1 /*
2 * Copyright (c) 2001, 2015, 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/systemDictionary.hpp"
29 #include "code/codeCache.hpp"
30 #include "gc/cms/cmsCollectorPolicy.hpp"
31 #include "gc/cms/cmsOopClosures.inline.hpp"
32 #include "gc/cms/compactibleFreeListSpace.hpp"
33 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp"
34 #include "gc/cms/concurrentMarkSweepThread.hpp"
35 #include "gc/cms/parNewGeneration.hpp"
36 #include "gc/cms/vmCMSOperations.hpp"
37 #include "gc/serial/genMarkSweep.hpp"
38 #include "gc/serial/tenuredGeneration.hpp"
39 #include "gc/shared/adaptiveSizePolicy.hpp"
40 #include "gc/shared/cardGeneration.inline.hpp"
41 #include "gc/shared/cardTableRS.hpp"
42 #include "gc/shared/collectedHeap.inline.hpp"
43 #include "gc/shared/collectorCounters.hpp"
44 #include "gc/shared/collectorPolicy.hpp"
45 #include "gc/shared/gcLocker.inline.hpp"
46 #include "gc/shared/gcTimer.hpp"
47 #include "gc/shared/gcTrace.hpp"
48 #include "gc/shared/gcTraceTime.hpp"
49 #include "gc/shared/genCollectedHeap.hpp"
50 #include "gc/shared/genOopClosures.inline.hpp"
51 #include "gc/shared/isGCActiveMark.hpp"
52 #include "gc/shared/referencePolicy.hpp"
53 #include "gc/shared/strongRootsScope.hpp"
54 #include "gc/shared/taskqueue.inline.hpp"
55 #include "memory/allocation.hpp"
56 #include "memory/iterator.inline.hpp"
57 #include "memory/padded.hpp"
58 #include "memory/resourceArea.hpp"
59 #include "oops/oop.inline.hpp"
60 #include "prims/jvmtiExport.hpp"
61 #include "runtime/atomic.inline.hpp"
62 #include "runtime/globals_extension.hpp"
63 #include "runtime/handles.inline.hpp"
64 #include "runtime/java.hpp"
65 #include "runtime/orderAccess.inline.hpp"
66 #include "runtime/vmThread.hpp"
67 #include "services/memoryService.hpp"
68 #include "services/runtimeService.hpp"
69 #include "utilities/stack.inline.hpp"
70
71 // statics
72 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
73 bool CMSCollector::_full_gc_requested = false;
74 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
75
76 //////////////////////////////////////////////////////////////////
77 // In support of CMS/VM thread synchronization
78 //////////////////////////////////////////////////////////////////
79 // We split use of the CGC_lock into 2 "levels".
80 // The low-level locking is of the usual CGC_lock monitor. We introduce
81 // a higher level "token" (hereafter "CMS token") built on top of the
82 // low level monitor (hereafter "CGC lock").
83 // The token-passing protocol gives priority to the VM thread. The
84 // CMS-lock doesn't provide any fairness guarantees, but clients
85 // should ensure that it is only held for very short, bounded
86 // durations.
87 //
88 // When either of the CMS thread or the VM thread is involved in
89 // collection operations during which it does not want the other
90 // thread to interfere, it obtains the CMS token.
91 //
92 // If either thread tries to get the token while the other has
93 // it, that thread waits. However, if the VM thread and CMS thread
94 // both want the token, then the VM thread gets priority while the
95 // CMS thread waits. This ensures, for instance, that the "concurrent"
96 // phases of the CMS thread's work do not block out the VM thread
97 // for long periods of time as the CMS thread continues to hog
98 // the token. (See bug 4616232).
99 //
100 // The baton-passing functions are, however, controlled by the
101 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
102 // and here the low-level CMS lock, not the high level token,
103 // ensures mutual exclusion.
104 //
105 // Two important conditions that we have to satisfy:
106 // 1. if a thread does a low-level wait on the CMS lock, then it
107 // relinquishes the CMS token if it were holding that token
108 // when it acquired the low-level CMS lock.
109 // 2. any low-level notifications on the low-level lock
110 // should only be sent when a thread has relinquished the token.
111 //
112 // In the absence of either property, we'd have potential deadlock.
113 //
114 // We protect each of the CMS (concurrent and sequential) phases
115 // with the CMS _token_, not the CMS _lock_.
116 //
117 // The only code protected by CMS lock is the token acquisition code
118 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
119 // baton-passing code.
120 //
121 // Unfortunately, i couldn't come up with a good abstraction to factor and
122 // hide the naked CGC_lock manipulation in the baton-passing code
123 // further below. That's something we should try to do. Also, the proof
124 // of correctness of this 2-level locking scheme is far from obvious,
125 // and potentially quite slippery. We have an uneasy suspicion, for instance,
126 // that there may be a theoretical possibility of delay/starvation in the
127 // low-level lock/wait/notify scheme used for the baton-passing because of
128 // potential interference with the priority scheme embodied in the
129 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
130 // invocation further below and marked with "XXX 20011219YSR".
131 // Indeed, as we note elsewhere, this may become yet more slippery
132 // in the presence of multiple CMS and/or multiple VM threads. XXX
133
134 class CMSTokenSync: public StackObj {
135 private:
136 bool _is_cms_thread;
137 public:
138 CMSTokenSync(bool is_cms_thread):
139 _is_cms_thread(is_cms_thread) {
140 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
141 "Incorrect argument to constructor");
142 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
143 }
144
145 ~CMSTokenSync() {
146 assert(_is_cms_thread ?
147 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
148 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
149 "Incorrect state");
150 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
151 }
152 };
153
154 // Convenience class that does a CMSTokenSync, and then acquires
155 // upto three locks.
156 class CMSTokenSyncWithLocks: public CMSTokenSync {
157 private:
158 // Note: locks are acquired in textual declaration order
159 // and released in the opposite order
160 MutexLockerEx _locker1, _locker2, _locker3;
161 public:
162 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
163 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
164 CMSTokenSync(is_cms_thread),
165 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
166 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
167 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
168 { }
169 };
170
171
172 //////////////////////////////////////////////////////////////////
173 // Concurrent Mark-Sweep Generation /////////////////////////////
174 //////////////////////////////////////////////////////////////////
175
176 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
177
178 // This struct contains per-thread things necessary to support parallel
179 // young-gen collection.
180 class CMSParGCThreadState: public CHeapObj<mtGC> {
181 public:
182 CFLS_LAB lab;
183 PromotionInfo promo;
184
185 // Constructor.
186 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
187 promo.setSpace(cfls);
188 }
189 };
190
191 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
192 ReservedSpace rs, size_t initial_byte_size, int level,
193 CardTableRS* ct, bool use_adaptive_freelists,
194 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) :
195 CardGeneration(rs, initial_byte_size, level, ct),
196 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
197 _did_compact(false)
198 {
199 HeapWord* bottom = (HeapWord*) _virtual_space.low();
200 HeapWord* end = (HeapWord*) _virtual_space.high();
201
202 _direct_allocated_words = 0;
203 NOT_PRODUCT(
204 _numObjectsPromoted = 0;
205 _numWordsPromoted = 0;
206 _numObjectsAllocated = 0;
207 _numWordsAllocated = 0;
208 )
209
210 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
211 use_adaptive_freelists,
212 dictionaryChoice);
213 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
214 _cmsSpace->_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 (int) ParallelGCThreads, // mt processing degree
288 _cmsGen->refs_discovery_is_mt(), // mt discovery
289 (int) 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 = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy();
307
308 // Generation Counters - generation 1, 1 subspace
309 _gen_counters = new GenerationCounters(gen_name, 1, 1,
310 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
311
312 _space_counters = new GSpaceCounters(gen_name, 0,
313 _virtual_space.reserved_size(),
314 this, _gen_counters);
315 }
316
317 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
318 _cms_gen(cms_gen)
319 {
320 assert(alpha <= 100, "bad value");
321 _saved_alpha = alpha;
322
323 // Initialize the alphas to the bootstrap value of 100.
324 _gc0_alpha = _cms_alpha = 100;
325
326 _cms_begin_time.update();
327 _cms_end_time.update();
328
329 _gc0_duration = 0.0;
330 _gc0_period = 0.0;
331 _gc0_promoted = 0;
332
333 _cms_duration = 0.0;
334 _cms_period = 0.0;
335 _cms_allocated = 0;
336
337 _cms_used_at_gc0_begin = 0;
338 _cms_used_at_gc0_end = 0;
339 _allow_duty_cycle_reduction = false;
340 _valid_bits = 0;
341 }
342
343 double CMSStats::cms_free_adjustment_factor(size_t free) const {
344 // TBD: CR 6909490
345 return 1.0;
346 }
347
348 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
349 }
350
351 // If promotion failure handling is on use
352 // the padded average size of the promotion for each
353 // young generation collection.
354 double CMSStats::time_until_cms_gen_full() const {
355 size_t cms_free = _cms_gen->cmsSpace()->free();
356 GenCollectedHeap* gch = GenCollectedHeap::heap();
357 size_t expected_promotion = MIN2(gch->young_gen()->capacity(),
358 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
359 if (cms_free > expected_promotion) {
360 // Start a cms collection if there isn't enough space to promote
361 // for the next minor collection. Use the padded average as
362 // a safety factor.
363 cms_free -= expected_promotion;
364
365 // Adjust by the safety factor.
366 double cms_free_dbl = (double)cms_free;
367 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
368 // Apply a further correction factor which tries to adjust
369 // for recent occurance of concurrent mode failures.
370 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
371 cms_free_dbl = cms_free_dbl * cms_adjustment;
372
373 if (PrintGCDetails && Verbose) {
374 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
375 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
376 cms_free, expected_promotion);
377 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
378 cms_free_dbl, cms_consumption_rate() + 1.0);
379 }
380 // Add 1 in case the consumption rate goes to zero.
381 return cms_free_dbl / (cms_consumption_rate() + 1.0);
382 }
383 return 0.0;
384 }
385
386 // Compare the duration of the cms collection to the
387 // time remaining before the cms generation is empty.
388 // Note that the time from the start of the cms collection
389 // to the start of the cms sweep (less than the total
390 // duration of the cms collection) can be used. This
391 // has been tried and some applications experienced
392 // promotion failures early in execution. This was
393 // possibly because the averages were not accurate
394 // enough at the beginning.
395 double CMSStats::time_until_cms_start() const {
396 // We add "gc0_period" to the "work" calculation
397 // below because this query is done (mostly) at the
398 // end of a scavenge, so we need to conservatively
399 // account for that much possible delay
400 // in the query so as to avoid concurrent mode failures
401 // due to starting the collection just a wee bit too
402 // late.
403 double work = cms_duration() + gc0_period();
404 double deadline = time_until_cms_gen_full();
405 // If a concurrent mode failure occurred recently, we want to be
406 // more conservative and halve our expected time_until_cms_gen_full()
407 if (work > deadline) {
408 if (Verbose && PrintGCDetails) {
409 gclog_or_tty->print(
410 " CMSCollector: collect because of anticipated promotion "
411 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
412 gc0_period(), time_until_cms_gen_full());
413 }
414 return 0.0;
415 }
416 return work - deadline;
417 }
418
419 #ifndef PRODUCT
420 void CMSStats::print_on(outputStream *st) const {
421 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
422 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
423 gc0_duration(), gc0_period(), gc0_promoted());
424 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
425 cms_duration(), cms_period(), cms_allocated());
426 st->print(",cms_since_beg=%g,cms_since_end=%g",
427 cms_time_since_begin(), cms_time_since_end());
428 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
429 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
430
431 if (valid()) {
432 st->print(",promo_rate=%g,cms_alloc_rate=%g",
433 promotion_rate(), cms_allocation_rate());
434 st->print(",cms_consumption_rate=%g,time_until_full=%g",
435 cms_consumption_rate(), time_until_cms_gen_full());
436 }
437 st->print(" ");
438 }
439 #endif // #ifndef PRODUCT
440
441 CMSCollector::CollectorState CMSCollector::_collectorState =
442 CMSCollector::Idling;
443 bool CMSCollector::_foregroundGCIsActive = false;
444 bool CMSCollector::_foregroundGCShouldWait = false;
445
446 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
447 CardTableRS* ct,
448 ConcurrentMarkSweepPolicy* cp):
449 _cmsGen(cmsGen),
450 _ct(ct),
451 _ref_processor(NULL), // will be set later
452 _conc_workers(NULL), // may be set later
453 _abort_preclean(false),
454 _start_sampling(false),
455 _between_prologue_and_epilogue(false),
456 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
457 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
458 -1 /* lock-free */, "No_lock" /* dummy */),
459 _modUnionClosurePar(&_modUnionTable),
460 // Adjust my span to cover old (cms) gen
461 _span(cmsGen->reserved()),
462 // Construct the is_alive_closure with _span & markBitMap
463 _is_alive_closure(_span, &_markBitMap),
464 _restart_addr(NULL),
465 _overflow_list(NULL),
466 _stats(cmsGen),
467 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
468 //verify that this lock should be acquired with safepoint check.
469 Monitor::_safepoint_check_sometimes)),
470 _eden_chunk_array(NULL), // may be set in ctor body
471 _eden_chunk_capacity(0), // -- ditto --
472 _eden_chunk_index(0), // -- ditto --
473 _survivor_plab_array(NULL), // -- ditto --
474 _survivor_chunk_array(NULL), // -- ditto --
475 _survivor_chunk_capacity(0), // -- ditto --
476 _survivor_chunk_index(0), // -- ditto --
477 _ser_pmc_preclean_ovflw(0),
478 _ser_kac_preclean_ovflw(0),
479 _ser_pmc_remark_ovflw(0),
480 _par_pmc_remark_ovflw(0),
481 _ser_kac_ovflw(0),
482 _par_kac_ovflw(0),
483 #ifndef PRODUCT
484 _num_par_pushes(0),
485 #endif
486 _collection_count_start(0),
487 _verifying(false),
488 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
489 _completed_initialization(false),
490 _collector_policy(cp),
491 _should_unload_classes(CMSClassUnloadingEnabled),
492 _concurrent_cycles_since_last_unload(0),
493 _roots_scanning_options(GenCollectedHeap::SO_None),
494 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
495 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
496 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
497 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
498 _cms_start_registered(false)
499 {
500 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
501 ExplicitGCInvokesConcurrent = true;
502 }
503 // Now expand the span and allocate the collection support structures
504 // (MUT, marking bit map etc.) to cover both generations subject to
505 // collection.
506
507 // For use by dirty card to oop closures.
508 _cmsGen->cmsSpace()->set_collector(this);
509
510 // Allocate MUT and marking bit map
511 {
512 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
513 if (!_markBitMap.allocate(_span)) {
514 warning("Failed to allocate CMS Bit Map");
515 return;
516 }
517 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
518 }
519 {
520 _modUnionTable.allocate(_span);
521 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
522 }
523
524 if (!_markStack.allocate(MarkStackSize)) {
525 warning("Failed to allocate CMS Marking Stack");
526 return;
527 }
528
529 // Support for multi-threaded concurrent phases
530 if (CMSConcurrentMTEnabled) {
531 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
532 // just for now
533 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
534 }
535 if (ConcGCThreads > 1) {
536 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
537 ConcGCThreads, true);
538 if (_conc_workers == NULL) {
539 warning("GC/CMS: _conc_workers allocation failure: "
540 "forcing -CMSConcurrentMTEnabled");
541 CMSConcurrentMTEnabled = false;
542 } else {
543 _conc_workers->initialize_workers();
544 }
545 } else {
546 CMSConcurrentMTEnabled = false;
547 }
548 }
549 if (!CMSConcurrentMTEnabled) {
550 ConcGCThreads = 0;
551 } else {
552 // Turn off CMSCleanOnEnter optimization temporarily for
553 // the MT case where it's not fixed yet; see 6178663.
554 CMSCleanOnEnter = false;
555 }
556 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
557 "Inconsistency");
558
559 // Parallel task queues; these are shared for the
560 // concurrent and stop-world phases of CMS, but
561 // are not shared with parallel scavenge (ParNew).
562 {
563 uint i;
564 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
565
566 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
567 || ParallelRefProcEnabled)
568 && num_queues > 0) {
569 _task_queues = new OopTaskQueueSet(num_queues);
570 if (_task_queues == NULL) {
571 warning("task_queues allocation failure.");
572 return;
573 }
574 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
575 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
576 for (i = 0; i < num_queues; i++) {
577 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
578 if (q == NULL) {
579 warning("work_queue allocation failure.");
580 return;
581 }
582 _task_queues->register_queue(i, q);
583 }
584 for (i = 0; i < num_queues; i++) {
585 _task_queues->queue(i)->initialize();
586 _hash_seed[i] = 17; // copied from ParNew
587 }
588 }
589 }
590
591 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
592
593 // Clip CMSBootstrapOccupancy between 0 and 100.
594 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100;
595
596 // Now tell CMS generations the identity of their collector
597 ConcurrentMarkSweepGeneration::set_collector(this);
598
599 // Create & start a CMS thread for this CMS collector
600 _cmsThread = ConcurrentMarkSweepThread::start(this);
601 assert(cmsThread() != NULL, "CMS Thread should have been created");
602 assert(cmsThread()->collector() == this,
603 "CMS Thread should refer to this gen");
604 assert(CGC_lock != NULL, "Where's the CGC_lock?");
605
606 // Support for parallelizing young gen rescan
607 GenCollectedHeap* gch = GenCollectedHeap::heap();
608 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
609 _young_gen = (ParNewGeneration*)gch->young_gen();
610 if (gch->supports_inline_contig_alloc()) {
611 _top_addr = gch->top_addr();
612 _end_addr = gch->end_addr();
613 assert(_young_gen != NULL, "no _young_gen");
614 _eden_chunk_index = 0;
615 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
616 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
617 }
618
619 // Support for parallelizing survivor space rescan
620 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
621 const size_t max_plab_samples =
622 ((DefNewGeneration*)_young_gen)->max_survivor_size() / plab_sample_minimum_size();
623
624 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
625 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC);
626 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
627 _survivor_chunk_capacity = 2*max_plab_samples;
628 for (uint i = 0; i < ParallelGCThreads; i++) {
629 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
630 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
631 assert(cur->end() == 0, "Should be 0");
632 assert(cur->array() == vec, "Should be vec");
633 assert(cur->capacity() == max_plab_samples, "Error");
634 }
635 }
636
637 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
638 _gc_counters = new CollectorCounters("CMS", 1);
639 _completed_initialization = true;
640 _inter_sweep_timer.start(); // start of time
641 }
642
643 size_t CMSCollector::plab_sample_minimum_size() {
644 // The default value of MinTLABSize is 2k, but there is
645 // no way to get the default value if the flag has been overridden.
646 return MAX2(ThreadLocalAllocBuffer::min_size() * HeapWordSize, 2 * K);
647 }
648
649 const char* ConcurrentMarkSweepGeneration::name() const {
650 return "concurrent mark-sweep generation";
651 }
652 void ConcurrentMarkSweepGeneration::update_counters() {
653 if (UsePerfData) {
654 _space_counters->update_all();
655 _gen_counters->update_all();
656 }
657 }
658
659 // this is an optimized version of update_counters(). it takes the
660 // used value as a parameter rather than computing it.
661 //
662 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
663 if (UsePerfData) {
664 _space_counters->update_used(used);
665 _space_counters->update_capacity();
666 _gen_counters->update_all();
667 }
668 }
669
670 void ConcurrentMarkSweepGeneration::print() const {
671 Generation::print();
672 cmsSpace()->print();
673 }
674
675 #ifndef PRODUCT
676 void ConcurrentMarkSweepGeneration::print_statistics() {
677 cmsSpace()->printFLCensus(0);
678 }
679 #endif
680
681 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
682 GenCollectedHeap* gch = GenCollectedHeap::heap();
683 if (PrintGCDetails) {
684 if (Verbose) {
685 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
686 level(), short_name(), s, used(), capacity());
687 } else {
688 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
689 level(), short_name(), s, used() / K, capacity() / K);
690 }
691 }
692 if (Verbose) {
693 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
694 gch->used(), gch->capacity());
695 } else {
696 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
697 gch->used() / K, gch->capacity() / K);
698 }
699 }
700
701 size_t
702 ConcurrentMarkSweepGeneration::contiguous_available() const {
703 // dld proposes an improvement in precision here. If the committed
704 // part of the space ends in a free block we should add that to
705 // uncommitted size in the calculation below. Will make this
706 // change later, staying with the approximation below for the
707 // time being. -- ysr.
708 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
709 }
710
711 size_t
712 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
713 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
714 }
715
716 size_t ConcurrentMarkSweepGeneration::max_available() const {
717 return free() + _virtual_space.uncommitted_size();
718 }
719
720 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
721 size_t available = max_available();
722 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
723 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
724 if (Verbose && PrintGCDetails) {
725 gclog_or_tty->print_cr(
726 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
727 "max_promo("SIZE_FORMAT")",
728 res? "":" not", available, res? ">=":"<",
729 av_promo, max_promotion_in_bytes);
730 }
731 return res;
732 }
733
734 // At a promotion failure dump information on block layout in heap
735 // (cms old generation).
736 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
737 if (CMSDumpAtPromotionFailure) {
738 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
739 }
740 }
741
742 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
743 // Clear the promotion information. These pointers can be adjusted
744 // along with all the other pointers into the heap but
745 // compaction is expected to be a rare event with
746 // a heap using cms so don't do it without seeing the need.
747 for (uint i = 0; i < ParallelGCThreads; i++) {
748 _par_gc_thread_states[i]->promo.reset();
749 }
750 }
751
752 void ConcurrentMarkSweepGeneration::compute_new_size() {
753 assert_locked_or_safepoint(Heap_lock);
754
755 // If incremental collection failed, we just want to expand
756 // to the limit.
757 if (incremental_collection_failed()) {
758 clear_incremental_collection_failed();
759 grow_to_reserved();
760 return;
761 }
762
763 // The heap has been compacted but not reset yet.
764 // Any metric such as free() or used() will be incorrect.
765
766 CardGeneration::compute_new_size();
767
768 // Reset again after a possible resizing
769 if (did_compact()) {
770 cmsSpace()->reset_after_compaction();
771 }
772 }
773
774 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
775 assert_locked_or_safepoint(Heap_lock);
776
777 // If incremental collection failed, we just want to expand
778 // to the limit.
779 if (incremental_collection_failed()) {
780 clear_incremental_collection_failed();
781 grow_to_reserved();
782 return;
783 }
784
785 double free_percentage = ((double) free()) / capacity();
786 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
787 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
788
789 // compute expansion delta needed for reaching desired free percentage
790 if (free_percentage < desired_free_percentage) {
791 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
792 assert(desired_capacity >= capacity(), "invalid expansion size");
793 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
794 if (PrintGCDetails && Verbose) {
795 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
796 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
797 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
798 gclog_or_tty->print_cr(" Desired free fraction %f",
799 desired_free_percentage);
800 gclog_or_tty->print_cr(" Maximum free fraction %f",
801 maximum_free_percentage);
802 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000);
803 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
804 desired_capacity/1000);
805 int prev_level = level() - 1;
806 if (prev_level >= 0) {
807 size_t prev_size = 0;
808 GenCollectedHeap* gch = GenCollectedHeap::heap();
809 Generation* prev_gen = gch->young_gen();
810 prev_size = prev_gen->capacity();
811 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
812 prev_size/1000);
813 }
814 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
815 unsafe_max_alloc_nogc()/1000);
816 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
817 contiguous_available()/1000);
818 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
819 expand_bytes);
820 }
821 // safe if expansion fails
822 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
823 if (PrintGCDetails && Verbose) {
824 gclog_or_tty->print_cr(" Expanded free fraction %f",
825 ((double) free()) / capacity());
826 }
827 } else {
828 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
829 assert(desired_capacity <= capacity(), "invalid expansion size");
830 size_t shrink_bytes = capacity() - desired_capacity;
831 // Don't shrink unless the delta is greater than the minimum shrink we want
832 if (shrink_bytes >= MinHeapDeltaBytes) {
833 shrink_free_list_by(shrink_bytes);
834 }
835 }
836 }
837
838 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
839 return cmsSpace()->freelistLock();
840 }
841
842 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
843 bool tlab) {
844 CMSSynchronousYieldRequest yr;
845 MutexLockerEx x(freelistLock(),
846 Mutex::_no_safepoint_check_flag);
847 return have_lock_and_allocate(size, tlab);
848 }
849
850 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
851 bool tlab /* ignored */) {
852 assert_lock_strong(freelistLock());
853 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
854 HeapWord* res = cmsSpace()->allocate(adjustedSize);
855 // Allocate the object live (grey) if the background collector has
856 // started marking. This is necessary because the marker may
857 // have passed this address and consequently this object will
858 // not otherwise be greyed and would be incorrectly swept up.
859 // Note that if this object contains references, the writing
860 // of those references will dirty the card containing this object
861 // allowing the object to be blackened (and its references scanned)
862 // either during a preclean phase or at the final checkpoint.
863 if (res != NULL) {
864 // We may block here with an uninitialized object with
865 // its mark-bit or P-bits not yet set. Such objects need
866 // to be safely navigable by block_start().
867 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
868 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
869 collector()->direct_allocated(res, adjustedSize);
870 _direct_allocated_words += adjustedSize;
871 // allocation counters
872 NOT_PRODUCT(
873 _numObjectsAllocated++;
874 _numWordsAllocated += (int)adjustedSize;
875 )
876 }
877 return res;
878 }
879
880 // In the case of direct allocation by mutators in a generation that
881 // is being concurrently collected, the object must be allocated
882 // live (grey) if the background collector has started marking.
883 // This is necessary because the marker may
884 // have passed this address and consequently this object will
885 // not otherwise be greyed and would be incorrectly swept up.
886 // Note that if this object contains references, the writing
887 // of those references will dirty the card containing this object
888 // allowing the object to be blackened (and its references scanned)
889 // either during a preclean phase or at the final checkpoint.
890 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
891 assert(_markBitMap.covers(start, size), "Out of bounds");
892 if (_collectorState >= Marking) {
893 MutexLockerEx y(_markBitMap.lock(),
894 Mutex::_no_safepoint_check_flag);
895 // [see comments preceding SweepClosure::do_blk() below for details]
896 //
897 // Can the P-bits be deleted now? JJJ
898 //
899 // 1. need to mark the object as live so it isn't collected
900 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
901 // 3. need to mark the end of the object so marking, precleaning or sweeping
902 // can skip over uninitialized or unparsable objects. An allocated
903 // object is considered uninitialized for our purposes as long as
904 // its klass word is NULL. All old gen objects are parsable
905 // as soon as they are initialized.)
906 _markBitMap.mark(start); // object is live
907 _markBitMap.mark(start + 1); // object is potentially uninitialized?
908 _markBitMap.mark(start + size - 1);
909 // mark end of object
910 }
911 // check that oop looks uninitialized
912 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
913 }
914
915 void CMSCollector::promoted(bool par, HeapWord* start,
916 bool is_obj_array, size_t obj_size) {
917 assert(_markBitMap.covers(start), "Out of bounds");
918 // See comment in direct_allocated() about when objects should
919 // be allocated live.
920 if (_collectorState >= Marking) {
921 // we already hold the marking bit map lock, taken in
922 // the prologue
923 if (par) {
924 _markBitMap.par_mark(start);
925 } else {
926 _markBitMap.mark(start);
927 }
928 // We don't need to mark the object as uninitialized (as
929 // in direct_allocated above) because this is being done with the
930 // world stopped and the object will be initialized by the
931 // time the marking, precleaning or sweeping get to look at it.
932 // But see the code for copying objects into the CMS generation,
933 // where we need to ensure that concurrent readers of the
934 // block offset table are able to safely navigate a block that
935 // is in flux from being free to being allocated (and in
936 // transition while being copied into) and subsequently
937 // becoming a bona-fide object when the copy/promotion is complete.
938 assert(SafepointSynchronize::is_at_safepoint(),
939 "expect promotion only at safepoints");
940
941 if (_collectorState < Sweeping) {
942 // Mark the appropriate cards in the modUnionTable, so that
943 // this object gets scanned before the sweep. If this is
944 // not done, CMS generation references in the object might
945 // not get marked.
946 // For the case of arrays, which are otherwise precisely
947 // marked, we need to dirty the entire array, not just its head.
948 if (is_obj_array) {
949 // The [par_]mark_range() method expects mr.end() below to
950 // be aligned to the granularity of a bit's representation
951 // in the heap. In the case of the MUT below, that's a
952 // card size.
953 MemRegion mr(start,
954 (HeapWord*)round_to((intptr_t)(start + obj_size),
955 CardTableModRefBS::card_size /* bytes */));
956 if (par) {
957 _modUnionTable.par_mark_range(mr);
958 } else {
959 _modUnionTable.mark_range(mr);
960 }
961 } else { // not an obj array; we can just mark the head
962 if (par) {
963 _modUnionTable.par_mark(start);
964 } else {
965 _modUnionTable.mark(start);
966 }
967 }
968 }
969 }
970 }
971
972 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
973 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
974 // allocate, copy and if necessary update promoinfo --
975 // delegate to underlying space.
976 assert_lock_strong(freelistLock());
977
978 #ifndef PRODUCT
979 if (GenCollectedHeap::heap()->promotion_should_fail()) {
980 return NULL;
981 }
982 #endif // #ifndef PRODUCT
983
984 oop res = _cmsSpace->promote(obj, obj_size);
985 if (res == NULL) {
986 // expand and retry
987 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
988 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
989 // Since this is the old generation, we don't try to promote
990 // into a more senior generation.
991 res = _cmsSpace->promote(obj, obj_size);
992 }
993 if (res != NULL) {
994 // See comment in allocate() about when objects should
995 // be allocated live.
996 assert(obj->is_oop(), "Will dereference klass pointer below");
997 collector()->promoted(false, // Not parallel
998 (HeapWord*)res, obj->is_objArray(), obj_size);
999 // promotion counters
1000 NOT_PRODUCT(
1001 _numObjectsPromoted++;
1002 _numWordsPromoted +=
1003 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1004 )
1005 }
1006 return res;
1007 }
1008
1009
1010 // IMPORTANT: Notes on object size recognition in CMS.
1011 // ---------------------------------------------------
1012 // A block of storage in the CMS generation is always in
1013 // one of three states. A free block (FREE), an allocated
1014 // object (OBJECT) whose size() method reports the correct size,
1015 // and an intermediate state (TRANSIENT) in which its size cannot
1016 // be accurately determined.
1017 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1018 // -----------------------------------------------------
1019 // FREE: klass_word & 1 == 1; mark_word holds block size
1020 //
1021 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1022 // obj->size() computes correct size
1023 //
1024 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1025 //
1026 // STATE IDENTIFICATION: (64 bit+COOPS)
1027 // ------------------------------------
1028 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1029 //
1030 // OBJECT: klass_word installed; klass_word != 0;
1031 // obj->size() computes correct size
1032 //
1033 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1034 //
1035 //
1036 // STATE TRANSITION DIAGRAM
1037 //
1038 // mut / parnew mut / parnew
1039 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1040 // ^ |
1041 // |------------------------ DEAD <------------------------------------|
1042 // sweep mut
1043 //
1044 // While a block is in TRANSIENT state its size cannot be determined
1045 // so readers will either need to come back later or stall until
1046 // the size can be determined. Note that for the case of direct
1047 // allocation, P-bits, when available, may be used to determine the
1048 // size of an object that may not yet have been initialized.
1049
1050 // Things to support parallel young-gen collection.
1051 oop
1052 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1053 oop old, markOop m,
1054 size_t word_sz) {
1055 #ifndef PRODUCT
1056 if (GenCollectedHeap::heap()->promotion_should_fail()) {
1057 return NULL;
1058 }
1059 #endif // #ifndef PRODUCT
1060
1061 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1062 PromotionInfo* promoInfo = &ps->promo;
1063 // if we are tracking promotions, then first ensure space for
1064 // promotion (including spooling space for saving header if necessary).
1065 // then allocate and copy, then track promoted info if needed.
1066 // When tracking (see PromotionInfo::track()), the mark word may
1067 // be displaced and in this case restoration of the mark word
1068 // occurs in the (oop_since_save_marks_)iterate phase.
1069 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1070 // Out of space for allocating spooling buffers;
1071 // try expanding and allocating spooling buffers.
1072 if (!expand_and_ensure_spooling_space(promoInfo)) {
1073 return NULL;
1074 }
1075 }
1076 assert(promoInfo->has_spooling_space(), "Control point invariant");
1077 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1078 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1079 if (obj_ptr == NULL) {
1080 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1081 if (obj_ptr == NULL) {
1082 return NULL;
1083 }
1084 }
1085 oop obj = oop(obj_ptr);
1086 OrderAccess::storestore();
1087 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1088 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1089 // IMPORTANT: See note on object initialization for CMS above.
1090 // Otherwise, copy the object. Here we must be careful to insert the
1091 // klass pointer last, since this marks the block as an allocated object.
1092 // Except with compressed oops it's the mark word.
1093 HeapWord* old_ptr = (HeapWord*)old;
1094 // Restore the mark word copied above.
1095 obj->set_mark(m);
1096 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1097 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1098 OrderAccess::storestore();
1099
1100 if (UseCompressedClassPointers) {
1101 // Copy gap missed by (aligned) header size calculation below
1102 obj->set_klass_gap(old->klass_gap());
1103 }
1104 if (word_sz > (size_t)oopDesc::header_size()) {
1105 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1106 obj_ptr + oopDesc::header_size(),
1107 word_sz - oopDesc::header_size());
1108 }
1109
1110 // Now we can track the promoted object, if necessary. We take care
1111 // to delay the transition from uninitialized to full object
1112 // (i.e., insertion of klass pointer) until after, so that it
1113 // atomically becomes a promoted object.
1114 if (promoInfo->tracking()) {
1115 promoInfo->track((PromotedObject*)obj, old->klass());
1116 }
1117 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1118 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1119 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1120
1121 // Finally, install the klass pointer (this should be volatile).
1122 OrderAccess::storestore();
1123 obj->set_klass(old->klass());
1124 // We should now be able to calculate the right size for this object
1125 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1126
1127 collector()->promoted(true, // parallel
1128 obj_ptr, old->is_objArray(), word_sz);
1129
1130 NOT_PRODUCT(
1131 Atomic::inc_ptr(&_numObjectsPromoted);
1132 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1133 )
1134
1135 return obj;
1136 }
1137
1138 void
1139 ConcurrentMarkSweepGeneration::
1140 par_promote_alloc_done(int thread_num) {
1141 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1142 ps->lab.retire(thread_num);
1143 }
1144
1145 void
1146 ConcurrentMarkSweepGeneration::
1147 par_oop_since_save_marks_iterate_done(int thread_num) {
1148 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1149 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1150 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1151 }
1152
1153 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1154 size_t size,
1155 bool tlab)
1156 {
1157 // We allow a STW collection only if a full
1158 // collection was requested.
1159 return full || should_allocate(size, tlab); // FIX ME !!!
1160 // This and promotion failure handling are connected at the
1161 // hip and should be fixed by untying them.
1162 }
1163
1164 bool CMSCollector::shouldConcurrentCollect() {
1165 if (_full_gc_requested) {
1166 if (Verbose && PrintGCDetails) {
1167 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1168 " gc request (or gc_locker)");
1169 }
1170 return true;
1171 }
1172
1173 FreelistLocker x(this);
1174 // ------------------------------------------------------------------
1175 // Print out lots of information which affects the initiation of
1176 // a collection.
1177 if (PrintCMSInitiationStatistics && stats().valid()) {
1178 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1179 gclog_or_tty->stamp();
1180 gclog_or_tty->cr();
1181 stats().print_on(gclog_or_tty);
1182 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1183 stats().time_until_cms_gen_full());
1184 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1185 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1186 _cmsGen->contiguous_available());
1187 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1188 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1189 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1190 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1191 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1192 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1193 gclog_or_tty->print_cr("metadata initialized %d",
1194 MetaspaceGC::should_concurrent_collect());
1195 }
1196 // ------------------------------------------------------------------
1197
1198 // If the estimated time to complete a cms collection (cms_duration())
1199 // is less than the estimated time remaining until the cms generation
1200 // is full, start a collection.
1201 if (!UseCMSInitiatingOccupancyOnly) {
1202 if (stats().valid()) {
1203 if (stats().time_until_cms_start() == 0.0) {
1204 return true;
1205 }
1206 } else {
1207 // We want to conservatively collect somewhat early in order
1208 // to try and "bootstrap" our CMS/promotion statistics;
1209 // this branch will not fire after the first successful CMS
1210 // collection because the stats should then be valid.
1211 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1212 if (Verbose && PrintGCDetails) {
1213 gclog_or_tty->print_cr(
1214 " CMSCollector: collect for bootstrapping statistics:"
1215 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1216 _bootstrap_occupancy);
1217 }
1218 return true;
1219 }
1220 }
1221 }
1222
1223 // Otherwise, we start a collection cycle if
1224 // old gen want a collection cycle started. Each may use
1225 // an appropriate criterion for making this decision.
1226 // XXX We need to make sure that the gen expansion
1227 // criterion dovetails well with this. XXX NEED TO FIX THIS
1228 if (_cmsGen->should_concurrent_collect()) {
1229 if (Verbose && PrintGCDetails) {
1230 gclog_or_tty->print_cr("CMS old gen initiated");
1231 }
1232 return true;
1233 }
1234
1235 // We start a collection if we believe an incremental collection may fail;
1236 // this is not likely to be productive in practice because it's probably too
1237 // late anyway.
1238 GenCollectedHeap* gch = GenCollectedHeap::heap();
1239 assert(gch->collector_policy()->is_generation_policy(),
1240 "You may want to check the correctness of the following");
1241 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1242 if (Verbose && PrintGCDetails) {
1243 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1244 }
1245 return true;
1246 }
1247
1248 if (MetaspaceGC::should_concurrent_collect()) {
1249 if (Verbose && PrintGCDetails) {
1250 gclog_or_tty->print("CMSCollector: collect for metadata allocation ");
1251 }
1252 return true;
1253 }
1254
1255 // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1256 if (CMSTriggerInterval >= 0) {
1257 if (CMSTriggerInterval == 0) {
1258 // Trigger always
1259 return true;
1260 }
1261
1262 // Check the CMS time since begin (we do not check the stats validity
1263 // as we want to be able to trigger the first CMS cycle as well)
1264 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1265 if (Verbose && PrintGCDetails) {
1266 if (stats().valid()) {
1267 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1268 stats().cms_time_since_begin());
1269 } else {
1270 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)");
1271 }
1272 }
1273 return true;
1274 }
1275 }
1276
1277 return false;
1278 }
1279
1280 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1281
1282 // Clear _expansion_cause fields of constituent generations
1283 void CMSCollector::clear_expansion_cause() {
1284 _cmsGen->clear_expansion_cause();
1285 }
1286
1287 // We should be conservative in starting a collection cycle. To
1288 // start too eagerly runs the risk of collecting too often in the
1289 // extreme. To collect too rarely falls back on full collections,
1290 // which works, even if not optimum in terms of concurrent work.
1291 // As a work around for too eagerly collecting, use the flag
1292 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1293 // giving the user an easily understandable way of controlling the
1294 // collections.
1295 // We want to start a new collection cycle if any of the following
1296 // conditions hold:
1297 // . our current occupancy exceeds the configured initiating occupancy
1298 // for this generation, or
1299 // . we recently needed to expand this space and have not, since that
1300 // expansion, done a collection of this generation, or
1301 // . the underlying space believes that it may be a good idea to initiate
1302 // a concurrent collection (this may be based on criteria such as the
1303 // following: the space uses linear allocation and linear allocation is
1304 // going to fail, or there is believed to be excessive fragmentation in
1305 // the generation, etc... or ...
1306 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1307 // the case of the old generation; see CR 6543076):
1308 // we may be approaching a point at which allocation requests may fail because
1309 // we will be out of sufficient free space given allocation rate estimates.]
1310 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1311
1312 assert_lock_strong(freelistLock());
1313 if (occupancy() > initiating_occupancy()) {
1314 if (PrintGCDetails && Verbose) {
1315 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1316 short_name(), occupancy(), initiating_occupancy());
1317 }
1318 return true;
1319 }
1320 if (UseCMSInitiatingOccupancyOnly) {
1321 return false;
1322 }
1323 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1324 if (PrintGCDetails && Verbose) {
1325 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1326 short_name());
1327 }
1328 return true;
1329 }
1330 if (_cmsSpace->should_concurrent_collect()) {
1331 if (PrintGCDetails && Verbose) {
1332 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1333 short_name());
1334 }
1335 return true;
1336 }
1337 return false;
1338 }
1339
1340 void ConcurrentMarkSweepGeneration::collect(bool full,
1341 bool clear_all_soft_refs,
1342 size_t size,
1343 bool tlab)
1344 {
1345 collector()->collect(full, clear_all_soft_refs, size, tlab);
1346 }
1347
1348 void CMSCollector::collect(bool full,
1349 bool clear_all_soft_refs,
1350 size_t size,
1351 bool tlab)
1352 {
1353 // The following "if" branch is present for defensive reasons.
1354 // In the current uses of this interface, it can be replaced with:
1355 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1356 // But I am not placing that assert here to allow future
1357 // generality in invoking this interface.
1358 if (GC_locker::is_active()) {
1359 // A consistency test for GC_locker
1360 assert(GC_locker::needs_gc(), "Should have been set already");
1361 // Skip this foreground collection, instead
1362 // expanding the heap if necessary.
1363 // Need the free list locks for the call to free() in compute_new_size()
1364 compute_new_size();
1365 return;
1366 }
1367 acquire_control_and_collect(full, clear_all_soft_refs);
1368 }
1369
1370 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1371 GenCollectedHeap* gch = GenCollectedHeap::heap();
1372 unsigned int gc_count = gch->total_full_collections();
1373 if (gc_count == full_gc_count) {
1374 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1375 _full_gc_requested = true;
1376 _full_gc_cause = cause;
1377 CGC_lock->notify(); // nudge CMS thread
1378 } else {
1379 assert(gc_count > full_gc_count, "Error: causal loop");
1380 }
1381 }
1382
1383 bool CMSCollector::is_external_interruption() {
1384 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1385 return GCCause::is_user_requested_gc(cause) ||
1386 GCCause::is_serviceability_requested_gc(cause);
1387 }
1388
1389 void CMSCollector::report_concurrent_mode_interruption() {
1390 if (is_external_interruption()) {
1391 if (PrintGCDetails) {
1392 gclog_or_tty->print(" (concurrent mode interrupted)");
1393 }
1394 } else {
1395 if (PrintGCDetails) {
1396 gclog_or_tty->print(" (concurrent mode failure)");
1397 }
1398 _gc_tracer_cm->report_concurrent_mode_failure();
1399 }
1400 }
1401
1402
1403 // The foreground and background collectors need to coordinate in order
1404 // to make sure that they do not mutually interfere with CMS collections.
1405 // When a background collection is active,
1406 // the foreground collector may need to take over (preempt) and
1407 // synchronously complete an ongoing collection. Depending on the
1408 // frequency of the background collections and the heap usage
1409 // of the application, this preemption can be seldom or frequent.
1410 // There are only certain
1411 // points in the background collection that the "collection-baton"
1412 // can be passed to the foreground collector.
1413 //
1414 // The foreground collector will wait for the baton before
1415 // starting any part of the collection. The foreground collector
1416 // will only wait at one location.
1417 //
1418 // The background collector will yield the baton before starting a new
1419 // phase of the collection (e.g., before initial marking, marking from roots,
1420 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1421 // of the loop which switches the phases. The background collector does some
1422 // of the phases (initial mark, final re-mark) with the world stopped.
1423 // Because of locking involved in stopping the world,
1424 // the foreground collector should not block waiting for the background
1425 // collector when it is doing a stop-the-world phase. The background
1426 // collector will yield the baton at an additional point just before
1427 // it enters a stop-the-world phase. Once the world is stopped, the
1428 // background collector checks the phase of the collection. If the
1429 // phase has not changed, it proceeds with the collection. If the
1430 // phase has changed, it skips that phase of the collection. See
1431 // the comments on the use of the Heap_lock in collect_in_background().
1432 //
1433 // Variable used in baton passing.
1434 // _foregroundGCIsActive - Set to true by the foreground collector when
1435 // it wants the baton. The foreground clears it when it has finished
1436 // the collection.
1437 // _foregroundGCShouldWait - Set to true by the background collector
1438 // when it is running. The foreground collector waits while
1439 // _foregroundGCShouldWait is true.
1440 // CGC_lock - monitor used to protect access to the above variables
1441 // and to notify the foreground and background collectors.
1442 // _collectorState - current state of the CMS collection.
1443 //
1444 // The foreground collector
1445 // acquires the CGC_lock
1446 // sets _foregroundGCIsActive
1447 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1448 // various locks acquired in preparation for the collection
1449 // are released so as not to block the background collector
1450 // that is in the midst of a collection
1451 // proceeds with the collection
1452 // clears _foregroundGCIsActive
1453 // returns
1454 //
1455 // The background collector in a loop iterating on the phases of the
1456 // collection
1457 // acquires the CGC_lock
1458 // sets _foregroundGCShouldWait
1459 // if _foregroundGCIsActive is set
1460 // clears _foregroundGCShouldWait, notifies _CGC_lock
1461 // waits on _CGC_lock for _foregroundGCIsActive to become false
1462 // and exits the loop.
1463 // otherwise
1464 // proceed with that phase of the collection
1465 // if the phase is a stop-the-world phase,
1466 // yield the baton once more just before enqueueing
1467 // the stop-world CMS operation (executed by the VM thread).
1468 // returns after all phases of the collection are done
1469 //
1470
1471 void CMSCollector::acquire_control_and_collect(bool full,
1472 bool clear_all_soft_refs) {
1473 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1474 assert(!Thread::current()->is_ConcurrentGC_thread(),
1475 "shouldn't try to acquire control from self!");
1476
1477 // Start the protocol for acquiring control of the
1478 // collection from the background collector (aka CMS thread).
1479 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1480 "VM thread should have CMS token");
1481 // Remember the possibly interrupted state of an ongoing
1482 // concurrent collection
1483 CollectorState first_state = _collectorState;
1484
1485 // Signal to a possibly ongoing concurrent collection that
1486 // we want to do a foreground collection.
1487 _foregroundGCIsActive = true;
1488
1489 // release locks and wait for a notify from the background collector
1490 // releasing the locks in only necessary for phases which
1491 // do yields to improve the granularity of the collection.
1492 assert_lock_strong(bitMapLock());
1493 // We need to lock the Free list lock for the space that we are
1494 // currently collecting.
1495 assert(haveFreelistLocks(), "Must be holding free list locks");
1496 bitMapLock()->unlock();
1497 releaseFreelistLocks();
1498 {
1499 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1500 if (_foregroundGCShouldWait) {
1501 // We are going to be waiting for action for the CMS thread;
1502 // it had better not be gone (for instance at shutdown)!
1503 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1504 "CMS thread must be running");
1505 // Wait here until the background collector gives us the go-ahead
1506 ConcurrentMarkSweepThread::clear_CMS_flag(
1507 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1508 // Get a possibly blocked CMS thread going:
1509 // Note that we set _foregroundGCIsActive true above,
1510 // without protection of the CGC_lock.
1511 CGC_lock->notify();
1512 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1513 "Possible deadlock");
1514 while (_foregroundGCShouldWait) {
1515 // wait for notification
1516 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1517 // Possibility of delay/starvation here, since CMS token does
1518 // not know to give priority to VM thread? Actually, i think
1519 // there wouldn't be any delay/starvation, but the proof of
1520 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1521 }
1522 ConcurrentMarkSweepThread::set_CMS_flag(
1523 ConcurrentMarkSweepThread::CMS_vm_has_token);
1524 }
1525 }
1526 // The CMS_token is already held. Get back the other locks.
1527 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1528 "VM thread should have CMS token");
1529 getFreelistLocks();
1530 bitMapLock()->lock_without_safepoint_check();
1531 if (TraceCMSState) {
1532 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1533 INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state);
1534 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1535 }
1536
1537 // Inform cms gen if this was due to partial collection failing.
1538 // The CMS gen may use this fact to determine its expansion policy.
1539 GenCollectedHeap* gch = GenCollectedHeap::heap();
1540 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1541 assert(!_cmsGen->incremental_collection_failed(),
1542 "Should have been noticed, reacted to and cleared");
1543 _cmsGen->set_incremental_collection_failed();
1544 }
1545
1546 if (first_state > Idling) {
1547 report_concurrent_mode_interruption();
1548 }
1549
1550 set_did_compact(true);
1551
1552 // If the collection is being acquired from the background
1553 // collector, there may be references on the discovered
1554 // references lists. Abandon those references, since some
1555 // of them may have become unreachable after concurrent
1556 // discovery; the STW compacting collector will redo discovery
1557 // more precisely, without being subject to floating garbage.
1558 // Leaving otherwise unreachable references in the discovered
1559 // lists would require special handling.
1560 ref_processor()->disable_discovery();
1561 ref_processor()->abandon_partial_discovery();
1562 ref_processor()->verify_no_references_recorded();
1563
1564 if (first_state > Idling) {
1565 save_heap_summary();
1566 }
1567
1568 do_compaction_work(clear_all_soft_refs);
1569
1570 // Has the GC time limit been exceeded?
1571 size_t max_eden_size = _young_gen->max_capacity() -
1572 _young_gen->to()->capacity() -
1573 _young_gen->from()->capacity();
1574 GCCause::Cause gc_cause = gch->gc_cause();
1575 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1576 _young_gen->eden()->used(),
1577 _cmsGen->max_capacity(),
1578 max_eden_size,
1579 full,
1580 gc_cause,
1581 gch->collector_policy());
1582
1583 // Reset the expansion cause, now that we just completed
1584 // a collection cycle.
1585 clear_expansion_cause();
1586 _foregroundGCIsActive = false;
1587 return;
1588 }
1589
1590 // Resize the tenured generation
1591 // after obtaining the free list locks for the
1592 // two generations.
1593 void CMSCollector::compute_new_size() {
1594 assert_locked_or_safepoint(Heap_lock);
1595 FreelistLocker z(this);
1596 MetaspaceGC::compute_new_size();
1597 _cmsGen->compute_new_size_free_list();
1598 }
1599
1600 // A work method used by the foreground collector to do
1601 // a mark-sweep-compact.
1602 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1603 GenCollectedHeap* gch = GenCollectedHeap::heap();
1604
1605 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1606 gc_timer->register_gc_start();
1607
1608 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1609 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
1610
1611 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id());
1612
1613 // Temporarily widen the span of the weak reference processing to
1614 // the entire heap.
1615 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1616 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1617 // Temporarily, clear the "is_alive_non_header" field of the
1618 // reference processor.
1619 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1620 // Temporarily make reference _processing_ single threaded (non-MT).
1621 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1622 // Temporarily make refs discovery atomic
1623 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1624 // Temporarily make reference _discovery_ single threaded (non-MT)
1625 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1626
1627 ref_processor()->set_enqueuing_is_done(false);
1628 ref_processor()->enable_discovery();
1629 ref_processor()->setup_policy(clear_all_soft_refs);
1630 // If an asynchronous collection finishes, the _modUnionTable is
1631 // all clear. If we are assuming the collection from an asynchronous
1632 // collection, clear the _modUnionTable.
1633 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1634 "_modUnionTable should be clear if the baton was not passed");
1635 _modUnionTable.clear_all();
1636 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
1637 "mod union for klasses should be clear if the baton was passed");
1638 _ct->klass_rem_set()->clear_mod_union();
1639
1640 // We must adjust the allocation statistics being maintained
1641 // in the free list space. We do so by reading and clearing
1642 // the sweep timer and updating the block flux rate estimates below.
1643 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1644 if (_inter_sweep_timer.is_active()) {
1645 _inter_sweep_timer.stop();
1646 // Note that we do not use this sample to update the _inter_sweep_estimate.
1647 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1648 _inter_sweep_estimate.padded_average(),
1649 _intra_sweep_estimate.padded_average());
1650 }
1651
1652 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
1653 ref_processor(), clear_all_soft_refs);
1654 #ifdef ASSERT
1655 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1656 size_t free_size = cms_space->free();
1657 assert(free_size ==
1658 pointer_delta(cms_space->end(), cms_space->compaction_top())
1659 * HeapWordSize,
1660 "All the free space should be compacted into one chunk at top");
1661 assert(cms_space->dictionary()->total_chunk_size(
1662 debug_only(cms_space->freelistLock())) == 0 ||
1663 cms_space->totalSizeInIndexedFreeLists() == 0,
1664 "All the free space should be in a single chunk");
1665 size_t num = cms_space->totalCount();
1666 assert((free_size == 0 && num == 0) ||
1667 (free_size > 0 && (num == 1 || num == 2)),
1668 "There should be at most 2 free chunks after compaction");
1669 #endif // ASSERT
1670 _collectorState = Resetting;
1671 assert(_restart_addr == NULL,
1672 "Should have been NULL'd before baton was passed");
1673 reset(false /* == !concurrent */);
1674 _cmsGen->reset_after_compaction();
1675 _concurrent_cycles_since_last_unload = 0;
1676
1677 // Clear any data recorded in the PLAB chunk arrays.
1678 if (_survivor_plab_array != NULL) {
1679 reset_survivor_plab_arrays();
1680 }
1681
1682 // Adjust the per-size allocation stats for the next epoch.
1683 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1684 // Restart the "inter sweep timer" for the next epoch.
1685 _inter_sweep_timer.reset();
1686 _inter_sweep_timer.start();
1687
1688 gc_timer->register_gc_end();
1689
1690 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1691
1692 // For a mark-sweep-compact, compute_new_size() will be called
1693 // in the heap's do_collection() method.
1694 }
1695
1696 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1697 ContiguousSpace* eden_space = _young_gen->eden();
1698 ContiguousSpace* from_space = _young_gen->from();
1699 ContiguousSpace* to_space = _young_gen->to();
1700 // Eden
1701 if (_eden_chunk_array != NULL) {
1702 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1703 p2i(eden_space->bottom()), p2i(eden_space->top()),
1704 p2i(eden_space->end()), eden_space->capacity());
1705 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", "
1706 "_eden_chunk_capacity=" SIZE_FORMAT,
1707 _eden_chunk_index, _eden_chunk_capacity);
1708 for (size_t i = 0; i < _eden_chunk_index; i++) {
1709 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1710 i, p2i(_eden_chunk_array[i]));
1711 }
1712 }
1713 // Survivor
1714 if (_survivor_chunk_array != NULL) {
1715 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1716 p2i(from_space->bottom()), p2i(from_space->top()),
1717 p2i(from_space->end()), from_space->capacity());
1718 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", "
1719 "_survivor_chunk_capacity=" SIZE_FORMAT,
1720 _survivor_chunk_index, _survivor_chunk_capacity);
1721 for (size_t i = 0; i < _survivor_chunk_index; i++) {
1722 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1723 i, p2i(_survivor_chunk_array[i]));
1724 }
1725 }
1726 }
1727
1728 void CMSCollector::getFreelistLocks() const {
1729 // Get locks for all free lists in all generations that this
1730 // collector is responsible for
1731 _cmsGen->freelistLock()->lock_without_safepoint_check();
1732 }
1733
1734 void CMSCollector::releaseFreelistLocks() const {
1735 // Release locks for all free lists in all generations that this
1736 // collector is responsible for
1737 _cmsGen->freelistLock()->unlock();
1738 }
1739
1740 bool CMSCollector::haveFreelistLocks() const {
1741 // Check locks for all free lists in all generations that this
1742 // collector is responsible for
1743 assert_lock_strong(_cmsGen->freelistLock());
1744 PRODUCT_ONLY(ShouldNotReachHere());
1745 return true;
1746 }
1747
1748 // A utility class that is used by the CMS collector to
1749 // temporarily "release" the foreground collector from its
1750 // usual obligation to wait for the background collector to
1751 // complete an ongoing phase before proceeding.
1752 class ReleaseForegroundGC: public StackObj {
1753 private:
1754 CMSCollector* _c;
1755 public:
1756 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1757 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1758 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1759 // allow a potentially blocked foreground collector to proceed
1760 _c->_foregroundGCShouldWait = false;
1761 if (_c->_foregroundGCIsActive) {
1762 CGC_lock->notify();
1763 }
1764 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1765 "Possible deadlock");
1766 }
1767
1768 ~ReleaseForegroundGC() {
1769 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1770 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1771 _c->_foregroundGCShouldWait = true;
1772 }
1773 };
1774
1775 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1776 assert(Thread::current()->is_ConcurrentGC_thread(),
1777 "A CMS asynchronous collection is only allowed on a CMS thread.");
1778
1779 GenCollectedHeap* gch = GenCollectedHeap::heap();
1780 {
1781 bool safepoint_check = Mutex::_no_safepoint_check_flag;
1782 MutexLockerEx hl(Heap_lock, safepoint_check);
1783 FreelistLocker fll(this);
1784 MutexLockerEx x(CGC_lock, safepoint_check);
1785 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
1786 // The foreground collector is active or we're
1787 // not using asynchronous collections. Skip this
1788 // background collection.
1789 assert(!_foregroundGCShouldWait, "Should be clear");
1790 return;
1791 } else {
1792 assert(_collectorState == Idling, "Should be idling before start.");
1793 _collectorState = InitialMarking;
1794 register_gc_start(cause);
1795 // Reset the expansion cause, now that we are about to begin
1796 // a new cycle.
1797 clear_expansion_cause();
1798
1799 // Clear the MetaspaceGC flag since a concurrent collection
1800 // is starting but also clear it after the collection.
1801 MetaspaceGC::set_should_concurrent_collect(false);
1802 }
1803 // Decide if we want to enable class unloading as part of the
1804 // ensuing concurrent GC cycle.
1805 update_should_unload_classes();
1806 _full_gc_requested = false; // acks all outstanding full gc requests
1807 _full_gc_cause = GCCause::_no_gc;
1808 // Signal that we are about to start a collection
1809 gch->increment_total_full_collections(); // ... starting a collection cycle
1810 _collection_count_start = gch->total_full_collections();
1811 }
1812
1813 // Used for PrintGC
1814 size_t prev_used;
1815 if (PrintGC && Verbose) {
1816 prev_used = _cmsGen->used();
1817 }
1818
1819 // The change of the collection state is normally done at this level;
1820 // the exceptions are phases that are executed while the world is
1821 // stopped. For those phases the change of state is done while the
1822 // world is stopped. For baton passing purposes this allows the
1823 // background collector to finish the phase and change state atomically.
1824 // The foreground collector cannot wait on a phase that is done
1825 // while the world is stopped because the foreground collector already
1826 // has the world stopped and would deadlock.
1827 while (_collectorState != Idling) {
1828 if (TraceCMSState) {
1829 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
1830 p2i(Thread::current()), _collectorState);
1831 }
1832 // The foreground collector
1833 // holds the Heap_lock throughout its collection.
1834 // holds the CMS token (but not the lock)
1835 // except while it is waiting for the background collector to yield.
1836 //
1837 // The foreground collector should be blocked (not for long)
1838 // if the background collector is about to start a phase
1839 // executed with world stopped. If the background
1840 // collector has already started such a phase, the
1841 // foreground collector is blocked waiting for the
1842 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
1843 // are executed in the VM thread.
1844 //
1845 // The locking order is
1846 // PendingListLock (PLL) -- if applicable (FinalMarking)
1847 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
1848 // CMS token (claimed in
1849 // stop_world_and_do() -->
1850 // safepoint_synchronize() -->
1851 // CMSThread::synchronize())
1852
1853 {
1854 // Check if the FG collector wants us to yield.
1855 CMSTokenSync x(true); // is cms thread
1856 if (waitForForegroundGC()) {
1857 // We yielded to a foreground GC, nothing more to be
1858 // done this round.
1859 assert(_foregroundGCShouldWait == false, "We set it to false in "
1860 "waitForForegroundGC()");
1861 if (TraceCMSState) {
1862 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1863 " exiting collection CMS state %d",
1864 p2i(Thread::current()), _collectorState);
1865 }
1866 return;
1867 } else {
1868 // The background collector can run but check to see if the
1869 // foreground collector has done a collection while the
1870 // background collector was waiting to get the CGC_lock
1871 // above. If yes, break so that _foregroundGCShouldWait
1872 // is cleared before returning.
1873 if (_collectorState == Idling) {
1874 break;
1875 }
1876 }
1877 }
1878
1879 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1880 "should be waiting");
1881
1882 switch (_collectorState) {
1883 case InitialMarking:
1884 {
1885 ReleaseForegroundGC x(this);
1886 stats().record_cms_begin();
1887 VM_CMS_Initial_Mark initial_mark_op(this);
1888 VMThread::execute(&initial_mark_op);
1889 }
1890 // The collector state may be any legal state at this point
1891 // since the background collector may have yielded to the
1892 // foreground collector.
1893 break;
1894 case Marking:
1895 // initial marking in checkpointRootsInitialWork has been completed
1896 if (markFromRoots()) { // we were successful
1897 assert(_collectorState == Precleaning, "Collector state should "
1898 "have changed");
1899 } else {
1900 assert(_foregroundGCIsActive, "Internal state inconsistency");
1901 }
1902 break;
1903 case Precleaning:
1904 // marking from roots in markFromRoots has been completed
1905 preclean();
1906 assert(_collectorState == AbortablePreclean ||
1907 _collectorState == FinalMarking,
1908 "Collector state should have changed");
1909 break;
1910 case AbortablePreclean:
1911 abortable_preclean();
1912 assert(_collectorState == FinalMarking, "Collector state should "
1913 "have changed");
1914 break;
1915 case FinalMarking:
1916 {
1917 ReleaseForegroundGC x(this);
1918
1919 VM_CMS_Final_Remark final_remark_op(this);
1920 VMThread::execute(&final_remark_op);
1921 }
1922 assert(_foregroundGCShouldWait, "block post-condition");
1923 break;
1924 case Sweeping:
1925 // final marking in checkpointRootsFinal has been completed
1926 sweep();
1927 assert(_collectorState == Resizing, "Collector state change "
1928 "to Resizing must be done under the free_list_lock");
1929
1930 case Resizing: {
1931 // Sweeping has been completed...
1932 // At this point the background collection has completed.
1933 // Don't move the call to compute_new_size() down
1934 // into code that might be executed if the background
1935 // collection was preempted.
1936 {
1937 ReleaseForegroundGC x(this); // unblock FG collection
1938 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
1939 CMSTokenSync z(true); // not strictly needed.
1940 if (_collectorState == Resizing) {
1941 compute_new_size();
1942 save_heap_summary();
1943 _collectorState = Resetting;
1944 } else {
1945 assert(_collectorState == Idling, "The state should only change"
1946 " because the foreground collector has finished the collection");
1947 }
1948 }
1949 break;
1950 }
1951 case Resetting:
1952 // CMS heap resizing has been completed
1953 reset(true);
1954 assert(_collectorState == Idling, "Collector state should "
1955 "have changed");
1956
1957 MetaspaceGC::set_should_concurrent_collect(false);
1958
1959 stats().record_cms_end();
1960 // Don't move the concurrent_phases_end() and compute_new_size()
1961 // calls to here because a preempted background collection
1962 // has it's state set to "Resetting".
1963 break;
1964 case Idling:
1965 default:
1966 ShouldNotReachHere();
1967 break;
1968 }
1969 if (TraceCMSState) {
1970 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
1971 p2i(Thread::current()), _collectorState);
1972 }
1973 assert(_foregroundGCShouldWait, "block post-condition");
1974 }
1975
1976 // Should this be in gc_epilogue?
1977 collector_policy()->counters()->update_counters();
1978
1979 {
1980 // Clear _foregroundGCShouldWait and, in the event that the
1981 // foreground collector is waiting, notify it, before
1982 // returning.
1983 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1984 _foregroundGCShouldWait = false;
1985 if (_foregroundGCIsActive) {
1986 CGC_lock->notify();
1987 }
1988 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1989 "Possible deadlock");
1990 }
1991 if (TraceCMSState) {
1992 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1993 " exiting collection CMS state %d",
1994 p2i(Thread::current()), _collectorState);
1995 }
1996 if (PrintGC && Verbose) {
1997 _cmsGen->print_heap_change(prev_used);
1998 }
1999 }
2000
2001 void CMSCollector::register_gc_start(GCCause::Cause cause) {
2002 _cms_start_registered = true;
2003 _gc_timer_cm->register_gc_start();
2004 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
2005 }
2006
2007 void CMSCollector::register_gc_end() {
2008 if (_cms_start_registered) {
2009 report_heap_summary(GCWhen::AfterGC);
2010
2011 _gc_timer_cm->register_gc_end();
2012 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
2013 _cms_start_registered = false;
2014 }
2015 }
2016
2017 void CMSCollector::save_heap_summary() {
2018 GenCollectedHeap* gch = GenCollectedHeap::heap();
2019 _last_heap_summary = gch->create_heap_summary();
2020 _last_metaspace_summary = gch->create_metaspace_summary();
2021 }
2022
2023 void CMSCollector::report_heap_summary(GCWhen::Type when) {
2024 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
2025 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
2026 }
2027
2028 bool CMSCollector::waitForForegroundGC() {
2029 bool res = false;
2030 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2031 "CMS thread should have CMS token");
2032 // Block the foreground collector until the
2033 // background collectors decides whether to
2034 // yield.
2035 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2036 _foregroundGCShouldWait = true;
2037 if (_foregroundGCIsActive) {
2038 // The background collector yields to the
2039 // foreground collector and returns a value
2040 // indicating that it has yielded. The foreground
2041 // collector can proceed.
2042 res = true;
2043 _foregroundGCShouldWait = false;
2044 ConcurrentMarkSweepThread::clear_CMS_flag(
2045 ConcurrentMarkSweepThread::CMS_cms_has_token);
2046 ConcurrentMarkSweepThread::set_CMS_flag(
2047 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2048 // Get a possibly blocked foreground thread going
2049 CGC_lock->notify();
2050 if (TraceCMSState) {
2051 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2052 p2i(Thread::current()), _collectorState);
2053 }
2054 while (_foregroundGCIsActive) {
2055 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2056 }
2057 ConcurrentMarkSweepThread::set_CMS_flag(
2058 ConcurrentMarkSweepThread::CMS_cms_has_token);
2059 ConcurrentMarkSweepThread::clear_CMS_flag(
2060 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2061 }
2062 if (TraceCMSState) {
2063 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2064 p2i(Thread::current()), _collectorState);
2065 }
2066 return res;
2067 }
2068
2069 // Because of the need to lock the free lists and other structures in
2070 // the collector, common to all the generations that the collector is
2071 // collecting, we need the gc_prologues of individual CMS generations
2072 // delegate to their collector. It may have been simpler had the
2073 // current infrastructure allowed one to call a prologue on a
2074 // collector. In the absence of that we have the generation's
2075 // prologue delegate to the collector, which delegates back
2076 // some "local" work to a worker method in the individual generations
2077 // that it's responsible for collecting, while itself doing any
2078 // work common to all generations it's responsible for. A similar
2079 // comment applies to the gc_epilogue()'s.
2080 // The role of the variable _between_prologue_and_epilogue is to
2081 // enforce the invocation protocol.
2082 void CMSCollector::gc_prologue(bool full) {
2083 // Call gc_prologue_work() for the CMSGen
2084 // we are responsible for.
2085
2086 // The following locking discipline assumes that we are only called
2087 // when the world is stopped.
2088 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2089
2090 // The CMSCollector prologue must call the gc_prologues for the
2091 // "generations" that it's responsible
2092 // for.
2093
2094 assert( Thread::current()->is_VM_thread()
2095 || ( CMSScavengeBeforeRemark
2096 && Thread::current()->is_ConcurrentGC_thread()),
2097 "Incorrect thread type for prologue execution");
2098
2099 if (_between_prologue_and_epilogue) {
2100 // We have already been invoked; this is a gc_prologue delegation
2101 // from yet another CMS generation that we are responsible for, just
2102 // ignore it since all relevant work has already been done.
2103 return;
2104 }
2105
2106 // set a bit saying prologue has been called; cleared in epilogue
2107 _between_prologue_and_epilogue = true;
2108 // Claim locks for common data structures, then call gc_prologue_work()
2109 // for each CMSGen.
2110
2111 getFreelistLocks(); // gets free list locks on constituent spaces
2112 bitMapLock()->lock_without_safepoint_check();
2113
2114 // Should call gc_prologue_work() for all cms gens we are responsible for
2115 bool duringMarking = _collectorState >= Marking
2116 && _collectorState < Sweeping;
2117
2118 // The young collections clear the modified oops state, which tells if
2119 // there are any modified oops in the class. The remark phase also needs
2120 // that information. Tell the young collection to save the union of all
2121 // modified klasses.
2122 if (duringMarking) {
2123 _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2124 }
2125
2126 bool registerClosure = duringMarking;
2127
2128 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2129
2130 if (!full) {
2131 stats().record_gc0_begin();
2132 }
2133 }
2134
2135 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2136
2137 _capacity_at_prologue = capacity();
2138 _used_at_prologue = used();
2139
2140 // Delegate to CMScollector which knows how to coordinate between
2141 // this and any other CMS generations that it is responsible for
2142 // collecting.
2143 collector()->gc_prologue(full);
2144 }
2145
2146 // This is a "private" interface for use by this generation's CMSCollector.
2147 // Not to be called directly by any other entity (for instance,
2148 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2149 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2150 bool registerClosure, ModUnionClosure* modUnionClosure) {
2151 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2152 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2153 "Should be NULL");
2154 if (registerClosure) {
2155 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2156 }
2157 cmsSpace()->gc_prologue();
2158 // Clear stat counters
2159 NOT_PRODUCT(
2160 assert(_numObjectsPromoted == 0, "check");
2161 assert(_numWordsPromoted == 0, "check");
2162 if (Verbose && PrintGC) {
2163 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2164 SIZE_FORMAT" bytes concurrently",
2165 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2166 }
2167 _numObjectsAllocated = 0;
2168 _numWordsAllocated = 0;
2169 )
2170 }
2171
2172 void CMSCollector::gc_epilogue(bool full) {
2173 // The following locking discipline assumes that we are only called
2174 // when the world is stopped.
2175 assert(SafepointSynchronize::is_at_safepoint(),
2176 "world is stopped assumption");
2177
2178 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2179 // if linear allocation blocks need to be appropriately marked to allow the
2180 // the blocks to be parsable. We also check here whether we need to nudge the
2181 // CMS collector thread to start a new cycle (if it's not already active).
2182 assert( Thread::current()->is_VM_thread()
2183 || ( CMSScavengeBeforeRemark
2184 && Thread::current()->is_ConcurrentGC_thread()),
2185 "Incorrect thread type for epilogue execution");
2186
2187 if (!_between_prologue_and_epilogue) {
2188 // We have already been invoked; this is a gc_epilogue delegation
2189 // from yet another CMS generation that we are responsible for, just
2190 // ignore it since all relevant work has already been done.
2191 return;
2192 }
2193 assert(haveFreelistLocks(), "must have freelist locks");
2194 assert_lock_strong(bitMapLock());
2195
2196 _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2197
2198 _cmsGen->gc_epilogue_work(full);
2199
2200 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2201 // in case sampling was not already enabled, enable it
2202 _start_sampling = true;
2203 }
2204 // reset _eden_chunk_array so sampling starts afresh
2205 _eden_chunk_index = 0;
2206
2207 size_t cms_used = _cmsGen->cmsSpace()->used();
2208
2209 // update performance counters - this uses a special version of
2210 // update_counters() that allows the utilization to be passed as a
2211 // parameter, avoiding multiple calls to used().
2212 //
2213 _cmsGen->update_counters(cms_used);
2214
2215 bitMapLock()->unlock();
2216 releaseFreelistLocks();
2217
2218 if (!CleanChunkPoolAsync) {
2219 Chunk::clean_chunk_pool();
2220 }
2221
2222 set_did_compact(false);
2223 _between_prologue_and_epilogue = false; // ready for next cycle
2224 }
2225
2226 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2227 collector()->gc_epilogue(full);
2228
2229 // Also reset promotion tracking in par gc thread states.
2230 for (uint i = 0; i < ParallelGCThreads; i++) {
2231 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2232 }
2233 }
2234
2235 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2236 assert(!incremental_collection_failed(), "Should have been cleared");
2237 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2238 cmsSpace()->gc_epilogue();
2239 // Print stat counters
2240 NOT_PRODUCT(
2241 assert(_numObjectsAllocated == 0, "check");
2242 assert(_numWordsAllocated == 0, "check");
2243 if (Verbose && PrintGC) {
2244 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2245 SIZE_FORMAT" bytes",
2246 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2247 }
2248 _numObjectsPromoted = 0;
2249 _numWordsPromoted = 0;
2250 )
2251
2252 if (PrintGC && Verbose) {
2253 // Call down the chain in contiguous_available needs the freelistLock
2254 // so print this out before releasing the freeListLock.
2255 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2256 contiguous_available());
2257 }
2258 }
2259
2260 #ifndef PRODUCT
2261 bool CMSCollector::have_cms_token() {
2262 Thread* thr = Thread::current();
2263 if (thr->is_VM_thread()) {
2264 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2265 } else if (thr->is_ConcurrentGC_thread()) {
2266 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2267 } else if (thr->is_GC_task_thread()) {
2268 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2269 ParGCRareEvent_lock->owned_by_self();
2270 }
2271 return false;
2272 }
2273 #endif
2274
2275 // Check reachability of the given heap address in CMS generation,
2276 // treating all other generations as roots.
2277 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2278 // We could "guarantee" below, rather than assert, but I'll
2279 // leave these as "asserts" so that an adventurous debugger
2280 // could try this in the product build provided some subset of
2281 // the conditions were met, provided they were interested in the
2282 // results and knew that the computation below wouldn't interfere
2283 // with other concurrent computations mutating the structures
2284 // being read or written.
2285 assert(SafepointSynchronize::is_at_safepoint(),
2286 "Else mutations in object graph will make answer suspect");
2287 assert(have_cms_token(), "Should hold cms token");
2288 assert(haveFreelistLocks(), "must hold free list locks");
2289 assert_lock_strong(bitMapLock());
2290
2291 // Clear the marking bit map array before starting, but, just
2292 // for kicks, first report if the given address is already marked
2293 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr),
2294 _markBitMap.isMarked(addr) ? "" : " not");
2295
2296 if (verify_after_remark()) {
2297 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2298 bool result = verification_mark_bm()->isMarked(addr);
2299 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr),
2300 result ? "IS" : "is NOT");
2301 return result;
2302 } else {
2303 gclog_or_tty->print_cr("Could not compute result");
2304 return false;
2305 }
2306 }
2307
2308
2309 void
2310 CMSCollector::print_on_error(outputStream* st) {
2311 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2312 if (collector != NULL) {
2313 CMSBitMap* bitmap = &collector->_markBitMap;
2314 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2315 bitmap->print_on_error(st, " Bits: ");
2316
2317 st->cr();
2318
2319 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2320 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2321 mut_bitmap->print_on_error(st, " Bits: ");
2322 }
2323 }
2324
2325 ////////////////////////////////////////////////////////
2326 // CMS Verification Support
2327 ////////////////////////////////////////////////////////
2328 // Following the remark phase, the following invariant
2329 // should hold -- each object in the CMS heap which is
2330 // marked in markBitMap() should be marked in the verification_mark_bm().
2331
2332 class VerifyMarkedClosure: public BitMapClosure {
2333 CMSBitMap* _marks;
2334 bool _failed;
2335
2336 public:
2337 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2338
2339 bool do_bit(size_t offset) {
2340 HeapWord* addr = _marks->offsetToHeapWord(offset);
2341 if (!_marks->isMarked(addr)) {
2342 oop(addr)->print_on(gclog_or_tty);
2343 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", p2i(addr));
2344 _failed = true;
2345 }
2346 return true;
2347 }
2348
2349 bool failed() { return _failed; }
2350 };
2351
2352 bool CMSCollector::verify_after_remark(bool silent) {
2353 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... ");
2354 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2355 static bool init = false;
2356
2357 assert(SafepointSynchronize::is_at_safepoint(),
2358 "Else mutations in object graph will make answer suspect");
2359 assert(have_cms_token(),
2360 "Else there may be mutual interference in use of "
2361 " verification data structures");
2362 assert(_collectorState > Marking && _collectorState <= Sweeping,
2363 "Else marking info checked here may be obsolete");
2364 assert(haveFreelistLocks(), "must hold free list locks");
2365 assert_lock_strong(bitMapLock());
2366
2367
2368 // Allocate marking bit map if not already allocated
2369 if (!init) { // first time
2370 if (!verification_mark_bm()->allocate(_span)) {
2371 return false;
2372 }
2373 init = true;
2374 }
2375
2376 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2377
2378 // Turn off refs discovery -- so we will be tracing through refs.
2379 // This is as intended, because by this time
2380 // GC must already have cleared any refs that need to be cleared,
2381 // and traced those that need to be marked; moreover,
2382 // the marking done here is not going to interfere in any
2383 // way with the marking information used by GC.
2384 NoRefDiscovery no_discovery(ref_processor());
2385
2386 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2387
2388 // Clear any marks from a previous round
2389 verification_mark_bm()->clear_all();
2390 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2391 verify_work_stacks_empty();
2392
2393 GenCollectedHeap* gch = GenCollectedHeap::heap();
2394 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2395 // Update the saved marks which may affect the root scans.
2396 gch->save_marks();
2397
2398 if (CMSRemarkVerifyVariant == 1) {
2399 // In this first variant of verification, we complete
2400 // all marking, then check if the new marks-vector is
2401 // a subset of the CMS marks-vector.
2402 verify_after_remark_work_1();
2403 } else if (CMSRemarkVerifyVariant == 2) {
2404 // In this second variant of verification, we flag an error
2405 // (i.e. an object reachable in the new marks-vector not reachable
2406 // in the CMS marks-vector) immediately, also indicating the
2407 // identify of an object (A) that references the unmarked object (B) --
2408 // presumably, a mutation to A failed to be picked up by preclean/remark?
2409 verify_after_remark_work_2();
2410 } else {
2411 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant",
2412 CMSRemarkVerifyVariant);
2413 }
2414 if (!silent) gclog_or_tty->print(" done] ");
2415 return true;
2416 }
2417
2418 void CMSCollector::verify_after_remark_work_1() {
2419 ResourceMark rm;
2420 HandleMark hm;
2421 GenCollectedHeap* gch = GenCollectedHeap::heap();
2422
2423 // Get a clear set of claim bits for the roots processing to work with.
2424 ClassLoaderDataGraph::clear_claimed_marks();
2425
2426 // Mark from roots one level into CMS
2427 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2428 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2429
2430 {
2431 StrongRootsScope srs(1);
2432
2433 gch->gen_process_roots(&srs,
2434 _cmsGen->level(),
2435 true, // younger gens are roots
2436 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2437 should_unload_classes(),
2438 ¬Older,
2439 NULL,
2440 NULL);
2441 }
2442
2443 // Now mark from the roots
2444 MarkFromRootsClosure markFromRootsClosure(this, _span,
2445 verification_mark_bm(), verification_mark_stack(),
2446 false /* don't yield */, true /* verifying */);
2447 assert(_restart_addr == NULL, "Expected pre-condition");
2448 verification_mark_bm()->iterate(&markFromRootsClosure);
2449 while (_restart_addr != NULL) {
2450 // Deal with stack overflow: by restarting at the indicated
2451 // address.
2452 HeapWord* ra = _restart_addr;
2453 markFromRootsClosure.reset(ra);
2454 _restart_addr = NULL;
2455 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2456 }
2457 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2458 verify_work_stacks_empty();
2459
2460 // Marking completed -- now verify that each bit marked in
2461 // verification_mark_bm() is also marked in markBitMap(); flag all
2462 // errors by printing corresponding objects.
2463 VerifyMarkedClosure vcl(markBitMap());
2464 verification_mark_bm()->iterate(&vcl);
2465 if (vcl.failed()) {
2466 gclog_or_tty->print("Verification failed");
2467 gch->print_on(gclog_or_tty);
2468 fatal("CMS: failed marking verification after remark");
2469 }
2470 }
2471
2472 class VerifyKlassOopsKlassClosure : public KlassClosure {
2473 class VerifyKlassOopsClosure : public OopClosure {
2474 CMSBitMap* _bitmap;
2475 public:
2476 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2477 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2478 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2479 } _oop_closure;
2480 public:
2481 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2482 void do_klass(Klass* k) {
2483 k->oops_do(&_oop_closure);
2484 }
2485 };
2486
2487 void CMSCollector::verify_after_remark_work_2() {
2488 ResourceMark rm;
2489 HandleMark hm;
2490 GenCollectedHeap* gch = GenCollectedHeap::heap();
2491
2492 // Get a clear set of claim bits for the roots processing to work with.
2493 ClassLoaderDataGraph::clear_claimed_marks();
2494
2495 // Mark from roots one level into CMS
2496 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2497 markBitMap());
2498 CLDToOopClosure cld_closure(¬Older, true);
2499
2500 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2501
2502 {
2503 StrongRootsScope srs(1);
2504
2505 gch->gen_process_roots(&srs,
2506 _cmsGen->level(),
2507 true, // younger gens are roots
2508 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2509 should_unload_classes(),
2510 ¬Older,
2511 NULL,
2512 &cld_closure);
2513 }
2514
2515 // Now mark from the roots
2516 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2517 verification_mark_bm(), markBitMap(), verification_mark_stack());
2518 assert(_restart_addr == NULL, "Expected pre-condition");
2519 verification_mark_bm()->iterate(&markFromRootsClosure);
2520 while (_restart_addr != NULL) {
2521 // Deal with stack overflow: by restarting at the indicated
2522 // address.
2523 HeapWord* ra = _restart_addr;
2524 markFromRootsClosure.reset(ra);
2525 _restart_addr = NULL;
2526 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2527 }
2528 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2529 verify_work_stacks_empty();
2530
2531 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
2532 ClassLoaderDataGraph::classes_do(&verify_klass_oops);
2533
2534 // Marking completed -- now verify that each bit marked in
2535 // verification_mark_bm() is also marked in markBitMap(); flag all
2536 // errors by printing corresponding objects.
2537 VerifyMarkedClosure vcl(markBitMap());
2538 verification_mark_bm()->iterate(&vcl);
2539 assert(!vcl.failed(), "Else verification above should not have succeeded");
2540 }
2541
2542 void ConcurrentMarkSweepGeneration::save_marks() {
2543 // delegate to CMS space
2544 cmsSpace()->save_marks();
2545 for (uint i = 0; i < ParallelGCThreads; i++) {
2546 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2547 }
2548 }
2549
2550 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2551 return cmsSpace()->no_allocs_since_save_marks();
2552 }
2553
2554 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2555 \
2556 void ConcurrentMarkSweepGeneration:: \
2557 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2558 cl->set_generation(this); \
2559 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2560 cl->reset_generation(); \
2561 save_marks(); \
2562 }
2563
2564 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2565
2566 void
2567 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2568 if (freelistLock()->owned_by_self()) {
2569 Generation::oop_iterate(cl);
2570 } else {
2571 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2572 Generation::oop_iterate(cl);
2573 }
2574 }
2575
2576 void
2577 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2578 if (freelistLock()->owned_by_self()) {
2579 Generation::object_iterate(cl);
2580 } else {
2581 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2582 Generation::object_iterate(cl);
2583 }
2584 }
2585
2586 void
2587 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2588 if (freelistLock()->owned_by_self()) {
2589 Generation::safe_object_iterate(cl);
2590 } else {
2591 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2592 Generation::safe_object_iterate(cl);
2593 }
2594 }
2595
2596 void
2597 ConcurrentMarkSweepGeneration::post_compact() {
2598 }
2599
2600 void
2601 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2602 // Fix the linear allocation blocks to look like free blocks.
2603
2604 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2605 // are not called when the heap is verified during universe initialization and
2606 // at vm shutdown.
2607 if (freelistLock()->owned_by_self()) {
2608 cmsSpace()->prepare_for_verify();
2609 } else {
2610 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2611 cmsSpace()->prepare_for_verify();
2612 }
2613 }
2614
2615 void
2616 ConcurrentMarkSweepGeneration::verify() {
2617 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2618 // are not called when the heap is verified during universe initialization and
2619 // at vm shutdown.
2620 if (freelistLock()->owned_by_self()) {
2621 cmsSpace()->verify();
2622 } else {
2623 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2624 cmsSpace()->verify();
2625 }
2626 }
2627
2628 void CMSCollector::verify() {
2629 _cmsGen->verify();
2630 }
2631
2632 #ifndef PRODUCT
2633 bool CMSCollector::overflow_list_is_empty() const {
2634 assert(_num_par_pushes >= 0, "Inconsistency");
2635 if (_overflow_list == NULL) {
2636 assert(_num_par_pushes == 0, "Inconsistency");
2637 }
2638 return _overflow_list == NULL;
2639 }
2640
2641 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2642 // merely consolidate assertion checks that appear to occur together frequently.
2643 void CMSCollector::verify_work_stacks_empty() const {
2644 assert(_markStack.isEmpty(), "Marking stack should be empty");
2645 assert(overflow_list_is_empty(), "Overflow list should be empty");
2646 }
2647
2648 void CMSCollector::verify_overflow_empty() const {
2649 assert(overflow_list_is_empty(), "Overflow list should be empty");
2650 assert(no_preserved_marks(), "No preserved marks");
2651 }
2652 #endif // PRODUCT
2653
2654 // Decide if we want to enable class unloading as part of the
2655 // ensuing concurrent GC cycle. We will collect and
2656 // unload classes if it's the case that:
2657 // (1) an explicit gc request has been made and the flag
2658 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
2659 // (2) (a) class unloading is enabled at the command line, and
2660 // (b) old gen is getting really full
2661 // NOTE: Provided there is no change in the state of the heap between
2662 // calls to this method, it should have idempotent results. Moreover,
2663 // its results should be monotonically increasing (i.e. going from 0 to 1,
2664 // but not 1 to 0) between successive calls between which the heap was
2665 // not collected. For the implementation below, it must thus rely on
2666 // the property that concurrent_cycles_since_last_unload()
2667 // will not decrease unless a collection cycle happened and that
2668 // _cmsGen->is_too_full() are
2669 // themselves also monotonic in that sense. See check_monotonicity()
2670 // below.
2671 void CMSCollector::update_should_unload_classes() {
2672 _should_unload_classes = false;
2673 // Condition 1 above
2674 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
2675 _should_unload_classes = true;
2676 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
2677 // Disjuncts 2.b.(i,ii,iii) above
2678 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2679 CMSClassUnloadingMaxInterval)
2680 || _cmsGen->is_too_full();
2681 }
2682 }
2683
2684 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2685 bool res = should_concurrent_collect();
2686 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2687 return res;
2688 }
2689
2690 void CMSCollector::setup_cms_unloading_and_verification_state() {
2691 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2692 || VerifyBeforeExit;
2693 const int rso = GenCollectedHeap::SO_AllCodeCache;
2694
2695 // We set the proper root for this CMS cycle here.
2696 if (should_unload_classes()) { // Should unload classes this cycle
2697 remove_root_scanning_option(rso); // Shrink the root set appropriately
2698 set_verifying(should_verify); // Set verification state for this cycle
2699 return; // Nothing else needs to be done at this time
2700 }
2701
2702 // Not unloading classes this cycle
2703 assert(!should_unload_classes(), "Inconsistency!");
2704
2705 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2706 // Include symbols, strings and code cache elements to prevent their resurrection.
2707 add_root_scanning_option(rso);
2708 set_verifying(true);
2709 } else if (verifying() && !should_verify) {
2710 // We were verifying, but some verification flags got disabled.
2711 set_verifying(false);
2712 // Exclude symbols, strings and code cache elements from root scanning to
2713 // reduce IM and RM pauses.
2714 remove_root_scanning_option(rso);
2715 }
2716 }
2717
2718
2719 #ifndef PRODUCT
2720 HeapWord* CMSCollector::block_start(const void* p) const {
2721 const HeapWord* addr = (HeapWord*)p;
2722 if (_span.contains(p)) {
2723 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2724 return _cmsGen->cmsSpace()->block_start(p);
2725 }
2726 }
2727 return NULL;
2728 }
2729 #endif
2730
2731 HeapWord*
2732 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2733 bool tlab,
2734 bool parallel) {
2735 CMSSynchronousYieldRequest yr;
2736 assert(!tlab, "Can't deal with TLAB allocation");
2737 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2738 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2739 if (GCExpandToAllocateDelayMillis > 0) {
2740 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2741 }
2742 return have_lock_and_allocate(word_size, tlab);
2743 }
2744
2745 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2746 size_t bytes,
2747 size_t expand_bytes,
2748 CMSExpansionCause::Cause cause)
2749 {
2750
2751 bool success = expand(bytes, expand_bytes);
2752
2753 // remember why we expanded; this information is used
2754 // by shouldConcurrentCollect() when making decisions on whether to start
2755 // a new CMS cycle.
2756 if (success) {
2757 set_expansion_cause(cause);
2758 if (PrintGCDetails && Verbose) {
2759 gclog_or_tty->print_cr("Expanded CMS gen for %s",
2760 CMSExpansionCause::to_string(cause));
2761 }
2762 }
2763 }
2764
2765 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2766 HeapWord* res = NULL;
2767 MutexLocker x(ParGCRareEvent_lock);
2768 while (true) {
2769 // Expansion by some other thread might make alloc OK now:
2770 res = ps->lab.alloc(word_sz);
2771 if (res != NULL) return res;
2772 // If there's not enough expansion space available, give up.
2773 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2774 return NULL;
2775 }
2776 // Otherwise, we try expansion.
2777 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2778 // Now go around the loop and try alloc again;
2779 // A competing par_promote might beat us to the expansion space,
2780 // so we may go around the loop again if promotion fails again.
2781 if (GCExpandToAllocateDelayMillis > 0) {
2782 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2783 }
2784 }
2785 }
2786
2787
2788 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2789 PromotionInfo* promo) {
2790 MutexLocker x(ParGCRareEvent_lock);
2791 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2792 while (true) {
2793 // Expansion by some other thread might make alloc OK now:
2794 if (promo->ensure_spooling_space()) {
2795 assert(promo->has_spooling_space(),
2796 "Post-condition of successful ensure_spooling_space()");
2797 return true;
2798 }
2799 // If there's not enough expansion space available, give up.
2800 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2801 return false;
2802 }
2803 // Otherwise, we try expansion.
2804 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2805 // Now go around the loop and try alloc again;
2806 // A competing allocation might beat us to the expansion space,
2807 // so we may go around the loop again if allocation fails again.
2808 if (GCExpandToAllocateDelayMillis > 0) {
2809 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2810 }
2811 }
2812 }
2813
2814 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2815 // Only shrink if a compaction was done so that all the free space
2816 // in the generation is in a contiguous block at the end.
2817 if (did_compact()) {
2818 CardGeneration::shrink(bytes);
2819 }
2820 }
2821
2822 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2823 assert_locked_or_safepoint(Heap_lock);
2824 }
2825
2826 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2827 assert_locked_or_safepoint(Heap_lock);
2828 assert_lock_strong(freelistLock());
2829 if (PrintGCDetails && Verbose) {
2830 warning("Shrinking of CMS not yet implemented");
2831 }
2832 return;
2833 }
2834
2835
2836 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2837 // phases.
2838 class CMSPhaseAccounting: public StackObj {
2839 public:
2840 CMSPhaseAccounting(CMSCollector *collector,
2841 const char *phase,
2842 const GCId gc_id,
2843 bool print_cr = true);
2844 ~CMSPhaseAccounting();
2845
2846 private:
2847 CMSCollector *_collector;
2848 const char *_phase;
2849 elapsedTimer _wallclock;
2850 bool _print_cr;
2851 const GCId _gc_id;
2852
2853 public:
2854 // Not MT-safe; so do not pass around these StackObj's
2855 // where they may be accessed by other threads.
2856 jlong wallclock_millis() {
2857 assert(_wallclock.is_active(), "Wall clock should not stop");
2858 _wallclock.stop(); // to record time
2859 jlong ret = _wallclock.milliseconds();
2860 _wallclock.start(); // restart
2861 return ret;
2862 }
2863 };
2864
2865 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2866 const char *phase,
2867 const GCId gc_id,
2868 bool print_cr) :
2869 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) {
2870
2871 if (PrintCMSStatistics != 0) {
2872 _collector->resetYields();
2873 }
2874 if (PrintGCDetails) {
2875 gclog_or_tty->gclog_stamp(_gc_id);
2876 gclog_or_tty->print_cr("[%s-concurrent-%s-start]",
2877 _collector->cmsGen()->short_name(), _phase);
2878 }
2879 _collector->resetTimer();
2880 _wallclock.start();
2881 _collector->startTimer();
2882 }
2883
2884 CMSPhaseAccounting::~CMSPhaseAccounting() {
2885 assert(_wallclock.is_active(), "Wall clock should not have stopped");
2886 _collector->stopTimer();
2887 _wallclock.stop();
2888 if (PrintGCDetails) {
2889 gclog_or_tty->gclog_stamp(_gc_id);
2890 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
2891 _collector->cmsGen()->short_name(),
2892 _phase, _collector->timerValue(), _wallclock.seconds());
2893 if (_print_cr) {
2894 gclog_or_tty->cr();
2895 }
2896 if (PrintCMSStatistics != 0) {
2897 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
2898 _collector->yields());
2899 }
2900 }
2901 }
2902
2903 // CMS work
2904
2905 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2906 class CMSParMarkTask : public AbstractGangTask {
2907 protected:
2908 CMSCollector* _collector;
2909 uint _n_workers;
2910 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2911 AbstractGangTask(name),
2912 _collector(collector),
2913 _n_workers(n_workers) {}
2914 // Work method in support of parallel rescan ... of young gen spaces
2915 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
2916 ContiguousSpace* space,
2917 HeapWord** chunk_array, size_t chunk_top);
2918 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
2919 };
2920
2921 // Parallel initial mark task
2922 class CMSParInitialMarkTask: public CMSParMarkTask {
2923 StrongRootsScope* _strong_roots_scope;
2924 public:
2925 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) :
2926 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers),
2927 _strong_roots_scope(strong_roots_scope) {}
2928 void work(uint worker_id);
2929 };
2930
2931 // Checkpoint the roots into this generation from outside
2932 // this generation. [Note this initial checkpoint need only
2933 // be approximate -- we'll do a catch up phase subsequently.]
2934 void CMSCollector::checkpointRootsInitial() {
2935 assert(_collectorState == InitialMarking, "Wrong collector state");
2936 check_correct_thread_executing();
2937 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
2938
2939 save_heap_summary();
2940 report_heap_summary(GCWhen::BeforeGC);
2941
2942 ReferenceProcessor* rp = ref_processor();
2943 assert(_restart_addr == NULL, "Control point invariant");
2944 {
2945 // acquire locks for subsequent manipulations
2946 MutexLockerEx x(bitMapLock(),
2947 Mutex::_no_safepoint_check_flag);
2948 checkpointRootsInitialWork();
2949 // enable ("weak") refs discovery
2950 rp->enable_discovery();
2951 _collectorState = Marking;
2952 }
2953 }
2954
2955 void CMSCollector::checkpointRootsInitialWork() {
2956 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2957 assert(_collectorState == InitialMarking, "just checking");
2958
2959 // If there has not been a GC[n-1] since last GC[n] cycle completed,
2960 // precede our marking with a collection of all
2961 // younger generations to keep floating garbage to a minimum.
2962 // XXX: we won't do this for now -- it's an optimization to be done later.
2963
2964 // already have locks
2965 assert_lock_strong(bitMapLock());
2966 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2967
2968 // Setup the verification and class unloading state for this
2969 // CMS collection cycle.
2970 setup_cms_unloading_and_verification_state();
2971
2972 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
2973 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
2974
2975 // Reset all the PLAB chunk arrays if necessary.
2976 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2977 reset_survivor_plab_arrays();
2978 }
2979
2980 ResourceMark rm;
2981 HandleMark hm;
2982
2983 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2984 GenCollectedHeap* gch = GenCollectedHeap::heap();
2985
2986 verify_work_stacks_empty();
2987 verify_overflow_empty();
2988
2989 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2990 // Update the saved marks which may affect the root scans.
2991 gch->save_marks();
2992
2993 // weak reference processing has not started yet.
2994 ref_processor()->set_enqueuing_is_done(false);
2995
2996 // Need to remember all newly created CLDs,
2997 // so that we can guarantee that the remark finds them.
2998 ClassLoaderDataGraph::remember_new_clds(true);
2999
3000 // Whenever a CLD is found, it will be claimed before proceeding to mark
3001 // the klasses. The claimed marks need to be cleared before marking starts.
3002 ClassLoaderDataGraph::clear_claimed_marks();
3003
3004 if (CMSPrintEdenSurvivorChunks) {
3005 print_eden_and_survivor_chunk_arrays();
3006 }
3007
3008 {
3009 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3010 if (CMSParallelInitialMarkEnabled) {
3011 // The parallel version.
3012 FlexibleWorkGang* workers = gch->workers();
3013 assert(workers != NULL, "Need parallel worker threads.");
3014 uint n_workers = workers->active_workers();
3015
3016 StrongRootsScope srs(n_workers);
3017
3018 CMSParInitialMarkTask tsk(this, &srs, n_workers);
3019 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
3020 if (n_workers > 1) {
3021 workers->run_task(&tsk);
3022 } else {
3023 tsk.work(0);
3024 }
3025 } else {
3026 // The serial version.
3027 CLDToOopClosure cld_closure(¬Older, true);
3028 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3029
3030 StrongRootsScope srs(1);
3031
3032 gch->gen_process_roots(&srs,
3033 _cmsGen->level(),
3034 true, // younger gens are roots
3035 GenCollectedHeap::ScanningOption(roots_scanning_options()),
3036 should_unload_classes(),
3037 ¬Older,
3038 NULL,
3039 &cld_closure);
3040 }
3041 }
3042
3043 // Clear mod-union table; it will be dirtied in the prologue of
3044 // CMS generation per each younger generation collection.
3045
3046 assert(_modUnionTable.isAllClear(),
3047 "Was cleared in most recent final checkpoint phase"
3048 " or no bits are set in the gc_prologue before the start of the next "
3049 "subsequent marking phase.");
3050
3051 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
3052
3053 // Save the end of the used_region of the constituent generations
3054 // to be used to limit the extent of sweep in each generation.
3055 save_sweep_limits();
3056 verify_overflow_empty();
3057 }
3058
3059 bool CMSCollector::markFromRoots() {
3060 // we might be tempted to assert that:
3061 // assert(!SafepointSynchronize::is_at_safepoint(),
3062 // "inconsistent argument?");
3063 // However that wouldn't be right, because it's possible that
3064 // a safepoint is indeed in progress as a younger generation
3065 // stop-the-world GC happens even as we mark in this generation.
3066 assert(_collectorState == Marking, "inconsistent state?");
3067 check_correct_thread_executing();
3068 verify_overflow_empty();
3069
3070 // Weak ref discovery note: We may be discovering weak
3071 // refs in this generation concurrent (but interleaved) with
3072 // weak ref discovery by a younger generation collector.
3073
3074 CMSTokenSyncWithLocks ts(true, bitMapLock());
3075 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3076 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3077 bool res = markFromRootsWork();
3078 if (res) {
3079 _collectorState = Precleaning;
3080 } else { // We failed and a foreground collection wants to take over
3081 assert(_foregroundGCIsActive, "internal state inconsistency");
3082 assert(_restart_addr == NULL, "foreground will restart from scratch");
3083 if (PrintGCDetails) {
3084 gclog_or_tty->print_cr("bailing out to foreground collection");
3085 }
3086 }
3087 verify_overflow_empty();
3088 return res;
3089 }
3090
3091 bool CMSCollector::markFromRootsWork() {
3092 // iterate over marked bits in bit map, doing a full scan and mark
3093 // from these roots using the following algorithm:
3094 // . if oop is to the right of the current scan pointer,
3095 // mark corresponding bit (we'll process it later)
3096 // . else (oop is to left of current scan pointer)
3097 // push oop on marking stack
3098 // . drain the marking stack
3099
3100 // Note that when we do a marking step we need to hold the
3101 // bit map lock -- recall that direct allocation (by mutators)
3102 // and promotion (by younger generation collectors) is also
3103 // marking the bit map. [the so-called allocate live policy.]
3104 // Because the implementation of bit map marking is not
3105 // robust wrt simultaneous marking of bits in the same word,
3106 // we need to make sure that there is no such interference
3107 // between concurrent such updates.
3108
3109 // already have locks
3110 assert_lock_strong(bitMapLock());
3111
3112 verify_work_stacks_empty();
3113 verify_overflow_empty();
3114 bool result = false;
3115 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3116 result = do_marking_mt();
3117 } else {
3118 result = do_marking_st();
3119 }
3120 return result;
3121 }
3122
3123 // Forward decl
3124 class CMSConcMarkingTask;
3125
3126 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3127 CMSCollector* _collector;
3128 CMSConcMarkingTask* _task;
3129 public:
3130 virtual void yield();
3131
3132 // "n_threads" is the number of threads to be terminated.
3133 // "queue_set" is a set of work queues of other threads.
3134 // "collector" is the CMS collector associated with this task terminator.
3135 // "yield" indicates whether we need the gang as a whole to yield.
3136 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3137 ParallelTaskTerminator(n_threads, queue_set),
3138 _collector(collector) { }
3139
3140 void set_task(CMSConcMarkingTask* task) {
3141 _task = task;
3142 }
3143 };
3144
3145 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3146 CMSConcMarkingTask* _task;
3147 public:
3148 bool should_exit_termination();
3149 void set_task(CMSConcMarkingTask* task) {
3150 _task = task;
3151 }
3152 };
3153
3154 // MT Concurrent Marking Task
3155 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3156 CMSCollector* _collector;
3157 uint _n_workers; // requested/desired # workers
3158 bool _result;
3159 CompactibleFreeListSpace* _cms_space;
3160 char _pad_front[64]; // padding to ...
3161 HeapWord* _global_finger; // ... avoid sharing cache line
3162 char _pad_back[64];
3163 HeapWord* _restart_addr;
3164
3165 // Exposed here for yielding support
3166 Mutex* const _bit_map_lock;
3167
3168 // The per thread work queues, available here for stealing
3169 OopTaskQueueSet* _task_queues;
3170
3171 // Termination (and yielding) support
3172 CMSConcMarkingTerminator _term;
3173 CMSConcMarkingTerminatorTerminator _term_term;
3174
3175 public:
3176 CMSConcMarkingTask(CMSCollector* collector,
3177 CompactibleFreeListSpace* cms_space,
3178 YieldingFlexibleWorkGang* workers,
3179 OopTaskQueueSet* task_queues):
3180 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3181 _collector(collector),
3182 _cms_space(cms_space),
3183 _n_workers(0), _result(true),
3184 _task_queues(task_queues),
3185 _term(_n_workers, task_queues, _collector),
3186 _bit_map_lock(collector->bitMapLock())
3187 {
3188 _requested_size = _n_workers;
3189 _term.set_task(this);
3190 _term_term.set_task(this);
3191 _restart_addr = _global_finger = _cms_space->bottom();
3192 }
3193
3194
3195 OopTaskQueueSet* task_queues() { return _task_queues; }
3196
3197 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3198
3199 HeapWord** global_finger_addr() { return &_global_finger; }
3200
3201 CMSConcMarkingTerminator* terminator() { return &_term; }
3202
3203 virtual void set_for_termination(uint active_workers) {
3204 terminator()->reset_for_reuse(active_workers);
3205 }
3206
3207 void work(uint worker_id);
3208 bool should_yield() {
3209 return ConcurrentMarkSweepThread::should_yield()
3210 && !_collector->foregroundGCIsActive();
3211 }
3212
3213 virtual void coordinator_yield(); // stuff done by coordinator
3214 bool result() { return _result; }
3215
3216 void reset(HeapWord* ra) {
3217 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3218 _restart_addr = _global_finger = ra;
3219 _term.reset_for_reuse();
3220 }
3221
3222 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3223 OopTaskQueue* work_q);
3224
3225 private:
3226 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3227 void do_work_steal(int i);
3228 void bump_global_finger(HeapWord* f);
3229 };
3230
3231 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3232 assert(_task != NULL, "Error");
3233 return _task->yielding();
3234 // Note that we do not need the disjunct || _task->should_yield() above
3235 // because we want terminating threads to yield only if the task
3236 // is already in the midst of yielding, which happens only after at least one
3237 // thread has yielded.
3238 }
3239
3240 void CMSConcMarkingTerminator::yield() {
3241 if (_task->should_yield()) {
3242 _task->yield();
3243 } else {
3244 ParallelTaskTerminator::yield();
3245 }
3246 }
3247
3248 ////////////////////////////////////////////////////////////////
3249 // Concurrent Marking Algorithm Sketch
3250 ////////////////////////////////////////////////////////////////
3251 // Until all tasks exhausted (both spaces):
3252 // -- claim next available chunk
3253 // -- bump global finger via CAS
3254 // -- find first object that starts in this chunk
3255 // and start scanning bitmap from that position
3256 // -- scan marked objects for oops
3257 // -- CAS-mark target, and if successful:
3258 // . if target oop is above global finger (volatile read)
3259 // nothing to do
3260 // . if target oop is in chunk and above local finger
3261 // then nothing to do
3262 // . else push on work-queue
3263 // -- Deal with possible overflow issues:
3264 // . local work-queue overflow causes stuff to be pushed on
3265 // global (common) overflow queue
3266 // . always first empty local work queue
3267 // . then get a batch of oops from global work queue if any
3268 // . then do work stealing
3269 // -- When all tasks claimed (both spaces)
3270 // and local work queue empty,
3271 // then in a loop do:
3272 // . check global overflow stack; steal a batch of oops and trace
3273 // . try to steal from other threads oif GOS is empty
3274 // . if neither is available, offer termination
3275 // -- Terminate and return result
3276 //
3277 void CMSConcMarkingTask::work(uint worker_id) {
3278 elapsedTimer _timer;
3279 ResourceMark rm;
3280 HandleMark hm;
3281
3282 DEBUG_ONLY(_collector->verify_overflow_empty();)
3283
3284 // Before we begin work, our work queue should be empty
3285 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3286 // Scan the bitmap covering _cms_space, tracing through grey objects.
3287 _timer.start();
3288 do_scan_and_mark(worker_id, _cms_space);
3289 _timer.stop();
3290 if (PrintCMSStatistics != 0) {
3291 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3292 worker_id, _timer.seconds());
3293 // XXX: need xxx/xxx type of notation, two timers
3294 }
3295
3296 // ... do work stealing
3297 _timer.reset();
3298 _timer.start();
3299 do_work_steal(worker_id);
3300 _timer.stop();
3301 if (PrintCMSStatistics != 0) {
3302 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3303 worker_id, _timer.seconds());
3304 // XXX: need xxx/xxx type of notation, two timers
3305 }
3306 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3307 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3308 // Note that under the current task protocol, the
3309 // following assertion is true even of the spaces
3310 // expanded since the completion of the concurrent
3311 // marking. XXX This will likely change under a strict
3312 // ABORT semantics.
3313 // After perm removal the comparison was changed to
3314 // greater than or equal to from strictly greater than.
3315 // Before perm removal the highest address sweep would
3316 // have been at the end of perm gen but now is at the
3317 // end of the tenured gen.
3318 assert(_global_finger >= _cms_space->end(),
3319 "All tasks have been completed");
3320 DEBUG_ONLY(_collector->verify_overflow_empty();)
3321 }
3322
3323 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3324 HeapWord* read = _global_finger;
3325 HeapWord* cur = read;
3326 while (f > read) {
3327 cur = read;
3328 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3329 if (cur == read) {
3330 // our cas succeeded
3331 assert(_global_finger >= f, "protocol consistency");
3332 break;
3333 }
3334 }
3335 }
3336
3337 // This is really inefficient, and should be redone by
3338 // using (not yet available) block-read and -write interfaces to the
3339 // stack and the work_queue. XXX FIX ME !!!
3340 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3341 OopTaskQueue* work_q) {
3342 // Fast lock-free check
3343 if (ovflw_stk->length() == 0) {
3344 return false;
3345 }
3346 assert(work_q->size() == 0, "Shouldn't steal");
3347 MutexLockerEx ml(ovflw_stk->par_lock(),
3348 Mutex::_no_safepoint_check_flag);
3349 // Grab up to 1/4 the size of the work queue
3350 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3351 (size_t)ParGCDesiredObjsFromOverflowList);
3352 num = MIN2(num, ovflw_stk->length());
3353 for (int i = (int) num; i > 0; i--) {
3354 oop cur = ovflw_stk->pop();
3355 assert(cur != NULL, "Counted wrong?");
3356 work_q->push(cur);
3357 }
3358 return num > 0;
3359 }
3360
3361 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3362 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3363 int n_tasks = pst->n_tasks();
3364 // We allow that there may be no tasks to do here because
3365 // we are restarting after a stack overflow.
3366 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3367 uint nth_task = 0;
3368
3369 HeapWord* aligned_start = sp->bottom();
3370 if (sp->used_region().contains(_restart_addr)) {
3371 // Align down to a card boundary for the start of 0th task
3372 // for this space.
3373 aligned_start =
3374 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3375 CardTableModRefBS::card_size);
3376 }
3377
3378 size_t chunk_size = sp->marking_task_size();
3379 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3380 // Having claimed the nth task in this space,
3381 // compute the chunk that it corresponds to:
3382 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3383 aligned_start + (nth_task+1)*chunk_size);
3384 // Try and bump the global finger via a CAS;
3385 // note that we need to do the global finger bump
3386 // _before_ taking the intersection below, because
3387 // the task corresponding to that region will be
3388 // deemed done even if the used_region() expands
3389 // because of allocation -- as it almost certainly will
3390 // during start-up while the threads yield in the
3391 // closure below.
3392 HeapWord* finger = span.end();
3393 bump_global_finger(finger); // atomically
3394 // There are null tasks here corresponding to chunks
3395 // beyond the "top" address of the space.
3396 span = span.intersection(sp->used_region());
3397 if (!span.is_empty()) { // Non-null task
3398 HeapWord* prev_obj;
3399 assert(!span.contains(_restart_addr) || nth_task == 0,
3400 "Inconsistency");
3401 if (nth_task == 0) {
3402 // For the 0th task, we'll not need to compute a block_start.
3403 if (span.contains(_restart_addr)) {
3404 // In the case of a restart because of stack overflow,
3405 // we might additionally skip a chunk prefix.
3406 prev_obj = _restart_addr;
3407 } else {
3408 prev_obj = span.start();
3409 }
3410 } else {
3411 // We want to skip the first object because
3412 // the protocol is to scan any object in its entirety
3413 // that _starts_ in this span; a fortiori, any
3414 // object starting in an earlier span is scanned
3415 // as part of an earlier claimed task.
3416 // Below we use the "careful" version of block_start
3417 // so we do not try to navigate uninitialized objects.
3418 prev_obj = sp->block_start_careful(span.start());
3419 // Below we use a variant of block_size that uses the
3420 // Printezis bits to avoid waiting for allocated
3421 // objects to become initialized/parsable.
3422 while (prev_obj < span.start()) {
3423 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3424 if (sz > 0) {
3425 prev_obj += sz;
3426 } else {
3427 // In this case we may end up doing a bit of redundant
3428 // scanning, but that appears unavoidable, short of
3429 // locking the free list locks; see bug 6324141.
3430 break;
3431 }
3432 }
3433 }
3434 if (prev_obj < span.end()) {
3435 MemRegion my_span = MemRegion(prev_obj, span.end());
3436 // Do the marking work within a non-empty span --
3437 // the last argument to the constructor indicates whether the
3438 // iteration should be incremental with periodic yields.
3439 Par_MarkFromRootsClosure cl(this, _collector, my_span,
3440 &_collector->_markBitMap,
3441 work_queue(i),
3442 &_collector->_markStack);
3443 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3444 } // else nothing to do for this task
3445 } // else nothing to do for this task
3446 }
3447 // We'd be tempted to assert here that since there are no
3448 // more tasks left to claim in this space, the global_finger
3449 // must exceed space->top() and a fortiori space->end(). However,
3450 // that would not quite be correct because the bumping of
3451 // global_finger occurs strictly after the claiming of a task,
3452 // so by the time we reach here the global finger may not yet
3453 // have been bumped up by the thread that claimed the last
3454 // task.
3455 pst->all_tasks_completed();
3456 }
3457
3458 class Par_ConcMarkingClosure: public MetadataAwareOopClosure {
3459 private:
3460 CMSCollector* _collector;
3461 CMSConcMarkingTask* _task;
3462 MemRegion _span;
3463 CMSBitMap* _bit_map;
3464 CMSMarkStack* _overflow_stack;
3465 OopTaskQueue* _work_queue;
3466 protected:
3467 DO_OOP_WORK_DEFN
3468 public:
3469 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3470 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3471 MetadataAwareOopClosure(collector->ref_processor()),
3472 _collector(collector),
3473 _task(task),
3474 _span(collector->_span),
3475 _work_queue(work_queue),
3476 _bit_map(bit_map),
3477 _overflow_stack(overflow_stack)
3478 { }
3479 virtual void do_oop(oop* p);
3480 virtual void do_oop(narrowOop* p);
3481
3482 void trim_queue(size_t max);
3483 void handle_stack_overflow(HeapWord* lost);
3484 void do_yield_check() {
3485 if (_task->should_yield()) {
3486 _task->yield();
3487 }
3488 }
3489 };
3490
3491 // Grey object scanning during work stealing phase --
3492 // the salient assumption here is that any references
3493 // that are in these stolen objects being scanned must
3494 // already have been initialized (else they would not have
3495 // been published), so we do not need to check for
3496 // uninitialized objects before pushing here.
3497 void Par_ConcMarkingClosure::do_oop(oop obj) {
3498 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
3499 HeapWord* addr = (HeapWord*)obj;
3500 // Check if oop points into the CMS generation
3501 // and is not marked
3502 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3503 // a white object ...
3504 // If we manage to "claim" the object, by being the
3505 // first thread to mark it, then we push it on our
3506 // marking stack
3507 if (_bit_map->par_mark(addr)) { // ... now grey
3508 // push on work queue (grey set)
3509 bool simulate_overflow = false;
3510 NOT_PRODUCT(
3511 if (CMSMarkStackOverflowALot &&
3512 _collector->simulate_overflow()) {
3513 // simulate a stack overflow
3514 simulate_overflow = true;
3515 }
3516 )
3517 if (simulate_overflow ||
3518 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3519 // stack overflow
3520 if (PrintCMSStatistics != 0) {
3521 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
3522 SIZE_FORMAT, _overflow_stack->capacity());
3523 }
3524 // We cannot assert that the overflow stack is full because
3525 // it may have been emptied since.
3526 assert(simulate_overflow ||
3527 _work_queue->size() == _work_queue->max_elems(),
3528 "Else push should have succeeded");
3529 handle_stack_overflow(addr);
3530 }
3531 } // Else, some other thread got there first
3532 do_yield_check();
3533 }
3534 }
3535
3536 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3537 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3538
3539 void Par_ConcMarkingClosure::trim_queue(size_t max) {
3540 while (_work_queue->size() > max) {
3541 oop new_oop;
3542 if (_work_queue->pop_local(new_oop)) {
3543 assert(new_oop->is_oop(), "Should be an oop");
3544 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3545 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3546 new_oop->oop_iterate(this); // do_oop() above
3547 do_yield_check();
3548 }
3549 }
3550 }
3551
3552 // Upon stack overflow, we discard (part of) the stack,
3553 // remembering the least address amongst those discarded
3554 // in CMSCollector's _restart_address.
3555 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3556 // We need to do this under a mutex to prevent other
3557 // workers from interfering with the work done below.
3558 MutexLockerEx ml(_overflow_stack->par_lock(),
3559 Mutex::_no_safepoint_check_flag);
3560 // Remember the least grey address discarded
3561 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3562 _collector->lower_restart_addr(ra);
3563 _overflow_stack->reset(); // discard stack contents
3564 _overflow_stack->expand(); // expand the stack if possible
3565 }
3566
3567
3568 void CMSConcMarkingTask::do_work_steal(int i) {
3569 OopTaskQueue* work_q = work_queue(i);
3570 oop obj_to_scan;
3571 CMSBitMap* bm = &(_collector->_markBitMap);
3572 CMSMarkStack* ovflw = &(_collector->_markStack);
3573 int* seed = _collector->hash_seed(i);
3574 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3575 while (true) {
3576 cl.trim_queue(0);
3577 assert(work_q->size() == 0, "Should have been emptied above");
3578 if (get_work_from_overflow_stack(ovflw, work_q)) {
3579 // Can't assert below because the work obtained from the
3580 // overflow stack may already have been stolen from us.
3581 // assert(work_q->size() > 0, "Work from overflow stack");
3582 continue;
3583 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3584 assert(obj_to_scan->is_oop(), "Should be an oop");
3585 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3586 obj_to_scan->oop_iterate(&cl);
3587 } else if (terminator()->offer_termination(&_term_term)) {
3588 assert(work_q->size() == 0, "Impossible!");
3589 break;
3590 } else if (yielding() || should_yield()) {
3591 yield();
3592 }
3593 }
3594 }
3595
3596 // This is run by the CMS (coordinator) thread.
3597 void CMSConcMarkingTask::coordinator_yield() {
3598 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3599 "CMS thread should hold CMS token");
3600 // First give up the locks, then yield, then re-lock
3601 // We should probably use a constructor/destructor idiom to
3602 // do this unlock/lock or modify the MutexUnlocker class to
3603 // serve our purpose. XXX
3604 assert_lock_strong(_bit_map_lock);
3605 _bit_map_lock->unlock();
3606 ConcurrentMarkSweepThread::desynchronize(true);
3607 _collector->stopTimer();
3608 if (PrintCMSStatistics != 0) {
3609 _collector->incrementYields();
3610 }
3611
3612 // It is possible for whichever thread initiated the yield request
3613 // not to get a chance to wake up and take the bitmap lock between
3614 // this thread releasing it and reacquiring it. So, while the
3615 // should_yield() flag is on, let's sleep for a bit to give the
3616 // other thread a chance to wake up. The limit imposed on the number
3617 // of iterations is defensive, to avoid any unforseen circumstances
3618 // putting us into an infinite loop. Since it's always been this
3619 // (coordinator_yield()) method that was observed to cause the
3620 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3621 // which is by default non-zero. For the other seven methods that
3622 // also perform the yield operation, as are using a different
3623 // parameter (CMSYieldSleepCount) which is by default zero. This way we
3624 // can enable the sleeping for those methods too, if necessary.
3625 // See 6442774.
3626 //
3627 // We really need to reconsider the synchronization between the GC
3628 // thread and the yield-requesting threads in the future and we
3629 // should really use wait/notify, which is the recommended
3630 // way of doing this type of interaction. Additionally, we should
3631 // consolidate the eight methods that do the yield operation and they
3632 // are almost identical into one for better maintainability and
3633 // readability. See 6445193.
3634 //
3635 // Tony 2006.06.29
3636 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3637 ConcurrentMarkSweepThread::should_yield() &&
3638 !CMSCollector::foregroundGCIsActive(); ++i) {
3639 os::sleep(Thread::current(), 1, false);
3640 }
3641
3642 ConcurrentMarkSweepThread::synchronize(true);
3643 _bit_map_lock->lock_without_safepoint_check();
3644 _collector->startTimer();
3645 }
3646
3647 bool CMSCollector::do_marking_mt() {
3648 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3649 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3650 conc_workers()->active_workers(),
3651 Threads::number_of_non_daemon_threads());
3652 conc_workers()->set_active_workers(num_workers);
3653
3654 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
3655
3656 CMSConcMarkingTask tsk(this,
3657 cms_space,
3658 conc_workers(),
3659 task_queues());
3660
3661 // Since the actual number of workers we get may be different
3662 // from the number we requested above, do we need to do anything different
3663 // below? In particular, may be we need to subclass the SequantialSubTasksDone
3664 // class?? XXX
3665 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3666
3667 // Refs discovery is already non-atomic.
3668 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3669 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3670 conc_workers()->start_task(&tsk);
3671 while (tsk.yielded()) {
3672 tsk.coordinator_yield();
3673 conc_workers()->continue_task(&tsk);
3674 }
3675 // If the task was aborted, _restart_addr will be non-NULL
3676 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3677 while (_restart_addr != NULL) {
3678 // XXX For now we do not make use of ABORTED state and have not
3679 // yet implemented the right abort semantics (even in the original
3680 // single-threaded CMS case). That needs some more investigation
3681 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3682 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3683 // If _restart_addr is non-NULL, a marking stack overflow
3684 // occurred; we need to do a fresh marking iteration from the
3685 // indicated restart address.
3686 if (_foregroundGCIsActive) {
3687 // We may be running into repeated stack overflows, having
3688 // reached the limit of the stack size, while making very
3689 // slow forward progress. It may be best to bail out and
3690 // let the foreground collector do its job.
3691 // Clear _restart_addr, so that foreground GC
3692 // works from scratch. This avoids the headache of
3693 // a "rescan" which would otherwise be needed because
3694 // of the dirty mod union table & card table.
3695 _restart_addr = NULL;
3696 return false;
3697 }
3698 // Adjust the task to restart from _restart_addr
3699 tsk.reset(_restart_addr);
3700 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3701 _restart_addr);
3702 _restart_addr = NULL;
3703 // Get the workers going again
3704 conc_workers()->start_task(&tsk);
3705 while (tsk.yielded()) {
3706 tsk.coordinator_yield();
3707 conc_workers()->continue_task(&tsk);
3708 }
3709 }
3710 assert(tsk.completed(), "Inconsistency");
3711 assert(tsk.result() == true, "Inconsistency");
3712 return true;
3713 }
3714
3715 bool CMSCollector::do_marking_st() {
3716 ResourceMark rm;
3717 HandleMark hm;
3718
3719 // Temporarily make refs discovery single threaded (non-MT)
3720 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3721 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3722 &_markStack, CMSYield);
3723 // the last argument to iterate indicates whether the iteration
3724 // should be incremental with periodic yields.
3725 _markBitMap.iterate(&markFromRootsClosure);
3726 // If _restart_addr is non-NULL, a marking stack overflow
3727 // occurred; we need to do a fresh iteration from the
3728 // indicated restart address.
3729 while (_restart_addr != NULL) {
3730 if (_foregroundGCIsActive) {
3731 // We may be running into repeated stack overflows, having
3732 // reached the limit of the stack size, while making very
3733 // slow forward progress. It may be best to bail out and
3734 // let the foreground collector do its job.
3735 // Clear _restart_addr, so that foreground GC
3736 // works from scratch. This avoids the headache of
3737 // a "rescan" which would otherwise be needed because
3738 // of the dirty mod union table & card table.
3739 _restart_addr = NULL;
3740 return false; // indicating failure to complete marking
3741 }
3742 // Deal with stack overflow:
3743 // we restart marking from _restart_addr
3744 HeapWord* ra = _restart_addr;
3745 markFromRootsClosure.reset(ra);
3746 _restart_addr = NULL;
3747 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3748 }
3749 return true;
3750 }
3751
3752 void CMSCollector::preclean() {
3753 check_correct_thread_executing();
3754 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3755 verify_work_stacks_empty();
3756 verify_overflow_empty();
3757 _abort_preclean = false;
3758 if (CMSPrecleaningEnabled) {
3759 if (!CMSEdenChunksRecordAlways) {
3760 _eden_chunk_index = 0;
3761 }
3762 size_t used = get_eden_used();
3763 size_t capacity = get_eden_capacity();
3764 // Don't start sampling unless we will get sufficiently
3765 // many samples.
3766 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
3767 * CMSScheduleRemarkEdenPenetration)) {
3768 _start_sampling = true;
3769 } else {
3770 _start_sampling = false;
3771 }
3772 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3773 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3774 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3775 }
3776 CMSTokenSync x(true); // is cms thread
3777 if (CMSPrecleaningEnabled) {
3778 sample_eden();
3779 _collectorState = AbortablePreclean;
3780 } else {
3781 _collectorState = FinalMarking;
3782 }
3783 verify_work_stacks_empty();
3784 verify_overflow_empty();
3785 }
3786
3787 // Try and schedule the remark such that young gen
3788 // occupancy is CMSScheduleRemarkEdenPenetration %.
3789 void CMSCollector::abortable_preclean() {
3790 check_correct_thread_executing();
3791 assert(CMSPrecleaningEnabled, "Inconsistent control state");
3792 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3793
3794 // If Eden's current occupancy is below this threshold,
3795 // immediately schedule the remark; else preclean
3796 // past the next scavenge in an effort to
3797 // schedule the pause as described above. By choosing
3798 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3799 // we will never do an actual abortable preclean cycle.
3800 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3801 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3802 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3803 // We need more smarts in the abortable preclean
3804 // loop below to deal with cases where allocation
3805 // in young gen is very very slow, and our precleaning
3806 // is running a losing race against a horde of
3807 // mutators intent on flooding us with CMS updates
3808 // (dirty cards).
3809 // One, admittedly dumb, strategy is to give up
3810 // after a certain number of abortable precleaning loops
3811 // or after a certain maximum time. We want to make
3812 // this smarter in the next iteration.
3813 // XXX FIX ME!!! YSR
3814 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3815 while (!(should_abort_preclean() ||
3816 ConcurrentMarkSweepThread::should_terminate())) {
3817 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3818 cumworkdone += workdone;
3819 loops++;
3820 // Voluntarily terminate abortable preclean phase if we have
3821 // been at it for too long.
3822 if ((CMSMaxAbortablePrecleanLoops != 0) &&
3823 loops >= CMSMaxAbortablePrecleanLoops) {
3824 if (PrintGCDetails) {
3825 gclog_or_tty->print(" CMS: abort preclean due to loops ");
3826 }
3827 break;
3828 }
3829 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3830 if (PrintGCDetails) {
3831 gclog_or_tty->print(" CMS: abort preclean due to time ");
3832 }
3833 break;
3834 }
3835 // If we are doing little work each iteration, we should
3836 // take a short break.
3837 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3838 // Sleep for some time, waiting for work to accumulate
3839 stopTimer();
3840 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3841 startTimer();
3842 waited++;
3843 }
3844 }
3845 if (PrintCMSStatistics > 0) {
3846 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3847 loops, waited, cumworkdone);
3848 }
3849 }
3850 CMSTokenSync x(true); // is cms thread
3851 if (_collectorState != Idling) {
3852 assert(_collectorState == AbortablePreclean,
3853 "Spontaneous state transition?");
3854 _collectorState = FinalMarking;
3855 } // Else, a foreground collection completed this CMS cycle.
3856 return;
3857 }
3858
3859 // Respond to an Eden sampling opportunity
3860 void CMSCollector::sample_eden() {
3861 // Make sure a young gc cannot sneak in between our
3862 // reading and recording of a sample.
3863 assert(Thread::current()->is_ConcurrentGC_thread(),
3864 "Only the cms thread may collect Eden samples");
3865 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3866 "Should collect samples while holding CMS token");
3867 if (!_start_sampling) {
3868 return;
3869 }
3870 // When CMSEdenChunksRecordAlways is true, the eden chunk array
3871 // is populated by the young generation.
3872 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3873 if (_eden_chunk_index < _eden_chunk_capacity) {
3874 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
3875 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3876 "Unexpected state of Eden");
3877 // We'd like to check that what we just sampled is an oop-start address;
3878 // however, we cannot do that here since the object may not yet have been
3879 // initialized. So we'll instead do the check when we _use_ this sample
3880 // later.
3881 if (_eden_chunk_index == 0 ||
3882 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3883 _eden_chunk_array[_eden_chunk_index-1])
3884 >= CMSSamplingGrain)) {
3885 _eden_chunk_index++; // commit sample
3886 }
3887 }
3888 }
3889 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3890 size_t used = get_eden_used();
3891 size_t capacity = get_eden_capacity();
3892 assert(used <= capacity, "Unexpected state of Eden");
3893 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3894 _abort_preclean = true;
3895 }
3896 }
3897 }
3898
3899
3900 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3901 assert(_collectorState == Precleaning ||
3902 _collectorState == AbortablePreclean, "incorrect state");
3903 ResourceMark rm;
3904 HandleMark hm;
3905
3906 // Precleaning is currently not MT but the reference processor
3907 // may be set for MT. Disable it temporarily here.
3908 ReferenceProcessor* rp = ref_processor();
3909 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3910
3911 // Do one pass of scrubbing the discovered reference lists
3912 // to remove any reference objects with strongly-reachable
3913 // referents.
3914 if (clean_refs) {
3915 CMSPrecleanRefsYieldClosure yield_cl(this);
3916 assert(rp->span().equals(_span), "Spans should be equal");
3917 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3918 &_markStack, true /* preclean */);
3919 CMSDrainMarkingStackClosure complete_trace(this,
3920 _span, &_markBitMap, &_markStack,
3921 &keep_alive, true /* preclean */);
3922
3923 // We don't want this step to interfere with a young
3924 // collection because we don't want to take CPU
3925 // or memory bandwidth away from the young GC threads
3926 // (which may be as many as there are CPUs).
3927 // Note that we don't need to protect ourselves from
3928 // interference with mutators because they can't
3929 // manipulate the discovered reference lists nor affect
3930 // the computed reachability of the referents, the
3931 // only properties manipulated by the precleaning
3932 // of these reference lists.
3933 stopTimer();
3934 CMSTokenSyncWithLocks x(true /* is cms thread */,
3935 bitMapLock());
3936 startTimer();
3937 sample_eden();
3938
3939 // The following will yield to allow foreground
3940 // collection to proceed promptly. XXX YSR:
3941 // The code in this method may need further
3942 // tweaking for better performance and some restructuring
3943 // for cleaner interfaces.
3944 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3945 rp->preclean_discovered_references(
3946 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3947 gc_timer, _gc_tracer_cm->gc_id());
3948 }
3949
3950 if (clean_survivor) { // preclean the active survivor space(s)
3951 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3952 &_markBitMap, &_modUnionTable,
3953 &_markStack, true /* precleaning phase */);
3954 stopTimer();
3955 CMSTokenSyncWithLocks ts(true /* is cms thread */,
3956 bitMapLock());
3957 startTimer();
3958 unsigned int before_count =
3959 GenCollectedHeap::heap()->total_collections();
3960 SurvivorSpacePrecleanClosure
3961 sss_cl(this, _span, &_markBitMap, &_markStack,
3962 &pam_cl, before_count, CMSYield);
3963 _young_gen->from()->object_iterate_careful(&sss_cl);
3964 _young_gen->to()->object_iterate_careful(&sss_cl);
3965 }
3966 MarkRefsIntoAndScanClosure
3967 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3968 &_markStack, this, CMSYield,
3969 true /* precleaning phase */);
3970 // CAUTION: The following closure has persistent state that may need to
3971 // be reset upon a decrease in the sequence of addresses it
3972 // processes.
3973 ScanMarkedObjectsAgainCarefullyClosure
3974 smoac_cl(this, _span,
3975 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3976
3977 // Preclean dirty cards in ModUnionTable and CardTable using
3978 // appropriate convergence criterion;
3979 // repeat CMSPrecleanIter times unless we find that
3980 // we are losing.
3981 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3982 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3983 "Bad convergence multiplier");
3984 assert(CMSPrecleanThreshold >= 100,
3985 "Unreasonably low CMSPrecleanThreshold");
3986
3987 size_t numIter, cumNumCards, lastNumCards, curNumCards;
3988 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3989 numIter < CMSPrecleanIter;
3990 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3991 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
3992 if (Verbose && PrintGCDetails) {
3993 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3994 }
3995 // Either there are very few dirty cards, so re-mark
3996 // pause will be small anyway, or our pre-cleaning isn't
3997 // that much faster than the rate at which cards are being
3998 // dirtied, so we might as well stop and re-mark since
3999 // precleaning won't improve our re-mark time by much.
4000 if (curNumCards <= CMSPrecleanThreshold ||
4001 (numIter > 0 &&
4002 (curNumCards * CMSPrecleanDenominator >
4003 lastNumCards * CMSPrecleanNumerator))) {
4004 numIter++;
4005 cumNumCards += curNumCards;
4006 break;
4007 }
4008 }
4009
4010 preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
4011
4012 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4013 cumNumCards += curNumCards;
4014 if (PrintGCDetails && PrintCMSStatistics != 0) {
4015 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
4016 curNumCards, cumNumCards, numIter);
4017 }
4018 return cumNumCards; // as a measure of useful work done
4019 }
4020
4021 // PRECLEANING NOTES:
4022 // Precleaning involves:
4023 // . reading the bits of the modUnionTable and clearing the set bits.
4024 // . For the cards corresponding to the set bits, we scan the
4025 // objects on those cards. This means we need the free_list_lock
4026 // so that we can safely iterate over the CMS space when scanning
4027 // for oops.
4028 // . When we scan the objects, we'll be both reading and setting
4029 // marks in the marking bit map, so we'll need the marking bit map.
4030 // . For protecting _collector_state transitions, we take the CGC_lock.
4031 // Note that any races in the reading of of card table entries by the
4032 // CMS thread on the one hand and the clearing of those entries by the
4033 // VM thread or the setting of those entries by the mutator threads on the
4034 // other are quite benign. However, for efficiency it makes sense to keep
4035 // the VM thread from racing with the CMS thread while the latter is
4036 // dirty card info to the modUnionTable. We therefore also use the
4037 // CGC_lock to protect the reading of the card table and the mod union
4038 // table by the CM thread.
4039 // . We run concurrently with mutator updates, so scanning
4040 // needs to be done carefully -- we should not try to scan
4041 // potentially uninitialized objects.
4042 //
4043 // Locking strategy: While holding the CGC_lock, we scan over and
4044 // reset a maximal dirty range of the mod union / card tables, then lock
4045 // the free_list_lock and bitmap lock to do a full marking, then
4046 // release these locks; and repeat the cycle. This allows for a
4047 // certain amount of fairness in the sharing of these locks between
4048 // the CMS collector on the one hand, and the VM thread and the
4049 // mutators on the other.
4050
4051 // NOTE: preclean_mod_union_table() and preclean_card_table()
4052 // further below are largely identical; if you need to modify
4053 // one of these methods, please check the other method too.
4054
4055 size_t CMSCollector::preclean_mod_union_table(
4056 ConcurrentMarkSweepGeneration* gen,
4057 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4058 verify_work_stacks_empty();
4059 verify_overflow_empty();
4060
4061 // strategy: starting with the first card, accumulate contiguous
4062 // ranges of dirty cards; clear these cards, then scan the region
4063 // covered by these cards.
4064
4065 // Since all of the MUT is committed ahead, we can just use
4066 // that, in case the generations expand while we are precleaning.
4067 // It might also be fine to just use the committed part of the
4068 // generation, but we might potentially miss cards when the
4069 // generation is rapidly expanding while we are in the midst
4070 // of precleaning.
4071 HeapWord* startAddr = gen->reserved().start();
4072 HeapWord* endAddr = gen->reserved().end();
4073
4074 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4075
4076 size_t numDirtyCards, cumNumDirtyCards;
4077 HeapWord *nextAddr, *lastAddr;
4078 for (cumNumDirtyCards = numDirtyCards = 0,
4079 nextAddr = lastAddr = startAddr;
4080 nextAddr < endAddr;
4081 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4082
4083 ResourceMark rm;
4084 HandleMark hm;
4085
4086 MemRegion dirtyRegion;
4087 {
4088 stopTimer();
4089 // Potential yield point
4090 CMSTokenSync ts(true);
4091 startTimer();
4092 sample_eden();
4093 // Get dirty region starting at nextOffset (inclusive),
4094 // simultaneously clearing it.
4095 dirtyRegion =
4096 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4097 assert(dirtyRegion.start() >= nextAddr,
4098 "returned region inconsistent?");
4099 }
4100 // Remember where the next search should begin.
4101 // The returned region (if non-empty) is a right open interval,
4102 // so lastOffset is obtained from the right end of that
4103 // interval.
4104 lastAddr = dirtyRegion.end();
4105 // Should do something more transparent and less hacky XXX
4106 numDirtyCards =
4107 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4108
4109 // We'll scan the cards in the dirty region (with periodic
4110 // yields for foreground GC as needed).
4111 if (!dirtyRegion.is_empty()) {
4112 assert(numDirtyCards > 0, "consistency check");
4113 HeapWord* stop_point = NULL;
4114 stopTimer();
4115 // Potential yield point
4116 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4117 bitMapLock());
4118 startTimer();
4119 {
4120 verify_work_stacks_empty();
4121 verify_overflow_empty();
4122 sample_eden();
4123 stop_point =
4124 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4125 }
4126 if (stop_point != NULL) {
4127 // The careful iteration stopped early either because it found an
4128 // uninitialized object, or because we were in the midst of an
4129 // "abortable preclean", which should now be aborted. Redirty
4130 // the bits corresponding to the partially-scanned or unscanned
4131 // cards. We'll either restart at the next block boundary or
4132 // abort the preclean.
4133 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4134 "Should only be AbortablePreclean.");
4135 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4136 if (should_abort_preclean()) {
4137 break; // out of preclean loop
4138 } else {
4139 // Compute the next address at which preclean should pick up;
4140 // might need bitMapLock in order to read P-bits.
4141 lastAddr = next_card_start_after_block(stop_point);
4142 }
4143 }
4144 } else {
4145 assert(lastAddr == endAddr, "consistency check");
4146 assert(numDirtyCards == 0, "consistency check");
4147 break;
4148 }
4149 }
4150 verify_work_stacks_empty();
4151 verify_overflow_empty();
4152 return cumNumDirtyCards;
4153 }
4154
4155 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4156 // below are largely identical; if you need to modify
4157 // one of these methods, please check the other method too.
4158
4159 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4160 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4161 // strategy: it's similar to precleamModUnionTable above, in that
4162 // we accumulate contiguous ranges of dirty cards, mark these cards
4163 // precleaned, then scan the region covered by these cards.
4164 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4165 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4166
4167 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4168
4169 size_t numDirtyCards, cumNumDirtyCards;
4170 HeapWord *lastAddr, *nextAddr;
4171
4172 for (cumNumDirtyCards = numDirtyCards = 0,
4173 nextAddr = lastAddr = startAddr;
4174 nextAddr < endAddr;
4175 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4176
4177 ResourceMark rm;
4178 HandleMark hm;
4179
4180 MemRegion dirtyRegion;
4181 {
4182 // See comments in "Precleaning notes" above on why we
4183 // do this locking. XXX Could the locking overheads be
4184 // too high when dirty cards are sparse? [I don't think so.]
4185 stopTimer();
4186 CMSTokenSync x(true); // is cms thread
4187 startTimer();
4188 sample_eden();
4189 // Get and clear dirty region from card table
4190 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4191 MemRegion(nextAddr, endAddr),
4192 true,
4193 CardTableModRefBS::precleaned_card_val());
4194
4195 assert(dirtyRegion.start() >= nextAddr,
4196 "returned region inconsistent?");
4197 }
4198 lastAddr = dirtyRegion.end();
4199 numDirtyCards =
4200 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4201
4202 if (!dirtyRegion.is_empty()) {
4203 stopTimer();
4204 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4205 startTimer();
4206 sample_eden();
4207 verify_work_stacks_empty();
4208 verify_overflow_empty();
4209 HeapWord* stop_point =
4210 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4211 if (stop_point != NULL) {
4212 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4213 "Should only be AbortablePreclean.");
4214 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4215 if (should_abort_preclean()) {
4216 break; // out of preclean loop
4217 } else {
4218 // Compute the next address at which preclean should pick up.
4219 lastAddr = next_card_start_after_block(stop_point);
4220 }
4221 }
4222 } else {
4223 break;
4224 }
4225 }
4226 verify_work_stacks_empty();
4227 verify_overflow_empty();
4228 return cumNumDirtyCards;
4229 }
4230
4231 class PrecleanKlassClosure : public KlassClosure {
4232 KlassToOopClosure _cm_klass_closure;
4233 public:
4234 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4235 void do_klass(Klass* k) {
4236 if (k->has_accumulated_modified_oops()) {
4237 k->clear_accumulated_modified_oops();
4238
4239 _cm_klass_closure.do_klass(k);
4240 }
4241 }
4242 };
4243
4244 // The freelist lock is needed to prevent asserts, is it really needed?
4245 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4246
4247 cl->set_freelistLock(freelistLock);
4248
4249 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4250
4251 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4252 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4253 PrecleanKlassClosure preclean_klass_closure(cl);
4254 ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
4255
4256 verify_work_stacks_empty();
4257 verify_overflow_empty();
4258 }
4259
4260 void CMSCollector::checkpointRootsFinal() {
4261 assert(_collectorState == FinalMarking, "incorrect state transition?");
4262 check_correct_thread_executing();
4263 // world is stopped at this checkpoint
4264 assert(SafepointSynchronize::is_at_safepoint(),
4265 "world should be stopped");
4266 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4267
4268 verify_work_stacks_empty();
4269 verify_overflow_empty();
4270
4271 if (PrintGCDetails) {
4272 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4273 _young_gen->used() / K,
4274 _young_gen->capacity() / K);
4275 }
4276 {
4277 if (CMSScavengeBeforeRemark) {
4278 GenCollectedHeap* gch = GenCollectedHeap::heap();
4279 // Temporarily set flag to false, GCH->do_collection will
4280 // expect it to be false and set to true
4281 FlagSetting fl(gch->_is_gc_active, false);
4282 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
4283 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4284 int level = _cmsGen->level() - 1;
4285 if (level >= 0) {
4286 gch->do_collection(true, // full (i.e. force, see below)
4287 false, // !clear_all_soft_refs
4288 0, // size
4289 false, // is_tlab
4290 level // max_level
4291 );
4292 }
4293 }
4294 FreelistLocker x(this);
4295 MutexLockerEx y(bitMapLock(),
4296 Mutex::_no_safepoint_check_flag);
4297 checkpointRootsFinalWork();
4298 }
4299 verify_work_stacks_empty();
4300 verify_overflow_empty();
4301 }
4302
4303 void CMSCollector::checkpointRootsFinalWork() {
4304 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4305
4306 assert(haveFreelistLocks(), "must have free list locks");
4307 assert_lock_strong(bitMapLock());
4308
4309 ResourceMark rm;
4310 HandleMark hm;
4311
4312 GenCollectedHeap* gch = GenCollectedHeap::heap();
4313
4314 if (should_unload_classes()) {
4315 CodeCache::gc_prologue();
4316 }
4317 assert(haveFreelistLocks(), "must have free list locks");
4318 assert_lock_strong(bitMapLock());
4319
4320 // We might assume that we need not fill TLAB's when
4321 // CMSScavengeBeforeRemark is set, because we may have just done
4322 // a scavenge which would have filled all TLAB's -- and besides
4323 // Eden would be empty. This however may not always be the case --
4324 // for instance although we asked for a scavenge, it may not have
4325 // happened because of a JNI critical section. We probably need
4326 // a policy for deciding whether we can in that case wait until
4327 // the critical section releases and then do the remark following
4328 // the scavenge, and skip it here. In the absence of that policy,
4329 // or of an indication of whether the scavenge did indeed occur,
4330 // we cannot rely on TLAB's having been filled and must do
4331 // so here just in case a scavenge did not happen.
4332 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4333 // Update the saved marks which may affect the root scans.
4334 gch->save_marks();
4335
4336 if (CMSPrintEdenSurvivorChunks) {
4337 print_eden_and_survivor_chunk_arrays();
4338 }
4339
4340 {
4341 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4342
4343 // Note on the role of the mod union table:
4344 // Since the marker in "markFromRoots" marks concurrently with
4345 // mutators, it is possible for some reachable objects not to have been
4346 // scanned. For instance, an only reference to an object A was
4347 // placed in object B after the marker scanned B. Unless B is rescanned,
4348 // A would be collected. Such updates to references in marked objects
4349 // are detected via the mod union table which is the set of all cards
4350 // dirtied since the first checkpoint in this GC cycle and prior to
4351 // the most recent young generation GC, minus those cleaned up by the
4352 // concurrent precleaning.
4353 if (CMSParallelRemarkEnabled) {
4354 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
4355 do_remark_parallel();
4356 } else {
4357 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4358 _gc_timer_cm, _gc_tracer_cm->gc_id());
4359 do_remark_non_parallel();
4360 }
4361 }
4362 verify_work_stacks_empty();
4363 verify_overflow_empty();
4364
4365 {
4366 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4367 refProcessingWork();
4368 }
4369 verify_work_stacks_empty();
4370 verify_overflow_empty();
4371
4372 if (should_unload_classes()) {
4373 CodeCache::gc_epilogue();
4374 }
4375 JvmtiExport::gc_epilogue();
4376
4377 // If we encountered any (marking stack / work queue) overflow
4378 // events during the current CMS cycle, take appropriate
4379 // remedial measures, where possible, so as to try and avoid
4380 // recurrence of that condition.
4381 assert(_markStack.isEmpty(), "No grey objects");
4382 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4383 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4384 if (ser_ovflw > 0) {
4385 if (PrintCMSStatistics != 0) {
4386 gclog_or_tty->print_cr("Marking stack overflow (benign) "
4387 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
4388 ", kac_preclean="SIZE_FORMAT")",
4389 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4390 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4391 }
4392 _markStack.expand();
4393 _ser_pmc_remark_ovflw = 0;
4394 _ser_pmc_preclean_ovflw = 0;
4395 _ser_kac_preclean_ovflw = 0;
4396 _ser_kac_ovflw = 0;
4397 }
4398 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4399 if (PrintCMSStatistics != 0) {
4400 gclog_or_tty->print_cr("Work queue overflow (benign) "
4401 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4402 _par_pmc_remark_ovflw, _par_kac_ovflw);
4403 }
4404 _par_pmc_remark_ovflw = 0;
4405 _par_kac_ovflw = 0;
4406 }
4407 if (PrintCMSStatistics != 0) {
4408 if (_markStack._hit_limit > 0) {
4409 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
4410 _markStack._hit_limit);
4411 }
4412 if (_markStack._failed_double > 0) {
4413 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
4414 " current capacity "SIZE_FORMAT,
4415 _markStack._failed_double,
4416 _markStack.capacity());
4417 }
4418 }
4419 _markStack._hit_limit = 0;
4420 _markStack._failed_double = 0;
4421
4422 if ((VerifyAfterGC || VerifyDuringGC) &&
4423 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4424 verify_after_remark();
4425 }
4426
4427 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4428
4429 // Change under the freelistLocks.
4430 _collectorState = Sweeping;
4431 // Call isAllClear() under bitMapLock
4432 assert(_modUnionTable.isAllClear(),
4433 "Should be clear by end of the final marking");
4434 assert(_ct->klass_rem_set()->mod_union_is_clear(),
4435 "Should be clear by end of the final marking");
4436 }
4437
4438 void CMSParInitialMarkTask::work(uint worker_id) {
4439 elapsedTimer _timer;
4440 ResourceMark rm;
4441 HandleMark hm;
4442
4443 // ---------- scan from roots --------------
4444 _timer.start();
4445 GenCollectedHeap* gch = GenCollectedHeap::heap();
4446 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4447
4448 // ---------- young gen roots --------------
4449 {
4450 work_on_young_gen_roots(worker_id, &par_mri_cl);
4451 _timer.stop();
4452 if (PrintCMSStatistics != 0) {
4453 gclog_or_tty->print_cr(
4454 "Finished young gen initial mark scan work in %dth thread: %3.3f sec",
4455 worker_id, _timer.seconds());
4456 }
4457 }
4458
4459 // ---------- remaining roots --------------
4460 _timer.reset();
4461 _timer.start();
4462
4463 CLDToOopClosure cld_closure(&par_mri_cl, true);
4464
4465 gch->gen_process_roots(_strong_roots_scope,
4466 _collector->_cmsGen->level(),
4467 false, // yg was scanned above
4468 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4469 _collector->should_unload_classes(),
4470 &par_mri_cl,
4471 NULL,
4472 &cld_closure);
4473 assert(_collector->should_unload_classes()
4474 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4475 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4476 _timer.stop();
4477 if (PrintCMSStatistics != 0) {
4478 gclog_or_tty->print_cr(
4479 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec",
4480 worker_id, _timer.seconds());
4481 }
4482 }
4483
4484 // Parallel remark task
4485 class CMSParRemarkTask: public CMSParMarkTask {
4486 CompactibleFreeListSpace* _cms_space;
4487
4488 // The per-thread work queues, available here for stealing.
4489 OopTaskQueueSet* _task_queues;
4490 ParallelTaskTerminator _term;
4491 StrongRootsScope* _strong_roots_scope;
4492
4493 public:
4494 // A value of 0 passed to n_workers will cause the number of
4495 // workers to be taken from the active workers in the work gang.
4496 CMSParRemarkTask(CMSCollector* collector,
4497 CompactibleFreeListSpace* cms_space,
4498 uint n_workers, FlexibleWorkGang* workers,
4499 OopTaskQueueSet* task_queues,
4500 StrongRootsScope* strong_roots_scope):
4501 CMSParMarkTask("Rescan roots and grey objects in parallel",
4502 collector, n_workers),
4503 _cms_space(cms_space),
4504 _task_queues(task_queues),
4505 _term(n_workers, task_queues),
4506 _strong_roots_scope(strong_roots_scope) { }
4507
4508 OopTaskQueueSet* task_queues() { return _task_queues; }
4509
4510 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4511
4512 ParallelTaskTerminator* terminator() { return &_term; }
4513 uint n_workers() { return _n_workers; }
4514
4515 void work(uint worker_id);
4516
4517 private:
4518 // ... of dirty cards in old space
4519 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4520 Par_MarkRefsIntoAndScanClosure* cl);
4521
4522 // ... work stealing for the above
4523 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4524 };
4525
4526 class RemarkKlassClosure : public KlassClosure {
4527 KlassToOopClosure _cm_klass_closure;
4528 public:
4529 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4530 void do_klass(Klass* k) {
4531 // Check if we have modified any oops in the Klass during the concurrent marking.
4532 if (k->has_accumulated_modified_oops()) {
4533 k->clear_accumulated_modified_oops();
4534
4535 // We could have transfered the current modified marks to the accumulated marks,
4536 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4537 } else if (k->has_modified_oops()) {
4538 // Don't clear anything, this info is needed by the next young collection.
4539 } else {
4540 // No modified oops in the Klass.
4541 return;
4542 }
4543
4544 // The klass has modified fields, need to scan the klass.
4545 _cm_klass_closure.do_klass(k);
4546 }
4547 };
4548
4549 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
4550 ParNewGeneration* young_gen = _collector->_young_gen;
4551 ContiguousSpace* eden_space = young_gen->eden();
4552 ContiguousSpace* from_space = young_gen->from();
4553 ContiguousSpace* to_space = young_gen->to();
4554
4555 HeapWord** eca = _collector->_eden_chunk_array;
4556 size_t ect = _collector->_eden_chunk_index;
4557 HeapWord** sca = _collector->_survivor_chunk_array;
4558 size_t sct = _collector->_survivor_chunk_index;
4559
4560 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4561 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4562
4563 do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
4564 do_young_space_rescan(worker_id, cl, from_space, sca, sct);
4565 do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
4566 }
4567
4568 // work_queue(i) is passed to the closure
4569 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
4570 // also is passed to do_dirty_card_rescan_tasks() and to
4571 // do_work_steal() to select the i-th task_queue.
4572
4573 void CMSParRemarkTask::work(uint worker_id) {
4574 elapsedTimer _timer;
4575 ResourceMark rm;
4576 HandleMark hm;
4577
4578 // ---------- rescan from roots --------------
4579 _timer.start();
4580 GenCollectedHeap* gch = GenCollectedHeap::heap();
4581 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4582 _collector->_span, _collector->ref_processor(),
4583 &(_collector->_markBitMap),
4584 work_queue(worker_id));
4585
4586 // Rescan young gen roots first since these are likely
4587 // coarsely partitioned and may, on that account, constitute
4588 // the critical path; thus, it's best to start off that
4589 // work first.
4590 // ---------- young gen roots --------------
4591 {
4592 work_on_young_gen_roots(worker_id, &par_mrias_cl);
4593 _timer.stop();
4594 if (PrintCMSStatistics != 0) {
4595 gclog_or_tty->print_cr(
4596 "Finished young gen rescan work in %dth thread: %3.3f sec",
4597 worker_id, _timer.seconds());
4598 }
4599 }
4600
4601 // ---------- remaining roots --------------
4602 _timer.reset();
4603 _timer.start();
4604 gch->gen_process_roots(_strong_roots_scope,
4605 _collector->_cmsGen->level(),
4606 false, // yg was scanned above
4607 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4608 _collector->should_unload_classes(),
4609 &par_mrias_cl,
4610 NULL,
4611 NULL); // The dirty klasses will be handled below
4612
4613 assert(_collector->should_unload_classes()
4614 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4615 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4616 _timer.stop();
4617 if (PrintCMSStatistics != 0) {
4618 gclog_or_tty->print_cr(
4619 "Finished remaining root rescan work in %dth thread: %3.3f sec",
4620 worker_id, _timer.seconds());
4621 }
4622
4623 // ---------- unhandled CLD scanning ----------
4624 if (worker_id == 0) { // Single threaded at the moment.
4625 _timer.reset();
4626 _timer.start();
4627
4628 // Scan all new class loader data objects and new dependencies that were
4629 // introduced during concurrent marking.
4630 ResourceMark rm;
4631 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4632 for (int i = 0; i < array->length(); i++) {
4633 par_mrias_cl.do_class_loader_data(array->at(i));
4634 }
4635
4636 // We don't need to keep track of new CLDs anymore.
4637 ClassLoaderDataGraph::remember_new_clds(false);
4638
4639 _timer.stop();
4640 if (PrintCMSStatistics != 0) {
4641 gclog_or_tty->print_cr(
4642 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec",
4643 worker_id, _timer.seconds());
4644 }
4645 }
4646
4647 // ---------- dirty klass scanning ----------
4648 if (worker_id == 0) { // Single threaded at the moment.
4649 _timer.reset();
4650 _timer.start();
4651
4652 // Scan all classes that was dirtied during the concurrent marking phase.
4653 RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
4654 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
4655
4656 _timer.stop();
4657 if (PrintCMSStatistics != 0) {
4658 gclog_or_tty->print_cr(
4659 "Finished dirty klass scanning work in %dth thread: %3.3f sec",
4660 worker_id, _timer.seconds());
4661 }
4662 }
4663
4664 // We might have added oops to ClassLoaderData::_handles during the
4665 // concurrent marking phase. These oops point to newly allocated objects
4666 // that are guaranteed to be kept alive. Either by the direct allocation
4667 // code, or when the young collector processes the roots. Hence,
4668 // we don't have to revisit the _handles block during the remark phase.
4669
4670 // ---------- rescan dirty cards ------------
4671 _timer.reset();
4672 _timer.start();
4673
4674 // Do the rescan tasks for each of the two spaces
4675 // (cms_space) in turn.
4676 // "worker_id" is passed to select the task_queue for "worker_id"
4677 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4678 _timer.stop();
4679 if (PrintCMSStatistics != 0) {
4680 gclog_or_tty->print_cr(
4681 "Finished dirty card rescan work in %dth thread: %3.3f sec",
4682 worker_id, _timer.seconds());
4683 }
4684
4685 // ---------- steal work from other threads ...
4686 // ---------- ... and drain overflow list.
4687 _timer.reset();
4688 _timer.start();
4689 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4690 _timer.stop();
4691 if (PrintCMSStatistics != 0) {
4692 gclog_or_tty->print_cr(
4693 "Finished work stealing in %dth thread: %3.3f sec",
4694 worker_id, _timer.seconds());
4695 }
4696 }
4697
4698 // Note that parameter "i" is not used.
4699 void
4700 CMSParMarkTask::do_young_space_rescan(uint worker_id,
4701 OopsInGenClosure* cl, ContiguousSpace* space,
4702 HeapWord** chunk_array, size_t chunk_top) {
4703 // Until all tasks completed:
4704 // . claim an unclaimed task
4705 // . compute region boundaries corresponding to task claimed
4706 // using chunk_array
4707 // . par_oop_iterate(cl) over that region
4708
4709 ResourceMark rm;
4710 HandleMark hm;
4711
4712 SequentialSubTasksDone* pst = space->par_seq_tasks();
4713
4714 uint nth_task = 0;
4715 uint n_tasks = pst->n_tasks();
4716
4717 if (n_tasks > 0) {
4718 assert(pst->valid(), "Uninitialized use?");
4719 HeapWord *start, *end;
4720 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4721 // We claimed task # nth_task; compute its boundaries.
4722 if (chunk_top == 0) { // no samples were taken
4723 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4724 start = space->bottom();
4725 end = space->top();
4726 } else if (nth_task == 0) {
4727 start = space->bottom();
4728 end = chunk_array[nth_task];
4729 } else if (nth_task < (uint)chunk_top) {
4730 assert(nth_task >= 1, "Control point invariant");
4731 start = chunk_array[nth_task - 1];
4732 end = chunk_array[nth_task];
4733 } else {
4734 assert(nth_task == (uint)chunk_top, "Control point invariant");
4735 start = chunk_array[chunk_top - 1];
4736 end = space->top();
4737 }
4738 MemRegion mr(start, end);
4739 // Verify that mr is in space
4740 assert(mr.is_empty() || space->used_region().contains(mr),
4741 "Should be in space");
4742 // Verify that "start" is an object boundary
4743 assert(mr.is_empty() || oop(mr.start())->is_oop(),
4744 "Should be an oop");
4745 space->par_oop_iterate(mr, cl);
4746 }
4747 pst->all_tasks_completed();
4748 }
4749 }
4750
4751 void
4752 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4753 CompactibleFreeListSpace* sp, int i,
4754 Par_MarkRefsIntoAndScanClosure* cl) {
4755 // Until all tasks completed:
4756 // . claim an unclaimed task
4757 // . compute region boundaries corresponding to task claimed
4758 // . transfer dirty bits ct->mut for that region
4759 // . apply rescanclosure to dirty mut bits for that region
4760
4761 ResourceMark rm;
4762 HandleMark hm;
4763
4764 OopTaskQueue* work_q = work_queue(i);
4765 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4766 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4767 // CAUTION: This closure has state that persists across calls to
4768 // the work method dirty_range_iterate_clear() in that it has
4769 // embedded in it a (subtype of) UpwardsObjectClosure. The
4770 // use of that state in the embedded UpwardsObjectClosure instance
4771 // assumes that the cards are always iterated (even if in parallel
4772 // by several threads) in monotonically increasing order per each
4773 // thread. This is true of the implementation below which picks
4774 // card ranges (chunks) in monotonically increasing order globally
4775 // and, a-fortiori, in monotonically increasing order per thread
4776 // (the latter order being a subsequence of the former).
4777 // If the work code below is ever reorganized into a more chaotic
4778 // work-partitioning form than the current "sequential tasks"
4779 // paradigm, the use of that persistent state will have to be
4780 // revisited and modified appropriately. See also related
4781 // bug 4756801 work on which should examine this code to make
4782 // sure that the changes there do not run counter to the
4783 // assumptions made here and necessary for correctness and
4784 // efficiency. Note also that this code might yield inefficient
4785 // behavior in the case of very large objects that span one or
4786 // more work chunks. Such objects would potentially be scanned
4787 // several times redundantly. Work on 4756801 should try and
4788 // address that performance anomaly if at all possible. XXX
4789 MemRegion full_span = _collector->_span;
4790 CMSBitMap* bm = &(_collector->_markBitMap); // shared
4791 MarkFromDirtyCardsClosure
4792 greyRescanClosure(_collector, full_span, // entire span of interest
4793 sp, bm, work_q, cl);
4794
4795 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4796 assert(pst->valid(), "Uninitialized use?");
4797 uint nth_task = 0;
4798 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
4799 MemRegion span = sp->used_region();
4800 HeapWord* start_addr = span.start();
4801 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
4802 alignment);
4803 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4804 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
4805 start_addr, "Check alignment");
4806 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
4807 chunk_size, "Check alignment");
4808
4809 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4810 // Having claimed the nth_task, compute corresponding mem-region,
4811 // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4812 // The alignment restriction ensures that we do not need any
4813 // synchronization with other gang-workers while setting or
4814 // clearing bits in thus chunk of the MUT.
4815 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4816 start_addr + (nth_task+1)*chunk_size);
4817 // The last chunk's end might be way beyond end of the
4818 // used region. In that case pull back appropriately.
4819 if (this_span.end() > end_addr) {
4820 this_span.set_end(end_addr);
4821 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4822 }
4823 // Iterate over the dirty cards covering this chunk, marking them
4824 // precleaned, and setting the corresponding bits in the mod union
4825 // table. Since we have been careful to partition at Card and MUT-word
4826 // boundaries no synchronization is needed between parallel threads.
4827 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
4828 &modUnionClosure);
4829
4830 // Having transferred these marks into the modUnionTable,
4831 // rescan the marked objects on the dirty cards in the modUnionTable.
4832 // Even if this is at a synchronous collection, the initial marking
4833 // may have been done during an asynchronous collection so there
4834 // may be dirty bits in the mod-union table.
4835 _collector->_modUnionTable.dirty_range_iterate_clear(
4836 this_span, &greyRescanClosure);
4837 _collector->_modUnionTable.verifyNoOneBitsInRange(
4838 this_span.start(),
4839 this_span.end());
4840 }
4841 pst->all_tasks_completed(); // declare that i am done
4842 }
4843
4844 // . see if we can share work_queues with ParNew? XXX
4845 void
4846 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
4847 int* seed) {
4848 OopTaskQueue* work_q = work_queue(i);
4849 NOT_PRODUCT(int num_steals = 0;)
4850 oop obj_to_scan;
4851 CMSBitMap* bm = &(_collector->_markBitMap);
4852
4853 while (true) {
4854 // Completely finish any left over work from (an) earlier round(s)
4855 cl->trim_queue(0);
4856 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4857 (size_t)ParGCDesiredObjsFromOverflowList);
4858 // Now check if there's any work in the overflow list
4859 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4860 // only affects the number of attempts made to get work from the
4861 // overflow list and does not affect the number of workers. Just
4862 // pass ParallelGCThreads so this behavior is unchanged.
4863 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4864 work_q,
4865 ParallelGCThreads)) {
4866 // found something in global overflow list;
4867 // not yet ready to go stealing work from others.
4868 // We'd like to assert(work_q->size() != 0, ...)
4869 // because we just took work from the overflow list,
4870 // but of course we can't since all of that could have
4871 // been already stolen from us.
4872 // "He giveth and He taketh away."
4873 continue;
4874 }
4875 // Verify that we have no work before we resort to stealing
4876 assert(work_q->size() == 0, "Have work, shouldn't steal");
4877 // Try to steal from other queues that have work
4878 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4879 NOT_PRODUCT(num_steals++;)
4880 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
4881 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4882 // Do scanning work
4883 obj_to_scan->oop_iterate(cl);
4884 // Loop around, finish this work, and try to steal some more
4885 } else if (terminator()->offer_termination()) {
4886 break; // nirvana from the infinite cycle
4887 }
4888 }
4889 NOT_PRODUCT(
4890 if (PrintCMSStatistics != 0) {
4891 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
4892 }
4893 )
4894 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4895 "Else our work is not yet done");
4896 }
4897
4898 // Record object boundaries in _eden_chunk_array by sampling the eden
4899 // top in the slow-path eden object allocation code path and record
4900 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4901 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4902 // sampling in sample_eden() that activates during the part of the
4903 // preclean phase.
4904 void CMSCollector::sample_eden_chunk() {
4905 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4906 if (_eden_chunk_lock->try_lock()) {
4907 // Record a sample. This is the critical section. The contents
4908 // of the _eden_chunk_array have to be non-decreasing in the
4909 // address order.
4910 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4911 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4912 "Unexpected state of Eden");
4913 if (_eden_chunk_index == 0 ||
4914 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4915 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4916 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4917 _eden_chunk_index++; // commit sample
4918 }
4919 _eden_chunk_lock->unlock();
4920 }
4921 }
4922 }
4923
4924 // Return a thread-local PLAB recording array, as appropriate.
4925 void* CMSCollector::get_data_recorder(int thr_num) {
4926 if (_survivor_plab_array != NULL &&
4927 (CMSPLABRecordAlways ||
4928 (_collectorState > Marking && _collectorState < FinalMarking))) {
4929 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4930 ChunkArray* ca = &_survivor_plab_array[thr_num];
4931 ca->reset(); // clear it so that fresh data is recorded
4932 return (void*) ca;
4933 } else {
4934 return NULL;
4935 }
4936 }
4937
4938 // Reset all the thread-local PLAB recording arrays
4939 void CMSCollector::reset_survivor_plab_arrays() {
4940 for (uint i = 0; i < ParallelGCThreads; i++) {
4941 _survivor_plab_array[i].reset();
4942 }
4943 }
4944
4945 // Merge the per-thread plab arrays into the global survivor chunk
4946 // array which will provide the partitioning of the survivor space
4947 // for CMS initial scan and rescan.
4948 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4949 int no_of_gc_threads) {
4950 assert(_survivor_plab_array != NULL, "Error");
4951 assert(_survivor_chunk_array != NULL, "Error");
4952 assert(_collectorState == FinalMarking ||
4953 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4954 for (int j = 0; j < no_of_gc_threads; j++) {
4955 _cursor[j] = 0;
4956 }
4957 HeapWord* top = surv->top();
4958 size_t i;
4959 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
4960 HeapWord* min_val = top; // Higher than any PLAB address
4961 uint min_tid = 0; // position of min_val this round
4962 for (int j = 0; j < no_of_gc_threads; j++) {
4963 ChunkArray* cur_sca = &_survivor_plab_array[j];
4964 if (_cursor[j] == cur_sca->end()) {
4965 continue;
4966 }
4967 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4968 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4969 assert(surv->used_region().contains(cur_val), "Out of bounds value");
4970 if (cur_val < min_val) {
4971 min_tid = j;
4972 min_val = cur_val;
4973 } else {
4974 assert(cur_val < top, "All recorded addresses should be less");
4975 }
4976 }
4977 // At this point min_val and min_tid are respectively
4978 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4979 // and the thread (j) that witnesses that address.
4980 // We record this address in the _survivor_chunk_array[i]
4981 // and increment _cursor[min_tid] prior to the next round i.
4982 if (min_val == top) {
4983 break;
4984 }
4985 _survivor_chunk_array[i] = min_val;
4986 _cursor[min_tid]++;
4987 }
4988 // We are all done; record the size of the _survivor_chunk_array
4989 _survivor_chunk_index = i; // exclusive: [0, i)
4990 if (PrintCMSStatistics > 0) {
4991 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4992 }
4993 // Verify that we used up all the recorded entries
4994 #ifdef ASSERT
4995 size_t total = 0;
4996 for (int j = 0; j < no_of_gc_threads; j++) {
4997 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4998 total += _cursor[j];
4999 }
5000 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5001 // Check that the merged array is in sorted order
5002 if (total > 0) {
5003 for (size_t i = 0; i < total - 1; i++) {
5004 if (PrintCMSStatistics > 0) {
5005 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5006 i, p2i(_survivor_chunk_array[i]));
5007 }
5008 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5009 "Not sorted");
5010 }
5011 }
5012 #endif // ASSERT
5013 }
5014
5015 // Set up the space's par_seq_tasks structure for work claiming
5016 // for parallel initial scan and rescan of young gen.
5017 // See ParRescanTask where this is currently used.
5018 void
5019 CMSCollector::
5020 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5021 assert(n_threads > 0, "Unexpected n_threads argument");
5022
5023 // Eden space
5024 if (!_young_gen->eden()->is_empty()) {
5025 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
5026 assert(!pst->valid(), "Clobbering existing data?");
5027 // Each valid entry in [0, _eden_chunk_index) represents a task.
5028 size_t n_tasks = _eden_chunk_index + 1;
5029 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5030 // Sets the condition for completion of the subtask (how many threads
5031 // need to finish in order to be done).
5032 pst->set_n_threads(n_threads);
5033 pst->set_n_tasks((int)n_tasks);
5034 }
5035
5036 // Merge the survivor plab arrays into _survivor_chunk_array
5037 if (_survivor_plab_array != NULL) {
5038 merge_survivor_plab_arrays(_young_gen->from(), n_threads);
5039 } else {
5040 assert(_survivor_chunk_index == 0, "Error");
5041 }
5042
5043 // To space
5044 {
5045 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
5046 assert(!pst->valid(), "Clobbering existing data?");
5047 // Sets the condition for completion of the subtask (how many threads
5048 // need to finish in order to be done).
5049 pst->set_n_threads(n_threads);
5050 pst->set_n_tasks(1);
5051 assert(pst->valid(), "Error");
5052 }
5053
5054 // From space
5055 {
5056 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
5057 assert(!pst->valid(), "Clobbering existing data?");
5058 size_t n_tasks = _survivor_chunk_index + 1;
5059 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5060 // Sets the condition for completion of the subtask (how many threads
5061 // need to finish in order to be done).
5062 pst->set_n_threads(n_threads);
5063 pst->set_n_tasks((int)n_tasks);
5064 assert(pst->valid(), "Error");
5065 }
5066 }
5067
5068 // Parallel version of remark
5069 void CMSCollector::do_remark_parallel() {
5070 GenCollectedHeap* gch = GenCollectedHeap::heap();
5071 FlexibleWorkGang* workers = gch->workers();
5072 assert(workers != NULL, "Need parallel worker threads.");
5073 // Choose to use the number of GC workers most recently set
5074 // into "active_workers". If active_workers is not set, set it
5075 // to ParallelGCThreads.
5076 uint n_workers = workers->active_workers();
5077 if (n_workers == 0) {
5078 assert(n_workers > 0, "Should have been set during scavenge");
5079 n_workers = ParallelGCThreads;
5080 workers->set_active_workers(n_workers);
5081 }
5082 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5083
5084 StrongRootsScope srs(n_workers);
5085
5086 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
5087
5088 // We won't be iterating over the cards in the card table updating
5089 // the younger_gen cards, so we shouldn't call the following else
5090 // the verification code as well as subsequent younger_refs_iterate
5091 // code would get confused. XXX
5092 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5093
5094 // The young gen rescan work will not be done as part of
5095 // process_roots (which currently doesn't know how to
5096 // parallelize such a scan), but rather will be broken up into
5097 // a set of parallel tasks (via the sampling that the [abortable]
5098 // preclean phase did of eden, plus the [two] tasks of
5099 // scanning the [two] survivor spaces. Further fine-grain
5100 // parallelization of the scanning of the survivor spaces
5101 // themselves, and of precleaning of the younger gen itself
5102 // is deferred to the future.
5103 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5104
5105 // The dirty card rescan work is broken up into a "sequence"
5106 // of parallel tasks (per constituent space) that are dynamically
5107 // claimed by the parallel threads.
5108 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5109
5110 // It turns out that even when we're using 1 thread, doing the work in a
5111 // separate thread causes wide variance in run times. We can't help this
5112 // in the multi-threaded case, but we special-case n=1 here to get
5113 // repeatable measurements of the 1-thread overhead of the parallel code.
5114 if (n_workers > 1) {
5115 // Make refs discovery MT-safe, if it isn't already: it may not
5116 // necessarily be so, since it's possible that we are doing
5117 // ST marking.
5118 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5119 workers->run_task(&tsk);
5120 } else {
5121 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5122 tsk.work(0);
5123 }
5124
5125 // restore, single-threaded for now, any preserved marks
5126 // as a result of work_q overflow
5127 restore_preserved_marks_if_any();
5128 }
5129
5130 // Non-parallel version of remark
5131 void CMSCollector::do_remark_non_parallel() {
5132 ResourceMark rm;
5133 HandleMark hm;
5134 GenCollectedHeap* gch = GenCollectedHeap::heap();
5135 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5136
5137 MarkRefsIntoAndScanClosure
5138 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
5139 &_markStack, this,
5140 false /* should_yield */, false /* not precleaning */);
5141 MarkFromDirtyCardsClosure
5142 markFromDirtyCardsClosure(this, _span,
5143 NULL, // space is set further below
5144 &_markBitMap, &_markStack, &mrias_cl);
5145 {
5146 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5147 // Iterate over the dirty cards, setting the corresponding bits in the
5148 // mod union table.
5149 {
5150 ModUnionClosure modUnionClosure(&_modUnionTable);
5151 _ct->ct_bs()->dirty_card_iterate(
5152 _cmsGen->used_region(),
5153 &modUnionClosure);
5154 }
5155 // Having transferred these marks into the modUnionTable, we just need
5156 // to rescan the marked objects on the dirty cards in the modUnionTable.
5157 // The initial marking may have been done during an asynchronous
5158 // collection so there may be dirty bits in the mod-union table.
5159 const int alignment =
5160 CardTableModRefBS::card_size * BitsPerWord;
5161 {
5162 // ... First handle dirty cards in CMS gen
5163 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5164 MemRegion ur = _cmsGen->used_region();
5165 HeapWord* lb = ur.start();
5166 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5167 MemRegion cms_span(lb, ub);
5168 _modUnionTable.dirty_range_iterate_clear(cms_span,
5169 &markFromDirtyCardsClosure);
5170 verify_work_stacks_empty();
5171 if (PrintCMSStatistics != 0) {
5172 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5173 markFromDirtyCardsClosure.num_dirty_cards());
5174 }
5175 }
5176 }
5177 if (VerifyDuringGC &&
5178 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5179 HandleMark hm; // Discard invalid handles created during verification
5180 Universe::verify();
5181 }
5182 {
5183 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5184
5185 verify_work_stacks_empty();
5186
5187 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5188 StrongRootsScope srs(1);
5189
5190 gch->gen_process_roots(&srs,
5191 _cmsGen->level(),
5192 true, // younger gens as roots
5193 GenCollectedHeap::ScanningOption(roots_scanning_options()),
5194 should_unload_classes(),
5195 &mrias_cl,
5196 NULL,
5197 NULL); // The dirty klasses will be handled below
5198
5199 assert(should_unload_classes()
5200 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
5201 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5202 }
5203
5204 {
5205 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5206
5207 verify_work_stacks_empty();
5208
5209 // Scan all class loader data objects that might have been introduced
5210 // during concurrent marking.
5211 ResourceMark rm;
5212 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5213 for (int i = 0; i < array->length(); i++) {
5214 mrias_cl.do_class_loader_data(array->at(i));
5215 }
5216
5217 // We don't need to keep track of new CLDs anymore.
5218 ClassLoaderDataGraph::remember_new_clds(false);
5219
5220 verify_work_stacks_empty();
5221 }
5222
5223 {
5224 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5225
5226 verify_work_stacks_empty();
5227
5228 RemarkKlassClosure remark_klass_closure(&mrias_cl);
5229 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5230
5231 verify_work_stacks_empty();
5232 }
5233
5234 // We might have added oops to ClassLoaderData::_handles during the
5235 // concurrent marking phase. These oops point to newly allocated objects
5236 // that are guaranteed to be kept alive. Either by the direct allocation
5237 // code, or when the young collector processes the roots. Hence,
5238 // we don't have to revisit the _handles block during the remark phase.
5239
5240 verify_work_stacks_empty();
5241 // Restore evacuated mark words, if any, used for overflow list links
5242 if (!CMSOverflowEarlyRestoration) {
5243 restore_preserved_marks_if_any();
5244 }
5245 verify_overflow_empty();
5246 }
5247
5248 ////////////////////////////////////////////////////////
5249 // Parallel Reference Processing Task Proxy Class
5250 ////////////////////////////////////////////////////////
5251 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5252 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5253 CMSCollector* _collector;
5254 CMSBitMap* _mark_bit_map;
5255 const MemRegion _span;
5256 ProcessTask& _task;
5257
5258 public:
5259 CMSRefProcTaskProxy(ProcessTask& task,
5260 CMSCollector* collector,
5261 const MemRegion& span,
5262 CMSBitMap* mark_bit_map,
5263 AbstractWorkGang* workers,
5264 OopTaskQueueSet* task_queues):
5265 // XXX Should superclass AGTWOQ also know about AWG since it knows
5266 // about the task_queues used by the AWG? Then it could initialize
5267 // the terminator() object. See 6984287. The set_for_termination()
5268 // below is a temporary band-aid for the regression in 6984287.
5269 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5270 task_queues),
5271 _task(task),
5272 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5273 {
5274 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5275 "Inconsistency in _span");
5276 set_for_termination(workers->active_workers());
5277 }
5278
5279 OopTaskQueueSet* task_queues() { return queues(); }
5280
5281 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5282
5283 void do_work_steal(int i,
5284 CMSParDrainMarkingStackClosure* drain,
5285 CMSParKeepAliveClosure* keep_alive,
5286 int* seed);
5287
5288 virtual void work(uint worker_id);
5289 };
5290
5291 void CMSRefProcTaskProxy::work(uint worker_id) {
5292 ResourceMark rm;
5293 HandleMark hm;
5294 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5295 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5296 _mark_bit_map,
5297 work_queue(worker_id));
5298 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5299 _mark_bit_map,
5300 work_queue(worker_id));
5301 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5302 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5303 if (_task.marks_oops_alive()) {
5304 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5305 _collector->hash_seed(worker_id));
5306 }
5307 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5308 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5309 }
5310
5311 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5312 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5313 EnqueueTask& _task;
5314
5315 public:
5316 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5317 : AbstractGangTask("Enqueue reference objects in parallel"),
5318 _task(task)
5319 { }
5320
5321 virtual void work(uint worker_id)
5322 {
5323 _task.work(worker_id);
5324 }
5325 };
5326
5327 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5328 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5329 _span(span),
5330 _bit_map(bit_map),
5331 _work_queue(work_queue),
5332 _mark_and_push(collector, span, bit_map, work_queue),
5333 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5334 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5335 { }
5336
5337 // . see if we can share work_queues with ParNew? XXX
5338 void CMSRefProcTaskProxy::do_work_steal(int i,
5339 CMSParDrainMarkingStackClosure* drain,
5340 CMSParKeepAliveClosure* keep_alive,
5341 int* seed) {
5342 OopTaskQueue* work_q = work_queue(i);
5343 NOT_PRODUCT(int num_steals = 0;)
5344 oop obj_to_scan;
5345
5346 while (true) {
5347 // Completely finish any left over work from (an) earlier round(s)
5348 drain->trim_queue(0);
5349 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5350 (size_t)ParGCDesiredObjsFromOverflowList);
5351 // Now check if there's any work in the overflow list
5352 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5353 // only affects the number of attempts made to get work from the
5354 // overflow list and does not affect the number of workers. Just
5355 // pass ParallelGCThreads so this behavior is unchanged.
5356 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5357 work_q,
5358 ParallelGCThreads)) {
5359 // Found something in global overflow list;
5360 // not yet ready to go stealing work from others.
5361 // We'd like to assert(work_q->size() != 0, ...)
5362 // because we just took work from the overflow list,
5363 // but of course we can't, since all of that might have
5364 // been already stolen from us.
5365 continue;
5366 }
5367 // Verify that we have no work before we resort to stealing
5368 assert(work_q->size() == 0, "Have work, shouldn't steal");
5369 // Try to steal from other queues that have work
5370 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5371 NOT_PRODUCT(num_steals++;)
5372 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5373 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5374 // Do scanning work
5375 obj_to_scan->oop_iterate(keep_alive);
5376 // Loop around, finish this work, and try to steal some more
5377 } else if (terminator()->offer_termination()) {
5378 break; // nirvana from the infinite cycle
5379 }
5380 }
5381 NOT_PRODUCT(
5382 if (PrintCMSStatistics != 0) {
5383 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5384 }
5385 )
5386 }
5387
5388 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5389 {
5390 GenCollectedHeap* gch = GenCollectedHeap::heap();
5391 FlexibleWorkGang* workers = gch->workers();
5392 assert(workers != NULL, "Need parallel worker threads.");
5393 CMSRefProcTaskProxy rp_task(task, &_collector,
5394 _collector.ref_processor()->span(),
5395 _collector.markBitMap(),
5396 workers, _collector.task_queues());
5397 workers->run_task(&rp_task);
5398 }
5399
5400 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5401 {
5402
5403 GenCollectedHeap* gch = GenCollectedHeap::heap();
5404 FlexibleWorkGang* workers = gch->workers();
5405 assert(workers != NULL, "Need parallel worker threads.");
5406 CMSRefEnqueueTaskProxy enq_task(task);
5407 workers->run_task(&enq_task);
5408 }
5409
5410 void CMSCollector::refProcessingWork() {
5411 ResourceMark rm;
5412 HandleMark hm;
5413
5414 ReferenceProcessor* rp = ref_processor();
5415 assert(rp->span().equals(_span), "Spans should be equal");
5416 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5417 // Process weak references.
5418 rp->setup_policy(false);
5419 verify_work_stacks_empty();
5420
5421 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5422 &_markStack, false /* !preclean */);
5423 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5424 _span, &_markBitMap, &_markStack,
5425 &cmsKeepAliveClosure, false /* !preclean */);
5426 {
5427 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5428
5429 ReferenceProcessorStats stats;
5430 if (rp->processing_is_mt()) {
5431 // Set the degree of MT here. If the discovery is done MT, there
5432 // may have been a different number of threads doing the discovery
5433 // and a different number of discovered lists may have Ref objects.
5434 // That is OK as long as the Reference lists are balanced (see
5435 // balance_all_queues() and balance_queues()).
5436 GenCollectedHeap* gch = GenCollectedHeap::heap();
5437 uint active_workers = ParallelGCThreads;
5438 FlexibleWorkGang* workers = gch->workers();
5439 if (workers != NULL) {
5440 active_workers = workers->active_workers();
5441 // The expectation is that active_workers will have already
5442 // been set to a reasonable value. If it has not been set,
5443 // investigate.
5444 assert(active_workers > 0, "Should have been set during scavenge");
5445 }
5446 rp->set_active_mt_degree(active_workers);
5447 CMSRefProcTaskExecutor task_executor(*this);
5448 stats = rp->process_discovered_references(&_is_alive_closure,
5449 &cmsKeepAliveClosure,
5450 &cmsDrainMarkingStackClosure,
5451 &task_executor,
5452 _gc_timer_cm,
5453 _gc_tracer_cm->gc_id());
5454 } else {
5455 stats = rp->process_discovered_references(&_is_alive_closure,
5456 &cmsKeepAliveClosure,
5457 &cmsDrainMarkingStackClosure,
5458 NULL,
5459 _gc_timer_cm,
5460 _gc_tracer_cm->gc_id());
5461 }
5462 _gc_tracer_cm->report_gc_reference_stats(stats);
5463
5464 }
5465
5466 // This is the point where the entire marking should have completed.
5467 verify_work_stacks_empty();
5468
5469 if (should_unload_classes()) {
5470 {
5471 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5472
5473 // Unload classes and purge the SystemDictionary.
5474 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5475
5476 // Unload nmethods.
5477 CodeCache::do_unloading(&_is_alive_closure, purged_class);
5478
5479 // Prune dead klasses from subklass/sibling/implementor lists.
5480 Klass::clean_weak_klass_links(&_is_alive_closure);
5481 }
5482
5483 {
5484 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5485 // Clean up unreferenced symbols in symbol table.
5486 SymbolTable::unlink();
5487 }
5488
5489 {
5490 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5491 // Delete entries for dead interned strings.
5492 StringTable::unlink(&_is_alive_closure);
5493 }
5494 }
5495
5496
5497 // Restore any preserved marks as a result of mark stack or
5498 // work queue overflow
5499 restore_preserved_marks_if_any(); // done single-threaded for now
5500
5501 rp->set_enqueuing_is_done(true);
5502 if (rp->processing_is_mt()) {
5503 rp->balance_all_queues();
5504 CMSRefProcTaskExecutor task_executor(*this);
5505 rp->enqueue_discovered_references(&task_executor);
5506 } else {
5507 rp->enqueue_discovered_references(NULL);
5508 }
5509 rp->verify_no_references_recorded();
5510 assert(!rp->discovery_enabled(), "should have been disabled");
5511 }
5512
5513 #ifndef PRODUCT
5514 void CMSCollector::check_correct_thread_executing() {
5515 Thread* t = Thread::current();
5516 // Only the VM thread or the CMS thread should be here.
5517 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5518 "Unexpected thread type");
5519 // If this is the vm thread, the foreground process
5520 // should not be waiting. Note that _foregroundGCIsActive is
5521 // true while the foreground collector is waiting.
5522 if (_foregroundGCShouldWait) {
5523 // We cannot be the VM thread
5524 assert(t->is_ConcurrentGC_thread(),
5525 "Should be CMS thread");
5526 } else {
5527 // We can be the CMS thread only if we are in a stop-world
5528 // phase of CMS collection.
5529 if (t->is_ConcurrentGC_thread()) {
5530 assert(_collectorState == InitialMarking ||
5531 _collectorState == FinalMarking,
5532 "Should be a stop-world phase");
5533 // The CMS thread should be holding the CMS_token.
5534 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5535 "Potential interference with concurrently "
5536 "executing VM thread");
5537 }
5538 }
5539 }
5540 #endif
5541
5542 void CMSCollector::sweep() {
5543 assert(_collectorState == Sweeping, "just checking");
5544 check_correct_thread_executing();
5545 verify_work_stacks_empty();
5546 verify_overflow_empty();
5547 increment_sweep_count();
5548 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5549
5550 _inter_sweep_timer.stop();
5551 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5552
5553 assert(!_intra_sweep_timer.is_active(), "Should not be active");
5554 _intra_sweep_timer.reset();
5555 _intra_sweep_timer.start();
5556 {
5557 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5558 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails);
5559 // First sweep the old gen
5560 {
5561 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5562 bitMapLock());
5563 sweepWork(_cmsGen);
5564 }
5565
5566 // Update Universe::_heap_*_at_gc figures.
5567 // We need all the free list locks to make the abstract state
5568 // transition from Sweeping to Resetting. See detailed note
5569 // further below.
5570 {
5571 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5572 // Update heap occupancy information which is used as
5573 // input to soft ref clearing policy at the next gc.
5574 Universe::update_heap_info_at_gc();
5575 _collectorState = Resizing;
5576 }
5577 }
5578 verify_work_stacks_empty();
5579 verify_overflow_empty();
5580
5581 if (should_unload_classes()) {
5582 // Delay purge to the beginning of the next safepoint. Metaspace::contains
5583 // requires that the virtual spaces are stable and not deleted.
5584 ClassLoaderDataGraph::set_should_purge(true);
5585 }
5586
5587 _intra_sweep_timer.stop();
5588 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5589
5590 _inter_sweep_timer.reset();
5591 _inter_sweep_timer.start();
5592
5593 // We need to use a monotonically non-decreasing time in ms
5594 // or we will see time-warp warnings and os::javaTimeMillis()
5595 // does not guarantee monotonicity.
5596 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5597 update_time_of_last_gc(now);
5598
5599 // NOTE on abstract state transitions:
5600 // Mutators allocate-live and/or mark the mod-union table dirty
5601 // based on the state of the collection. The former is done in
5602 // the interval [Marking, Sweeping] and the latter in the interval
5603 // [Marking, Sweeping). Thus the transitions into the Marking state
5604 // and out of the Sweeping state must be synchronously visible
5605 // globally to the mutators.
5606 // The transition into the Marking state happens with the world
5607 // stopped so the mutators will globally see it. Sweeping is
5608 // done asynchronously by the background collector so the transition
5609 // from the Sweeping state to the Resizing state must be done
5610 // under the freelistLock (as is the check for whether to
5611 // allocate-live and whether to dirty the mod-union table).
5612 assert(_collectorState == Resizing, "Change of collector state to"
5613 " Resizing must be done under the freelistLocks (plural)");
5614
5615 // Now that sweeping has been completed, we clear
5616 // the incremental_collection_failed flag,
5617 // thus inviting a younger gen collection to promote into
5618 // this generation. If such a promotion may still fail,
5619 // the flag will be set again when a young collection is
5620 // attempted.
5621 GenCollectedHeap* gch = GenCollectedHeap::heap();
5622 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
5623 gch->update_full_collections_completed(_collection_count_start);
5624 }
5625
5626 // FIX ME!!! Looks like this belongs in CFLSpace, with
5627 // CMSGen merely delegating to it.
5628 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5629 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5630 HeapWord* minAddr = _cmsSpace->bottom();
5631 HeapWord* largestAddr =
5632 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5633 if (largestAddr == NULL) {
5634 // The dictionary appears to be empty. In this case
5635 // try to coalesce at the end of the heap.
5636 largestAddr = _cmsSpace->end();
5637 }
5638 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5639 size_t nearLargestOffset =
5640 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5641 if (PrintFLSStatistics != 0) {
5642 gclog_or_tty->print_cr(
5643 "CMS: Large Block: " PTR_FORMAT ";"
5644 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5645 p2i(largestAddr),
5646 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5647 }
5648 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5649 }
5650
5651 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5652 return addr >= _cmsSpace->nearLargestChunk();
5653 }
5654
5655 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5656 return _cmsSpace->find_chunk_at_end();
5657 }
5658
5659 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
5660 bool full) {
5661 // The next lower level has been collected. Gather any statistics
5662 // that are of interest at this point.
5663 if (!full && (current_level + 1) == level()) {
5664 // Gather statistics on the young generation collection.
5665 collector()->stats().record_gc0_end(used());
5666 }
5667 }
5668
5669 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen) {
5670 // We iterate over the space(s) underlying this generation,
5671 // checking the mark bit map to see if the bits corresponding
5672 // to specific blocks are marked or not. Blocks that are
5673 // marked are live and are not swept up. All remaining blocks
5674 // are swept up, with coalescing on-the-fly as we sweep up
5675 // contiguous free and/or garbage blocks:
5676 // We need to ensure that the sweeper synchronizes with allocators
5677 // and stop-the-world collectors. In particular, the following
5678 // locks are used:
5679 // . CMS token: if this is held, a stop the world collection cannot occur
5680 // . freelistLock: if this is held no allocation can occur from this
5681 // generation by another thread
5682 // . bitMapLock: if this is held, no other thread can access or update
5683 //
5684
5685 // Note that we need to hold the freelistLock if we use
5686 // block iterate below; else the iterator might go awry if
5687 // a mutator (or promotion) causes block contents to change
5688 // (for instance if the allocator divvies up a block).
5689 // If we hold the free list lock, for all practical purposes
5690 // young generation GC's can't occur (they'll usually need to
5691 // promote), so we might as well prevent all young generation
5692 // GC's while we do a sweeping step. For the same reason, we might
5693 // as well take the bit map lock for the entire duration
5694
5695 // check that we hold the requisite locks
5696 assert(have_cms_token(), "Should hold cms token");
5697 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5698 assert_lock_strong(gen->freelistLock());
5699 assert_lock_strong(bitMapLock());
5700
5701 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5702 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
5703 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5704 _inter_sweep_estimate.padded_average(),
5705 _intra_sweep_estimate.padded_average());
5706 gen->setNearLargestChunk();
5707
5708 {
5709 SweepClosure sweepClosure(this, gen, &_markBitMap, CMSYield);
5710 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5711 // We need to free-up/coalesce garbage/blocks from a
5712 // co-terminal free run. This is done in the SweepClosure
5713 // destructor; so, do not remove this scope, else the
5714 // end-of-sweep-census below will be off by a little bit.
5715 }
5716 gen->cmsSpace()->sweep_completed();
5717 gen->cmsSpace()->endSweepFLCensus(sweep_count());
5718 if (should_unload_classes()) { // unloaded classes this cycle,
5719 _concurrent_cycles_since_last_unload = 0; // ... reset count
5720 } else { // did not unload classes,
5721 _concurrent_cycles_since_last_unload++; // ... increment count
5722 }
5723 }
5724
5725 // Reset CMS data structures (for now just the marking bit map)
5726 // preparatory for the next cycle.
5727 void CMSCollector::reset(bool concurrent) {
5728 if (concurrent) {
5729 CMSTokenSyncWithLocks ts(true, bitMapLock());
5730
5731 // If the state is not "Resetting", the foreground thread
5732 // has done a collection and the resetting.
5733 if (_collectorState != Resetting) {
5734 assert(_collectorState == Idling, "The state should only change"
5735 " because the foreground collector has finished the collection");
5736 return;
5737 }
5738
5739 // Clear the mark bitmap (no grey objects to start with)
5740 // for the next cycle.
5741 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5742 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails);
5743
5744 HeapWord* curAddr = _markBitMap.startWord();
5745 while (curAddr < _markBitMap.endWord()) {
5746 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
5747 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5748 _markBitMap.clear_large_range(chunk);
5749 if (ConcurrentMarkSweepThread::should_yield() &&
5750 !foregroundGCIsActive() &&
5751 CMSYield) {
5752 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5753 "CMS thread should hold CMS token");
5754 assert_lock_strong(bitMapLock());
5755 bitMapLock()->unlock();
5756 ConcurrentMarkSweepThread::desynchronize(true);
5757 stopTimer();
5758 if (PrintCMSStatistics != 0) {
5759 incrementYields();
5760 }
5761
5762 // See the comment in coordinator_yield()
5763 for (unsigned i = 0; i < CMSYieldSleepCount &&
5764 ConcurrentMarkSweepThread::should_yield() &&
5765 !CMSCollector::foregroundGCIsActive(); ++i) {
5766 os::sleep(Thread::current(), 1, false);
5767 }
5768
5769 ConcurrentMarkSweepThread::synchronize(true);
5770 bitMapLock()->lock_without_safepoint_check();
5771 startTimer();
5772 }
5773 curAddr = chunk.end();
5774 }
5775 // A successful mostly concurrent collection has been done.
5776 // Because only the full (i.e., concurrent mode failure) collections
5777 // are being measured for gc overhead limits, clean the "near" flag
5778 // and count.
5779 size_policy()->reset_gc_overhead_limit_count();
5780 _collectorState = Idling;
5781 } else {
5782 // already have the lock
5783 assert(_collectorState == Resetting, "just checking");
5784 assert_lock_strong(bitMapLock());
5785 _markBitMap.clear_all();
5786 _collectorState = Idling;
5787 }
5788
5789 register_gc_end();
5790 }
5791
5792 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5793 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5794 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id());
5795 TraceCollectorStats tcs(counters());
5796
5797 switch (op) {
5798 case CMS_op_checkpointRootsInitial: {
5799 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5800 checkpointRootsInitial();
5801 if (PrintGC) {
5802 _cmsGen->printOccupancy("initial-mark");
5803 }
5804 break;
5805 }
5806 case CMS_op_checkpointRootsFinal: {
5807 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5808 checkpointRootsFinal();
5809 if (PrintGC) {
5810 _cmsGen->printOccupancy("remark");
5811 }
5812 break;
5813 }
5814 default:
5815 fatal("No such CMS_op");
5816 }
5817 }
5818
5819 #ifndef PRODUCT
5820 size_t const CMSCollector::skip_header_HeapWords() {
5821 return FreeChunk::header_size();
5822 }
5823
5824 // Try and collect here conditions that should hold when
5825 // CMS thread is exiting. The idea is that the foreground GC
5826 // thread should not be blocked if it wants to terminate
5827 // the CMS thread and yet continue to run the VM for a while
5828 // after that.
5829 void CMSCollector::verify_ok_to_terminate() const {
5830 assert(Thread::current()->is_ConcurrentGC_thread(),
5831 "should be called by CMS thread");
5832 assert(!_foregroundGCShouldWait, "should be false");
5833 // We could check here that all the various low-level locks
5834 // are not held by the CMS thread, but that is overkill; see
5835 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5836 // is checked.
5837 }
5838 #endif
5839
5840 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5841 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5842 "missing Printezis mark?");
5843 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5844 size_t size = pointer_delta(nextOneAddr + 1, addr);
5845 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5846 "alignment problem");
5847 assert(size >= 3, "Necessary for Printezis marks to work");
5848 return size;
5849 }
5850
5851 // A variant of the above (block_size_using_printezis_bits()) except
5852 // that we return 0 if the P-bits are not yet set.
5853 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5854 if (_markBitMap.isMarked(addr + 1)) {
5855 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5856 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5857 size_t size = pointer_delta(nextOneAddr + 1, addr);
5858 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5859 "alignment problem");
5860 assert(size >= 3, "Necessary for Printezis marks to work");
5861 return size;
5862 }
5863 return 0;
5864 }
5865
5866 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5867 size_t sz = 0;
5868 oop p = (oop)addr;
5869 if (p->klass_or_null() != NULL) {
5870 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5871 } else {
5872 sz = block_size_using_printezis_bits(addr);
5873 }
5874 assert(sz > 0, "size must be nonzero");
5875 HeapWord* next_block = addr + sz;
5876 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
5877 CardTableModRefBS::card_size);
5878 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
5879 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
5880 "must be different cards");
5881 return next_card;
5882 }
5883
5884
5885 // CMS Bit Map Wrapper /////////////////////////////////////////
5886
5887 // Construct a CMS bit map infrastructure, but don't create the
5888 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5889 // further below.
5890 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5891 _bm(),
5892 _shifter(shifter),
5893 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5894 Monitor::_safepoint_check_sometimes) : NULL)
5895 {
5896 _bmStartWord = 0;
5897 _bmWordSize = 0;
5898 }
5899
5900 bool CMSBitMap::allocate(MemRegion mr) {
5901 _bmStartWord = mr.start();
5902 _bmWordSize = mr.word_size();
5903 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5904 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5905 if (!brs.is_reserved()) {
5906 warning("CMS bit map allocation failure");
5907 return false;
5908 }
5909 // For now we'll just commit all of the bit map up front.
5910 // Later on we'll try to be more parsimonious with swap.
5911 if (!_virtual_space.initialize(brs, brs.size())) {
5912 warning("CMS bit map backing store failure");
5913 return false;
5914 }
5915 assert(_virtual_space.committed_size() == brs.size(),
5916 "didn't reserve backing store for all of CMS bit map?");
5917 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
5918 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5919 _bmWordSize, "inconsistency in bit map sizing");
5920 _bm.set_size(_bmWordSize >> _shifter);
5921
5922 // bm.clear(); // can we rely on getting zero'd memory? verify below
5923 assert(isAllClear(),
5924 "Expected zero'd memory from ReservedSpace constructor");
5925 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5926 "consistency check");
5927 return true;
5928 }
5929
5930 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5931 HeapWord *next_addr, *end_addr, *last_addr;
5932 assert_locked();
5933 assert(covers(mr), "out-of-range error");
5934 // XXX assert that start and end are appropriately aligned
5935 for (next_addr = mr.start(), end_addr = mr.end();
5936 next_addr < end_addr; next_addr = last_addr) {
5937 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5938 last_addr = dirty_region.end();
5939 if (!dirty_region.is_empty()) {
5940 cl->do_MemRegion(dirty_region);
5941 } else {
5942 assert(last_addr == end_addr, "program logic");
5943 return;
5944 }
5945 }
5946 }
5947
5948 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5949 _bm.print_on_error(st, prefix);
5950 }
5951
5952 #ifndef PRODUCT
5953 void CMSBitMap::assert_locked() const {
5954 CMSLockVerifier::assert_locked(lock());
5955 }
5956
5957 bool CMSBitMap::covers(MemRegion mr) const {
5958 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5959 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5960 "size inconsistency");
5961 return (mr.start() >= _bmStartWord) &&
5962 (mr.end() <= endWord());
5963 }
5964
5965 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5966 return (start >= _bmStartWord && (start + size) <= endWord());
5967 }
5968
5969 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5970 // verify that there are no 1 bits in the interval [left, right)
5971 FalseBitMapClosure falseBitMapClosure;
5972 iterate(&falseBitMapClosure, left, right);
5973 }
5974
5975 void CMSBitMap::region_invariant(MemRegion mr)
5976 {
5977 assert_locked();
5978 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5979 assert(!mr.is_empty(), "unexpected empty region");
5980 assert(covers(mr), "mr should be covered by bit map");
5981 // convert address range into offset range
5982 size_t start_ofs = heapWordToOffset(mr.start());
5983 // Make sure that end() is appropriately aligned
5984 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
5985 (1 << (_shifter+LogHeapWordSize))),
5986 "Misaligned mr.end()");
5987 size_t end_ofs = heapWordToOffset(mr.end());
5988 assert(end_ofs > start_ofs, "Should mark at least one bit");
5989 }
5990
5991 #endif
5992
5993 bool CMSMarkStack::allocate(size_t size) {
5994 // allocate a stack of the requisite depth
5995 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5996 size * sizeof(oop)));
5997 if (!rs.is_reserved()) {
5998 warning("CMSMarkStack allocation failure");
5999 return false;
6000 }
6001 if (!_virtual_space.initialize(rs, rs.size())) {
6002 warning("CMSMarkStack backing store failure");
6003 return false;
6004 }
6005 assert(_virtual_space.committed_size() == rs.size(),
6006 "didn't reserve backing store for all of CMS stack?");
6007 _base = (oop*)(_virtual_space.low());
6008 _index = 0;
6009 _capacity = size;
6010 NOT_PRODUCT(_max_depth = 0);
6011 return true;
6012 }
6013
6014 // XXX FIX ME !!! In the MT case we come in here holding a
6015 // leaf lock. For printing we need to take a further lock
6016 // which has lower rank. We need to recalibrate the two
6017 // lock-ranks involved in order to be able to print the
6018 // messages below. (Or defer the printing to the caller.
6019 // For now we take the expedient path of just disabling the
6020 // messages for the problematic case.)
6021 void CMSMarkStack::expand() {
6022 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6023 if (_capacity == MarkStackSizeMax) {
6024 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6025 // We print a warning message only once per CMS cycle.
6026 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6027 }
6028 return;
6029 }
6030 // Double capacity if possible
6031 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6032 // Do not give up existing stack until we have managed to
6033 // get the double capacity that we desired.
6034 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6035 new_capacity * sizeof(oop)));
6036 if (rs.is_reserved()) {
6037 // Release the backing store associated with old stack
6038 _virtual_space.release();
6039 // Reinitialize virtual space for new stack
6040 if (!_virtual_space.initialize(rs, rs.size())) {
6041 fatal("Not enough swap for expanded marking stack");
6042 }
6043 _base = (oop*)(_virtual_space.low());
6044 _index = 0;
6045 _capacity = new_capacity;
6046 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6047 // Failed to double capacity, continue;
6048 // we print a detail message only once per CMS cycle.
6049 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6050 SIZE_FORMAT"K",
6051 _capacity / K, new_capacity / K);
6052 }
6053 }
6054
6055
6056 // Closures
6057 // XXX: there seems to be a lot of code duplication here;
6058 // should refactor and consolidate common code.
6059
6060 // This closure is used to mark refs into the CMS generation in
6061 // the CMS bit map. Called at the first checkpoint. This closure
6062 // assumes that we do not need to re-mark dirty cards; if the CMS
6063 // generation on which this is used is not an oldest
6064 // generation then this will lose younger_gen cards!
6065
6066 MarkRefsIntoClosure::MarkRefsIntoClosure(
6067 MemRegion span, CMSBitMap* bitMap):
6068 _span(span),
6069 _bitMap(bitMap)
6070 {
6071 assert(_ref_processor == NULL, "deliberately left NULL");
6072 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6073 }
6074
6075 void MarkRefsIntoClosure::do_oop(oop obj) {
6076 // if p points into _span, then mark corresponding bit in _markBitMap
6077 assert(obj->is_oop(), "expected an oop");
6078 HeapWord* addr = (HeapWord*)obj;
6079 if (_span.contains(addr)) {
6080 // this should be made more efficient
6081 _bitMap->mark(addr);
6082 }
6083 }
6084
6085 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6086 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6087
6088 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
6089 MemRegion span, CMSBitMap* bitMap):
6090 _span(span),
6091 _bitMap(bitMap)
6092 {
6093 assert(_ref_processor == NULL, "deliberately left NULL");
6094 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6095 }
6096
6097 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
6098 // if p points into _span, then mark corresponding bit in _markBitMap
6099 assert(obj->is_oop(), "expected an oop");
6100 HeapWord* addr = (HeapWord*)obj;
6101 if (_span.contains(addr)) {
6102 // this should be made more efficient
6103 _bitMap->par_mark(addr);
6104 }
6105 }
6106
6107 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6108 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6109
6110 // A variant of the above, used for CMS marking verification.
6111 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6112 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6113 _span(span),
6114 _verification_bm(verification_bm),
6115 _cms_bm(cms_bm)
6116 {
6117 assert(_ref_processor == NULL, "deliberately left NULL");
6118 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6119 }
6120
6121 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6122 // if p points into _span, then mark corresponding bit in _markBitMap
6123 assert(obj->is_oop(), "expected an oop");
6124 HeapWord* addr = (HeapWord*)obj;
6125 if (_span.contains(addr)) {
6126 _verification_bm->mark(addr);
6127 if (!_cms_bm->isMarked(addr)) {
6128 oop(addr)->print();
6129 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6130 fatal("... aborting");
6131 }
6132 }
6133 }
6134
6135 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6136 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6137
6138 //////////////////////////////////////////////////
6139 // MarkRefsIntoAndScanClosure
6140 //////////////////////////////////////////////////
6141
6142 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6143 ReferenceProcessor* rp,
6144 CMSBitMap* bit_map,
6145 CMSBitMap* mod_union_table,
6146 CMSMarkStack* mark_stack,
6147 CMSCollector* collector,
6148 bool should_yield,
6149 bool concurrent_precleaning):
6150 _collector(collector),
6151 _span(span),
6152 _bit_map(bit_map),
6153 _mark_stack(mark_stack),
6154 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6155 mark_stack, concurrent_precleaning),
6156 _yield(should_yield),
6157 _concurrent_precleaning(concurrent_precleaning),
6158 _freelistLock(NULL)
6159 {
6160 _ref_processor = rp;
6161 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6162 }
6163
6164 // This closure is used to mark refs into the CMS generation at the
6165 // second (final) checkpoint, and to scan and transitively follow
6166 // the unmarked oops. It is also used during the concurrent precleaning
6167 // phase while scanning objects on dirty cards in the CMS generation.
6168 // The marks are made in the marking bit map and the marking stack is
6169 // used for keeping the (newly) grey objects during the scan.
6170 // The parallel version (Par_...) appears further below.
6171 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6172 if (obj != NULL) {
6173 assert(obj->is_oop(), "expected an oop");
6174 HeapWord* addr = (HeapWord*)obj;
6175 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6176 assert(_collector->overflow_list_is_empty(),
6177 "overflow list should be empty");
6178 if (_span.contains(addr) &&
6179 !_bit_map->isMarked(addr)) {
6180 // mark bit map (object is now grey)
6181 _bit_map->mark(addr);
6182 // push on marking stack (stack should be empty), and drain the
6183 // stack by applying this closure to the oops in the oops popped
6184 // from the stack (i.e. blacken the grey objects)
6185 bool res = _mark_stack->push(obj);
6186 assert(res, "Should have space to push on empty stack");
6187 do {
6188 oop new_oop = _mark_stack->pop();
6189 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6190 assert(_bit_map->isMarked((HeapWord*)new_oop),
6191 "only grey objects on this stack");
6192 // iterate over the oops in this oop, marking and pushing
6193 // the ones in CMS heap (i.e. in _span).
6194 new_oop->oop_iterate(&_pushAndMarkClosure);
6195 // check if it's time to yield
6196 do_yield_check();
6197 } while (!_mark_stack->isEmpty() ||
6198 (!_concurrent_precleaning && take_from_overflow_list()));
6199 // if marking stack is empty, and we are not doing this
6200 // during precleaning, then check the overflow list
6201 }
6202 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6203 assert(_collector->overflow_list_is_empty(),
6204 "overflow list was drained above");
6205 // We could restore evacuated mark words, if any, used for
6206 // overflow list links here because the overflow list is
6207 // provably empty here. That would reduce the maximum
6208 // size requirements for preserved_{oop,mark}_stack.
6209 // But we'll just postpone it until we are all done
6210 // so we can just stream through.
6211 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6212 _collector->restore_preserved_marks_if_any();
6213 assert(_collector->no_preserved_marks(), "No preserved marks");
6214 }
6215 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6216 "All preserved marks should have been restored above");
6217 }
6218 }
6219
6220 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6221 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6222
6223 void MarkRefsIntoAndScanClosure::do_yield_work() {
6224 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6225 "CMS thread should hold CMS token");
6226 assert_lock_strong(_freelistLock);
6227 assert_lock_strong(_bit_map->lock());
6228 // relinquish the free_list_lock and bitMaplock()
6229 _bit_map->lock()->unlock();
6230 _freelistLock->unlock();
6231 ConcurrentMarkSweepThread::desynchronize(true);
6232 _collector->stopTimer();
6233 if (PrintCMSStatistics != 0) {
6234 _collector->incrementYields();
6235 }
6236
6237 // See the comment in coordinator_yield()
6238 for (unsigned i = 0;
6239 i < CMSYieldSleepCount &&
6240 ConcurrentMarkSweepThread::should_yield() &&
6241 !CMSCollector::foregroundGCIsActive();
6242 ++i) {
6243 os::sleep(Thread::current(), 1, false);
6244 }
6245
6246 ConcurrentMarkSweepThread::synchronize(true);
6247 _freelistLock->lock_without_safepoint_check();
6248 _bit_map->lock()->lock_without_safepoint_check();
6249 _collector->startTimer();
6250 }
6251
6252 ///////////////////////////////////////////////////////////
6253 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6254 // MarkRefsIntoAndScanClosure
6255 ///////////////////////////////////////////////////////////
6256 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6257 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6258 CMSBitMap* bit_map, OopTaskQueue* work_queue):
6259 _span(span),
6260 _bit_map(bit_map),
6261 _work_queue(work_queue),
6262 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6263 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6264 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
6265 {
6266 _ref_processor = rp;
6267 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6268 }
6269
6270 // This closure is used to mark refs into the CMS generation at the
6271 // second (final) checkpoint, and to scan and transitively follow
6272 // the unmarked oops. The marks are made in the marking bit map and
6273 // the work_queue is used for keeping the (newly) grey objects during
6274 // the scan phase whence they are also available for stealing by parallel
6275 // threads. Since the marking bit map is shared, updates are
6276 // synchronized (via CAS).
6277 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6278 if (obj != NULL) {
6279 // Ignore mark word because this could be an already marked oop
6280 // that may be chained at the end of the overflow list.
6281 assert(obj->is_oop(true), "expected an oop");
6282 HeapWord* addr = (HeapWord*)obj;
6283 if (_span.contains(addr) &&
6284 !_bit_map->isMarked(addr)) {
6285 // mark bit map (object will become grey):
6286 // It is possible for several threads to be
6287 // trying to "claim" this object concurrently;
6288 // the unique thread that succeeds in marking the
6289 // object first will do the subsequent push on
6290 // to the work queue (or overflow list).
6291 if (_bit_map->par_mark(addr)) {
6292 // push on work_queue (which may not be empty), and trim the
6293 // queue to an appropriate length by applying this closure to
6294 // the oops in the oops popped from the stack (i.e. blacken the
6295 // grey objects)
6296 bool res = _work_queue->push(obj);
6297 assert(res, "Low water mark should be less than capacity?");
6298 trim_queue(_low_water_mark);
6299 } // Else, another thread claimed the object
6300 }
6301 }
6302 }
6303
6304 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6305 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6306
6307 // This closure is used to rescan the marked objects on the dirty cards
6308 // in the mod union table and the card table proper.
6309 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6310 oop p, MemRegion mr) {
6311
6312 size_t size = 0;
6313 HeapWord* addr = (HeapWord*)p;
6314 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6315 assert(_span.contains(addr), "we are scanning the CMS generation");
6316 // check if it's time to yield
6317 if (do_yield_check()) {
6318 // We yielded for some foreground stop-world work,
6319 // and we have been asked to abort this ongoing preclean cycle.
6320 return 0;
6321 }
6322 if (_bitMap->isMarked(addr)) {
6323 // it's marked; is it potentially uninitialized?
6324 if (p->klass_or_null() != NULL) {
6325 // an initialized object; ignore mark word in verification below
6326 // since we are running concurrent with mutators
6327 assert(p->is_oop(true), "should be an oop");
6328 if (p->is_objArray()) {
6329 // objArrays are precisely marked; restrict scanning
6330 // to dirty cards only.
6331 size = CompactibleFreeListSpace::adjustObjectSize(
6332 p->oop_iterate(_scanningClosure, mr));
6333 } else {
6334 // A non-array may have been imprecisely marked; we need
6335 // to scan object in its entirety.
6336 size = CompactibleFreeListSpace::adjustObjectSize(
6337 p->oop_iterate(_scanningClosure));
6338 }
6339 #ifdef ASSERT
6340 size_t direct_size =
6341 CompactibleFreeListSpace::adjustObjectSize(p->size());
6342 assert(size == direct_size, "Inconsistency in size");
6343 assert(size >= 3, "Necessary for Printezis marks to work");
6344 if (!_bitMap->isMarked(addr+1)) {
6345 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6346 } else {
6347 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6348 assert(_bitMap->isMarked(addr+size-1),
6349 "inconsistent Printezis mark");
6350 }
6351 #endif // ASSERT
6352 } else {
6353 // An uninitialized object.
6354 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6355 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6356 size = pointer_delta(nextOneAddr + 1, addr);
6357 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6358 "alignment problem");
6359 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6360 // will dirty the card when the klass pointer is installed in the
6361 // object (signaling the completion of initialization).
6362 }
6363 } else {
6364 // Either a not yet marked object or an uninitialized object
6365 if (p->klass_or_null() == NULL) {
6366 // An uninitialized object, skip to the next card, since
6367 // we may not be able to read its P-bits yet.
6368 assert(size == 0, "Initial value");
6369 } else {
6370 // An object not (yet) reached by marking: we merely need to
6371 // compute its size so as to go look at the next block.
6372 assert(p->is_oop(true), "should be an oop");
6373 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6374 }
6375 }
6376 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6377 return size;
6378 }
6379
6380 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6381 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6382 "CMS thread should hold CMS token");
6383 assert_lock_strong(_freelistLock);
6384 assert_lock_strong(_bitMap->lock());
6385 // relinquish the free_list_lock and bitMaplock()
6386 _bitMap->lock()->unlock();
6387 _freelistLock->unlock();
6388 ConcurrentMarkSweepThread::desynchronize(true);
6389 _collector->stopTimer();
6390 if (PrintCMSStatistics != 0) {
6391 _collector->incrementYields();
6392 }
6393
6394 // See the comment in coordinator_yield()
6395 for (unsigned i = 0; i < CMSYieldSleepCount &&
6396 ConcurrentMarkSweepThread::should_yield() &&
6397 !CMSCollector::foregroundGCIsActive(); ++i) {
6398 os::sleep(Thread::current(), 1, false);
6399 }
6400
6401 ConcurrentMarkSweepThread::synchronize(true);
6402 _freelistLock->lock_without_safepoint_check();
6403 _bitMap->lock()->lock_without_safepoint_check();
6404 _collector->startTimer();
6405 }
6406
6407
6408 //////////////////////////////////////////////////////////////////
6409 // SurvivorSpacePrecleanClosure
6410 //////////////////////////////////////////////////////////////////
6411 // This (single-threaded) closure is used to preclean the oops in
6412 // the survivor spaces.
6413 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6414
6415 HeapWord* addr = (HeapWord*)p;
6416 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6417 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6418 assert(p->klass_or_null() != NULL, "object should be initialized");
6419 // an initialized object; ignore mark word in verification below
6420 // since we are running concurrent with mutators
6421 assert(p->is_oop(true), "should be an oop");
6422 // Note that we do not yield while we iterate over
6423 // the interior oops of p, pushing the relevant ones
6424 // on our marking stack.
6425 size_t size = p->oop_iterate(_scanning_closure);
6426 do_yield_check();
6427 // Observe that below, we do not abandon the preclean
6428 // phase as soon as we should; rather we empty the
6429 // marking stack before returning. This is to satisfy
6430 // some existing assertions. In general, it may be a
6431 // good idea to abort immediately and complete the marking
6432 // from the grey objects at a later time.
6433 while (!_mark_stack->isEmpty()) {
6434 oop new_oop = _mark_stack->pop();
6435 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6436 assert(_bit_map->isMarked((HeapWord*)new_oop),
6437 "only grey objects on this stack");
6438 // iterate over the oops in this oop, marking and pushing
6439 // the ones in CMS heap (i.e. in _span).
6440 new_oop->oop_iterate(_scanning_closure);
6441 // check if it's time to yield
6442 do_yield_check();
6443 }
6444 unsigned int after_count =
6445 GenCollectedHeap::heap()->total_collections();
6446 bool abort = (_before_count != after_count) ||
6447 _collector->should_abort_preclean();
6448 return abort ? 0 : size;
6449 }
6450
6451 void SurvivorSpacePrecleanClosure::do_yield_work() {
6452 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6453 "CMS thread should hold CMS token");
6454 assert_lock_strong(_bit_map->lock());
6455 // Relinquish the bit map lock
6456 _bit_map->lock()->unlock();
6457 ConcurrentMarkSweepThread::desynchronize(true);
6458 _collector->stopTimer();
6459 if (PrintCMSStatistics != 0) {
6460 _collector->incrementYields();
6461 }
6462
6463 // See the comment in coordinator_yield()
6464 for (unsigned i = 0; i < CMSYieldSleepCount &&
6465 ConcurrentMarkSweepThread::should_yield() &&
6466 !CMSCollector::foregroundGCIsActive(); ++i) {
6467 os::sleep(Thread::current(), 1, false);
6468 }
6469
6470 ConcurrentMarkSweepThread::synchronize(true);
6471 _bit_map->lock()->lock_without_safepoint_check();
6472 _collector->startTimer();
6473 }
6474
6475 // This closure is used to rescan the marked objects on the dirty cards
6476 // in the mod union table and the card table proper. In the parallel
6477 // case, although the bitMap is shared, we do a single read so the
6478 // isMarked() query is "safe".
6479 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6480 // Ignore mark word because we are running concurrent with mutators
6481 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p)));
6482 HeapWord* addr = (HeapWord*)p;
6483 assert(_span.contains(addr), "we are scanning the CMS generation");
6484 bool is_obj_array = false;
6485 #ifdef ASSERT
6486 if (!_parallel) {
6487 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6488 assert(_collector->overflow_list_is_empty(),
6489 "overflow list should be empty");
6490
6491 }
6492 #endif // ASSERT
6493 if (_bit_map->isMarked(addr)) {
6494 // Obj arrays are precisely marked, non-arrays are not;
6495 // so we scan objArrays precisely and non-arrays in their
6496 // entirety.
6497 if (p->is_objArray()) {
6498 is_obj_array = true;
6499 if (_parallel) {
6500 p->oop_iterate(_par_scan_closure, mr);
6501 } else {
6502 p->oop_iterate(_scan_closure, mr);
6503 }
6504 } else {
6505 if (_parallel) {
6506 p->oop_iterate(_par_scan_closure);
6507 } else {
6508 p->oop_iterate(_scan_closure);
6509 }
6510 }
6511 }
6512 #ifdef ASSERT
6513 if (!_parallel) {
6514 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6515 assert(_collector->overflow_list_is_empty(),
6516 "overflow list should be empty");
6517
6518 }
6519 #endif // ASSERT
6520 return is_obj_array;
6521 }
6522
6523 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6524 MemRegion span,
6525 CMSBitMap* bitMap, CMSMarkStack* markStack,
6526 bool should_yield, bool verifying):
6527 _collector(collector),
6528 _span(span),
6529 _bitMap(bitMap),
6530 _mut(&collector->_modUnionTable),
6531 _markStack(markStack),
6532 _yield(should_yield),
6533 _skipBits(0)
6534 {
6535 assert(_markStack->isEmpty(), "stack should be empty");
6536 _finger = _bitMap->startWord();
6537 _threshold = _finger;
6538 assert(_collector->_restart_addr == NULL, "Sanity check");
6539 assert(_span.contains(_finger), "Out of bounds _finger?");
6540 DEBUG_ONLY(_verifying = verifying;)
6541 }
6542
6543 void MarkFromRootsClosure::reset(HeapWord* addr) {
6544 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6545 assert(_span.contains(addr), "Out of bounds _finger?");
6546 _finger = addr;
6547 _threshold = (HeapWord*)round_to(
6548 (intptr_t)_finger, CardTableModRefBS::card_size);
6549 }
6550
6551 // Should revisit to see if this should be restructured for
6552 // greater efficiency.
6553 bool MarkFromRootsClosure::do_bit(size_t offset) {
6554 if (_skipBits > 0) {
6555 _skipBits--;
6556 return true;
6557 }
6558 // convert offset into a HeapWord*
6559 HeapWord* addr = _bitMap->startWord() + offset;
6560 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6561 "address out of range");
6562 assert(_bitMap->isMarked(addr), "tautology");
6563 if (_bitMap->isMarked(addr+1)) {
6564 // this is an allocated but not yet initialized object
6565 assert(_skipBits == 0, "tautology");
6566 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6567 oop p = oop(addr);
6568 if (p->klass_or_null() == NULL) {
6569 DEBUG_ONLY(if (!_verifying) {)
6570 // We re-dirty the cards on which this object lies and increase
6571 // the _threshold so that we'll come back to scan this object
6572 // during the preclean or remark phase. (CMSCleanOnEnter)
6573 if (CMSCleanOnEnter) {
6574 size_t sz = _collector->block_size_using_printezis_bits(addr);
6575 HeapWord* end_card_addr = (HeapWord*)round_to(
6576 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6577 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6578 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6579 // Bump _threshold to end_card_addr; note that
6580 // _threshold cannot possibly exceed end_card_addr, anyhow.
6581 // This prevents future clearing of the card as the scan proceeds
6582 // to the right.
6583 assert(_threshold <= end_card_addr,
6584 "Because we are just scanning into this object");
6585 if (_threshold < end_card_addr) {
6586 _threshold = end_card_addr;
6587 }
6588 if (p->klass_or_null() != NULL) {
6589 // Redirty the range of cards...
6590 _mut->mark_range(redirty_range);
6591 } // ...else the setting of klass will dirty the card anyway.
6592 }
6593 DEBUG_ONLY(})
6594 return true;
6595 }
6596 }
6597 scanOopsInOop(addr);
6598 return true;
6599 }
6600
6601 // We take a break if we've been at this for a while,
6602 // so as to avoid monopolizing the locks involved.
6603 void MarkFromRootsClosure::do_yield_work() {
6604 // First give up the locks, then yield, then re-lock
6605 // We should probably use a constructor/destructor idiom to
6606 // do this unlock/lock or modify the MutexUnlocker class to
6607 // serve our purpose. XXX
6608 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6609 "CMS thread should hold CMS token");
6610 assert_lock_strong(_bitMap->lock());
6611 _bitMap->lock()->unlock();
6612 ConcurrentMarkSweepThread::desynchronize(true);
6613 _collector->stopTimer();
6614 if (PrintCMSStatistics != 0) {
6615 _collector->incrementYields();
6616 }
6617
6618 // See the comment in coordinator_yield()
6619 for (unsigned i = 0; i < CMSYieldSleepCount &&
6620 ConcurrentMarkSweepThread::should_yield() &&
6621 !CMSCollector::foregroundGCIsActive(); ++i) {
6622 os::sleep(Thread::current(), 1, false);
6623 }
6624
6625 ConcurrentMarkSweepThread::synchronize(true);
6626 _bitMap->lock()->lock_without_safepoint_check();
6627 _collector->startTimer();
6628 }
6629
6630 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6631 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6632 assert(_markStack->isEmpty(),
6633 "should drain stack to limit stack usage");
6634 // convert ptr to an oop preparatory to scanning
6635 oop obj = oop(ptr);
6636 // Ignore mark word in verification below, since we
6637 // may be running concurrent with mutators.
6638 assert(obj->is_oop(true), "should be an oop");
6639 assert(_finger <= ptr, "_finger runneth ahead");
6640 // advance the finger to right end of this object
6641 _finger = ptr + obj->size();
6642 assert(_finger > ptr, "we just incremented it above");
6643 // On large heaps, it may take us some time to get through
6644 // the marking phase. During
6645 // this time it's possible that a lot of mutations have
6646 // accumulated in the card table and the mod union table --
6647 // these mutation records are redundant until we have
6648 // actually traced into the corresponding card.
6649 // Here, we check whether advancing the finger would make
6650 // us cross into a new card, and if so clear corresponding
6651 // cards in the MUT (preclean them in the card-table in the
6652 // future).
6653
6654 DEBUG_ONLY(if (!_verifying) {)
6655 // The clean-on-enter optimization is disabled by default,
6656 // until we fix 6178663.
6657 if (CMSCleanOnEnter && (_finger > _threshold)) {
6658 // [_threshold, _finger) represents the interval
6659 // of cards to be cleared in MUT (or precleaned in card table).
6660 // The set of cards to be cleared is all those that overlap
6661 // with the interval [_threshold, _finger); note that
6662 // _threshold is always kept card-aligned but _finger isn't
6663 // always card-aligned.
6664 HeapWord* old_threshold = _threshold;
6665 assert(old_threshold == (HeapWord*)round_to(
6666 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6667 "_threshold should always be card-aligned");
6668 _threshold = (HeapWord*)round_to(
6669 (intptr_t)_finger, CardTableModRefBS::card_size);
6670 MemRegion mr(old_threshold, _threshold);
6671 assert(!mr.is_empty(), "Control point invariant");
6672 assert(_span.contains(mr), "Should clear within span");
6673 _mut->clear_range(mr);
6674 }
6675 DEBUG_ONLY(})
6676 // Note: the finger doesn't advance while we drain
6677 // the stack below.
6678 PushOrMarkClosure pushOrMarkClosure(_collector,
6679 _span, _bitMap, _markStack,
6680 _finger, this);
6681 bool res = _markStack->push(obj);
6682 assert(res, "Empty non-zero size stack should have space for single push");
6683 while (!_markStack->isEmpty()) {
6684 oop new_oop = _markStack->pop();
6685 // Skip verifying header mark word below because we are
6686 // running concurrent with mutators.
6687 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6688 // now scan this oop's oops
6689 new_oop->oop_iterate(&pushOrMarkClosure);
6690 do_yield_check();
6691 }
6692 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6693 }
6694
6695 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
6696 CMSCollector* collector, MemRegion span,
6697 CMSBitMap* bit_map,
6698 OopTaskQueue* work_queue,
6699 CMSMarkStack* overflow_stack):
6700 _collector(collector),
6701 _whole_span(collector->_span),
6702 _span(span),
6703 _bit_map(bit_map),
6704 _mut(&collector->_modUnionTable),
6705 _work_queue(work_queue),
6706 _overflow_stack(overflow_stack),
6707 _skip_bits(0),
6708 _task(task)
6709 {
6710 assert(_work_queue->size() == 0, "work_queue should be empty");
6711 _finger = span.start();
6712 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
6713 assert(_span.contains(_finger), "Out of bounds _finger?");
6714 }
6715
6716 // Should revisit to see if this should be restructured for
6717 // greater efficiency.
6718 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
6719 if (_skip_bits > 0) {
6720 _skip_bits--;
6721 return true;
6722 }
6723 // convert offset into a HeapWord*
6724 HeapWord* addr = _bit_map->startWord() + offset;
6725 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6726 "address out of range");
6727 assert(_bit_map->isMarked(addr), "tautology");
6728 if (_bit_map->isMarked(addr+1)) {
6729 // this is an allocated object that might not yet be initialized
6730 assert(_skip_bits == 0, "tautology");
6731 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
6732 oop p = oop(addr);
6733 if (p->klass_or_null() == NULL) {
6734 // in the case of Clean-on-Enter optimization, redirty card
6735 // and avoid clearing card by increasing the threshold.
6736 return true;
6737 }
6738 }
6739 scan_oops_in_oop(addr);
6740 return true;
6741 }
6742
6743 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6744 assert(_bit_map->isMarked(ptr), "expected bit to be set");
6745 // Should we assert that our work queue is empty or
6746 // below some drain limit?
6747 assert(_work_queue->size() == 0,
6748 "should drain stack to limit stack usage");
6749 // convert ptr to an oop preparatory to scanning
6750 oop obj = oop(ptr);
6751 // Ignore mark word in verification below, since we
6752 // may be running concurrent with mutators.
6753 assert(obj->is_oop(true), "should be an oop");
6754 assert(_finger <= ptr, "_finger runneth ahead");
6755 // advance the finger to right end of this object
6756 _finger = ptr + obj->size();
6757 assert(_finger > ptr, "we just incremented it above");
6758 // On large heaps, it may take us some time to get through
6759 // the marking phase. During
6760 // this time it's possible that a lot of mutations have
6761 // accumulated in the card table and the mod union table --
6762 // these mutation records are redundant until we have
6763 // actually traced into the corresponding card.
6764 // Here, we check whether advancing the finger would make
6765 // us cross into a new card, and if so clear corresponding
6766 // cards in the MUT (preclean them in the card-table in the
6767 // future).
6768
6769 // The clean-on-enter optimization is disabled by default,
6770 // until we fix 6178663.
6771 if (CMSCleanOnEnter && (_finger > _threshold)) {
6772 // [_threshold, _finger) represents the interval
6773 // of cards to be cleared in MUT (or precleaned in card table).
6774 // The set of cards to be cleared is all those that overlap
6775 // with the interval [_threshold, _finger); note that
6776 // _threshold is always kept card-aligned but _finger isn't
6777 // always card-aligned.
6778 HeapWord* old_threshold = _threshold;
6779 assert(old_threshold == (HeapWord*)round_to(
6780 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6781 "_threshold should always be card-aligned");
6782 _threshold = (HeapWord*)round_to(
6783 (intptr_t)_finger, CardTableModRefBS::card_size);
6784 MemRegion mr(old_threshold, _threshold);
6785 assert(!mr.is_empty(), "Control point invariant");
6786 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6787 _mut->clear_range(mr);
6788 }
6789
6790 // Note: the local finger doesn't advance while we drain
6791 // the stack below, but the global finger sure can and will.
6792 HeapWord** gfa = _task->global_finger_addr();
6793 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
6794 _span, _bit_map,
6795 _work_queue,
6796 _overflow_stack,
6797 _finger,
6798 gfa, this);
6799 bool res = _work_queue->push(obj); // overflow could occur here
6800 assert(res, "Will hold once we use workqueues");
6801 while (true) {
6802 oop new_oop;
6803 if (!_work_queue->pop_local(new_oop)) {
6804 // We emptied our work_queue; check if there's stuff that can
6805 // be gotten from the overflow stack.
6806 if (CMSConcMarkingTask::get_work_from_overflow_stack(
6807 _overflow_stack, _work_queue)) {
6808 do_yield_check();
6809 continue;
6810 } else { // done
6811 break;
6812 }
6813 }
6814 // Skip verifying header mark word below because we are
6815 // running concurrent with mutators.
6816 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6817 // now scan this oop's oops
6818 new_oop->oop_iterate(&pushOrMarkClosure);
6819 do_yield_check();
6820 }
6821 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6822 }
6823
6824 // Yield in response to a request from VM Thread or
6825 // from mutators.
6826 void Par_MarkFromRootsClosure::do_yield_work() {
6827 assert(_task != NULL, "sanity");
6828 _task->yield();
6829 }
6830
6831 // A variant of the above used for verifying CMS marking work.
6832 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6833 MemRegion span,
6834 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6835 CMSMarkStack* mark_stack):
6836 _collector(collector),
6837 _span(span),
6838 _verification_bm(verification_bm),
6839 _cms_bm(cms_bm),
6840 _mark_stack(mark_stack),
6841 _pam_verify_closure(collector, span, verification_bm, cms_bm,
6842 mark_stack)
6843 {
6844 assert(_mark_stack->isEmpty(), "stack should be empty");
6845 _finger = _verification_bm->startWord();
6846 assert(_collector->_restart_addr == NULL, "Sanity check");
6847 assert(_span.contains(_finger), "Out of bounds _finger?");
6848 }
6849
6850 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6851 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6852 assert(_span.contains(addr), "Out of bounds _finger?");
6853 _finger = addr;
6854 }
6855
6856 // Should revisit to see if this should be restructured for
6857 // greater efficiency.
6858 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6859 // convert offset into a HeapWord*
6860 HeapWord* addr = _verification_bm->startWord() + offset;
6861 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6862 "address out of range");
6863 assert(_verification_bm->isMarked(addr), "tautology");
6864 assert(_cms_bm->isMarked(addr), "tautology");
6865
6866 assert(_mark_stack->isEmpty(),
6867 "should drain stack to limit stack usage");
6868 // convert addr to an oop preparatory to scanning
6869 oop obj = oop(addr);
6870 assert(obj->is_oop(), "should be an oop");
6871 assert(_finger <= addr, "_finger runneth ahead");
6872 // advance the finger to right end of this object
6873 _finger = addr + obj->size();
6874 assert(_finger > addr, "we just incremented it above");
6875 // Note: the finger doesn't advance while we drain
6876 // the stack below.
6877 bool res = _mark_stack->push(obj);
6878 assert(res, "Empty non-zero size stack should have space for single push");
6879 while (!_mark_stack->isEmpty()) {
6880 oop new_oop = _mark_stack->pop();
6881 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
6882 // now scan this oop's oops
6883 new_oop->oop_iterate(&_pam_verify_closure);
6884 }
6885 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6886 return true;
6887 }
6888
6889 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6890 CMSCollector* collector, MemRegion span,
6891 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6892 CMSMarkStack* mark_stack):
6893 MetadataAwareOopClosure(collector->ref_processor()),
6894 _collector(collector),
6895 _span(span),
6896 _verification_bm(verification_bm),
6897 _cms_bm(cms_bm),
6898 _mark_stack(mark_stack)
6899 { }
6900
6901 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6902 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6903
6904 // Upon stack overflow, we discard (part of) the stack,
6905 // remembering the least address amongst those discarded
6906 // in CMSCollector's _restart_address.
6907 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6908 // Remember the least grey address discarded
6909 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6910 _collector->lower_restart_addr(ra);
6911 _mark_stack->reset(); // discard stack contents
6912 _mark_stack->expand(); // expand the stack if possible
6913 }
6914
6915 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6916 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
6917 HeapWord* addr = (HeapWord*)obj;
6918 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6919 // Oop lies in _span and isn't yet grey or black
6920 _verification_bm->mark(addr); // now grey
6921 if (!_cms_bm->isMarked(addr)) {
6922 oop(addr)->print();
6923 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
6924 p2i(addr));
6925 fatal("... aborting");
6926 }
6927
6928 if (!_mark_stack->push(obj)) { // stack overflow
6929 if (PrintCMSStatistics != 0) {
6930 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
6931 SIZE_FORMAT, _mark_stack->capacity());
6932 }
6933 assert(_mark_stack->isFull(), "Else push should have succeeded");
6934 handle_stack_overflow(addr);
6935 }
6936 // anything including and to the right of _finger
6937 // will be scanned as we iterate over the remainder of the
6938 // bit map
6939 }
6940 }
6941
6942 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6943 MemRegion span,
6944 CMSBitMap* bitMap, CMSMarkStack* markStack,
6945 HeapWord* finger, MarkFromRootsClosure* parent) :
6946 MetadataAwareOopClosure(collector->ref_processor()),
6947 _collector(collector),
6948 _span(span),
6949 _bitMap(bitMap),
6950 _markStack(markStack),
6951 _finger(finger),
6952 _parent(parent)
6953 { }
6954
6955 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
6956 MemRegion span,
6957 CMSBitMap* bit_map,
6958 OopTaskQueue* work_queue,
6959 CMSMarkStack* overflow_stack,
6960 HeapWord* finger,
6961 HeapWord** global_finger_addr,
6962 Par_MarkFromRootsClosure* parent) :
6963 MetadataAwareOopClosure(collector->ref_processor()),
6964 _collector(collector),
6965 _whole_span(collector->_span),
6966 _span(span),
6967 _bit_map(bit_map),
6968 _work_queue(work_queue),
6969 _overflow_stack(overflow_stack),
6970 _finger(finger),
6971 _global_finger_addr(global_finger_addr),
6972 _parent(parent)
6973 { }
6974
6975 // Assumes thread-safe access by callers, who are
6976 // responsible for mutual exclusion.
6977 void CMSCollector::lower_restart_addr(HeapWord* low) {
6978 assert(_span.contains(low), "Out of bounds addr");
6979 if (_restart_addr == NULL) {
6980 _restart_addr = low;
6981 } else {
6982 _restart_addr = MIN2(_restart_addr, low);
6983 }
6984 }
6985
6986 // Upon stack overflow, we discard (part of) the stack,
6987 // remembering the least address amongst those discarded
6988 // in CMSCollector's _restart_address.
6989 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6990 // Remember the least grey address discarded
6991 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6992 _collector->lower_restart_addr(ra);
6993 _markStack->reset(); // discard stack contents
6994 _markStack->expand(); // expand the stack if possible
6995 }
6996
6997 // Upon stack overflow, we discard (part of) the stack,
6998 // remembering the least address amongst those discarded
6999 // in CMSCollector's _restart_address.
7000 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7001 // We need to do this under a mutex to prevent other
7002 // workers from interfering with the work done below.
7003 MutexLockerEx ml(_overflow_stack->par_lock(),
7004 Mutex::_no_safepoint_check_flag);
7005 // Remember the least grey address discarded
7006 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7007 _collector->lower_restart_addr(ra);
7008 _overflow_stack->reset(); // discard stack contents
7009 _overflow_stack->expand(); // expand the stack if possible
7010 }
7011
7012 void PushOrMarkClosure::do_oop(oop obj) {
7013 // Ignore mark word because we are running concurrent with mutators.
7014 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7015 HeapWord* addr = (HeapWord*)obj;
7016 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7017 // Oop lies in _span and isn't yet grey or black
7018 _bitMap->mark(addr); // now grey
7019 if (addr < _finger) {
7020 // the bit map iteration has already either passed, or
7021 // sampled, this bit in the bit map; we'll need to
7022 // use the marking stack to scan this oop's oops.
7023 bool simulate_overflow = false;
7024 NOT_PRODUCT(
7025 if (CMSMarkStackOverflowALot &&
7026 _collector->simulate_overflow()) {
7027 // simulate a stack overflow
7028 simulate_overflow = true;
7029 }
7030 )
7031 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7032 if (PrintCMSStatistics != 0) {
7033 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7034 SIZE_FORMAT, _markStack->capacity());
7035 }
7036 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7037 handle_stack_overflow(addr);
7038 }
7039 }
7040 // anything including and to the right of _finger
7041 // will be scanned as we iterate over the remainder of the
7042 // bit map
7043 do_yield_check();
7044 }
7045 }
7046
7047 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7048 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7049
7050 void Par_PushOrMarkClosure::do_oop(oop obj) {
7051 // Ignore mark word because we are running concurrent with mutators.
7052 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7053 HeapWord* addr = (HeapWord*)obj;
7054 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7055 // Oop lies in _span and isn't yet grey or black
7056 // We read the global_finger (volatile read) strictly after marking oop
7057 bool res = _bit_map->par_mark(addr); // now grey
7058 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7059 // Should we push this marked oop on our stack?
7060 // -- if someone else marked it, nothing to do
7061 // -- if target oop is above global finger nothing to do
7062 // -- if target oop is in chunk and above local finger
7063 // then nothing to do
7064 // -- else push on work queue
7065 if ( !res // someone else marked it, they will deal with it
7066 || (addr >= *gfa) // will be scanned in a later task
7067 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7068 return;
7069 }
7070 // the bit map iteration has already either passed, or
7071 // sampled, this bit in the bit map; we'll need to
7072 // use the marking stack to scan this oop's oops.
7073 bool simulate_overflow = false;
7074 NOT_PRODUCT(
7075 if (CMSMarkStackOverflowALot &&
7076 _collector->simulate_overflow()) {
7077 // simulate a stack overflow
7078 simulate_overflow = true;
7079 }
7080 )
7081 if (simulate_overflow ||
7082 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7083 // stack overflow
7084 if (PrintCMSStatistics != 0) {
7085 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7086 SIZE_FORMAT, _overflow_stack->capacity());
7087 }
7088 // We cannot assert that the overflow stack is full because
7089 // it may have been emptied since.
7090 assert(simulate_overflow ||
7091 _work_queue->size() == _work_queue->max_elems(),
7092 "Else push should have succeeded");
7093 handle_stack_overflow(addr);
7094 }
7095 do_yield_check();
7096 }
7097 }
7098
7099 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7100 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7101
7102 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7103 MemRegion span,
7104 ReferenceProcessor* rp,
7105 CMSBitMap* bit_map,
7106 CMSBitMap* mod_union_table,
7107 CMSMarkStack* mark_stack,
7108 bool concurrent_precleaning):
7109 MetadataAwareOopClosure(rp),
7110 _collector(collector),
7111 _span(span),
7112 _bit_map(bit_map),
7113 _mod_union_table(mod_union_table),
7114 _mark_stack(mark_stack),
7115 _concurrent_precleaning(concurrent_precleaning)
7116 {
7117 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7118 }
7119
7120 // Grey object rescan during pre-cleaning and second checkpoint phases --
7121 // the non-parallel version (the parallel version appears further below.)
7122 void PushAndMarkClosure::do_oop(oop obj) {
7123 // Ignore mark word verification. If during concurrent precleaning,
7124 // the object monitor may be locked. If during the checkpoint
7125 // phases, the object may already have been reached by a different
7126 // path and may be at the end of the global overflow list (so
7127 // the mark word may be NULL).
7128 assert(obj->is_oop_or_null(true /* ignore mark word */),
7129 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7130 HeapWord* addr = (HeapWord*)obj;
7131 // Check if oop points into the CMS generation
7132 // and is not marked
7133 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7134 // a white object ...
7135 _bit_map->mark(addr); // ... now grey
7136 // push on the marking stack (grey set)
7137 bool simulate_overflow = false;
7138 NOT_PRODUCT(
7139 if (CMSMarkStackOverflowALot &&
7140 _collector->simulate_overflow()) {
7141 // simulate a stack overflow
7142 simulate_overflow = true;
7143 }
7144 )
7145 if (simulate_overflow || !_mark_stack->push(obj)) {
7146 if (_concurrent_precleaning) {
7147 // During precleaning we can just dirty the appropriate card(s)
7148 // in the mod union table, thus ensuring that the object remains
7149 // in the grey set and continue. In the case of object arrays
7150 // we need to dirty all of the cards that the object spans,
7151 // since the rescan of object arrays will be limited to the
7152 // dirty cards.
7153 // Note that no one can be interfering with us in this action
7154 // of dirtying the mod union table, so no locking or atomics
7155 // are required.
7156 if (obj->is_objArray()) {
7157 size_t sz = obj->size();
7158 HeapWord* end_card_addr = (HeapWord*)round_to(
7159 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7160 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7161 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7162 _mod_union_table->mark_range(redirty_range);
7163 } else {
7164 _mod_union_table->mark(addr);
7165 }
7166 _collector->_ser_pmc_preclean_ovflw++;
7167 } else {
7168 // During the remark phase, we need to remember this oop
7169 // in the overflow list.
7170 _collector->push_on_overflow_list(obj);
7171 _collector->_ser_pmc_remark_ovflw++;
7172 }
7173 }
7174 }
7175 }
7176
7177 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7178 MemRegion span,
7179 ReferenceProcessor* rp,
7180 CMSBitMap* bit_map,
7181 OopTaskQueue* work_queue):
7182 MetadataAwareOopClosure(rp),
7183 _collector(collector),
7184 _span(span),
7185 _bit_map(bit_map),
7186 _work_queue(work_queue)
7187 {
7188 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7189 }
7190
7191 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7192 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7193
7194 // Grey object rescan during second checkpoint phase --
7195 // the parallel version.
7196 void Par_PushAndMarkClosure::do_oop(oop obj) {
7197 // In the assert below, we ignore the mark word because
7198 // this oop may point to an already visited object that is
7199 // on the overflow stack (in which case the mark word has
7200 // been hijacked for chaining into the overflow stack --
7201 // if this is the last object in the overflow stack then
7202 // its mark word will be NULL). Because this object may
7203 // have been subsequently popped off the global overflow
7204 // stack, and the mark word possibly restored to the prototypical
7205 // value, by the time we get to examined this failing assert in
7206 // the debugger, is_oop_or_null(false) may subsequently start
7207 // to hold.
7208 assert(obj->is_oop_or_null(true),
7209 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7210 HeapWord* addr = (HeapWord*)obj;
7211 // Check if oop points into the CMS generation
7212 // and is not marked
7213 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7214 // a white object ...
7215 // If we manage to "claim" the object, by being the
7216 // first thread to mark it, then we push it on our
7217 // marking stack
7218 if (_bit_map->par_mark(addr)) { // ... now grey
7219 // push on work queue (grey set)
7220 bool simulate_overflow = false;
7221 NOT_PRODUCT(
7222 if (CMSMarkStackOverflowALot &&
7223 _collector->par_simulate_overflow()) {
7224 // simulate a stack overflow
7225 simulate_overflow = true;
7226 }
7227 )
7228 if (simulate_overflow || !_work_queue->push(obj)) {
7229 _collector->par_push_on_overflow_list(obj);
7230 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7231 }
7232 } // Else, some other thread got there first
7233 }
7234 }
7235
7236 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7237 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7238
7239 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7240 Mutex* bml = _collector->bitMapLock();
7241 assert_lock_strong(bml);
7242 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7243 "CMS thread should hold CMS token");
7244
7245 bml->unlock();
7246 ConcurrentMarkSweepThread::desynchronize(true);
7247
7248 _collector->stopTimer();
7249 if (PrintCMSStatistics != 0) {
7250 _collector->incrementYields();
7251 }
7252
7253 // See the comment in coordinator_yield()
7254 for (unsigned i = 0; i < CMSYieldSleepCount &&
7255 ConcurrentMarkSweepThread::should_yield() &&
7256 !CMSCollector::foregroundGCIsActive(); ++i) {
7257 os::sleep(Thread::current(), 1, false);
7258 }
7259
7260 ConcurrentMarkSweepThread::synchronize(true);
7261 bml->lock();
7262
7263 _collector->startTimer();
7264 }
7265
7266 bool CMSPrecleanRefsYieldClosure::should_return() {
7267 if (ConcurrentMarkSweepThread::should_yield()) {
7268 do_yield_work();
7269 }
7270 return _collector->foregroundGCIsActive();
7271 }
7272
7273 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7274 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7275 "mr should be aligned to start at a card boundary");
7276 // We'd like to assert:
7277 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7278 // "mr should be a range of cards");
7279 // However, that would be too strong in one case -- the last
7280 // partition ends at _unallocated_block which, in general, can be
7281 // an arbitrary boundary, not necessarily card aligned.
7282 if (PrintCMSStatistics != 0) {
7283 _num_dirty_cards +=
7284 mr.word_size()/CardTableModRefBS::card_size_in_words;
7285 }
7286 _space->object_iterate_mem(mr, &_scan_cl);
7287 }
7288
7289 SweepClosure::SweepClosure(CMSCollector* collector,
7290 ConcurrentMarkSweepGeneration* g,
7291 CMSBitMap* bitMap, bool should_yield) :
7292 _collector(collector),
7293 _g(g),
7294 _sp(g->cmsSpace()),
7295 _limit(_sp->sweep_limit()),
7296 _freelistLock(_sp->freelistLock()),
7297 _bitMap(bitMap),
7298 _yield(should_yield),
7299 _inFreeRange(false), // No free range at beginning of sweep
7300 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7301 _lastFreeRangeCoalesced(false),
7302 _freeFinger(g->used_region().start())
7303 {
7304 NOT_PRODUCT(
7305 _numObjectsFreed = 0;
7306 _numWordsFreed = 0;
7307 _numObjectsLive = 0;
7308 _numWordsLive = 0;
7309 _numObjectsAlreadyFree = 0;
7310 _numWordsAlreadyFree = 0;
7311 _last_fc = NULL;
7312
7313 _sp->initializeIndexedFreeListArrayReturnedBytes();
7314 _sp->dictionary()->initialize_dict_returned_bytes();
7315 )
7316 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7317 "sweep _limit out of bounds");
7318 if (CMSTraceSweeper) {
7319 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
7320 p2i(_limit));
7321 }
7322 }
7323
7324 void SweepClosure::print_on(outputStream* st) const {
7325 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7326 p2i(_sp->bottom()), p2i(_sp->end()));
7327 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7328 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7329 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7330 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7331 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7332 }
7333
7334 #ifndef PRODUCT
7335 // Assertion checking only: no useful work in product mode --
7336 // however, if any of the flags below become product flags,
7337 // you may need to review this code to see if it needs to be
7338 // enabled in product mode.
7339 SweepClosure::~SweepClosure() {
7340 assert_lock_strong(_freelistLock);
7341 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7342 "sweep _limit out of bounds");
7343 if (inFreeRange()) {
7344 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
7345 print();
7346 ShouldNotReachHere();
7347 }
7348 if (Verbose && PrintGC) {
7349 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes",
7350 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7351 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
7352 SIZE_FORMAT" bytes "
7353 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
7354 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7355 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7356 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
7357 * sizeof(HeapWord);
7358 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
7359
7360 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7361 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7362 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7363 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7364 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes);
7365 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
7366 indexListReturnedBytes);
7367 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
7368 dict_returned_bytes);
7369 }
7370 }
7371 if (CMSTraceSweeper) {
7372 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
7373 p2i(_limit));
7374 }
7375 }
7376 #endif // PRODUCT
7377
7378 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7379 bool freeRangeInFreeLists) {
7380 if (CMSTraceSweeper) {
7381 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n",
7382 p2i(freeFinger), freeRangeInFreeLists);
7383 }
7384 assert(!inFreeRange(), "Trampling existing free range");
7385 set_inFreeRange(true);
7386 set_lastFreeRangeCoalesced(false);
7387
7388 set_freeFinger(freeFinger);
7389 set_freeRangeInFreeLists(freeRangeInFreeLists);
7390 if (CMSTestInFreeList) {
7391 if (freeRangeInFreeLists) {
7392 FreeChunk* fc = (FreeChunk*) freeFinger;
7393 assert(fc->is_free(), "A chunk on the free list should be free.");
7394 assert(fc->size() > 0, "Free range should have a size");
7395 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7396 }
7397 }
7398 }
7399
7400 // Note that the sweeper runs concurrently with mutators. Thus,
7401 // it is possible for direct allocation in this generation to happen
7402 // in the middle of the sweep. Note that the sweeper also coalesces
7403 // contiguous free blocks. Thus, unless the sweeper and the allocator
7404 // synchronize appropriately freshly allocated blocks may get swept up.
7405 // This is accomplished by the sweeper locking the free lists while
7406 // it is sweeping. Thus blocks that are determined to be free are
7407 // indeed free. There is however one additional complication:
7408 // blocks that have been allocated since the final checkpoint and
7409 // mark, will not have been marked and so would be treated as
7410 // unreachable and swept up. To prevent this, the allocator marks
7411 // the bit map when allocating during the sweep phase. This leads,
7412 // however, to a further complication -- objects may have been allocated
7413 // but not yet initialized -- in the sense that the header isn't yet
7414 // installed. The sweeper can not then determine the size of the block
7415 // in order to skip over it. To deal with this case, we use a technique
7416 // (due to Printezis) to encode such uninitialized block sizes in the
7417 // bit map. Since the bit map uses a bit per every HeapWord, but the
7418 // CMS generation has a minimum object size of 3 HeapWords, it follows
7419 // that "normal marks" won't be adjacent in the bit map (there will
7420 // always be at least two 0 bits between successive 1 bits). We make use
7421 // of these "unused" bits to represent uninitialized blocks -- the bit
7422 // corresponding to the start of the uninitialized object and the next
7423 // bit are both set. Finally, a 1 bit marks the end of the object that
7424 // started with the two consecutive 1 bits to indicate its potentially
7425 // uninitialized state.
7426
7427 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7428 FreeChunk* fc = (FreeChunk*)addr;
7429 size_t res;
7430
7431 // Check if we are done sweeping. Below we check "addr >= _limit" rather
7432 // than "addr == _limit" because although _limit was a block boundary when
7433 // we started the sweep, it may no longer be one because heap expansion
7434 // may have caused us to coalesce the block ending at the address _limit
7435 // with a newly expanded chunk (this happens when _limit was set to the
7436 // previous _end of the space), so we may have stepped past _limit:
7437 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7438 if (addr >= _limit) { // we have swept up to or past the limit: finish up
7439 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7440 "sweep _limit out of bounds");
7441 assert(addr < _sp->end(), "addr out of bounds");
7442 // Flush any free range we might be holding as a single
7443 // coalesced chunk to the appropriate free list.
7444 if (inFreeRange()) {
7445 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7446 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", p2i(freeFinger())));
7447 flush_cur_free_chunk(freeFinger(),
7448 pointer_delta(addr, freeFinger()));
7449 if (CMSTraceSweeper) {
7450 gclog_or_tty->print("Sweep: last chunk: ");
7451 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") "
7452 "[coalesced:%d]\n",
7453 p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7454 lastFreeRangeCoalesced() ? 1 : 0);
7455 }
7456 }
7457
7458 // help the iterator loop finish
7459 return pointer_delta(_sp->end(), addr);
7460 }
7461
7462 assert(addr < _limit, "sweep invariant");
7463 // check if we should yield
7464 do_yield_check(addr);
7465 if (fc->is_free()) {
7466 // Chunk that is already free
7467 res = fc->size();
7468 do_already_free_chunk(fc);
7469 debug_only(_sp->verifyFreeLists());
7470 // If we flush the chunk at hand in lookahead_and_flush()
7471 // and it's coalesced with a preceding chunk, then the
7472 // process of "mangling" the payload of the coalesced block
7473 // will cause erasure of the size information from the
7474 // (erstwhile) header of all the coalesced blocks but the
7475 // first, so the first disjunct in the assert will not hold
7476 // in that specific case (in which case the second disjunct
7477 // will hold).
7478 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7479 "Otherwise the size info doesn't change at this step");
7480 NOT_PRODUCT(
7481 _numObjectsAlreadyFree++;
7482 _numWordsAlreadyFree += res;
7483 )
7484 NOT_PRODUCT(_last_fc = fc;)
7485 } else if (!_bitMap->isMarked(addr)) {
7486 // Chunk is fresh garbage
7487 res = do_garbage_chunk(fc);
7488 debug_only(_sp->verifyFreeLists());
7489 NOT_PRODUCT(
7490 _numObjectsFreed++;
7491 _numWordsFreed += res;
7492 )
7493 } else {
7494 // Chunk that is alive.
7495 res = do_live_chunk(fc);
7496 debug_only(_sp->verifyFreeLists());
7497 NOT_PRODUCT(
7498 _numObjectsLive++;
7499 _numWordsLive += res;
7500 )
7501 }
7502 return res;
7503 }
7504
7505 // For the smart allocation, record following
7506 // split deaths - a free chunk is removed from its free list because
7507 // it is being split into two or more chunks.
7508 // split birth - a free chunk is being added to its free list because
7509 // a larger free chunk has been split and resulted in this free chunk.
7510 // coal death - a free chunk is being removed from its free list because
7511 // it is being coalesced into a large free chunk.
7512 // coal birth - a free chunk is being added to its free list because
7513 // it was created when two or more free chunks where coalesced into
7514 // this free chunk.
7515 //
7516 // These statistics are used to determine the desired number of free
7517 // chunks of a given size. The desired number is chosen to be relative
7518 // to the end of a CMS sweep. The desired number at the end of a sweep
7519 // is the
7520 // count-at-end-of-previous-sweep (an amount that was enough)
7521 // - count-at-beginning-of-current-sweep (the excess)
7522 // + split-births (gains in this size during interval)
7523 // - split-deaths (demands on this size during interval)
7524 // where the interval is from the end of one sweep to the end of the
7525 // next.
7526 //
7527 // When sweeping the sweeper maintains an accumulated chunk which is
7528 // the chunk that is made up of chunks that have been coalesced. That
7529 // will be termed the left-hand chunk. A new chunk of garbage that
7530 // is being considered for coalescing will be referred to as the
7531 // right-hand chunk.
7532 //
7533 // When making a decision on whether to coalesce a right-hand chunk with
7534 // the current left-hand chunk, the current count vs. the desired count
7535 // of the left-hand chunk is considered. Also if the right-hand chunk
7536 // is near the large chunk at the end of the heap (see
7537 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7538 // left-hand chunk is coalesced.
7539 //
7540 // When making a decision about whether to split a chunk, the desired count
7541 // vs. the current count of the candidate to be split is also considered.
7542 // If the candidate is underpopulated (currently fewer chunks than desired)
7543 // a chunk of an overpopulated (currently more chunks than desired) size may
7544 // be chosen. The "hint" associated with a free list, if non-null, points
7545 // to a free list which may be overpopulated.
7546 //
7547
7548 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7549 const size_t size = fc->size();
7550 // Chunks that cannot be coalesced are not in the
7551 // free lists.
7552 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7553 assert(_sp->verify_chunk_in_free_list(fc),
7554 "free chunk should be in free lists");
7555 }
7556 // a chunk that is already free, should not have been
7557 // marked in the bit map
7558 HeapWord* const addr = (HeapWord*) fc;
7559 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7560 // Verify that the bit map has no bits marked between
7561 // addr and purported end of this block.
7562 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7563
7564 // Some chunks cannot be coalesced under any circumstances.
7565 // See the definition of cantCoalesce().
7566 if (!fc->cantCoalesce()) {
7567 // This chunk can potentially be coalesced.
7568 if (_sp->adaptive_freelists()) {
7569 // All the work is done in
7570 do_post_free_or_garbage_chunk(fc, size);
7571 } else { // Not adaptive free lists
7572 // this is a free chunk that can potentially be coalesced by the sweeper;
7573 if (!inFreeRange()) {
7574 // if the next chunk is a free block that can't be coalesced
7575 // it doesn't make sense to remove this chunk from the free lists
7576 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
7577 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
7578 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
7579 nextChunk->is_free() && // ... which is free...
7580 nextChunk->cantCoalesce()) { // ... but can't be coalesced
7581 // nothing to do
7582 } else {
7583 // Potentially the start of a new free range:
7584 // Don't eagerly remove it from the free lists.
7585 // No need to remove it if it will just be put
7586 // back again. (Also from a pragmatic point of view
7587 // if it is a free block in a region that is beyond
7588 // any allocated blocks, an assertion will fail)
7589 // Remember the start of a free run.
7590 initialize_free_range(addr, true);
7591 // end - can coalesce with next chunk
7592 }
7593 } else {
7594 // the midst of a free range, we are coalescing
7595 print_free_block_coalesced(fc);
7596 if (CMSTraceSweeper) {
7597 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size);
7598 }
7599 // remove it from the free lists
7600 _sp->removeFreeChunkFromFreeLists(fc);
7601 set_lastFreeRangeCoalesced(true);
7602 // If the chunk is being coalesced and the current free range is
7603 // in the free lists, remove the current free range so that it
7604 // will be returned to the free lists in its entirety - all
7605 // the coalesced pieces included.
7606 if (freeRangeInFreeLists()) {
7607 FreeChunk* ffc = (FreeChunk*) freeFinger();
7608 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7609 "Size of free range is inconsistent with chunk size.");
7610 if (CMSTestInFreeList) {
7611 assert(_sp->verify_chunk_in_free_list(ffc),
7612 "free range is not in free lists");
7613 }
7614 _sp->removeFreeChunkFromFreeLists(ffc);
7615 set_freeRangeInFreeLists(false);
7616 }
7617 }
7618 }
7619 // Note that if the chunk is not coalescable (the else arm
7620 // below), we unconditionally flush, without needing to do
7621 // a "lookahead," as we do below.
7622 if (inFreeRange()) lookahead_and_flush(fc, size);
7623 } else {
7624 // Code path common to both original and adaptive free lists.
7625
7626 // cant coalesce with previous block; this should be treated
7627 // as the end of a free run if any
7628 if (inFreeRange()) {
7629 // we kicked some butt; time to pick up the garbage
7630 assert(freeFinger() < addr, "freeFinger points too high");
7631 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7632 }
7633 // else, nothing to do, just continue
7634 }
7635 }
7636
7637 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7638 // This is a chunk of garbage. It is not in any free list.
7639 // Add it to a free list or let it possibly be coalesced into
7640 // a larger chunk.
7641 HeapWord* const addr = (HeapWord*) fc;
7642 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7643
7644 if (_sp->adaptive_freelists()) {
7645 // Verify that the bit map has no bits marked between
7646 // addr and purported end of just dead object.
7647 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7648
7649 do_post_free_or_garbage_chunk(fc, size);
7650 } else {
7651 if (!inFreeRange()) {
7652 // start of a new free range
7653 assert(size > 0, "A free range should have a size");
7654 initialize_free_range(addr, false);
7655 } else {
7656 // this will be swept up when we hit the end of the
7657 // free range
7658 if (CMSTraceSweeper) {
7659 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size);
7660 }
7661 // If the chunk is being coalesced and the current free range is
7662 // in the free lists, remove the current free range so that it
7663 // will be returned to the free lists in its entirety - all
7664 // the coalesced pieces included.
7665 if (freeRangeInFreeLists()) {
7666 FreeChunk* ffc = (FreeChunk*)freeFinger();
7667 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7668 "Size of free range is inconsistent with chunk size.");
7669 if (CMSTestInFreeList) {
7670 assert(_sp->verify_chunk_in_free_list(ffc),
7671 "free range is not in free lists");
7672 }
7673 _sp->removeFreeChunkFromFreeLists(ffc);
7674 set_freeRangeInFreeLists(false);
7675 }
7676 set_lastFreeRangeCoalesced(true);
7677 }
7678 // this will be swept up when we hit the end of the free range
7679
7680 // Verify that the bit map has no bits marked between
7681 // addr and purported end of just dead object.
7682 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7683 }
7684 assert(_limit >= addr + size,
7685 "A freshly garbage chunk can't possibly straddle over _limit");
7686 if (inFreeRange()) lookahead_and_flush(fc, size);
7687 return size;
7688 }
7689
7690 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7691 HeapWord* addr = (HeapWord*) fc;
7692 // The sweeper has just found a live object. Return any accumulated
7693 // left hand chunk to the free lists.
7694 if (inFreeRange()) {
7695 assert(freeFinger() < addr, "freeFinger points too high");
7696 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7697 }
7698
7699 // This object is live: we'd normally expect this to be
7700 // an oop, and like to assert the following:
7701 // assert(oop(addr)->is_oop(), "live block should be an oop");
7702 // However, as we commented above, this may be an object whose
7703 // header hasn't yet been initialized.
7704 size_t size;
7705 assert(_bitMap->isMarked(addr), "Tautology for this control point");
7706 if (_bitMap->isMarked(addr + 1)) {
7707 // Determine the size from the bit map, rather than trying to
7708 // compute it from the object header.
7709 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7710 size = pointer_delta(nextOneAddr + 1, addr);
7711 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7712 "alignment problem");
7713
7714 #ifdef ASSERT
7715 if (oop(addr)->klass_or_null() != NULL) {
7716 // Ignore mark word because we are running concurrent with mutators
7717 assert(oop(addr)->is_oop(true), "live block should be an oop");
7718 assert(size ==
7719 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7720 "P-mark and computed size do not agree");
7721 }
7722 #endif
7723
7724 } else {
7725 // This should be an initialized object that's alive.
7726 assert(oop(addr)->klass_or_null() != NULL,
7727 "Should be an initialized object");
7728 // Ignore mark word because we are running concurrent with mutators
7729 assert(oop(addr)->is_oop(true), "live block should be an oop");
7730 // Verify that the bit map has no bits marked between
7731 // addr and purported end of this block.
7732 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7733 assert(size >= 3, "Necessary for Printezis marks to work");
7734 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7735 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7736 }
7737 return size;
7738 }
7739
7740 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7741 size_t chunkSize) {
7742 // do_post_free_or_garbage_chunk() should only be called in the case
7743 // of the adaptive free list allocator.
7744 const bool fcInFreeLists = fc->is_free();
7745 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
7746 assert((HeapWord*)fc <= _limit, "sweep invariant");
7747 if (CMSTestInFreeList && fcInFreeLists) {
7748 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7749 }
7750
7751 if (CMSTraceSweeper) {
7752 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7753 }
7754
7755 HeapWord* const fc_addr = (HeapWord*) fc;
7756
7757 bool coalesce;
7758 const size_t left = pointer_delta(fc_addr, freeFinger());
7759 const size_t right = chunkSize;
7760 switch (FLSCoalescePolicy) {
7761 // numeric value forms a coalition aggressiveness metric
7762 case 0: { // never coalesce
7763 coalesce = false;
7764 break;
7765 }
7766 case 1: { // coalesce if left & right chunks on overpopulated lists
7767 coalesce = _sp->coalOverPopulated(left) &&
7768 _sp->coalOverPopulated(right);
7769 break;
7770 }
7771 case 2: { // coalesce if left chunk on overpopulated list (default)
7772 coalesce = _sp->coalOverPopulated(left);
7773 break;
7774 }
7775 case 3: { // coalesce if left OR right chunk on overpopulated list
7776 coalesce = _sp->coalOverPopulated(left) ||
7777 _sp->coalOverPopulated(right);
7778 break;
7779 }
7780 case 4: { // always coalesce
7781 coalesce = true;
7782 break;
7783 }
7784 default:
7785 ShouldNotReachHere();
7786 }
7787
7788 // Should the current free range be coalesced?
7789 // If the chunk is in a free range and either we decided to coalesce above
7790 // or the chunk is near the large block at the end of the heap
7791 // (isNearLargestChunk() returns true), then coalesce this chunk.
7792 const bool doCoalesce = inFreeRange()
7793 && (coalesce || _g->isNearLargestChunk(fc_addr));
7794 if (doCoalesce) {
7795 // Coalesce the current free range on the left with the new
7796 // chunk on the right. If either is on a free list,
7797 // it must be removed from the list and stashed in the closure.
7798 if (freeRangeInFreeLists()) {
7799 FreeChunk* const ffc = (FreeChunk*)freeFinger();
7800 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7801 "Size of free range is inconsistent with chunk size.");
7802 if (CMSTestInFreeList) {
7803 assert(_sp->verify_chunk_in_free_list(ffc),
7804 "Chunk is not in free lists");
7805 }
7806 _sp->coalDeath(ffc->size());
7807 _sp->removeFreeChunkFromFreeLists(ffc);
7808 set_freeRangeInFreeLists(false);
7809 }
7810 if (fcInFreeLists) {
7811 _sp->coalDeath(chunkSize);
7812 assert(fc->size() == chunkSize,
7813 "The chunk has the wrong size or is not in the free lists");
7814 _sp->removeFreeChunkFromFreeLists(fc);
7815 }
7816 set_lastFreeRangeCoalesced(true);
7817 print_free_block_coalesced(fc);
7818 } else { // not in a free range and/or should not coalesce
7819 // Return the current free range and start a new one.
7820 if (inFreeRange()) {
7821 // In a free range but cannot coalesce with the right hand chunk.
7822 // Put the current free range into the free lists.
7823 flush_cur_free_chunk(freeFinger(),
7824 pointer_delta(fc_addr, freeFinger()));
7825 }
7826 // Set up for new free range. Pass along whether the right hand
7827 // chunk is in the free lists.
7828 initialize_free_range((HeapWord*)fc, fcInFreeLists);
7829 }
7830 }
7831
7832 // Lookahead flush:
7833 // If we are tracking a free range, and this is the last chunk that
7834 // we'll look at because its end crosses past _limit, we'll preemptively
7835 // flush it along with any free range we may be holding on to. Note that
7836 // this can be the case only for an already free or freshly garbage
7837 // chunk. If this block is an object, it can never straddle
7838 // over _limit. The "straddling" occurs when _limit is set at
7839 // the previous end of the space when this cycle started, and
7840 // a subsequent heap expansion caused the previously co-terminal
7841 // free block to be coalesced with the newly expanded portion,
7842 // thus rendering _limit a non-block-boundary making it dangerous
7843 // for the sweeper to step over and examine.
7844 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7845 assert(inFreeRange(), "Should only be called if currently in a free range.");
7846 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7847 assert(_sp->used_region().contains(eob - 1),
7848 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7849 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7850 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7851 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size));
7852 if (eob >= _limit) {
7853 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7854 if (CMSTraceSweeper) {
7855 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
7856 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7857 "[" PTR_FORMAT "," PTR_FORMAT ")",
7858 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7859 }
7860 // Return the storage we are tracking back into the free lists.
7861 if (CMSTraceSweeper) {
7862 gclog_or_tty->print_cr("Flushing ... ");
7863 }
7864 assert(freeFinger() < eob, "Error");
7865 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7866 }
7867 }
7868
7869 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7870 assert(inFreeRange(), "Should only be called if currently in a free range.");
7871 assert(size > 0,
7872 "A zero sized chunk cannot be added to the free lists.");
7873 if (!freeRangeInFreeLists()) {
7874 if (CMSTestInFreeList) {
7875 FreeChunk* fc = (FreeChunk*) chunk;
7876 fc->set_size(size);
7877 assert(!_sp->verify_chunk_in_free_list(fc),
7878 "chunk should not be in free lists yet");
7879 }
7880 if (CMSTraceSweeper) {
7881 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists",
7882 p2i(chunk), size);
7883 }
7884 // A new free range is going to be starting. The current
7885 // free range has not been added to the free lists yet or
7886 // was removed so add it back.
7887 // If the current free range was coalesced, then the death
7888 // of the free range was recorded. Record a birth now.
7889 if (lastFreeRangeCoalesced()) {
7890 _sp->coalBirth(size);
7891 }
7892 _sp->addChunkAndRepairOffsetTable(chunk, size,
7893 lastFreeRangeCoalesced());
7894 } else if (CMSTraceSweeper) {
7895 gclog_or_tty->print_cr("Already in free list: nothing to flush");
7896 }
7897 set_inFreeRange(false);
7898 set_freeRangeInFreeLists(false);
7899 }
7900
7901 // We take a break if we've been at this for a while,
7902 // so as to avoid monopolizing the locks involved.
7903 void SweepClosure::do_yield_work(HeapWord* addr) {
7904 // Return current free chunk being used for coalescing (if any)
7905 // to the appropriate freelist. After yielding, the next
7906 // free block encountered will start a coalescing range of
7907 // free blocks. If the next free block is adjacent to the
7908 // chunk just flushed, they will need to wait for the next
7909 // sweep to be coalesced.
7910 if (inFreeRange()) {
7911 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7912 }
7913
7914 // First give up the locks, then yield, then re-lock.
7915 // We should probably use a constructor/destructor idiom to
7916 // do this unlock/lock or modify the MutexUnlocker class to
7917 // serve our purpose. XXX
7918 assert_lock_strong(_bitMap->lock());
7919 assert_lock_strong(_freelistLock);
7920 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7921 "CMS thread should hold CMS token");
7922 _bitMap->lock()->unlock();
7923 _freelistLock->unlock();
7924 ConcurrentMarkSweepThread::desynchronize(true);
7925 _collector->stopTimer();
7926 if (PrintCMSStatistics != 0) {
7927 _collector->incrementYields();
7928 }
7929
7930 // See the comment in coordinator_yield()
7931 for (unsigned i = 0; i < CMSYieldSleepCount &&
7932 ConcurrentMarkSweepThread::should_yield() &&
7933 !CMSCollector::foregroundGCIsActive(); ++i) {
7934 os::sleep(Thread::current(), 1, false);
7935 }
7936
7937 ConcurrentMarkSweepThread::synchronize(true);
7938 _freelistLock->lock();
7939 _bitMap->lock()->lock_without_safepoint_check();
7940 _collector->startTimer();
7941 }
7942
7943 #ifndef PRODUCT
7944 // This is actually very useful in a product build if it can
7945 // be called from the debugger. Compile it into the product
7946 // as needed.
7947 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7948 return debug_cms_space->verify_chunk_in_free_list(fc);
7949 }
7950 #endif
7951
7952 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7953 if (CMSTraceSweeper) {
7954 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7955 p2i(fc), fc->size());
7956 }
7957 }
7958
7959 // CMSIsAliveClosure
7960 bool CMSIsAliveClosure::do_object_b(oop obj) {
7961 HeapWord* addr = (HeapWord*)obj;
7962 return addr != NULL &&
7963 (!_span.contains(addr) || _bit_map->isMarked(addr));
7964 }
7965
7966
7967 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7968 MemRegion span,
7969 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7970 bool cpc):
7971 _collector(collector),
7972 _span(span),
7973 _bit_map(bit_map),
7974 _mark_stack(mark_stack),
7975 _concurrent_precleaning(cpc) {
7976 assert(!_span.is_empty(), "Empty span could spell trouble");
7977 }
7978
7979
7980 // CMSKeepAliveClosure: the serial version
7981 void CMSKeepAliveClosure::do_oop(oop obj) {
7982 HeapWord* addr = (HeapWord*)obj;
7983 if (_span.contains(addr) &&
7984 !_bit_map->isMarked(addr)) {
7985 _bit_map->mark(addr);
7986 bool simulate_overflow = false;
7987 NOT_PRODUCT(
7988 if (CMSMarkStackOverflowALot &&
7989 _collector->simulate_overflow()) {
7990 // simulate a stack overflow
7991 simulate_overflow = true;
7992 }
7993 )
7994 if (simulate_overflow || !_mark_stack->push(obj)) {
7995 if (_concurrent_precleaning) {
7996 // We dirty the overflown object and let the remark
7997 // phase deal with it.
7998 assert(_collector->overflow_list_is_empty(), "Error");
7999 // In the case of object arrays, we need to dirty all of
8000 // the cards that the object spans. No locking or atomics
8001 // are needed since no one else can be mutating the mod union
8002 // table.
8003 if (obj->is_objArray()) {
8004 size_t sz = obj->size();
8005 HeapWord* end_card_addr =
8006 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
8007 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8008 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8009 _collector->_modUnionTable.mark_range(redirty_range);
8010 } else {
8011 _collector->_modUnionTable.mark(addr);
8012 }
8013 _collector->_ser_kac_preclean_ovflw++;
8014 } else {
8015 _collector->push_on_overflow_list(obj);
8016 _collector->_ser_kac_ovflw++;
8017 }
8018 }
8019 }
8020 }
8021
8022 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8023 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8024
8025 // CMSParKeepAliveClosure: a parallel version of the above.
8026 // The work queues are private to each closure (thread),
8027 // but (may be) available for stealing by other threads.
8028 void CMSParKeepAliveClosure::do_oop(oop obj) {
8029 HeapWord* addr = (HeapWord*)obj;
8030 if (_span.contains(addr) &&
8031 !_bit_map->isMarked(addr)) {
8032 // In general, during recursive tracing, several threads
8033 // may be concurrently getting here; the first one to
8034 // "tag" it, claims it.
8035 if (_bit_map->par_mark(addr)) {
8036 bool res = _work_queue->push(obj);
8037 assert(res, "Low water mark should be much less than capacity");
8038 // Do a recursive trim in the hope that this will keep
8039 // stack usage lower, but leave some oops for potential stealers
8040 trim_queue(_low_water_mark);
8041 } // Else, another thread got there first
8042 }
8043 }
8044
8045 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8046 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8047
8048 void CMSParKeepAliveClosure::trim_queue(uint max) {
8049 while (_work_queue->size() > max) {
8050 oop new_oop;
8051 if (_work_queue->pop_local(new_oop)) {
8052 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8053 assert(_bit_map->isMarked((HeapWord*)new_oop),
8054 "no white objects on this stack!");
8055 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8056 // iterate over the oops in this oop, marking and pushing
8057 // the ones in CMS heap (i.e. in _span).
8058 new_oop->oop_iterate(&_mark_and_push);
8059 }
8060 }
8061 }
8062
8063 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8064 CMSCollector* collector,
8065 MemRegion span, CMSBitMap* bit_map,
8066 OopTaskQueue* work_queue):
8067 _collector(collector),
8068 _span(span),
8069 _bit_map(bit_map),
8070 _work_queue(work_queue) { }
8071
8072 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8073 HeapWord* addr = (HeapWord*)obj;
8074 if (_span.contains(addr) &&
8075 !_bit_map->isMarked(addr)) {
8076 if (_bit_map->par_mark(addr)) {
8077 bool simulate_overflow = false;
8078 NOT_PRODUCT(
8079 if (CMSMarkStackOverflowALot &&
8080 _collector->par_simulate_overflow()) {
8081 // simulate a stack overflow
8082 simulate_overflow = true;
8083 }
8084 )
8085 if (simulate_overflow || !_work_queue->push(obj)) {
8086 _collector->par_push_on_overflow_list(obj);
8087 _collector->_par_kac_ovflw++;
8088 }
8089 } // Else another thread got there already
8090 }
8091 }
8092
8093 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8094 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8095
8096 //////////////////////////////////////////////////////////////////
8097 // CMSExpansionCause /////////////////////////////
8098 //////////////////////////////////////////////////////////////////
8099 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8100 switch (cause) {
8101 case _no_expansion:
8102 return "No expansion";
8103 case _satisfy_free_ratio:
8104 return "Free ratio";
8105 case _satisfy_promotion:
8106 return "Satisfy promotion";
8107 case _satisfy_allocation:
8108 return "allocation";
8109 case _allocate_par_lab:
8110 return "Par LAB";
8111 case _allocate_par_spooling_space:
8112 return "Par Spooling Space";
8113 case _adaptive_size_policy:
8114 return "Ergonomics";
8115 default:
8116 return "unknown";
8117 }
8118 }
8119
8120 void CMSDrainMarkingStackClosure::do_void() {
8121 // the max number to take from overflow list at a time
8122 const size_t num = _mark_stack->capacity()/4;
8123 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8124 "Overflow list should be NULL during concurrent phases");
8125 while (!_mark_stack->isEmpty() ||
8126 // if stack is empty, check the overflow list
8127 _collector->take_from_overflow_list(num, _mark_stack)) {
8128 oop obj = _mark_stack->pop();
8129 HeapWord* addr = (HeapWord*)obj;
8130 assert(_span.contains(addr), "Should be within span");
8131 assert(_bit_map->isMarked(addr), "Should be marked");
8132 assert(obj->is_oop(), "Should be an oop");
8133 obj->oop_iterate(_keep_alive);
8134 }
8135 }
8136
8137 void CMSParDrainMarkingStackClosure::do_void() {
8138 // drain queue
8139 trim_queue(0);
8140 }
8141
8142 // Trim our work_queue so its length is below max at return
8143 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8144 while (_work_queue->size() > max) {
8145 oop new_oop;
8146 if (_work_queue->pop_local(new_oop)) {
8147 assert(new_oop->is_oop(), "Expected an oop");
8148 assert(_bit_map->isMarked((HeapWord*)new_oop),
8149 "no white objects on this stack!");
8150 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8151 // iterate over the oops in this oop, marking and pushing
8152 // the ones in CMS heap (i.e. in _span).
8153 new_oop->oop_iterate(&_mark_and_push);
8154 }
8155 }
8156 }
8157
8158 ////////////////////////////////////////////////////////////////////
8159 // Support for Marking Stack Overflow list handling and related code
8160 ////////////////////////////////////////////////////////////////////
8161 // Much of the following code is similar in shape and spirit to the
8162 // code used in ParNewGC. We should try and share that code
8163 // as much as possible in the future.
8164
8165 #ifndef PRODUCT
8166 // Debugging support for CMSStackOverflowALot
8167
8168 // It's OK to call this multi-threaded; the worst thing
8169 // that can happen is that we'll get a bunch of closely
8170 // spaced simulated overflows, but that's OK, in fact
8171 // probably good as it would exercise the overflow code
8172 // under contention.
8173 bool CMSCollector::simulate_overflow() {
8174 if (_overflow_counter-- <= 0) { // just being defensive
8175 _overflow_counter = CMSMarkStackOverflowInterval;
8176 return true;
8177 } else {
8178 return false;
8179 }
8180 }
8181
8182 bool CMSCollector::par_simulate_overflow() {
8183 return simulate_overflow();
8184 }
8185 #endif
8186
8187 // Single-threaded
8188 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8189 assert(stack->isEmpty(), "Expected precondition");
8190 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8191 size_t i = num;
8192 oop cur = _overflow_list;
8193 const markOop proto = markOopDesc::prototype();
8194 NOT_PRODUCT(ssize_t n = 0;)
8195 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8196 next = oop(cur->mark());
8197 cur->set_mark(proto); // until proven otherwise
8198 assert(cur->is_oop(), "Should be an oop");
8199 bool res = stack->push(cur);
8200 assert(res, "Bit off more than can chew?");
8201 NOT_PRODUCT(n++;)
8202 }
8203 _overflow_list = cur;
8204 #ifndef PRODUCT
8205 assert(_num_par_pushes >= n, "Too many pops?");
8206 _num_par_pushes -=n;
8207 #endif
8208 return !stack->isEmpty();
8209 }
8210
8211 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
8212 // (MT-safe) Get a prefix of at most "num" from the list.
8213 // The overflow list is chained through the mark word of
8214 // each object in the list. We fetch the entire list,
8215 // break off a prefix of the right size and return the
8216 // remainder. If other threads try to take objects from
8217 // the overflow list at that time, they will wait for
8218 // some time to see if data becomes available. If (and
8219 // only if) another thread places one or more object(s)
8220 // on the global list before we have returned the suffix
8221 // to the global list, we will walk down our local list
8222 // to find its end and append the global list to
8223 // our suffix before returning it. This suffix walk can
8224 // prove to be expensive (quadratic in the amount of traffic)
8225 // when there are many objects in the overflow list and
8226 // there is much producer-consumer contention on the list.
8227 // *NOTE*: The overflow list manipulation code here and
8228 // in ParNewGeneration:: are very similar in shape,
8229 // except that in the ParNew case we use the old (from/eden)
8230 // copy of the object to thread the list via its klass word.
8231 // Because of the common code, if you make any changes in
8232 // the code below, please check the ParNew version to see if
8233 // similar changes might be needed.
8234 // CR 6797058 has been filed to consolidate the common code.
8235 bool CMSCollector::par_take_from_overflow_list(size_t num,
8236 OopTaskQueue* work_q,
8237 int no_of_gc_threads) {
8238 assert(work_q->size() == 0, "First empty local work queue");
8239 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8240 if (_overflow_list == NULL) {
8241 return false;
8242 }
8243 // Grab the entire list; we'll put back a suffix
8244 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8245 Thread* tid = Thread::current();
8246 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
8247 // set to ParallelGCThreads.
8248 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
8249 size_t sleep_time_millis = MAX2((size_t)1, num/100);
8250 // If the list is busy, we spin for a short while,
8251 // sleeping between attempts to get the list.
8252 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
8253 os::sleep(tid, sleep_time_millis, false);
8254 if (_overflow_list == NULL) {
8255 // Nothing left to take
8256 return false;
8257 } else if (_overflow_list != BUSY) {
8258 // Try and grab the prefix
8259 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8260 }
8261 }
8262 // If the list was found to be empty, or we spun long
8263 // enough, we give up and return empty-handed. If we leave
8264 // the list in the BUSY state below, it must be the case that
8265 // some other thread holds the overflow list and will set it
8266 // to a non-BUSY state in the future.
8267 if (prefix == NULL || prefix == BUSY) {
8268 // Nothing to take or waited long enough
8269 if (prefix == NULL) {
8270 // Write back the NULL in case we overwrote it with BUSY above
8271 // and it is still the same value.
8272 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8273 }
8274 return false;
8275 }
8276 assert(prefix != NULL && prefix != BUSY, "Error");
8277 size_t i = num;
8278 oop cur = prefix;
8279 // Walk down the first "num" objects, unless we reach the end.
8280 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8281 if (cur->mark() == NULL) {
8282 // We have "num" or fewer elements in the list, so there
8283 // is nothing to return to the global list.
8284 // Write back the NULL in lieu of the BUSY we wrote
8285 // above, if it is still the same value.
8286 if (_overflow_list == BUSY) {
8287 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8288 }
8289 } else {
8290 // Chop off the suffix and return it to the global list.
8291 assert(cur->mark() != BUSY, "Error");
8292 oop suffix_head = cur->mark(); // suffix will be put back on global list
8293 cur->set_mark(NULL); // break off suffix
8294 // It's possible that the list is still in the empty(busy) state
8295 // we left it in a short while ago; in that case we may be
8296 // able to place back the suffix without incurring the cost
8297 // of a walk down the list.
8298 oop observed_overflow_list = _overflow_list;
8299 oop cur_overflow_list = observed_overflow_list;
8300 bool attached = false;
8301 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
8302 observed_overflow_list =
8303 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8304 if (cur_overflow_list == observed_overflow_list) {
8305 attached = true;
8306 break;
8307 } else cur_overflow_list = observed_overflow_list;
8308 }
8309 if (!attached) {
8310 // Too bad, someone else sneaked in (at least) an element; we'll need
8311 // to do a splice. Find tail of suffix so we can prepend suffix to global
8312 // list.
8313 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8314 oop suffix_tail = cur;
8315 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8316 "Tautology");
8317 observed_overflow_list = _overflow_list;
8318 do {
8319 cur_overflow_list = observed_overflow_list;
8320 if (cur_overflow_list != BUSY) {
8321 // Do the splice ...
8322 suffix_tail->set_mark(markOop(cur_overflow_list));
8323 } else { // cur_overflow_list == BUSY
8324 suffix_tail->set_mark(NULL);
8325 }
8326 // ... and try to place spliced list back on overflow_list ...
8327 observed_overflow_list =
8328 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8329 } while (cur_overflow_list != observed_overflow_list);
8330 // ... until we have succeeded in doing so.
8331 }
8332 }
8333
8334 // Push the prefix elements on work_q
8335 assert(prefix != NULL, "control point invariant");
8336 const markOop proto = markOopDesc::prototype();
8337 oop next;
8338 NOT_PRODUCT(ssize_t n = 0;)
8339 for (cur = prefix; cur != NULL; cur = next) {
8340 next = oop(cur->mark());
8341 cur->set_mark(proto); // until proven otherwise
8342 assert(cur->is_oop(), "Should be an oop");
8343 bool res = work_q->push(cur);
8344 assert(res, "Bit off more than we can chew?");
8345 NOT_PRODUCT(n++;)
8346 }
8347 #ifndef PRODUCT
8348 assert(_num_par_pushes >= n, "Too many pops?");
8349 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8350 #endif
8351 return true;
8352 }
8353
8354 // Single-threaded
8355 void CMSCollector::push_on_overflow_list(oop p) {
8356 NOT_PRODUCT(_num_par_pushes++;)
8357 assert(p->is_oop(), "Not an oop");
8358 preserve_mark_if_necessary(p);
8359 p->set_mark((markOop)_overflow_list);
8360 _overflow_list = p;
8361 }
8362
8363 // Multi-threaded; use CAS to prepend to overflow list
8364 void CMSCollector::par_push_on_overflow_list(oop p) {
8365 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8366 assert(p->is_oop(), "Not an oop");
8367 par_preserve_mark_if_necessary(p);
8368 oop observed_overflow_list = _overflow_list;
8369 oop cur_overflow_list;
8370 do {
8371 cur_overflow_list = observed_overflow_list;
8372 if (cur_overflow_list != BUSY) {
8373 p->set_mark(markOop(cur_overflow_list));
8374 } else {
8375 p->set_mark(NULL);
8376 }
8377 observed_overflow_list =
8378 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8379 } while (cur_overflow_list != observed_overflow_list);
8380 }
8381 #undef BUSY
8382
8383 // Single threaded
8384 // General Note on GrowableArray: pushes may silently fail
8385 // because we are (temporarily) out of C-heap for expanding
8386 // the stack. The problem is quite ubiquitous and affects
8387 // a lot of code in the JVM. The prudent thing for GrowableArray
8388 // to do (for now) is to exit with an error. However, that may
8389 // be too draconian in some cases because the caller may be
8390 // able to recover without much harm. For such cases, we
8391 // should probably introduce a "soft_push" method which returns
8392 // an indication of success or failure with the assumption that
8393 // the caller may be able to recover from a failure; code in
8394 // the VM can then be changed, incrementally, to deal with such
8395 // failures where possible, thus, incrementally hardening the VM
8396 // in such low resource situations.
8397 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8398 _preserved_oop_stack.push(p);
8399 _preserved_mark_stack.push(m);
8400 assert(m == p->mark(), "Mark word changed");
8401 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8402 "bijection");
8403 }
8404
8405 // Single threaded
8406 void CMSCollector::preserve_mark_if_necessary(oop p) {
8407 markOop m = p->mark();
8408 if (m->must_be_preserved(p)) {
8409 preserve_mark_work(p, m);
8410 }
8411 }
8412
8413 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8414 markOop m = p->mark();
8415 if (m->must_be_preserved(p)) {
8416 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8417 // Even though we read the mark word without holding
8418 // the lock, we are assured that it will not change
8419 // because we "own" this oop, so no other thread can
8420 // be trying to push it on the overflow list; see
8421 // the assertion in preserve_mark_work() that checks
8422 // that m == p->mark().
8423 preserve_mark_work(p, m);
8424 }
8425 }
8426
8427 // We should be able to do this multi-threaded,
8428 // a chunk of stack being a task (this is
8429 // correct because each oop only ever appears
8430 // once in the overflow list. However, it's
8431 // not very easy to completely overlap this with
8432 // other operations, so will generally not be done
8433 // until all work's been completed. Because we
8434 // expect the preserved oop stack (set) to be small,
8435 // it's probably fine to do this single-threaded.
8436 // We can explore cleverer concurrent/overlapped/parallel
8437 // processing of preserved marks if we feel the
8438 // need for this in the future. Stack overflow should
8439 // be so rare in practice and, when it happens, its
8440 // effect on performance so great that this will
8441 // likely just be in the noise anyway.
8442 void CMSCollector::restore_preserved_marks_if_any() {
8443 assert(SafepointSynchronize::is_at_safepoint(),
8444 "world should be stopped");
8445 assert(Thread::current()->is_ConcurrentGC_thread() ||
8446 Thread::current()->is_VM_thread(),
8447 "should be single-threaded");
8448 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8449 "bijection");
8450
8451 while (!_preserved_oop_stack.is_empty()) {
8452 oop p = _preserved_oop_stack.pop();
8453 assert(p->is_oop(), "Should be an oop");
8454 assert(_span.contains(p), "oop should be in _span");
8455 assert(p->mark() == markOopDesc::prototype(),
8456 "Set when taken from overflow list");
8457 markOop m = _preserved_mark_stack.pop();
8458 p->set_mark(m);
8459 }
8460 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8461 "stacks were cleared above");
8462 }
8463
8464 #ifndef PRODUCT
8465 bool CMSCollector::no_preserved_marks() const {
8466 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8467 }
8468 #endif
8469
8470 // Transfer some number of overflown objects to usual marking
8471 // stack. Return true if some objects were transferred.
8472 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8473 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8474 (size_t)ParGCDesiredObjsFromOverflowList);
8475
8476 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8477 assert(_collector->overflow_list_is_empty() || res,
8478 "If list is not empty, we should have taken something");
8479 assert(!res || !_mark_stack->isEmpty(),
8480 "If we took something, it should now be on our stack");
8481 return res;
8482 }
8483
8484 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8485 size_t res = _sp->block_size_no_stall(addr, _collector);
8486 if (_sp->block_is_obj(addr)) {
8487 if (_live_bit_map->isMarked(addr)) {
8488 // It can't have been dead in a previous cycle
8489 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8490 } else {
8491 _dead_bit_map->mark(addr); // mark the dead object
8492 }
8493 }
8494 // Could be 0, if the block size could not be computed without stalling.
8495 return res;
8496 }
8497
8498 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8499
8500 switch (phase) {
8501 case CMSCollector::InitialMarking:
8502 initialize(true /* fullGC */ ,
8503 cause /* cause of the GC */,
8504 true /* recordGCBeginTime */,
8505 true /* recordPreGCUsage */,
8506 false /* recordPeakUsage */,
8507 false /* recordPostGCusage */,
8508 true /* recordAccumulatedGCTime */,
8509 false /* recordGCEndTime */,
8510 false /* countCollection */ );
8511 break;
8512
8513 case CMSCollector::FinalMarking:
8514 initialize(true /* fullGC */ ,
8515 cause /* cause of the GC */,
8516 false /* recordGCBeginTime */,
8517 false /* recordPreGCUsage */,
8518 false /* recordPeakUsage */,
8519 false /* recordPostGCusage */,
8520 true /* recordAccumulatedGCTime */,
8521 false /* recordGCEndTime */,
8522 false /* countCollection */ );
8523 break;
8524
8525 case CMSCollector::Sweeping:
8526 initialize(true /* fullGC */ ,
8527 cause /* cause of the GC */,
8528 false /* recordGCBeginTime */,
8529 false /* recordPreGCUsage */,
8530 true /* recordPeakUsage */,
8531 true /* recordPostGCusage */,
8532 false /* recordAccumulatedGCTime */,
8533 true /* recordGCEndTime */,
8534 true /* countCollection */ );
8535 break;
8536
8537 default:
8538 ShouldNotReachHere();
8539 }
8540 }