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