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