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