection because the stats should then be valid.
1162 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1163 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f",
1164 _cmsGen->occupancy(), _bootstrap_occupancy);
1165 return true;
1166 }
1167 }
1168 }
1169
1170 // Otherwise, we start a collection cycle if
1171 // old gen want a collection cycle started. Each may use
1172 // an appropriate criterion for making this decision.
1173 // XXX We need to make sure that the gen expansion
1174 // criterion dovetails well with this. XXX NEED TO FIX THIS
1175 if (_cmsGen->should_concurrent_collect()) {
1176 log.print("CMS old gen initiated");
1177 return true;
1178 }
1179
1180 // We start a collection if we believe an incremental collection may fail;
1181 // this is not likely to be productive in practice because it's probably too
1182 // late anyway.
1183 GenCollectedHeap* gch = GenCollectedHeap::heap();
1184 assert(gch->collector_policy()->is_generation_policy(),
1185 "You may want to check the correctness of the following");
1186 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1187 log.print("CMSCollector: collect because incremental collection will fail ");
1188 return true;
1189 }
1190
1191 if (MetaspaceGC::should_concurrent_collect()) {
1192 log.print("CMSCollector: collect for metadata allocation ");
1193 return true;
1194 }
1195
1196 // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1197 if (CMSTriggerInterval >= 0) {
1198 if (CMSTriggerInterval == 0) {
1199 // Trigger always
1200 return true;
1201 }
1202
1203 // Check the CMS time since begin (we do not check the stats validity
1204 // as we want to be able to trigger the first CMS cycle as well)
1205 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1206 if (stats().valid()) {
1207 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1208 stats().cms_time_since_begin());
1209 } else {
1210 log.print("CMSCollector: collect because of trigger interval (first collection)");
1211 }
1212 return true;
1213 }
1214 }
1215
1216 return false;
1217 }
1218
1219 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1220
1221 // Clear _expansion_cause fields of constituent generations
1222 void CMSCollector::clear_expansion_cause() {
1223 _cmsGen->clear_expansion_cause();
1224 }
1225
1226 // We should be conservative in starting a collection cycle. To
1227 // start too eagerly runs the risk of collecting too often in the
1228 // extreme. To collect too rarely falls back on full collections,
1229 // which works, even if not optimum in terms of concurrent work.
1230 // As a work around for too eagerly collecting, use the flag
1231 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1232 // giving the user an easily understandable way of controlling the
1233 // collections.
1234 // We want to start a new collection cycle if any of the following
1235 // conditions hold:
1236 // . our current occupancy exceeds the configured initiating occupancy
1237 // for this generation, or
1238 // . we recently needed to expand this space and have not, since that
1239 // expansion, done a collection of this generation, or
1240 // . the underlying space believes that it may be a good idea to initiate
1241 // a concurrent collection (this may be based on criteria such as the
1242 // following: the space uses linear allocation and linear allocation is
1243 // going to fail, or there is believed to be excessive fragmentation in
1244 // the generation, etc... or ...
1245 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1246 // the case of the old generation; see CR 6543076):
1247 // we may be approaching a point at which allocation requests may fail because
1248 // we will be out of sufficient free space given allocation rate estimates.]
1249 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1250
1251 assert_lock_strong(freelistLock());
1252 if (occupancy() > initiating_occupancy()) {
1253 log_trace(gc)(" %s: collect because of occupancy %f / %f ",
1254 short_name(), occupancy(), initiating_occupancy());
1255 return true;
1256 }
1257 if (UseCMSInitiatingOccupancyOnly) {
1258 return false;
1259 }
1260 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1261 log_trace(gc)(" %s: collect because expanded for allocation ", short_name());
1262 return true;
1263 }
1264 return false;
1265 }
1266
1267 void ConcurrentMarkSweepGeneration::collect(bool full,
1268 bool clear_all_soft_refs,
1269 size_t size,
1270 bool tlab)
1271 {
1272 collector()->collect(full, clear_all_soft_refs, size, tlab);
1273 }
1274
1275 void CMSCollector::collect(bool full,
1276 bool clear_all_soft_refs,
1277 size_t size,
1278 bool tlab)
1279 {
1280 // The following "if" branch is present for defensive reasons.
1281 // In the current uses of this interface, it can be replaced with:
1282 // assert(!GCLocker.is_active(), "Can't be called otherwise");
1283 // But I am not placing that assert here to allow future
1284 // generality in invoking this interface.
1285 if (GCLocker::is_active()) {
1286 // A consistency test for GCLocker
1287 assert(GCLocker::needs_gc(), "Should have been set already");
1288 // Skip this foreground collection, instead
1289 // expanding the heap if necessary.
1290 // Need the free list locks for the call to free() in compute_new_size()
1291 compute_new_size();
1292 return;
1293 }
1294 acquire_control_and_collect(full, clear_all_soft_refs);
1295 }
1296
1297 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1298 GenCollectedHeap* gch = GenCollectedHeap::heap();
1299 unsigned int gc_count = gch->total_full_collections();
1300 if (gc_count == full_gc_count) {
1301 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1302 _full_gc_requested = true;
1303 _full_gc_cause = cause;
1304 CGC_lock->notify(); // nudge CMS thread
1305 } else {
1306 assert(gc_count > full_gc_count, "Error: causal loop");
1307 }
1308 }
1309
1310 bool CMSCollector::is_external_interruption() {
1311 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1312 return GCCause::is_user_requested_gc(cause) ||
1313 GCCause::is_serviceability_requested_gc(cause);
1314 }
1315
1316 void CMSCollector::report_concurrent_mode_interruption() {
1317 if (is_external_interruption()) {
1318 log_debug(gc)("Concurrent mode interrupted");
1319 } else {
1320 log_debug(gc)("Concurrent mode failure");
1321 _gc_tracer_cm->report_concurrent_mode_failure();
1322 }
1323 }
1324
1325
1326 // The foreground and background collectors need to coordinate in order
1327 // to make sure that they do not mutually interfere with CMS collections.
1328 // When a background collection is active,
1329 // the foreground collector may need to take over (preempt) and
1330 // synchronously complete an ongoing collection. Depending on the
1331 // frequency of the background collections and the heap usage
1332 // of the application, this preemption can be seldom or frequent.
1333 // There are only certain
1334 // points in the background collection that the "collection-baton"
1335 // can be passed to the foreground collector.
1336 //
1337 // The foreground collector will wait for the baton before
1338 // starting any part of the collection. The foreground collector
1339 // will only wait at one location.
1340 //
1341 // The background collector will yield the baton before starting a new
1342 // phase of the collection (e.g., before initial marking, marking from roots,
1343 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1344 // of the loop which switches the phases. The background collector does some
1345 // of the phases (initial mark, final re-mark) with the world stopped.
1346 // Because of locking involved in stopping the world,
1347 // the foreground collector should not block waiting for the background
1348 // collector when it is doing a stop-the-world phase. The background
1349 // collector will yield the baton at an additional point just before
1350 // it enters a stop-the-world phase. Once the world is stopped, the
1351 // background collector checks the phase of the collection. If the
1352 // phase has not changed, it proceeds with the collection. If the
1353 // phase has changed, it skips that phase of the collection. See
1354 // the comments on the use of the Heap_lock in collect_in_background().
1355 //
1356 // Variable used in baton passing.
1357 // _foregroundGCIsActive - Set to true by the foreground collector when
1358 // it wants the baton. The foreground clears it when it has finished
1359 // the collection.
1360 // _foregroundGCShouldWait - Set to true by the background collector
1361 // when it is running. The foreground collector waits while
1362 // _foregroundGCShouldWait is true.
1363 // CGC_lock - monitor used to protect access to the above variables
1364 // and to notify the foreground and background collectors.
1365 // _collectorState - current state of the CMS collection.
1366 //
1367 // The foreground collector
1368 // acquires the CGC_lock
1369 // sets _foregroundGCIsActive
1370 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1371 // various locks acquired in preparation for the collection
1372 // are released so as not to block the background collector
1373 // that is in the midst of a collection
1374 // proceeds with the collection
1375 // clears _foregroundGCIsActive
1376 // returns
1377 //
1378 // The background collector in a loop iterating on the phases of the
1379 // collection
1380 // acquires the CGC_lock
1381 // sets _foregroundGCShouldWait
1382 // if _foregroundGCIsActive is set
1383 // clears _foregroundGCShouldWait, notifies _CGC_lock
1384 // waits on _CGC_lock for _foregroundGCIsActive to become false
1385 // and exits the loop.
1386 // otherwise
1387 // proceed with that phase of the collection
1388 // if the phase is a stop-the-world phase,
1389 // yield the baton once more just before enqueueing
1390 // the stop-world CMS operation (executed by the VM thread).
1391 // returns after all phases of the collection are done
1392 //
1393
1394 void CMSCollector::acquire_control_and_collect(bool full,
1395 bool clear_all_soft_refs) {
1396 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1397 assert(!Thread::current()->is_ConcurrentGC_thread(),
1398 "shouldn't try to acquire control from self!");
1399
1400 // Start the protocol for acquiring control of the
1401 // collection from the background collector (aka CMS thread).
1402 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1403 "VM thread should have CMS token");
1404 // Remember the possibly interrupted state of an ongoing
1405 // concurrent collection
1406 CollectorState first_state = _collectorState;
1407
1408 // Signal to a possibly ongoing concurrent collection that
1409 // we want to do a foreground collection.
1410 _foregroundGCIsActive = true;
1411
1412 // release locks and wait for a notify from the background collector
1413 // releasing the locks in only necessary for phases which
1414 // do yields to improve the granularity of the collection.
1415 assert_lock_strong(bitMapLock());
1416 // We need to lock the Free list lock for the space that we are
1417 // currently collecting.
1418 assert(haveFreelistLocks(), "Must be holding free list locks");
1419 bitMapLock()->unlock();
1420 releaseFreelistLocks();
1421 {
1422 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1423 if (_foregroundGCShouldWait) {
1424 // We are going to be waiting for action for the CMS thread;
1425 // it had better not be gone (for instance at shutdown)!
1426 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(),
1427 "CMS thread must be running");
1428 // Wait here until the background collector gives us the go-ahead
1429 ConcurrentMarkSweepThread::clear_CMS_flag(
1430 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1431 // Get a possibly blocked CMS thread going:
1432 // Note that we set _foregroundGCIsActive true above,
1433 // without protection of the CGC_lock.
1434 CGC_lock->notify();
1435 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1436 "Possible deadlock");
1437 while (_foregroundGCShouldWait) {
1438 // wait for notification
1439 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1440 // Possibility of delay/starvation here, since CMS token does
1441 // not know to give priority to VM thread? Actually, i think
1442 // there wouldn't be any delay/starvation, but the proof of
1443 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1444 }
1445 ConcurrentMarkSweepThread::set_CMS_flag(
1446 ConcurrentMarkSweepThread::CMS_vm_has_token);
1447 }
1448 }
1449 // The CMS_token is already held. Get back the other locks.
1450 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1451 "VM thread should have CMS token");
1452 getFreelistLocks();
1453 bitMapLock()->lock_without_safepoint_check();
1454 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d",
1455 p2i(Thread::current()), first_state);
1456 log_debug(gc, state)(" gets control with state %d", _collectorState);
1457
1458 // Inform cms gen if this was due to partial collection failing.
1459 // The CMS gen may use this fact to determine its expansion policy.
1460 GenCollectedHeap* gch = GenCollectedHeap::heap();
1461 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1462 assert(!_cmsGen->incremental_collection_failed(),
1463 "Should have been noticed, reacted to and cleared");
1464 _cmsGen->set_incremental_collection_failed();
1465 }
1466
1467 if (first_state > Idling) {
1468 report_concurrent_mode_interruption();
1469 }
1470
1471 set_did_compact(true);
1472
1473 // If the collection is being acquired from the background
1474 // collector, there may be references on the discovered
1475 // references lists. Abandon those references, since some
1476 // of them may have become unreachable after concurrent
1477 // discovery; the STW compacting collector will redo discovery
1478 // more precisely, without being subject to floating garbage.
1479 // Leaving otherwise unreachable references in the discovered
1480 // lists would require special handling.
1481 ref_processor()->disable_discovery();
1482 ref_processor()->abandon_partial_discovery();
1483 ref_processor()->verify_no_references_recorded();
1484
1485 if (first_state > Idling) {
1486 save_heap_summary();
1487 }
1488
1489 do_compaction_work(clear_all_soft_refs);
1490
1491 // Has the GC time limit been exceeded?
1492 size_t max_eden_size = _young_gen->max_eden_size();
1493 GCCause::Cause gc_cause = gch->gc_cause();
1494 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1495 _young_gen->eden()->used(),
1496 _cmsGen->max_capacity(),
1497 max_eden_size,
1498 full,
1499 gc_cause,
1500 gch->collector_policy());
1501
1502 // Reset the expansion cause, now that we just completed
1503 // a collection cycle.
1504 clear_expansion_cause();
1505 _foregroundGCIsActive = false;
1506 return;
1507 }
1508
1509 // Resize the tenured generation
1510 // after obtaining the free list locks for the
1511 // two generations.
1512 void CMSCollector::compute_new_size() {
1513 assert_locked_or_safepoint(Heap_lock);
1514 FreelistLocker z(this);
1515 MetaspaceGC::compute_new_size();
1516 _cmsGen->compute_new_size_free_list();
1517 }
1518
1519 // A work method used by the foreground collector to do
1520 // a mark-sweep-compact.
1521 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1522 GenCollectedHeap* gch = GenCollectedHeap::heap();
1523
1524 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1525 gc_timer->register_gc_start();
1526
1527 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1528 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
1529
1530 gch->pre_full_gc_dump(gc_timer);
1531
1532 GCTraceTime(Trace, gc, phases) t("CMS:MSC");
1533
1534 // Temporarily widen the span of the weak reference processing to
1535 // the entire heap.
1536 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1537 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1538 // Temporarily, clear the "is_alive_non_header" field of the
1539 // reference processor.
1540 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1541 // Temporarily make reference _processing_ single threaded (non-MT).
1542 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1543 // Temporarily make refs discovery atomic
1544 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1545 // Temporarily make reference _discovery_ single threaded (non-MT)
1546 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1547
1548 ref_processor()->set_enqueuing_is_done(false);
1549 ref_processor()->enable_discovery();
1550 ref_processor()->setup_policy(clear_all_soft_refs);
1551 // If an asynchronous collection finishes, the _modUnionTable is
1552 // all clear. If we are assuming the collection from an asynchronous
1553 // collection, clear the _modUnionTable.
1554 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1555 "_modUnionTable should be clear if the baton was not passed");
1556 _modUnionTable.clear_all();
1557 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
1558 "mod union for klasses should be clear if the baton was passed");
1559 _ct->klass_rem_set()->clear_mod_union();
1560
1561 // We must adjust the allocation statistics being maintained
1562 // in the free list space. We do so by reading and clearing
1563 // the sweep timer and updating the block flux rate estimates below.
1564 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1565 if (_inter_sweep_timer.is_active()) {
1566 _inter_sweep_timer.stop();
1567 // Note that we do not use this sample to update the _inter_sweep_estimate.
1568 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1569 _inter_sweep_estimate.padded_average(),
1570 _intra_sweep_estimate.padded_average());
1571 }
1572
1573 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs);
1574 #ifdef ASSERT
1575 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1576 size_t free_size = cms_space->free();
1577 assert(free_size ==
1578 pointer_delta(cms_space->end(), cms_space->compaction_top())
1579 * HeapWordSize,
1580 "All the free space should be compacted into one chunk at top");
1581 assert(cms_space->dictionary()->total_chunk_size(
1582 debug_only(cms_space->freelistLock())) == 0 ||
1583 cms_space->totalSizeInIndexedFreeLists() == 0,
1584 "All the free space should be in a single chunk");
1585 size_t num = cms_space->totalCount();
1586 assert((free_size == 0 && num == 0) ||
1587 (free_size > 0 && (num == 1 || num == 2)),
1588 "There should be at most 2 free chunks after compaction");
1589 #endif // ASSERT
1590 _collectorState = Resetting;
1591 assert(_restart_addr == NULL,
1592 "Should have been NULL'd before baton was passed");
1593 reset_stw();
1594 _cmsGen->reset_after_compaction();
1595 _concurrent_cycles_since_last_unload = 0;
1596
1597 // Clear any data recorded in the PLAB chunk arrays.
1598 if (_survivor_plab_array != NULL) {
1599 reset_survivor_plab_arrays();
1600 }
1601
1602 // Adjust the per-size allocation stats for the next epoch.
1603 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1604 // Restart the "inter sweep timer" for the next epoch.
1605 _inter_sweep_timer.reset();
1606 _inter_sweep_timer.start();
1607
1608 // No longer a need to do a concurrent collection for Metaspace.
1609 MetaspaceGC::set_should_concurrent_collect(false);
1610
1611 gch->post_full_gc_dump(gc_timer);
1612
1613 gc_timer->register_gc_end();
1614
1615 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1616
1617 // For a mark-sweep-compact, compute_new_size() will be called
1618 // in the heap's do_collection() method.
1619 }
1620
1621 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1622 Log(gc, heap) log;
1623 if (!log.is_trace()) {
1624 return;
1625 }
1626
1627 ContiguousSpace* eden_space = _young_gen->eden();
1628 ContiguousSpace* from_space = _young_gen->from();
1629 ContiguousSpace* to_space = _young_gen->to();
1630 // Eden
1631 if (_eden_chunk_array != NULL) {
1632 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1633 p2i(eden_space->bottom()), p2i(eden_space->top()),
1634 p2i(eden_space->end()), eden_space->capacity());
1635 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT,
1636 _eden_chunk_index, _eden_chunk_capacity);
1637 for (size_t i = 0; i < _eden_chunk_index; i++) {
1638 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i]));
1639 }
1640 }
1641 // Survivor
1642 if (_survivor_chunk_array != NULL) {
1643 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1644 p2i(from_space->bottom()), p2i(from_space->top()),
1645 p2i(from_space->end()), from_space->capacity());
1646 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT,
1647 _survivor_chunk_index, _survivor_chunk_capacity);
1648 for (size_t i = 0; i < _survivor_chunk_index; i++) {
1649 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i]));
1650 }
1651 }
1652 }
1653
1654 void CMSCollector::getFreelistLocks() const {
1655 // Get locks for all free lists in all generations that this
1656 // collector is responsible for
1657 _cmsGen->freelistLock()->lock_without_safepoint_check();
1658 }
1659
1660 void CMSCollector::releaseFreelistLocks() const {
1661 // Release locks for all free lists in all generations that this
1662 // collector is responsible for
1663 _cmsGen->freelistLock()->unlock();
1664 }
1665
1666 bool CMSCollector::haveFreelistLocks() const {
1667 // Check locks for all free lists in all generations that this
1668 // collector is responsible for
1669 assert_lock_strong(_cmsGen->freelistLock());
1670 PRODUCT_ONLY(ShouldNotReachHere());
1671 return true;
1672 }
1673
1674 // A utility class that is used by the CMS collector to
1675 // temporarily "release" the foreground collector from its
1676 // usual obligation to wait for the background collector to
1677 // complete an ongoing phase before proceeding.
1678 class ReleaseForegroundGC: public StackObj {
1679 private:
1680 CMSCollector* _c;
1681 public:
1682 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1683 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1684 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1685 // allow a potentially blocked foreground collector to proceed
1686 _c->_foregroundGCShouldWait = false;
1687 if (_c->_foregroundGCIsActive) {
1688 CGC_lock->notify();
1689 }
1690 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1691 "Possible deadlock");
1692 }
1693
1694 ~ReleaseForegroundGC() {
1695 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1696 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1697 _c->_foregroundGCShouldWait = true;
1698 }
1699 };
1700
1701 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1702 assert(Thread::current()->is_ConcurrentGC_thread(),
1703 "A CMS asynchronous collection is only allowed on a CMS thread.");
1704
1705 GenCollectedHeap* gch = GenCollectedHeap::heap();
1706 {
1707 bool safepoint_check = Mutex::_no_safepoint_check_flag;
1708 MutexLockerEx hl(Heap_lock, safepoint_check);
1709 FreelistLocker fll(this);
1710 MutexLockerEx x(CGC_lock, safepoint_check);
1711 if (_foregroundGCIsActive) {
1712 // The foreground collector is. Skip this
1713 // background collection.
1714 assert(!_foregroundGCShouldWait, "Should be clear");
1715 return;
1716 } else {
1717 assert(_collectorState == Idling, "Should be idling before start.");
1718 _collectorState = InitialMarking;
1719 register_gc_start(cause);
1720 // Reset the expansion cause, now that we are about to begin
1721 // a new cycle.
1722 clear_expansion_cause();
1723
1724 // Clear the MetaspaceGC flag since a concurrent collection
1725 // is starting but also clear it after the collection.
1726 MetaspaceGC::set_should_concurrent_collect(false);
1727 }
1728 // Decide if we want to enable class unloading as part of the
1729 // ensuing concurrent GC cycle.
1730 update_should_unload_classes();
1731 _full_gc_requested = false; // acks all outstanding full gc requests
1732 _full_gc_cause = GCCause::_no_gc;
1733 // Signal that we are about to start a collection
1734 gch->increment_total_full_collections(); // ... starting a collection cycle
1735 _collection_count_start = gch->total_full_collections();
1736 }
1737
1738 size_t prev_used = _cmsGen->used();
1739
1740 // The change of the collection state is normally done at this level;
1741 // the exceptions are phases that are executed while the world is
1742 // stopped. For those phases the change of state is done while the
1743 // world is stopped. For baton passing purposes this allows the
1744 // background collector to finish the phase and change state atomically.
1745 // The foreground collector cannot wait on a phase that is done
1746 // while the world is stopped because the foreground collector already
1747 // has the world stopped and would deadlock.
1748 while (_collectorState != Idling) {
1749 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d",
1750 p2i(Thread::current()), _collectorState);
1751 // The foreground collector
1752 // holds the Heap_lock throughout its collection.
1753 // holds the CMS token (but not the lock)
1754 // except while it is waiting for the background collector to yield.
1755 //
1756 // The foreground collector should be blocked (not for long)
1757 // if the background collector is about to start a phase
1758 // executed with world stopped. If the background
1759 // collector has already started such a phase, the
1760 // foreground collector is blocked waiting for the
1761 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
1762 // are executed in the VM thread.
1763 //
1764 // The locking order is
1765 // PendingListLock (PLL) -- if applicable (FinalMarking)
1766 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
1767 // CMS token (claimed in
1768 // stop_world_and_do() -->
1769 // safepoint_synchronize() -->
1770 // CMSThread::synchronize())
1771
1772 {
1773 // Check if the FG collector wants us to yield.
1774 CMSTokenSync x(true); // is cms thread
1775 if (waitForForegroundGC()) {
1776 // We yielded to a foreground GC, nothing more to be
1777 // done this round.
1778 assert(_foregroundGCShouldWait == false, "We set it to false in "
1779 "waitForForegroundGC()");
1780 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1781 p2i(Thread::current()), _collectorState);
1782 return;
1783 } else {
1784 // The background collector can run but check to see if the
1785 // foreground collector has done a collection while the
1786 // background collector was waiting to get the CGC_lock
1787 // above. If yes, break so that _foregroundGCShouldWait
1788 // is cleared before returning.
1789 if (_collectorState == Idling) {
1790 break;
1791 }
1792 }
1793 }
1794
1795 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1796 "should be waiting");
1797
1798 switch (_collectorState) {
1799 case InitialMarking:
1800 {
1801 ReleaseForegroundGC x(this);
1802 stats().record_cms_begin();
1803 VM_CMS_Initial_Mark initial_mark_op(this);
1804 VMThread::execute(&initial_mark_op);
1805 }
1806 // The collector state may be any legal state at this point
1807 // since the background collector may have yielded to the
1808 // foreground collector.
1809 break;
1810 case Marking:
1811 // initial marking in checkpointRootsInitialWork has been completed
1812 if (markFromRoots()) { // we were successful
1813 assert(_collectorState == Precleaning, "Collector state should "
1814 "have changed");
1815 } else {
1816 assert(_foregroundGCIsActive, "Internal state inconsistency");
1817 }
1818 break;
1819 case Precleaning:
1820 // marking from roots in markFromRoots has been completed
1821 preclean();
1822 assert(_collectorState == AbortablePreclean ||
1823 _collectorState == FinalMarking,
1824 "Collector state should have changed");
1825 break;
1826 case AbortablePreclean:
1827 abortable_preclean();
1828 assert(_collectorState == FinalMarking, "Collector state should "
1829 "have changed");
1830 break;
1831 case FinalMarking:
1832 {
1833 ReleaseForegroundGC x(this);
1834
1835 VM_CMS_Final_Remark final_remark_op(this);
1836 VMThread::execute(&final_remark_op);
1837 }
1838 assert(_foregroundGCShouldWait, "block post-condition");
1839 break;
1840 case Sweeping:
1841 // final marking in checkpointRootsFinal has been completed
1842 sweep();
1843 assert(_collectorState == Resizing, "Collector state change "
1844 "to Resizing must be done under the free_list_lock");
1845
1846 case Resizing: {
1847 // Sweeping has been completed...
1848 // At this point the background collection has completed.
1849 // Don't move the call to compute_new_size() down
1850 // into code that might be executed if the background
1851 // collection was preempted.
1852 {
1853 ReleaseForegroundGC x(this); // unblock FG collection
1854 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
1855 CMSTokenSync z(true); // not strictly needed.
1856 if (_collectorState == Resizing) {
1857 compute_new_size();
1858 save_heap_summary();
1859 _collectorState = Resetting;
1860 } else {
1861 assert(_collectorState == Idling, "The state should only change"
1862 " because the foreground collector has finished the collection");
1863 }
1864 }
1865 break;
1866 }
1867 case Resetting:
1868 // CMS heap resizing has been completed
1869 reset_concurrent();
1870 assert(_collectorState == Idling, "Collector state should "
1871 "have changed");
1872
1873 MetaspaceGC::set_should_concurrent_collect(false);
1874
1875 stats().record_cms_end();
1876 // Don't move the concurrent_phases_end() and compute_new_size()
1877 // calls to here because a preempted background collection
1878 // has it's state set to "Resetting".
1879 break;
1880 case Idling:
1881 default:
1882 ShouldNotReachHere();
1883 break;
1884 }
1885 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d",
1886 p2i(Thread::current()), _collectorState);
1887 assert(_foregroundGCShouldWait, "block post-condition");
1888 }
1889
1890 // Should this be in gc_epilogue?
1891 collector_policy()->counters()->update_counters();
1892
1893 {
1894 // Clear _foregroundGCShouldWait and, in the event that the
1895 // foreground collector is waiting, notify it, before
1896 // returning.
1897 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1898 _foregroundGCShouldWait = false;
1899 if (_foregroundGCIsActive) {
1900 CGC_lock->notify();
1901 }
1902 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1903 "Possible deadlock");
1904 }
1905 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1906 p2i(Thread::current()), _collectorState);
1907 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)",
1908 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K);
1909 }
1910
1911 void CMSCollector::register_gc_start(GCCause::Cause cause) {
1912 _cms_start_registered = true;
1913 _gc_timer_cm->register_gc_start();
1914 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
1915 }
1916
1917 void CMSCollector::register_gc_end() {
1918 if (_cms_start_registered) {
1919 report_heap_summary(GCWhen::AfterGC);
1920
1921 _gc_timer_cm->register_gc_end();
1922 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1923 _cms_start_registered = false;
1924 }
1925 }
1926
1927 void CMSCollector::save_heap_summary() {
1928 GenCollectedHeap* gch = GenCollectedHeap::heap();
1929 _last_heap_summary = gch->create_heap_summary();
1930 _last_metaspace_summary = gch->create_metaspace_summary();
1931 }
1932
1933 void CMSCollector::report_heap_summary(GCWhen::Type when) {
1934 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
1935 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
1936 }
1937
1938 bool CMSCollector::waitForForegroundGC() {
1939 bool res = false;
1940 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1941 "CMS thread should have CMS token");
1942 // Block the foreground collector until the
1943 // background collectors decides whether to
1944 // yield.
1945 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1946 _foregroundGCShouldWait = true;
1947 if (_foregroundGCIsActive) {
1948 // The background collector yields to the
1949 // foreground collector and returns a value
1950 // indicating that it has yielded. The foreground
1951 // collector can proceed.
1952 res = true;
1953 _foregroundGCShouldWait = false;
1954 ConcurrentMarkSweepThread::clear_CMS_flag(
1955 ConcurrentMarkSweepThread::CMS_cms_has_token);
1956 ConcurrentMarkSweepThread::set_CMS_flag(
1957 ConcurrentMarkSweepThread::CMS_cms_wants_token);
1958 // Get a possibly blocked foreground thread going
1959 CGC_lock->notify();
1960 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
1961 p2i(Thread::current()), _collectorState);
1962 while (_foregroundGCIsActive) {
1963 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1964 }
1965 ConcurrentMarkSweepThread::set_CMS_flag(
1966 ConcurrentMarkSweepThread::CMS_cms_has_token);
1967 ConcurrentMarkSweepThread::clear_CMS_flag(
1968 ConcurrentMarkSweepThread::CMS_cms_wants_token);
1969 }
1970 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
1971 p2i(Thread::current()), _collectorState);
1972 return res;
1973 }
1974
1975 // Because of the need to lock the free lists and other structures in
1976 // the collector, common to all the generations that the collector is
1977 // collecting, we need the gc_prologues of individual CMS generations
1978 // delegate to their collector. It may have been simpler had the
1979 // current infrastructure allowed one to call a prologue on a
1980 // collector. In the absence of that we have the generation's
1981 // prologue delegate to the collector, which delegates back
1982 // some "local" work to a worker method in the individual generations
1983 // that it's responsible for collecting, while itself doing any
1984 // work common to all generations it's responsible for. A similar
1985 // comment applies to the gc_epilogue()'s.
1986 // The role of the variable _between_prologue_and_epilogue is to
1987 // enforce the invocation protocol.
1988 void CMSCollector::gc_prologue(bool full) {
1989 // Call gc_prologue_work() for the CMSGen
1990 // we are responsible for.
1991
1992 // The following locking discipline assumes that we are only called
1993 // when the world is stopped.
1994 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
1995
1996 // The CMSCollector prologue must call the gc_prologues for the
1997 // "generations" that it's responsible
1998 // for.
1999
2000 assert( Thread::current()->is_VM_thread()
2001 || ( CMSScavengeBeforeRemark
2002 && Thread::current()->is_ConcurrentGC_thread()),
2003 "Incorrect thread type for prologue execution");
2004
2005 if (_between_prologue_and_epilogue) {
2006 // We have already been invoked; this is a gc_prologue delegation
2007 // from yet another CMS generation that we are responsible for, just
2008 // ignore it since all relevant work has already been done.
2009 return;
2010 }
2011
2012 // set a bit saying prologue has been called; cleared in epilogue
2013 _between_prologue_and_epilogue = true;
2014 // Claim locks for common data structures, then call gc_prologue_work()
2015 // for each CMSGen.
2016
2017 getFreelistLocks(); // gets free list locks on constituent spaces
2018 bitMapLock()->lock_without_safepoint_check();
2019
2020 // Should call gc_prologue_work() for all cms gens we are responsible for
2021 bool duringMarking = _collectorState >= Marking
2022 && _collectorState < Sweeping;
2023
2024 // The young collections clear the modified oops state, which tells if
2025 // there are any modified oops in the class. The remark phase also needs
2026 // that information. Tell the young collection to save the union of all
2027 // modified klasses.
2028 if (duringMarking) {
2029 _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2030 }
2031
2032 bool registerClosure = duringMarking;
2033
2034 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2035
2036 if (!full) {
2037 stats().record_gc0_begin();
2038 }
2039 }
2040
2041 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2042
2043 _capacity_at_prologue = capacity();
2044 _used_at_prologue = used();
2045
2046 // We enable promotion tracking so that card-scanning can recognize
2047 // which objects have been promoted during this GC and skip them.
2048 for (uint i = 0; i < ParallelGCThreads; i++) {
2049 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2050 }
2051
2052 // Delegate to CMScollector which knows how to coordinate between
2053 // this and any other CMS generations that it is responsible for
2054 // collecting.
2055 collector()->gc_prologue(full);
2056 }
2057
2058 // This is a "private" interface for use by this generation's CMSCollector.
2059 // Not to be called directly by any other entity (for instance,
2060 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2061 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2062 bool registerClosure, ModUnionClosure* modUnionClosure) {
2063 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2064 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2065 "Should be NULL");
2066 if (registerClosure) {
2067 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2068 }
2069 cmsSpace()->gc_prologue();
2070 // Clear stat counters
2071 NOT_PRODUCT(
2072 assert(_numObjectsPromoted == 0, "check");
2073 assert(_numWordsPromoted == 0, "check");
2074 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently",
2075 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2076 _numObjectsAllocated = 0;
2077 _numWordsAllocated = 0;
2078 )
2079 }
2080
2081 void CMSCollector::gc_epilogue(bool full) {
2082 // The following locking discipline assumes that we are only called
2083 // when the world is stopped.
2084 assert(SafepointSynchronize::is_at_safepoint(),
2085 "world is stopped assumption");
2086
2087 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2088 // if linear allocation blocks need to be appropriately marked to allow the
2089 // the blocks to be parsable. We also check here whether we need to nudge the
2090 // CMS collector thread to start a new cycle (if it's not already active).
2091 assert( Thread::current()->is_VM_thread()
2092 || ( CMSScavengeBeforeRemark
2093 && Thread::current()->is_ConcurrentGC_thread()),
2094 "Incorrect thread type for epilogue execution");
2095
2096 if (!_between_prologue_and_epilogue) {
2097 // We have already been invoked; this is a gc_epilogue delegation
2098 // from yet another CMS generation that we are responsible for, just
2099 // ignore it since all relevant work has already been done.
2100 return;
2101 }
2102 assert(haveFreelistLocks(), "must have freelist locks");
2103 assert_lock_strong(bitMapLock());
2104
2105 _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2106
2107 _cmsGen->gc_epilogue_work(full);
2108
2109 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2110 // in case sampling was not already enabled, enable it
2111 _start_sampling = true;
2112 }
2113 // reset _eden_chunk_array so sampling starts afresh
2114 _eden_chunk_index = 0;
2115
2116 size_t cms_used = _cmsGen->cmsSpace()->used();
2117
2118 // update performance counters - this uses a special version of
2119 // update_counters() that allows the utilization to be passed as a
2120 // parameter, avoiding multiple calls to used().
2121 //
2122 _cmsGen->update_counters(cms_used);
2123
2124 bitMapLock()->unlock();
2125 releaseFreelistLocks();
2126
2127 if (!CleanChunkPoolAsync) {
2128 Chunk::clean_chunk_pool();
2129 }
2130
2131 set_did_compact(false);
2132 _between_prologue_and_epilogue = false; // ready for next cycle
2133 }
2134
2135 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2136 collector()->gc_epilogue(full);
2137
2138 // When using ParNew, promotion tracking should have already been
2139 // disabled. However, the prologue (which enables promotion
2140 // tracking) and epilogue are called irrespective of the type of
2141 // GC. So they will also be called before and after Full GCs, during
2142 // which promotion tracking will not be explicitly disabled. So,
2143 // it's safer to also disable it here too (to be symmetric with
2144 // enabling it in the prologue).
2145 for (uint i = 0; i < ParallelGCThreads; i++) {
2146 _par_gc_thread_states[i]->promo.stopTrackingPromotions();
2147 }
2148 }
2149
2150 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2151 assert(!incremental_collection_failed(), "Should have been cleared");
2152 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2153 cmsSpace()->gc_epilogue();
2154 // Print stat counters
2155 NOT_PRODUCT(
2156 assert(_numObjectsAllocated == 0, "check");
2157 assert(_numWordsAllocated == 0, "check");
2158 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
2159 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2160 _numObjectsPromoted = 0;
2161 _numWordsPromoted = 0;
2162 )
2163
2164 // Call down the chain in contiguous_available needs the freelistLock
2165 // so print this out before releasing the freeListLock.
2166 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available());
2167 }
2168
2169 #ifndef PRODUCT
2170 bool CMSCollector::have_cms_token() {
2171 Thread* thr = Thread::current();
2172 if (thr->is_VM_thread()) {
2173 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2174 } else if (thr->is_ConcurrentGC_thread()) {
2175 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2176 } else if (thr->is_GC_task_thread()) {
2177 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2178 ParGCRareEvent_lock->owned_by_self();
2179 }
2180 return false;
2181 }
2182
2183 // Check reachability of the given heap address in CMS generation,
2184 // treating all other generations as roots.
2185 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2186 // We could "guarantee" below, rather than assert, but I'll
2187 // leave these as "asserts" so that an adventurous debugger
2188 // could try this in the product build provided some subset of
2189 // the conditions were met, provided they were interested in the
2190 // results and knew that the computation below wouldn't interfere
2191 // with other concurrent computations mutating the structures
2192 // being read or written.
2193 assert(SafepointSynchronize::is_at_safepoint(),
2194 "Else mutations in object graph will make answer suspect");
2195 assert(have_cms_token(), "Should hold cms token");
2196 assert(haveFreelistLocks(), "must hold free list locks");
2197 assert_lock_strong(bitMapLock());
2198
2199 // Clear the marking bit map array before starting, but, just
2200 // for kicks, first report if the given address is already marked
2201 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr),
2202 _markBitMap.isMarked(addr) ? "" : " not");
2203
2204 if (verify_after_remark()) {
2205 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2206 bool result = verification_mark_bm()->isMarked(addr);
2207 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr),
2208 result ? "IS" : "is NOT");
2209 return result;
2210 } else {
2211 tty->print_cr("Could not compute result");
2212 return false;
2213 }
2214 }
2215 #endif
2216
2217 void
2218 CMSCollector::print_on_error(outputStream* st) {
2219 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2220 if (collector != NULL) {
2221 CMSBitMap* bitmap = &collector->_markBitMap;
2222 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2223 bitmap->print_on_error(st, " Bits: ");
2224
2225 st->cr();
2226
2227 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2228 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2229 mut_bitmap->print_on_error(st, " Bits: ");
2230 }
2231 }
2232
2233 ////////////////////////////////////////////////////////
2234 // CMS Verification Support
2235 ////////////////////////////////////////////////////////
2236 // Following the remark phase, the following invariant
2237 // should hold -- each object in the CMS heap which is
2238 // marked in markBitMap() should be marked in the verification_mark_bm().
2239
2240 class VerifyMarkedClosure: public BitMapClosure {
2241 CMSBitMap* _marks;
2242 bool _failed;
2243
2244 public:
2245 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2246
2247 bool do_bit(size_t offset) {
2248 HeapWord* addr = _marks->offsetToHeapWord(offset);
2249 if (!_marks->isMarked(addr)) {
2250 Log(gc, verify) log;
2251 ResourceMark rm;
2252 oop(addr)->print_on(log.error_stream());
2253 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
2254 _failed = true;
2255 }
2256 return true;
2257 }
2258
2259 bool failed() { return _failed; }
2260 };
2261
2262 bool CMSCollector::verify_after_remark() {
2263 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking.");
2264 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2265 static bool init = false;
2266
2267 assert(SafepointSynchronize::is_at_safepoint(),
2268 "Else mutations in object graph will make answer suspect");
2269 assert(have_cms_token(),
2270 "Else there may be mutual interference in use of "
2271 " verification data structures");
2272 assert(_collectorState > Marking && _collectorState <= Sweeping,
2273 "Else marking info checked here may be obsolete");
2274 assert(haveFreelistLocks(), "must hold free list locks");
2275 assert_lock_strong(bitMapLock());
2276
2277
2278 // Allocate marking bit map if not already allocated
2279 if (!init) { // first time
2280 if (!verification_mark_bm()->allocate(_span)) {
2281 return false;
2282 }
2283 init = true;
2284 }
2285
2286 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2287
2288 // Turn off refs discovery -- so we will be tracing through refs.
2289 // This is as intended, because by this time
2290 // GC must already have cleared any refs that need to be cleared,
2291 // and traced those that need to be marked; moreover,
2292 // the marking done here is not going to interfere in any
2293 // way with the marking information used by GC.
2294 NoRefDiscovery no_discovery(ref_processor());
2295
2296 #if defined(COMPILER2) || INCLUDE_JVMCI
2297 DerivedPointerTableDeactivate dpt_deact;
2298 #endif
2299
2300 // Clear any marks from a previous round
2301 verification_mark_bm()->clear_all();
2302 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2303 verify_work_stacks_empty();
2304
2305 GenCollectedHeap* gch = GenCollectedHeap::heap();
2306 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2307 // Update the saved marks which may affect the root scans.
2308 gch->save_marks();
2309
2310 if (CMSRemarkVerifyVariant == 1) {
2311 // In this first variant of verification, we complete
2312 // all marking, then check if the new marks-vector is
2313 // a subset of the CMS marks-vector.
2314 verify_after_remark_work_1();
2315 } else {
2316 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2");
2317 // In this second variant of verification, we flag an error
2318 // (i.e. an object reachable in the new marks-vector not reachable
2319 // in the CMS marks-vector) immediately, also indicating the
2320 // identify of an object (A) that references the unmarked object (B) --
2321 // presumably, a mutation to A failed to be picked up by preclean/remark?
2322 verify_after_remark_work_2();
2323 }
2324
2325 return true;
2326 }
2327
2328 void CMSCollector::verify_after_remark_work_1() {
2329 ResourceMark rm;
2330 HandleMark hm;
2331 GenCollectedHeap* gch = GenCollectedHeap::heap();
2332
2333 // Get a clear set of claim bits for the roots processing to work with.
2334 ClassLoaderDataGraph::clear_claimed_marks();
2335
2336 // Mark from roots one level into CMS
2337 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2338 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2339
2340 {
2341 StrongRootsScope srs(1);
2342
2343 gch->gen_process_roots(&srs,
2344 GenCollectedHeap::OldGen,
2345 true, // young gen as roots
2346 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2347 should_unload_classes(),
2348 ¬Older,
2349 NULL,
2350 NULL);
2351 }
2352
2353 // Now mark from the roots
2354 MarkFromRootsClosure markFromRootsClosure(this, _span,
2355 verification_mark_bm(), verification_mark_stack(),
2356 false /* don't yield */, true /* verifying */);
2357 assert(_restart_addr == NULL, "Expected pre-condition");
2358 verification_mark_bm()->iterate(&markFromRootsClosure);
2359 while (_restart_addr != NULL) {
2360 // Deal with stack overflow: by restarting at the indicated
2361 // address.
2362 HeapWord* ra = _restart_addr;
2363 markFromRootsClosure.reset(ra);
2364 _restart_addr = NULL;
2365 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2366 }
2367 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2368 verify_work_stacks_empty();
2369
2370 // Marking completed -- now verify that each bit marked in
2371 // verification_mark_bm() is also marked in markBitMap(); flag all
2372 // errors by printing corresponding objects.
2373 VerifyMarkedClosure vcl(markBitMap());
2374 verification_mark_bm()->iterate(&vcl);
2375 if (vcl.failed()) {
2376 Log(gc, verify) log;
2377 log.error("Failed marking verification after remark");
2378 ResourceMark rm;
2379 gch->print_on(log.error_stream());
2380 fatal("CMS: failed marking verification after remark");
2381 }
2382 }
2383
2384 class VerifyKlassOopsKlassClosure : public KlassClosure {
2385 class VerifyKlassOopsClosure : public OopClosure {
2386 CMSBitMap* _bitmap;
2387 public:
2388 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2389 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2390 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2391 } _oop_closure;
2392 public:
2393 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2394 void do_klass(Klass* k) {
2395 k->oops_do(&_oop_closure);
2396 }
2397 };
2398
2399 void CMSCollector::verify_after_remark_work_2() {
2400 ResourceMark rm;
2401 HandleMark hm;
2402 GenCollectedHeap* gch = GenCollectedHeap::heap();
2403
2404 // Get a clear set of claim bits for the roots processing to work with.
2405 ClassLoaderDataGraph::clear_claimed_marks();
2406
2407 // Mark from roots one level into CMS
2408 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2409 markBitMap());
2410 CLDToOopClosure cld_closure(¬Older, true);
2411
2412 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2413
2414 {
2415 StrongRootsScope srs(1);
2416
2417 gch->gen_process_roots(&srs,
2418 GenCollectedHeap::OldGen,
2419 true, // young gen as roots
2420 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2421 should_unload_classes(),
2422 ¬Older,
2423 NULL,
2424 &cld_closure);
2425 }
2426
2427 // Now mark from the roots
2428 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2429 verification_mark_bm(), markBitMap(), verification_mark_stack());
2430 assert(_restart_addr == NULL, "Expected pre-condition");
2431 verification_mark_bm()->iterate(&markFromRootsClosure);
2432 while (_restart_addr != NULL) {
2433 // Deal with stack overflow: by restarting at the indicated
2434 // address.
2435 HeapWord* ra = _restart_addr;
2436 markFromRootsClosure.reset(ra);
2437 _restart_addr = NULL;
2438 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2439 }
2440 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2441 verify_work_stacks_empty();
2442
2443 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
2444 ClassLoaderDataGraph::classes_do(&verify_klass_oops);
2445
2446 // Marking completed -- now verify that each bit marked in
2447 // verification_mark_bm() is also marked in markBitMap(); flag all
2448 // errors by printing corresponding objects.
2449 VerifyMarkedClosure vcl(markBitMap());
2450 verification_mark_bm()->iterate(&vcl);
2451 assert(!vcl.failed(), "Else verification above should not have succeeded");
2452 }
2453
2454 void ConcurrentMarkSweepGeneration::save_marks() {
2455 // delegate to CMS space
2456 cmsSpace()->save_marks();
2457 }
2458
2459 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2460 return cmsSpace()->no_allocs_since_save_marks();
2461 }
2462
2463 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2464 \
2465 void ConcurrentMarkSweepGeneration:: \
2466 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2467 cl->set_generation(this); \
2468 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2469 cl->reset_generation(); \
2470 save_marks(); \
2471 }
2472
2473 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2474
2475 void
2476 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2477 if (freelistLock()->owned_by_self()) {
2478 Generation::oop_iterate(cl);
2479 } else {
2480 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2481 Generation::oop_iterate(cl);
2482 }
2483 }
2484
2485 void
2486 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2487 if (freelistLock()->owned_by_self()) {
2488 Generation::object_iterate(cl);
2489 } else {
2490 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2491 Generation::object_iterate(cl);
2492 }
2493 }
2494
2495 void
2496 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2497 if (freelistLock()->owned_by_self()) {
2498 Generation::safe_object_iterate(cl);
2499 } else {
2500 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2501 Generation::safe_object_iterate(cl);
2502 }
2503 }
2504
2505 void
2506 ConcurrentMarkSweepGeneration::post_compact() {
2507 }
2508
2509 void
2510 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2511 // Fix the linear allocation blocks to look like free blocks.
2512
2513 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2514 // are not called when the heap is verified during universe initialization and
2515 // at vm shutdown.
2516 if (freelistLock()->owned_by_self()) {
2517 cmsSpace()->prepare_for_verify();
2518 } else {
2519 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2520 cmsSpace()->prepare_for_verify();
2521 }
2522 }
2523
2524 void
2525 ConcurrentMarkSweepGeneration::verify() {
2526 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2527 // are not called when the heap is verified during universe initialization and
2528 // at vm shutdown.
2529 if (freelistLock()->owned_by_self()) {
2530 cmsSpace()->verify();
2531 } else {
2532 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2533 cmsSpace()->verify();
2534 }
2535 }
2536
2537 void CMSCollector::verify() {
2538 _cmsGen->verify();
2539 }
2540
2541 #ifndef PRODUCT
2542 bool CMSCollector::overflow_list_is_empty() const {
2543 assert(_num_par_pushes >= 0, "Inconsistency");
2544 if (_overflow_list == NULL) {
2545 assert(_num_par_pushes == 0, "Inconsistency");
2546 }
2547 return _overflow_list == NULL;
2548 }
2549
2550 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2551 // merely consolidate assertion checks that appear to occur together frequently.
2552 void CMSCollector::verify_work_stacks_empty() const {
2553 assert(_markStack.isEmpty(), "Marking stack should be empty");
2554 assert(overflow_list_is_empty(), "Overflow list should be empty");
2555 }
2556
2557 void CMSCollector::verify_overflow_empty() const {
2558 assert(overflow_list_is_empty(), "Overflow list should be empty");
2559 assert(no_preserved_marks(), "No preserved marks");
2560 }
2561 #endif // PRODUCT
2562
2563 // Decide if we want to enable class unloading as part of the
2564 // ensuing concurrent GC cycle. We will collect and
2565 // unload classes if it's the case that:
2566 // (1) an explicit gc request has been made and the flag
2567 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
2568 // (2) (a) class unloading is enabled at the command line, and
2569 // (b) old gen is getting really full
2570 // NOTE: Provided there is no change in the state of the heap between
2571 // calls to this method, it should have idempotent results. Moreover,
2572 // its results should be monotonically increasing (i.e. going from 0 to 1,
2573 // but not 1 to 0) between successive calls between which the heap was
2574 // not collected. For the implementation below, it must thus rely on
2575 // the property that concurrent_cycles_since_last_unload()
2576 // will not decrease unless a collection cycle happened and that
2577 // _cmsGen->is_too_full() are
2578 // themselves also monotonic in that sense. See check_monotonicity()
2579 // below.
2580 void CMSCollector::update_should_unload_classes() {
2581 _should_unload_classes = false;
2582 // Condition 1 above
2583 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
2584 _should_unload_classes = true;
2585 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
2586 // Disjuncts 2.b.(i,ii,iii) above
2587 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2588 CMSClassUnloadingMaxInterval)
2589 || _cmsGen->is_too_full();
2590 }
2591 }
2592
2593 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2594 bool res = should_concurrent_collect();
2595 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2596 return res;
2597 }
2598
2599 void CMSCollector::setup_cms_unloading_and_verification_state() {
2600 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2601 || VerifyBeforeExit;
2602 const int rso = GenCollectedHeap::SO_AllCodeCache;
2603
2604 // We set the proper root for this CMS cycle here.
2605 if (should_unload_classes()) { // Should unload classes this cycle
2606 remove_root_scanning_option(rso); // Shrink the root set appropriately
2607 set_verifying(should_verify); // Set verification state for this cycle
2608 return; // Nothing else needs to be done at this time
2609 }
2610
2611 // Not unloading classes this cycle
2612 assert(!should_unload_classes(), "Inconsistency!");
2613
2614 // If we are not unloading classes then add SO_AllCodeCache to root
2615 // scanning options.
2616 add_root_scanning_option(rso);
2617
2618 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2619 set_verifying(true);
2620 } else if (verifying() && !should_verify) {
2621 // We were verifying, but some verification flags got disabled.
2622 set_verifying(false);
2623 // Exclude symbols, strings and code cache elements from root scanning to
2624 // reduce IM and RM pauses.
2625 remove_root_scanning_option(rso);
2626 }
2627 }
2628
2629
2630 #ifndef PRODUCT
2631 HeapWord* CMSCollector::block_start(const void* p) const {
2632 const HeapWord* addr = (HeapWord*)p;
2633 if (_span.contains(p)) {
2634 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2635 return _cmsGen->cmsSpace()->block_start(p);
2636 }
2637 }
2638 return NULL;
2639 }
2640 #endif
2641
2642 HeapWord*
2643 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2644 bool tlab,
2645 bool parallel) {
2646 CMSSynchronousYieldRequest yr;
2647 assert(!tlab, "Can't deal with TLAB allocation");
2648 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2649 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2650 if (GCExpandToAllocateDelayMillis > 0) {
2651 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2652 }
2653 return have_lock_and_allocate(word_size, tlab);
2654 }
2655
2656 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2657 size_t bytes,
2658 size_t expand_bytes,
2659 CMSExpansionCause::Cause cause)
2660 {
2661
2662 bool success = expand(bytes, expand_bytes);
2663
2664 // remember why we expanded; this information is used
2665 // by shouldConcurrentCollect() when making decisions on whether to start
2666 // a new CMS cycle.
2667 if (success) {
2668 set_expansion_cause(cause);
2669 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause));
2670 }
2671 }
2672
2673 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2674 HeapWord* res = NULL;
2675 MutexLocker x(ParGCRareEvent_lock);
2676 while (true) {
2677 // Expansion by some other thread might make alloc OK now:
2678 res = ps->lab.alloc(word_sz);
2679 if (res != NULL) return res;
2680 // If there's not enough expansion space available, give up.
2681 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2682 return NULL;
2683 }
2684 // Otherwise, we try expansion.
2685 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2686 // Now go around the loop and try alloc again;
2687 // A competing par_promote might beat us to the expansion space,
2688 // so we may go around the loop again if promotion fails again.
2689 if (GCExpandToAllocateDelayMillis > 0) {
2690 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2691 }
2692 }
2693 }
2694
2695
2696 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2697 PromotionInfo* promo) {
2698 MutexLocker x(ParGCRareEvent_lock);
2699 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2700 while (true) {
2701 // Expansion by some other thread might make alloc OK now:
2702 if (promo->ensure_spooling_space()) {
2703 assert(promo->has_spooling_space(),
2704 "Post-condition of successful ensure_spooling_space()");
2705 return true;
2706 }
2707 // If there's not enough expansion space available, give up.
2708 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2709 return false;
2710 }
2711 // Otherwise, we try expansion.
2712 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2713 // Now go around the loop and try alloc again;
2714 // A competing allocation might beat us to the expansion space,
2715 // so we may go around the loop again if allocation fails again.
2716 if (GCExpandToAllocateDelayMillis > 0) {
2717 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2718 }
2719 }
2720 }
2721
2722 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2723 // Only shrink if a compaction was done so that all the free space
2724 // in the generation is in a contiguous block at the end.
2725 if (did_compact()) {
2726 CardGeneration::shrink(bytes);
2727 }
2728 }
2729
2730 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2731 assert_locked_or_safepoint(Heap_lock);
2732 }
2733
2734 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2735 assert_locked_or_safepoint(Heap_lock);
2736 assert_lock_strong(freelistLock());
2737 log_trace(gc)("Shrinking of CMS not yet implemented");
2738 return;
2739 }
2740
2741
2742 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2743 // phases.
2744 class CMSPhaseAccounting: public StackObj {
2745 public:
2746 CMSPhaseAccounting(CMSCollector *collector,
2747 const char *title);
2748 ~CMSPhaseAccounting();
2749
2750 private:
2751 CMSCollector *_collector;
2752 const char *_title;
2753 GCTraceConcTime(Info, gc) _trace_time;
2754
2755 public:
2756 // Not MT-safe; so do not pass around these StackObj's
2757 // where they may be accessed by other threads.
2758 double wallclock_millis() {
2759 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time());
2760 }
2761 };
2762
2763 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2764 const char *title) :
2765 _collector(collector), _title(title), _trace_time(title) {
2766
2767 _collector->resetYields();
2768 _collector->resetTimer();
2769 _collector->startTimer();
2770 _collector->gc_timer_cm()->register_gc_concurrent_start(title);
2771 }
2772
2773 CMSPhaseAccounting::~CMSPhaseAccounting() {
2774 _collector->gc_timer_cm()->register_gc_concurrent_end();
2775 _collector->stopTimer();
2776 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks()));
2777 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields());
2778 }
2779
2780 // CMS work
2781
2782 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2783 class CMSParMarkTask : public AbstractGangTask {
2784 protected:
2785 CMSCollector* _collector;
2786 uint _n_workers;
2787 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2788 AbstractGangTask(name),
2789 _collector(collector),
2790 _n_workers(n_workers) {}
2791 // Work method in support of parallel rescan ... of young gen spaces
2792 void do_young_space_rescan(OopsInGenClosure* cl,
2793 ContiguousSpace* space,
2794 HeapWord** chunk_array, size_t chunk_top);
2795 void work_on_young_gen_roots(OopsInGenClosure* cl);
2796 };
2797
2798 // Parallel initial mark task
2799 class CMSParInitialMarkTask: public CMSParMarkTask {
2800 StrongRootsScope* _strong_roots_scope;
2801 public:
2802 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) :
2803 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers),
2804 _strong_roots_scope(strong_roots_scope) {}
2805 void work(uint worker_id);
2806 };
2807
2808 // Checkpoint the roots into this generation from outside
2809 // this generation. [Note this initial checkpoint need only
2810 // be approximate -- we'll do a catch up phase subsequently.]
2811 void CMSCollector::checkpointRootsInitial() {
2812 assert(_collectorState == InitialMarking, "Wrong collector state");
2813 check_correct_thread_executing();
2814 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
2815
2816 save_heap_summary();
2817 report_heap_summary(GCWhen::BeforeGC);
2818
2819 ReferenceProcessor* rp = ref_processor();
2820 assert(_restart_addr == NULL, "Control point invariant");
2821 {
2822 // acquire locks for subsequent manipulations
2823 MutexLockerEx x(bitMapLock(),
2824 Mutex::_no_safepoint_check_flag);
2825 checkpointRootsInitialWork();
2826 // enable ("weak") refs discovery
2827 rp->enable_discovery();
2828 _collectorState = Marking;
2829 }
2830 }
2831
2832 void CMSCollector::checkpointRootsInitialWork() {
2833 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2834 assert(_collectorState == InitialMarking, "just checking");
2835
2836 // Already have locks.
2837 assert_lock_strong(bitMapLock());
2838 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2839
2840 // Setup the verification and class unloading state for this
2841 // CMS collection cycle.
2842 setup_cms_unloading_and_verification_state();
2843
2844 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm);
2845
2846 // Reset all the PLAB chunk arrays if necessary.
2847 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2848 reset_survivor_plab_arrays();
2849 }
2850
2851 ResourceMark rm;
2852 HandleMark hm;
2853
2854 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2855 GenCollectedHeap* gch = GenCollectedHeap::heap();
2856
2857 verify_work_stacks_empty();
2858 verify_overflow_empty();
2859
2860 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2861 // Update the saved marks which may affect the root scans.
2862 gch->save_marks();
2863
2864 // weak reference processing has not started yet.
2865 ref_processor()->set_enqueuing_is_done(false);
2866
2867 // Need to remember all newly created CLDs,
2868 // so that we can guarantee that the remark finds them.
2869 ClassLoaderDataGraph::remember_new_clds(true);
2870
2871 // Whenever a CLD is found, it will be claimed before proceeding to mark
2872 // the klasses. The claimed marks need to be cleared before marking starts.
2873 ClassLoaderDataGraph::clear_claimed_marks();
2874
2875 print_eden_and_survivor_chunk_arrays();
2876
2877 {
2878 #if defined(COMPILER2) || INCLUDE_JVMCI
2879 DerivedPointerTableDeactivate dpt_deact;
2880 #endif
2881 if (CMSParallelInitialMarkEnabled) {
2882 // The parallel version.
2883 WorkGang* workers = gch->workers();
2884 assert(workers != NULL, "Need parallel worker threads.");
2885 uint n_workers = workers->active_workers();
2886
2887 StrongRootsScope srs(n_workers);
2888
2889 CMSParInitialMarkTask tsk(this, &srs, n_workers);
2890 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
2891 // If the total workers is greater than 1, then multiple workers
2892 // may be used at some time and the initialization has been set
2893 // such that the single threaded path cannot be used.
2894 if (workers->total_workers() > 1) {
2895 workers->run_task(&tsk);
2896 } else {
2897 tsk.work(0);
2898 }
2899 } else {
2900 // The serial version.
2901 CLDToOopClosure cld_closure(¬Older, true);
2902 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2903
2904 StrongRootsScope srs(1);
2905
2906 gch->gen_process_roots(&srs,
2907 GenCollectedHeap::OldGen,
2908 true, // young gen as roots
2909 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2910 should_unload_classes(),
2911 ¬Older,
2912 NULL,
2913 &cld_closure);
2914 }
2915 }
2916
2917 // Clear mod-union table; it will be dirtied in the prologue of
2918 // CMS generation per each young generation collection.
2919
2920 assert(_modUnionTable.isAllClear(),
2921 "Was cleared in most recent final checkpoint phase"
2922 " or no bits are set in the gc_prologue before the start of the next "
2923 "subsequent marking phase.");
2924
2925 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
2926
2927 // Save the end of the used_region of the constituent generations
2928 // to be used to limit the extent of sweep in each generation.
2929 save_sweep_limits();
2930 verify_overflow_empty();
2931 }
2932
2933 bool CMSCollector::markFromRoots() {
2934 // we might be tempted to assert that:
2935 // assert(!SafepointSynchronize::is_at_safepoint(),
2936 // "inconsistent argument?");
2937 // However that wouldn't be right, because it's possible that
2938 // a safepoint is indeed in progress as a young generation
2939 // stop-the-world GC happens even as we mark in this generation.
2940 assert(_collectorState == Marking, "inconsistent state?");
2941 check_correct_thread_executing();
2942 verify_overflow_empty();
2943
2944 // Weak ref discovery note: We may be discovering weak
2945 // refs in this generation concurrent (but interleaved) with
2946 // weak ref discovery by the young generation collector.
2947
2948 CMSTokenSyncWithLocks ts(true, bitMapLock());
2949 GCTraceCPUTime tcpu;
2950 CMSPhaseAccounting pa(this, "Concurrent Mark");
2951 bool res = markFromRootsWork();
2952 if (res) {
2953 _collectorState = Precleaning;
2954 } else { // We failed and a foreground collection wants to take over
2955 assert(_foregroundGCIsActive, "internal state inconsistency");
2956 assert(_restart_addr == NULL, "foreground will restart from scratch");
2957 log_debug(gc)("bailing out to foreground collection");
2958 }
2959 verify_overflow_empty();
2960 return res;
2961 }
2962
2963 bool CMSCollector::markFromRootsWork() {
2964 // iterate over marked bits in bit map, doing a full scan and mark
2965 // from these roots using the following algorithm:
2966 // . if oop is to the right of the current scan pointer,
2967 // mark corresponding bit (we'll process it later)
2968 // . else (oop is to left of current scan pointer)
2969 // push oop on marking stack
2970 // . drain the marking stack
2971
2972 // Note that when we do a marking step we need to hold the
2973 // bit map lock -- recall that direct allocation (by mutators)
2974 // and promotion (by the young generation collector) is also
2975 // marking the bit map. [the so-called allocate live policy.]
2976 // Because the implementation of bit map marking is not
2977 // robust wrt simultaneous marking of bits in the same word,
2978 // we need to make sure that there is no such interference
2979 // between concurrent such updates.
2980
2981 // already have locks
2982 assert_lock_strong(bitMapLock());
2983
2984 verify_work_stacks_empty();
2985 verify_overflow_empty();
2986 bool result = false;
2987 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
2988 result = do_marking_mt();
2989 } else {
2990 result = do_marking_st();
2991 }
2992 return result;
2993 }
2994
2995 // Forward decl
2996 class CMSConcMarkingTask;
2997
2998 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
2999 CMSCollector* _collector;
3000 CMSConcMarkingTask* _task;
3001 public:
3002 virtual void yield();
3003
3004 // "n_threads" is the number of threads to be terminated.
3005 // "queue_set" is a set of work queues of other threads.
3006 // "collector" is the CMS collector associated with this task terminator.
3007 // "yield" indicates whether we need the gang as a whole to yield.
3008 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3009 ParallelTaskTerminator(n_threads, queue_set),
3010 _collector(collector) { }
3011
3012 void set_task(CMSConcMarkingTask* task) {
3013 _task = task;
3014 }
3015 };
3016
3017 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3018 CMSConcMarkingTask* _task;
3019 public:
3020 bool should_exit_termination();
3021 void set_task(CMSConcMarkingTask* task) {
3022 _task = task;
3023 }
3024 };
3025
3026 // MT Concurrent Marking Task
3027 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3028 CMSCollector* _collector;
3029 uint _n_workers; // requested/desired # workers
3030 bool _result;
3031 CompactibleFreeListSpace* _cms_space;
3032 char _pad_front[64]; // padding to ...
3033 HeapWord* _global_finger; // ... avoid sharing cache line
3034 char _pad_back[64];
3035 HeapWord* _restart_addr;
3036
3037 // Exposed here for yielding support
3038 Mutex* const _bit_map_lock;
3039
3040 // The per thread work queues, available here for stealing
3041 OopTaskQueueSet* _task_queues;
3042
3043 // Termination (and yielding) support
3044 CMSConcMarkingTerminator _term;
3045 CMSConcMarkingTerminatorTerminator _term_term;
3046
3047 public:
3048 CMSConcMarkingTask(CMSCollector* collector,
3049 CompactibleFreeListSpace* cms_space,
3050 YieldingFlexibleWorkGang* workers,
3051 OopTaskQueueSet* task_queues):
3052 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3053 _collector(collector),
3054 _cms_space(cms_space),
3055 _n_workers(0), _result(true),
3056 _task_queues(task_queues),
3057 _term(_n_workers, task_queues, _collector),
3058 _bit_map_lock(collector->bitMapLock())
3059 {
3060 _requested_size = _n_workers;
3061 _term.set_task(this);
3062 _term_term.set_task(this);
3063 _restart_addr = _global_finger = _cms_space->bottom();
3064 }
3065
3066
3067 OopTaskQueueSet* task_queues() { return _task_queues; }
3068
3069 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3070
3071 HeapWord** global_finger_addr() { return &_global_finger; }
3072
3073 CMSConcMarkingTerminator* terminator() { return &_term; }
3074
3075 virtual void set_for_termination(uint active_workers) {
3076 terminator()->reset_for_reuse(active_workers);
3077 }
3078
3079 void work(uint worker_id);
3080 bool should_yield() {
3081 return ConcurrentMarkSweepThread::should_yield()
3082 && !_collector->foregroundGCIsActive();
3083 }
3084
3085 virtual void coordinator_yield(); // stuff done by coordinator
3086 bool result() { return _result; }
3087
3088 void reset(HeapWord* ra) {
3089 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3090 _restart_addr = _global_finger = ra;
3091 _term.reset_for_reuse();
3092 }
3093
3094 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3095 OopTaskQueue* work_q);
3096
3097 private:
3098 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3099 void do_work_steal(int i);
3100 void bump_global_finger(HeapWord* f);
3101 };
3102
3103 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3104 assert(_task != NULL, "Error");
3105 return _task->yielding();
3106 // Note that we do not need the disjunct || _task->should_yield() above
3107 // because we want terminating threads to yield only if the task
3108 // is already in the midst of yielding, which happens only after at least one
3109 // thread has yielded.
3110 }
3111
3112 void CMSConcMarkingTerminator::yield() {
3113 if (_task->should_yield()) {
3114 _task->yield();
3115 } else {
3116 ParallelTaskTerminator::yield();
3117 }
3118 }
3119
3120 ////////////////////////////////////////////////////////////////
3121 // Concurrent Marking Algorithm Sketch
3122 ////////////////////////////////////////////////////////////////
3123 // Until all tasks exhausted (both spaces):
3124 // -- claim next available chunk
3125 // -- bump global finger via CAS
3126 // -- find first object that starts in this chunk
3127 // and start scanning bitmap from that position
3128 // -- scan marked objects for oops
3129 // -- CAS-mark target, and if successful:
3130 // . if target oop is above global finger (volatile read)
3131 // nothing to do
3132 // . if target oop is in chunk and above local finger
3133 // then nothing to do
3134 // . else push on work-queue
3135 // -- Deal with possible overflow issues:
3136 // . local work-queue overflow causes stuff to be pushed on
3137 // global (common) overflow queue
3138 // . always first empty local work queue
3139 // . then get a batch of oops from global work queue if any
3140 // . then do work stealing
3141 // -- When all tasks claimed (both spaces)
3142 // and local work queue empty,
3143 // then in a loop do:
3144 // . check global overflow stack; steal a batch of oops and trace
3145 // . try to steal from other threads oif GOS is empty
3146 // . if neither is available, offer termination
3147 // -- Terminate and return result
3148 //
3149 void CMSConcMarkingTask::work(uint worker_id) {
3150 elapsedTimer _timer;
3151 ResourceMark rm;
3152 HandleMark hm;
3153
3154 DEBUG_ONLY(_collector->verify_overflow_empty();)
3155
3156 // Before we begin work, our work queue should be empty
3157 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3158 // Scan the bitmap covering _cms_space, tracing through grey objects.
3159 _timer.start();
3160 do_scan_and_mark(worker_id, _cms_space);
3161 _timer.stop();
3162 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3163
3164 // ... do work stealing
3165 _timer.reset();
3166 _timer.start();
3167 do_work_steal(worker_id);
3168 _timer.stop();
3169 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3170 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3171 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3172 // Note that under the current task protocol, the
3173 // following assertion is true even of the spaces
3174 // expanded since the completion of the concurrent
3175 // marking. XXX This will likely change under a strict
3176 // ABORT semantics.
3177 // After perm removal the comparison was changed to
3178 // greater than or equal to from strictly greater than.
3179 // Before perm removal the highest address sweep would
3180 // have been at the end of perm gen but now is at the
3181 // end of the tenured gen.
3182 assert(_global_finger >= _cms_space->end(),
3183 "All tasks have been completed");
3184 DEBUG_ONLY(_collector->verify_overflow_empty();)
3185 }
3186
3187 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3188 HeapWord* read = _global_finger;
3189 HeapWord* cur = read;
3190 while (f > read) {
3191 cur = read;
3192 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3193 if (cur == read) {
3194 // our cas succeeded
3195 assert(_global_finger >= f, "protocol consistency");
3196 break;
3197 }
3198 }
3199 }
3200
3201 // This is really inefficient, and should be redone by
3202 // using (not yet available) block-read and -write interfaces to the
3203 // stack and the work_queue. XXX FIX ME !!!
3204 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3205 OopTaskQueue* work_q) {
3206 // Fast lock-free check
3207 if (ovflw_stk->length() == 0) {
3208 return false;
3209 }
3210 assert(work_q->size() == 0, "Shouldn't steal");
3211 MutexLockerEx ml(ovflw_stk->par_lock(),
3212 Mutex::_no_safepoint_check_flag);
3213 // Grab up to 1/4 the size of the work queue
3214 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3215 (size_t)ParGCDesiredObjsFromOverflowList);
3216 num = MIN2(num, ovflw_stk->length());
3217 for (int i = (int) num; i > 0; i--) {
3218 oop cur = ovflw_stk->pop();
3219 assert(cur != NULL, "Counted wrong?");
3220 work_q->push(cur);
3221 }
3222 return num > 0;
3223 }
3224
3225 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3226 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3227 int n_tasks = pst->n_tasks();
3228 // We allow that there may be no tasks to do here because
3229 // we are restarting after a stack overflow.
3230 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3231 uint nth_task = 0;
3232
3233 HeapWord* aligned_start = sp->bottom();
3234 if (sp->used_region().contains(_restart_addr)) {
3235 // Align down to a card boundary for the start of 0th task
3236 // for this space.
3237 aligned_start =
3238 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3239 CardTableModRefBS::card_size);
3240 }
3241
3242 size_t chunk_size = sp->marking_task_size();
3243 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3244 // Having claimed the nth task in this space,
3245 // compute the chunk that it corresponds to:
3246 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3247 aligned_start + (nth_task+1)*chunk_size);
3248 // Try and bump the global finger via a CAS;
3249 // note that we need to do the global finger bump
3250 // _before_ taking the intersection below, because
3251 // the task corresponding to that region will be
3252 // deemed done even if the used_region() expands
3253 // because of allocation -- as it almost certainly will
3254 // during start-up while the threads yield in the
3255 // closure below.
3256 HeapWord* finger = span.end();
3257 bump_global_finger(finger); // atomically
3258 // There are null tasks here corresponding to chunks
3259 // beyond the "top" address of the space.
3260 span = span.intersection(sp->used_region());
3261 if (!span.is_empty()) { // Non-null task
3262 HeapWord* prev_obj;
3263 assert(!span.contains(_restart_addr) || nth_task == 0,
3264 "Inconsistency");
3265 if (nth_task == 0) {
3266 // For the 0th task, we'll not need to compute a block_start.
3267 if (span.contains(_restart_addr)) {
3268 // In the case of a restart because of stack overflow,
3269 // we might additionally skip a chunk prefix.
3270 prev_obj = _restart_addr;
3271 } else {
3272 prev_obj = span.start();
3273 }
3274 } else {
3275 // We want to skip the first object because
3276 // the protocol is to scan any object in its entirety
3277 // that _starts_ in this span; a fortiori, any
3278 // object starting in an earlier span is scanned
3279 // as part of an earlier claimed task.
3280 // Below we use the "careful" version of block_start
3281 // so we do not try to navigate uninitialized objects.
3282 prev_obj = sp->block_start_careful(span.start());
3283 // Below we use a variant of block_size that uses the
3284 // Printezis bits to avoid waiting for allocated
3285 // objects to become initialized/parsable.
3286 while (prev_obj < span.start()) {
3287 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3288 if (sz > 0) {
3289 prev_obj += sz;
3290 } else {
3291 // In this case we may end up doing a bit of redundant
3292 // scanning, but that appears unavoidable, short of
3293 // locking the free list locks; see bug 6324141.
3294 break;
3295 }
3296 }
3297 }
3298 if (prev_obj < span.end()) {
3299 MemRegion my_span = MemRegion(prev_obj, span.end());
3300 // Do the marking work within a non-empty span --
3301 // the last argument to the constructor indicates whether the
3302 // iteration should be incremental with periodic yields.
3303 ParMarkFromRootsClosure cl(this, _collector, my_span,
3304 &_collector->_markBitMap,
3305 work_queue(i),
3306 &_collector->_markStack);
3307 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3308 } // else nothing to do for this task
3309 } // else nothing to do for this task
3310 }
3311 // We'd be tempted to assert here that since there are no
3312 // more tasks left to claim in this space, the global_finger
3313 // must exceed space->top() and a fortiori space->end(). However,
3314 // that would not quite be correct because the bumping of
3315 // global_finger occurs strictly after the claiming of a task,
3316 // so by the time we reach here the global finger may not yet
3317 // have been bumped up by the thread that claimed the last
3318 // task.
3319 pst->all_tasks_completed();
3320 }
3321
3322 class ParConcMarkingClosure: public MetadataAwareOopClosure {
3323 private:
3324 CMSCollector* _collector;
3325 CMSConcMarkingTask* _task;
3326 MemRegion _span;
3327 CMSBitMap* _bit_map;
3328 CMSMarkStack* _overflow_stack;
3329 OopTaskQueue* _work_queue;
3330 protected:
3331 DO_OOP_WORK_DEFN
3332 public:
3333 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3334 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3335 MetadataAwareOopClosure(collector->ref_processor()),
3336 _collector(collector),
3337 _task(task),
3338 _span(collector->_span),
3339 _work_queue(work_queue),
3340 _bit_map(bit_map),
3341 _overflow_stack(overflow_stack)
3342 { }
3343 virtual void do_oop(oop* p);
3344 virtual void do_oop(narrowOop* p);
3345
3346 void trim_queue(size_t max);
3347 void handle_stack_overflow(HeapWord* lost);
3348 void do_yield_check() {
3349 if (_task->should_yield()) {
3350 _task->yield();
3351 }
3352 }
3353 };
3354
3355 DO_OOP_WORK_IMPL(ParConcMarkingClosure)
3356
3357 // Grey object scanning during work stealing phase --
3358 // the salient assumption here is that any references
3359 // that are in these stolen objects being scanned must
3360 // already have been initialized (else they would not have
3361 // been published), so we do not need to check for
3362 // uninitialized objects before pushing here.
3363 void ParConcMarkingClosure::do_oop(oop obj) {
3364 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
3365 HeapWord* addr = (HeapWord*)obj;
3366 // Check if oop points into the CMS generation
3367 // and is not marked
3368 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3369 // a white object ...
3370 // If we manage to "claim" the object, by being the
3371 // first thread to mark it, then we push it on our
3372 // marking stack
3373 if (_bit_map->par_mark(addr)) { // ... now grey
3374 // push on work queue (grey set)
3375 bool simulate_overflow = false;
3376 NOT_PRODUCT(
3377 if (CMSMarkStackOverflowALot &&
3378 _collector->simulate_overflow()) {
3379 // simulate a stack overflow
3380 simulate_overflow = true;
3381 }
3382 )
3383 if (simulate_overflow ||
3384 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3385 // stack overflow
3386 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
3387 // We cannot assert that the overflow stack is full because
3388 // it may have been emptied since.
3389 assert(simulate_overflow ||
3390 _work_queue->size() == _work_queue->max_elems(),
3391 "Else push should have succeeded");
3392 handle_stack_overflow(addr);
3393 }
3394 } // Else, some other thread got there first
3395 do_yield_check();
3396 }
3397 }
3398
3399 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); }
3400 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); }
3401
3402 void ParConcMarkingClosure::trim_queue(size_t max) {
3403 while (_work_queue->size() > max) {
3404 oop new_oop;
3405 if (_work_queue->pop_local(new_oop)) {
3406 assert(new_oop->is_oop(), "Should be an oop");
3407 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3408 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3409 new_oop->oop_iterate(this); // do_oop() above
3410 do_yield_check();
3411 }
3412 }
3413 }
3414
3415 // Upon stack overflow, we discard (part of) the stack,
3416 // remembering the least address amongst those discarded
3417 // in CMSCollector's _restart_address.
3418 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3419 // We need to do this under a mutex to prevent other
3420 // workers from interfering with the work done below.
3421 MutexLockerEx ml(_overflow_stack->par_lock(),
3422 Mutex::_no_safepoint_check_flag);
3423 // Remember the least grey address discarded
3424 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3425 _collector->lower_restart_addr(ra);
3426 _overflow_stack->reset(); // discard stack contents
3427 _overflow_stack->expand(); // expand the stack if possible
3428 }
3429
3430
3431 void CMSConcMarkingTask::do_work_steal(int i) {
3432 OopTaskQueue* work_q = work_queue(i);
3433 oop obj_to_scan;
3434 CMSBitMap* bm = &(_collector->_markBitMap);
3435 CMSMarkStack* ovflw = &(_collector->_markStack);
3436 int* seed = _collector->hash_seed(i);
3437 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3438 while (true) {
3439 cl.trim_queue(0);
3440 assert(work_q->size() == 0, "Should have been emptied above");
3441 if (get_work_from_overflow_stack(ovflw, work_q)) {
3442 // Can't assert below because the work obtained from the
3443 // overflow stack may already have been stolen from us.
3444 // assert(work_q->size() > 0, "Work from overflow stack");
3445 continue;
3446 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3447 assert(obj_to_scan->is_oop(), "Should be an oop");
3448 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3449 obj_to_scan->oop_iterate(&cl);
3450 } else if (terminator()->offer_termination(&_term_term)) {
3451 assert(work_q->size() == 0, "Impossible!");
3452 break;
3453 } else if (yielding() || should_yield()) {
3454 yield();
3455 }
3456 }
3457 }
3458
3459 // This is run by the CMS (coordinator) thread.
3460 void CMSConcMarkingTask::coordinator_yield() {
3461 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3462 "CMS thread should hold CMS token");
3463 // First give up the locks, then yield, then re-lock
3464 // We should probably use a constructor/destructor idiom to
3465 // do this unlock/lock or modify the MutexUnlocker class to
3466 // serve our purpose. XXX
3467 assert_lock_strong(_bit_map_lock);
3468 _bit_map_lock->unlock();
3469 ConcurrentMarkSweepThread::desynchronize(true);
3470 _collector->stopTimer();
3471 _collector->incrementYields();
3472
3473 // It is possible for whichever thread initiated the yield request
3474 // not to get a chance to wake up and take the bitmap lock between
3475 // this thread releasing it and reacquiring it. So, while the
3476 // should_yield() flag is on, let's sleep for a bit to give the
3477 // other thread a chance to wake up. The limit imposed on the number
3478 // of iterations is defensive, to avoid any unforseen circumstances
3479 // putting us into an infinite loop. Since it's always been this
3480 // (coordinator_yield()) method that was observed to cause the
3481 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3482 // which is by default non-zero. For the other seven methods that
3483 // also perform the yield operation, as are using a different
3484 // parameter (CMSYieldSleepCount) which is by default zero. This way we
3485 // can enable the sleeping for those methods too, if necessary.
3486 // See 6442774.
3487 //
3488 // We really need to reconsider the synchronization between the GC
3489 // thread and the yield-requesting threads in the future and we
3490 // should really use wait/notify, which is the recommended
3491 // way of doing this type of interaction. Additionally, we should
3492 // consolidate the eight methods that do the yield operation and they
3493 // are almost identical into one for better maintainability and
3494 // readability. See 6445193.
3495 //
3496 // Tony 2006.06.29
3497 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3498 ConcurrentMarkSweepThread::should_yield() &&
3499 !CMSCollector::foregroundGCIsActive(); ++i) {
3500 os::sleep(Thread::current(), 1, false);
3501 }
3502
3503 ConcurrentMarkSweepThread::synchronize(true);
3504 _bit_map_lock->lock_without_safepoint_check();
3505 _collector->startTimer();
3506 }
3507
3508 bool CMSCollector::do_marking_mt() {
3509 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3510 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3511 conc_workers()->active_workers(),
3512 Threads::number_of_non_daemon_threads());
3513 num_workers = conc_workers()->update_active_workers(num_workers);
3514
3515 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
3516
3517 CMSConcMarkingTask tsk(this,
3518 cms_space,
3519 conc_workers(),
3520 task_queues());
3521
3522 // Since the actual number of workers we get may be different
3523 // from the number we requested above, do we need to do anything different
3524 // below? In particular, may be we need to subclass the SequantialSubTasksDone
3525 // class?? XXX
3526 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3527
3528 // Refs discovery is already non-atomic.
3529 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3530 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3531 conc_workers()->start_task(&tsk);
3532 while (tsk.yielded()) {
3533 tsk.coordinator_yield();
3534 conc_workers()->continue_task(&tsk);
3535 }
3536 // If the task was aborted, _restart_addr will be non-NULL
3537 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3538 while (_restart_addr != NULL) {
3539 // XXX For now we do not make use of ABORTED state and have not
3540 // yet implemented the right abort semantics (even in the original
3541 // single-threaded CMS case). That needs some more investigation
3542 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3543 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3544 // If _restart_addr is non-NULL, a marking stack overflow
3545 // occurred; we need to do a fresh marking iteration from the
3546 // indicated restart address.
3547 if (_foregroundGCIsActive) {
3548 // We may be running into repeated stack overflows, having
3549 // reached the limit of the stack size, while making very
3550 // slow forward progress. It may be best to bail out and
3551 // let the foreground collector do its job.
3552 // Clear _restart_addr, so that foreground GC
3553 // works from scratch. This avoids the headache of
3554 // a "rescan" which would otherwise be needed because
3555 // of the dirty mod union table & card table.
3556 _restart_addr = NULL;
3557 return false;
3558 }
3559 // Adjust the task to restart from _restart_addr
3560 tsk.reset(_restart_addr);
3561 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3562 _restart_addr);
3563 _restart_addr = NULL;
3564 // Get the workers going again
3565 conc_workers()->start_task(&tsk);
3566 while (tsk.yielded()) {
3567 tsk.coordinator_yield();
3568 conc_workers()->continue_task(&tsk);
3569 }
3570 }
3571 assert(tsk.completed(), "Inconsistency");
3572 assert(tsk.result() == true, "Inconsistency");
3573 return true;
3574 }
3575
3576 bool CMSCollector::do_marking_st() {
3577 ResourceMark rm;
3578 HandleMark hm;
3579
3580 // Temporarily make refs discovery single threaded (non-MT)
3581 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3582 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3583 &_markStack, CMSYield);
3584 // the last argument to iterate indicates whether the iteration
3585 // should be incremental with periodic yields.
3586 _markBitMap.iterate(&markFromRootsClosure);
3587 // If _restart_addr is non-NULL, a marking stack overflow
3588 // occurred; we need to do a fresh iteration from the
3589 // indicated restart address.
3590 while (_restart_addr != NULL) {
3591 if (_foregroundGCIsActive) {
3592 // We may be running into repeated stack overflows, having
3593 // reached the limit of the stack size, while making very
3594 // slow forward progress. It may be best to bail out and
3595 // let the foreground collector do its job.
3596 // Clear _restart_addr, so that foreground GC
3597 // works from scratch. This avoids the headache of
3598 // a "rescan" which would otherwise be needed because
3599 // of the dirty mod union table & card table.
3600 _restart_addr = NULL;
3601 return false; // indicating failure to complete marking
3602 }
3603 // Deal with stack overflow:
3604 // we restart marking from _restart_addr
3605 HeapWord* ra = _restart_addr;
3606 markFromRootsClosure.reset(ra);
3607 _restart_addr = NULL;
3608 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3609 }
3610 return true;
3611 }
3612
3613 void CMSCollector::preclean() {
3614 check_correct_thread_executing();
3615 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3616 verify_work_stacks_empty();
3617 verify_overflow_empty();
3618 _abort_preclean = false;
3619 if (CMSPrecleaningEnabled) {
3620 if (!CMSEdenChunksRecordAlways) {
3621 _eden_chunk_index = 0;
3622 }
3623 size_t used = get_eden_used();
3624 size_t capacity = get_eden_capacity();
3625 // Don't start sampling unless we will get sufficiently
3626 // many samples.
3627 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100)
3628 * CMSScheduleRemarkEdenPenetration)) {
3629 _start_sampling = true;
3630 } else {
3631 _start_sampling = false;
3632 }
3633 GCTraceCPUTime tcpu;
3634 CMSPhaseAccounting pa(this, "Concurrent Preclean");
3635 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3636 }
3637 CMSTokenSync x(true); // is cms thread
3638 if (CMSPrecleaningEnabled) {
3639 sample_eden();
3640 _collectorState = AbortablePreclean;
3641 } else {
3642 _collectorState = FinalMarking;
3643 }
3644 verify_work_stacks_empty();
3645 verify_overflow_empty();
3646 }
3647
3648 // Try and schedule the remark such that young gen
3649 // occupancy is CMSScheduleRemarkEdenPenetration %.
3650 void CMSCollector::abortable_preclean() {
3651 check_correct_thread_executing();
3652 assert(CMSPrecleaningEnabled, "Inconsistent control state");
3653 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3654
3655 // If Eden's current occupancy is below this threshold,
3656 // immediately schedule the remark; else preclean
3657 // past the next scavenge in an effort to
3658 // schedule the pause as described above. By choosing
3659 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3660 // we will never do an actual abortable preclean cycle.
3661 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3662 GCTraceCPUTime tcpu;
3663 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean");
3664 // We need more smarts in the abortable preclean
3665 // loop below to deal with cases where allocation
3666 // in young gen is very very slow, and our precleaning
3667 // is running a losing race against a horde of
3668 // mutators intent on flooding us with CMS updates
3669 // (dirty cards).
3670 // One, admittedly dumb, strategy is to give up
3671 // after a certain number of abortable precleaning loops
3672 // or after a certain maximum time. We want to make
3673 // this smarter in the next iteration.
3674 // XXX FIX ME!!! YSR
3675 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3676 while (!(should_abort_preclean() ||
3677 ConcurrentMarkSweepThread::cmst()->should_terminate())) {
3678 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3679 cumworkdone += workdone;
3680 loops++;
3681 // Voluntarily terminate abortable preclean phase if we have
3682 // been at it for too long.
3683 if ((CMSMaxAbortablePrecleanLoops != 0) &&
3684 loops >= CMSMaxAbortablePrecleanLoops) {
3685 log_debug(gc)(" CMS: abort preclean due to loops ");
3686 break;
3687 }
3688 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3689 log_debug(gc)(" CMS: abort preclean due to time ");
3690 break;
3691 }
3692 // If we are doing little work each iteration, we should
3693 // take a short break.
3694 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3695 // Sleep for some time, waiting for work to accumulate
3696 stopTimer();
3697 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3698 startTimer();
3699 waited++;
3700 }
3701 }
3702 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3703 loops, waited, cumworkdone);
3704 }
3705 CMSTokenSync x(true); // is cms thread
3706 if (_collectorState != Idling) {
3707 assert(_collectorState == AbortablePreclean,
3708 "Spontaneous state transition?");
3709 _collectorState = FinalMarking;
3710 } // Else, a foreground collection completed this CMS cycle.
3711 return;
3712 }
3713
3714 // Respond to an Eden sampling opportunity
3715 void CMSCollector::sample_eden() {
3716 // Make sure a young gc cannot sneak in between our
3717 // reading and recording of a sample.
3718 assert(Thread::current()->is_ConcurrentGC_thread(),
3719 "Only the cms thread may collect Eden samples");
3720 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3721 "Should collect samples while holding CMS token");
3722 if (!_start_sampling) {
3723 return;
3724 }
3725 // When CMSEdenChunksRecordAlways is true, the eden chunk array
3726 // is populated by the young generation.
3727 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3728 if (_eden_chunk_index < _eden_chunk_capacity) {
3729 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
3730 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3731 "Unexpected state of Eden");
3732 // We'd like to check that what we just sampled is an oop-start address;
3733 // however, we cannot do that here since the object may not yet have been
3734 // initialized. So we'll instead do the check when we _use_ this sample
3735 // later.
3736 if (_eden_chunk_index == 0 ||
3737 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3738 _eden_chunk_array[_eden_chunk_index-1])
3739 >= CMSSamplingGrain)) {
3740 _eden_chunk_index++; // commit sample
3741 }
3742 }
3743 }
3744 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3745 size_t used = get_eden_used();
3746 size_t capacity = get_eden_capacity();
3747 assert(used <= capacity, "Unexpected state of Eden");
3748 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3749 _abort_preclean = true;
3750 }
3751 }
3752 }
3753
3754
3755 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3756 assert(_collectorState == Precleaning ||
3757 _collectorState == AbortablePreclean, "incorrect state");
3758 ResourceMark rm;
3759 HandleMark hm;
3760
3761 // Precleaning is currently not MT but the reference processor
3762 // may be set for MT. Disable it temporarily here.
3763 ReferenceProcessor* rp = ref_processor();
3764 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3765
3766 // Do one pass of scrubbing the discovered reference lists
3767 // to remove any reference objects with strongly-reachable
3768 // referents.
3769 if (clean_refs) {
3770 CMSPrecleanRefsYieldClosure yield_cl(this);
3771 assert(rp->span().equals(_span), "Spans should be equal");
3772 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3773 &_markStack, true /* preclean */);
3774 CMSDrainMarkingStackClosure complete_trace(this,
3775 _span, &_markBitMap, &_markStack,
3776 &keep_alive, true /* preclean */);
3777
3778 // We don't want this step to interfere with a young
3779 // collection because we don't want to take CPU
3780 // or memory bandwidth away from the young GC threads
3781 // (which may be as many as there are CPUs).
3782 // Note that we don't need to protect ourselves from
3783 // interference with mutators because they can't
3784 // manipulate the discovered reference lists nor affect
3785 // the computed reachability of the referents, the
3786 // only properties manipulated by the precleaning
3787 // of these reference lists.
3788 stopTimer();
3789 CMSTokenSyncWithLocks x(true /* is cms thread */,
3790 bitMapLock());
3791 startTimer();
3792 sample_eden();
3793
3794 // The following will yield to allow foreground
3795 // collection to proceed promptly. XXX YSR:
3796 // The code in this method may need further
3797 // tweaking for better performance and some restructuring
3798 // for cleaner interfaces.
3799 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3800 rp->preclean_discovered_references(
3801 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3802 gc_timer);
3803 }
3804
3805 if (clean_survivor) { // preclean the active survivor space(s)
3806 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3807 &_markBitMap, &_modUnionTable,
3808 &_markStack, true /* precleaning phase */);
3809 stopTimer();
3810 CMSTokenSyncWithLocks ts(true /* is cms thread */,
3811 bitMapLock());
3812 startTimer();
3813 unsigned int before_count =
3814 GenCollectedHeap::heap()->total_collections();
3815 SurvivorSpacePrecleanClosure
3816 sss_cl(this, _span, &_markBitMap, &_markStack,
3817 &pam_cl, before_count, CMSYield);
3818 _young_gen->from()->object_iterate_careful(&sss_cl);
3819 _young_gen->to()->object_iterate_careful(&sss_cl);
3820 }
3821 MarkRefsIntoAndScanClosure
3822 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3823 &_markStack, this, CMSYield,
3824 true /* precleaning phase */);
3825 // CAUTION: The following closure has persistent state that may need to
3826 // be reset upon a decrease in the sequence of addresses it
3827 // processes.
3828 ScanMarkedObjectsAgainCarefullyClosure
3829 smoac_cl(this, _span,
3830 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3831
3832 // Preclean dirty cards in ModUnionTable and CardTable using
3833 // appropriate convergence criterion;
3834 // repeat CMSPrecleanIter times unless we find that
3835 // we are losing.
3836 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3837 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3838 "Bad convergence multiplier");
3839 assert(CMSPrecleanThreshold >= 100,
3840 "Unreasonably low CMSPrecleanThreshold");
3841
3842 size_t numIter, cumNumCards, lastNumCards, curNumCards;
3843 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3844 numIter < CMSPrecleanIter;
3845 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3846 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
3847 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3848 // Either there are very few dirty cards, so re-mark
3849 // pause will be small anyway, or our pre-cleaning isn't
3850 // that much faster than the rate at which cards are being
3851 // dirtied, so we might as well stop and re-mark since
3852 // precleaning won't improve our re-mark time by much.
3853 if (curNumCards <= CMSPrecleanThreshold ||
3854 (numIter > 0 &&
3855 (curNumCards * CMSPrecleanDenominator >
3856 lastNumCards * CMSPrecleanNumerator))) {
3857 numIter++;
3858 cumNumCards += curNumCards;
3859 break;
3860 }
3861 }
3862
3863 preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
3864
3865 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
3866 cumNumCards += curNumCards;
3867 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
3868 curNumCards, cumNumCards, numIter);
3869 return cumNumCards; // as a measure of useful work done
3870 }
3871
3872 // PRECLEANING NOTES:
3873 // Precleaning involves:
3874 // . reading the bits of the modUnionTable and clearing the set bits.
3875 // . For the cards corresponding to the set bits, we scan the
3876 // objects on those cards. This means we need the free_list_lock
3877 // so that we can safely iterate over the CMS space when scanning
3878 // for oops.
3879 // . When we scan the objects, we'll be both reading and setting
3880 // marks in the marking bit map, so we'll need the marking bit map.
3881 // . For protecting _collector_state transitions, we take the CGC_lock.
3882 // Note that any races in the reading of of card table entries by the
3883 // CMS thread on the one hand and the clearing of those entries by the
3884 // VM thread or the setting of those entries by the mutator threads on the
3885 // other are quite benign. However, for efficiency it makes sense to keep
3886 // the VM thread from racing with the CMS thread while the latter is
3887 // dirty card info to the modUnionTable. We therefore also use the
3888 // CGC_lock to protect the reading of the card table and the mod union
3889 // table by the CM thread.
3890 // . We run concurrently with mutator updates, so scanning
3891 // needs to be done carefully -- we should not try to scan
3892 // potentially uninitialized objects.
3893 //
3894 // Locking strategy: While holding the CGC_lock, we scan over and
3895 // reset a maximal dirty range of the mod union / card tables, then lock
3896 // the free_list_lock and bitmap lock to do a full marking, then
3897 // release these locks; and repeat the cycle. This allows for a
3898 // certain amount of fairness in the sharing of these locks between
3899 // the CMS collector on the one hand, and the VM thread and the
3900 // mutators on the other.
3901
3902 // NOTE: preclean_mod_union_table() and preclean_card_table()
3903 // further below are largely identical; if you need to modify
3904 // one of these methods, please check the other method too.
3905
3906 size_t CMSCollector::preclean_mod_union_table(
3907 ConcurrentMarkSweepGeneration* old_gen,
3908 ScanMarkedObjectsAgainCarefullyClosure* cl) {
3909 verify_work_stacks_empty();
3910 verify_overflow_empty();
3911
3912 // strategy: starting with the first card, accumulate contiguous
3913 // ranges of dirty cards; clear these cards, then scan the region
3914 // covered by these cards.
3915
3916 // Since all of the MUT is committed ahead, we can just use
3917 // that, in case the generations expand while we are precleaning.
3918 // It might also be fine to just use the committed part of the
3919 // generation, but we might potentially miss cards when the
3920 // generation is rapidly expanding while we are in the midst
3921 // of precleaning.
3922 HeapWord* startAddr = old_gen->reserved().start();
3923 HeapWord* endAddr = old_gen->reserved().end();
3924
3925 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
3926
3927 size_t numDirtyCards, cumNumDirtyCards;
3928 HeapWord *nextAddr, *lastAddr;
3929 for (cumNumDirtyCards = numDirtyCards = 0,
3930 nextAddr = lastAddr = startAddr;
3931 nextAddr < endAddr;
3932 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
3933
3934 ResourceMark rm;
3935 HandleMark hm;
3936
3937 MemRegion dirtyRegion;
3938 {
3939 stopTimer();
3940 // Potential yield point
3941 CMSTokenSync ts(true);
3942 startTimer();
3943 sample_eden();
3944 // Get dirty region starting at nextOffset (inclusive),
3945 // simultaneously clearing it.
3946 dirtyRegion =
3947 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
3948 assert(dirtyRegion.start() >= nextAddr,
3949 "returned region inconsistent?");
3950 }
3951 // Remember where the next search should begin.
3952 // The returned region (if non-empty) is a right open interval,
3953 // so lastOffset is obtained from the right end of that
3954 // interval.
3955 lastAddr = dirtyRegion.end();
3956 // Should do something more transparent and less hacky XXX
3957 numDirtyCards =
3958 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
3959
3960 // We'll scan the cards in the dirty region (with periodic
3961 // yields for foreground GC as needed).
3962 if (!dirtyRegion.is_empty()) {
3963 assert(numDirtyCards > 0, "consistency check");
3964 HeapWord* stop_point = NULL;
3965 stopTimer();
3966 // Potential yield point
3967 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(),
3968 bitMapLock());
3969 startTimer();
3970 {
3971 verify_work_stacks_empty();
3972 verify_overflow_empty();
3973 sample_eden();
3974 stop_point =
3975 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
3976 }
3977 if (stop_point != NULL) {
3978 // The careful iteration stopped early either because it found an
3979 // uninitialized object, or because we were in the midst of an
3980 // "abortable preclean", which should now be aborted. Redirty
3981 // the bits corresponding to the partially-scanned or unscanned
3982 // cards. We'll either restart at the next block boundary or
3983 // abort the preclean.
3984 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
3985 "Should only be AbortablePreclean.");
3986 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
3987 if (should_abort_preclean()) {
3988 break; // out of preclean loop
3989 } else {
3990 // Compute the next address at which preclean should pick up;
3991 // might need bitMapLock in order to read P-bits.
3992 lastAddr = next_card_start_after_block(stop_point);
3993 }
3994 }
3995 } else {
3996 assert(lastAddr == endAddr, "consistency check");
3997 assert(numDirtyCards == 0, "consistency check");
3998 break;
3999 }
4000 }
4001 verify_work_stacks_empty();
4002 verify_overflow_empty();
4003 return cumNumDirtyCards;
4004 }
4005
4006 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4007 // below are largely identical; if you need to modify
4008 // one of these methods, please check the other method too.
4009
4010 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen,
4011 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4012 // strategy: it's similar to precleamModUnionTable above, in that
4013 // we accumulate contiguous ranges of dirty cards, mark these cards
4014 // precleaned, then scan the region covered by these cards.
4015 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high());
4016 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low());
4017
4018 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
4019
4020 size_t numDirtyCards, cumNumDirtyCards;
4021 HeapWord *lastAddr, *nextAddr;
4022
4023 for (cumNumDirtyCards = numDirtyCards = 0,
4024 nextAddr = lastAddr = startAddr;
4025 nextAddr < endAddr;
4026 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4027
4028 ResourceMark rm;
4029 HandleMark hm;
4030
4031 MemRegion dirtyRegion;
4032 {
4033 // See comments in "Precleaning notes" above on why we
4034 // do this locking. XXX Could the locking overheads be
4035 // too high when dirty cards are sparse? [I don't think so.]
4036 stopTimer();
4037 CMSTokenSync x(true); // is cms thread
4038 startTimer();
4039 sample_eden();
4040 // Get and clear dirty region from card table
4041 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4042 MemRegion(nextAddr, endAddr),
4043 true,
4044 CardTableModRefBS::precleaned_card_val());
4045
4046 assert(dirtyRegion.start() >= nextAddr,
4047 "returned region inconsistent?");
4048 }
4049 lastAddr = dirtyRegion.end();
4050 numDirtyCards =
4051 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4052
4053 if (!dirtyRegion.is_empty()) {
4054 stopTimer();
4055 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock());
4056 startTimer();
4057 sample_eden();
4058 verify_work_stacks_empty();
4059 verify_overflow_empty();
4060 HeapWord* stop_point =
4061 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4062 if (stop_point != NULL) {
4063 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4064 "Should only be AbortablePreclean.");
4065 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4066 if (should_abort_preclean()) {
4067 break; // out of preclean loop
4068 } else {
4069 // Compute the next address at which preclean should pick up.
4070 lastAddr = next_card_start_after_block(stop_point);
4071 }
4072 }
4073 } else {
4074 break;
4075 }
4076 }
4077 verify_work_stacks_empty();
4078 verify_overflow_empty();
4079 return cumNumDirtyCards;
4080 }
4081
4082 class PrecleanKlassClosure : public KlassClosure {
4083 KlassToOopClosure _cm_klass_closure;
4084 public:
4085 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4086 void do_klass(Klass* k) {
4087 if (k->has_accumulated_modified_oops()) {
4088 k->clear_accumulated_modified_oops();
4089
4090 _cm_klass_closure.do_klass(k);
4091 }
4092 }
4093 };
4094
4095 // The freelist lock is needed to prevent asserts, is it really needed?
4096 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4097
4098 cl->set_freelistLock(freelistLock);
4099
4100 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4101
4102 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4103 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4104 PrecleanKlassClosure preclean_klass_closure(cl);
4105 ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
4106
4107 verify_work_stacks_empty();
4108 verify_overflow_empty();
4109 }
4110
4111 void CMSCollector::checkpointRootsFinal() {
4112 assert(_collectorState == FinalMarking, "incorrect state transition?");
4113 check_correct_thread_executing();
4114 // world is stopped at this checkpoint
4115 assert(SafepointSynchronize::is_at_safepoint(),
4116 "world should be stopped");
4117 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4118
4119 verify_work_stacks_empty();
4120 verify_overflow_empty();
4121
4122 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)",
4123 _young_gen->used() / K, _young_gen->capacity() / K);
4124 {
4125 if (CMSScavengeBeforeRemark) {
4126 GenCollectedHeap* gch = GenCollectedHeap::heap();
4127 // Temporarily set flag to false, GCH->do_collection will
4128 // expect it to be false and set to true
4129 FlagSetting fl(gch->_is_gc_active, false);
4130
4131 gch->do_collection(true, // full (i.e. force, see below)
4132 false, // !clear_all_soft_refs
4133 0, // size
4134 false, // is_tlab
4135 GenCollectedHeap::YoungGen // type
4136 );
4137 }
4138 FreelistLocker x(this);
4139 MutexLockerEx y(bitMapLock(),
4140 Mutex::_no_safepoint_check_flag);
4141 checkpointRootsFinalWork();
4142 }
4143 verify_work_stacks_empty();
4144 verify_overflow_empty();
4145 }
4146
4147 void CMSCollector::checkpointRootsFinalWork() {
4148 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm);
4149
4150 assert(haveFreelistLocks(), "must have free list locks");
4151 assert_lock_strong(bitMapLock());
4152
4153 ResourceMark rm;
4154 HandleMark hm;
4155
4156 GenCollectedHeap* gch = GenCollectedHeap::heap();
4157
4158 if (should_unload_classes()) {
4159 CodeCache::gc_prologue();
4160 }
4161 assert(haveFreelistLocks(), "must have free list locks");
4162 assert_lock_strong(bitMapLock());
4163
4164 // We might assume that we need not fill TLAB's when
4165 // CMSScavengeBeforeRemark is set, because we may have just done
4166 // a scavenge which would have filled all TLAB's -- and besides
4167 // Eden would be empty. This however may not always be the case --
4168 // for instance although we asked for a scavenge, it may not have
4169 // happened because of a JNI critical section. We probably need
4170 // a policy for deciding whether we can in that case wait until
4171 // the critical section releases and then do the remark following
4172 // the scavenge, and skip it here. In the absence of that policy,
4173 // or of an indication of whether the scavenge did indeed occur,
4174 // we cannot rely on TLAB's having been filled and must do
4175 // so here just in case a scavenge did not happen.
4176 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4177 // Update the saved marks which may affect the root scans.
4178 gch->save_marks();
4179
4180 print_eden_and_survivor_chunk_arrays();
4181
4182 {
4183 #if defined(COMPILER2) || INCLUDE_JVMCI
4184 DerivedPointerTableDeactivate dpt_deact;
4185 #endif
4186
4187 // Note on the role of the mod union table:
4188 // Since the marker in "markFromRoots" marks concurrently with
4189 // mutators, it is possible for some reachable objects not to have been
4190 // scanned. For instance, an only reference to an object A was
4191 // placed in object B after the marker scanned B. Unless B is rescanned,
4192 // A would be collected. Such updates to references in marked objects
4193 // are detected via the mod union table which is the set of all cards
4194 // dirtied since the first checkpoint in this GC cycle and prior to
4195 // the most recent young generation GC, minus those cleaned up by the
4196 // concurrent precleaning.
4197 if (CMSParallelRemarkEnabled) {
4198 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm);
4199 do_remark_parallel();
4200 } else {
4201 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm);
4202 do_remark_non_parallel();
4203 }
4204 }
4205 verify_work_stacks_empty();
4206 verify_overflow_empty();
4207
4208 {
4209 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm);
4210 refProcessingWork();
4211 }
4212 verify_work_stacks_empty();
4213 verify_overflow_empty();
4214
4215 if (should_unload_classes()) {
4216 CodeCache::gc_epilogue();
4217 }
4218 JvmtiExport::gc_epilogue();
4219
4220 // If we encountered any (marking stack / work queue) overflow
4221 // events during the current CMS cycle, take appropriate
4222 // remedial measures, where possible, so as to try and avoid
4223 // recurrence of that condition.
4224 assert(_markStack.isEmpty(), "No grey objects");
4225 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4226 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4227 if (ser_ovflw > 0) {
4228 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")",
4229 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4230 _markStack.expand();
4231 _ser_pmc_remark_ovflw = 0;
4232 _ser_pmc_preclean_ovflw = 0;
4233 _ser_kac_preclean_ovflw = 0;
4234 _ser_kac_ovflw = 0;
4235 }
4236 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4237 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")",
4238 _par_pmc_remark_ovflw, _par_kac_ovflw);
4239 _par_pmc_remark_ovflw = 0;
4240 _par_kac_ovflw = 0;
4241 }
4242 if (_markStack._hit_limit > 0) {
4243 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")",
4244 _markStack._hit_limit);
4245 }
4246 if (_markStack._failed_double > 0) {
4247 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT,
4248 _markStack._failed_double, _markStack.capacity());
4249 }
4250 _markStack._hit_limit = 0;
4251 _markStack._failed_double = 0;
4252
4253 if ((VerifyAfterGC || VerifyDuringGC) &&
4254 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4255 verify_after_remark();
4256 }
4257
4258 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4259
4260 // Change under the freelistLocks.
4261 _collectorState = Sweeping;
4262 // Call isAllClear() under bitMapLock
4263 assert(_modUnionTable.isAllClear(),
4264 "Should be clear by end of the final marking");
4265 assert(_ct->klass_rem_set()->mod_union_is_clear(),
4266 "Should be clear by end of the final marking");
4267 }
4268
4269 void CMSParInitialMarkTask::work(uint worker_id) {
4270 elapsedTimer _timer;
4271 ResourceMark rm;
4272 HandleMark hm;
4273
4274 // ---------- scan from roots --------------
4275 _timer.start();
4276 GenCollectedHeap* gch = GenCollectedHeap::heap();
4277 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4278
4279 // ---------- young gen roots --------------
4280 {
4281 work_on_young_gen_roots(&par_mri_cl);
4282 _timer.stop();
4283 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4284 }
4285
4286 // ---------- remaining roots --------------
4287 _timer.reset();
4288 _timer.start();
4289
4290 CLDToOopClosure cld_closure(&par_mri_cl, true);
4291
4292 gch->gen_process_roots(_strong_roots_scope,
4293 GenCollectedHeap::OldGen,
4294 false, // yg was scanned above
4295 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4296 _collector->should_unload_classes(),
4297 &par_mri_cl,
4298 NULL,
4299 &cld_closure);
4300 assert(_collector->should_unload_classes()
4301 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4302 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4303 _timer.stop();
4304 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4305 }
4306
4307 // Parallel remark task
4308 class CMSParRemarkTask: public CMSParMarkTask {
4309 CompactibleFreeListSpace* _cms_space;
4310
4311 // The per-thread work queues, available here for stealing.
4312 OopTaskQueueSet* _task_queues;
4313 ParallelTaskTerminator _term;
4314 StrongRootsScope* _strong_roots_scope;
4315
4316 public:
4317 // A value of 0 passed to n_workers will cause the number of
4318 // workers to be taken from the active workers in the work gang.
4319 CMSParRemarkTask(CMSCollector* collector,
4320 CompactibleFreeListSpace* cms_space,
4321 uint n_workers, WorkGang* workers,
4322 OopTaskQueueSet* task_queues,
4323 StrongRootsScope* strong_roots_scope):
4324 CMSParMarkTask("Rescan roots and grey objects in parallel",
4325 collector, n_workers),
4326 _cms_space(cms_space),
4327 _task_queues(task_queues),
4328 _term(n_workers, task_queues),
4329 _strong_roots_scope(strong_roots_scope) { }
4330
4331 OopTaskQueueSet* task_queues() { return _task_queues; }
4332
4333 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4334
4335 ParallelTaskTerminator* terminator() { return &_term; }
4336 uint n_workers() { return _n_workers; }
4337
4338 void work(uint worker_id);
4339
4340 private:
4341 // ... of dirty cards in old space
4342 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4343 ParMarkRefsIntoAndScanClosure* cl);
4344
4345 // ... work stealing for the above
4346 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed);
4347 };
4348
4349 class RemarkKlassClosure : public KlassClosure {
4350 KlassToOopClosure _cm_klass_closure;
4351 public:
4352 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4353 void do_klass(Klass* k) {
4354 // Check if we have modified any oops in the Klass during the concurrent marking.
4355 if (k->has_accumulated_modified_oops()) {
4356 k->clear_accumulated_modified_oops();
4357
4358 // We could have transfered the current modified marks to the accumulated marks,
4359 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4360 } else if (k->has_modified_oops()) {
4361 // Don't clear anything, this info is needed by the next young collection.
4362 } else {
4363 // No modified oops in the Klass.
4364 return;
4365 }
4366
4367 // The klass has modified fields, need to scan the klass.
4368 _cm_klass_closure.do_klass(k);
4369 }
4370 };
4371
4372 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) {
4373 ParNewGeneration* young_gen = _collector->_young_gen;
4374 ContiguousSpace* eden_space = young_gen->eden();
4375 ContiguousSpace* from_space = young_gen->from();
4376 ContiguousSpace* to_space = young_gen->to();
4377
4378 HeapWord** eca = _collector->_eden_chunk_array;
4379 size_t ect = _collector->_eden_chunk_index;
4380 HeapWord** sca = _collector->_survivor_chunk_array;
4381 size_t sct = _collector->_survivor_chunk_index;
4382
4383 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4384 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4385
4386 do_young_space_rescan(cl, to_space, NULL, 0);
4387 do_young_space_rescan(cl, from_space, sca, sct);
4388 do_young_space_rescan(cl, eden_space, eca, ect);
4389 }
4390
4391 // work_queue(i) is passed to the closure
4392 // ParMarkRefsIntoAndScanClosure. The "i" parameter
4393 // also is passed to do_dirty_card_rescan_tasks() and to
4394 // do_work_steal() to select the i-th task_queue.
4395
4396 void CMSParRemarkTask::work(uint worker_id) {
4397 elapsedTimer _timer;
4398 ResourceMark rm;
4399 HandleMark hm;
4400
4401 // ---------- rescan from roots --------------
4402 _timer.start();
4403 GenCollectedHeap* gch = GenCollectedHeap::heap();
4404 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4405 _collector->_span, _collector->ref_processor(),
4406 &(_collector->_markBitMap),
4407 work_queue(worker_id));
4408
4409 // Rescan young gen roots first since these are likely
4410 // coarsely partitioned and may, on that account, constitute
4411 // the critical path; thus, it's best to start off that
4412 // work first.
4413 // ---------- young gen roots --------------
4414 {
4415 work_on_young_gen_roots(&par_mrias_cl);
4416 _timer.stop();
4417 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4418 }
4419
4420 // ---------- remaining roots --------------
4421 _timer.reset();
4422 _timer.start();
4423 gch->gen_process_roots(_strong_roots_scope,
4424 GenCollectedHeap::OldGen,
4425 false, // yg was scanned above
4426 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4427 _collector->should_unload_classes(),
4428 &par_mrias_cl,
4429 NULL,
4430 NULL); // The dirty klasses will be handled below
4431
4432 assert(_collector->should_unload_classes()
4433 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4434 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4435 _timer.stop();
4436 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4437
4438 // ---------- unhandled CLD scanning ----------
4439 if (worker_id == 0) { // Single threaded at the moment.
4440 _timer.reset();
4441 _timer.start();
4442
4443 // Scan all new class loader data objects and new dependencies that were
4444 // introduced during concurrent marking.
4445 ResourceMark rm;
4446 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4447 for (int i = 0; i < array->length(); i++) {
4448 par_mrias_cl.do_cld_nv(array->at(i));
4449 }
4450
4451 // We don't need to keep track of new CLDs anymore.
4452 ClassLoaderDataGraph::remember_new_clds(false);
4453
4454 _timer.stop();
4455 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4456 }
4457
4458 // ---------- dirty klass scanning ----------
4459 if (worker_id == 0) { // Single threaded at the moment.
4460 _timer.reset();
4461 _timer.start();
4462
4463 // Scan all classes that was dirtied during the concurrent marking phase.
4464 RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
4465 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
4466
4467 _timer.stop();
4468 log_trace(gc, task)("Finished dirty klass scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4469 }
4470
4471 // We might have added oops to ClassLoaderData::_handles during the
4472 // concurrent marking phase. These oops point to newly allocated objects
4473 // that are guaranteed to be kept alive. Either by the direct allocation
4474 // code, or when the young collector processes the roots. Hence,
4475 // we don't have to revisit the _handles block during the remark phase.
4476
4477 // ---------- rescan dirty cards ------------
4478 _timer.reset();
4479 _timer.start();
4480
4481 // Do the rescan tasks for each of the two spaces
4482 // (cms_space) in turn.
4483 // "worker_id" is passed to select the task_queue for "worker_id"
4484 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4485 _timer.stop();
4486 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4487
4488 // ---------- steal work from other threads ...
4489 // ---------- ... and drain overflow list.
4490 _timer.reset();
4491 _timer.start();
4492 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4493 _timer.stop();
4494 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4495 }
4496
4497 void
4498 CMSParMarkTask::do_young_space_rescan(
4499 OopsInGenClosure* cl, ContiguousSpace* space,
4500 HeapWord** chunk_array, size_t chunk_top) {
4501 // Until all tasks completed:
4502 // . claim an unclaimed task
4503 // . compute region boundaries corresponding to task claimed
4504 // using chunk_array
4505 // . par_oop_iterate(cl) over that region
4506
4507 ResourceMark rm;
4508 HandleMark hm;
4509
4510 SequentialSubTasksDone* pst = space->par_seq_tasks();
4511
4512 uint nth_task = 0;
4513 uint n_tasks = pst->n_tasks();
4514
4515 if (n_tasks > 0) {
4516 assert(pst->valid(), "Uninitialized use?");
4517 HeapWord *start, *end;
4518 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4519 // We claimed task # nth_task; compute its boundaries.
4520 if (chunk_top == 0) { // no samples were taken
4521 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4522 start = space->bottom();
4523 end = space->top();
4524 } else if (nth_task == 0) {
4525 start = space->bottom();
4526 end = chunk_array[nth_task];
4527 } else if (nth_task < (uint)chunk_top) {
4528 assert(nth_task >= 1, "Control point invariant");
4529 start = chunk_array[nth_task - 1];
4530 end = chunk_array[nth_task];
4531 } else {
4532 assert(nth_task == (uint)chunk_top, "Control point invariant");
4533 start = chunk_array[chunk_top - 1];
4534 end = space->top();
4535 }
4536 MemRegion mr(start, end);
4537 // Verify that mr is in space
4538 assert(mr.is_empty() || space->used_region().contains(mr),
4539 "Should be in space");
4540 // Verify that "start" is an object boundary
4541 assert(mr.is_empty() || oop(mr.start())->is_oop(),
4542 "Should be an oop");
4543 space->par_oop_iterate(mr, cl);
4544 }
4545 pst->all_tasks_completed();
4546 }
4547 }
4548
4549 void
4550 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4551 CompactibleFreeListSpace* sp, int i,
4552 ParMarkRefsIntoAndScanClosure* cl) {
4553 // Until all tasks completed:
4554 // . claim an unclaimed task
4555 // . compute region boundaries corresponding to task claimed
4556 // . transfer dirty bits ct->mut for that region
4557 // . apply rescanclosure to dirty mut bits for that region
4558
4559 ResourceMark rm;
4560 HandleMark hm;
4561
4562 OopTaskQueue* work_q = work_queue(i);
4563 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4564 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4565 // CAUTION: This closure has state that persists across calls to
4566 // the work method dirty_range_iterate_clear() in that it has
4567 // embedded in it a (subtype of) UpwardsObjectClosure. The
4568 // use of that state in the embedded UpwardsObjectClosure instance
4569 // assumes that the cards are always iterated (even if in parallel
4570 // by several threads) in monotonically increasing order per each
4571 // thread. This is true of the implementation below which picks
4572 // card ranges (chunks) in monotonically increasing order globally
4573 // and, a-fortiori, in monotonically increasing order per thread
4574 // (the latter order being a subsequence of the former).
4575 // If the work code below is ever reorganized into a more chaotic
4576 // work-partitioning form than the current "sequential tasks"
4577 // paradigm, the use of that persistent state will have to be
4578 // revisited and modified appropriately. See also related
4579 // bug 4756801 work on which should examine this code to make
4580 // sure that the changes there do not run counter to the
4581 // assumptions made here and necessary for correctness and
4582 // efficiency. Note also that this code might yield inefficient
4583 // behavior in the case of very large objects that span one or
4584 // more work chunks. Such objects would potentially be scanned
4585 // several times redundantly. Work on 4756801 should try and
4586 // address that performance anomaly if at all possible. XXX
4587 MemRegion full_span = _collector->_span;
4588 CMSBitMap* bm = &(_collector->_markBitMap); // shared
4589 MarkFromDirtyCardsClosure
4590 greyRescanClosure(_collector, full_span, // entire span of interest
4591 sp, bm, work_q, cl);
4592
4593 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4594 assert(pst->valid(), "Uninitialized use?");
4595 uint nth_task = 0;
4596 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
4597 MemRegion span = sp->used_region();
4598 HeapWord* start_addr = span.start();
4599 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
4600 alignment);
4601 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4602 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
4603 start_addr, "Check alignment");
4604 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
4605 chunk_size, "Check alignment");
4606
4607 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4608 // Having claimed the nth_task, compute corresponding mem-region,
4609 // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4610 // The alignment restriction ensures that we do not need any
4611 // synchronization with other gang-workers while setting or
4612 // clearing bits in thus chunk of the MUT.
4613 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4614 start_addr + (nth_task+1)*chunk_size);
4615 // The last chunk's end might be way beyond end of the
4616 // used region. In that case pull back appropriately.
4617 if (this_span.end() > end_addr) {
4618 this_span.set_end(end_addr);
4619 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4620 }
4621 // Iterate over the dirty cards covering this chunk, marking them
4622 // precleaned, and setting the corresponding bits in the mod union
4623 // table. Since we have been careful to partition at Card and MUT-word
4624 // boundaries no synchronization is needed between parallel threads.
4625 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
4626 &modUnionClosure);
4627
4628 // Having transferred these marks into the modUnionTable,
4629 // rescan the marked objects on the dirty cards in the modUnionTable.
4630 // Even if this is at a synchronous collection, the initial marking
4631 // may have been done during an asynchronous collection so there
4632 // may be dirty bits in the mod-union table.
4633 _collector->_modUnionTable.dirty_range_iterate_clear(
4634 this_span, &greyRescanClosure);
4635 _collector->_modUnionTable.verifyNoOneBitsInRange(
4636 this_span.start(),
4637 this_span.end());
4638 }
4639 pst->all_tasks_completed(); // declare that i am done
4640 }
4641
4642 // . see if we can share work_queues with ParNew? XXX
4643 void
4644 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl,
4645 int* seed) {
4646 OopTaskQueue* work_q = work_queue(i);
4647 NOT_PRODUCT(int num_steals = 0;)
4648 oop obj_to_scan;
4649 CMSBitMap* bm = &(_collector->_markBitMap);
4650
4651 while (true) {
4652 // Completely finish any left over work from (an) earlier round(s)
4653 cl->trim_queue(0);
4654 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4655 (size_t)ParGCDesiredObjsFromOverflowList);
4656 // Now check if there's any work in the overflow list
4657 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4658 // only affects the number of attempts made to get work from the
4659 // overflow list and does not affect the number of workers. Just
4660 // pass ParallelGCThreads so this behavior is unchanged.
4661 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4662 work_q,
4663 ParallelGCThreads)) {
4664 // found something in global overflow list;
4665 // not yet ready to go stealing work from others.
4666 // We'd like to assert(work_q->size() != 0, ...)
4667 // because we just took work from the overflow list,
4668 // but of course we can't since all of that could have
4669 // been already stolen from us.
4670 // "He giveth and He taketh away."
4671 continue;
4672 }
4673 // Verify that we have no work before we resort to stealing
4674 assert(work_q->size() == 0, "Have work, shouldn't steal");
4675 // Try to steal from other queues that have work
4676 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4677 NOT_PRODUCT(num_steals++;)
4678 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
4679 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4680 // Do scanning work
4681 obj_to_scan->oop_iterate(cl);
4682 // Loop around, finish this work, and try to steal some more
4683 } else if (terminator()->offer_termination()) {
4684 break; // nirvana from the infinite cycle
4685 }
4686 }
4687 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
4688 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4689 "Else our work is not yet done");
4690 }
4691
4692 // Record object boundaries in _eden_chunk_array by sampling the eden
4693 // top in the slow-path eden object allocation code path and record
4694 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4695 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4696 // sampling in sample_eden() that activates during the part of the
4697 // preclean phase.
4698 void CMSCollector::sample_eden_chunk() {
4699 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4700 if (_eden_chunk_lock->try_lock()) {
4701 // Record a sample. This is the critical section. The contents
4702 // of the _eden_chunk_array have to be non-decreasing in the
4703 // address order.
4704 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4705 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4706 "Unexpected state of Eden");
4707 if (_eden_chunk_index == 0 ||
4708 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4709 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4710 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4711 _eden_chunk_index++; // commit sample
4712 }
4713 _eden_chunk_lock->unlock();
4714 }
4715 }
4716 }
4717
4718 // Return a thread-local PLAB recording array, as appropriate.
4719 void* CMSCollector::get_data_recorder(int thr_num) {
4720 if (_survivor_plab_array != NULL &&
4721 (CMSPLABRecordAlways ||
4722 (_collectorState > Marking && _collectorState < FinalMarking))) {
4723 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4724 ChunkArray* ca = &_survivor_plab_array[thr_num];
4725 ca->reset(); // clear it so that fresh data is recorded
4726 return (void*) ca;
4727 } else {
4728 return NULL;
4729 }
4730 }
4731
4732 // Reset all the thread-local PLAB recording arrays
4733 void CMSCollector::reset_survivor_plab_arrays() {
4734 for (uint i = 0; i < ParallelGCThreads; i++) {
4735 _survivor_plab_array[i].reset();
4736 }
4737 }
4738
4739 // Merge the per-thread plab arrays into the global survivor chunk
4740 // array which will provide the partitioning of the survivor space
4741 // for CMS initial scan and rescan.
4742 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4743 int no_of_gc_threads) {
4744 assert(_survivor_plab_array != NULL, "Error");
4745 assert(_survivor_chunk_array != NULL, "Error");
4746 assert(_collectorState == FinalMarking ||
4747 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4748 for (int j = 0; j < no_of_gc_threads; j++) {
4749 _cursor[j] = 0;
4750 }
4751 HeapWord* top = surv->top();
4752 size_t i;
4753 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
4754 HeapWord* min_val = top; // Higher than any PLAB address
4755 uint min_tid = 0; // position of min_val this round
4756 for (int j = 0; j < no_of_gc_threads; j++) {
4757 ChunkArray* cur_sca = &_survivor_plab_array[j];
4758 if (_cursor[j] == cur_sca->end()) {
4759 continue;
4760 }
4761 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4762 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4763 assert(surv->used_region().contains(cur_val), "Out of bounds value");
4764 if (cur_val < min_val) {
4765 min_tid = j;
4766 min_val = cur_val;
4767 } else {
4768 assert(cur_val < top, "All recorded addresses should be less");
4769 }
4770 }
4771 // At this point min_val and min_tid are respectively
4772 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4773 // and the thread (j) that witnesses that address.
4774 // We record this address in the _survivor_chunk_array[i]
4775 // and increment _cursor[min_tid] prior to the next round i.
4776 if (min_val == top) {
4777 break;
4778 }
4779 _survivor_chunk_array[i] = min_val;
4780 _cursor[min_tid]++;
4781 }
4782 // We are all done; record the size of the _survivor_chunk_array
4783 _survivor_chunk_index = i; // exclusive: [0, i)
4784 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4785 // Verify that we used up all the recorded entries
4786 #ifdef ASSERT
4787 size_t total = 0;
4788 for (int j = 0; j < no_of_gc_threads; j++) {
4789 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4790 total += _cursor[j];
4791 }
4792 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
4793 // Check that the merged array is in sorted order
4794 if (total > 0) {
4795 for (size_t i = 0; i < total - 1; i++) {
4796 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
4797 i, p2i(_survivor_chunk_array[i]));
4798 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
4799 "Not sorted");
4800 }
4801 }
4802 #endif // ASSERT
4803 }
4804
4805 // Set up the space's par_seq_tasks structure for work claiming
4806 // for parallel initial scan and rescan of young gen.
4807 // See ParRescanTask where this is currently used.
4808 void
4809 CMSCollector::
4810 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
4811 assert(n_threads > 0, "Unexpected n_threads argument");
4812
4813 // Eden space
4814 if (!_young_gen->eden()->is_empty()) {
4815 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
4816 assert(!pst->valid(), "Clobbering existing data?");
4817 // Each valid entry in [0, _eden_chunk_index) represents a task.
4818 size_t n_tasks = _eden_chunk_index + 1;
4819 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
4820 // Sets the condition for completion of the subtask (how many threads
4821 // need to finish in order to be done).
4822 pst->set_n_threads(n_threads);
4823 pst->set_n_tasks((int)n_tasks);
4824 }
4825
4826 // Merge the survivor plab arrays into _survivor_chunk_array
4827 if (_survivor_plab_array != NULL) {
4828 merge_survivor_plab_arrays(_young_gen->from(), n_threads);
4829 } else {
4830 assert(_survivor_chunk_index == 0, "Error");
4831 }
4832
4833 // To space
4834 {
4835 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
4836 assert(!pst->valid(), "Clobbering existing data?");
4837 // Sets the condition for completion of the subtask (how many threads
4838 // need to finish in order to be done).
4839 pst->set_n_threads(n_threads);
4840 pst->set_n_tasks(1);
4841 assert(pst->valid(), "Error");
4842 }
4843
4844 // From space
4845 {
4846 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
4847 assert(!pst->valid(), "Clobbering existing data?");
4848 size_t n_tasks = _survivor_chunk_index + 1;
4849 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
4850 // Sets the condition for completion of the subtask (how many threads
4851 // need to finish in order to be done).
4852 pst->set_n_threads(n_threads);
4853 pst->set_n_tasks((int)n_tasks);
4854 assert(pst->valid(), "Error");
4855 }
4856 }
4857
4858 // Parallel version of remark
4859 void CMSCollector::do_remark_parallel() {
4860 GenCollectedHeap* gch = GenCollectedHeap::heap();
4861 WorkGang* workers = gch->workers();
4862 assert(workers != NULL, "Need parallel worker threads.");
4863 // Choose to use the number of GC workers most recently set
4864 // into "active_workers".
4865 uint n_workers = workers->active_workers();
4866
4867 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4868
4869 StrongRootsScope srs(n_workers);
4870
4871 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
4872
4873 // We won't be iterating over the cards in the card table updating
4874 // the younger_gen cards, so we shouldn't call the following else
4875 // the verification code as well as subsequent younger_refs_iterate
4876 // code would get confused. XXX
4877 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
4878
4879 // The young gen rescan work will not be done as part of
4880 // process_roots (which currently doesn't know how to
4881 // parallelize such a scan), but rather will be broken up into
4882 // a set of parallel tasks (via the sampling that the [abortable]
4883 // preclean phase did of eden, plus the [two] tasks of
4884 // scanning the [two] survivor spaces. Further fine-grain
4885 // parallelization of the scanning of the survivor spaces
4886 // themselves, and of precleaning of the young gen itself
4887 // is deferred to the future.
4888 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
4889
4890 // The dirty card rescan work is broken up into a "sequence"
4891 // of parallel tasks (per constituent space) that are dynamically
4892 // claimed by the parallel threads.
4893 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
4894
4895 // It turns out that even when we're using 1 thread, doing the work in a
4896 // separate thread causes wide variance in run times. We can't help this
4897 // in the multi-threaded case, but we special-case n=1 here to get
4898 // repeatable measurements of the 1-thread overhead of the parallel code.
4899 if (n_workers > 1) {
4900 // Make refs discovery MT-safe, if it isn't already: it may not
4901 // necessarily be so, since it's possible that we are doing
4902 // ST marking.
4903 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
4904 workers->run_task(&tsk);
4905 } else {
4906 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4907 tsk.work(0);
4908 }
4909
4910 // restore, single-threaded for now, any preserved marks
4911 // as a result of work_q overflow
4912 restore_preserved_marks_if_any();
4913 }
4914
4915 // Non-parallel version of remark
4916 void CMSCollector::do_remark_non_parallel() {
4917 ResourceMark rm;
4918 HandleMark hm;
4919 GenCollectedHeap* gch = GenCollectedHeap::heap();
4920 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4921
4922 MarkRefsIntoAndScanClosure
4923 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
4924 &_markStack, this,
4925 false /* should_yield */, false /* not precleaning */);
4926 MarkFromDirtyCardsClosure
4927 markFromDirtyCardsClosure(this, _span,
4928 NULL, // space is set further below
4929 &_markBitMap, &_markStack, &mrias_cl);
4930 {
4931 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm);
4932 // Iterate over the dirty cards, setting the corresponding bits in the
4933 // mod union table.
4934 {
4935 ModUnionClosure modUnionClosure(&_modUnionTable);
4936 _ct->ct_bs()->dirty_card_iterate(
4937 _cmsGen->used_region(),
4938 &modUnionClosure);
4939 }
4940 // Having transferred these marks into the modUnionTable, we just need
4941 // to rescan the marked objects on the dirty cards in the modUnionTable.
4942 // The initial marking may have been done during an asynchronous
4943 // collection so there may be dirty bits in the mod-union table.
4944 const int alignment =
4945 CardTableModRefBS::card_size * BitsPerWord;
4946 {
4947 // ... First handle dirty cards in CMS gen
4948 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
4949 MemRegion ur = _cmsGen->used_region();
4950 HeapWord* lb = ur.start();
4951 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
4952 MemRegion cms_span(lb, ub);
4953 _modUnionTable.dirty_range_iterate_clear(cms_span,
4954 &markFromDirtyCardsClosure);
4955 verify_work_stacks_empty();
4956 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards());
4957 }
4958 }
4959 if (VerifyDuringGC &&
4960 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4961 HandleMark hm; // Discard invalid handles created during verification
4962 Universe::verify();
4963 }
4964 {
4965 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm);
4966
4967 verify_work_stacks_empty();
4968
4969 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
4970 StrongRootsScope srs(1);
4971
4972 gch->gen_process_roots(&srs,
4973 GenCollectedHeap::OldGen,
4974 true, // young gen as roots
4975 GenCollectedHeap::ScanningOption(roots_scanning_options()),
4976 should_unload_classes(),
4977 &mrias_cl,
4978 NULL,
4979 NULL); // The dirty klasses will be handled below
4980
4981 assert(should_unload_classes()
4982 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4983 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4984 }
4985
4986 {
4987 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm);
4988
4989 verify_work_stacks_empty();
4990
4991 // Scan all class loader data objects that might have been introduced
4992 // during concurrent marking.
4993 ResourceMark rm;
4994 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4995 for (int i = 0; i < array->length(); i++) {
4996 mrias_cl.do_cld_nv(array->at(i));
4997 }
4998
4999 // We don't need to keep track of new CLDs anymore.
5000 ClassLoaderDataGraph::remember_new_clds(false);
5001
5002 verify_work_stacks_empty();
5003 }
5004
5005 {
5006 GCTraceTime(Trace, gc, phases) t("Dirty Klass Scan", _gc_timer_cm);
5007
5008 verify_work_stacks_empty();
5009
5010 RemarkKlassClosure remark_klass_closure(&mrias_cl);
5011 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5012
5013 verify_work_stacks_empty();
5014 }
5015
5016 // We might have added oops to ClassLoaderData::_handles during the
5017 // concurrent marking phase. These oops point to newly allocated objects
5018 // that are guaranteed to be kept alive. Either by the direct allocation
5019 // code, or when the young collector processes the roots. Hence,
5020 // we don't have to revisit the _handles block during the remark phase.
5021
5022 verify_work_stacks_empty();
5023 // Restore evacuated mark words, if any, used for overflow list links
5024 restore_preserved_marks_if_any();
5025
5026 verify_overflow_empty();
5027 }
5028
5029 ////////////////////////////////////////////////////////
5030 // Parallel Reference Processing Task Proxy Class
5031 ////////////////////////////////////////////////////////
5032 class AbstractGangTaskWOopQueues : public AbstractGangTask {
5033 OopTaskQueueSet* _queues;
5034 ParallelTaskTerminator _terminator;
5035 public:
5036 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) :
5037 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {}
5038 ParallelTaskTerminator* terminator() { return &_terminator; }
5039 OopTaskQueueSet* queues() { return _queues; }
5040 };
5041
5042 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5043 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5044 CMSCollector* _collector;
5045 CMSBitMap* _mark_bit_map;
5046 const MemRegion _span;
5047 ProcessTask& _task;
5048
5049 public:
5050 CMSRefProcTaskProxy(ProcessTask& task,
5051 CMSCollector* collector,
5052 const MemRegion& span,
5053 CMSBitMap* mark_bit_map,
5054 AbstractWorkGang* workers,
5055 OopTaskQueueSet* task_queues):
5056 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5057 task_queues,
5058 workers->active_workers()),
5059 _task(task),
5060 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5061 {
5062 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5063 "Inconsistency in _span");
5064 }
5065
5066 OopTaskQueueSet* task_queues() { return queues(); }
5067
5068 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5069
5070 void do_work_steal(int i,
5071 CMSParDrainMarkingStackClosure* drain,
5072 CMSParKeepAliveClosure* keep_alive,
5073 int* seed);
5074
5075 virtual void work(uint worker_id);
5076 };
5077
5078 void CMSRefProcTaskProxy::work(uint worker_id) {
5079 ResourceMark rm;
5080 HandleMark hm;
5081 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5082 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5083 _mark_bit_map,
5084 work_queue(worker_id));
5085 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5086 _mark_bit_map,
5087 work_queue(worker_id));
5088 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5089 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5090 if (_task.marks_oops_alive()) {
5091 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5092 _collector->hash_seed(worker_id));
5093 }
5094 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5095 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5096 }
5097
5098 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5099 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5100 EnqueueTask& _task;
5101
5102 public:
5103 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5104 : AbstractGangTask("Enqueue reference objects in parallel"),
5105 _task(task)
5106 { }
5107
5108 virtual void work(uint worker_id)
5109 {
5110 _task.work(worker_id);
5111 }
5112 };
5113
5114 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5115 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5116 _span(span),
5117 _bit_map(bit_map),
5118 _work_queue(work_queue),
5119 _mark_and_push(collector, span, bit_map, work_queue),
5120 _low_water_mark(MIN2((work_queue->max_elems()/4),
5121 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5122 { }
5123
5124 // . see if we can share work_queues with ParNew? XXX
5125 void CMSRefProcTaskProxy::do_work_steal(int i,
5126 CMSParDrainMarkingStackClosure* drain,
5127 CMSParKeepAliveClosure* keep_alive,
5128 int* seed) {
5129 OopTaskQueue* work_q = work_queue(i);
5130 NOT_PRODUCT(int num_steals = 0;)
5131 oop obj_to_scan;
5132
5133 while (true) {
5134 // Completely finish any left over work from (an) earlier round(s)
5135 drain->trim_queue(0);
5136 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5137 (size_t)ParGCDesiredObjsFromOverflowList);
5138 // Now check if there's any work in the overflow list
5139 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5140 // only affects the number of attempts made to get work from the
5141 // overflow list and does not affect the number of workers. Just
5142 // pass ParallelGCThreads so this behavior is unchanged.
5143 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5144 work_q,
5145 ParallelGCThreads)) {
5146 // Found something in global overflow list;
5147 // not yet ready to go stealing work from others.
5148 // We'd like to assert(work_q->size() != 0, ...)
5149 // because we just took work from the overflow list,
5150 // but of course we can't, since all of that might have
5151 // been already stolen from us.
5152 continue;
5153 }
5154 // Verify that we have no work before we resort to stealing
5155 assert(work_q->size() == 0, "Have work, shouldn't steal");
5156 // Try to steal from other queues that have work
5157 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5158 NOT_PRODUCT(num_steals++;)
5159 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5160 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5161 // Do scanning work
5162 obj_to_scan->oop_iterate(keep_alive);
5163 // Loop around, finish this work, and try to steal some more
5164 } else if (terminator()->offer_termination()) {
5165 break; // nirvana from the infinite cycle
5166 }
5167 }
5168 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
5169 }
5170
5171 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5172 {
5173 GenCollectedHeap* gch = GenCollectedHeap::heap();
5174 WorkGang* workers = gch->workers();
5175 assert(workers != NULL, "Need parallel worker threads.");
5176 CMSRefProcTaskProxy rp_task(task, &_collector,
5177 _collector.ref_processor()->span(),
5178 _collector.markBitMap(),
5179 workers, _collector.task_queues());
5180 workers->run_task(&rp_task);
5181 }
5182
5183 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5184 {
5185
5186 GenCollectedHeap* gch = GenCollectedHeap::heap();
5187 WorkGang* workers = gch->workers();
5188 assert(workers != NULL, "Need parallel worker threads.");
5189 CMSRefEnqueueTaskProxy enq_task(task);
5190 workers->run_task(&enq_task);
5191 }
5192
5193 void CMSCollector::refProcessingWork() {
5194 ResourceMark rm;
5195 HandleMark hm;
5196
5197 ReferenceProcessor* rp = ref_processor();
5198 assert(rp->span().equals(_span), "Spans should be equal");
5199 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5200 // Process weak references.
5201 rp->setup_policy(false);
5202 verify_work_stacks_empty();
5203
5204 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5205 &_markStack, false /* !preclean */);
5206 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5207 _span, &_markBitMap, &_markStack,
5208 &cmsKeepAliveClosure, false /* !preclean */);
5209 {
5210 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm);
5211
5212 ReferenceProcessorStats stats;
5213 if (rp->processing_is_mt()) {
5214 // Set the degree of MT here. If the discovery is done MT, there
5215 // may have been a different number of threads doing the discovery
5216 // and a different number of discovered lists may have Ref objects.
5217 // That is OK as long as the Reference lists are balanced (see
5218 // balance_all_queues() and balance_queues()).
5219 GenCollectedHeap* gch = GenCollectedHeap::heap();
5220 uint active_workers = ParallelGCThreads;
5221 WorkGang* workers = gch->workers();
5222 if (workers != NULL) {
5223 active_workers = workers->active_workers();
5224 // The expectation is that active_workers will have already
5225 // been set to a reasonable value. If it has not been set,
5226 // investigate.
5227 assert(active_workers > 0, "Should have been set during scavenge");
5228 }
5229 rp->set_active_mt_degree(active_workers);
5230 CMSRefProcTaskExecutor task_executor(*this);
5231 stats = rp->process_discovered_references(&_is_alive_closure,
5232 &cmsKeepAliveClosure,
5233 &cmsDrainMarkingStackClosure,
5234 &task_executor,
5235 _gc_timer_cm);
5236 } else {
5237 stats = rp->process_discovered_references(&_is_alive_closure,
5238 &cmsKeepAliveClosure,
5239 &cmsDrainMarkingStackClosure,
5240 NULL,
5241 _gc_timer_cm);
5242 }
5243 _gc_tracer_cm->report_gc_reference_stats(stats);
5244
5245 }
5246
5247 // This is the point where the entire marking should have completed.
5248 verify_work_stacks_empty();
5249
5250 if (should_unload_classes()) {
5251 {
5252 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm);
5253
5254 // Unload classes and purge the SystemDictionary.
5255 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5256
5257 // Unload nmethods.
5258 CodeCache::do_unloading(&_is_alive_closure, purged_class);
5259
5260 // Prune dead klasses from subklass/sibling/implementor lists.
5261 Klass::clean_weak_klass_links(&_is_alive_closure);
5262 }
5263
5264 {
5265 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm);
5266 // Clean up unreferenced symbols in symbol table.
5267 SymbolTable::unlink();
5268 }
5269
5270 {
5271 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm);
5272 // Delete entries for dead interned strings.
5273 StringTable::unlink(&_is_alive_closure);
5274 }
5275 }
5276
5277
5278 // Restore any preserved marks as a result of mark stack or
5279 // work queue overflow
5280 restore_preserved_marks_if_any(); // done single-threaded for now
5281
5282 rp->set_enqueuing_is_done(true);
5283 if (rp->processing_is_mt()) {
5284 rp->balance_all_queues();
5285 CMSRefProcTaskExecutor task_executor(*this);
5286 rp->enqueue_discovered_references(&task_executor);
5287 } else {
5288 rp->enqueue_discovered_references(NULL);
5289 }
5290 rp->verify_no_references_recorded();
5291 assert(!rp->discovery_enabled(), "should have been disabled");
5292 }
5293
5294 #ifndef PRODUCT
5295 void CMSCollector::check_correct_thread_executing() {
5296 Thread* t = Thread::current();
5297 // Only the VM thread or the CMS thread should be here.
5298 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5299 "Unexpected thread type");
5300 // If this is the vm thread, the foreground process
5301 // should not be waiting. Note that _foregroundGCIsActive is
5302 // true while the foreground collector is waiting.
5303 if (_foregroundGCShouldWait) {
5304 // We cannot be the VM thread
5305 assert(t->is_ConcurrentGC_thread(),
5306 "Should be CMS thread");
5307 } else {
5308 // We can be the CMS thread only if we are in a stop-world
5309 // phase of CMS collection.
5310 if (t->is_ConcurrentGC_thread()) {
5311 assert(_collectorState == InitialMarking ||
5312 _collectorState == FinalMarking,
5313 "Should be a stop-world phase");
5314 // The CMS thread should be holding the CMS_token.
5315 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5316 "Potential interference with concurrently "
5317 "executing VM thread");
5318 }
5319 }
5320 }
5321 #endif
5322
5323 void CMSCollector::sweep() {
5324 assert(_collectorState == Sweeping, "just checking");
5325 check_correct_thread_executing();
5326 verify_work_stacks_empty();
5327 verify_overflow_empty();
5328 increment_sweep_count();
5329 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5330
5331 _inter_sweep_timer.stop();
5332 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5333
5334 assert(!_intra_sweep_timer.is_active(), "Should not be active");
5335 _intra_sweep_timer.reset();
5336 _intra_sweep_timer.start();
5337 {
5338 GCTraceCPUTime tcpu;
5339 CMSPhaseAccounting pa(this, "Concurrent Sweep");
5340 // First sweep the old gen
5341 {
5342 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5343 bitMapLock());
5344 sweepWork(_cmsGen);
5345 }
5346
5347 // Update Universe::_heap_*_at_gc figures.
5348 // We need all the free list locks to make the abstract state
5349 // transition from Sweeping to Resetting. See detailed note
5350 // further below.
5351 {
5352 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5353 // Update heap occupancy information which is used as
5354 // input to soft ref clearing policy at the next gc.
5355 Universe::update_heap_info_at_gc();
5356 _collectorState = Resizing;
5357 }
5358 }
5359 verify_work_stacks_empty();
5360 verify_overflow_empty();
5361
5362 if (should_unload_classes()) {
5363 // Delay purge to the beginning of the next safepoint. Metaspace::contains
5364 // requires that the virtual spaces are stable and not deleted.
5365 ClassLoaderDataGraph::set_should_purge(true);
5366 }
5367
5368 _intra_sweep_timer.stop();
5369 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5370
5371 _inter_sweep_timer.reset();
5372 _inter_sweep_timer.start();
5373
5374 // We need to use a monotonically non-decreasing time in ms
5375 // or we will see time-warp warnings and os::javaTimeMillis()
5376 // does not guarantee monotonicity.
5377 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5378 update_time_of_last_gc(now);
5379
5380 // NOTE on abstract state transitions:
5381 // Mutators allocate-live and/or mark the mod-union table dirty
5382 // based on the state of the collection. The former is done in
5383 // the interval [Marking, Sweeping] and the latter in the interval
5384 // [Marking, Sweeping). Thus the transitions into the Marking state
5385 // and out of the Sweeping state must be synchronously visible
5386 // globally to the mutators.
5387 // The transition into the Marking state happens with the world
5388 // stopped so the mutators will globally see it. Sweeping is
5389 // done asynchronously by the background collector so the transition
5390 // from the Sweeping state to the Resizing state must be done
5391 // under the freelistLock (as is the check for whether to
5392 // allocate-live and whether to dirty the mod-union table).
5393 assert(_collectorState == Resizing, "Change of collector state to"
5394 " Resizing must be done under the freelistLocks (plural)");
5395
5396 // Now that sweeping has been completed, we clear
5397 // the incremental_collection_failed flag,
5398 // thus inviting a younger gen collection to promote into
5399 // this generation. If such a promotion may still fail,
5400 // the flag will be set again when a young collection is
5401 // attempted.
5402 GenCollectedHeap* gch = GenCollectedHeap::heap();
5403 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
5404 gch->update_full_collections_completed(_collection_count_start);
5405 }
5406
5407 // FIX ME!!! Looks like this belongs in CFLSpace, with
5408 // CMSGen merely delegating to it.
5409 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5410 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5411 HeapWord* minAddr = _cmsSpace->bottom();
5412 HeapWord* largestAddr =
5413 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5414 if (largestAddr == NULL) {
5415 // The dictionary appears to be empty. In this case
5416 // try to coalesce at the end of the heap.
5417 largestAddr = _cmsSpace->end();
5418 }
5419 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5420 size_t nearLargestOffset =
5421 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5422 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5423 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5424 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5425 }
5426
5427 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5428 return addr >= _cmsSpace->nearLargestChunk();
5429 }
5430
5431 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5432 return _cmsSpace->find_chunk_at_end();
5433 }
5434
5435 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation,
5436 bool full) {
5437 // If the young generation has been collected, gather any statistics
5438 // that are of interest at this point.
5439 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation);
5440 if (!full && current_is_young) {
5441 // Gather statistics on the young generation collection.
5442 collector()->stats().record_gc0_end(used());
5443 }
5444 }
5445
5446 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) {
5447 // We iterate over the space(s) underlying this generation,
5448 // checking the mark bit map to see if the bits corresponding
5449 // to specific blocks are marked or not. Blocks that are
5450 // marked are live and are not swept up. All remaining blocks
5451 // are swept up, with coalescing on-the-fly as we sweep up
5452 // contiguous free and/or garbage blocks:
5453 // We need to ensure that the sweeper synchronizes with allocators
5454 // and stop-the-world collectors. In particular, the following
5455 // locks are used:
5456 // . CMS token: if this is held, a stop the world collection cannot occur
5457 // . freelistLock: if this is held no allocation can occur from this
5458 // generation by another thread
5459 // . bitMapLock: if this is held, no other thread can access or update
5460 //
5461
5462 // Note that we need to hold the freelistLock if we use
5463 // block iterate below; else the iterator might go awry if
5464 // a mutator (or promotion) causes block contents to change
5465 // (for instance if the allocator divvies up a block).
5466 // If we hold the free list lock, for all practical purposes
5467 // young generation GC's can't occur (they'll usually need to
5468 // promote), so we might as well prevent all young generation
5469 // GC's while we do a sweeping step. For the same reason, we might
5470 // as well take the bit map lock for the entire duration
5471
5472 // check that we hold the requisite locks
5473 assert(have_cms_token(), "Should hold cms token");
5474 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5475 assert_lock_strong(old_gen->freelistLock());
5476 assert_lock_strong(bitMapLock());
5477
5478 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5479 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
5480 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5481 _inter_sweep_estimate.padded_average(),
5482 _intra_sweep_estimate.padded_average());
5483 old_gen->setNearLargestChunk();
5484
5485 {
5486 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield);
5487 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5488 // We need to free-up/coalesce garbage/blocks from a
5489 // co-terminal free run. This is done in the SweepClosure
5490 // destructor; so, do not remove this scope, else the
5491 // end-of-sweep-census below will be off by a little bit.
5492 }
5493 old_gen->cmsSpace()->sweep_completed();
5494 old_gen->cmsSpace()->endSweepFLCensus(sweep_count());
5495 if (should_unload_classes()) { // unloaded classes this cycle,
5496 _concurrent_cycles_since_last_unload = 0; // ... reset count
5497 } else { // did not unload classes,
5498 _concurrent_cycles_since_last_unload++; // ... increment count
5499 }
5500 }
5501
5502 // Reset CMS data structures (for now just the marking bit map)
5503 // preparatory for the next cycle.
5504 void CMSCollector::reset_concurrent() {
5505 CMSTokenSyncWithLocks ts(true, bitMapLock());
5506
5507 // If the state is not "Resetting", the foreground thread
5508 // has done a collection and the resetting.
5509 if (_collectorState != Resetting) {
5510 assert(_collectorState == Idling, "The state should only change"
5511 " because the foreground collector has finished the collection");
5512 return;
5513 }
5514
5515 {
5516 // Clear the mark bitmap (no grey objects to start with)
5517 // for the next cycle.
5518 GCTraceCPUTime tcpu;
5519 CMSPhaseAccounting cmspa(this, "Concurrent Reset");
5520
5521 HeapWord* curAddr = _markBitMap.startWord();
5522 while (curAddr < _markBitMap.endWord()) {
5523 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
5524 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5525 _markBitMap.clear_large_range(chunk);
5526 if (ConcurrentMarkSweepThread::should_yield() &&
5527 !foregroundGCIsActive() &&
5528 CMSYield) {
5529 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5530 "CMS thread should hold CMS token");
5531 assert_lock_strong(bitMapLock());
5532 bitMapLock()->unlock();
5533 ConcurrentMarkSweepThread::desynchronize(true);
5534 stopTimer();
5535 incrementYields();
5536
5537 // See the comment in coordinator_yield()
5538 for (unsigned i = 0; i < CMSYieldSleepCount &&
5539 ConcurrentMarkSweepThread::should_yield() &&
5540 !CMSCollector::foregroundGCIsActive(); ++i) {
5541 os::sleep(Thread::current(), 1, false);
5542 }
5543
5544 ConcurrentMarkSweepThread::synchronize(true);
5545 bitMapLock()->lock_without_safepoint_check();
5546 startTimer();
5547 }
5548 curAddr = chunk.end();
5549 }
5550 // A successful mostly concurrent collection has been done.
5551 // Because only the full (i.e., concurrent mode failure) collections
5552 // are being measured for gc overhead limits, clean the "near" flag
5553 // and count.
5554 size_policy()->reset_gc_overhead_limit_count();
5555 _collectorState = Idling;
5556 }
5557
5558 register_gc_end();
5559 }
5560
5561 // Same as above but for STW paths
5562 void CMSCollector::reset_stw() {
5563 // already have the lock
5564 assert(_collectorState == Resetting, "just checking");
5565 assert_lock_strong(bitMapLock());
5566 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id());
5567 _markBitMap.clear_all();
5568 _collectorState = Idling;
5569 register_gc_end();
5570 }
5571
5572 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5573 GCTraceCPUTime tcpu;
5574 TraceCollectorStats tcs(counters());
5575
5576 switch (op) {
5577 case CMS_op_checkpointRootsInitial: {
5578 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true);
5579 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5580 checkpointRootsInitial();
5581 break;
5582 }
5583 case CMS_op_checkpointRootsFinal: {
5584 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true);
5585 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5586 checkpointRootsFinal();
5587 break;
5588 }
5589 default:
5590 fatal("No such CMS_op");
5591 }
5592 }
5593
5594 #ifndef PRODUCT
5595 size_t const CMSCollector::skip_header_HeapWords() {
5596 return FreeChunk::header_size();
5597 }
5598
5599 // Try and collect here conditions that should hold when
5600 // CMS thread is exiting. The idea is that the foreground GC
5601 // thread should not be blocked if it wants to terminate
5602 // the CMS thread and yet continue to run the VM for a while
5603 // after that.
5604 void CMSCollector::verify_ok_to_terminate() const {
5605 assert(Thread::current()->is_ConcurrentGC_thread(),
5606 "should be called by CMS thread");
5607 assert(!_foregroundGCShouldWait, "should be false");
5608 // We could check here that all the various low-level locks
5609 // are not held by the CMS thread, but that is overkill; see
5610 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5611 // is checked.
5612 }
5613 #endif
5614
5615 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5616 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5617 "missing Printezis mark?");
5618 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5619 size_t size = pointer_delta(nextOneAddr + 1, addr);
5620 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5621 "alignment problem");
5622 assert(size >= 3, "Necessary for Printezis marks to work");
5623 return size;
5624 }
5625
5626 // A variant of the above (block_size_using_printezis_bits()) except
5627 // that we return 0 if the P-bits are not yet set.
5628 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5629 if (_markBitMap.isMarked(addr + 1)) {
5630 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5631 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5632 size_t size = pointer_delta(nextOneAddr + 1, addr);
5633 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5634 "alignment problem");
5635 assert(size >= 3, "Necessary for Printezis marks to work");
5636 return size;
5637 }
5638 return 0;
5639 }
5640
5641 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5642 size_t sz = 0;
5643 oop p = (oop)addr;
5644 if (p->klass_or_null() != NULL) {
5645 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5646 } else {
5647 sz = block_size_using_printezis_bits(addr);
5648 }
5649 assert(sz > 0, "size must be nonzero");
5650 HeapWord* next_block = addr + sz;
5651 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
5652 CardTableModRefBS::card_size);
5653 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
5654 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
5655 "must be different cards");
5656 return next_card;
5657 }
5658
5659
5660 // CMS Bit Map Wrapper /////////////////////////////////////////
5661
5662 // Construct a CMS bit map infrastructure, but don't create the
5663 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5664 // further below.
5665 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5666 _bm(),
5667 _shifter(shifter),
5668 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5669 Monitor::_safepoint_check_sometimes) : NULL)
5670 {
5671 _bmStartWord = 0;
5672 _bmWordSize = 0;
5673 }
5674
5675 bool CMSBitMap::allocate(MemRegion mr) {
5676 _bmStartWord = mr.start();
5677 _bmWordSize = mr.word_size();
5678 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5679 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5680 if (!brs.is_reserved()) {
5681 log_warning(gc)("CMS bit map allocation failure");
5682 return false;
5683 }
5684 // For now we'll just commit all of the bit map up front.
5685 // Later on we'll try to be more parsimonious with swap.
5686 if (!_virtual_space.initialize(brs, brs.size())) {
5687 log_warning(gc)("CMS bit map backing store failure");
5688 return false;
5689 }
5690 assert(_virtual_space.committed_size() == brs.size(),
5691 "didn't reserve backing store for all of CMS bit map?");
5692 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5693 _bmWordSize, "inconsistency in bit map sizing");
5694 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter);
5695
5696 // bm.clear(); // can we rely on getting zero'd memory? verify below
5697 assert(isAllClear(),
5698 "Expected zero'd memory from ReservedSpace constructor");
5699 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5700 "consistency check");
5701 return true;
5702 }
5703
5704 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5705 HeapWord *next_addr, *end_addr, *last_addr;
5706 assert_locked();
5707 assert(covers(mr), "out-of-range error");
5708 // XXX assert that start and end are appropriately aligned
5709 for (next_addr = mr.start(), end_addr = mr.end();
5710 next_addr < end_addr; next_addr = last_addr) {
5711 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5712 last_addr = dirty_region.end();
5713 if (!dirty_region.is_empty()) {
5714 cl->do_MemRegion(dirty_region);
5715 } else {
5716 assert(last_addr == end_addr, "program logic");
5717 return;
5718 }
5719 }
5720 }
5721
5722 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5723 _bm.print_on_error(st, prefix);
5724 }
5725
5726 #ifndef PRODUCT
5727 void CMSBitMap::assert_locked() const {
5728 CMSLockVerifier::assert_locked(lock());
5729 }
5730
5731 bool CMSBitMap::covers(MemRegion mr) const {
5732 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5733 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5734 "size inconsistency");
5735 return (mr.start() >= _bmStartWord) &&
5736 (mr.end() <= endWord());
5737 }
5738
5739 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5740 return (start >= _bmStartWord && (start + size) <= endWord());
5741 }
5742
5743 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5744 // verify that there are no 1 bits in the interval [left, right)
5745 FalseBitMapClosure falseBitMapClosure;
5746 iterate(&falseBitMapClosure, left, right);
5747 }
5748
5749 void CMSBitMap::region_invariant(MemRegion mr)
5750 {
5751 assert_locked();
5752 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5753 assert(!mr.is_empty(), "unexpected empty region");
5754 assert(covers(mr), "mr should be covered by bit map");
5755 // convert address range into offset range
5756 size_t start_ofs = heapWordToOffset(mr.start());
5757 // Make sure that end() is appropriately aligned
5758 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
5759 (1 << (_shifter+LogHeapWordSize))),
5760 "Misaligned mr.end()");
5761 size_t end_ofs = heapWordToOffset(mr.end());
5762 assert(end_ofs > start_ofs, "Should mark at least one bit");
5763 }
5764
5765 #endif
5766
5767 bool CMSMarkStack::allocate(size_t size) {
5768 // allocate a stack of the requisite depth
5769 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5770 size * sizeof(oop)));
5771 if (!rs.is_reserved()) {
5772 log_warning(gc)("CMSMarkStack allocation failure");
5773 return false;
5774 }
5775 if (!_virtual_space.initialize(rs, rs.size())) {
5776 log_warning(gc)("CMSMarkStack backing store failure");
5777 return false;
5778 }
5779 assert(_virtual_space.committed_size() == rs.size(),
5780 "didn't reserve backing store for all of CMS stack?");
5781 _base = (oop*)(_virtual_space.low());
5782 _index = 0;
5783 _capacity = size;
5784 NOT_PRODUCT(_max_depth = 0);
5785 return true;
5786 }
5787
5788 // XXX FIX ME !!! In the MT case we come in here holding a
5789 // leaf lock. For printing we need to take a further lock
5790 // which has lower rank. We need to recalibrate the two
5791 // lock-ranks involved in order to be able to print the
5792 // messages below. (Or defer the printing to the caller.
5793 // For now we take the expedient path of just disabling the
5794 // messages for the problematic case.)
5795 void CMSMarkStack::expand() {
5796 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
5797 if (_capacity == MarkStackSizeMax) {
5798 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) {
5799 // We print a warning message only once per CMS cycle.
5800 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit");
5801 }
5802 return;
5803 }
5804 // Double capacity if possible
5805 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
5806 // Do not give up existing stack until we have managed to
5807 // get the double capacity that we desired.
5808 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5809 new_capacity * sizeof(oop)));
5810 if (rs.is_reserved()) {
5811 // Release the backing store associated with old stack
5812 _virtual_space.release();
5813 // Reinitialize virtual space for new stack
5814 if (!_virtual_space.initialize(rs, rs.size())) {
5815 fatal("Not enough swap for expanded marking stack");
5816 }
5817 _base = (oop*)(_virtual_space.low());
5818 _index = 0;
5819 _capacity = new_capacity;
5820 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) {
5821 // Failed to double capacity, continue;
5822 // we print a detail message only once per CMS cycle.
5823 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
5824 _capacity / K, new_capacity / K);
5825 }
5826 }
5827
5828
5829 // Closures
5830 // XXX: there seems to be a lot of code duplication here;
5831 // should refactor and consolidate common code.
5832
5833 // This closure is used to mark refs into the CMS generation in
5834 // the CMS bit map. Called at the first checkpoint. This closure
5835 // assumes that we do not need to re-mark dirty cards; if the CMS
5836 // generation on which this is used is not an oldest
5837 // generation then this will lose younger_gen cards!
5838
5839 MarkRefsIntoClosure::MarkRefsIntoClosure(
5840 MemRegion span, CMSBitMap* bitMap):
5841 _span(span),
5842 _bitMap(bitMap)
5843 {
5844 assert(ref_processor() == NULL, "deliberately left NULL");
5845 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5846 }
5847
5848 void MarkRefsIntoClosure::do_oop(oop obj) {
5849 // if p points into _span, then mark corresponding bit in _markBitMap
5850 assert(obj->is_oop(), "expected an oop");
5851 HeapWord* addr = (HeapWord*)obj;
5852 if (_span.contains(addr)) {
5853 // this should be made more efficient
5854 _bitMap->mark(addr);
5855 }
5856 }
5857
5858 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
5859 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
5860
5861 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure(
5862 MemRegion span, CMSBitMap* bitMap):
5863 _span(span),
5864 _bitMap(bitMap)
5865 {
5866 assert(ref_processor() == NULL, "deliberately left NULL");
5867 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5868 }
5869
5870 void ParMarkRefsIntoClosure::do_oop(oop obj) {
5871 // if p points into _span, then mark corresponding bit in _markBitMap
5872 assert(obj->is_oop(), "expected an oop");
5873 HeapWord* addr = (HeapWord*)obj;
5874 if (_span.contains(addr)) {
5875 // this should be made more efficient
5876 _bitMap->par_mark(addr);
5877 }
5878 }
5879
5880 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); }
5881 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); }
5882
5883 // A variant of the above, used for CMS marking verification.
5884 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
5885 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
5886 _span(span),
5887 _verification_bm(verification_bm),
5888 _cms_bm(cms_bm)
5889 {
5890 assert(ref_processor() == NULL, "deliberately left NULL");
5891 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
5892 }
5893
5894 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
5895 // if p points into _span, then mark corresponding bit in _markBitMap
5896 assert(obj->is_oop(), "expected an oop");
5897 HeapWord* addr = (HeapWord*)obj;
5898 if (_span.contains(addr)) {
5899 _verification_bm->mark(addr);
5900 if (!_cms_bm->isMarked(addr)) {
5901 Log(gc, verify) log;
5902 ResourceMark rm;
5903 oop(addr)->print_on(log.error_stream());
5904 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
5905 fatal("... aborting");
5906 }
5907 }
5908 }
5909
5910 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5911 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5912
5913 //////////////////////////////////////////////////
5914 // MarkRefsIntoAndScanClosure
5915 //////////////////////////////////////////////////
5916
5917 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
5918 ReferenceProcessor* rp,
5919 CMSBitMap* bit_map,
5920 CMSBitMap* mod_union_table,
5921 CMSMarkStack* mark_stack,
5922 CMSCollector* collector,
5923 bool should_yield,
5924 bool concurrent_precleaning):
5925 _collector(collector),
5926 _span(span),
5927 _bit_map(bit_map),
5928 _mark_stack(mark_stack),
5929 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
5930 mark_stack, concurrent_precleaning),
5931 _yield(should_yield),
5932 _concurrent_precleaning(concurrent_precleaning),
5933 _freelistLock(NULL)
5934 {
5935 // FIXME: Should initialize in base class constructor.
5936 assert(rp != NULL, "ref_processor shouldn't be NULL");
5937 set_ref_processor_internal(rp);
5938 }
5939
5940 // This closure is used to mark refs into the CMS generation at the
5941 // second (final) checkpoint, and to scan and transitively follow
5942 // the unmarked oops. It is also used during the concurrent precleaning
5943 // phase while scanning objects on dirty cards in the CMS generation.
5944 // The marks are made in the marking bit map and the marking stack is
5945 // used for keeping the (newly) grey objects during the scan.
5946 // The parallel version (Par_...) appears further below.
5947 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
5948 if (obj != NULL) {
5949 assert(obj->is_oop(), "expected an oop");
5950 HeapWord* addr = (HeapWord*)obj;
5951 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
5952 assert(_collector->overflow_list_is_empty(),
5953 "overflow list should be empty");
5954 if (_span.contains(addr) &&
5955 !_bit_map->isMarked(addr)) {
5956 // mark bit map (object is now grey)
5957 _bit_map->mark(addr);
5958 // push on marking stack (stack should be empty), and drain the
5959 // stack by applying this closure to the oops in the oops popped
5960 // from the stack (i.e. blacken the grey objects)
5961 bool res = _mark_stack->push(obj);
5962 assert(res, "Should have space to push on empty stack");
5963 do {
5964 oop new_oop = _mark_stack->pop();
5965 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
5966 assert(_bit_map->isMarked((HeapWord*)new_oop),
5967 "only grey objects on this stack");
5968 // iterate over the oops in this oop, marking and pushing
5969 // the ones in CMS heap (i.e. in _span).
5970 new_oop->oop_iterate(&_pushAndMarkClosure);
5971 // check if it's time to yield
5972 do_yield_check();
5973 } while (!_mark_stack->isEmpty() ||
5974 (!_concurrent_precleaning && take_from_overflow_list()));
5975 // if marking stack is empty, and we are not doing this
5976 // during precleaning, then check the overflow list
5977 }
5978 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
5979 assert(_collector->overflow_list_is_empty(),
5980 "overflow list was drained above");
5981
5982 assert(_collector->no_preserved_marks(),
5983 "All preserved marks should have been restored above");
5984 }
5985 }
5986
5987 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5988 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5989
5990 void MarkRefsIntoAndScanClosure::do_yield_work() {
5991 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5992 "CMS thread should hold CMS token");
5993 assert_lock_strong(_freelistLock);
5994 assert_lock_strong(_bit_map->lock());
5995 // relinquish the free_list_lock and bitMaplock()
5996 _bit_map->lock()->unlock();
5997 _freelistLock->unlock();
5998 ConcurrentMarkSweepThread::desynchronize(true);
5999 _collector->stopTimer();
6000 _collector->incrementYields();
6001
6002 // See the comment in coordinator_yield()
6003 for (unsigned i = 0;
6004 i < CMSYieldSleepCount &&
6005 ConcurrentMarkSweepThread::should_yield() &&
6006 !CMSCollector::foregroundGCIsActive();
6007 ++i) {
6008 os::sleep(Thread::current(), 1, false);
6009 }
6010
6011 ConcurrentMarkSweepThread::synchronize(true);
6012 _freelistLock->lock_without_safepoint_check();
6013 _bit_map->lock()->lock_without_safepoint_check();
6014 _collector->startTimer();
6015 }
6016
6017 ///////////////////////////////////////////////////////////
6018 // ParMarkRefsIntoAndScanClosure: a parallel version of
6019 // MarkRefsIntoAndScanClosure
6020 ///////////////////////////////////////////////////////////
6021 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure(
6022 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6023 CMSBitMap* bit_map, OopTaskQueue* work_queue):
6024 _span(span),
6025 _bit_map(bit_map),
6026 _work_queue(work_queue),
6027 _low_water_mark(MIN2((work_queue->max_elems()/4),
6028 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6029 _parPushAndMarkClosure(collector, span, rp, bit_map, work_queue)
6030 {
6031 // FIXME: Should initialize in base class constructor.
6032 assert(rp != NULL, "ref_processor shouldn't be NULL");
6033 set_ref_processor_internal(rp);
6034 }
6035
6036 // This closure is used to mark refs into the CMS generation at the
6037 // second (final) checkpoint, and to scan and transitively follow
6038 // the unmarked oops. The marks are made in the marking bit map and
6039 // the work_queue is used for keeping the (newly) grey objects during
6040 // the scan phase whence they are also available for stealing by parallel
6041 // threads. Since the marking bit map is shared, updates are
6042 // synchronized (via CAS).
6043 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) {
6044 if (obj != NULL) {
6045 // Ignore mark word because this could be an already marked oop
6046 // that may be chained at the end of the overflow list.
6047 assert(obj->is_oop(true), "expected an oop");
6048 HeapWord* addr = (HeapWord*)obj;
6049 if (_span.contains(addr) &&
6050 !_bit_map->isMarked(addr)) {
6051 // mark bit map (object will become grey):
6052 // It is possible for several threads to be
6053 // trying to "claim" this object concurrently;
6054 // the unique thread that succeeds in marking the
6055 // object first will do the subsequent push on
6056 // to the work queue (or overflow list).
6057 if (_bit_map->par_mark(addr)) {
6058 // push on work_queue (which may not be empty), and trim the
6059 // queue to an appropriate length by applying this closure to
6060 // the oops in the oops popped from the stack (i.e. blacken the
6061 // grey objects)
6062 bool res = _work_queue->push(obj);
6063 assert(res, "Low water mark should be less than capacity?");
6064 trim_queue(_low_water_mark);
6065 } // Else, another thread claimed the object
6066 }
6067 }
6068 }
6069
6070 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6071 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6072
6073 // This closure is used to rescan the marked objects on the dirty cards
6074 // in the mod union table and the card table proper.
6075 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6076 oop p, MemRegion mr) {
6077
6078 size_t size = 0;
6079 HeapWord* addr = (HeapWord*)p;
6080 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6081 assert(_span.contains(addr), "we are scanning the CMS generation");
6082 // check if it's time to yield
6083 if (do_yield_check()) {
6084 // We yielded for some foreground stop-world work,
6085 // and we have been asked to abort this ongoing preclean cycle.
6086 return 0;
6087 }
6088 if (_bitMap->isMarked(addr)) {
6089 // it's marked; is it potentially uninitialized?
6090 if (p->klass_or_null() != NULL) {
6091 // an initialized object; ignore mark word in verification below
6092 // since we are running concurrent with mutators
6093 assert(p->is_oop(true), "should be an oop");
6094 if (p->is_objArray()) {
6095 // objArrays are precisely marked; restrict scanning
6096 // to dirty cards only.
6097 size = CompactibleFreeListSpace::adjustObjectSize(
6098 p->oop_iterate_size(_scanningClosure, mr));
6099 } else {
6100 // A non-array may have been imprecisely marked; we need
6101 // to scan object in its entirety.
6102 size = CompactibleFreeListSpace::adjustObjectSize(
6103 p->oop_iterate_size(_scanningClosure));
6104 }
6105 #ifdef ASSERT
6106 size_t direct_size =
6107 CompactibleFreeListSpace::adjustObjectSize(p->size());
6108 assert(size == direct_size, "Inconsistency in size");
6109 assert(size >= 3, "Necessary for Printezis marks to work");
6110 HeapWord* start_pbit = addr + 1;
6111 HeapWord* end_pbit = addr + size - 1;
6112 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit),
6113 "inconsistent Printezis mark");
6114 // Verify inner mark bits (between Printezis bits) are clear,
6115 // but don't repeat if there are multiple dirty regions for
6116 // the same object, to avoid potential O(N^2) performance.
6117 if (addr != _last_scanned_object) {
6118 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit);
6119 _last_scanned_object = addr;
6120 }
6121 #endif // ASSERT
6122 } else {
6123 // An uninitialized object.
6124 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6125 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6126 size = pointer_delta(nextOneAddr + 1, addr);
6127 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6128 "alignment problem");
6129 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6130 // will dirty the card when the klass pointer is installed in the
6131 // object (signaling the completion of initialization).
6132 }
6133 } else {
6134 // Either a not yet marked object or an uninitialized object
6135 if (p->klass_or_null() == NULL) {
6136 // An uninitialized object, skip to the next card, since
6137 // we may not be able to read its P-bits yet.
6138 assert(size == 0, "Initial value");
6139 } else {
6140 // An object not (yet) reached by marking: we merely need to
6141 // compute its size so as to go look at the next block.
6142 assert(p->is_oop(true), "should be an oop");
6143 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6144 }
6145 }
6146 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6147 return size;
6148 }
6149
6150 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6151 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6152 "CMS thread should hold CMS token");
6153 assert_lock_strong(_freelistLock);
6154 assert_lock_strong(_bitMap->lock());
6155 // relinquish the free_list_lock and bitMaplock()
6156 _bitMap->lock()->unlock();
6157 _freelistLock->unlock();
6158 ConcurrentMarkSweepThread::desynchronize(true);
6159 _collector->stopTimer();
6160 _collector->incrementYields();
6161
6162 // See the comment in coordinator_yield()
6163 for (unsigned i = 0; i < CMSYieldSleepCount &&
6164 ConcurrentMarkSweepThread::should_yield() &&
6165 !CMSCollector::foregroundGCIsActive(); ++i) {
6166 os::sleep(Thread::current(), 1, false);
6167 }
6168
6169 ConcurrentMarkSweepThread::synchronize(true);
6170 _freelistLock->lock_without_safepoint_check();
6171 _bitMap->lock()->lock_without_safepoint_check();
6172 _collector->startTimer();
6173 }
6174
6175
6176 //////////////////////////////////////////////////////////////////
6177 // SurvivorSpacePrecleanClosure
6178 //////////////////////////////////////////////////////////////////
6179 // This (single-threaded) closure is used to preclean the oops in
6180 // the survivor spaces.
6181 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6182
6183 HeapWord* addr = (HeapWord*)p;
6184 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6185 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6186 assert(p->klass_or_null() != NULL, "object should be initialized");
6187 // an initialized object; ignore mark word in verification below
6188 // since we are running concurrent with mutators
6189 assert(p->is_oop(true), "should be an oop");
6190 // Note that we do not yield while we iterate over
6191 // the interior oops of p, pushing the relevant ones
6192 // on our marking stack.
6193 size_t size = p->oop_iterate_size(_scanning_closure);
6194 do_yield_check();
6195 // Observe that below, we do not abandon the preclean
6196 // phase as soon as we should; rather we empty the
6197 // marking stack before returning. This is to satisfy
6198 // some existing assertions. In general, it may be a
6199 // good idea to abort immediately and complete the marking
6200 // from the grey objects at a later time.
6201 while (!_mark_stack->isEmpty()) {
6202 oop new_oop = _mark_stack->pop();
6203 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6204 assert(_bit_map->isMarked((HeapWord*)new_oop),
6205 "only grey objects on this stack");
6206 // iterate over the oops in this oop, marking and pushing
6207 // the ones in CMS heap (i.e. in _span).
6208 new_oop->oop_iterate(_scanning_closure);
6209 // check if it's time to yield
6210 do_yield_check();
6211 }
6212 unsigned int after_count =
6213 GenCollectedHeap::heap()->total_collections();
6214 bool abort = (_before_count != after_count) ||
6215 _collector->should_abort_preclean();
6216 return abort ? 0 : size;
6217 }
6218
6219 void SurvivorSpacePrecleanClosure::do_yield_work() {
6220 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6221 "CMS thread should hold CMS token");
6222 assert_lock_strong(_bit_map->lock());
6223 // Relinquish the bit map lock
6224 _bit_map->lock()->unlock();
6225 ConcurrentMarkSweepThread::desynchronize(true);
6226 _collector->stopTimer();
6227 _collector->incrementYields();
6228
6229 // See the comment in coordinator_yield()
6230 for (unsigned i = 0; i < CMSYieldSleepCount &&
6231 ConcurrentMarkSweepThread::should_yield() &&
6232 !CMSCollector::foregroundGCIsActive(); ++i) {
6233 os::sleep(Thread::current(), 1, false);
6234 }
6235
6236 ConcurrentMarkSweepThread::synchronize(true);
6237 _bit_map->lock()->lock_without_safepoint_check();
6238 _collector->startTimer();
6239 }
6240
6241 // This closure is used to rescan the marked objects on the dirty cards
6242 // in the mod union table and the card table proper. In the parallel
6243 // case, although the bitMap is shared, we do a single read so the
6244 // isMarked() query is "safe".
6245 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6246 // Ignore mark word because we are running concurrent with mutators
6247 assert(p->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p));
6248 HeapWord* addr = (HeapWord*)p;
6249 assert(_span.contains(addr), "we are scanning the CMS generation");
6250 bool is_obj_array = false;
6251 #ifdef ASSERT
6252 if (!_parallel) {
6253 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6254 assert(_collector->overflow_list_is_empty(),
6255 "overflow list should be empty");
6256
6257 }
6258 #endif // ASSERT
6259 if (_bit_map->isMarked(addr)) {
6260 // Obj arrays are precisely marked, non-arrays are not;
6261 // so we scan objArrays precisely and non-arrays in their
6262 // entirety.
6263 if (p->is_objArray()) {
6264 is_obj_array = true;
6265 if (_parallel) {
6266 p->oop_iterate(_par_scan_closure, mr);
6267 } else {
6268 p->oop_iterate(_scan_closure, mr);
6269 }
6270 } else {
6271 if (_parallel) {
6272 p->oop_iterate(_par_scan_closure);
6273 } else {
6274 p->oop_iterate(_scan_closure);
6275 }
6276 }
6277 }
6278 #ifdef ASSERT
6279 if (!_parallel) {
6280 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6281 assert(_collector->overflow_list_is_empty(),
6282 "overflow list should be empty");
6283
6284 }
6285 #endif // ASSERT
6286 return is_obj_array;
6287 }
6288
6289 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6290 MemRegion span,
6291 CMSBitMap* bitMap, CMSMarkStack* markStack,
6292 bool should_yield, bool verifying):
6293 _collector(collector),
6294 _span(span),
6295 _bitMap(bitMap),
6296 _mut(&collector->_modUnionTable),
6297 _markStack(markStack),
6298 _yield(should_yield),
6299 _skipBits(0)
6300 {
6301 assert(_markStack->isEmpty(), "stack should be empty");
6302 _finger = _bitMap->startWord();
6303 _threshold = _finger;
6304 assert(_collector->_restart_addr == NULL, "Sanity check");
6305 assert(_span.contains(_finger), "Out of bounds _finger?");
6306 DEBUG_ONLY(_verifying = verifying;)
6307 }
6308
6309 void MarkFromRootsClosure::reset(HeapWord* addr) {
6310 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6311 assert(_span.contains(addr), "Out of bounds _finger?");
6312 _finger = addr;
6313 _threshold = (HeapWord*)round_to(
6314 (intptr_t)_finger, CardTableModRefBS::card_size);
6315 }
6316
6317 // Should revisit to see if this should be restructured for
6318 // greater efficiency.
6319 bool MarkFromRootsClosure::do_bit(size_t offset) {
6320 if (_skipBits > 0) {
6321 _skipBits--;
6322 return true;
6323 }
6324 // convert offset into a HeapWord*
6325 HeapWord* addr = _bitMap->startWord() + offset;
6326 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6327 "address out of range");
6328 assert(_bitMap->isMarked(addr), "tautology");
6329 if (_bitMap->isMarked(addr+1)) {
6330 // this is an allocated but not yet initialized object
6331 assert(_skipBits == 0, "tautology");
6332 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6333 oop p = oop(addr);
6334 if (p->klass_or_null() == NULL) {
6335 DEBUG_ONLY(if (!_verifying) {)
6336 // We re-dirty the cards on which this object lies and increase
6337 // the _threshold so that we'll come back to scan this object
6338 // during the preclean or remark phase. (CMSCleanOnEnter)
6339 if (CMSCleanOnEnter) {
6340 size_t sz = _collector->block_size_using_printezis_bits(addr);
6341 HeapWord* end_card_addr = (HeapWord*)round_to(
6342 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6343 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6344 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6345 // Bump _threshold to end_card_addr; note that
6346 // _threshold cannot possibly exceed end_card_addr, anyhow.
6347 // This prevents future clearing of the card as the scan proceeds
6348 // to the right.
6349 assert(_threshold <= end_card_addr,
6350 "Because we are just scanning into this object");
6351 if (_threshold < end_card_addr) {
6352 _threshold = end_card_addr;
6353 }
6354 if (p->klass_or_null() != NULL) {
6355 // Redirty the range of cards...
6356 _mut->mark_range(redirty_range);
6357 } // ...else the setting of klass will dirty the card anyway.
6358 }
6359 DEBUG_ONLY(})
6360 return true;
6361 }
6362 }
6363 scanOopsInOop(addr);
6364 return true;
6365 }
6366
6367 // We take a break if we've been at this for a while,
6368 // so as to avoid monopolizing the locks involved.
6369 void MarkFromRootsClosure::do_yield_work() {
6370 // First give up the locks, then yield, then re-lock
6371 // We should probably use a constructor/destructor idiom to
6372 // do this unlock/lock or modify the MutexUnlocker class to
6373 // serve our purpose. XXX
6374 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6375 "CMS thread should hold CMS token");
6376 assert_lock_strong(_bitMap->lock());
6377 _bitMap->lock()->unlock();
6378 ConcurrentMarkSweepThread::desynchronize(true);
6379 _collector->stopTimer();
6380 _collector->incrementYields();
6381
6382 // See the comment in coordinator_yield()
6383 for (unsigned i = 0; i < CMSYieldSleepCount &&
6384 ConcurrentMarkSweepThread::should_yield() &&
6385 !CMSCollector::foregroundGCIsActive(); ++i) {
6386 os::sleep(Thread::current(), 1, false);
6387 }
6388
6389 ConcurrentMarkSweepThread::synchronize(true);
6390 _bitMap->lock()->lock_without_safepoint_check();
6391 _collector->startTimer();
6392 }
6393
6394 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6395 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6396 assert(_markStack->isEmpty(),
6397 "should drain stack to limit stack usage");
6398 // convert ptr to an oop preparatory to scanning
6399 oop obj = oop(ptr);
6400 // Ignore mark word in verification below, since we
6401 // may be running concurrent with mutators.
6402 assert(obj->is_oop(true), "should be an oop");
6403 assert(_finger <= ptr, "_finger runneth ahead");
6404 // advance the finger to right end of this object
6405 _finger = ptr + obj->size();
6406 assert(_finger > ptr, "we just incremented it above");
6407 // On large heaps, it may take us some time to get through
6408 // the marking phase. During
6409 // this time it's possible that a lot of mutations have
6410 // accumulated in the card table and the mod union table --
6411 // these mutation records are redundant until we have
6412 // actually traced into the corresponding card.
6413 // Here, we check whether advancing the finger would make
6414 // us cross into a new card, and if so clear corresponding
6415 // cards in the MUT (preclean them in the card-table in the
6416 // future).
6417
6418 DEBUG_ONLY(if (!_verifying) {)
6419 // The clean-on-enter optimization is disabled by default,
6420 // until we fix 6178663.
6421 if (CMSCleanOnEnter && (_finger > _threshold)) {
6422 // [_threshold, _finger) represents the interval
6423 // of cards to be cleared in MUT (or precleaned in card table).
6424 // The set of cards to be cleared is all those that overlap
6425 // with the interval [_threshold, _finger); note that
6426 // _threshold is always kept card-aligned but _finger isn't
6427 // always card-aligned.
6428 HeapWord* old_threshold = _threshold;
6429 assert(old_threshold == (HeapWord*)round_to(
6430 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6431 "_threshold should always be card-aligned");
6432 _threshold = (HeapWord*)round_to(
6433 (intptr_t)_finger, CardTableModRefBS::card_size);
6434 MemRegion mr(old_threshold, _threshold);
6435 assert(!mr.is_empty(), "Control point invariant");
6436 assert(_span.contains(mr), "Should clear within span");
6437 _mut->clear_range(mr);
6438 }
6439 DEBUG_ONLY(})
6440 // Note: the finger doesn't advance while we drain
6441 // the stack below.
6442 PushOrMarkClosure pushOrMarkClosure(_collector,
6443 _span, _bitMap, _markStack,
6444 _finger, this);
6445 bool res = _markStack->push(obj);
6446 assert(res, "Empty non-zero size stack should have space for single push");
6447 while (!_markStack->isEmpty()) {
6448 oop new_oop = _markStack->pop();
6449 // Skip verifying header mark word below because we are
6450 // running concurrent with mutators.
6451 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6452 // now scan this oop's oops
6453 new_oop->oop_iterate(&pushOrMarkClosure);
6454 do_yield_check();
6455 }
6456 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6457 }
6458
6459 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task,
6460 CMSCollector* collector, MemRegion span,
6461 CMSBitMap* bit_map,
6462 OopTaskQueue* work_queue,
6463 CMSMarkStack* overflow_stack):
6464 _collector(collector),
6465 _whole_span(collector->_span),
6466 _span(span),
6467 _bit_map(bit_map),
6468 _mut(&collector->_modUnionTable),
6469 _work_queue(work_queue),
6470 _overflow_stack(overflow_stack),
6471 _skip_bits(0),
6472 _task(task)
6473 {
6474 assert(_work_queue->size() == 0, "work_queue should be empty");
6475 _finger = span.start();
6476 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
6477 assert(_span.contains(_finger), "Out of bounds _finger?");
6478 }
6479
6480 // Should revisit to see if this should be restructured for
6481 // greater efficiency.
6482 bool ParMarkFromRootsClosure::do_bit(size_t offset) {
6483 if (_skip_bits > 0) {
6484 _skip_bits--;
6485 return true;
6486 }
6487 // convert offset into a HeapWord*
6488 HeapWord* addr = _bit_map->startWord() + offset;
6489 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6490 "address out of range");
6491 assert(_bit_map->isMarked(addr), "tautology");
6492 if (_bit_map->isMarked(addr+1)) {
6493 // this is an allocated object that might not yet be initialized
6494 assert(_skip_bits == 0, "tautology");
6495 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
6496 oop p = oop(addr);
6497 if (p->klass_or_null() == NULL) {
6498 // in the case of Clean-on-Enter optimization, redirty card
6499 // and avoid clearing card by increasing the threshold.
6500 return true;
6501 }
6502 }
6503 scan_oops_in_oop(addr);
6504 return true;
6505 }
6506
6507 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6508 assert(_bit_map->isMarked(ptr), "expected bit to be set");
6509 // Should we assert that our work queue is empty or
6510 // below some drain limit?
6511 assert(_work_queue->size() == 0,
6512 "should drain stack to limit stack usage");
6513 // convert ptr to an oop preparatory to scanning
6514 oop obj = oop(ptr);
6515 // Ignore mark word in verification below, since we
6516 // may be running concurrent with mutators.
6517 assert(obj->is_oop(true), "should be an oop");
6518 assert(_finger <= ptr, "_finger runneth ahead");
6519 // advance the finger to right end of this object
6520 _finger = ptr + obj->size();
6521 assert(_finger > ptr, "we just incremented it above");
6522 // On large heaps, it may take us some time to get through
6523 // the marking phase. During
6524 // this time it's possible that a lot of mutations have
6525 // accumulated in the card table and the mod union table --
6526 // these mutation records are redundant until we have
6527 // actually traced into the corresponding card.
6528 // Here, we check whether advancing the finger would make
6529 // us cross into a new card, and if so clear corresponding
6530 // cards in the MUT (preclean them in the card-table in the
6531 // future).
6532
6533 // The clean-on-enter optimization is disabled by default,
6534 // until we fix 6178663.
6535 if (CMSCleanOnEnter && (_finger > _threshold)) {
6536 // [_threshold, _finger) represents the interval
6537 // of cards to be cleared in MUT (or precleaned in card table).
6538 // The set of cards to be cleared is all those that overlap
6539 // with the interval [_threshold, _finger); note that
6540 // _threshold is always kept card-aligned but _finger isn't
6541 // always card-aligned.
6542 HeapWord* old_threshold = _threshold;
6543 assert(old_threshold == (HeapWord*)round_to(
6544 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6545 "_threshold should always be card-aligned");
6546 _threshold = (HeapWord*)round_to(
6547 (intptr_t)_finger, CardTableModRefBS::card_size);
6548 MemRegion mr(old_threshold, _threshold);
6549 assert(!mr.is_empty(), "Control point invariant");
6550 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6551 _mut->clear_range(mr);
6552 }
6553
6554 // Note: the local finger doesn't advance while we drain
6555 // the stack below, but the global finger sure can and will.
6556 HeapWord** gfa = _task->global_finger_addr();
6557 ParPushOrMarkClosure pushOrMarkClosure(_collector,
6558 _span, _bit_map,
6559 _work_queue,
6560 _overflow_stack,
6561 _finger,
6562 gfa, this);
6563 bool res = _work_queue->push(obj); // overflow could occur here
6564 assert(res, "Will hold once we use workqueues");
6565 while (true) {
6566 oop new_oop;
6567 if (!_work_queue->pop_local(new_oop)) {
6568 // We emptied our work_queue; check if there's stuff that can
6569 // be gotten from the overflow stack.
6570 if (CMSConcMarkingTask::get_work_from_overflow_stack(
6571 _overflow_stack, _work_queue)) {
6572 do_yield_check();
6573 continue;
6574 } else { // done
6575 break;
6576 }
6577 }
6578 // Skip verifying header mark word below because we are
6579 // running concurrent with mutators.
6580 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6581 // now scan this oop's oops
6582 new_oop->oop_iterate(&pushOrMarkClosure);
6583 do_yield_check();
6584 }
6585 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6586 }
6587
6588 // Yield in response to a request from VM Thread or
6589 // from mutators.
6590 void ParMarkFromRootsClosure::do_yield_work() {
6591 assert(_task != NULL, "sanity");
6592 _task->yield();
6593 }
6594
6595 // A variant of the above used for verifying CMS marking work.
6596 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6597 MemRegion span,
6598 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6599 CMSMarkStack* mark_stack):
6600 _collector(collector),
6601 _span(span),
6602 _verification_bm(verification_bm),
6603 _cms_bm(cms_bm),
6604 _mark_stack(mark_stack),
6605 _pam_verify_closure(collector, span, verification_bm, cms_bm,
6606 mark_stack)
6607 {
6608 assert(_mark_stack->isEmpty(), "stack should be empty");
6609 _finger = _verification_bm->startWord();
6610 assert(_collector->_restart_addr == NULL, "Sanity check");
6611 assert(_span.contains(_finger), "Out of bounds _finger?");
6612 }
6613
6614 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6615 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6616 assert(_span.contains(addr), "Out of bounds _finger?");
6617 _finger = addr;
6618 }
6619
6620 // Should revisit to see if this should be restructured for
6621 // greater efficiency.
6622 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6623 // convert offset into a HeapWord*
6624 HeapWord* addr = _verification_bm->startWord() + offset;
6625 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6626 "address out of range");
6627 assert(_verification_bm->isMarked(addr), "tautology");
6628 assert(_cms_bm->isMarked(addr), "tautology");
6629
6630 assert(_mark_stack->isEmpty(),
6631 "should drain stack to limit stack usage");
6632 // convert addr to an oop preparatory to scanning
6633 oop obj = oop(addr);
6634 assert(obj->is_oop(), "should be an oop");
6635 assert(_finger <= addr, "_finger runneth ahead");
6636 // advance the finger to right end of this object
6637 _finger = addr + obj->size();
6638 assert(_finger > addr, "we just incremented it above");
6639 // Note: the finger doesn't advance while we drain
6640 // the stack below.
6641 bool res = _mark_stack->push(obj);
6642 assert(res, "Empty non-zero size stack should have space for single push");
6643 while (!_mark_stack->isEmpty()) {
6644 oop new_oop = _mark_stack->pop();
6645 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
6646 // now scan this oop's oops
6647 new_oop->oop_iterate(&_pam_verify_closure);
6648 }
6649 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6650 return true;
6651 }
6652
6653 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6654 CMSCollector* collector, MemRegion span,
6655 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6656 CMSMarkStack* mark_stack):
6657 MetadataAwareOopClosure(collector->ref_processor()),
6658 _collector(collector),
6659 _span(span),
6660 _verification_bm(verification_bm),
6661 _cms_bm(cms_bm),
6662 _mark_stack(mark_stack)
6663 { }
6664
6665 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6666 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6667
6668 // Upon stack overflow, we discard (part of) the stack,
6669 // remembering the least address amongst those discarded
6670 // in CMSCollector's _restart_address.
6671 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6672 // Remember the least grey address discarded
6673 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6674 _collector->lower_restart_addr(ra);
6675 _mark_stack->reset(); // discard stack contents
6676 _mark_stack->expand(); // expand the stack if possible
6677 }
6678
6679 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6680 assert(obj->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6681 HeapWord* addr = (HeapWord*)obj;
6682 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6683 // Oop lies in _span and isn't yet grey or black
6684 _verification_bm->mark(addr); // now grey
6685 if (!_cms_bm->isMarked(addr)) {
6686 Log(gc, verify) log;
6687 ResourceMark rm;
6688 oop(addr)->print_on(log.error_stream());
6689 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6690 fatal("... aborting");
6691 }
6692
6693 if (!_mark_stack->push(obj)) { // stack overflow
6694 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity());
6695 assert(_mark_stack->isFull(), "Else push should have succeeded");
6696 handle_stack_overflow(addr);
6697 }
6698 // anything including and to the right of _finger
6699 // will be scanned as we iterate over the remainder of the
6700 // bit map
6701 }
6702 }
6703
6704 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6705 MemRegion span,
6706 CMSBitMap* bitMap, CMSMarkStack* markStack,
6707 HeapWord* finger, MarkFromRootsClosure* parent) :
6708 MetadataAwareOopClosure(collector->ref_processor()),
6709 _collector(collector),
6710 _span(span),
6711 _bitMap(bitMap),
6712 _markStack(markStack),
6713 _finger(finger),
6714 _parent(parent)
6715 { }
6716
6717 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector,
6718 MemRegion span,
6719 CMSBitMap* bit_map,
6720 OopTaskQueue* work_queue,
6721 CMSMarkStack* overflow_stack,
6722 HeapWord* finger,
6723 HeapWord** global_finger_addr,
6724 ParMarkFromRootsClosure* parent) :
6725 MetadataAwareOopClosure(collector->ref_processor()),
6726 _collector(collector),
6727 _whole_span(collector->_span),
6728 _span(span),
6729 _bit_map(bit_map),
6730 _work_queue(work_queue),
6731 _overflow_stack(overflow_stack),
6732 _finger(finger),
6733 _global_finger_addr(global_finger_addr),
6734 _parent(parent)
6735 { }
6736
6737 // Assumes thread-safe access by callers, who are
6738 // responsible for mutual exclusion.
6739 void CMSCollector::lower_restart_addr(HeapWord* low) {
6740 assert(_span.contains(low), "Out of bounds addr");
6741 if (_restart_addr == NULL) {
6742 _restart_addr = low;
6743 } else {
6744 _restart_addr = MIN2(_restart_addr, low);
6745 }
6746 }
6747
6748 // Upon stack overflow, we discard (part of) the stack,
6749 // remembering the least address amongst those discarded
6750 // in CMSCollector's _restart_address.
6751 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6752 // Remember the least grey address discarded
6753 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6754 _collector->lower_restart_addr(ra);
6755 _markStack->reset(); // discard stack contents
6756 _markStack->expand(); // expand the stack if possible
6757 }
6758
6759 // Upon stack overflow, we discard (part of) the stack,
6760 // remembering the least address amongst those discarded
6761 // in CMSCollector's _restart_address.
6762 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6763 // We need to do this under a mutex to prevent other
6764 // workers from interfering with the work done below.
6765 MutexLockerEx ml(_overflow_stack->par_lock(),
6766 Mutex::_no_safepoint_check_flag);
6767 // Remember the least grey address discarded
6768 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
6769 _collector->lower_restart_addr(ra);
6770 _overflow_stack->reset(); // discard stack contents
6771 _overflow_stack->expand(); // expand the stack if possible
6772 }
6773
6774 void PushOrMarkClosure::do_oop(oop obj) {
6775 // Ignore mark word because we are running concurrent with mutators.
6776 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6777 HeapWord* addr = (HeapWord*)obj;
6778 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
6779 // Oop lies in _span and isn't yet grey or black
6780 _bitMap->mark(addr); // now grey
6781 if (addr < _finger) {
6782 // the bit map iteration has already either passed, or
6783 // sampled, this bit in the bit map; we'll need to
6784 // use the marking stack to scan this oop's oops.
6785 bool simulate_overflow = false;
6786 NOT_PRODUCT(
6787 if (CMSMarkStackOverflowALot &&
6788 _collector->simulate_overflow()) {
6789 // simulate a stack overflow
6790 simulate_overflow = true;
6791 }
6792 )
6793 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
6794 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity());
6795 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
6796 handle_stack_overflow(addr);
6797 }
6798 }
6799 // anything including and to the right of _finger
6800 // will be scanned as we iterate over the remainder of the
6801 // bit map
6802 do_yield_check();
6803 }
6804 }
6805
6806 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
6807 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
6808
6809 void ParPushOrMarkClosure::do_oop(oop obj) {
6810 // Ignore mark word because we are running concurrent with mutators.
6811 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6812 HeapWord* addr = (HeapWord*)obj;
6813 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
6814 // Oop lies in _span and isn't yet grey or black
6815 // We read the global_finger (volatile read) strictly after marking oop
6816 bool res = _bit_map->par_mark(addr); // now grey
6817 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
6818 // Should we push this marked oop on our stack?
6819 // -- if someone else marked it, nothing to do
6820 // -- if target oop is above global finger nothing to do
6821 // -- if target oop is in chunk and above local finger
6822 // then nothing to do
6823 // -- else push on work queue
6824 if ( !res // someone else marked it, they will deal with it
6825 || (addr >= *gfa) // will be scanned in a later task
6826 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
6827 return;
6828 }
6829 // the bit map iteration has already either passed, or
6830 // sampled, this bit in the bit map; we'll need to
6831 // use the marking stack to scan this oop's oops.
6832 bool simulate_overflow = false;
6833 NOT_PRODUCT(
6834 if (CMSMarkStackOverflowALot &&
6835 _collector->simulate_overflow()) {
6836 // simulate a stack overflow
6837 simulate_overflow = true;
6838 }
6839 )
6840 if (simulate_overflow ||
6841 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
6842 // stack overflow
6843 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
6844 // We cannot assert that the overflow stack is full because
6845 // it may have been emptied since.
6846 assert(simulate_overflow ||
6847 _work_queue->size() == _work_queue->max_elems(),
6848 "Else push should have succeeded");
6849 handle_stack_overflow(addr);
6850 }
6851 do_yield_check();
6852 }
6853 }
6854
6855 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); }
6856 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); }
6857
6858 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
6859 MemRegion span,
6860 ReferenceProcessor* rp,
6861 CMSBitMap* bit_map,
6862 CMSBitMap* mod_union_table,
6863 CMSMarkStack* mark_stack,
6864 bool concurrent_precleaning):
6865 MetadataAwareOopClosure(rp),
6866 _collector(collector),
6867 _span(span),
6868 _bit_map(bit_map),
6869 _mod_union_table(mod_union_table),
6870 _mark_stack(mark_stack),
6871 _concurrent_precleaning(concurrent_precleaning)
6872 {
6873 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL");
6874 }
6875
6876 // Grey object rescan during pre-cleaning and second checkpoint phases --
6877 // the non-parallel version (the parallel version appears further below.)
6878 void PushAndMarkClosure::do_oop(oop obj) {
6879 // Ignore mark word verification. If during concurrent precleaning,
6880 // the object monitor may be locked. If during the checkpoint
6881 // phases, the object may already have been reached by a different
6882 // path and may be at the end of the global overflow list (so
6883 // the mark word may be NULL).
6884 assert(obj->is_oop_or_null(true /* ignore mark word */),
6885 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6886 HeapWord* addr = (HeapWord*)obj;
6887 // Check if oop points into the CMS generation
6888 // and is not marked
6889 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6890 // a white object ...
6891 _bit_map->mark(addr); // ... now grey
6892 // push on the marking stack (grey set)
6893 bool simulate_overflow = false;
6894 NOT_PRODUCT(
6895 if (CMSMarkStackOverflowALot &&
6896 _collector->simulate_overflow()) {
6897 // simulate a stack overflow
6898 simulate_overflow = true;
6899 }
6900 )
6901 if (simulate_overflow || !_mark_stack->push(obj)) {
6902 if (_concurrent_precleaning) {
6903 // During precleaning we can just dirty the appropriate card(s)
6904 // in the mod union table, thus ensuring that the object remains
6905 // in the grey set and continue. In the case of object arrays
6906 // we need to dirty all of the cards that the object spans,
6907 // since the rescan of object arrays will be limited to the
6908 // dirty cards.
6909 // Note that no one can be interfering with us in this action
6910 // of dirtying the mod union table, so no locking or atomics
6911 // are required.
6912 if (obj->is_objArray()) {
6913 size_t sz = obj->size();
6914 HeapWord* end_card_addr = (HeapWord*)round_to(
6915 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6916 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6917 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6918 _mod_union_table->mark_range(redirty_range);
6919 } else {
6920 _mod_union_table->mark(addr);
6921 }
6922 _collector->_ser_pmc_preclean_ovflw++;
6923 } else {
6924 // During the remark phase, we need to remember this oop
6925 // in the overflow list.
6926 _collector->push_on_overflow_list(obj);
6927 _collector->_ser_pmc_remark_ovflw++;
6928 }
6929 }
6930 }
6931 }
6932
6933 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector,
6934 MemRegion span,
6935 ReferenceProcessor* rp,
6936 CMSBitMap* bit_map,
6937 OopTaskQueue* work_queue):
6938 MetadataAwareOopClosure(rp),
6939 _collector(collector),
6940 _span(span),
6941 _bit_map(bit_map),
6942 _work_queue(work_queue)
6943 {
6944 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL");
6945 }
6946
6947 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
6948 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
6949
6950 // Grey object rescan during second checkpoint phase --
6951 // the parallel version.
6952 void ParPushAndMarkClosure::do_oop(oop obj) {
6953 // In the assert below, we ignore the mark word because
6954 // this oop may point to an already visited object that is
6955 // on the overflow stack (in which case the mark word has
6956 // been hijacked for chaining into the overflow stack --
6957 // if this is the last object in the overflow stack then
6958 // its mark word will be NULL). Because this object may
6959 // have been subsequently popped off the global overflow
6960 // stack, and the mark word possibly restored to the prototypical
6961 // value, by the time we get to examined this failing assert in
6962 // the debugger, is_oop_or_null(false) may subsequently start
6963 // to hold.
6964 assert(obj->is_oop_or_null(true),
6965 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6966 HeapWord* addr = (HeapWord*)obj;
6967 // Check if oop points into the CMS generation
6968 // and is not marked
6969 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6970 // a white object ...
6971 // If we manage to "claim" the object, by being the
6972 // first thread to mark it, then we push it on our
6973 // marking stack
6974 if (_bit_map->par_mark(addr)) { // ... now grey
6975 // push on work queue (grey set)
6976 bool simulate_overflow = false;
6977 NOT_PRODUCT(
6978 if (CMSMarkStackOverflowALot &&
6979 _collector->par_simulate_overflow()) {
6980 // simulate a stack overflow
6981 simulate_overflow = true;
6982 }
6983 )
6984 if (simulate_overflow || !_work_queue->push(obj)) {
6985 _collector->par_push_on_overflow_list(obj);
6986 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
6987 }
6988 } // Else, some other thread got there first
6989 }
6990 }
6991
6992 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); }
6993 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); }
6994
6995 void CMSPrecleanRefsYieldClosure::do_yield_work() {
6996 Mutex* bml = _collector->bitMapLock();
6997 assert_lock_strong(bml);
6998 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6999 "CMS thread should hold CMS token");
7000
7001 bml->unlock();
7002 ConcurrentMarkSweepThread::desynchronize(true);
7003
7004 _collector->stopTimer();
7005 _collector->incrementYields();
7006
7007 // See the comment in coordinator_yield()
7008 for (unsigned i = 0; i < CMSYieldSleepCount &&
7009 ConcurrentMarkSweepThread::should_yield() &&
7010 !CMSCollector::foregroundGCIsActive(); ++i) {
7011 os::sleep(Thread::current(), 1, false);
7012 }
7013
7014 ConcurrentMarkSweepThread::synchronize(true);
7015 bml->lock();
7016
7017 _collector->startTimer();
7018 }
7019
7020 bool CMSPrecleanRefsYieldClosure::should_return() {
7021 if (ConcurrentMarkSweepThread::should_yield()) {
7022 do_yield_work();
7023 }
7024 return _collector->foregroundGCIsActive();
7025 }
7026
7027 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7028 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7029 "mr should be aligned to start at a card boundary");
7030 // We'd like to assert:
7031 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7032 // "mr should be a range of cards");
7033 // However, that would be too strong in one case -- the last
7034 // partition ends at _unallocated_block which, in general, can be
7035 // an arbitrary boundary, not necessarily card aligned.
7036 _num_dirty_cards += mr.word_size()/CardTableModRefBS::card_size_in_words;
7037 _space->object_iterate_mem(mr, &_scan_cl);
7038 }
7039
7040 SweepClosure::SweepClosure(CMSCollector* collector,
7041 ConcurrentMarkSweepGeneration* g,
7042 CMSBitMap* bitMap, bool should_yield) :
7043 _collector(collector),
7044 _g(g),
7045 _sp(g->cmsSpace()),
7046 _limit(_sp->sweep_limit()),
7047 _freelistLock(_sp->freelistLock()),
7048 _bitMap(bitMap),
7049 _yield(should_yield),
7050 _inFreeRange(false), // No free range at beginning of sweep
7051 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7052 _lastFreeRangeCoalesced(false),
7053 _freeFinger(g->used_region().start())
7054 {
7055 NOT_PRODUCT(
7056 _numObjectsFreed = 0;
7057 _numWordsFreed = 0;
7058 _numObjectsLive = 0;
7059 _numWordsLive = 0;
7060 _numObjectsAlreadyFree = 0;
7061 _numWordsAlreadyFree = 0;
7062 _last_fc = NULL;
7063
7064 _sp->initializeIndexedFreeListArrayReturnedBytes();
7065 _sp->dictionary()->initialize_dict_returned_bytes();
7066 )
7067 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7068 "sweep _limit out of bounds");
7069 log_develop_trace(gc, sweep)("====================");
7070 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit));
7071 }
7072
7073 void SweepClosure::print_on(outputStream* st) const {
7074 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7075 p2i(_sp->bottom()), p2i(_sp->end()));
7076 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7077 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7078 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7079 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7080 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7081 }
7082
7083 #ifndef PRODUCT
7084 // Assertion checking only: no useful work in product mode --
7085 // however, if any of the flags below become product flags,
7086 // you may need to review this code to see if it needs to be
7087 // enabled in product mode.
7088 SweepClosure::~SweepClosure() {
7089 assert_lock_strong(_freelistLock);
7090 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7091 "sweep _limit out of bounds");
7092 if (inFreeRange()) {
7093 Log(gc, sweep) log;
7094 log.error("inFreeRange() should have been reset; dumping state of SweepClosure");
7095 ResourceMark rm;
7096 print_on(log.error_stream());
7097 ShouldNotReachHere();
7098 }
7099
7100 if (log_is_enabled(Debug, gc, sweep)) {
7101 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7102 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7103 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7104 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7105 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord);
7106 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes);
7107 }
7108
7109 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) {
7110 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7111 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7112 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7113 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes",
7114 returned_bytes, indexListReturnedBytes, dict_returned_bytes);
7115 }
7116 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit));
7117 log_develop_trace(gc, sweep)("================");
7118 }
7119 #endif // PRODUCT
7120
7121 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7122 bool freeRangeInFreeLists) {
7123 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)",
7124 p2i(freeFinger), freeRangeInFreeLists);
7125 assert(!inFreeRange(), "Trampling existing free range");
7126 set_inFreeRange(true);
7127 set_lastFreeRangeCoalesced(false);
7128
7129 set_freeFinger(freeFinger);
7130 set_freeRangeInFreeLists(freeRangeInFreeLists);
7131 if (CMSTestInFreeList) {
7132 if (freeRangeInFreeLists) {
7133 FreeChunk* fc = (FreeChunk*) freeFinger;
7134 assert(fc->is_free(), "A chunk on the free list should be free.");
7135 assert(fc->size() > 0, "Free range should have a size");
7136 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7137 }
7138 }
7139 }
7140
7141 // Note that the sweeper runs concurrently with mutators. Thus,
7142 // it is possible for direct allocation in this generation to happen
7143 // in the middle of the sweep. Note that the sweeper also coalesces
7144 // contiguous free blocks. Thus, unless the sweeper and the allocator
7145 // synchronize appropriately freshly allocated blocks may get swept up.
7146 // This is accomplished by the sweeper locking the free lists while
7147 // it is sweeping. Thus blocks that are determined to be free are
7148 // indeed free. There is however one additional complication:
7149 // blocks that have been allocated since the final checkpoint and
7150 // mark, will not have been marked and so would be treated as
7151 // unreachable and swept up. To prevent this, the allocator marks
7152 // the bit map when allocating during the sweep phase. This leads,
7153 // however, to a further complication -- objects may have been allocated
7154 // but not yet initialized -- in the sense that the header isn't yet
7155 // installed. The sweeper can not then determine the size of the block
7156 // in order to skip over it. To deal with this case, we use a technique
7157 // (due to Printezis) to encode such uninitialized block sizes in the
7158 // bit map. Since the bit map uses a bit per every HeapWord, but the
7159 // CMS generation has a minimum object size of 3 HeapWords, it follows
7160 // that "normal marks" won't be adjacent in the bit map (there will
7161 // always be at least two 0 bits between successive 1 bits). We make use
7162 // of these "unused" bits to represent uninitialized blocks -- the bit
7163 // corresponding to the start of the uninitialized object and the next
7164 // bit are both set. Finally, a 1 bit marks the end of the object that
7165 // started with the two consecutive 1 bits to indicate its potentially
7166 // uninitialized state.
7167
7168 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7169 FreeChunk* fc = (FreeChunk*)addr;
7170 size_t res;
7171
7172 // Check if we are done sweeping. Below we check "addr >= _limit" rather
7173 // than "addr == _limit" because although _limit was a block boundary when
7174 // we started the sweep, it may no longer be one because heap expansion
7175 // may have caused us to coalesce the block ending at the address _limit
7176 // with a newly expanded chunk (this happens when _limit was set to the
7177 // previous _end of the space), so we may have stepped past _limit:
7178 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7179 if (addr >= _limit) { // we have swept up to or past the limit: finish up
7180 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7181 "sweep _limit out of bounds");
7182 assert(addr < _sp->end(), "addr out of bounds");
7183 // Flush any free range we might be holding as a single
7184 // coalesced chunk to the appropriate free list.
7185 if (inFreeRange()) {
7186 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7187 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger()));
7188 flush_cur_free_chunk(freeFinger(),
7189 pointer_delta(addr, freeFinger()));
7190 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]",
7191 p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7192 lastFreeRangeCoalesced() ? 1 : 0);
7193 }
7194
7195 // help the iterator loop finish
7196 return pointer_delta(_sp->end(), addr);
7197 }
7198
7199 assert(addr < _limit, "sweep invariant");
7200 // check if we should yield
7201 do_yield_check(addr);
7202 if (fc->is_free()) {
7203 // Chunk that is already free
7204 res = fc->size();
7205 do_already_free_chunk(fc);
7206 debug_only(_sp->verifyFreeLists());
7207 // If we flush the chunk at hand in lookahead_and_flush()
7208 // and it's coalesced with a preceding chunk, then the
7209 // process of "mangling" the payload of the coalesced block
7210 // will cause erasure of the size information from the
7211 // (erstwhile) header of all the coalesced blocks but the
7212 // first, so the first disjunct in the assert will not hold
7213 // in that specific case (in which case the second disjunct
7214 // will hold).
7215 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7216 "Otherwise the size info doesn't change at this step");
7217 NOT_PRODUCT(
7218 _numObjectsAlreadyFree++;
7219 _numWordsAlreadyFree += res;
7220 )
7221 NOT_PRODUCT(_last_fc = fc;)
7222 } else if (!_bitMap->isMarked(addr)) {
7223 // Chunk is fresh garbage
7224 res = do_garbage_chunk(fc);
7225 debug_only(_sp->verifyFreeLists());
7226 NOT_PRODUCT(
7227 _numObjectsFreed++;
7228 _numWordsFreed += res;
7229 )
7230 } else {
7231 // Chunk that is alive.
7232 res = do_live_chunk(fc);
7233 debug_only(_sp->verifyFreeLists());
7234 NOT_PRODUCT(
7235 _numObjectsLive++;
7236 _numWordsLive += res;
7237 )
7238 }
7239 return res;
7240 }
7241
7242 // For the smart allocation, record following
7243 // split deaths - a free chunk is removed from its free list because
7244 // it is being split into two or more chunks.
7245 // split birth - a free chunk is being added to its free list because
7246 // a larger free chunk has been split and resulted in this free chunk.
7247 // coal death - a free chunk is being removed from its free list because
7248 // it is being coalesced into a large free chunk.
7249 // coal birth - a free chunk is being added to its free list because
7250 // it was created when two or more free chunks where coalesced into
7251 // this free chunk.
7252 //
7253 // These statistics are used to determine the desired number of free
7254 // chunks of a given size. The desired number is chosen to be relative
7255 // to the end of a CMS sweep. The desired number at the end of a sweep
7256 // is the
7257 // count-at-end-of-previous-sweep (an amount that was enough)
7258 // - count-at-beginning-of-current-sweep (the excess)
7259 // + split-births (gains in this size during interval)
7260 // - split-deaths (demands on this size during interval)
7261 // where the interval is from the end of one sweep to the end of the
7262 // next.
7263 //
7264 // When sweeping the sweeper maintains an accumulated chunk which is
7265 // the chunk that is made up of chunks that have been coalesced. That
7266 // will be termed the left-hand chunk. A new chunk of garbage that
7267 // is being considered for coalescing will be referred to as the
7268 // right-hand chunk.
7269 //
7270 // When making a decision on whether to coalesce a right-hand chunk with
7271 // the current left-hand chunk, the current count vs. the desired count
7272 // of the left-hand chunk is considered. Also if the right-hand chunk
7273 // is near the large chunk at the end of the heap (see
7274 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7275 // left-hand chunk is coalesced.
7276 //
7277 // When making a decision about whether to split a chunk, the desired count
7278 // vs. the current count of the candidate to be split is also considered.
7279 // If the candidate is underpopulated (currently fewer chunks than desired)
7280 // a chunk of an overpopulated (currently more chunks than desired) size may
7281 // be chosen. The "hint" associated with a free list, if non-null, points
7282 // to a free list which may be overpopulated.
7283 //
7284
7285 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7286 const size_t size = fc->size();
7287 // Chunks that cannot be coalesced are not in the
7288 // free lists.
7289 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7290 assert(_sp->verify_chunk_in_free_list(fc),
7291 "free chunk should be in free lists");
7292 }
7293 // a chunk that is already free, should not have been
7294 // marked in the bit map
7295 HeapWord* const addr = (HeapWord*) fc;
7296 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7297 // Verify that the bit map has no bits marked between
7298 // addr and purported end of this block.
7299 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7300
7301 // Some chunks cannot be coalesced under any circumstances.
7302 // See the definition of cantCoalesce().
7303 if (!fc->cantCoalesce()) {
7304 // This chunk can potentially be coalesced.
7305 // All the work is done in
7306 do_post_free_or_garbage_chunk(fc, size);
7307 // Note that if the chunk is not coalescable (the else arm
7308 // below), we unconditionally flush, without needing to do
7309 // a "lookahead," as we do below.
7310 if (inFreeRange()) lookahead_and_flush(fc, size);
7311 } else {
7312 // Code path common to both original and adaptive free lists.
7313
7314 // cant coalesce with previous block; this should be treated
7315 // as the end of a free run if any
7316 if (inFreeRange()) {
7317 // we kicked some butt; time to pick up the garbage
7318 assert(freeFinger() < addr, "freeFinger points too high");
7319 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7320 }
7321 // else, nothing to do, just continue
7322 }
7323 }
7324
7325 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7326 // This is a chunk of garbage. It is not in any free list.
7327 // Add it to a free list or let it possibly be coalesced into
7328 // a larger chunk.
7329 HeapWord* const addr = (HeapWord*) fc;
7330 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7331
7332 // Verify that the bit map has no bits marked between
7333 // addr and purported end of just dead object.
7334 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7335 do_post_free_or_garbage_chunk(fc, size);
7336
7337 assert(_limit >= addr + size,
7338 "A freshly garbage chunk can't possibly straddle over _limit");
7339 if (inFreeRange()) lookahead_and_flush(fc, size);
7340 return size;
7341 }
7342
7343 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7344 HeapWord* addr = (HeapWord*) fc;
7345 // The sweeper has just found a live object. Return any accumulated
7346 // left hand chunk to the free lists.
7347 if (inFreeRange()) {
7348 assert(freeFinger() < addr, "freeFinger points too high");
7349 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7350 }
7351
7352 // This object is live: we'd normally expect this to be
7353 // an oop, and like to assert the following:
7354 // assert(oop(addr)->is_oop(), "live block should be an oop");
7355 // However, as we commented above, this may be an object whose
7356 // header hasn't yet been initialized.
7357 size_t size;
7358 assert(_bitMap->isMarked(addr), "Tautology for this control point");
7359 if (_bitMap->isMarked(addr + 1)) {
7360 // Determine the size from the bit map, rather than trying to
7361 // compute it from the object header.
7362 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7363 size = pointer_delta(nextOneAddr + 1, addr);
7364 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7365 "alignment problem");
7366
7367 #ifdef ASSERT
7368 if (oop(addr)->klass_or_null() != NULL) {
7369 // Ignore mark word because we are running concurrent with mutators
7370 assert(oop(addr)->is_oop(true), "live block should be an oop");
7371 assert(size ==
7372 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7373 "P-mark and computed size do not agree");
7374 }
7375 #endif
7376
7377 } else {
7378 // This should be an initialized object that's alive.
7379 assert(oop(addr)->klass_or_null() != NULL,
7380 "Should be an initialized object");
7381 // Ignore mark word because we are running concurrent with mutators
7382 assert(oop(addr)->is_oop(true), "live block should be an oop");
7383 // Verify that the bit map has no bits marked between
7384 // addr and purported end of this block.
7385 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7386 assert(size >= 3, "Necessary for Printezis marks to work");
7387 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7388 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7389 }
7390 return size;
7391 }
7392
7393 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7394 size_t chunkSize) {
7395 // do_post_free_or_garbage_chunk() should only be called in the case
7396 // of the adaptive free list allocator.
7397 const bool fcInFreeLists = fc->is_free();
7398 assert((HeapWord*)fc <= _limit, "sweep invariant");
7399 if (CMSTestInFreeList && fcInFreeLists) {
7400 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7401 }
7402
7403 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7404
7405 HeapWord* const fc_addr = (HeapWord*) fc;
7406
7407 bool coalesce = false;
7408 const size_t left = pointer_delta(fc_addr, freeFinger());
7409 const size_t right = chunkSize;
7410 switch (FLSCoalescePolicy) {
7411 // numeric value forms a coalition aggressiveness metric
7412 case 0: { // never coalesce
7413 coalesce = false;
7414 break;
7415 }
7416 case 1: { // coalesce if left & right chunks on overpopulated lists
7417 coalesce = _sp->coalOverPopulated(left) &&
7418 _sp->coalOverPopulated(right);
7419 break;
7420 }
7421 case 2: { // coalesce if left chunk on overpopulated list (default)
7422 coalesce = _sp->coalOverPopulated(left);
7423 break;
7424 }
7425 case 3: { // coalesce if left OR right chunk on overpopulated list
7426 coalesce = _sp->coalOverPopulated(left) ||
7427 _sp->coalOverPopulated(right);
7428 break;
7429 }
7430 case 4: { // always coalesce
7431 coalesce = true;
7432 break;
7433 }
7434 default:
7435 ShouldNotReachHere();
7436 }
7437
7438 // Should the current free range be coalesced?
7439 // If the chunk is in a free range and either we decided to coalesce above
7440 // or the chunk is near the large block at the end of the heap
7441 // (isNearLargestChunk() returns true), then coalesce this chunk.
7442 const bool doCoalesce = inFreeRange()
7443 && (coalesce || _g->isNearLargestChunk(fc_addr));
7444 if (doCoalesce) {
7445 // Coalesce the current free range on the left with the new
7446 // chunk on the right. If either is on a free list,
7447 // it must be removed from the list and stashed in the closure.
7448 if (freeRangeInFreeLists()) {
7449 FreeChunk* const ffc = (FreeChunk*)freeFinger();
7450 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7451 "Size of free range is inconsistent with chunk size.");
7452 if (CMSTestInFreeList) {
7453 assert(_sp->verify_chunk_in_free_list(ffc),
7454 "Chunk is not in free lists");
7455 }
7456 _sp->coalDeath(ffc->size());
7457 _sp->removeFreeChunkFromFreeLists(ffc);
7458 set_freeRangeInFreeLists(false);
7459 }
7460 if (fcInFreeLists) {
7461 _sp->coalDeath(chunkSize);
7462 assert(fc->size() == chunkSize,
7463 "The chunk has the wrong size or is not in the free lists");
7464 _sp->removeFreeChunkFromFreeLists(fc);
7465 }
7466 set_lastFreeRangeCoalesced(true);
7467 print_free_block_coalesced(fc);
7468 } else { // not in a free range and/or should not coalesce
7469 // Return the current free range and start a new one.
7470 if (inFreeRange()) {
7471 // In a free range but cannot coalesce with the right hand chunk.
7472 // Put the current free range into the free lists.
7473 flush_cur_free_chunk(freeFinger(),
7474 pointer_delta(fc_addr, freeFinger()));
7475 }
7476 // Set up for new free range. Pass along whether the right hand
7477 // chunk is in the free lists.
7478 initialize_free_range((HeapWord*)fc, fcInFreeLists);
7479 }
7480 }
7481
7482 // Lookahead flush:
7483 // If we are tracking a free range, and this is the last chunk that
7484 // we'll look at because its end crosses past _limit, we'll preemptively
7485 // flush it along with any free range we may be holding on to. Note that
7486 // this can be the case only for an already free or freshly garbage
7487 // chunk. If this block is an object, it can never straddle
7488 // over _limit. The "straddling" occurs when _limit is set at
7489 // the previous end of the space when this cycle started, and
7490 // a subsequent heap expansion caused the previously co-terminal
7491 // free block to be coalesced with the newly expanded portion,
7492 // thus rendering _limit a non-block-boundary making it dangerous
7493 // for the sweeper to step over and examine.
7494 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7495 assert(inFreeRange(), "Should only be called if currently in a free range.");
7496 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7497 assert(_sp->used_region().contains(eob - 1),
7498 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7499 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7500 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7501 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size);
7502 if (eob >= _limit) {
7503 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7504 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block "
7505 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7506 "[" PTR_FORMAT "," PTR_FORMAT ")",
7507 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7508 // Return the storage we are tracking back into the free lists.
7509 log_develop_trace(gc, sweep)("Flushing ... ");
7510 assert(freeFinger() < eob, "Error");
7511 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7512 }
7513 }
7514
7515 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7516 assert(inFreeRange(), "Should only be called if currently in a free range.");
7517 assert(size > 0,
7518 "A zero sized chunk cannot be added to the free lists.");
7519 if (!freeRangeInFreeLists()) {
7520 if (CMSTestInFreeList) {
7521 FreeChunk* fc = (FreeChunk*) chunk;
7522 fc->set_size(size);
7523 assert(!_sp->verify_chunk_in_free_list(fc),
7524 "chunk should not be in free lists yet");
7525 }
7526 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size);
7527 // A new free range is going to be starting. The current
7528 // free range has not been added to the free lists yet or
7529 // was removed so add it back.
7530 // If the current free range was coalesced, then the death
7531 // of the free range was recorded. Record a birth now.
7532 if (lastFreeRangeCoalesced()) {
7533 _sp->coalBirth(size);
7534 }
7535 _sp->addChunkAndRepairOffsetTable(chunk, size,
7536 lastFreeRangeCoalesced());
7537 } else {
7538 log_develop_trace(gc, sweep)("Already in free list: nothing to flush");
7539 }
7540 set_inFreeRange(false);
7541 set_freeRangeInFreeLists(false);
7542 }
7543
7544 // We take a break if we've been at this for a while,
7545 // so as to avoid monopolizing the locks involved.
7546 void SweepClosure::do_yield_work(HeapWord* addr) {
7547 // Return current free chunk being used for coalescing (if any)
7548 // to the appropriate freelist. After yielding, the next
7549 // free block encountered will start a coalescing range of
7550 // free blocks. If the next free block is adjacent to the
7551 // chunk just flushed, they will need to wait for the next
7552 // sweep to be coalesced.
7553 if (inFreeRange()) {
7554 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7555 }
7556
7557 // First give up the locks, then yield, then re-lock.
7558 // We should probably use a constructor/destructor idiom to
7559 // do this unlock/lock or modify the MutexUnlocker class to
7560 // serve our purpose. XXX
7561 assert_lock_strong(_bitMap->lock());
7562 assert_lock_strong(_freelistLock);
7563 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7564 "CMS thread should hold CMS token");
7565 _bitMap->lock()->unlock();
7566 _freelistLock->unlock();
7567 ConcurrentMarkSweepThread::desynchronize(true);
7568 _collector->stopTimer();
7569 _collector->incrementYields();
7570
7571 // See the comment in coordinator_yield()
7572 for (unsigned i = 0; i < CMSYieldSleepCount &&
7573 ConcurrentMarkSweepThread::should_yield() &&
7574 !CMSCollector::foregroundGCIsActive(); ++i) {
7575 os::sleep(Thread::current(), 1, false);
7576 }
7577
7578 ConcurrentMarkSweepThread::synchronize(true);
7579 _freelistLock->lock();
7580 _bitMap->lock()->lock_without_safepoint_check();
7581 _collector->startTimer();
7582 }
7583
7584 #ifndef PRODUCT
7585 // This is actually very useful in a product build if it can
7586 // be called from the debugger. Compile it into the product
7587 // as needed.
7588 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7589 return debug_cms_space->verify_chunk_in_free_list(fc);
7590 }
7591 #endif
7592
7593 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7594 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7595 p2i(fc), fc->size());
7596 }
7597
7598 // CMSIsAliveClosure
7599 bool CMSIsAliveClosure::do_object_b(oop obj) {
7600 HeapWord* addr = (HeapWord*)obj;
7601 return addr != NULL &&
7602 (!_span.contains(addr) || _bit_map->isMarked(addr));
7603 }
7604
7605
7606 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7607 MemRegion span,
7608 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7609 bool cpc):
7610 _collector(collector),
7611 _span(span),
7612 _bit_map(bit_map),
7613 _mark_stack(mark_stack),
7614 _concurrent_precleaning(cpc) {
7615 assert(!_span.is_empty(), "Empty span could spell trouble");
7616 }
7617
7618
7619 // CMSKeepAliveClosure: the serial version
7620 void CMSKeepAliveClosure::do_oop(oop obj) {
7621 HeapWord* addr = (HeapWord*)obj;
7622 if (_span.contains(addr) &&
7623 !_bit_map->isMarked(addr)) {
7624 _bit_map->mark(addr);
7625 bool simulate_overflow = false;
7626 NOT_PRODUCT(
7627 if (CMSMarkStackOverflowALot &&
7628 _collector->simulate_overflow()) {
7629 // simulate a stack overflow
7630 simulate_overflow = true;
7631 }
7632 )
7633 if (simulate_overflow || !_mark_stack->push(obj)) {
7634 if (_concurrent_precleaning) {
7635 // We dirty the overflown object and let the remark
7636 // phase deal with it.
7637 assert(_collector->overflow_list_is_empty(), "Error");
7638 // In the case of object arrays, we need to dirty all of
7639 // the cards that the object spans. No locking or atomics
7640 // are needed since no one else can be mutating the mod union
7641 // table.
7642 if (obj->is_objArray()) {
7643 size_t sz = obj->size();
7644 HeapWord* end_card_addr =
7645 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
7646 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7647 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7648 _collector->_modUnionTable.mark_range(redirty_range);
7649 } else {
7650 _collector->_modUnionTable.mark(addr);
7651 }
7652 _collector->_ser_kac_preclean_ovflw++;
7653 } else {
7654 _collector->push_on_overflow_list(obj);
7655 _collector->_ser_kac_ovflw++;
7656 }
7657 }
7658 }
7659 }
7660
7661 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7662 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7663
7664 // CMSParKeepAliveClosure: a parallel version of the above.
7665 // The work queues are private to each closure (thread),
7666 // but (may be) available for stealing by other threads.
7667 void CMSParKeepAliveClosure::do_oop(oop obj) {
7668 HeapWord* addr = (HeapWord*)obj;
7669 if (_span.contains(addr) &&
7670 !_bit_map->isMarked(addr)) {
7671 // In general, during recursive tracing, several threads
7672 // may be concurrently getting here; the first one to
7673 // "tag" it, claims it.
7674 if (_bit_map->par_mark(addr)) {
7675 bool res = _work_queue->push(obj);
7676 assert(res, "Low water mark should be much less than capacity");
7677 // Do a recursive trim in the hope that this will keep
7678 // stack usage lower, but leave some oops for potential stealers
7679 trim_queue(_low_water_mark);
7680 } // Else, another thread got there first
7681 }
7682 }
7683
7684 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7685 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7686
7687 void CMSParKeepAliveClosure::trim_queue(uint max) {
7688 while (_work_queue->size() > max) {
7689 oop new_oop;
7690 if (_work_queue->pop_local(new_oop)) {
7691 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
7692 assert(_bit_map->isMarked((HeapWord*)new_oop),
7693 "no white objects on this stack!");
7694 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7695 // iterate over the oops in this oop, marking and pushing
7696 // the ones in CMS heap (i.e. in _span).
7697 new_oop->oop_iterate(&_mark_and_push);
7698 }
7699 }
7700 }
7701
7702 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
7703 CMSCollector* collector,
7704 MemRegion span, CMSBitMap* bit_map,
7705 OopTaskQueue* work_queue):
7706 _collector(collector),
7707 _span(span),
7708 _bit_map(bit_map),
7709 _work_queue(work_queue) { }
7710
7711 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
7712 HeapWord* addr = (HeapWord*)obj;
7713 if (_span.contains(addr) &&
7714 !_bit_map->isMarked(addr)) {
7715 if (_bit_map->par_mark(addr)) {
7716 bool simulate_overflow = false;
7717 NOT_PRODUCT(
7718 if (CMSMarkStackOverflowALot &&
7719 _collector->par_simulate_overflow()) {
7720 // simulate a stack overflow
7721 simulate_overflow = true;
7722 }
7723 )
7724 if (simulate_overflow || !_work_queue->push(obj)) {
7725 _collector->par_push_on_overflow_list(obj);
7726 _collector->_par_kac_ovflw++;
7727 }
7728 } // Else another thread got there already
7729 }
7730 }
7731
7732 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7733 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7734
7735 //////////////////////////////////////////////////////////////////
7736 // CMSExpansionCause /////////////////////////////
7737 //////////////////////////////////////////////////////////////////
7738 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
7739 switch (cause) {
7740 case _no_expansion:
7741 return "No expansion";
7742 case _satisfy_free_ratio:
7743 return "Free ratio";
7744 case _satisfy_promotion:
7745 return "Satisfy promotion";
7746 case _satisfy_allocation:
7747 return "allocation";
7748 case _allocate_par_lab:
7749 return "Par LAB";
7750 case _allocate_par_spooling_space:
7751 return "Par Spooling Space";
7752 case _adaptive_size_policy:
7753 return "Ergonomics";
7754 default:
7755 return "unknown";
7756 }
7757 }
7758
7759 void CMSDrainMarkingStackClosure::do_void() {
7760 // the max number to take from overflow list at a time
7761 const size_t num = _mark_stack->capacity()/4;
7762 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
7763 "Overflow list should be NULL during concurrent phases");
7764 while (!_mark_stack->isEmpty() ||
7765 // if stack is empty, check the overflow list
7766 _collector->take_from_overflow_list(num, _mark_stack)) {
7767 oop obj = _mark_stack->pop();
7768 HeapWord* addr = (HeapWord*)obj;
7769 assert(_span.contains(addr), "Should be within span");
7770 assert(_bit_map->isMarked(addr), "Should be marked");
7771 assert(obj->is_oop(), "Should be an oop");
7772 obj->oop_iterate(_keep_alive);
7773 }
7774 }
7775
7776 void CMSParDrainMarkingStackClosure::do_void() {
7777 // drain queue
7778 trim_queue(0);
7779 }
7780
7781 // Trim our work_queue so its length is below max at return
7782 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
7783 while (_work_queue->size() > max) {
7784 oop new_oop;
7785 if (_work_queue->pop_local(new_oop)) {
7786 assert(new_oop->is_oop(), "Expected an oop");
7787 assert(_bit_map->isMarked((HeapWord*)new_oop),
7788 "no white objects on this stack!");
7789 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7790 // iterate over the oops in this oop, marking and pushing
7791 // the ones in CMS heap (i.e. in _span).
7792 new_oop->oop_iterate(&_mark_and_push);
7793 }
7794 }
7795 }
7796
7797 ////////////////////////////////////////////////////////////////////
7798 // Support for Marking Stack Overflow list handling and related code
7799 ////////////////////////////////////////////////////////////////////
7800 // Much of the following code is similar in shape and spirit to the
7801 // code used in ParNewGC. We should try and share that code
7802 // as much as possible in the future.
7803
7804 #ifndef PRODUCT
7805 // Debugging support for CMSStackOverflowALot
7806
7807 // It's OK to call this multi-threaded; the worst thing
7808 // that can happen is that we'll get a bunch of closely
7809 // spaced simulated overflows, but that's OK, in fact
7810 // probably good as it would exercise the overflow code
7811 // under contention.
7812 bool CMSCollector::simulate_overflow() {
7813 if (_overflow_counter-- <= 0) { // just being defensive
7814 _overflow_counter = CMSMarkStackOverflowInterval;
7815 return true;
7816 } else {
7817 return false;
7818 }
7819 }
7820
7821 bool CMSCollector::par_simulate_overflow() {
7822 return simulate_overflow();
7823 }
7824 #endif
7825
7826 // Single-threaded
7827 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
7828 assert(stack->isEmpty(), "Expected precondition");
7829 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
7830 size_t i = num;
7831 oop cur = _overflow_list;
7832 const markOop proto = markOopDesc::prototype();
7833 NOT_PRODUCT(ssize_t n = 0;)
7834 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
7835 next = oop(cur->mark());
7836 cur->set_mark(proto); // until proven otherwise
7837 assert(cur->is_oop(), "Should be an oop");
7838 bool res = stack->push(cur);
7839 assert(res, "Bit off more than can chew?");
7840 NOT_PRODUCT(n++;)
7841 }
7842 _overflow_list = cur;
7843 #ifndef PRODUCT
7844 assert(_num_par_pushes >= n, "Too many pops?");
7845 _num_par_pushes -=n;
7846 #endif
7847 return !stack->isEmpty();
7848 }
7849
7850 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
7851 // (MT-safe) Get a prefix of at most "num" from the list.
7852 // The overflow list is chained through the mark word of
7853 // each object in the list. We fetch the entire list,
7854 // break off a prefix of the right size and return the
7855 // remainder. If other threads try to take objects from
7856 // the overflow list at that time, they will wait for
7857 // some time to see if data becomes available. If (and
7858 // only if) another thread places one or more object(s)
7859 // on the global list before we have returned the suffix
7860 // to the global list, we will walk down our local list
7861 // to find its end and append the global list to
7862 // our suffix before returning it. This suffix walk can
7863 // prove to be expensive (quadratic in the amount of traffic)
7864 // when there are many objects in the overflow list and
7865 // there is much producer-consumer contention on the list.
7866 // *NOTE*: The overflow list manipulation code here and
7867 // in ParNewGeneration:: are very similar in shape,
7868 // except that in the ParNew case we use the old (from/eden)
7869 // copy of the object to thread the list via its klass word.
7870 // Because of the common code, if you make any changes in
7871 // the code below, please check the ParNew version to see if
7872 // similar changes might be needed.
7873 // CR 6797058 has been filed to consolidate the common code.
7874 bool CMSCollector::par_take_from_overflow_list(size_t num,
7875 OopTaskQueue* work_q,
7876 int no_of_gc_threads) {
7877 assert(work_q->size() == 0, "First empty local work queue");
7878 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
7879 if (_overflow_list == NULL) {
7880 return false;
7881 }
7882 // Grab the entire list; we'll put back a suffix
7883 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
7884 Thread* tid = Thread::current();
7885 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
7886 // set to ParallelGCThreads.
7887 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
7888 size_t sleep_time_millis = MAX2((size_t)1, num/100);
7889 // If the list is busy, we spin for a short while,
7890 // sleeping between attempts to get the list.
7891 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
7892 os::sleep(tid, sleep_time_millis, false);
7893 if (_overflow_list == NULL) {
7894 // Nothing left to take
7895 return false;
7896 } else if (_overflow_list != BUSY) {
7897 // Try and grab the prefix
7898 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
7899 }
7900 }
7901 // If the list was found to be empty, or we spun long
7902 // enough, we give up and return empty-handed. If we leave
7903 // the list in the BUSY state below, it must be the case that
7904 // some other thread holds the overflow list and will set it
7905 // to a non-BUSY state in the future.
7906 if (prefix == NULL || prefix == BUSY) {
7907 // Nothing to take or waited long enough
7908 if (prefix == NULL) {
7909 // Write back the NULL in case we overwrote it with BUSY above
7910 // and it is still the same value.
7911 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
7912 }
7913 return false;
7914 }
7915 assert(prefix != NULL && prefix != BUSY, "Error");
7916 size_t i = num;
7917 oop cur = prefix;
7918 // Walk down the first "num" objects, unless we reach the end.
7919 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
7920 if (cur->mark() == NULL) {
7921 // We have "num" or fewer elements in the list, so there
7922 // is nothing to return to the global list.
7923 // Write back the NULL in lieu of the BUSY we wrote
7924 // above, if it is still the same value.
7925 if (_overflow_list == BUSY) {
7926 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
7927 }
7928 } else {
7929 // Chop off the suffix and return it to the global list.
7930 assert(cur->mark() != BUSY, "Error");
7931 oop suffix_head = cur->mark(); // suffix will be put back on global list
7932 cur->set_mark(NULL); // break off suffix
7933 // It's possible that the list is still in the empty(busy) state
7934 // we left it in a short while ago; in that case we may be
7935 // able to place back the suffix without incurring the cost
7936 // of a walk down the list.
7937 oop observed_overflow_list = _overflow_list;
7938 oop cur_overflow_list = observed_overflow_list;
7939 bool attached = false;
7940 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
7941 observed_overflow_list =
7942 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
7943 if (cur_overflow_list == observed_overflow_list) {
7944 attached = true;
7945 break;
7946 } else cur_overflow_list = observed_overflow_list;
7947 }
7948 if (!attached) {
7949 // Too bad, someone else sneaked in (at least) an element; we'll need
7950 // to do a splice. Find tail of suffix so we can prepend suffix to global
7951 // list.
7952 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
7953 oop suffix_tail = cur;
7954 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
7955 "Tautology");
7956 observed_overflow_list = _overflow_list;
7957 do {
7958 cur_overflow_list = observed_overflow_list;
7959 if (cur_overflow_list != BUSY) {
7960 // Do the splice ...
7961 suffix_tail->set_mark(markOop(cur_overflow_list));
7962 } else { // cur_overflow_list == BUSY
7963 suffix_tail->set_mark(NULL);
7964 }
7965 // ... and try to place spliced list back on overflow_list ...
7966 observed_overflow_list =
7967 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
7968 } while (cur_overflow_list != observed_overflow_list);
7969 // ... until we have succeeded in doing so.
7970 }
7971 }
7972
7973 // Push the prefix elements on work_q
7974 assert(prefix != NULL, "control point invariant");
7975 const markOop proto = markOopDesc::prototype();
7976 oop next;
7977 NOT_PRODUCT(ssize_t n = 0;)
7978 for (cur = prefix; cur != NULL; cur = next) {
7979 next = oop(cur->mark());
7980 cur->set_mark(proto); // until proven otherwise
7981 assert(cur->is_oop(), "Should be an oop");
7982 bool res = work_q->push(cur);
7983 assert(res, "Bit off more than we can chew?");
7984 NOT_PRODUCT(n++;)
7985 }
7986 #ifndef PRODUCT
7987 assert(_num_par_pushes >= n, "Too many pops?");
7988 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
7989 #endif
7990 return true;
7991 }
7992
7993 // Single-threaded
7994 void CMSCollector::push_on_overflow_list(oop p) {
7995 NOT_PRODUCT(_num_par_pushes++;)
7996 assert(p->is_oop(), "Not an oop");
7997 preserve_mark_if_necessary(p);
7998 p->set_mark((markOop)_overflow_list);
7999 _overflow_list = p;
8000 }
8001
8002 // Multi-threaded; use CAS to prepend to overflow list
8003 void CMSCollector::par_push_on_overflow_list(oop p) {
8004 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8005 assert(p->is_oop(), "Not an oop");
8006 par_preserve_mark_if_necessary(p);
8007 oop observed_overflow_list = _overflow_list;
8008 oop cur_overflow_list;
8009 do {
8010 cur_overflow_list = observed_overflow_list;
8011 if (cur_overflow_list != BUSY) {
8012 p->set_mark(markOop(cur_overflow_list));
8013 } else {
8014 p->set_mark(NULL);
8015 }
8016 observed_overflow_list =
8017 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8018 } while (cur_overflow_list != observed_overflow_list);
8019 }
8020 #undef BUSY
8021
8022 // Single threaded
8023 // General Note on GrowableArray: pushes may silently fail
8024 // because we are (temporarily) out of C-heap for expanding
8025 // the stack. The problem is quite ubiquitous and affects
8026 // a lot of code in the JVM. The prudent thing for GrowableArray
8027 // to do (for now) is to exit with an error. However, that may
8028 // be too draconian in some cases because the caller may be
8029 // able to recover without much harm. For such cases, we
8030 // should probably introduce a "soft_push" method which returns
8031 // an indication of success or failure with the assumption that
8032 // the caller may be able to recover from a failure; code in
8033 // the VM can then be changed, incrementally, to deal with such
8034 // failures where possible, thus, incrementally hardening the VM
8035 // in such low resource situations.
8036 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8037 _preserved_oop_stack.push(p);
8038 _preserved_mark_stack.push(m);
8039 assert(m == p->mark(), "Mark word changed");
8040 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8041 "bijection");
8042 }
8043
8044 // Single threaded
8045 void CMSCollector::preserve_mark_if_necessary(oop p) {
8046 markOop m = p->mark();
8047 if (m->must_be_preserved(p)) {
8048 preserve_mark_work(p, m);
8049 }
8050 }
8051
8052 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8053 markOop m = p->mark();
8054 if (m->must_be_preserved(p)) {
8055 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8056 // Even though we read the mark word without holding
8057 // the lock, we are assured that it will not change
8058 // because we "own" this oop, so no other thread can
8059 // be trying to push it on the overflow list; see
8060 // the assertion in preserve_mark_work() that checks
8061 // that m == p->mark().
8062 preserve_mark_work(p, m);
8063 }
8064 }
8065
8066 // We should be able to do this multi-threaded,
8067 // a chunk of stack being a task (this is
8068 // correct because each oop only ever appears
8069 // once in the overflow list. However, it's
8070 // not very easy to completely overlap this with
8071 // other operations, so will generally not be done
8072 // until all work's been completed. Because we
8073 // expect the preserved oop stack (set) to be small,
8074 // it's probably fine to do this single-threaded.
8075 // We can explore cleverer concurrent/overlapped/parallel
8076 // processing of preserved marks if we feel the
8077 // need for this in the future. Stack overflow should
8078 // be so rare in practice and, when it happens, its
8079 // effect on performance so great that this will
8080 // likely just be in the noise anyway.
8081 void CMSCollector::restore_preserved_marks_if_any() {
8082 assert(SafepointSynchronize::is_at_safepoint(),
8083 "world should be stopped");
8084 assert(Thread::current()->is_ConcurrentGC_thread() ||
8085 Thread::current()->is_VM_thread(),
8086 "should be single-threaded");
8087 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8088 "bijection");
8089
8090 while (!_preserved_oop_stack.is_empty()) {
8091 oop p = _preserved_oop_stack.pop();
8092 assert(p->is_oop(), "Should be an oop");
8093 assert(_span.contains(p), "oop should be in _span");
8094 assert(p->mark() == markOopDesc::prototype(),
8095 "Set when taken from overflow list");
8096 markOop m = _preserved_mark_stack.pop();
8097 p->set_mark(m);
8098 }
8099 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8100 "stacks were cleared above");
8101 }
8102
8103 #ifndef PRODUCT
8104 bool CMSCollector::no_preserved_marks() const {
8105 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8106 }
8107 #endif
8108
8109 // Transfer some number of overflown objects to usual marking
8110 // stack. Return true if some objects were transferred.
8111 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8112 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8113 (size_t)ParGCDesiredObjsFromOverflowList);
8114
8115 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8116 assert(_collector->overflow_list_is_empty() || res,
8117 "If list is not empty, we should have taken something");
8118 assert(!res || !_mark_stack->isEmpty(),
8119 "If we took something, it should now be on our stack");
8120 return res;
8121 }
8122
8123 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8124 size_t res = _sp->block_size_no_stall(addr, _collector);
8125 if (_sp->block_is_obj(addr)) {
8126 if (_live_bit_map->isMarked(addr)) {
8127 // It can't have been dead in a previous cycle
8128 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8129 } else {
8130 _dead_bit_map->mark(addr); // mark the dead object
8131 }
8132 }
8133 // Could be 0, if the block size could not be computed without stalling.
8134 return res;
8135 }
8136
8137 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8138
8139 switch (phase) {
8140 case CMSCollector::InitialMarking:
8141 initialize(true /* fullGC */ ,
8142 cause /* cause of the GC */,
8143 true /* recordGCBeginTime */,
8144 true /* recordPreGCUsage */,
8145 false /* recordPeakUsage */,
8146 false /* recordPostGCusage */,
8147 true /* recordAccumulatedGCTime */,
8148 false /* recordGCEndTime */,
8149 false /* countCollection */ );
8150 break;
8151
8152 case CMSCollector::FinalMarking:
8153 initialize(true /* fullGC */ ,
8154 cause /* cause of the GC */,
8155 false /* recordGCBeginTime */,
8156 false /* recordPreGCUsage */,
8157 false /* recordPeakUsage */,
8158 false /* recordPostGCusage */,
8159 true /* recordAccumulatedGCTime */,
8160 false /* recordGCEndTime */,
8161 false /* countCollection */ );
8162 break;
8163
8164 case CMSCollector::Sweeping:
8165 initialize(true /* fullGC */ ,
8166 cause /* cause of the GC */,
8167 false /* recordGCBeginTime */,
8168 false /* recordPreGCUsage */,
8169 true /* recordPeakUsage */,
8170 true /* recordPostGCusage */,
8171 false /* recordAccumulatedGCTime */,
8172 true /* recordGCEndTime */,
8173 true /* countCollection */ );
8174 break;
8175
8176 default:
8177 ShouldNotReachHere();
8178 }
8179 }