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/cmsHeap.hpp"
33 #include "gc/cms/cmsOopClosures.inline.hpp"
34 #include "gc/cms/compactibleFreeListSpace.hpp"
35 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp"
36 #include "gc/cms/concurrentMarkSweepThread.hpp"
37 #include "gc/cms/parNewGeneration.hpp"
38 #include "gc/cms/vmCMSOperations.hpp"
39 #include "gc/serial/genMarkSweep.hpp"
40 #include "gc/serial/tenuredGeneration.hpp"
41 #include "gc/shared/adaptiveSizePolicy.hpp"
42 #include "gc/shared/cardGeneration.inline.hpp"
43 #include "gc/shared/cardTableRS.hpp"
44 #include "gc/shared/collectedHeap.inline.hpp"
45 #include "gc/shared/collectorCounters.hpp"
46 #include "gc/shared/collectorPolicy.hpp"
47 #include "gc/shared/gcLocker.inline.hpp"
48 #include "gc/shared/gcPolicyCounters.hpp"
49 #include "gc/shared/gcTimer.hpp"
50 #include "gc/shared/gcTrace.hpp"
51 #include "gc/shared/gcTraceTime.inline.hpp"
52 #include "gc/shared/genCollectedHeap.hpp"
53 #include "gc/shared/genOopClosures.inline.hpp"
54 #include "gc/shared/isGCActiveMark.hpp"
55 #include "gc/shared/referencePolicy.hpp"
56 #include "gc/shared/strongRootsScope.hpp"
57 #include "gc/shared/taskqueue.inline.hpp"
58 #include "gc/shared/weakProcessor.hpp"
59 #include "logging/log.hpp"
60 #include "logging/logStream.hpp"
61 #include "memory/allocation.hpp"
62 #include "memory/iterator.inline.hpp"
63 #include "memory/padded.hpp"
64 #include "memory/resourceArea.hpp"
65 #include "oops/oop.inline.hpp"
66 #include "prims/jvmtiExport.hpp"
67 #include "runtime/atomic.hpp"
68 #include "runtime/globals_extension.hpp"
69 #include "runtime/handles.inline.hpp"
70 #include "runtime/java.hpp"
71 #include "runtime/orderAccess.inline.hpp"
72 #include "runtime/timer.hpp"
73 #include "runtime/vmThread.hpp"
74 #include "services/memoryService.hpp"
75 #include "services/runtimeService.hpp"
76 #include "utilities/align.hpp"
77 #include "utilities/stack.inline.hpp"
78
79 // statics
80 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
81 bool CMSCollector::_full_gc_requested = false;
82 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
83
84 //////////////////////////////////////////////////////////////////
85 // In support of CMS/VM thread synchronization
86 //////////////////////////////////////////////////////////////////
87 // We split use of the CGC_lock into 2 "levels".
88 // The low-level locking is of the usual CGC_lock monitor. We introduce
89 // a higher level "token" (hereafter "CMS token") built on top of the
90 // low level monitor (hereafter "CGC lock").
91 // The token-passing protocol gives priority to the VM thread. The
92 // CMS-lock doesn't provide any fairness guarantees, but clients
93 // should ensure that it is only held for very short, bounded
94 // durations.
95 //
96 // When either of the CMS thread or the VM thread is involved in
97 // collection operations during which it does not want the other
98 // thread to interfere, it obtains the CMS token.
99 //
100 // If either thread tries to get the token while the other has
101 // it, that thread waits. However, if the VM thread and CMS thread
102 // both want the token, then the VM thread gets priority while the
103 // CMS thread waits. This ensures, for instance, that the "concurrent"
104 // phases of the CMS thread's work do not block out the VM thread
105 // for long periods of time as the CMS thread continues to hog
106 // the token. (See bug 4616232).
107 //
108 // The baton-passing functions are, however, controlled by the
109 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
110 // and here the low-level CMS lock, not the high level token,
111 // ensures mutual exclusion.
112 //
113 // Two important conditions that we have to satisfy:
114 // 1. if a thread does a low-level wait on the CMS lock, then it
115 // relinquishes the CMS token if it were holding that token
116 // when it acquired the low-level CMS lock.
117 // 2. any low-level notifications on the low-level lock
118 // should only be sent when a thread has relinquished the token.
119 //
120 // In the absence of either property, we'd have potential deadlock.
121 //
122 // We protect each of the CMS (concurrent and sequential) phases
123 // with the CMS _token_, not the CMS _lock_.
124 //
125 // The only code protected by CMS lock is the token acquisition code
126 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
127 // baton-passing code.
128 //
129 // Unfortunately, i couldn't come up with a good abstraction to factor and
130 // hide the naked CGC_lock manipulation in the baton-passing code
131 // further below. That's something we should try to do. Also, the proof
132 // of correctness of this 2-level locking scheme is far from obvious,
133 // and potentially quite slippery. We have an uneasy suspicion, for instance,
134 // that there may be a theoretical possibility of delay/starvation in the
135 // low-level lock/wait/notify scheme used for the baton-passing because of
136 // potential interference with the priority scheme embodied in the
137 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
138 // invocation further below and marked with "XXX 20011219YSR".
139 // Indeed, as we note elsewhere, this may become yet more slippery
140 // in the presence of multiple CMS and/or multiple VM threads. XXX
141
142 class CMSTokenSync: public StackObj {
143 private:
144 bool _is_cms_thread;
145 public:
146 CMSTokenSync(bool is_cms_thread):
147 _is_cms_thread(is_cms_thread) {
148 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
149 "Incorrect argument to constructor");
150 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
151 }
152
153 ~CMSTokenSync() {
154 assert(_is_cms_thread ?
155 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
156 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
157 "Incorrect state");
158 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
159 }
160 };
161
162 // Convenience class that does a CMSTokenSync, and then acquires
163 // upto three locks.
164 class CMSTokenSyncWithLocks: public CMSTokenSync {
165 private:
166 // Note: locks are acquired in textual declaration order
167 // and released in the opposite order
168 MutexLockerEx _locker1, _locker2, _locker3;
169 public:
170 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
171 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
172 CMSTokenSync(is_cms_thread),
173 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
174 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
175 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
176 { }
177 };
178
179
180 //////////////////////////////////////////////////////////////////
181 // Concurrent Mark-Sweep Generation /////////////////////////////
182 //////////////////////////////////////////////////////////////////
183
184 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
185
186 // This struct contains per-thread things necessary to support parallel
187 // young-gen collection.
188 class CMSParGCThreadState: public CHeapObj<mtGC> {
189 public:
190 CompactibleFreeListSpaceLAB lab;
191 PromotionInfo promo;
192
193 // Constructor.
194 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
195 promo.setSpace(cfls);
196 }
197 };
198
199 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
200 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) :
201 CardGeneration(rs, initial_byte_size, ct),
202 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
203 _did_compact(false)
204 {
205 HeapWord* bottom = (HeapWord*) _virtual_space.low();
206 HeapWord* end = (HeapWord*) _virtual_space.high();
207
208 _direct_allocated_words = 0;
209 NOT_PRODUCT(
210 _numObjectsPromoted = 0;
211 _numWordsPromoted = 0;
212 _numObjectsAllocated = 0;
213 _numWordsAllocated = 0;
214 )
215
216 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end));
217 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
218 _cmsSpace->_old_gen = this;
219
220 _gc_stats = new CMSGCStats();
221
222 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
223 // offsets match. The ability to tell free chunks from objects
224 // depends on this property.
225 debug_only(
226 FreeChunk* junk = NULL;
227 assert(UseCompressedClassPointers ||
228 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
229 "Offset of FreeChunk::_prev within FreeChunk must match"
230 " that of OopDesc::_klass within OopDesc");
231 )
232
233 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC);
234 for (uint i = 0; i < ParallelGCThreads; i++) {
235 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
236 }
237
238 _incremental_collection_failed = false;
239 // The "dilatation_factor" is the expansion that can occur on
240 // account of the fact that the minimum object size in the CMS
241 // generation may be larger than that in, say, a contiguous young
242 // generation.
243 // Ideally, in the calculation below, we'd compute the dilatation
244 // factor as: MinChunkSize/(promoting_gen's min object size)
245 // Since we do not have such a general query interface for the
246 // promoting generation, we'll instead just use the minimum
247 // object size (which today is a header's worth of space);
248 // note that all arithmetic is in units of HeapWords.
249 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
250 assert(_dilatation_factor >= 1.0, "from previous assert");
251 }
252
253
254 // The field "_initiating_occupancy" represents the occupancy percentage
255 // at which we trigger a new collection cycle. Unless explicitly specified
256 // via CMSInitiatingOccupancyFraction (argument "io" below), it
257 // is calculated by:
258 //
259 // Let "f" be MinHeapFreeRatio in
260 //
261 // _initiating_occupancy = 100-f +
262 // f * (CMSTriggerRatio/100)
263 // where CMSTriggerRatio is the argument "tr" below.
264 //
265 // That is, if we assume the heap is at its desired maximum occupancy at the
266 // end of a collection, we let CMSTriggerRatio of the (purported) free
267 // space be allocated before initiating a new collection cycle.
268 //
269 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
270 assert(io <= 100 && tr <= 100, "Check the arguments");
271 if (io >= 0) {
272 _initiating_occupancy = (double)io / 100.0;
273 } else {
274 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
275 (double)(tr * MinHeapFreeRatio) / 100.0)
276 / 100.0;
277 }
278 }
279
280 void ConcurrentMarkSweepGeneration::ref_processor_init() {
281 assert(collector() != NULL, "no collector");
282 collector()->ref_processor_init();
283 }
284
285 void CMSCollector::ref_processor_init() {
286 if (_ref_processor == NULL) {
287 // Allocate and initialize a reference processor
288 _ref_processor =
289 new ReferenceProcessor(_span, // span
290 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
291 ParallelGCThreads, // mt processing degree
292 _cmsGen->refs_discovery_is_mt(), // mt discovery
293 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
294 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
295 &_is_alive_closure); // closure for liveness info
296 // Initialize the _ref_processor field of CMSGen
297 _cmsGen->set_ref_processor(_ref_processor);
298
299 }
300 }
301
302 AdaptiveSizePolicy* CMSCollector::size_policy() {
303 CMSHeap* heap = CMSHeap::heap();
304 return heap->gen_policy()->size_policy();
305 }
306
307 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
308
309 const char* gen_name = "old";
310 GenCollectorPolicy* gcp = CMSHeap::heap()->gen_policy();
311 // Generation Counters - generation 1, 1 subspace
312 _gen_counters = new GenerationCounters(gen_name, 1, 1,
313 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
314
315 _space_counters = new GSpaceCounters(gen_name, 0,
316 _virtual_space.reserved_size(),
317 this, _gen_counters);
318 }
319
320 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
321 _cms_gen(cms_gen)
322 {
323 assert(alpha <= 100, "bad value");
324 _saved_alpha = alpha;
325
326 // Initialize the alphas to the bootstrap value of 100.
327 _gc0_alpha = _cms_alpha = 100;
328
329 _cms_begin_time.update();
330 _cms_end_time.update();
331
332 _gc0_duration = 0.0;
333 _gc0_period = 0.0;
334 _gc0_promoted = 0;
335
336 _cms_duration = 0.0;
337 _cms_period = 0.0;
338 _cms_allocated = 0;
339
340 _cms_used_at_gc0_begin = 0;
341 _cms_used_at_gc0_end = 0;
342 _allow_duty_cycle_reduction = false;
343 _valid_bits = 0;
344 }
345
346 double CMSStats::cms_free_adjustment_factor(size_t free) const {
347 // TBD: CR 6909490
348 return 1.0;
349 }
350
351 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
352 }
353
354 // If promotion failure handling is on use
355 // the padded average size of the promotion for each
356 // young generation collection.
357 double CMSStats::time_until_cms_gen_full() const {
358 size_t cms_free = _cms_gen->cmsSpace()->free();
359 CMSHeap* heap = CMSHeap::heap();
360 size_t expected_promotion = MIN2(heap->young_gen()->capacity(),
361 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
362 if (cms_free > expected_promotion) {
363 // Start a cms collection if there isn't enough space to promote
364 // for the next young collection. Use the padded average as
365 // a safety factor.
366 cms_free -= expected_promotion;
367
368 // Adjust by the safety factor.
369 double cms_free_dbl = (double)cms_free;
370 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0;
371 // Apply a further correction factor which tries to adjust
372 // for recent occurance of concurrent mode failures.
373 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
374 cms_free_dbl = cms_free_dbl * cms_adjustment;
375
376 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
377 cms_free, expected_promotion);
378 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0);
379 // Add 1 in case the consumption rate goes to zero.
380 return cms_free_dbl / (cms_consumption_rate() + 1.0);
381 }
382 return 0.0;
383 }
384
385 // Compare the duration of the cms collection to the
386 // time remaining before the cms generation is empty.
387 // Note that the time from the start of the cms collection
388 // to the start of the cms sweep (less than the total
389 // duration of the cms collection) can be used. This
390 // has been tried and some applications experienced
391 // promotion failures early in execution. This was
392 // possibly because the averages were not accurate
393 // enough at the beginning.
394 double CMSStats::time_until_cms_start() const {
395 // We add "gc0_period" to the "work" calculation
396 // below because this query is done (mostly) at the
397 // end of a scavenge, so we need to conservatively
398 // account for that much possible delay
399 // in the query so as to avoid concurrent mode failures
400 // due to starting the collection just a wee bit too
401 // late.
402 double work = cms_duration() + gc0_period();
403 double deadline = time_until_cms_gen_full();
404 // If a concurrent mode failure occurred recently, we want to be
405 // more conservative and halve our expected time_until_cms_gen_full()
406 if (work > deadline) {
407 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ",
408 cms_duration(), gc0_period(), time_until_cms_gen_full());
409 return 0.0;
410 }
411 return work - deadline;
412 }
413
414 #ifndef PRODUCT
415 void CMSStats::print_on(outputStream *st) const {
416 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
417 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
418 gc0_duration(), gc0_period(), gc0_promoted());
419 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
420 cms_duration(), cms_period(), cms_allocated());
421 st->print(",cms_since_beg=%g,cms_since_end=%g",
422 cms_time_since_begin(), cms_time_since_end());
423 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
424 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
425
426 if (valid()) {
427 st->print(",promo_rate=%g,cms_alloc_rate=%g",
428 promotion_rate(), cms_allocation_rate());
429 st->print(",cms_consumption_rate=%g,time_until_full=%g",
430 cms_consumption_rate(), time_until_cms_gen_full());
431 }
432 st->cr();
433 }
434 #endif // #ifndef PRODUCT
435
436 CMSCollector::CollectorState CMSCollector::_collectorState =
437 CMSCollector::Idling;
438 bool CMSCollector::_foregroundGCIsActive = false;
439 bool CMSCollector::_foregroundGCShouldWait = false;
440
441 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
442 CardTableRS* ct,
443 ConcurrentMarkSweepPolicy* cp):
444 _cmsGen(cmsGen),
445 _ct(ct),
446 _ref_processor(NULL), // will be set later
447 _conc_workers(NULL), // may be set later
448 _abort_preclean(false),
449 _start_sampling(false),
450 _between_prologue_and_epilogue(false),
451 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
452 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
453 -1 /* lock-free */, "No_lock" /* dummy */),
454 _modUnionClosurePar(&_modUnionTable),
455 // Adjust my span to cover old (cms) gen
456 _span(cmsGen->reserved()),
457 // Construct the is_alive_closure with _span & markBitMap
458 _is_alive_closure(_span, &_markBitMap),
459 _restart_addr(NULL),
460 _overflow_list(NULL),
461 _stats(cmsGen),
462 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
463 //verify that this lock should be acquired with safepoint check.
464 Monitor::_safepoint_check_sometimes)),
465 _eden_chunk_array(NULL), // may be set in ctor body
466 _eden_chunk_capacity(0), // -- ditto --
467 _eden_chunk_index(0), // -- ditto --
468 _survivor_plab_array(NULL), // -- ditto --
469 _survivor_chunk_array(NULL), // -- ditto --
470 _survivor_chunk_capacity(0), // -- ditto --
471 _survivor_chunk_index(0), // -- ditto --
472 _ser_pmc_preclean_ovflw(0),
473 _ser_kac_preclean_ovflw(0),
474 _ser_pmc_remark_ovflw(0),
475 _par_pmc_remark_ovflw(0),
476 _ser_kac_ovflw(0),
477 _par_kac_ovflw(0),
478 #ifndef PRODUCT
479 _num_par_pushes(0),
480 #endif
481 _collection_count_start(0),
482 _verifying(false),
483 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
484 _completed_initialization(false),
485 _collector_policy(cp),
486 _should_unload_classes(CMSClassUnloadingEnabled),
487 _concurrent_cycles_since_last_unload(0),
488 _roots_scanning_options(GenCollectedHeap::SO_None),
489 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
490 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
491 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
492 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
493 _cms_start_registered(false)
494 {
495 // Now expand the span and allocate the collection support structures
496 // (MUT, marking bit map etc.) to cover both generations subject to
497 // collection.
498
499 // For use by dirty card to oop closures.
500 _cmsGen->cmsSpace()->set_collector(this);
501
502 // Allocate MUT and marking bit map
503 {
504 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
505 if (!_markBitMap.allocate(_span)) {
506 log_warning(gc)("Failed to allocate CMS Bit Map");
507 return;
508 }
509 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
510 }
511 {
512 _modUnionTable.allocate(_span);
513 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
514 }
515
516 if (!_markStack.allocate(MarkStackSize)) {
517 log_warning(gc)("Failed to allocate CMS Marking Stack");
518 return;
519 }
520
521 // Support for multi-threaded concurrent phases
522 if (CMSConcurrentMTEnabled) {
523 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
524 // just for now
525 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4);
526 }
527 if (ConcGCThreads > 1) {
528 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
529 ConcGCThreads, true);
530 if (_conc_workers == NULL) {
531 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled");
532 CMSConcurrentMTEnabled = false;
533 } else {
534 _conc_workers->initialize_workers();
535 }
536 } else {
537 CMSConcurrentMTEnabled = false;
538 }
539 }
540 if (!CMSConcurrentMTEnabled) {
541 ConcGCThreads = 0;
542 } else {
543 // Turn off CMSCleanOnEnter optimization temporarily for
544 // the MT case where it's not fixed yet; see 6178663.
545 CMSCleanOnEnter = false;
546 }
547 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
548 "Inconsistency");
549 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads);
550 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
551
552 // Parallel task queues; these are shared for the
553 // concurrent and stop-world phases of CMS, but
554 // are not shared with parallel scavenge (ParNew).
555 {
556 uint i;
557 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads);
558
559 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
560 || ParallelRefProcEnabled)
561 && num_queues > 0) {
562 _task_queues = new OopTaskQueueSet(num_queues);
563 if (_task_queues == NULL) {
564 log_warning(gc)("task_queues allocation failure.");
565 return;
566 }
567 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
568 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
569 for (i = 0; i < num_queues; i++) {
570 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
571 if (q == NULL) {
572 log_warning(gc)("work_queue allocation failure.");
573 return;
574 }
575 _task_queues->register_queue(i, q);
576 }
577 for (i = 0; i < num_queues; i++) {
578 _task_queues->queue(i)->initialize();
579 _hash_seed[i] = 17; // copied from ParNew
580 }
581 }
582 }
583
584 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
585
586 // Clip CMSBootstrapOccupancy between 0 and 100.
587 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0;
588
589 // Now tell CMS generations the identity of their collector
590 ConcurrentMarkSweepGeneration::set_collector(this);
591
592 // Create & start a CMS thread for this CMS collector
593 _cmsThread = ConcurrentMarkSweepThread::start(this);
594 assert(cmsThread() != NULL, "CMS Thread should have been created");
595 assert(cmsThread()->collector() == this,
596 "CMS Thread should refer to this gen");
597 assert(CGC_lock != NULL, "Where's the CGC_lock?");
598
599 // Support for parallelizing young gen rescan
600 CMSHeap* heap = CMSHeap::heap();
601 assert(heap->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
602 _young_gen = (ParNewGeneration*)heap->young_gen();
603 if (heap->supports_inline_contig_alloc()) {
604 _top_addr = heap->top_addr();
605 _end_addr = heap->end_addr();
606 assert(_young_gen != NULL, "no _young_gen");
607 _eden_chunk_index = 0;
608 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain;
609 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
610 }
611
612 // Support for parallelizing survivor space rescan
613 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
614 const size_t max_plab_samples =
615 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize);
616
617 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
618 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
619 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
620 _survivor_chunk_capacity = max_plab_samples;
621 for (uint i = 0; i < ParallelGCThreads; i++) {
622 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
623 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
624 assert(cur->end() == 0, "Should be 0");
625 assert(cur->array() == vec, "Should be vec");
626 assert(cur->capacity() == max_plab_samples, "Error");
627 }
628 }
629
630 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
631 _gc_counters = new CollectorCounters("CMS", 1);
632 _completed_initialization = true;
633 _inter_sweep_timer.start(); // start of time
634 }
635
636 const char* ConcurrentMarkSweepGeneration::name() const {
637 return "concurrent mark-sweep generation";
638 }
639 void ConcurrentMarkSweepGeneration::update_counters() {
640 if (UsePerfData) {
641 _space_counters->update_all();
642 _gen_counters->update_all();
643 }
644 }
645
646 // this is an optimized version of update_counters(). it takes the
647 // used value as a parameter rather than computing it.
648 //
649 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
650 if (UsePerfData) {
651 _space_counters->update_used(used);
652 _space_counters->update_capacity();
653 _gen_counters->update_all();
654 }
655 }
656
657 void ConcurrentMarkSweepGeneration::print() const {
658 Generation::print();
659 cmsSpace()->print();
660 }
661
662 #ifndef PRODUCT
663 void ConcurrentMarkSweepGeneration::print_statistics() {
664 cmsSpace()->printFLCensus(0);
665 }
666 #endif
667
668 size_t
669 ConcurrentMarkSweepGeneration::contiguous_available() const {
670 // dld proposes an improvement in precision here. If the committed
671 // part of the space ends in a free block we should add that to
672 // uncommitted size in the calculation below. Will make this
673 // change later, staying with the approximation below for the
674 // time being. -- ysr.
675 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
676 }
677
678 size_t
679 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
680 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
681 }
682
683 size_t ConcurrentMarkSweepGeneration::max_available() const {
684 return free() + _virtual_space.uncommitted_size();
685 }
686
687 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
688 size_t available = max_available();
689 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
690 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
691 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")",
692 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes);
693 return res;
694 }
695
696 // At a promotion failure dump information on block layout in heap
697 // (cms old generation).
698 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
699 Log(gc, promotion) log;
700 if (log.is_trace()) {
701 LogStream ls(log.trace());
702 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls);
703 }
704 }
705
706 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
707 // Clear the promotion information. These pointers can be adjusted
708 // along with all the other pointers into the heap but
709 // compaction is expected to be a rare event with
710 // a heap using cms so don't do it without seeing the need.
711 for (uint i = 0; i < ParallelGCThreads; i++) {
712 _par_gc_thread_states[i]->promo.reset();
713 }
714 }
715
716 void ConcurrentMarkSweepGeneration::compute_new_size() {
717 assert_locked_or_safepoint(Heap_lock);
718
719 // If incremental collection failed, we just want to expand
720 // to the limit.
721 if (incremental_collection_failed()) {
722 clear_incremental_collection_failed();
723 grow_to_reserved();
724 return;
725 }
726
727 // The heap has been compacted but not reset yet.
728 // Any metric such as free() or used() will be incorrect.
729
730 CardGeneration::compute_new_size();
731
732 // Reset again after a possible resizing
733 if (did_compact()) {
734 cmsSpace()->reset_after_compaction();
735 }
736 }
737
738 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
739 assert_locked_or_safepoint(Heap_lock);
740
741 // If incremental collection failed, we just want to expand
742 // to the limit.
743 if (incremental_collection_failed()) {
744 clear_incremental_collection_failed();
745 grow_to_reserved();
746 return;
747 }
748
749 double free_percentage = ((double) free()) / capacity();
750 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
751 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
752
753 // compute expansion delta needed for reaching desired free percentage
754 if (free_percentage < desired_free_percentage) {
755 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
756 assert(desired_capacity >= capacity(), "invalid expansion size");
757 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
758 Log(gc) log;
759 if (log.is_trace()) {
760 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
761 log.trace("From compute_new_size: ");
762 log.trace(" Free fraction %f", free_percentage);
763 log.trace(" Desired free fraction %f", desired_free_percentage);
764 log.trace(" Maximum free fraction %f", maximum_free_percentage);
765 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000);
766 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000);
767 CMSHeap* heap = CMSHeap::heap();
768 assert(heap->is_old_gen(this), "The CMS generation should always be the old generation");
769 size_t young_size = heap->young_gen()->capacity();
770 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000);
771 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000);
772 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000);
773 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes);
774 }
775 // safe if expansion fails
776 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
777 log.trace(" Expanded free fraction %f", ((double) free()) / capacity());
778 } else {
779 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
780 assert(desired_capacity <= capacity(), "invalid expansion size");
781 size_t shrink_bytes = capacity() - desired_capacity;
782 // Don't shrink unless the delta is greater than the minimum shrink we want
783 if (shrink_bytes >= MinHeapDeltaBytes) {
784 shrink_free_list_by(shrink_bytes);
785 }
786 }
787 }
788
789 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
790 return cmsSpace()->freelistLock();
791 }
792
793 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) {
794 CMSSynchronousYieldRequest yr;
795 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
796 return have_lock_and_allocate(size, tlab);
797 }
798
799 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
800 bool tlab /* ignored */) {
801 assert_lock_strong(freelistLock());
802 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
803 HeapWord* res = cmsSpace()->allocate(adjustedSize);
804 // Allocate the object live (grey) if the background collector has
805 // started marking. This is necessary because the marker may
806 // have passed this address and consequently this object will
807 // not otherwise be greyed and would be incorrectly swept up.
808 // Note that if this object contains references, the writing
809 // of those references will dirty the card containing this object
810 // allowing the object to be blackened (and its references scanned)
811 // either during a preclean phase or at the final checkpoint.
812 if (res != NULL) {
813 // We may block here with an uninitialized object with
814 // its mark-bit or P-bits not yet set. Such objects need
815 // to be safely navigable by block_start().
816 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
817 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
818 collector()->direct_allocated(res, adjustedSize);
819 _direct_allocated_words += adjustedSize;
820 // allocation counters
821 NOT_PRODUCT(
822 _numObjectsAllocated++;
823 _numWordsAllocated += (int)adjustedSize;
824 )
825 }
826 return res;
827 }
828
829 // In the case of direct allocation by mutators in a generation that
830 // is being concurrently collected, the object must be allocated
831 // live (grey) if the background collector has started marking.
832 // This is necessary because the marker may
833 // have passed this address and consequently this object will
834 // not otherwise be greyed and would be incorrectly swept up.
835 // Note that if this object contains references, the writing
836 // of those references will dirty the card containing this object
837 // allowing the object to be blackened (and its references scanned)
838 // either during a preclean phase or at the final checkpoint.
839 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
840 assert(_markBitMap.covers(start, size), "Out of bounds");
841 if (_collectorState >= Marking) {
842 MutexLockerEx y(_markBitMap.lock(),
843 Mutex::_no_safepoint_check_flag);
844 // [see comments preceding SweepClosure::do_blk() below for details]
845 //
846 // Can the P-bits be deleted now? JJJ
847 //
848 // 1. need to mark the object as live so it isn't collected
849 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
850 // 3. need to mark the end of the object so marking, precleaning or sweeping
851 // can skip over uninitialized or unparsable objects. An allocated
852 // object is considered uninitialized for our purposes as long as
853 // its klass word is NULL. All old gen objects are parsable
854 // as soon as they are initialized.)
855 _markBitMap.mark(start); // object is live
856 _markBitMap.mark(start + 1); // object is potentially uninitialized?
857 _markBitMap.mark(start + size - 1);
858 // mark end of object
859 }
860 // check that oop looks uninitialized
861 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
862 }
863
864 void CMSCollector::promoted(bool par, HeapWord* start,
865 bool is_obj_array, size_t obj_size) {
866 assert(_markBitMap.covers(start), "Out of bounds");
867 // See comment in direct_allocated() about when objects should
868 // be allocated live.
869 if (_collectorState >= Marking) {
870 // we already hold the marking bit map lock, taken in
871 // the prologue
872 if (par) {
873 _markBitMap.par_mark(start);
874 } else {
875 _markBitMap.mark(start);
876 }
877 // We don't need to mark the object as uninitialized (as
878 // in direct_allocated above) because this is being done with the
879 // world stopped and the object will be initialized by the
880 // time the marking, precleaning or sweeping get to look at it.
881 // But see the code for copying objects into the CMS generation,
882 // where we need to ensure that concurrent readers of the
883 // block offset table are able to safely navigate a block that
884 // is in flux from being free to being allocated (and in
885 // transition while being copied into) and subsequently
886 // becoming a bona-fide object when the copy/promotion is complete.
887 assert(SafepointSynchronize::is_at_safepoint(),
888 "expect promotion only at safepoints");
889
890 if (_collectorState < Sweeping) {
891 // Mark the appropriate cards in the modUnionTable, so that
892 // this object gets scanned before the sweep. If this is
893 // not done, CMS generation references in the object might
894 // not get marked.
895 // For the case of arrays, which are otherwise precisely
896 // marked, we need to dirty the entire array, not just its head.
897 if (is_obj_array) {
898 // The [par_]mark_range() method expects mr.end() below to
899 // be aligned to the granularity of a bit's representation
900 // in the heap. In the case of the MUT below, that's a
901 // card size.
902 MemRegion mr(start,
903 align_up(start + obj_size,
904 CardTableModRefBS::card_size /* bytes */));
905 if (par) {
906 _modUnionTable.par_mark_range(mr);
907 } else {
908 _modUnionTable.mark_range(mr);
909 }
910 } else { // not an obj array; we can just mark the head
911 if (par) {
912 _modUnionTable.par_mark(start);
913 } else {
914 _modUnionTable.mark(start);
915 }
916 }
917 }
918 }
919 }
920
921 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
922 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
923 // allocate, copy and if necessary update promoinfo --
924 // delegate to underlying space.
925 assert_lock_strong(freelistLock());
926
927 #ifndef PRODUCT
928 if (CMSHeap::heap()->promotion_should_fail()) {
929 return NULL;
930 }
931 #endif // #ifndef PRODUCT
932
933 oop res = _cmsSpace->promote(obj, obj_size);
934 if (res == NULL) {
935 // expand and retry
936 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
937 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
938 // Since this is the old generation, we don't try to promote
939 // into a more senior generation.
940 res = _cmsSpace->promote(obj, obj_size);
941 }
942 if (res != NULL) {
943 // See comment in allocate() about when objects should
944 // be allocated live.
945 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below");
946 collector()->promoted(false, // Not parallel
947 (HeapWord*)res, obj->is_objArray(), obj_size);
948 // promotion counters
949 NOT_PRODUCT(
950 _numObjectsPromoted++;
951 _numWordsPromoted +=
952 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
953 )
954 }
955 return res;
956 }
957
958
959 // IMPORTANT: Notes on object size recognition in CMS.
960 // ---------------------------------------------------
961 // A block of storage in the CMS generation is always in
962 // one of three states. A free block (FREE), an allocated
963 // object (OBJECT) whose size() method reports the correct size,
964 // and an intermediate state (TRANSIENT) in which its size cannot
965 // be accurately determined.
966 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
967 // -----------------------------------------------------
968 // FREE: klass_word & 1 == 1; mark_word holds block size
969 //
970 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
971 // obj->size() computes correct size
972 //
973 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
974 //
975 // STATE IDENTIFICATION: (64 bit+COOPS)
976 // ------------------------------------
977 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
978 //
979 // OBJECT: klass_word installed; klass_word != 0;
980 // obj->size() computes correct size
981 //
982 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
983 //
984 //
985 // STATE TRANSITION DIAGRAM
986 //
987 // mut / parnew mut / parnew
988 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
989 // ^ |
990 // |------------------------ DEAD <------------------------------------|
991 // sweep mut
992 //
993 // While a block is in TRANSIENT state its size cannot be determined
994 // so readers will either need to come back later or stall until
995 // the size can be determined. Note that for the case of direct
996 // allocation, P-bits, when available, may be used to determine the
997 // size of an object that may not yet have been initialized.
998
999 // Things to support parallel young-gen collection.
1000 oop
1001 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1002 oop old, markOop m,
1003 size_t word_sz) {
1004 #ifndef PRODUCT
1005 if (CMSHeap::heap()->promotion_should_fail()) {
1006 return NULL;
1007 }
1008 #endif // #ifndef PRODUCT
1009
1010 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1011 PromotionInfo* promoInfo = &ps->promo;
1012 // if we are tracking promotions, then first ensure space for
1013 // promotion (including spooling space for saving header if necessary).
1014 // then allocate and copy, then track promoted info if needed.
1015 // When tracking (see PromotionInfo::track()), the mark word may
1016 // be displaced and in this case restoration of the mark word
1017 // occurs in the (oop_since_save_marks_)iterate phase.
1018 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1019 // Out of space for allocating spooling buffers;
1020 // try expanding and allocating spooling buffers.
1021 if (!expand_and_ensure_spooling_space(promoInfo)) {
1022 return NULL;
1023 }
1024 }
1025 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant");
1026 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1027 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1028 if (obj_ptr == NULL) {
1029 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1030 if (obj_ptr == NULL) {
1031 return NULL;
1032 }
1033 }
1034 oop obj = oop(obj_ptr);
1035 OrderAccess::storestore();
1036 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1037 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1038 // IMPORTANT: See note on object initialization for CMS above.
1039 // Otherwise, copy the object. Here we must be careful to insert the
1040 // klass pointer last, since this marks the block as an allocated object.
1041 // Except with compressed oops it's the mark word.
1042 HeapWord* old_ptr = (HeapWord*)old;
1043 // Restore the mark word copied above.
1044 obj->set_mark(m);
1045 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1046 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1047 OrderAccess::storestore();
1048
1049 if (UseCompressedClassPointers) {
1050 // Copy gap missed by (aligned) header size calculation below
1051 obj->set_klass_gap(old->klass_gap());
1052 }
1053 if (word_sz > (size_t)oopDesc::header_size()) {
1054 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1055 obj_ptr + oopDesc::header_size(),
1056 word_sz - oopDesc::header_size());
1057 }
1058
1059 // Now we can track the promoted object, if necessary. We take care
1060 // to delay the transition from uninitialized to full object
1061 // (i.e., insertion of klass pointer) until after, so that it
1062 // atomically becomes a promoted object.
1063 if (promoInfo->tracking()) {
1064 promoInfo->track((PromotedObject*)obj, old->klass());
1065 }
1066 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1067 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1068 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below");
1069
1070 // Finally, install the klass pointer (this should be volatile).
1071 OrderAccess::storestore();
1072 obj->set_klass(old->klass());
1073 // We should now be able to calculate the right size for this object
1074 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1075
1076 collector()->promoted(true, // parallel
1077 obj_ptr, old->is_objArray(), word_sz);
1078
1079 NOT_PRODUCT(
1080 Atomic::inc(&_numObjectsPromoted);
1081 Atomic::add(alloc_sz, &_numWordsPromoted);
1082 )
1083
1084 return obj;
1085 }
1086
1087 void
1088 ConcurrentMarkSweepGeneration::
1089 par_promote_alloc_done(int thread_num) {
1090 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1091 ps->lab.retire(thread_num);
1092 }
1093
1094 void
1095 ConcurrentMarkSweepGeneration::
1096 par_oop_since_save_marks_iterate_done(int thread_num) {
1097 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1098 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1099 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1100
1101 // Because card-scanning has been completed, subsequent phases
1102 // (e.g., reference processing) will not need to recognize which
1103 // objects have been promoted during this GC. So, we can now disable
1104 // promotion tracking.
1105 ps->promo.stopTrackingPromotions();
1106 }
1107
1108 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1109 size_t size,
1110 bool tlab)
1111 {
1112 // We allow a STW collection only if a full
1113 // collection was requested.
1114 return full || should_allocate(size, tlab); // FIX ME !!!
1115 // This and promotion failure handling are connected at the
1116 // hip and should be fixed by untying them.
1117 }
1118
1119 bool CMSCollector::shouldConcurrentCollect() {
1120 LogTarget(Trace, gc) log;
1121
1122 if (_full_gc_requested) {
1123 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)");
1124 return true;
1125 }
1126
1127 FreelistLocker x(this);
1128 // ------------------------------------------------------------------
1129 // Print out lots of information which affects the initiation of
1130 // a collection.
1131 if (log.is_enabled() && stats().valid()) {
1132 log.print("CMSCollector shouldConcurrentCollect: ");
1133
1134 LogStream out(log);
1135 stats().print_on(&out);
1136
1137 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full());
1138 log.print("free=" SIZE_FORMAT, _cmsGen->free());
1139 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available());
1140 log.print("promotion_rate=%g", stats().promotion_rate());
1141 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate());
1142 log.print("occupancy=%3.7f", _cmsGen->occupancy());
1143 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1144 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1145 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1146 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect());
1147 }
1148 // ------------------------------------------------------------------
1149
1150 // If the estimated time to complete a cms collection (cms_duration())
1151 // is less than the estimated time remaining until the cms generation
1152 // is full, start a collection.
1153 if (!UseCMSInitiatingOccupancyOnly) {
1154 if (stats().valid()) {
1155 if (stats().time_until_cms_start() == 0.0) {
1156 return true;
1157 }
1158 } else {
1159 // We want to conservatively collect somewhat early in order
1160 // to try and "bootstrap" our CMS/promotion statistics;
1161 // this branch will not fire after the first successful CMS
1162 // collection because the stats should then be valid.
1163 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1164 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f",
1165 _cmsGen->occupancy(), _bootstrap_occupancy);
1166 return true;
1167 }
1168 }
1169 }
1170
1171 // Otherwise, we start a collection cycle if
1172 // old gen want a collection cycle started. Each may use
1173 // an appropriate criterion for making this decision.
1174 // XXX We need to make sure that the gen expansion
1175 // criterion dovetails well with this. XXX NEED TO FIX THIS
1176 if (_cmsGen->should_concurrent_collect()) {
1177 log.print("CMS old gen initiated");
1178 return true;
1179 }
1180
1181 // We start a collection if we believe an incremental collection may fail;
1182 // this is not likely to be productive in practice because it's probably too
1183 // late anyway.
1184 CMSHeap* heap = CMSHeap::heap();
1185 if (heap->incremental_collection_will_fail(true /* consult_young */)) {
1186 log.print("CMSCollector: collect because incremental collection will fail ");
1187 return true;
1188 }
1189
1190 if (MetaspaceGC::should_concurrent_collect()) {
1191 log.print("CMSCollector: collect for metadata allocation ");
1192 return true;
1193 }
1194
1195 // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1196 if (CMSTriggerInterval >= 0) {
1197 if (CMSTriggerInterval == 0) {
1198 // Trigger always
1199 return true;
1200 }
1201
1202 // Check the CMS time since begin (we do not check the stats validity
1203 // as we want to be able to trigger the first CMS cycle as well)
1204 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1205 if (stats().valid()) {
1206 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1207 stats().cms_time_since_begin());
1208 } else {
1209 log.print("CMSCollector: collect because of trigger interval (first collection)");
1210 }
1211 return true;
1212 }
1213 }
1214
1215 return false;
1216 }
1217
1218 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1219
1220 // Clear _expansion_cause fields of constituent generations
1221 void CMSCollector::clear_expansion_cause() {
1222 _cmsGen->clear_expansion_cause();
1223 }
1224
1225 // We should be conservative in starting a collection cycle. To
1226 // start too eagerly runs the risk of collecting too often in the
1227 // extreme. To collect too rarely falls back on full collections,
1228 // which works, even if not optimum in terms of concurrent work.
1229 // As a work around for too eagerly collecting, use the flag
1230 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1231 // giving the user an easily understandable way of controlling the
1232 // collections.
1233 // We want to start a new collection cycle if any of the following
1234 // conditions hold:
1235 // . our current occupancy exceeds the configured initiating occupancy
1236 // for this generation, or
1237 // . we recently needed to expand this space and have not, since that
1238 // expansion, done a collection of this generation, or
1239 // . the underlying space believes that it may be a good idea to initiate
1240 // a concurrent collection (this may be based on criteria such as the
1241 // following: the space uses linear allocation and linear allocation is
1242 // going to fail, or there is believed to be excessive fragmentation in
1243 // the generation, etc... or ...
1244 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1245 // the case of the old generation; see CR 6543076):
1246 // we may be approaching a point at which allocation requests may fail because
1247 // we will be out of sufficient free space given allocation rate estimates.]
1248 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1249
1250 assert_lock_strong(freelistLock());
1251 if (occupancy() > initiating_occupancy()) {
1252 log_trace(gc)(" %s: collect because of occupancy %f / %f ",
1253 short_name(), occupancy(), initiating_occupancy());
1254 return true;
1255 }
1256 if (UseCMSInitiatingOccupancyOnly) {
1257 return false;
1258 }
1259 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1260 log_trace(gc)(" %s: collect because expanded for allocation ", short_name());
1261 return true;
1262 }
1263 return false;
1264 }
1265
1266 void ConcurrentMarkSweepGeneration::collect(bool full,
1267 bool clear_all_soft_refs,
1268 size_t size,
1269 bool tlab)
1270 {
1271 collector()->collect(full, clear_all_soft_refs, size, tlab);
1272 }
1273
1274 void CMSCollector::collect(bool full,
1275 bool clear_all_soft_refs,
1276 size_t size,
1277 bool tlab)
1278 {
1279 // The following "if" branch is present for defensive reasons.
1280 // In the current uses of this interface, it can be replaced with:
1281 // assert(!GCLocker.is_active(), "Can't be called otherwise");
1282 // But I am not placing that assert here to allow future
1283 // generality in invoking this interface.
1284 if (GCLocker::is_active()) {
1285 // A consistency test for GCLocker
1286 assert(GCLocker::needs_gc(), "Should have been set already");
1287 // Skip this foreground collection, instead
1288 // expanding the heap if necessary.
1289 // Need the free list locks for the call to free() in compute_new_size()
1290 compute_new_size();
1291 return;
1292 }
1293 acquire_control_and_collect(full, clear_all_soft_refs);
1294 }
1295
1296 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1297 CMSHeap* heap = CMSHeap::heap();
1298 unsigned int gc_count = heap->total_full_collections();
1299 if (gc_count == full_gc_count) {
1300 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1301 _full_gc_requested = true;
1302 _full_gc_cause = cause;
1303 CGC_lock->notify(); // nudge CMS thread
1304 } else {
1305 assert(gc_count > full_gc_count, "Error: causal loop");
1306 }
1307 }
1308
1309 bool CMSCollector::is_external_interruption() {
1310 GCCause::Cause cause = CMSHeap::heap()->gc_cause();
1311 return GCCause::is_user_requested_gc(cause) ||
1312 GCCause::is_serviceability_requested_gc(cause);
1313 }
1314
1315 void CMSCollector::report_concurrent_mode_interruption() {
1316 if (is_external_interruption()) {
1317 log_debug(gc)("Concurrent mode interrupted");
1318 } else {
1319 log_debug(gc)("Concurrent mode failure");
1320 _gc_tracer_cm->report_concurrent_mode_failure();
1321 }
1322 }
1323
1324
1325 // The foreground and background collectors need to coordinate in order
1326 // to make sure that they do not mutually interfere with CMS collections.
1327 // When a background collection is active,
1328 // the foreground collector may need to take over (preempt) and
1329 // synchronously complete an ongoing collection. Depending on the
1330 // frequency of the background collections and the heap usage
1331 // of the application, this preemption can be seldom or frequent.
1332 // There are only certain
1333 // points in the background collection that the "collection-baton"
1334 // can be passed to the foreground collector.
1335 //
1336 // The foreground collector will wait for the baton before
1337 // starting any part of the collection. The foreground collector
1338 // will only wait at one location.
1339 //
1340 // The background collector will yield the baton before starting a new
1341 // phase of the collection (e.g., before initial marking, marking from roots,
1342 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1343 // of the loop which switches the phases. The background collector does some
1344 // of the phases (initial mark, final re-mark) with the world stopped.
1345 // Because of locking involved in stopping the world,
1346 // the foreground collector should not block waiting for the background
1347 // collector when it is doing a stop-the-world phase. The background
1348 // collector will yield the baton at an additional point just before
1349 // it enters a stop-the-world phase. Once the world is stopped, the
1350 // background collector checks the phase of the collection. If the
1351 // phase has not changed, it proceeds with the collection. If the
1352 // phase has changed, it skips that phase of the collection. See
1353 // the comments on the use of the Heap_lock in collect_in_background().
1354 //
1355 // Variable used in baton passing.
1356 // _foregroundGCIsActive - Set to true by the foreground collector when
1357 // it wants the baton. The foreground clears it when it has finished
1358 // the collection.
1359 // _foregroundGCShouldWait - Set to true by the background collector
1360 // when it is running. The foreground collector waits while
1361 // _foregroundGCShouldWait is true.
1362 // CGC_lock - monitor used to protect access to the above variables
1363 // and to notify the foreground and background collectors.
1364 // _collectorState - current state of the CMS collection.
1365 //
1366 // The foreground collector
1367 // acquires the CGC_lock
1368 // sets _foregroundGCIsActive
1369 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1370 // various locks acquired in preparation for the collection
1371 // are released so as not to block the background collector
1372 // that is in the midst of a collection
1373 // proceeds with the collection
1374 // clears _foregroundGCIsActive
1375 // returns
1376 //
1377 // The background collector in a loop iterating on the phases of the
1378 // collection
1379 // acquires the CGC_lock
1380 // sets _foregroundGCShouldWait
1381 // if _foregroundGCIsActive is set
1382 // clears _foregroundGCShouldWait, notifies _CGC_lock
1383 // waits on _CGC_lock for _foregroundGCIsActive to become false
1384 // and exits the loop.
1385 // otherwise
1386 // proceed with that phase of the collection
1387 // if the phase is a stop-the-world phase,
1388 // yield the baton once more just before enqueueing
1389 // the stop-world CMS operation (executed by the VM thread).
1390 // returns after all phases of the collection are done
1391 //
1392
1393 void CMSCollector::acquire_control_and_collect(bool full,
1394 bool clear_all_soft_refs) {
1395 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1396 assert(!Thread::current()->is_ConcurrentGC_thread(),
1397 "shouldn't try to acquire control from self!");
1398
1399 // Start the protocol for acquiring control of the
1400 // collection from the background collector (aka CMS thread).
1401 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1402 "VM thread should have CMS token");
1403 // Remember the possibly interrupted state of an ongoing
1404 // concurrent collection
1405 CollectorState first_state = _collectorState;
1406
1407 // Signal to a possibly ongoing concurrent collection that
1408 // we want to do a foreground collection.
1409 _foregroundGCIsActive = true;
1410
1411 // release locks and wait for a notify from the background collector
1412 // releasing the locks in only necessary for phases which
1413 // do yields to improve the granularity of the collection.
1414 assert_lock_strong(bitMapLock());
1415 // We need to lock the Free list lock for the space that we are
1416 // currently collecting.
1417 assert(haveFreelistLocks(), "Must be holding free list locks");
1418 bitMapLock()->unlock();
1419 releaseFreelistLocks();
1420 {
1421 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1422 if (_foregroundGCShouldWait) {
1423 // We are going to be waiting for action for the CMS thread;
1424 // it had better not be gone (for instance at shutdown)!
1425 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(),
1426 "CMS thread must be running");
1427 // Wait here until the background collector gives us the go-ahead
1428 ConcurrentMarkSweepThread::clear_CMS_flag(
1429 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1430 // Get a possibly blocked CMS thread going:
1431 // Note that we set _foregroundGCIsActive true above,
1432 // without protection of the CGC_lock.
1433 CGC_lock->notify();
1434 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1435 "Possible deadlock");
1436 while (_foregroundGCShouldWait) {
1437 // wait for notification
1438 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1439 // Possibility of delay/starvation here, since CMS token does
1440 // not know to give priority to VM thread? Actually, i think
1441 // there wouldn't be any delay/starvation, but the proof of
1442 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1443 }
1444 ConcurrentMarkSweepThread::set_CMS_flag(
1445 ConcurrentMarkSweepThread::CMS_vm_has_token);
1446 }
1447 }
1448 // The CMS_token is already held. Get back the other locks.
1449 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1450 "VM thread should have CMS token");
1451 getFreelistLocks();
1452 bitMapLock()->lock_without_safepoint_check();
1453 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d",
1454 p2i(Thread::current()), first_state);
1455 log_debug(gc, state)(" gets control with state %d", _collectorState);
1456
1457 // Inform cms gen if this was due to partial collection failing.
1458 // The CMS gen may use this fact to determine its expansion policy.
1459 CMSHeap* heap = CMSHeap::heap();
1460 if (heap->incremental_collection_will_fail(false /* don't consult_young */)) {
1461 assert(!_cmsGen->incremental_collection_failed(),
1462 "Should have been noticed, reacted to and cleared");
1463 _cmsGen->set_incremental_collection_failed();
1464 }
1465
1466 if (first_state > Idling) {
1467 report_concurrent_mode_interruption();
1468 }
1469
1470 set_did_compact(true);
1471
1472 // If the collection is being acquired from the background
1473 // collector, there may be references on the discovered
1474 // references lists. Abandon those references, since some
1475 // of them may have become unreachable after concurrent
1476 // discovery; the STW compacting collector will redo discovery
1477 // more precisely, without being subject to floating garbage.
1478 // Leaving otherwise unreachable references in the discovered
1479 // lists would require special handling.
1480 ref_processor()->disable_discovery();
1481 ref_processor()->abandon_partial_discovery();
1482 ref_processor()->verify_no_references_recorded();
1483
1484 if (first_state > Idling) {
1485 save_heap_summary();
1486 }
1487
1488 do_compaction_work(clear_all_soft_refs);
1489
1490 // Has the GC time limit been exceeded?
1491 size_t max_eden_size = _young_gen->max_eden_size();
1492 GCCause::Cause gc_cause = heap->gc_cause();
1493 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1494 _young_gen->eden()->used(),
1495 _cmsGen->max_capacity(),
1496 max_eden_size,
1497 full,
1498 gc_cause,
1499 heap->collector_policy());
1500
1501 // Reset the expansion cause, now that we just completed
1502 // a collection cycle.
1503 clear_expansion_cause();
1504 _foregroundGCIsActive = false;
1505 return;
1506 }
1507
1508 // Resize the tenured generation
1509 // after obtaining the free list locks for the
1510 // two generations.
1511 void CMSCollector::compute_new_size() {
1512 assert_locked_or_safepoint(Heap_lock);
1513 FreelistLocker z(this);
1514 MetaspaceGC::compute_new_size();
1515 _cmsGen->compute_new_size_free_list();
1516 }
1517
1518 // A work method used by the foreground collector to do
1519 // a mark-sweep-compact.
1520 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1521 CMSHeap* heap = CMSHeap::heap();
1522
1523 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1524 gc_timer->register_gc_start();
1525
1526 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1527 gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start());
1528
1529 heap->pre_full_gc_dump(gc_timer);
1530
1531 GCTraceTime(Trace, gc, phases) t("CMS:MSC");
1532
1533 // Temporarily widen the span of the weak reference processing to
1534 // the entire heap.
1535 MemRegion new_span(CMSHeap::heap()->reserved_region());
1536 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1537 // Temporarily, clear the "is_alive_non_header" field of the
1538 // reference processor.
1539 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1540 // Temporarily make reference _processing_ single threaded (non-MT).
1541 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1542 // Temporarily make refs discovery atomic
1543 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1544 // Temporarily make reference _discovery_ single threaded (non-MT)
1545 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1546
1547 ref_processor()->set_enqueuing_is_done(false);
1548 ref_processor()->enable_discovery();
1549 ref_processor()->setup_policy(clear_all_soft_refs);
1550 // If an asynchronous collection finishes, the _modUnionTable is
1551 // all clear. If we are assuming the collection from an asynchronous
1552 // collection, clear the _modUnionTable.
1553 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1554 "_modUnionTable should be clear if the baton was not passed");
1555 _modUnionTable.clear_all();
1556 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(),
1557 "mod union for klasses should be clear if the baton was passed");
1558 _ct->cld_rem_set()->clear_mod_union();
1559
1560
1561 // We must adjust the allocation statistics being maintained
1562 // in the free list space. We do so by reading and clearing
1563 // the sweep timer and updating the block flux rate estimates below.
1564 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1565 if (_inter_sweep_timer.is_active()) {
1566 _inter_sweep_timer.stop();
1567 // Note that we do not use this sample to update the _inter_sweep_estimate.
1568 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1569 _inter_sweep_estimate.padded_average(),
1570 _intra_sweep_estimate.padded_average());
1571 }
1572
1573 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs);
1574 #ifdef ASSERT
1575 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1576 size_t free_size = cms_space->free();
1577 assert(free_size ==
1578 pointer_delta(cms_space->end(), cms_space->compaction_top())
1579 * HeapWordSize,
1580 "All the free space should be compacted into one chunk at top");
1581 assert(cms_space->dictionary()->total_chunk_size(
1582 debug_only(cms_space->freelistLock())) == 0 ||
1583 cms_space->totalSizeInIndexedFreeLists() == 0,
1584 "All the free space should be in a single chunk");
1585 size_t num = cms_space->totalCount();
1586 assert((free_size == 0 && num == 0) ||
1587 (free_size > 0 && (num == 1 || num == 2)),
1588 "There should be at most 2 free chunks after compaction");
1589 #endif // ASSERT
1590 _collectorState = Resetting;
1591 assert(_restart_addr == NULL,
1592 "Should have been NULL'd before baton was passed");
1593 reset_stw();
1594 _cmsGen->reset_after_compaction();
1595 _concurrent_cycles_since_last_unload = 0;
1596
1597 // Clear any data recorded in the PLAB chunk arrays.
1598 if (_survivor_plab_array != NULL) {
1599 reset_survivor_plab_arrays();
1600 }
1601
1602 // Adjust the per-size allocation stats for the next epoch.
1603 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1604 // Restart the "inter sweep timer" for the next epoch.
1605 _inter_sweep_timer.reset();
1606 _inter_sweep_timer.start();
1607
1608 // No longer a need to do a concurrent collection for Metaspace.
1609 MetaspaceGC::set_should_concurrent_collect(false);
1610
1611 heap->post_full_gc_dump(gc_timer);
1612
1613 gc_timer->register_gc_end();
1614
1615 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1616
1617 // For a mark-sweep-compact, compute_new_size() will be called
1618 // in the heap's do_collection() method.
1619 }
1620
1621 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1622 Log(gc, heap) log;
1623 if (!log.is_trace()) {
1624 return;
1625 }
1626
1627 ContiguousSpace* eden_space = _young_gen->eden();
1628 ContiguousSpace* from_space = _young_gen->from();
1629 ContiguousSpace* to_space = _young_gen->to();
1630 // Eden
1631 if (_eden_chunk_array != NULL) {
1632 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1633 p2i(eden_space->bottom()), p2i(eden_space->top()),
1634 p2i(eden_space->end()), eden_space->capacity());
1635 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT,
1636 _eden_chunk_index, _eden_chunk_capacity);
1637 for (size_t i = 0; i < _eden_chunk_index; i++) {
1638 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i]));
1639 }
1640 }
1641 // Survivor
1642 if (_survivor_chunk_array != NULL) {
1643 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1644 p2i(from_space->bottom()), p2i(from_space->top()),
1645 p2i(from_space->end()), from_space->capacity());
1646 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT,
1647 _survivor_chunk_index, _survivor_chunk_capacity);
1648 for (size_t i = 0; i < _survivor_chunk_index; i++) {
1649 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i]));
1650 }
1651 }
1652 }
1653
1654 void CMSCollector::getFreelistLocks() const {
1655 // Get locks for all free lists in all generations that this
1656 // collector is responsible for
1657 _cmsGen->freelistLock()->lock_without_safepoint_check();
1658 }
1659
1660 void CMSCollector::releaseFreelistLocks() const {
1661 // Release locks for all free lists in all generations that this
1662 // collector is responsible for
1663 _cmsGen->freelistLock()->unlock();
1664 }
1665
1666 bool CMSCollector::haveFreelistLocks() const {
1667 // Check locks for all free lists in all generations that this
1668 // collector is responsible for
1669 assert_lock_strong(_cmsGen->freelistLock());
1670 PRODUCT_ONLY(ShouldNotReachHere());
1671 return true;
1672 }
1673
1674 // A utility class that is used by the CMS collector to
1675 // temporarily "release" the foreground collector from its
1676 // usual obligation to wait for the background collector to
1677 // complete an ongoing phase before proceeding.
1678 class ReleaseForegroundGC: public StackObj {
1679 private:
1680 CMSCollector* _c;
1681 public:
1682 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1683 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1684 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1685 // allow a potentially blocked foreground collector to proceed
1686 _c->_foregroundGCShouldWait = false;
1687 if (_c->_foregroundGCIsActive) {
1688 CGC_lock->notify();
1689 }
1690 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1691 "Possible deadlock");
1692 }
1693
1694 ~ReleaseForegroundGC() {
1695 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1696 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1697 _c->_foregroundGCShouldWait = true;
1698 }
1699 };
1700
1701 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1702 assert(Thread::current()->is_ConcurrentGC_thread(),
1703 "A CMS asynchronous collection is only allowed on a CMS thread.");
1704
1705 CMSHeap* heap = CMSHeap::heap();
1706 {
1707 bool safepoint_check = Mutex::_no_safepoint_check_flag;
1708 MutexLockerEx hl(Heap_lock, safepoint_check);
1709 FreelistLocker fll(this);
1710 MutexLockerEx x(CGC_lock, safepoint_check);
1711 if (_foregroundGCIsActive) {
1712 // The foreground collector is. Skip this
1713 // background collection.
1714 assert(!_foregroundGCShouldWait, "Should be clear");
1715 return;
1716 } else {
1717 assert(_collectorState == Idling, "Should be idling before start.");
1718 _collectorState = InitialMarking;
1719 register_gc_start(cause);
1720 // Reset the expansion cause, now that we are about to begin
1721 // a new cycle.
1722 clear_expansion_cause();
1723
1724 // Clear the MetaspaceGC flag since a concurrent collection
1725 // is starting but also clear it after the collection.
1726 MetaspaceGC::set_should_concurrent_collect(false);
1727 }
1728 // Decide if we want to enable class unloading as part of the
1729 // ensuing concurrent GC cycle.
1730 update_should_unload_classes();
1731 _full_gc_requested = false; // acks all outstanding full gc requests
1732 _full_gc_cause = GCCause::_no_gc;
1733 // Signal that we are about to start a collection
1734 heap->increment_total_full_collections(); // ... starting a collection cycle
1735 _collection_count_start = heap->total_full_collections();
1736 }
1737
1738 size_t prev_used = _cmsGen->used();
1739
1740 // The change of the collection state is normally done at this level;
1741 // the exceptions are phases that are executed while the world is
1742 // stopped. For those phases the change of state is done while the
1743 // world is stopped. For baton passing purposes this allows the
1744 // background collector to finish the phase and change state atomically.
1745 // The foreground collector cannot wait on a phase that is done
1746 // while the world is stopped because the foreground collector already
1747 // has the world stopped and would deadlock.
1748 while (_collectorState != Idling) {
1749 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d",
1750 p2i(Thread::current()), _collectorState);
1751 // The foreground collector
1752 // holds the Heap_lock throughout its collection.
1753 // holds the CMS token (but not the lock)
1754 // except while it is waiting for the background collector to yield.
1755 //
1756 // The foreground collector should be blocked (not for long)
1757 // if the background collector is about to start a phase
1758 // executed with world stopped. If the background
1759 // collector has already started such a phase, the
1760 // foreground collector is blocked waiting for the
1761 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
1762 // are executed in the VM thread.
1763 //
1764 // The locking order is
1765 // PendingListLock (PLL) -- if applicable (FinalMarking)
1766 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
1767 // CMS token (claimed in
1768 // stop_world_and_do() -->
1769 // safepoint_synchronize() -->
1770 // CMSThread::synchronize())
1771
1772 {
1773 // Check if the FG collector wants us to yield.
1774 CMSTokenSync x(true); // is cms thread
1775 if (waitForForegroundGC()) {
1776 // We yielded to a foreground GC, nothing more to be
1777 // done this round.
1778 assert(_foregroundGCShouldWait == false, "We set it to false in "
1779 "waitForForegroundGC()");
1780 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1781 p2i(Thread::current()), _collectorState);
1782 return;
1783 } else {
1784 // The background collector can run but check to see if the
1785 // foreground collector has done a collection while the
1786 // background collector was waiting to get the CGC_lock
1787 // above. If yes, break so that _foregroundGCShouldWait
1788 // is cleared before returning.
1789 if (_collectorState == Idling) {
1790 break;
1791 }
1792 }
1793 }
1794
1795 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1796 "should be waiting");
1797
1798 switch (_collectorState) {
1799 case InitialMarking:
1800 {
1801 ReleaseForegroundGC x(this);
1802 stats().record_cms_begin();
1803 VM_CMS_Initial_Mark initial_mark_op(this);
1804 VMThread::execute(&initial_mark_op);
1805 }
1806 // The collector state may be any legal state at this point
1807 // since the background collector may have yielded to the
1808 // foreground collector.
1809 break;
1810 case Marking:
1811 // initial marking in checkpointRootsInitialWork has been completed
1812 if (markFromRoots()) { // we were successful
1813 assert(_collectorState == Precleaning, "Collector state should "
1814 "have changed");
1815 } else {
1816 assert(_foregroundGCIsActive, "Internal state inconsistency");
1817 }
1818 break;
1819 case Precleaning:
1820 // marking from roots in markFromRoots has been completed
1821 preclean();
1822 assert(_collectorState == AbortablePreclean ||
1823 _collectorState == FinalMarking,
1824 "Collector state should have changed");
1825 break;
1826 case AbortablePreclean:
1827 abortable_preclean();
1828 assert(_collectorState == FinalMarking, "Collector state should "
1829 "have changed");
1830 break;
1831 case FinalMarking:
1832 {
1833 ReleaseForegroundGC x(this);
1834
1835 VM_CMS_Final_Remark final_remark_op(this);
1836 VMThread::execute(&final_remark_op);
1837 }
1838 assert(_foregroundGCShouldWait, "block post-condition");
1839 break;
1840 case Sweeping:
1841 // final marking in checkpointRootsFinal has been completed
1842 sweep();
1843 assert(_collectorState == Resizing, "Collector state change "
1844 "to Resizing must be done under the free_list_lock");
1845
1846 case Resizing: {
1847 // Sweeping has been completed...
1848 // At this point the background collection has completed.
1849 // Don't move the call to compute_new_size() down
1850 // into code that might be executed if the background
1851 // collection was preempted.
1852 {
1853 ReleaseForegroundGC x(this); // unblock FG collection
1854 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
1855 CMSTokenSync z(true); // not strictly needed.
1856 if (_collectorState == Resizing) {
1857 compute_new_size();
1858 save_heap_summary();
1859 _collectorState = Resetting;
1860 } else {
1861 assert(_collectorState == Idling, "The state should only change"
1862 " because the foreground collector has finished the collection");
1863 }
1864 }
1865 break;
1866 }
1867 case Resetting:
1868 // CMS heap resizing has been completed
1869 reset_concurrent();
1870 assert(_collectorState == Idling, "Collector state should "
1871 "have changed");
1872
1873 MetaspaceGC::set_should_concurrent_collect(false);
1874
1875 stats().record_cms_end();
1876 // Don't move the concurrent_phases_end() and compute_new_size()
1877 // calls to here because a preempted background collection
1878 // has it's state set to "Resetting".
1879 break;
1880 case Idling:
1881 default:
1882 ShouldNotReachHere();
1883 break;
1884 }
1885 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d",
1886 p2i(Thread::current()), _collectorState);
1887 assert(_foregroundGCShouldWait, "block post-condition");
1888 }
1889
1890 // Should this be in gc_epilogue?
1891 collector_policy()->counters()->update_counters();
1892
1893 {
1894 // Clear _foregroundGCShouldWait and, in the event that the
1895 // foreground collector is waiting, notify it, before
1896 // returning.
1897 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1898 _foregroundGCShouldWait = false;
1899 if (_foregroundGCIsActive) {
1900 CGC_lock->notify();
1901 }
1902 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1903 "Possible deadlock");
1904 }
1905 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1906 p2i(Thread::current()), _collectorState);
1907 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)",
1908 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K);
1909 }
1910
1911 void CMSCollector::register_gc_start(GCCause::Cause cause) {
1912 _cms_start_registered = true;
1913 _gc_timer_cm->register_gc_start();
1914 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
1915 }
1916
1917 void CMSCollector::register_gc_end() {
1918 if (_cms_start_registered) {
1919 report_heap_summary(GCWhen::AfterGC);
1920
1921 _gc_timer_cm->register_gc_end();
1922 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1923 _cms_start_registered = false;
1924 }
1925 }
1926
1927 void CMSCollector::save_heap_summary() {
1928 CMSHeap* heap = CMSHeap::heap();
1929 _last_heap_summary = heap->create_heap_summary();
1930 _last_metaspace_summary = heap->create_metaspace_summary();
1931 }
1932
1933 void CMSCollector::report_heap_summary(GCWhen::Type when) {
1934 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
1935 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
1936 }
1937
1938 bool CMSCollector::waitForForegroundGC() {
1939 bool res = false;
1940 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1941 "CMS thread should have CMS token");
1942 // Block the foreground collector until the
1943 // background collectors decides whether to
1944 // yield.
1945 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1946 _foregroundGCShouldWait = true;
1947 if (_foregroundGCIsActive) {
1948 // The background collector yields to the
1949 // foreground collector and returns a value
1950 // indicating that it has yielded. The foreground
1951 // collector can proceed.
1952 res = true;
1953 _foregroundGCShouldWait = false;
1954 ConcurrentMarkSweepThread::clear_CMS_flag(
1955 ConcurrentMarkSweepThread::CMS_cms_has_token);
1956 ConcurrentMarkSweepThread::set_CMS_flag(
1957 ConcurrentMarkSweepThread::CMS_cms_wants_token);
1958 // Get a possibly blocked foreground thread going
1959 CGC_lock->notify();
1960 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
1961 p2i(Thread::current()), _collectorState);
1962 while (_foregroundGCIsActive) {
1963 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1964 }
1965 ConcurrentMarkSweepThread::set_CMS_flag(
1966 ConcurrentMarkSweepThread::CMS_cms_has_token);
1967 ConcurrentMarkSweepThread::clear_CMS_flag(
1968 ConcurrentMarkSweepThread::CMS_cms_wants_token);
1969 }
1970 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
1971 p2i(Thread::current()), _collectorState);
1972 return res;
1973 }
1974
1975 // Because of the need to lock the free lists and other structures in
1976 // the collector, common to all the generations that the collector is
1977 // collecting, we need the gc_prologues of individual CMS generations
1978 // delegate to their collector. It may have been simpler had the
1979 // current infrastructure allowed one to call a prologue on a
1980 // collector. In the absence of that we have the generation's
1981 // prologue delegate to the collector, which delegates back
1982 // some "local" work to a worker method in the individual generations
1983 // that it's responsible for collecting, while itself doing any
1984 // work common to all generations it's responsible for. A similar
1985 // comment applies to the gc_epilogue()'s.
1986 // The role of the variable _between_prologue_and_epilogue is to
1987 // enforce the invocation protocol.
1988 void CMSCollector::gc_prologue(bool full) {
1989 // Call gc_prologue_work() for the CMSGen
1990 // we are responsible for.
1991
1992 // The following locking discipline assumes that we are only called
1993 // when the world is stopped.
1994 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
1995
1996 // The CMSCollector prologue must call the gc_prologues for the
1997 // "generations" that it's responsible
1998 // for.
1999
2000 assert( Thread::current()->is_VM_thread()
2001 || ( CMSScavengeBeforeRemark
2002 && Thread::current()->is_ConcurrentGC_thread()),
2003 "Incorrect thread type for prologue execution");
2004
2005 if (_between_prologue_and_epilogue) {
2006 // We have already been invoked; this is a gc_prologue delegation
2007 // from yet another CMS generation that we are responsible for, just
2008 // ignore it since all relevant work has already been done.
2009 return;
2010 }
2011
2012 // set a bit saying prologue has been called; cleared in epilogue
2013 _between_prologue_and_epilogue = true;
2014 // Claim locks for common data structures, then call gc_prologue_work()
2015 // for each CMSGen.
2016
2017 getFreelistLocks(); // gets free list locks on constituent spaces
2018 bitMapLock()->lock_without_safepoint_check();
2019
2020 // Should call gc_prologue_work() for all cms gens we are responsible for
2021 bool duringMarking = _collectorState >= Marking
2022 && _collectorState < Sweeping;
2023
2024 // The young collections clear the modified oops state, which tells if
2025 // there are any modified oops in the class. The remark phase also needs
2026 // that information. Tell the young collection to save the union of all
2027 // modified klasses.
2028 if (duringMarking) {
2029 _ct->cld_rem_set()->set_accumulate_modified_oops(true);
2030 }
2031
2032 bool registerClosure = duringMarking;
2033
2034 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2035
2036 if (!full) {
2037 stats().record_gc0_begin();
2038 }
2039 }
2040
2041 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2042
2043 _capacity_at_prologue = capacity();
2044 _used_at_prologue = used();
2045
2046 // We enable promotion tracking so that card-scanning can recognize
2047 // which objects have been promoted during this GC and skip them.
2048 for (uint i = 0; i < ParallelGCThreads; i++) {
2049 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2050 }
2051
2052 // Delegate to CMScollector which knows how to coordinate between
2053 // this and any other CMS generations that it is responsible for
2054 // collecting.
2055 collector()->gc_prologue(full);
2056 }
2057
2058 // This is a "private" interface for use by this generation's CMSCollector.
2059 // Not to be called directly by any other entity (for instance,
2060 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2061 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2062 bool registerClosure, ModUnionClosure* modUnionClosure) {
2063 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2064 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2065 "Should be NULL");
2066 if (registerClosure) {
2067 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2068 }
2069 cmsSpace()->gc_prologue();
2070 // Clear stat counters
2071 NOT_PRODUCT(
2072 assert(_numObjectsPromoted == 0, "check");
2073 assert(_numWordsPromoted == 0, "check");
2074 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently",
2075 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2076 _numObjectsAllocated = 0;
2077 _numWordsAllocated = 0;
2078 )
2079 }
2080
2081 void CMSCollector::gc_epilogue(bool full) {
2082 // The following locking discipline assumes that we are only called
2083 // when the world is stopped.
2084 assert(SafepointSynchronize::is_at_safepoint(),
2085 "world is stopped assumption");
2086
2087 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2088 // if linear allocation blocks need to be appropriately marked to allow the
2089 // the blocks to be parsable. We also check here whether we need to nudge the
2090 // CMS collector thread to start a new cycle (if it's not already active).
2091 assert( Thread::current()->is_VM_thread()
2092 || ( CMSScavengeBeforeRemark
2093 && Thread::current()->is_ConcurrentGC_thread()),
2094 "Incorrect thread type for epilogue execution");
2095
2096 if (!_between_prologue_and_epilogue) {
2097 // We have already been invoked; this is a gc_epilogue delegation
2098 // from yet another CMS generation that we are responsible for, just
2099 // ignore it since all relevant work has already been done.
2100 return;
2101 }
2102 assert(haveFreelistLocks(), "must have freelist locks");
2103 assert_lock_strong(bitMapLock());
2104
2105 _ct->cld_rem_set()->set_accumulate_modified_oops(false);
2106
2107 _cmsGen->gc_epilogue_work(full);
2108
2109 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2110 // in case sampling was not already enabled, enable it
2111 _start_sampling = true;
2112 }
2113 // reset _eden_chunk_array so sampling starts afresh
2114 _eden_chunk_index = 0;
2115
2116 size_t cms_used = _cmsGen->cmsSpace()->used();
2117
2118 // update performance counters - this uses a special version of
2119 // update_counters() that allows the utilization to be passed as a
2120 // parameter, avoiding multiple calls to used().
2121 //
2122 _cmsGen->update_counters(cms_used);
2123
2124 bitMapLock()->unlock();
2125 releaseFreelistLocks();
2126
2127 if (!CleanChunkPoolAsync) {
2128 Chunk::clean_chunk_pool();
2129 }
2130
2131 set_did_compact(false);
2132 _between_prologue_and_epilogue = false; // ready for next cycle
2133 }
2134
2135 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2136 collector()->gc_epilogue(full);
2137
2138 // When using ParNew, promotion tracking should have already been
2139 // disabled. However, the prologue (which enables promotion
2140 // tracking) and epilogue are called irrespective of the type of
2141 // GC. So they will also be called before and after Full GCs, during
2142 // which promotion tracking will not be explicitly disabled. So,
2143 // it's safer to also disable it here too (to be symmetric with
2144 // enabling it in the prologue).
2145 for (uint i = 0; i < ParallelGCThreads; i++) {
2146 _par_gc_thread_states[i]->promo.stopTrackingPromotions();
2147 }
2148 }
2149
2150 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2151 assert(!incremental_collection_failed(), "Should have been cleared");
2152 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2153 cmsSpace()->gc_epilogue();
2154 // Print stat counters
2155 NOT_PRODUCT(
2156 assert(_numObjectsAllocated == 0, "check");
2157 assert(_numWordsAllocated == 0, "check");
2158 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
2159 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2160 _numObjectsPromoted = 0;
2161 _numWordsPromoted = 0;
2162 )
2163
2164 // Call down the chain in contiguous_available needs the freelistLock
2165 // so print this out before releasing the freeListLock.
2166 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available());
2167 }
2168
2169 #ifndef PRODUCT
2170 bool CMSCollector::have_cms_token() {
2171 Thread* thr = Thread::current();
2172 if (thr->is_VM_thread()) {
2173 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2174 } else if (thr->is_ConcurrentGC_thread()) {
2175 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2176 } else if (thr->is_GC_task_thread()) {
2177 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2178 ParGCRareEvent_lock->owned_by_self();
2179 }
2180 return false;
2181 }
2182
2183 // Check reachability of the given heap address in CMS generation,
2184 // treating all other generations as roots.
2185 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2186 // We could "guarantee" below, rather than assert, but I'll
2187 // leave these as "asserts" so that an adventurous debugger
2188 // could try this in the product build provided some subset of
2189 // the conditions were met, provided they were interested in the
2190 // results and knew that the computation below wouldn't interfere
2191 // with other concurrent computations mutating the structures
2192 // being read or written.
2193 assert(SafepointSynchronize::is_at_safepoint(),
2194 "Else mutations in object graph will make answer suspect");
2195 assert(have_cms_token(), "Should hold cms token");
2196 assert(haveFreelistLocks(), "must hold free list locks");
2197 assert_lock_strong(bitMapLock());
2198
2199 // Clear the marking bit map array before starting, but, just
2200 // for kicks, first report if the given address is already marked
2201 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr),
2202 _markBitMap.isMarked(addr) ? "" : " not");
2203
2204 if (verify_after_remark()) {
2205 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2206 bool result = verification_mark_bm()->isMarked(addr);
2207 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr),
2208 result ? "IS" : "is NOT");
2209 return result;
2210 } else {
2211 tty->print_cr("Could not compute result");
2212 return false;
2213 }
2214 }
2215 #endif
2216
2217 void
2218 CMSCollector::print_on_error(outputStream* st) {
2219 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2220 if (collector != NULL) {
2221 CMSBitMap* bitmap = &collector->_markBitMap;
2222 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2223 bitmap->print_on_error(st, " Bits: ");
2224
2225 st->cr();
2226
2227 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2228 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2229 mut_bitmap->print_on_error(st, " Bits: ");
2230 }
2231 }
2232
2233 ////////////////////////////////////////////////////////
2234 // CMS Verification Support
2235 ////////////////////////////////////////////////////////
2236 // Following the remark phase, the following invariant
2237 // should hold -- each object in the CMS heap which is
2238 // marked in markBitMap() should be marked in the verification_mark_bm().
2239
2240 class VerifyMarkedClosure: public BitMapClosure {
2241 CMSBitMap* _marks;
2242 bool _failed;
2243
2244 public:
2245 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2246
2247 bool do_bit(size_t offset) {
2248 HeapWord* addr = _marks->offsetToHeapWord(offset);
2249 if (!_marks->isMarked(addr)) {
2250 Log(gc, verify) log;
2251 ResourceMark rm;
2252 LogStream ls(log.error());
2253 oop(addr)->print_on(&ls);
2254 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
2255 _failed = true;
2256 }
2257 return true;
2258 }
2259
2260 bool failed() { return _failed; }
2261 };
2262
2263 bool CMSCollector::verify_after_remark() {
2264 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking.");
2265 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2266 static bool init = false;
2267
2268 assert(SafepointSynchronize::is_at_safepoint(),
2269 "Else mutations in object graph will make answer suspect");
2270 assert(have_cms_token(),
2271 "Else there may be mutual interference in use of "
2272 " verification data structures");
2273 assert(_collectorState > Marking && _collectorState <= Sweeping,
2274 "Else marking info checked here may be obsolete");
2275 assert(haveFreelistLocks(), "must hold free list locks");
2276 assert_lock_strong(bitMapLock());
2277
2278
2279 // Allocate marking bit map if not already allocated
2280 if (!init) { // first time
2281 if (!verification_mark_bm()->allocate(_span)) {
2282 return false;
2283 }
2284 init = true;
2285 }
2286
2287 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2288
2289 // Turn off refs discovery -- so we will be tracing through refs.
2290 // This is as intended, because by this time
2291 // GC must already have cleared any refs that need to be cleared,
2292 // and traced those that need to be marked; moreover,
2293 // the marking done here is not going to interfere in any
2294 // way with the marking information used by GC.
2295 NoRefDiscovery no_discovery(ref_processor());
2296
2297 #if COMPILER2_OR_JVMCI
2298 DerivedPointerTableDeactivate dpt_deact;
2299 #endif
2300
2301 // Clear any marks from a previous round
2302 verification_mark_bm()->clear_all();
2303 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2304 verify_work_stacks_empty();
2305
2306 CMSHeap* heap = CMSHeap::heap();
2307 heap->ensure_parsability(false); // fill TLABs, but no need to retire them
2308 // Update the saved marks which may affect the root scans.
2309 heap->save_marks();
2310
2311 if (CMSRemarkVerifyVariant == 1) {
2312 // In this first variant of verification, we complete
2313 // all marking, then check if the new marks-vector is
2314 // a subset of the CMS marks-vector.
2315 verify_after_remark_work_1();
2316 } else {
2317 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2");
2318 // In this second variant of verification, we flag an error
2319 // (i.e. an object reachable in the new marks-vector not reachable
2320 // in the CMS marks-vector) immediately, also indicating the
2321 // identify of an object (A) that references the unmarked object (B) --
2322 // presumably, a mutation to A failed to be picked up by preclean/remark?
2323 verify_after_remark_work_2();
2324 }
2325
2326 return true;
2327 }
2328
2329 void CMSCollector::verify_after_remark_work_1() {
2330 ResourceMark rm;
2331 HandleMark hm;
2332 CMSHeap* heap = CMSHeap::heap();
2333
2334 // Get a clear set of claim bits for the roots processing to work with.
2335 ClassLoaderDataGraph::clear_claimed_marks();
2336
2337 // Mark from roots one level into CMS
2338 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2339 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2340
2341 {
2342 StrongRootsScope srs(1);
2343
2344 heap->cms_process_roots(&srs,
2345 true, // young gen as roots
2346 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2347 should_unload_classes(),
2348 ¬Older,
2349 NULL);
2350 }
2351
2352 // Now mark from the roots
2353 MarkFromRootsClosure markFromRootsClosure(this, _span,
2354 verification_mark_bm(), verification_mark_stack(),
2355 false /* don't yield */, true /* verifying */);
2356 assert(_restart_addr == NULL, "Expected pre-condition");
2357 verification_mark_bm()->iterate(&markFromRootsClosure);
2358 while (_restart_addr != NULL) {
2359 // Deal with stack overflow: by restarting at the indicated
2360 // address.
2361 HeapWord* ra = _restart_addr;
2362 markFromRootsClosure.reset(ra);
2363 _restart_addr = NULL;
2364 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2365 }
2366 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2367 verify_work_stacks_empty();
2368
2369 // Marking completed -- now verify that each bit marked in
2370 // verification_mark_bm() is also marked in markBitMap(); flag all
2371 // errors by printing corresponding objects.
2372 VerifyMarkedClosure vcl(markBitMap());
2373 verification_mark_bm()->iterate(&vcl);
2374 if (vcl.failed()) {
2375 Log(gc, verify) log;
2376 log.error("Failed marking verification after remark");
2377 ResourceMark rm;
2378 LogStream ls(log.error());
2379 heap->print_on(&ls);
2380 fatal("CMS: failed marking verification after remark");
2381 }
2382 }
2383
2384 class VerifyCLDOopsCLDClosure : public CLDClosure {
2385 class VerifyCLDOopsClosure : public OopClosure {
2386 CMSBitMap* _bitmap;
2387 public:
2388 VerifyCLDOopsClosure(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 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2394 void do_cld(ClassLoaderData* cld) {
2395 cld->oops_do(&_oop_closure, false, false);
2396 }
2397 };
2398
2399 void CMSCollector::verify_after_remark_work_2() {
2400 ResourceMark rm;
2401 HandleMark hm;
2402 CMSHeap* heap = CMSHeap::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 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2413
2414 {
2415 StrongRootsScope srs(1);
2416
2417 heap->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 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm());
2442 ClassLoaderDataGraph::cld_do(&verify_cld_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, CMSHeap::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 CMSHeap* heap = CMSHeap::heap();
2848
2849 verify_work_stacks_empty();
2850 verify_overflow_empty();
2851
2852 heap->ensure_parsability(false); // fill TLABs, but no need to retire them
2853 // Update the saved marks which may affect the root scans.
2854 heap->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 COMPILER2_OR_JVMCI
2871 DerivedPointerTableDeactivate dpt_deact;
2872 #endif
2873 if (CMSParallelInitialMarkEnabled) {
2874 // The parallel version.
2875 WorkGang* workers = heap->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 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2895
2896 StrongRootsScope srs(1);
2897
2898 heap->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->cld_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 = Atomic::cmpxchg(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 = align_down(_restart_addr, CardTableModRefBS::card_size);
3228 }
3229
3230 size_t chunk_size = sp->marking_task_size();
3231 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3232 // Having claimed the nth task in this space,
3233 // compute the chunk that it corresponds to:
3234 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3235 aligned_start + (nth_task+1)*chunk_size);
3236 // Try and bump the global finger via a CAS;
3237 // note that we need to do the global finger bump
3238 // _before_ taking the intersection below, because
3239 // the task corresponding to that region will be
3240 // deemed done even if the used_region() expands
3241 // because of allocation -- as it almost certainly will
3242 // during start-up while the threads yield in the
3243 // closure below.
3244 HeapWord* finger = span.end();
3245 bump_global_finger(finger); // atomically
3246 // There are null tasks here corresponding to chunks
3247 // beyond the "top" address of the space.
3248 span = span.intersection(sp->used_region());
3249 if (!span.is_empty()) { // Non-null task
3250 HeapWord* prev_obj;
3251 assert(!span.contains(_restart_addr) || nth_task == 0,
3252 "Inconsistency");
3253 if (nth_task == 0) {
3254 // For the 0th task, we'll not need to compute a block_start.
3255 if (span.contains(_restart_addr)) {
3256 // In the case of a restart because of stack overflow,
3257 // we might additionally skip a chunk prefix.
3258 prev_obj = _restart_addr;
3259 } else {
3260 prev_obj = span.start();
3261 }
3262 } else {
3263 // We want to skip the first object because
3264 // the protocol is to scan any object in its entirety
3265 // that _starts_ in this span; a fortiori, any
3266 // object starting in an earlier span is scanned
3267 // as part of an earlier claimed task.
3268 // Below we use the "careful" version of block_start
3269 // so we do not try to navigate uninitialized objects.
3270 prev_obj = sp->block_start_careful(span.start());
3271 // Below we use a variant of block_size that uses the
3272 // Printezis bits to avoid waiting for allocated
3273 // objects to become initialized/parsable.
3274 while (prev_obj < span.start()) {
3275 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3276 if (sz > 0) {
3277 prev_obj += sz;
3278 } else {
3279 // In this case we may end up doing a bit of redundant
3280 // scanning, but that appears unavoidable, short of
3281 // locking the free list locks; see bug 6324141.
3282 break;
3283 }
3284 }
3285 }
3286 if (prev_obj < span.end()) {
3287 MemRegion my_span = MemRegion(prev_obj, span.end());
3288 // Do the marking work within a non-empty span --
3289 // the last argument to the constructor indicates whether the
3290 // iteration should be incremental with periodic yields.
3291 ParMarkFromRootsClosure cl(this, _collector, my_span,
3292 &_collector->_markBitMap,
3293 work_queue(i),
3294 &_collector->_markStack);
3295 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3296 } // else nothing to do for this task
3297 } // else nothing to do for this task
3298 }
3299 // We'd be tempted to assert here that since there are no
3300 // more tasks left to claim in this space, the global_finger
3301 // must exceed space->top() and a fortiori space->end(). However,
3302 // that would not quite be correct because the bumping of
3303 // global_finger occurs strictly after the claiming of a task,
3304 // so by the time we reach here the global finger may not yet
3305 // have been bumped up by the thread that claimed the last
3306 // task.
3307 pst->all_tasks_completed();
3308 }
3309
3310 class ParConcMarkingClosure: public MetadataAwareOopClosure {
3311 private:
3312 CMSCollector* _collector;
3313 CMSConcMarkingTask* _task;
3314 MemRegion _span;
3315 CMSBitMap* _bit_map;
3316 CMSMarkStack* _overflow_stack;
3317 OopTaskQueue* _work_queue;
3318 protected:
3319 DO_OOP_WORK_DEFN
3320 public:
3321 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3322 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3323 MetadataAwareOopClosure(collector->ref_processor()),
3324 _collector(collector),
3325 _task(task),
3326 _span(collector->_span),
3327 _work_queue(work_queue),
3328 _bit_map(bit_map),
3329 _overflow_stack(overflow_stack)
3330 { }
3331 virtual void do_oop(oop* p);
3332 virtual void do_oop(narrowOop* p);
3333
3334 void trim_queue(size_t max);
3335 void handle_stack_overflow(HeapWord* lost);
3336 void do_yield_check() {
3337 if (_task->should_yield()) {
3338 _task->yield();
3339 }
3340 }
3341 };
3342
3343 DO_OOP_WORK_IMPL(ParConcMarkingClosure)
3344
3345 // Grey object scanning during work stealing phase --
3346 // the salient assumption here is that any references
3347 // that are in these stolen objects being scanned must
3348 // already have been initialized (else they would not have
3349 // been published), so we do not need to check for
3350 // uninitialized objects before pushing here.
3351 void ParConcMarkingClosure::do_oop(oop obj) {
3352 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
3353 HeapWord* addr = (HeapWord*)obj;
3354 // Check if oop points into the CMS generation
3355 // and is not marked
3356 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3357 // a white object ...
3358 // If we manage to "claim" the object, by being the
3359 // first thread to mark it, then we push it on our
3360 // marking stack
3361 if (_bit_map->par_mark(addr)) { // ... now grey
3362 // push on work queue (grey set)
3363 bool simulate_overflow = false;
3364 NOT_PRODUCT(
3365 if (CMSMarkStackOverflowALot &&
3366 _collector->simulate_overflow()) {
3367 // simulate a stack overflow
3368 simulate_overflow = true;
3369 }
3370 )
3371 if (simulate_overflow ||
3372 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3373 // stack overflow
3374 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
3375 // We cannot assert that the overflow stack is full because
3376 // it may have been emptied since.
3377 assert(simulate_overflow ||
3378 _work_queue->size() == _work_queue->max_elems(),
3379 "Else push should have succeeded");
3380 handle_stack_overflow(addr);
3381 }
3382 } // Else, some other thread got there first
3383 do_yield_check();
3384 }
3385 }
3386
3387 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); }
3388 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); }
3389
3390 void ParConcMarkingClosure::trim_queue(size_t max) {
3391 while (_work_queue->size() > max) {
3392 oop new_oop;
3393 if (_work_queue->pop_local(new_oop)) {
3394 assert(oopDesc::is_oop(new_oop), "Should be an oop");
3395 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3396 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3397 new_oop->oop_iterate(this); // do_oop() above
3398 do_yield_check();
3399 }
3400 }
3401 }
3402
3403 // Upon stack overflow, we discard (part of) the stack,
3404 // remembering the least address amongst those discarded
3405 // in CMSCollector's _restart_address.
3406 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3407 // We need to do this under a mutex to prevent other
3408 // workers from interfering with the work done below.
3409 MutexLockerEx ml(_overflow_stack->par_lock(),
3410 Mutex::_no_safepoint_check_flag);
3411 // Remember the least grey address discarded
3412 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3413 _collector->lower_restart_addr(ra);
3414 _overflow_stack->reset(); // discard stack contents
3415 _overflow_stack->expand(); // expand the stack if possible
3416 }
3417
3418
3419 void CMSConcMarkingTask::do_work_steal(int i) {
3420 OopTaskQueue* work_q = work_queue(i);
3421 oop obj_to_scan;
3422 CMSBitMap* bm = &(_collector->_markBitMap);
3423 CMSMarkStack* ovflw = &(_collector->_markStack);
3424 int* seed = _collector->hash_seed(i);
3425 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3426 while (true) {
3427 cl.trim_queue(0);
3428 assert(work_q->size() == 0, "Should have been emptied above");
3429 if (get_work_from_overflow_stack(ovflw, work_q)) {
3430 // Can't assert below because the work obtained from the
3431 // overflow stack may already have been stolen from us.
3432 // assert(work_q->size() > 0, "Work from overflow stack");
3433 continue;
3434 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3435 assert(oopDesc::is_oop(obj_to_scan), "Should be an oop");
3436 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3437 obj_to_scan->oop_iterate(&cl);
3438 } else if (terminator()->offer_termination(&_term_term)) {
3439 assert(work_q->size() == 0, "Impossible!");
3440 break;
3441 } else if (yielding() || should_yield()) {
3442 yield();
3443 }
3444 }
3445 }
3446
3447 // This is run by the CMS (coordinator) thread.
3448 void CMSConcMarkingTask::coordinator_yield() {
3449 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3450 "CMS thread should hold CMS token");
3451 // First give up the locks, then yield, then re-lock
3452 // We should probably use a constructor/destructor idiom to
3453 // do this unlock/lock or modify the MutexUnlocker class to
3454 // serve our purpose. XXX
3455 assert_lock_strong(_bit_map_lock);
3456 _bit_map_lock->unlock();
3457 ConcurrentMarkSweepThread::desynchronize(true);
3458 _collector->stopTimer();
3459 _collector->incrementYields();
3460
3461 // It is possible for whichever thread initiated the yield request
3462 // not to get a chance to wake up and take the bitmap lock between
3463 // this thread releasing it and reacquiring it. So, while the
3464 // should_yield() flag is on, let's sleep for a bit to give the
3465 // other thread a chance to wake up. The limit imposed on the number
3466 // of iterations is defensive, to avoid any unforseen circumstances
3467 // putting us into an infinite loop. Since it's always been this
3468 // (coordinator_yield()) method that was observed to cause the
3469 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3470 // which is by default non-zero. For the other seven methods that
3471 // also perform the yield operation, as are using a different
3472 // parameter (CMSYieldSleepCount) which is by default zero. This way we
3473 // can enable the sleeping for those methods too, if necessary.
3474 // See 6442774.
3475 //
3476 // We really need to reconsider the synchronization between the GC
3477 // thread and the yield-requesting threads in the future and we
3478 // should really use wait/notify, which is the recommended
3479 // way of doing this type of interaction. Additionally, we should
3480 // consolidate the eight methods that do the yield operation and they
3481 // are almost identical into one for better maintainability and
3482 // readability. See 6445193.
3483 //
3484 // Tony 2006.06.29
3485 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3486 ConcurrentMarkSweepThread::should_yield() &&
3487 !CMSCollector::foregroundGCIsActive(); ++i) {
3488 os::sleep(Thread::current(), 1, false);
3489 }
3490
3491 ConcurrentMarkSweepThread::synchronize(true);
3492 _bit_map_lock->lock_without_safepoint_check();
3493 _collector->startTimer();
3494 }
3495
3496 bool CMSCollector::do_marking_mt() {
3497 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3498 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3499 conc_workers()->active_workers(),
3500 Threads::number_of_non_daemon_threads());
3501 num_workers = conc_workers()->update_active_workers(num_workers);
3502 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers());
3503
3504 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
3505
3506 CMSConcMarkingTask tsk(this,
3507 cms_space,
3508 conc_workers(),
3509 task_queues());
3510
3511 // Since the actual number of workers we get may be different
3512 // from the number we requested above, do we need to do anything different
3513 // below? In particular, may be we need to subclass the SequantialSubTasksDone
3514 // class?? XXX
3515 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3516
3517 // Refs discovery is already non-atomic.
3518 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3519 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3520 conc_workers()->start_task(&tsk);
3521 while (tsk.yielded()) {
3522 tsk.coordinator_yield();
3523 conc_workers()->continue_task(&tsk);
3524 }
3525 // If the task was aborted, _restart_addr will be non-NULL
3526 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3527 while (_restart_addr != NULL) {
3528 // XXX For now we do not make use of ABORTED state and have not
3529 // yet implemented the right abort semantics (even in the original
3530 // single-threaded CMS case). That needs some more investigation
3531 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3532 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3533 // If _restart_addr is non-NULL, a marking stack overflow
3534 // occurred; we need to do a fresh marking iteration from the
3535 // indicated restart address.
3536 if (_foregroundGCIsActive) {
3537 // We may be running into repeated stack overflows, having
3538 // reached the limit of the stack size, while making very
3539 // slow forward progress. It may be best to bail out and
3540 // let the foreground collector do its job.
3541 // Clear _restart_addr, so that foreground GC
3542 // works from scratch. This avoids the headache of
3543 // a "rescan" which would otherwise be needed because
3544 // of the dirty mod union table & card table.
3545 _restart_addr = NULL;
3546 return false;
3547 }
3548 // Adjust the task to restart from _restart_addr
3549 tsk.reset(_restart_addr);
3550 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3551 _restart_addr);
3552 _restart_addr = NULL;
3553 // Get the workers going again
3554 conc_workers()->start_task(&tsk);
3555 while (tsk.yielded()) {
3556 tsk.coordinator_yield();
3557 conc_workers()->continue_task(&tsk);
3558 }
3559 }
3560 assert(tsk.completed(), "Inconsistency");
3561 assert(tsk.result() == true, "Inconsistency");
3562 return true;
3563 }
3564
3565 bool CMSCollector::do_marking_st() {
3566 ResourceMark rm;
3567 HandleMark hm;
3568
3569 // Temporarily make refs discovery single threaded (non-MT)
3570 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3571 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3572 &_markStack, CMSYield);
3573 // the last argument to iterate indicates whether the iteration
3574 // should be incremental with periodic yields.
3575 _markBitMap.iterate(&markFromRootsClosure);
3576 // If _restart_addr is non-NULL, a marking stack overflow
3577 // occurred; we need to do a fresh iteration from the
3578 // indicated restart address.
3579 while (_restart_addr != NULL) {
3580 if (_foregroundGCIsActive) {
3581 // We may be running into repeated stack overflows, having
3582 // reached the limit of the stack size, while making very
3583 // slow forward progress. It may be best to bail out and
3584 // let the foreground collector do its job.
3585 // Clear _restart_addr, so that foreground GC
3586 // works from scratch. This avoids the headache of
3587 // a "rescan" which would otherwise be needed because
3588 // of the dirty mod union table & card table.
3589 _restart_addr = NULL;
3590 return false; // indicating failure to complete marking
3591 }
3592 // Deal with stack overflow:
3593 // we restart marking from _restart_addr
3594 HeapWord* ra = _restart_addr;
3595 markFromRootsClosure.reset(ra);
3596 _restart_addr = NULL;
3597 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3598 }
3599 return true;
3600 }
3601
3602 void CMSCollector::preclean() {
3603 check_correct_thread_executing();
3604 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3605 verify_work_stacks_empty();
3606 verify_overflow_empty();
3607 _abort_preclean = false;
3608 if (CMSPrecleaningEnabled) {
3609 if (!CMSEdenChunksRecordAlways) {
3610 _eden_chunk_index = 0;
3611 }
3612 size_t used = get_eden_used();
3613 size_t capacity = get_eden_capacity();
3614 // Don't start sampling unless we will get sufficiently
3615 // many samples.
3616 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100)
3617 * CMSScheduleRemarkEdenPenetration)) {
3618 _start_sampling = true;
3619 } else {
3620 _start_sampling = false;
3621 }
3622 GCTraceCPUTime tcpu;
3623 CMSPhaseAccounting pa(this, "Concurrent Preclean");
3624 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3625 }
3626 CMSTokenSync x(true); // is cms thread
3627 if (CMSPrecleaningEnabled) {
3628 sample_eden();
3629 _collectorState = AbortablePreclean;
3630 } else {
3631 _collectorState = FinalMarking;
3632 }
3633 verify_work_stacks_empty();
3634 verify_overflow_empty();
3635 }
3636
3637 // Try and schedule the remark such that young gen
3638 // occupancy is CMSScheduleRemarkEdenPenetration %.
3639 void CMSCollector::abortable_preclean() {
3640 check_correct_thread_executing();
3641 assert(CMSPrecleaningEnabled, "Inconsistent control state");
3642 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3643
3644 // If Eden's current occupancy is below this threshold,
3645 // immediately schedule the remark; else preclean
3646 // past the next scavenge in an effort to
3647 // schedule the pause as described above. By choosing
3648 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3649 // we will never do an actual abortable preclean cycle.
3650 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3651 GCTraceCPUTime tcpu;
3652 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean");
3653 // We need more smarts in the abortable preclean
3654 // loop below to deal with cases where allocation
3655 // in young gen is very very slow, and our precleaning
3656 // is running a losing race against a horde of
3657 // mutators intent on flooding us with CMS updates
3658 // (dirty cards).
3659 // One, admittedly dumb, strategy is to give up
3660 // after a certain number of abortable precleaning loops
3661 // or after a certain maximum time. We want to make
3662 // this smarter in the next iteration.
3663 // XXX FIX ME!!! YSR
3664 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3665 while (!(should_abort_preclean() ||
3666 ConcurrentMarkSweepThread::cmst()->should_terminate())) {
3667 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3668 cumworkdone += workdone;
3669 loops++;
3670 // Voluntarily terminate abortable preclean phase if we have
3671 // been at it for too long.
3672 if ((CMSMaxAbortablePrecleanLoops != 0) &&
3673 loops >= CMSMaxAbortablePrecleanLoops) {
3674 log_debug(gc)(" CMS: abort preclean due to loops ");
3675 break;
3676 }
3677 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3678 log_debug(gc)(" CMS: abort preclean due to time ");
3679 break;
3680 }
3681 // If we are doing little work each iteration, we should
3682 // take a short break.
3683 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3684 // Sleep for some time, waiting for work to accumulate
3685 stopTimer();
3686 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3687 startTimer();
3688 waited++;
3689 }
3690 }
3691 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3692 loops, waited, cumworkdone);
3693 }
3694 CMSTokenSync x(true); // is cms thread
3695 if (_collectorState != Idling) {
3696 assert(_collectorState == AbortablePreclean,
3697 "Spontaneous state transition?");
3698 _collectorState = FinalMarking;
3699 } // Else, a foreground collection completed this CMS cycle.
3700 return;
3701 }
3702
3703 // Respond to an Eden sampling opportunity
3704 void CMSCollector::sample_eden() {
3705 // Make sure a young gc cannot sneak in between our
3706 // reading and recording of a sample.
3707 assert(Thread::current()->is_ConcurrentGC_thread(),
3708 "Only the cms thread may collect Eden samples");
3709 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3710 "Should collect samples while holding CMS token");
3711 if (!_start_sampling) {
3712 return;
3713 }
3714 // When CMSEdenChunksRecordAlways is true, the eden chunk array
3715 // is populated by the young generation.
3716 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3717 if (_eden_chunk_index < _eden_chunk_capacity) {
3718 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
3719 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3720 "Unexpected state of Eden");
3721 // We'd like to check that what we just sampled is an oop-start address;
3722 // however, we cannot do that here since the object may not yet have been
3723 // initialized. So we'll instead do the check when we _use_ this sample
3724 // later.
3725 if (_eden_chunk_index == 0 ||
3726 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3727 _eden_chunk_array[_eden_chunk_index-1])
3728 >= CMSSamplingGrain)) {
3729 _eden_chunk_index++; // commit sample
3730 }
3731 }
3732 }
3733 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3734 size_t used = get_eden_used();
3735 size_t capacity = get_eden_capacity();
3736 assert(used <= capacity, "Unexpected state of Eden");
3737 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3738 _abort_preclean = true;
3739 }
3740 }
3741 }
3742
3743
3744 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3745 assert(_collectorState == Precleaning ||
3746 _collectorState == AbortablePreclean, "incorrect state");
3747 ResourceMark rm;
3748 HandleMark hm;
3749
3750 // Precleaning is currently not MT but the reference processor
3751 // may be set for MT. Disable it temporarily here.
3752 ReferenceProcessor* rp = ref_processor();
3753 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3754
3755 // Do one pass of scrubbing the discovered reference lists
3756 // to remove any reference objects with strongly-reachable
3757 // referents.
3758 if (clean_refs) {
3759 CMSPrecleanRefsYieldClosure yield_cl(this);
3760 assert(rp->span().equals(_span), "Spans should be equal");
3761 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3762 &_markStack, true /* preclean */);
3763 CMSDrainMarkingStackClosure complete_trace(this,
3764 _span, &_markBitMap, &_markStack,
3765 &keep_alive, true /* preclean */);
3766
3767 // We don't want this step to interfere with a young
3768 // collection because we don't want to take CPU
3769 // or memory bandwidth away from the young GC threads
3770 // (which may be as many as there are CPUs).
3771 // Note that we don't need to protect ourselves from
3772 // interference with mutators because they can't
3773 // manipulate the discovered reference lists nor affect
3774 // the computed reachability of the referents, the
3775 // only properties manipulated by the precleaning
3776 // of these reference lists.
3777 stopTimer();
3778 CMSTokenSyncWithLocks x(true /* is cms thread */,
3779 bitMapLock());
3780 startTimer();
3781 sample_eden();
3782
3783 // The following will yield to allow foreground
3784 // collection to proceed promptly. XXX YSR:
3785 // The code in this method may need further
3786 // tweaking for better performance and some restructuring
3787 // for cleaner interfaces.
3788 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3789 rp->preclean_discovered_references(
3790 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3791 gc_timer);
3792 }
3793
3794 if (clean_survivor) { // preclean the active survivor space(s)
3795 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3796 &_markBitMap, &_modUnionTable,
3797 &_markStack, true /* precleaning phase */);
3798 stopTimer();
3799 CMSTokenSyncWithLocks ts(true /* is cms thread */,
3800 bitMapLock());
3801 startTimer();
3802 unsigned int before_count =
3803 CMSHeap::heap()->total_collections();
3804 SurvivorSpacePrecleanClosure
3805 sss_cl(this, _span, &_markBitMap, &_markStack,
3806 &pam_cl, before_count, CMSYield);
3807 _young_gen->from()->object_iterate_careful(&sss_cl);
3808 _young_gen->to()->object_iterate_careful(&sss_cl);
3809 }
3810 MarkRefsIntoAndScanClosure
3811 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3812 &_markStack, this, CMSYield,
3813 true /* precleaning phase */);
3814 // CAUTION: The following closure has persistent state that may need to
3815 // be reset upon a decrease in the sequence of addresses it
3816 // processes.
3817 ScanMarkedObjectsAgainCarefullyClosure
3818 smoac_cl(this, _span,
3819 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3820
3821 // Preclean dirty cards in ModUnionTable and CardTable using
3822 // appropriate convergence criterion;
3823 // repeat CMSPrecleanIter times unless we find that
3824 // we are losing.
3825 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3826 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3827 "Bad convergence multiplier");
3828 assert(CMSPrecleanThreshold >= 100,
3829 "Unreasonably low CMSPrecleanThreshold");
3830
3831 size_t numIter, cumNumCards, lastNumCards, curNumCards;
3832 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3833 numIter < CMSPrecleanIter;
3834 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3835 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
3836 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3837 // Either there are very few dirty cards, so re-mark
3838 // pause will be small anyway, or our pre-cleaning isn't
3839 // that much faster than the rate at which cards are being
3840 // dirtied, so we might as well stop and re-mark since
3841 // precleaning won't improve our re-mark time by much.
3842 if (curNumCards <= CMSPrecleanThreshold ||
3843 (numIter > 0 &&
3844 (curNumCards * CMSPrecleanDenominator >
3845 lastNumCards * CMSPrecleanNumerator))) {
3846 numIter++;
3847 cumNumCards += curNumCards;
3848 break;
3849 }
3850 }
3851
3852 preclean_cld(&mrias_cl, _cmsGen->freelistLock());
3853
3854 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
3855 cumNumCards += curNumCards;
3856 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
3857 curNumCards, cumNumCards, numIter);
3858 return cumNumCards; // as a measure of useful work done
3859 }
3860
3861 // PRECLEANING NOTES:
3862 // Precleaning involves:
3863 // . reading the bits of the modUnionTable and clearing the set bits.
3864 // . For the cards corresponding to the set bits, we scan the
3865 // objects on those cards. This means we need the free_list_lock
3866 // so that we can safely iterate over the CMS space when scanning
3867 // for oops.
3868 // . When we scan the objects, we'll be both reading and setting
3869 // marks in the marking bit map, so we'll need the marking bit map.
3870 // . For protecting _collector_state transitions, we take the CGC_lock.
3871 // Note that any races in the reading of of card table entries by the
3872 // CMS thread on the one hand and the clearing of those entries by the
3873 // VM thread or the setting of those entries by the mutator threads on the
3874 // other are quite benign. However, for efficiency it makes sense to keep
3875 // the VM thread from racing with the CMS thread while the latter is
3876 // dirty card info to the modUnionTable. We therefore also use the
3877 // CGC_lock to protect the reading of the card table and the mod union
3878 // table by the CM thread.
3879 // . We run concurrently with mutator updates, so scanning
3880 // needs to be done carefully -- we should not try to scan
3881 // potentially uninitialized objects.
3882 //
3883 // Locking strategy: While holding the CGC_lock, we scan over and
3884 // reset a maximal dirty range of the mod union / card tables, then lock
3885 // the free_list_lock and bitmap lock to do a full marking, then
3886 // release these locks; and repeat the cycle. This allows for a
3887 // certain amount of fairness in the sharing of these locks between
3888 // the CMS collector on the one hand, and the VM thread and the
3889 // mutators on the other.
3890
3891 // NOTE: preclean_mod_union_table() and preclean_card_table()
3892 // further below are largely identical; if you need to modify
3893 // one of these methods, please check the other method too.
3894
3895 size_t CMSCollector::preclean_mod_union_table(
3896 ConcurrentMarkSweepGeneration* old_gen,
3897 ScanMarkedObjectsAgainCarefullyClosure* cl) {
3898 verify_work_stacks_empty();
3899 verify_overflow_empty();
3900
3901 // strategy: starting with the first card, accumulate contiguous
3902 // ranges of dirty cards; clear these cards, then scan the region
3903 // covered by these cards.
3904
3905 // Since all of the MUT is committed ahead, we can just use
3906 // that, in case the generations expand while we are precleaning.
3907 // It might also be fine to just use the committed part of the
3908 // generation, but we might potentially miss cards when the
3909 // generation is rapidly expanding while we are in the midst
3910 // of precleaning.
3911 HeapWord* startAddr = old_gen->reserved().start();
3912 HeapWord* endAddr = old_gen->reserved().end();
3913
3914 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
3915
3916 size_t numDirtyCards, cumNumDirtyCards;
3917 HeapWord *nextAddr, *lastAddr;
3918 for (cumNumDirtyCards = numDirtyCards = 0,
3919 nextAddr = lastAddr = startAddr;
3920 nextAddr < endAddr;
3921 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
3922
3923 ResourceMark rm;
3924 HandleMark hm;
3925
3926 MemRegion dirtyRegion;
3927 {
3928 stopTimer();
3929 // Potential yield point
3930 CMSTokenSync ts(true);
3931 startTimer();
3932 sample_eden();
3933 // Get dirty region starting at nextOffset (inclusive),
3934 // simultaneously clearing it.
3935 dirtyRegion =
3936 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
3937 assert(dirtyRegion.start() >= nextAddr,
3938 "returned region inconsistent?");
3939 }
3940 // Remember where the next search should begin.
3941 // The returned region (if non-empty) is a right open interval,
3942 // so lastOffset is obtained from the right end of that
3943 // interval.
3944 lastAddr = dirtyRegion.end();
3945 // Should do something more transparent and less hacky XXX
3946 numDirtyCards =
3947 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
3948
3949 // We'll scan the cards in the dirty region (with periodic
3950 // yields for foreground GC as needed).
3951 if (!dirtyRegion.is_empty()) {
3952 assert(numDirtyCards > 0, "consistency check");
3953 HeapWord* stop_point = NULL;
3954 stopTimer();
3955 // Potential yield point
3956 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(),
3957 bitMapLock());
3958 startTimer();
3959 {
3960 verify_work_stacks_empty();
3961 verify_overflow_empty();
3962 sample_eden();
3963 stop_point =
3964 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
3965 }
3966 if (stop_point != NULL) {
3967 // The careful iteration stopped early either because it found an
3968 // uninitialized object, or because we were in the midst of an
3969 // "abortable preclean", which should now be aborted. Redirty
3970 // the bits corresponding to the partially-scanned or unscanned
3971 // cards. We'll either restart at the next block boundary or
3972 // abort the preclean.
3973 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
3974 "Should only be AbortablePreclean.");
3975 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
3976 if (should_abort_preclean()) {
3977 break; // out of preclean loop
3978 } else {
3979 // Compute the next address at which preclean should pick up;
3980 // might need bitMapLock in order to read P-bits.
3981 lastAddr = next_card_start_after_block(stop_point);
3982 }
3983 }
3984 } else {
3985 assert(lastAddr == endAddr, "consistency check");
3986 assert(numDirtyCards == 0, "consistency check");
3987 break;
3988 }
3989 }
3990 verify_work_stacks_empty();
3991 verify_overflow_empty();
3992 return cumNumDirtyCards;
3993 }
3994
3995 // NOTE: preclean_mod_union_table() above and preclean_card_table()
3996 // below are largely identical; if you need to modify
3997 // one of these methods, please check the other method too.
3998
3999 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen,
4000 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4001 // strategy: it's similar to precleamModUnionTable above, in that
4002 // we accumulate contiguous ranges of dirty cards, mark these cards
4003 // precleaned, then scan the region covered by these cards.
4004 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high());
4005 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low());
4006
4007 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
4008
4009 size_t numDirtyCards, cumNumDirtyCards;
4010 HeapWord *lastAddr, *nextAddr;
4011
4012 for (cumNumDirtyCards = numDirtyCards = 0,
4013 nextAddr = lastAddr = startAddr;
4014 nextAddr < endAddr;
4015 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4016
4017 ResourceMark rm;
4018 HandleMark hm;
4019
4020 MemRegion dirtyRegion;
4021 {
4022 // See comments in "Precleaning notes" above on why we
4023 // do this locking. XXX Could the locking overheads be
4024 // too high when dirty cards are sparse? [I don't think so.]
4025 stopTimer();
4026 CMSTokenSync x(true); // is cms thread
4027 startTimer();
4028 sample_eden();
4029 // Get and clear dirty region from card table
4030 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4031 MemRegion(nextAddr, endAddr),
4032 true,
4033 CardTableModRefBS::precleaned_card_val());
4034
4035 assert(dirtyRegion.start() >= nextAddr,
4036 "returned region inconsistent?");
4037 }
4038 lastAddr = dirtyRegion.end();
4039 numDirtyCards =
4040 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4041
4042 if (!dirtyRegion.is_empty()) {
4043 stopTimer();
4044 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock());
4045 startTimer();
4046 sample_eden();
4047 verify_work_stacks_empty();
4048 verify_overflow_empty();
4049 HeapWord* stop_point =
4050 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4051 if (stop_point != NULL) {
4052 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4053 "Should only be AbortablePreclean.");
4054 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4055 if (should_abort_preclean()) {
4056 break; // out of preclean loop
4057 } else {
4058 // Compute the next address at which preclean should pick up.
4059 lastAddr = next_card_start_after_block(stop_point);
4060 }
4061 }
4062 } else {
4063 break;
4064 }
4065 }
4066 verify_work_stacks_empty();
4067 verify_overflow_empty();
4068 return cumNumDirtyCards;
4069 }
4070
4071 class PrecleanCLDClosure : public CLDClosure {
4072 MetadataAwareOopsInGenClosure* _cm_closure;
4073 public:
4074 PrecleanCLDClosure(MetadataAwareOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {}
4075 void do_cld(ClassLoaderData* cld) {
4076 if (cld->has_accumulated_modified_oops()) {
4077 cld->clear_accumulated_modified_oops();
4078
4079 _cm_closure->do_cld(cld);
4080 }
4081 }
4082 };
4083
4084 // The freelist lock is needed to prevent asserts, is it really needed?
4085 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4086
4087 cl->set_freelistLock(freelistLock);
4088
4089 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4090
4091 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4092 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4093 PrecleanCLDClosure preclean_closure(cl);
4094 ClassLoaderDataGraph::cld_do(&preclean_closure);
4095
4096 verify_work_stacks_empty();
4097 verify_overflow_empty();
4098 }
4099
4100 void CMSCollector::checkpointRootsFinal() {
4101 assert(_collectorState == FinalMarking, "incorrect state transition?");
4102 check_correct_thread_executing();
4103 // world is stopped at this checkpoint
4104 assert(SafepointSynchronize::is_at_safepoint(),
4105 "world should be stopped");
4106 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
4107
4108 verify_work_stacks_empty();
4109 verify_overflow_empty();
4110
4111 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)",
4112 _young_gen->used() / K, _young_gen->capacity() / K);
4113 {
4114 if (CMSScavengeBeforeRemark) {
4115 CMSHeap* heap = CMSHeap::heap();
4116 // Temporarily set flag to false, GCH->do_collection will
4117 // expect it to be false and set to true
4118 FlagSetting fl(heap->_is_gc_active, false);
4119
4120 heap->do_collection(true, // full (i.e. force, see below)
4121 false, // !clear_all_soft_refs
4122 0, // size
4123 false, // is_tlab
4124 GenCollectedHeap::YoungGen // type
4125 );
4126 }
4127 FreelistLocker x(this);
4128 MutexLockerEx y(bitMapLock(),
4129 Mutex::_no_safepoint_check_flag);
4130 checkpointRootsFinalWork();
4131 }
4132 verify_work_stacks_empty();
4133 verify_overflow_empty();
4134 }
4135
4136 void CMSCollector::checkpointRootsFinalWork() {
4137 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm);
4138
4139 assert(haveFreelistLocks(), "must have free list locks");
4140 assert_lock_strong(bitMapLock());
4141
4142 ResourceMark rm;
4143 HandleMark hm;
4144
4145 CMSHeap* heap = CMSHeap::heap();
4146
4147 if (should_unload_classes()) {
4148 CodeCache::gc_prologue();
4149 }
4150 assert(haveFreelistLocks(), "must have free list locks");
4151 assert_lock_strong(bitMapLock());
4152
4153 // We might assume that we need not fill TLAB's when
4154 // CMSScavengeBeforeRemark is set, because we may have just done
4155 // a scavenge which would have filled all TLAB's -- and besides
4156 // Eden would be empty. This however may not always be the case --
4157 // for instance although we asked for a scavenge, it may not have
4158 // happened because of a JNI critical section. We probably need
4159 // a policy for deciding whether we can in that case wait until
4160 // the critical section releases and then do the remark following
4161 // the scavenge, and skip it here. In the absence of that policy,
4162 // or of an indication of whether the scavenge did indeed occur,
4163 // we cannot rely on TLAB's having been filled and must do
4164 // so here just in case a scavenge did not happen.
4165 heap->ensure_parsability(false); // fill TLAB's, but no need to retire them
4166 // Update the saved marks which may affect the root scans.
4167 heap->save_marks();
4168
4169 print_eden_and_survivor_chunk_arrays();
4170
4171 {
4172 #if COMPILER2_OR_JVMCI
4173 DerivedPointerTableDeactivate dpt_deact;
4174 #endif
4175
4176 // Note on the role of the mod union table:
4177 // Since the marker in "markFromRoots" marks concurrently with
4178 // mutators, it is possible for some reachable objects not to have been
4179 // scanned. For instance, an only reference to an object A was
4180 // placed in object B after the marker scanned B. Unless B is rescanned,
4181 // A would be collected. Such updates to references in marked objects
4182 // are detected via the mod union table which is the set of all cards
4183 // dirtied since the first checkpoint in this GC cycle and prior to
4184 // the most recent young generation GC, minus those cleaned up by the
4185 // concurrent precleaning.
4186 if (CMSParallelRemarkEnabled) {
4187 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm);
4188 do_remark_parallel();
4189 } else {
4190 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm);
4191 do_remark_non_parallel();
4192 }
4193 }
4194 verify_work_stacks_empty();
4195 verify_overflow_empty();
4196
4197 {
4198 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm);
4199 refProcessingWork();
4200 }
4201 verify_work_stacks_empty();
4202 verify_overflow_empty();
4203
4204 if (should_unload_classes()) {
4205 CodeCache::gc_epilogue();
4206 }
4207 JvmtiExport::gc_epilogue();
4208
4209 // If we encountered any (marking stack / work queue) overflow
4210 // events during the current CMS cycle, take appropriate
4211 // remedial measures, where possible, so as to try and avoid
4212 // recurrence of that condition.
4213 assert(_markStack.isEmpty(), "No grey objects");
4214 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4215 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4216 if (ser_ovflw > 0) {
4217 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")",
4218 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4219 _markStack.expand();
4220 _ser_pmc_remark_ovflw = 0;
4221 _ser_pmc_preclean_ovflw = 0;
4222 _ser_kac_preclean_ovflw = 0;
4223 _ser_kac_ovflw = 0;
4224 }
4225 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4226 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")",
4227 _par_pmc_remark_ovflw, _par_kac_ovflw);
4228 _par_pmc_remark_ovflw = 0;
4229 _par_kac_ovflw = 0;
4230 }
4231 if (_markStack._hit_limit > 0) {
4232 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")",
4233 _markStack._hit_limit);
4234 }
4235 if (_markStack._failed_double > 0) {
4236 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT,
4237 _markStack._failed_double, _markStack.capacity());
4238 }
4239 _markStack._hit_limit = 0;
4240 _markStack._failed_double = 0;
4241
4242 if ((VerifyAfterGC || VerifyDuringGC) &&
4243 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) {
4244 verify_after_remark();
4245 }
4246
4247 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4248
4249 // Change under the freelistLocks.
4250 _collectorState = Sweeping;
4251 // Call isAllClear() under bitMapLock
4252 assert(_modUnionTable.isAllClear(),
4253 "Should be clear by end of the final marking");
4254 assert(_ct->cld_rem_set()->mod_union_is_clear(),
4255 "Should be clear by end of the final marking");
4256 }
4257
4258 void CMSParInitialMarkTask::work(uint worker_id) {
4259 elapsedTimer _timer;
4260 ResourceMark rm;
4261 HandleMark hm;
4262
4263 // ---------- scan from roots --------------
4264 _timer.start();
4265 CMSHeap* heap = CMSHeap::heap();
4266 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4267
4268 // ---------- young gen roots --------------
4269 {
4270 work_on_young_gen_roots(&par_mri_cl);
4271 _timer.stop();
4272 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4273 }
4274
4275 // ---------- remaining roots --------------
4276 _timer.reset();
4277 _timer.start();
4278
4279 CLDToOopClosure cld_closure(&par_mri_cl, true);
4280
4281 heap->cms_process_roots(_strong_roots_scope,
4282 false, // yg was scanned above
4283 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4284 _collector->should_unload_classes(),
4285 &par_mri_cl,
4286 &cld_closure);
4287 assert(_collector->should_unload_classes()
4288 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4289 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4290 _timer.stop();
4291 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4292 }
4293
4294 // Parallel remark task
4295 class CMSParRemarkTask: public CMSParMarkTask {
4296 CompactibleFreeListSpace* _cms_space;
4297
4298 // The per-thread work queues, available here for stealing.
4299 OopTaskQueueSet* _task_queues;
4300 ParallelTaskTerminator _term;
4301 StrongRootsScope* _strong_roots_scope;
4302
4303 public:
4304 // A value of 0 passed to n_workers will cause the number of
4305 // workers to be taken from the active workers in the work gang.
4306 CMSParRemarkTask(CMSCollector* collector,
4307 CompactibleFreeListSpace* cms_space,
4308 uint n_workers, WorkGang* workers,
4309 OopTaskQueueSet* task_queues,
4310 StrongRootsScope* strong_roots_scope):
4311 CMSParMarkTask("Rescan roots and grey objects in parallel",
4312 collector, n_workers),
4313 _cms_space(cms_space),
4314 _task_queues(task_queues),
4315 _term(n_workers, task_queues),
4316 _strong_roots_scope(strong_roots_scope) { }
4317
4318 OopTaskQueueSet* task_queues() { return _task_queues; }
4319
4320 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4321
4322 ParallelTaskTerminator* terminator() { return &_term; }
4323 uint n_workers() { return _n_workers; }
4324
4325 void work(uint worker_id);
4326
4327 private:
4328 // ... of dirty cards in old space
4329 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4330 ParMarkRefsIntoAndScanClosure* cl);
4331
4332 // ... work stealing for the above
4333 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed);
4334 };
4335
4336 class RemarkCLDClosure : public CLDClosure {
4337 CLDToOopClosure _cm_closure;
4338 public:
4339 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure) {}
4340 void do_cld(ClassLoaderData* cld) {
4341 // Check if we have modified any oops in the CLD during the concurrent marking.
4342 if (cld->has_accumulated_modified_oops()) {
4343 cld->clear_accumulated_modified_oops();
4344
4345 // We could have transfered the current modified marks to the accumulated marks,
4346 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4347 } else if (cld->has_modified_oops()) {
4348 // Don't clear anything, this info is needed by the next young collection.
4349 } else {
4350 // No modified oops in the ClassLoaderData.
4351 return;
4352 }
4353
4354 // The klass has modified fields, need to scan the klass.
4355 _cm_closure.do_cld(cld);
4356 }
4357 };
4358
4359 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) {
4360 ParNewGeneration* young_gen = _collector->_young_gen;
4361 ContiguousSpace* eden_space = young_gen->eden();
4362 ContiguousSpace* from_space = young_gen->from();
4363 ContiguousSpace* to_space = young_gen->to();
4364
4365 HeapWord** eca = _collector->_eden_chunk_array;
4366 size_t ect = _collector->_eden_chunk_index;
4367 HeapWord** sca = _collector->_survivor_chunk_array;
4368 size_t sct = _collector->_survivor_chunk_index;
4369
4370 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4371 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4372
4373 do_young_space_rescan(cl, to_space, NULL, 0);
4374 do_young_space_rescan(cl, from_space, sca, sct);
4375 do_young_space_rescan(cl, eden_space, eca, ect);
4376 }
4377
4378 // work_queue(i) is passed to the closure
4379 // ParMarkRefsIntoAndScanClosure. The "i" parameter
4380 // also is passed to do_dirty_card_rescan_tasks() and to
4381 // do_work_steal() to select the i-th task_queue.
4382
4383 void CMSParRemarkTask::work(uint worker_id) {
4384 elapsedTimer _timer;
4385 ResourceMark rm;
4386 HandleMark hm;
4387
4388 // ---------- rescan from roots --------------
4389 _timer.start();
4390 CMSHeap* heap = CMSHeap::heap();
4391 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4392 _collector->_span, _collector->ref_processor(),
4393 &(_collector->_markBitMap),
4394 work_queue(worker_id));
4395
4396 // Rescan young gen roots first since these are likely
4397 // coarsely partitioned and may, on that account, constitute
4398 // the critical path; thus, it's best to start off that
4399 // work first.
4400 // ---------- young gen roots --------------
4401 {
4402 work_on_young_gen_roots(&par_mrias_cl);
4403 _timer.stop();
4404 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4405 }
4406
4407 // ---------- remaining roots --------------
4408 _timer.reset();
4409 _timer.start();
4410 heap->cms_process_roots(_strong_roots_scope,
4411 false, // yg was scanned above
4412 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4413 _collector->should_unload_classes(),
4414 &par_mrias_cl,
4415 NULL); // The dirty klasses will be handled below
4416
4417 assert(_collector->should_unload_classes()
4418 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4419 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4420 _timer.stop();
4421 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4422
4423 // ---------- unhandled CLD scanning ----------
4424 if (worker_id == 0) { // Single threaded at the moment.
4425 _timer.reset();
4426 _timer.start();
4427
4428 // Scan all new class loader data objects and new dependencies that were
4429 // introduced during concurrent marking.
4430 ResourceMark rm;
4431 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4432 for (int i = 0; i < array->length(); i++) {
4433 par_mrias_cl.do_cld_nv(array->at(i));
4434 }
4435
4436 // We don't need to keep track of new CLDs anymore.
4437 ClassLoaderDataGraph::remember_new_clds(false);
4438
4439 _timer.stop();
4440 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4441 }
4442
4443 // We might have added oops to ClassLoaderData::_handles during the
4444 // concurrent marking phase. These oops do not always point to newly allocated objects
4445 // that are guaranteed to be kept alive. Hence,
4446 // we do have to revisit the _handles block during the remark phase.
4447
4448 // ---------- dirty CLD scanning ----------
4449 if (worker_id == 0) { // Single threaded at the moment.
4450 _timer.reset();
4451 _timer.start();
4452
4453 // Scan all classes that was dirtied during the concurrent marking phase.
4454 RemarkCLDClosure remark_closure(&par_mrias_cl);
4455 ClassLoaderDataGraph::cld_do(&remark_closure);
4456
4457 _timer.stop();
4458 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4459 }
4460
4461
4462 // ---------- rescan dirty cards ------------
4463 _timer.reset();
4464 _timer.start();
4465
4466 // Do the rescan tasks for each of the two spaces
4467 // (cms_space) in turn.
4468 // "worker_id" is passed to select the task_queue for "worker_id"
4469 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4470 _timer.stop();
4471 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4472
4473 // ---------- steal work from other threads ...
4474 // ---------- ... and drain overflow list.
4475 _timer.reset();
4476 _timer.start();
4477 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4478 _timer.stop();
4479 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4480 }
4481
4482 void
4483 CMSParMarkTask::do_young_space_rescan(
4484 OopsInGenClosure* cl, ContiguousSpace* space,
4485 HeapWord** chunk_array, size_t chunk_top) {
4486 // Until all tasks completed:
4487 // . claim an unclaimed task
4488 // . compute region boundaries corresponding to task claimed
4489 // using chunk_array
4490 // . par_oop_iterate(cl) over that region
4491
4492 ResourceMark rm;
4493 HandleMark hm;
4494
4495 SequentialSubTasksDone* pst = space->par_seq_tasks();
4496
4497 uint nth_task = 0;
4498 uint n_tasks = pst->n_tasks();
4499
4500 if (n_tasks > 0) {
4501 assert(pst->valid(), "Uninitialized use?");
4502 HeapWord *start, *end;
4503 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4504 // We claimed task # nth_task; compute its boundaries.
4505 if (chunk_top == 0) { // no samples were taken
4506 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4507 start = space->bottom();
4508 end = space->top();
4509 } else if (nth_task == 0) {
4510 start = space->bottom();
4511 end = chunk_array[nth_task];
4512 } else if (nth_task < (uint)chunk_top) {
4513 assert(nth_task >= 1, "Control point invariant");
4514 start = chunk_array[nth_task - 1];
4515 end = chunk_array[nth_task];
4516 } else {
4517 assert(nth_task == (uint)chunk_top, "Control point invariant");
4518 start = chunk_array[chunk_top - 1];
4519 end = space->top();
4520 }
4521 MemRegion mr(start, end);
4522 // Verify that mr is in space
4523 assert(mr.is_empty() || space->used_region().contains(mr),
4524 "Should be in space");
4525 // Verify that "start" is an object boundary
4526 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())),
4527 "Should be an oop");
4528 space->par_oop_iterate(mr, cl);
4529 }
4530 pst->all_tasks_completed();
4531 }
4532 }
4533
4534 void
4535 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4536 CompactibleFreeListSpace* sp, int i,
4537 ParMarkRefsIntoAndScanClosure* cl) {
4538 // Until all tasks completed:
4539 // . claim an unclaimed task
4540 // . compute region boundaries corresponding to task claimed
4541 // . transfer dirty bits ct->mut for that region
4542 // . apply rescanclosure to dirty mut bits for that region
4543
4544 ResourceMark rm;
4545 HandleMark hm;
4546
4547 OopTaskQueue* work_q = work_queue(i);
4548 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4549 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4550 // CAUTION: This closure has state that persists across calls to
4551 // the work method dirty_range_iterate_clear() in that it has
4552 // embedded in it a (subtype of) UpwardsObjectClosure. The
4553 // use of that state in the embedded UpwardsObjectClosure instance
4554 // assumes that the cards are always iterated (even if in parallel
4555 // by several threads) in monotonically increasing order per each
4556 // thread. This is true of the implementation below which picks
4557 // card ranges (chunks) in monotonically increasing order globally
4558 // and, a-fortiori, in monotonically increasing order per thread
4559 // (the latter order being a subsequence of the former).
4560 // If the work code below is ever reorganized into a more chaotic
4561 // work-partitioning form than the current "sequential tasks"
4562 // paradigm, the use of that persistent state will have to be
4563 // revisited and modified appropriately. See also related
4564 // bug 4756801 work on which should examine this code to make
4565 // sure that the changes there do not run counter to the
4566 // assumptions made here and necessary for correctness and
4567 // efficiency. Note also that this code might yield inefficient
4568 // behavior in the case of very large objects that span one or
4569 // more work chunks. Such objects would potentially be scanned
4570 // several times redundantly. Work on 4756801 should try and
4571 // address that performance anomaly if at all possible. XXX
4572 MemRegion full_span = _collector->_span;
4573 CMSBitMap* bm = &(_collector->_markBitMap); // shared
4574 MarkFromDirtyCardsClosure
4575 greyRescanClosure(_collector, full_span, // entire span of interest
4576 sp, bm, work_q, cl);
4577
4578 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4579 assert(pst->valid(), "Uninitialized use?");
4580 uint nth_task = 0;
4581 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
4582 MemRegion span = sp->used_region();
4583 HeapWord* start_addr = span.start();
4584 HeapWord* end_addr = align_up(span.end(), alignment);
4585 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4586 assert(is_aligned(start_addr, alignment), "Check alignment");
4587 assert(is_aligned(chunk_size, alignment), "Check alignment");
4588
4589 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4590 // Having claimed the nth_task, compute corresponding mem-region,
4591 // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4592 // The alignment restriction ensures that we do not need any
4593 // synchronization with other gang-workers while setting or
4594 // clearing bits in thus chunk of the MUT.
4595 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4596 start_addr + (nth_task+1)*chunk_size);
4597 // The last chunk's end might be way beyond end of the
4598 // used region. In that case pull back appropriately.
4599 if (this_span.end() > end_addr) {
4600 this_span.set_end(end_addr);
4601 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4602 }
4603 // Iterate over the dirty cards covering this chunk, marking them
4604 // precleaned, and setting the corresponding bits in the mod union
4605 // table. Since we have been careful to partition at Card and MUT-word
4606 // boundaries no synchronization is needed between parallel threads.
4607 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
4608 &modUnionClosure);
4609
4610 // Having transferred these marks into the modUnionTable,
4611 // rescan the marked objects on the dirty cards in the modUnionTable.
4612 // Even if this is at a synchronous collection, the initial marking
4613 // may have been done during an asynchronous collection so there
4614 // may be dirty bits in the mod-union table.
4615 _collector->_modUnionTable.dirty_range_iterate_clear(
4616 this_span, &greyRescanClosure);
4617 _collector->_modUnionTable.verifyNoOneBitsInRange(
4618 this_span.start(),
4619 this_span.end());
4620 }
4621 pst->all_tasks_completed(); // declare that i am done
4622 }
4623
4624 // . see if we can share work_queues with ParNew? XXX
4625 void
4626 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl,
4627 int* seed) {
4628 OopTaskQueue* work_q = work_queue(i);
4629 NOT_PRODUCT(int num_steals = 0;)
4630 oop obj_to_scan;
4631 CMSBitMap* bm = &(_collector->_markBitMap);
4632
4633 while (true) {
4634 // Completely finish any left over work from (an) earlier round(s)
4635 cl->trim_queue(0);
4636 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4637 (size_t)ParGCDesiredObjsFromOverflowList);
4638 // Now check if there's any work in the overflow list
4639 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4640 // only affects the number of attempts made to get work from the
4641 // overflow list and does not affect the number of workers. Just
4642 // pass ParallelGCThreads so this behavior is unchanged.
4643 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4644 work_q,
4645 ParallelGCThreads)) {
4646 // found something in global overflow list;
4647 // not yet ready to go stealing work from others.
4648 // We'd like to assert(work_q->size() != 0, ...)
4649 // because we just took work from the overflow list,
4650 // but of course we can't since all of that could have
4651 // been already stolen from us.
4652 // "He giveth and He taketh away."
4653 continue;
4654 }
4655 // Verify that we have no work before we resort to stealing
4656 assert(work_q->size() == 0, "Have work, shouldn't steal");
4657 // Try to steal from other queues that have work
4658 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4659 NOT_PRODUCT(num_steals++;)
4660 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!");
4661 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4662 // Do scanning work
4663 obj_to_scan->oop_iterate(cl);
4664 // Loop around, finish this work, and try to steal some more
4665 } else if (terminator()->offer_termination()) {
4666 break; // nirvana from the infinite cycle
4667 }
4668 }
4669 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
4670 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4671 "Else our work is not yet done");
4672 }
4673
4674 // Record object boundaries in _eden_chunk_array by sampling the eden
4675 // top in the slow-path eden object allocation code path and record
4676 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4677 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4678 // sampling in sample_eden() that activates during the part of the
4679 // preclean phase.
4680 void CMSCollector::sample_eden_chunk() {
4681 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4682 if (_eden_chunk_lock->try_lock()) {
4683 // Record a sample. This is the critical section. The contents
4684 // of the _eden_chunk_array have to be non-decreasing in the
4685 // address order.
4686 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4687 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4688 "Unexpected state of Eden");
4689 if (_eden_chunk_index == 0 ||
4690 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4691 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4692 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4693 _eden_chunk_index++; // commit sample
4694 }
4695 _eden_chunk_lock->unlock();
4696 }
4697 }
4698 }
4699
4700 // Return a thread-local PLAB recording array, as appropriate.
4701 void* CMSCollector::get_data_recorder(int thr_num) {
4702 if (_survivor_plab_array != NULL &&
4703 (CMSPLABRecordAlways ||
4704 (_collectorState > Marking && _collectorState < FinalMarking))) {
4705 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4706 ChunkArray* ca = &_survivor_plab_array[thr_num];
4707 ca->reset(); // clear it so that fresh data is recorded
4708 return (void*) ca;
4709 } else {
4710 return NULL;
4711 }
4712 }
4713
4714 // Reset all the thread-local PLAB recording arrays
4715 void CMSCollector::reset_survivor_plab_arrays() {
4716 for (uint i = 0; i < ParallelGCThreads; i++) {
4717 _survivor_plab_array[i].reset();
4718 }
4719 }
4720
4721 // Merge the per-thread plab arrays into the global survivor chunk
4722 // array which will provide the partitioning of the survivor space
4723 // for CMS initial scan and rescan.
4724 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4725 int no_of_gc_threads) {
4726 assert(_survivor_plab_array != NULL, "Error");
4727 assert(_survivor_chunk_array != NULL, "Error");
4728 assert(_collectorState == FinalMarking ||
4729 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4730 for (int j = 0; j < no_of_gc_threads; j++) {
4731 _cursor[j] = 0;
4732 }
4733 HeapWord* top = surv->top();
4734 size_t i;
4735 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
4736 HeapWord* min_val = top; // Higher than any PLAB address
4737 uint min_tid = 0; // position of min_val this round
4738 for (int j = 0; j < no_of_gc_threads; j++) {
4739 ChunkArray* cur_sca = &_survivor_plab_array[j];
4740 if (_cursor[j] == cur_sca->end()) {
4741 continue;
4742 }
4743 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4744 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4745 assert(surv->used_region().contains(cur_val), "Out of bounds value");
4746 if (cur_val < min_val) {
4747 min_tid = j;
4748 min_val = cur_val;
4749 } else {
4750 assert(cur_val < top, "All recorded addresses should be less");
4751 }
4752 }
4753 // At this point min_val and min_tid are respectively
4754 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4755 // and the thread (j) that witnesses that address.
4756 // We record this address in the _survivor_chunk_array[i]
4757 // and increment _cursor[min_tid] prior to the next round i.
4758 if (min_val == top) {
4759 break;
4760 }
4761 _survivor_chunk_array[i] = min_val;
4762 _cursor[min_tid]++;
4763 }
4764 // We are all done; record the size of the _survivor_chunk_array
4765 _survivor_chunk_index = i; // exclusive: [0, i)
4766 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4767 // Verify that we used up all the recorded entries
4768 #ifdef ASSERT
4769 size_t total = 0;
4770 for (int j = 0; j < no_of_gc_threads; j++) {
4771 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4772 total += _cursor[j];
4773 }
4774 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
4775 // Check that the merged array is in sorted order
4776 if (total > 0) {
4777 for (size_t i = 0; i < total - 1; i++) {
4778 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
4779 i, p2i(_survivor_chunk_array[i]));
4780 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
4781 "Not sorted");
4782 }
4783 }
4784 #endif // ASSERT
4785 }
4786
4787 // Set up the space's par_seq_tasks structure for work claiming
4788 // for parallel initial scan and rescan of young gen.
4789 // See ParRescanTask where this is currently used.
4790 void
4791 CMSCollector::
4792 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
4793 assert(n_threads > 0, "Unexpected n_threads argument");
4794
4795 // Eden space
4796 if (!_young_gen->eden()->is_empty()) {
4797 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
4798 assert(!pst->valid(), "Clobbering existing data?");
4799 // Each valid entry in [0, _eden_chunk_index) represents a task.
4800 size_t n_tasks = _eden_chunk_index + 1;
4801 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
4802 // Sets the condition for completion of the subtask (how many threads
4803 // need to finish in order to be done).
4804 pst->set_n_threads(n_threads);
4805 pst->set_n_tasks((int)n_tasks);
4806 }
4807
4808 // Merge the survivor plab arrays into _survivor_chunk_array
4809 if (_survivor_plab_array != NULL) {
4810 merge_survivor_plab_arrays(_young_gen->from(), n_threads);
4811 } else {
4812 assert(_survivor_chunk_index == 0, "Error");
4813 }
4814
4815 // To space
4816 {
4817 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
4818 assert(!pst->valid(), "Clobbering existing data?");
4819 // Sets the condition for completion of the subtask (how many threads
4820 // need to finish in order to be done).
4821 pst->set_n_threads(n_threads);
4822 pst->set_n_tasks(1);
4823 assert(pst->valid(), "Error");
4824 }
4825
4826 // From space
4827 {
4828 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
4829 assert(!pst->valid(), "Clobbering existing data?");
4830 size_t n_tasks = _survivor_chunk_index + 1;
4831 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
4832 // Sets the condition for completion of the subtask (how many threads
4833 // need to finish in order to be done).
4834 pst->set_n_threads(n_threads);
4835 pst->set_n_tasks((int)n_tasks);
4836 assert(pst->valid(), "Error");
4837 }
4838 }
4839
4840 // Parallel version of remark
4841 void CMSCollector::do_remark_parallel() {
4842 CMSHeap* heap = CMSHeap::heap();
4843 WorkGang* workers = heap->workers();
4844 assert(workers != NULL, "Need parallel worker threads.");
4845 // Choose to use the number of GC workers most recently set
4846 // into "active_workers".
4847 uint n_workers = workers->active_workers();
4848
4849 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4850
4851 StrongRootsScope srs(n_workers);
4852
4853 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
4854
4855 // We won't be iterating over the cards in the card table updating
4856 // the younger_gen cards, so we shouldn't call the following else
4857 // the verification code as well as subsequent younger_refs_iterate
4858 // code would get confused. XXX
4859 // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
4860
4861 // The young gen rescan work will not be done as part of
4862 // process_roots (which currently doesn't know how to
4863 // parallelize such a scan), but rather will be broken up into
4864 // a set of parallel tasks (via the sampling that the [abortable]
4865 // preclean phase did of eden, plus the [two] tasks of
4866 // scanning the [two] survivor spaces. Further fine-grain
4867 // parallelization of the scanning of the survivor spaces
4868 // themselves, and of precleaning of the young gen itself
4869 // is deferred to the future.
4870 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
4871
4872 // The dirty card rescan work is broken up into a "sequence"
4873 // of parallel tasks (per constituent space) that are dynamically
4874 // claimed by the parallel threads.
4875 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
4876
4877 // It turns out that even when we're using 1 thread, doing the work in a
4878 // separate thread causes wide variance in run times. We can't help this
4879 // in the multi-threaded case, but we special-case n=1 here to get
4880 // repeatable measurements of the 1-thread overhead of the parallel code.
4881 if (n_workers > 1) {
4882 // Make refs discovery MT-safe, if it isn't already: it may not
4883 // necessarily be so, since it's possible that we are doing
4884 // ST marking.
4885 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
4886 workers->run_task(&tsk);
4887 } else {
4888 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4889 tsk.work(0);
4890 }
4891
4892 // restore, single-threaded for now, any preserved marks
4893 // as a result of work_q overflow
4894 restore_preserved_marks_if_any();
4895 }
4896
4897 // Non-parallel version of remark
4898 void CMSCollector::do_remark_non_parallel() {
4899 ResourceMark rm;
4900 HandleMark hm;
4901 CMSHeap* heap = CMSHeap::heap();
4902 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4903
4904 MarkRefsIntoAndScanClosure
4905 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
4906 &_markStack, this,
4907 false /* should_yield */, false /* not precleaning */);
4908 MarkFromDirtyCardsClosure
4909 markFromDirtyCardsClosure(this, _span,
4910 NULL, // space is set further below
4911 &_markBitMap, &_markStack, &mrias_cl);
4912 {
4913 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm);
4914 // Iterate over the dirty cards, setting the corresponding bits in the
4915 // mod union table.
4916 {
4917 ModUnionClosure modUnionClosure(&_modUnionTable);
4918 _ct->ct_bs()->dirty_card_iterate(
4919 _cmsGen->used_region(),
4920 &modUnionClosure);
4921 }
4922 // Having transferred these marks into the modUnionTable, we just need
4923 // to rescan the marked objects on the dirty cards in the modUnionTable.
4924 // The initial marking may have been done during an asynchronous
4925 // collection so there may be dirty bits in the mod-union table.
4926 const int alignment =
4927 CardTableModRefBS::card_size * BitsPerWord;
4928 {
4929 // ... First handle dirty cards in CMS gen
4930 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
4931 MemRegion ur = _cmsGen->used_region();
4932 HeapWord* lb = ur.start();
4933 HeapWord* ub = align_up(ur.end(), alignment);
4934 MemRegion cms_span(lb, ub);
4935 _modUnionTable.dirty_range_iterate_clear(cms_span,
4936 &markFromDirtyCardsClosure);
4937 verify_work_stacks_empty();
4938 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards());
4939 }
4940 }
4941 if (VerifyDuringGC &&
4942 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) {
4943 HandleMark hm; // Discard invalid handles created during verification
4944 Universe::verify();
4945 }
4946 {
4947 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm);
4948
4949 verify_work_stacks_empty();
4950
4951 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
4952 StrongRootsScope srs(1);
4953
4954 heap->cms_process_roots(&srs,
4955 true, // young gen as roots
4956 GenCollectedHeap::ScanningOption(roots_scanning_options()),
4957 should_unload_classes(),
4958 &mrias_cl,
4959 NULL); // The dirty klasses will be handled below
4960
4961 assert(should_unload_classes()
4962 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4963 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4964 }
4965
4966 {
4967 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm);
4968
4969 verify_work_stacks_empty();
4970
4971 // Scan all class loader data objects that might have been introduced
4972 // during concurrent marking.
4973 ResourceMark rm;
4974 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4975 for (int i = 0; i < array->length(); i++) {
4976 mrias_cl.do_cld_nv(array->at(i));
4977 }
4978
4979 // We don't need to keep track of new CLDs anymore.
4980 ClassLoaderDataGraph::remember_new_clds(false);
4981
4982 verify_work_stacks_empty();
4983 }
4984
4985 // We might have added oops to ClassLoaderData::_handles during the
4986 // concurrent marking phase. These oops do not point to newly allocated objects
4987 // that are guaranteed to be kept alive. Hence,
4988 // we do have to revisit the _handles block during the remark phase.
4989 {
4990 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm);
4991
4992 verify_work_stacks_empty();
4993
4994 RemarkCLDClosure remark_closure(&mrias_cl);
4995 ClassLoaderDataGraph::cld_do(&remark_closure);
4996
4997 verify_work_stacks_empty();
4998 }
4999
5000 verify_work_stacks_empty();
5001 // Restore evacuated mark words, if any, used for overflow list links
5002 restore_preserved_marks_if_any();
5003
5004 verify_overflow_empty();
5005 }
5006
5007 ////////////////////////////////////////////////////////
5008 // Parallel Reference Processing Task Proxy Class
5009 ////////////////////////////////////////////////////////
5010 class AbstractGangTaskWOopQueues : public AbstractGangTask {
5011 OopTaskQueueSet* _queues;
5012 ParallelTaskTerminator _terminator;
5013 public:
5014 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) :
5015 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {}
5016 ParallelTaskTerminator* terminator() { return &_terminator; }
5017 OopTaskQueueSet* queues() { return _queues; }
5018 };
5019
5020 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5021 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5022 CMSCollector* _collector;
5023 CMSBitMap* _mark_bit_map;
5024 const MemRegion _span;
5025 ProcessTask& _task;
5026
5027 public:
5028 CMSRefProcTaskProxy(ProcessTask& task,
5029 CMSCollector* collector,
5030 const MemRegion& span,
5031 CMSBitMap* mark_bit_map,
5032 AbstractWorkGang* workers,
5033 OopTaskQueueSet* task_queues):
5034 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5035 task_queues,
5036 workers->active_workers()),
5037 _task(task),
5038 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5039 {
5040 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5041 "Inconsistency in _span");
5042 }
5043
5044 OopTaskQueueSet* task_queues() { return queues(); }
5045
5046 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5047
5048 void do_work_steal(int i,
5049 CMSParDrainMarkingStackClosure* drain,
5050 CMSParKeepAliveClosure* keep_alive,
5051 int* seed);
5052
5053 virtual void work(uint worker_id);
5054 };
5055
5056 void CMSRefProcTaskProxy::work(uint worker_id) {
5057 ResourceMark rm;
5058 HandleMark hm;
5059 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5060 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5061 _mark_bit_map,
5062 work_queue(worker_id));
5063 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5064 _mark_bit_map,
5065 work_queue(worker_id));
5066 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5067 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5068 if (_task.marks_oops_alive()) {
5069 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5070 _collector->hash_seed(worker_id));
5071 }
5072 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5073 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5074 }
5075
5076 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5077 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5078 EnqueueTask& _task;
5079
5080 public:
5081 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5082 : AbstractGangTask("Enqueue reference objects in parallel"),
5083 _task(task)
5084 { }
5085
5086 virtual void work(uint worker_id)
5087 {
5088 _task.work(worker_id);
5089 }
5090 };
5091
5092 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5093 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5094 _span(span),
5095 _bit_map(bit_map),
5096 _work_queue(work_queue),
5097 _mark_and_push(collector, span, bit_map, work_queue),
5098 _low_water_mark(MIN2((work_queue->max_elems()/4),
5099 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5100 { }
5101
5102 // . see if we can share work_queues with ParNew? XXX
5103 void CMSRefProcTaskProxy::do_work_steal(int i,
5104 CMSParDrainMarkingStackClosure* drain,
5105 CMSParKeepAliveClosure* keep_alive,
5106 int* seed) {
5107 OopTaskQueue* work_q = work_queue(i);
5108 NOT_PRODUCT(int num_steals = 0;)
5109 oop obj_to_scan;
5110
5111 while (true) {
5112 // Completely finish any left over work from (an) earlier round(s)
5113 drain->trim_queue(0);
5114 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5115 (size_t)ParGCDesiredObjsFromOverflowList);
5116 // Now check if there's any work in the overflow list
5117 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5118 // only affects the number of attempts made to get work from the
5119 // overflow list and does not affect the number of workers. Just
5120 // pass ParallelGCThreads so this behavior is unchanged.
5121 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5122 work_q,
5123 ParallelGCThreads)) {
5124 // Found something in global overflow list;
5125 // not yet ready to go stealing work from others.
5126 // We'd like to assert(work_q->size() != 0, ...)
5127 // because we just took work from the overflow list,
5128 // but of course we can't, since all of that might have
5129 // been already stolen from us.
5130 continue;
5131 }
5132 // Verify that we have no work before we resort to stealing
5133 assert(work_q->size() == 0, "Have work, shouldn't steal");
5134 // Try to steal from other queues that have work
5135 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5136 NOT_PRODUCT(num_steals++;)
5137 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!");
5138 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5139 // Do scanning work
5140 obj_to_scan->oop_iterate(keep_alive);
5141 // Loop around, finish this work, and try to steal some more
5142 } else if (terminator()->offer_termination()) {
5143 break; // nirvana from the infinite cycle
5144 }
5145 }
5146 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
5147 }
5148
5149 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5150 {
5151 CMSHeap* heap = CMSHeap::heap();
5152 WorkGang* workers = heap->workers();
5153 assert(workers != NULL, "Need parallel worker threads.");
5154 CMSRefProcTaskProxy rp_task(task, &_collector,
5155 _collector.ref_processor()->span(),
5156 _collector.markBitMap(),
5157 workers, _collector.task_queues());
5158 workers->run_task(&rp_task);
5159 }
5160
5161 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5162 {
5163
5164 CMSHeap* heap = CMSHeap::heap();
5165 WorkGang* workers = heap->workers();
5166 assert(workers != NULL, "Need parallel worker threads.");
5167 CMSRefEnqueueTaskProxy enq_task(task);
5168 workers->run_task(&enq_task);
5169 }
5170
5171 void CMSCollector::refProcessingWork() {
5172 ResourceMark rm;
5173 HandleMark hm;
5174
5175 ReferenceProcessor* rp = ref_processor();
5176 assert(rp->span().equals(_span), "Spans should be equal");
5177 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5178 // Process weak references.
5179 rp->setup_policy(false);
5180 verify_work_stacks_empty();
5181
5182 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->num_q());
5183 {
5184 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm);
5185
5186 // Setup keep_alive and complete closures.
5187 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5188 &_markStack, false /* !preclean */);
5189 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5190 _span, &_markBitMap, &_markStack,
5191 &cmsKeepAliveClosure, false /* !preclean */);
5192
5193 ReferenceProcessorStats stats;
5194 if (rp->processing_is_mt()) {
5195 // Set the degree of MT here. If the discovery is done MT, there
5196 // may have been a different number of threads doing the discovery
5197 // and a different number of discovered lists may have Ref objects.
5198 // That is OK as long as the Reference lists are balanced (see
5199 // balance_all_queues() and balance_queues()).
5200 CMSHeap* heap = CMSHeap::heap();
5201 uint active_workers = ParallelGCThreads;
5202 WorkGang* workers = heap->workers();
5203 if (workers != NULL) {
5204 active_workers = workers->active_workers();
5205 // The expectation is that active_workers will have already
5206 // been set to a reasonable value. If it has not been set,
5207 // investigate.
5208 assert(active_workers > 0, "Should have been set during scavenge");
5209 }
5210 rp->set_active_mt_degree(active_workers);
5211 CMSRefProcTaskExecutor task_executor(*this);
5212 stats = rp->process_discovered_references(&_is_alive_closure,
5213 &cmsKeepAliveClosure,
5214 &cmsDrainMarkingStackClosure,
5215 &task_executor,
5216 &pt);
5217 } else {
5218 stats = rp->process_discovered_references(&_is_alive_closure,
5219 &cmsKeepAliveClosure,
5220 &cmsDrainMarkingStackClosure,
5221 NULL,
5222 &pt);
5223 }
5224 _gc_tracer_cm->report_gc_reference_stats(stats);
5225 pt.print_all_references();
5226 }
5227
5228 // This is the point where the entire marking should have completed.
5229 verify_work_stacks_empty();
5230
5231 {
5232 GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm);
5233 WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl);
5234 }
5235
5236 if (should_unload_classes()) {
5237 {
5238 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm);
5239
5240 // Unload classes and purge the SystemDictionary.
5241 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure, _gc_timer_cm);
5242
5243 // Unload nmethods.
5244 CodeCache::do_unloading(&_is_alive_closure, purged_class);
5245
5246 // Prune dead klasses from subklass/sibling/implementor lists.
5247 Klass::clean_weak_klass_links(&_is_alive_closure);
5248 }
5249
5250 {
5251 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm);
5252 // Clean up unreferenced symbols in symbol table.
5253 SymbolTable::unlink();
5254 }
5255
5256 {
5257 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm);
5258 // Delete entries for dead interned strings.
5259 StringTable::unlink(&_is_alive_closure);
5260 }
5261 }
5262
5263 // Restore any preserved marks as a result of mark stack or
5264 // work queue overflow
5265 restore_preserved_marks_if_any(); // done single-threaded for now
5266
5267 rp->set_enqueuing_is_done(true);
5268 if (rp->processing_is_mt()) {
5269 rp->balance_all_queues();
5270 CMSRefProcTaskExecutor task_executor(*this);
5271 rp->enqueue_discovered_references(&task_executor, &pt);
5272 } else {
5273 rp->enqueue_discovered_references(NULL, &pt);
5274 }
5275 rp->verify_no_references_recorded();
5276 pt.print_enqueue_phase();
5277 assert(!rp->discovery_enabled(), "should have been disabled");
5278 }
5279
5280 #ifndef PRODUCT
5281 void CMSCollector::check_correct_thread_executing() {
5282 Thread* t = Thread::current();
5283 // Only the VM thread or the CMS thread should be here.
5284 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5285 "Unexpected thread type");
5286 // If this is the vm thread, the foreground process
5287 // should not be waiting. Note that _foregroundGCIsActive is
5288 // true while the foreground collector is waiting.
5289 if (_foregroundGCShouldWait) {
5290 // We cannot be the VM thread
5291 assert(t->is_ConcurrentGC_thread(),
5292 "Should be CMS thread");
5293 } else {
5294 // We can be the CMS thread only if we are in a stop-world
5295 // phase of CMS collection.
5296 if (t->is_ConcurrentGC_thread()) {
5297 assert(_collectorState == InitialMarking ||
5298 _collectorState == FinalMarking,
5299 "Should be a stop-world phase");
5300 // The CMS thread should be holding the CMS_token.
5301 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5302 "Potential interference with concurrently "
5303 "executing VM thread");
5304 }
5305 }
5306 }
5307 #endif
5308
5309 void CMSCollector::sweep() {
5310 assert(_collectorState == Sweeping, "just checking");
5311 check_correct_thread_executing();
5312 verify_work_stacks_empty();
5313 verify_overflow_empty();
5314 increment_sweep_count();
5315 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
5316
5317 _inter_sweep_timer.stop();
5318 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5319
5320 assert(!_intra_sweep_timer.is_active(), "Should not be active");
5321 _intra_sweep_timer.reset();
5322 _intra_sweep_timer.start();
5323 {
5324 GCTraceCPUTime tcpu;
5325 CMSPhaseAccounting pa(this, "Concurrent Sweep");
5326 // First sweep the old gen
5327 {
5328 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5329 bitMapLock());
5330 sweepWork(_cmsGen);
5331 }
5332
5333 // Update Universe::_heap_*_at_gc figures.
5334 // We need all the free list locks to make the abstract state
5335 // transition from Sweeping to Resetting. See detailed note
5336 // further below.
5337 {
5338 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5339 // Update heap occupancy information which is used as
5340 // input to soft ref clearing policy at the next gc.
5341 Universe::update_heap_info_at_gc();
5342 _collectorState = Resizing;
5343 }
5344 }
5345 verify_work_stacks_empty();
5346 verify_overflow_empty();
5347
5348 if (should_unload_classes()) {
5349 // Delay purge to the beginning of the next safepoint. Metaspace::contains
5350 // requires that the virtual spaces are stable and not deleted.
5351 ClassLoaderDataGraph::set_should_purge(true);
5352 }
5353
5354 _intra_sweep_timer.stop();
5355 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5356
5357 _inter_sweep_timer.reset();
5358 _inter_sweep_timer.start();
5359
5360 // We need to use a monotonically non-decreasing time in ms
5361 // or we will see time-warp warnings and os::javaTimeMillis()
5362 // does not guarantee monotonicity.
5363 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5364 update_time_of_last_gc(now);
5365
5366 // NOTE on abstract state transitions:
5367 // Mutators allocate-live and/or mark the mod-union table dirty
5368 // based on the state of the collection. The former is done in
5369 // the interval [Marking, Sweeping] and the latter in the interval
5370 // [Marking, Sweeping). Thus the transitions into the Marking state
5371 // and out of the Sweeping state must be synchronously visible
5372 // globally to the mutators.
5373 // The transition into the Marking state happens with the world
5374 // stopped so the mutators will globally see it. Sweeping is
5375 // done asynchronously by the background collector so the transition
5376 // from the Sweeping state to the Resizing state must be done
5377 // under the freelistLock (as is the check for whether to
5378 // allocate-live and whether to dirty the mod-union table).
5379 assert(_collectorState == Resizing, "Change of collector state to"
5380 " Resizing must be done under the freelistLocks (plural)");
5381
5382 // Now that sweeping has been completed, we clear
5383 // the incremental_collection_failed flag,
5384 // thus inviting a younger gen collection to promote into
5385 // this generation. If such a promotion may still fail,
5386 // the flag will be set again when a young collection is
5387 // attempted.
5388 CMSHeap* heap = CMSHeap::heap();
5389 heap->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
5390 heap->update_full_collections_completed(_collection_count_start);
5391 }
5392
5393 // FIX ME!!! Looks like this belongs in CFLSpace, with
5394 // CMSGen merely delegating to it.
5395 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5396 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5397 HeapWord* minAddr = _cmsSpace->bottom();
5398 HeapWord* largestAddr =
5399 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5400 if (largestAddr == NULL) {
5401 // The dictionary appears to be empty. In this case
5402 // try to coalesce at the end of the heap.
5403 largestAddr = _cmsSpace->end();
5404 }
5405 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5406 size_t nearLargestOffset =
5407 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5408 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5409 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5410 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5411 }
5412
5413 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5414 return addr >= _cmsSpace->nearLargestChunk();
5415 }
5416
5417 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5418 return _cmsSpace->find_chunk_at_end();
5419 }
5420
5421 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation,
5422 bool full) {
5423 // If the young generation has been collected, gather any statistics
5424 // that are of interest at this point.
5425 bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation);
5426 if (!full && current_is_young) {
5427 // Gather statistics on the young generation collection.
5428 collector()->stats().record_gc0_end(used());
5429 }
5430 }
5431
5432 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) {
5433 // We iterate over the space(s) underlying this generation,
5434 // checking the mark bit map to see if the bits corresponding
5435 // to specific blocks are marked or not. Blocks that are
5436 // marked are live and are not swept up. All remaining blocks
5437 // are swept up, with coalescing on-the-fly as we sweep up
5438 // contiguous free and/or garbage blocks:
5439 // We need to ensure that the sweeper synchronizes with allocators
5440 // and stop-the-world collectors. In particular, the following
5441 // locks are used:
5442 // . CMS token: if this is held, a stop the world collection cannot occur
5443 // . freelistLock: if this is held no allocation can occur from this
5444 // generation by another thread
5445 // . bitMapLock: if this is held, no other thread can access or update
5446 //
5447
5448 // Note that we need to hold the freelistLock if we use
5449 // block iterate below; else the iterator might go awry if
5450 // a mutator (or promotion) causes block contents to change
5451 // (for instance if the allocator divvies up a block).
5452 // If we hold the free list lock, for all practical purposes
5453 // young generation GC's can't occur (they'll usually need to
5454 // promote), so we might as well prevent all young generation
5455 // GC's while we do a sweeping step. For the same reason, we might
5456 // as well take the bit map lock for the entire duration
5457
5458 // check that we hold the requisite locks
5459 assert(have_cms_token(), "Should hold cms token");
5460 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5461 assert_lock_strong(old_gen->freelistLock());
5462 assert_lock_strong(bitMapLock());
5463
5464 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5465 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
5466 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5467 _inter_sweep_estimate.padded_average(),
5468 _intra_sweep_estimate.padded_average());
5469 old_gen->setNearLargestChunk();
5470
5471 {
5472 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield);
5473 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5474 // We need to free-up/coalesce garbage/blocks from a
5475 // co-terminal free run. This is done in the SweepClosure
5476 // destructor; so, do not remove this scope, else the
5477 // end-of-sweep-census below will be off by a little bit.
5478 }
5479 old_gen->cmsSpace()->sweep_completed();
5480 old_gen->cmsSpace()->endSweepFLCensus(sweep_count());
5481 if (should_unload_classes()) { // unloaded classes this cycle,
5482 _concurrent_cycles_since_last_unload = 0; // ... reset count
5483 } else { // did not unload classes,
5484 _concurrent_cycles_since_last_unload++; // ... increment count
5485 }
5486 }
5487
5488 // Reset CMS data structures (for now just the marking bit map)
5489 // preparatory for the next cycle.
5490 void CMSCollector::reset_concurrent() {
5491 CMSTokenSyncWithLocks ts(true, bitMapLock());
5492
5493 // If the state is not "Resetting", the foreground thread
5494 // has done a collection and the resetting.
5495 if (_collectorState != Resetting) {
5496 assert(_collectorState == Idling, "The state should only change"
5497 " because the foreground collector has finished the collection");
5498 return;
5499 }
5500
5501 {
5502 // Clear the mark bitmap (no grey objects to start with)
5503 // for the next cycle.
5504 GCTraceCPUTime tcpu;
5505 CMSPhaseAccounting cmspa(this, "Concurrent Reset");
5506
5507 HeapWord* curAddr = _markBitMap.startWord();
5508 while (curAddr < _markBitMap.endWord()) {
5509 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
5510 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5511 _markBitMap.clear_large_range(chunk);
5512 if (ConcurrentMarkSweepThread::should_yield() &&
5513 !foregroundGCIsActive() &&
5514 CMSYield) {
5515 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5516 "CMS thread should hold CMS token");
5517 assert_lock_strong(bitMapLock());
5518 bitMapLock()->unlock();
5519 ConcurrentMarkSweepThread::desynchronize(true);
5520 stopTimer();
5521 incrementYields();
5522
5523 // See the comment in coordinator_yield()
5524 for (unsigned i = 0; i < CMSYieldSleepCount &&
5525 ConcurrentMarkSweepThread::should_yield() &&
5526 !CMSCollector::foregroundGCIsActive(); ++i) {
5527 os::sleep(Thread::current(), 1, false);
5528 }
5529
5530 ConcurrentMarkSweepThread::synchronize(true);
5531 bitMapLock()->lock_without_safepoint_check();
5532 startTimer();
5533 }
5534 curAddr = chunk.end();
5535 }
5536 // A successful mostly concurrent collection has been done.
5537 // Because only the full (i.e., concurrent mode failure) collections
5538 // are being measured for gc overhead limits, clean the "near" flag
5539 // and count.
5540 size_policy()->reset_gc_overhead_limit_count();
5541 _collectorState = Idling;
5542 }
5543
5544 register_gc_end();
5545 }
5546
5547 // Same as above but for STW paths
5548 void CMSCollector::reset_stw() {
5549 // already have the lock
5550 assert(_collectorState == Resetting, "just checking");
5551 assert_lock_strong(bitMapLock());
5552 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id());
5553 _markBitMap.clear_all();
5554 _collectorState = Idling;
5555 register_gc_end();
5556 }
5557
5558 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5559 GCTraceCPUTime tcpu;
5560 TraceCollectorStats tcs(counters());
5561
5562 switch (op) {
5563 case CMS_op_checkpointRootsInitial: {
5564 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true);
5565 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5566 checkpointRootsInitial();
5567 break;
5568 }
5569 case CMS_op_checkpointRootsFinal: {
5570 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true);
5571 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5572 checkpointRootsFinal();
5573 break;
5574 }
5575 default:
5576 fatal("No such CMS_op");
5577 }
5578 }
5579
5580 #ifndef PRODUCT
5581 size_t const CMSCollector::skip_header_HeapWords() {
5582 return FreeChunk::header_size();
5583 }
5584
5585 // Try and collect here conditions that should hold when
5586 // CMS thread is exiting. The idea is that the foreground GC
5587 // thread should not be blocked if it wants to terminate
5588 // the CMS thread and yet continue to run the VM for a while
5589 // after that.
5590 void CMSCollector::verify_ok_to_terminate() const {
5591 assert(Thread::current()->is_ConcurrentGC_thread(),
5592 "should be called by CMS thread");
5593 assert(!_foregroundGCShouldWait, "should be false");
5594 // We could check here that all the various low-level locks
5595 // are not held by the CMS thread, but that is overkill; see
5596 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5597 // is checked.
5598 }
5599 #endif
5600
5601 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5602 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5603 "missing Printezis mark?");
5604 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5605 size_t size = pointer_delta(nextOneAddr + 1, addr);
5606 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5607 "alignment problem");
5608 assert(size >= 3, "Necessary for Printezis marks to work");
5609 return size;
5610 }
5611
5612 // A variant of the above (block_size_using_printezis_bits()) except
5613 // that we return 0 if the P-bits are not yet set.
5614 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5615 if (_markBitMap.isMarked(addr + 1)) {
5616 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5617 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5618 size_t size = pointer_delta(nextOneAddr + 1, addr);
5619 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5620 "alignment problem");
5621 assert(size >= 3, "Necessary for Printezis marks to work");
5622 return size;
5623 }
5624 return 0;
5625 }
5626
5627 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5628 size_t sz = 0;
5629 oop p = (oop)addr;
5630 if (p->klass_or_null_acquire() != NULL) {
5631 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5632 } else {
5633 sz = block_size_using_printezis_bits(addr);
5634 }
5635 assert(sz > 0, "size must be nonzero");
5636 HeapWord* next_block = addr + sz;
5637 HeapWord* next_card = align_up(next_block, CardTableModRefBS::card_size);
5638 assert(align_down((uintptr_t)addr, CardTableModRefBS::card_size) <
5639 align_down((uintptr_t)next_card, CardTableModRefBS::card_size),
5640 "must be different cards");
5641 return next_card;
5642 }
5643
5644
5645 // CMS Bit Map Wrapper /////////////////////////////////////////
5646
5647 // Construct a CMS bit map infrastructure, but don't create the
5648 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5649 // further below.
5650 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5651 _bm(),
5652 _shifter(shifter),
5653 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5654 Monitor::_safepoint_check_sometimes) : NULL)
5655 {
5656 _bmStartWord = 0;
5657 _bmWordSize = 0;
5658 }
5659
5660 bool CMSBitMap::allocate(MemRegion mr) {
5661 _bmStartWord = mr.start();
5662 _bmWordSize = mr.word_size();
5663 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5664 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5665 if (!brs.is_reserved()) {
5666 log_warning(gc)("CMS bit map allocation failure");
5667 return false;
5668 }
5669 // For now we'll just commit all of the bit map up front.
5670 // Later on we'll try to be more parsimonious with swap.
5671 if (!_virtual_space.initialize(brs, brs.size())) {
5672 log_warning(gc)("CMS bit map backing store failure");
5673 return false;
5674 }
5675 assert(_virtual_space.committed_size() == brs.size(),
5676 "didn't reserve backing store for all of CMS bit map?");
5677 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5678 _bmWordSize, "inconsistency in bit map sizing");
5679 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter);
5680
5681 // bm.clear(); // can we rely on getting zero'd memory? verify below
5682 assert(isAllClear(),
5683 "Expected zero'd memory from ReservedSpace constructor");
5684 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5685 "consistency check");
5686 return true;
5687 }
5688
5689 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5690 HeapWord *next_addr, *end_addr, *last_addr;
5691 assert_locked();
5692 assert(covers(mr), "out-of-range error");
5693 // XXX assert that start and end are appropriately aligned
5694 for (next_addr = mr.start(), end_addr = mr.end();
5695 next_addr < end_addr; next_addr = last_addr) {
5696 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5697 last_addr = dirty_region.end();
5698 if (!dirty_region.is_empty()) {
5699 cl->do_MemRegion(dirty_region);
5700 } else {
5701 assert(last_addr == end_addr, "program logic");
5702 return;
5703 }
5704 }
5705 }
5706
5707 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5708 _bm.print_on_error(st, prefix);
5709 }
5710
5711 #ifndef PRODUCT
5712 void CMSBitMap::assert_locked() const {
5713 CMSLockVerifier::assert_locked(lock());
5714 }
5715
5716 bool CMSBitMap::covers(MemRegion mr) const {
5717 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5718 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5719 "size inconsistency");
5720 return (mr.start() >= _bmStartWord) &&
5721 (mr.end() <= endWord());
5722 }
5723
5724 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5725 return (start >= _bmStartWord && (start + size) <= endWord());
5726 }
5727
5728 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5729 // verify that there are no 1 bits in the interval [left, right)
5730 FalseBitMapClosure falseBitMapClosure;
5731 iterate(&falseBitMapClosure, left, right);
5732 }
5733
5734 void CMSBitMap::region_invariant(MemRegion mr)
5735 {
5736 assert_locked();
5737 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5738 assert(!mr.is_empty(), "unexpected empty region");
5739 assert(covers(mr), "mr should be covered by bit map");
5740 // convert address range into offset range
5741 size_t start_ofs = heapWordToOffset(mr.start());
5742 // Make sure that end() is appropriately aligned
5743 assert(mr.end() == align_up(mr.end(), (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(oopDesc::is_oop(obj), "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(oopDesc::is_oop(obj), "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(oopDesc::is_oop(obj), "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 LogStream ls(log.error());
5888 oop(addr)->print_on(&ls);
5889 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
5890 fatal("... aborting");
5891 }
5892 }
5893 }
5894
5895 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5896 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5897
5898 //////////////////////////////////////////////////
5899 // MarkRefsIntoAndScanClosure
5900 //////////////////////////////////////////////////
5901
5902 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
5903 ReferenceProcessor* rp,
5904 CMSBitMap* bit_map,
5905 CMSBitMap* mod_union_table,
5906 CMSMarkStack* mark_stack,
5907 CMSCollector* collector,
5908 bool should_yield,
5909 bool concurrent_precleaning):
5910 _collector(collector),
5911 _span(span),
5912 _bit_map(bit_map),
5913 _mark_stack(mark_stack),
5914 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
5915 mark_stack, concurrent_precleaning),
5916 _yield(should_yield),
5917 _concurrent_precleaning(concurrent_precleaning),
5918 _freelistLock(NULL)
5919 {
5920 // FIXME: Should initialize in base class constructor.
5921 assert(rp != NULL, "ref_processor shouldn't be NULL");
5922 set_ref_processor_internal(rp);
5923 }
5924
5925 // This closure is used to mark refs into the CMS generation at the
5926 // second (final) checkpoint, and to scan and transitively follow
5927 // the unmarked oops. It is also used during the concurrent precleaning
5928 // phase while scanning objects on dirty cards in the CMS generation.
5929 // The marks are made in the marking bit map and the marking stack is
5930 // used for keeping the (newly) grey objects during the scan.
5931 // The parallel version (Par_...) appears further below.
5932 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
5933 if (obj != NULL) {
5934 assert(oopDesc::is_oop(obj), "expected an oop");
5935 HeapWord* addr = (HeapWord*)obj;
5936 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
5937 assert(_collector->overflow_list_is_empty(),
5938 "overflow list should be empty");
5939 if (_span.contains(addr) &&
5940 !_bit_map->isMarked(addr)) {
5941 // mark bit map (object is now grey)
5942 _bit_map->mark(addr);
5943 // push on marking stack (stack should be empty), and drain the
5944 // stack by applying this closure to the oops in the oops popped
5945 // from the stack (i.e. blacken the grey objects)
5946 bool res = _mark_stack->push(obj);
5947 assert(res, "Should have space to push on empty stack");
5948 do {
5949 oop new_oop = _mark_stack->pop();
5950 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
5951 assert(_bit_map->isMarked((HeapWord*)new_oop),
5952 "only grey objects on this stack");
5953 // iterate over the oops in this oop, marking and pushing
5954 // the ones in CMS heap (i.e. in _span).
5955 new_oop->oop_iterate(&_pushAndMarkClosure);
5956 // check if it's time to yield
5957 do_yield_check();
5958 } while (!_mark_stack->isEmpty() ||
5959 (!_concurrent_precleaning && take_from_overflow_list()));
5960 // if marking stack is empty, and we are not doing this
5961 // during precleaning, then check the overflow list
5962 }
5963 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
5964 assert(_collector->overflow_list_is_empty(),
5965 "overflow list was drained above");
5966
5967 assert(_collector->no_preserved_marks(),
5968 "All preserved marks should have been restored above");
5969 }
5970 }
5971
5972 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5973 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5974
5975 void MarkRefsIntoAndScanClosure::do_yield_work() {
5976 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5977 "CMS thread should hold CMS token");
5978 assert_lock_strong(_freelistLock);
5979 assert_lock_strong(_bit_map->lock());
5980 // relinquish the free_list_lock and bitMaplock()
5981 _bit_map->lock()->unlock();
5982 _freelistLock->unlock();
5983 ConcurrentMarkSweepThread::desynchronize(true);
5984 _collector->stopTimer();
5985 _collector->incrementYields();
5986
5987 // See the comment in coordinator_yield()
5988 for (unsigned i = 0;
5989 i < CMSYieldSleepCount &&
5990 ConcurrentMarkSweepThread::should_yield() &&
5991 !CMSCollector::foregroundGCIsActive();
5992 ++i) {
5993 os::sleep(Thread::current(), 1, false);
5994 }
5995
5996 ConcurrentMarkSweepThread::synchronize(true);
5997 _freelistLock->lock_without_safepoint_check();
5998 _bit_map->lock()->lock_without_safepoint_check();
5999 _collector->startTimer();
6000 }
6001
6002 ///////////////////////////////////////////////////////////
6003 // ParMarkRefsIntoAndScanClosure: a parallel version of
6004 // MarkRefsIntoAndScanClosure
6005 ///////////////////////////////////////////////////////////
6006 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure(
6007 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6008 CMSBitMap* bit_map, OopTaskQueue* work_queue):
6009 _span(span),
6010 _bit_map(bit_map),
6011 _work_queue(work_queue),
6012 _low_water_mark(MIN2((work_queue->max_elems()/4),
6013 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6014 _parPushAndMarkClosure(collector, span, rp, bit_map, work_queue)
6015 {
6016 // FIXME: Should initialize in base class constructor.
6017 assert(rp != NULL, "ref_processor shouldn't be NULL");
6018 set_ref_processor_internal(rp);
6019 }
6020
6021 // This closure is used to mark refs into the CMS generation at the
6022 // second (final) checkpoint, and to scan and transitively follow
6023 // the unmarked oops. The marks are made in the marking bit map and
6024 // the work_queue is used for keeping the (newly) grey objects during
6025 // the scan phase whence they are also available for stealing by parallel
6026 // threads. Since the marking bit map is shared, updates are
6027 // synchronized (via CAS).
6028 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) {
6029 if (obj != NULL) {
6030 // Ignore mark word because this could be an already marked oop
6031 // that may be chained at the end of the overflow list.
6032 assert(oopDesc::is_oop(obj, true), "expected an oop");
6033 HeapWord* addr = (HeapWord*)obj;
6034 if (_span.contains(addr) &&
6035 !_bit_map->isMarked(addr)) {
6036 // mark bit map (object will become grey):
6037 // It is possible for several threads to be
6038 // trying to "claim" this object concurrently;
6039 // the unique thread that succeeds in marking the
6040 // object first will do the subsequent push on
6041 // to the work queue (or overflow list).
6042 if (_bit_map->par_mark(addr)) {
6043 // push on work_queue (which may not be empty), and trim the
6044 // queue to an appropriate length by applying this closure to
6045 // the oops in the oops popped from the stack (i.e. blacken the
6046 // grey objects)
6047 bool res = _work_queue->push(obj);
6048 assert(res, "Low water mark should be less than capacity?");
6049 trim_queue(_low_water_mark);
6050 } // Else, another thread claimed the object
6051 }
6052 }
6053 }
6054
6055 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6056 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6057
6058 // This closure is used to rescan the marked objects on the dirty cards
6059 // in the mod union table and the card table proper.
6060 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6061 oop p, MemRegion mr) {
6062
6063 size_t size = 0;
6064 HeapWord* addr = (HeapWord*)p;
6065 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6066 assert(_span.contains(addr), "we are scanning the CMS generation");
6067 // check if it's time to yield
6068 if (do_yield_check()) {
6069 // We yielded for some foreground stop-world work,
6070 // and we have been asked to abort this ongoing preclean cycle.
6071 return 0;
6072 }
6073 if (_bitMap->isMarked(addr)) {
6074 // it's marked; is it potentially uninitialized?
6075 if (p->klass_or_null_acquire() != NULL) {
6076 // an initialized object; ignore mark word in verification below
6077 // since we are running concurrent with mutators
6078 assert(oopDesc::is_oop(p, true), "should be an oop");
6079 if (p->is_objArray()) {
6080 // objArrays are precisely marked; restrict scanning
6081 // to dirty cards only.
6082 size = CompactibleFreeListSpace::adjustObjectSize(
6083 p->oop_iterate_size(_scanningClosure, mr));
6084 } else {
6085 // A non-array may have been imprecisely marked; we need
6086 // to scan object in its entirety.
6087 size = CompactibleFreeListSpace::adjustObjectSize(
6088 p->oop_iterate_size(_scanningClosure));
6089 }
6090 #ifdef ASSERT
6091 size_t direct_size =
6092 CompactibleFreeListSpace::adjustObjectSize(p->size());
6093 assert(size == direct_size, "Inconsistency in size");
6094 assert(size >= 3, "Necessary for Printezis marks to work");
6095 HeapWord* start_pbit = addr + 1;
6096 HeapWord* end_pbit = addr + size - 1;
6097 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit),
6098 "inconsistent Printezis mark");
6099 // Verify inner mark bits (between Printezis bits) are clear,
6100 // but don't repeat if there are multiple dirty regions for
6101 // the same object, to avoid potential O(N^2) performance.
6102 if (addr != _last_scanned_object) {
6103 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit);
6104 _last_scanned_object = addr;
6105 }
6106 #endif // ASSERT
6107 } else {
6108 // An uninitialized object.
6109 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6110 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6111 size = pointer_delta(nextOneAddr + 1, addr);
6112 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6113 "alignment problem");
6114 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6115 // will dirty the card when the klass pointer is installed in the
6116 // object (signaling the completion of initialization).
6117 }
6118 } else {
6119 // Either a not yet marked object or an uninitialized object
6120 if (p->klass_or_null_acquire() == NULL) {
6121 // An uninitialized object, skip to the next card, since
6122 // we may not be able to read its P-bits yet.
6123 assert(size == 0, "Initial value");
6124 } else {
6125 // An object not (yet) reached by marking: we merely need to
6126 // compute its size so as to go look at the next block.
6127 assert(oopDesc::is_oop(p, true), "should be an oop");
6128 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6129 }
6130 }
6131 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6132 return size;
6133 }
6134
6135 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6136 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6137 "CMS thread should hold CMS token");
6138 assert_lock_strong(_freelistLock);
6139 assert_lock_strong(_bitMap->lock());
6140 // relinquish the free_list_lock and bitMaplock()
6141 _bitMap->lock()->unlock();
6142 _freelistLock->unlock();
6143 ConcurrentMarkSweepThread::desynchronize(true);
6144 _collector->stopTimer();
6145 _collector->incrementYields();
6146
6147 // See the comment in coordinator_yield()
6148 for (unsigned i = 0; i < CMSYieldSleepCount &&
6149 ConcurrentMarkSweepThread::should_yield() &&
6150 !CMSCollector::foregroundGCIsActive(); ++i) {
6151 os::sleep(Thread::current(), 1, false);
6152 }
6153
6154 ConcurrentMarkSweepThread::synchronize(true);
6155 _freelistLock->lock_without_safepoint_check();
6156 _bitMap->lock()->lock_without_safepoint_check();
6157 _collector->startTimer();
6158 }
6159
6160
6161 //////////////////////////////////////////////////////////////////
6162 // SurvivorSpacePrecleanClosure
6163 //////////////////////////////////////////////////////////////////
6164 // This (single-threaded) closure is used to preclean the oops in
6165 // the survivor spaces.
6166 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6167
6168 HeapWord* addr = (HeapWord*)p;
6169 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6170 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6171 assert(p->klass_or_null() != NULL, "object should be initialized");
6172 // an initialized object; ignore mark word in verification below
6173 // since we are running concurrent with mutators
6174 assert(oopDesc::is_oop(p, true), "should be an oop");
6175 // Note that we do not yield while we iterate over
6176 // the interior oops of p, pushing the relevant ones
6177 // on our marking stack.
6178 size_t size = p->oop_iterate_size(_scanning_closure);
6179 do_yield_check();
6180 // Observe that below, we do not abandon the preclean
6181 // phase as soon as we should; rather we empty the
6182 // marking stack before returning. This is to satisfy
6183 // some existing assertions. In general, it may be a
6184 // good idea to abort immediately and complete the marking
6185 // from the grey objects at a later time.
6186 while (!_mark_stack->isEmpty()) {
6187 oop new_oop = _mark_stack->pop();
6188 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
6189 assert(_bit_map->isMarked((HeapWord*)new_oop),
6190 "only grey objects on this stack");
6191 // iterate over the oops in this oop, marking and pushing
6192 // the ones in CMS heap (i.e. in _span).
6193 new_oop->oop_iterate(_scanning_closure);
6194 // check if it's time to yield
6195 do_yield_check();
6196 }
6197 unsigned int after_count =
6198 CMSHeap::heap()->total_collections();
6199 bool abort = (_before_count != after_count) ||
6200 _collector->should_abort_preclean();
6201 return abort ? 0 : size;
6202 }
6203
6204 void SurvivorSpacePrecleanClosure::do_yield_work() {
6205 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6206 "CMS thread should hold CMS token");
6207 assert_lock_strong(_bit_map->lock());
6208 // Relinquish the bit map lock
6209 _bit_map->lock()->unlock();
6210 ConcurrentMarkSweepThread::desynchronize(true);
6211 _collector->stopTimer();
6212 _collector->incrementYields();
6213
6214 // See the comment in coordinator_yield()
6215 for (unsigned i = 0; i < CMSYieldSleepCount &&
6216 ConcurrentMarkSweepThread::should_yield() &&
6217 !CMSCollector::foregroundGCIsActive(); ++i) {
6218 os::sleep(Thread::current(), 1, false);
6219 }
6220
6221 ConcurrentMarkSweepThread::synchronize(true);
6222 _bit_map->lock()->lock_without_safepoint_check();
6223 _collector->startTimer();
6224 }
6225
6226 // This closure is used to rescan the marked objects on the dirty cards
6227 // in the mod union table and the card table proper. In the parallel
6228 // case, although the bitMap is shared, we do a single read so the
6229 // isMarked() query is "safe".
6230 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6231 // Ignore mark word because we are running concurrent with mutators
6232 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p));
6233 HeapWord* addr = (HeapWord*)p;
6234 assert(_span.contains(addr), "we are scanning the CMS generation");
6235 bool is_obj_array = false;
6236 #ifdef ASSERT
6237 if (!_parallel) {
6238 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6239 assert(_collector->overflow_list_is_empty(),
6240 "overflow list should be empty");
6241
6242 }
6243 #endif // ASSERT
6244 if (_bit_map->isMarked(addr)) {
6245 // Obj arrays are precisely marked, non-arrays are not;
6246 // so we scan objArrays precisely and non-arrays in their
6247 // entirety.
6248 if (p->is_objArray()) {
6249 is_obj_array = true;
6250 if (_parallel) {
6251 p->oop_iterate(_par_scan_closure, mr);
6252 } else {
6253 p->oop_iterate(_scan_closure, mr);
6254 }
6255 } else {
6256 if (_parallel) {
6257 p->oop_iterate(_par_scan_closure);
6258 } else {
6259 p->oop_iterate(_scan_closure);
6260 }
6261 }
6262 }
6263 #ifdef ASSERT
6264 if (!_parallel) {
6265 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6266 assert(_collector->overflow_list_is_empty(),
6267 "overflow list should be empty");
6268
6269 }
6270 #endif // ASSERT
6271 return is_obj_array;
6272 }
6273
6274 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6275 MemRegion span,
6276 CMSBitMap* bitMap, CMSMarkStack* markStack,
6277 bool should_yield, bool verifying):
6278 _collector(collector),
6279 _span(span),
6280 _bitMap(bitMap),
6281 _mut(&collector->_modUnionTable),
6282 _markStack(markStack),
6283 _yield(should_yield),
6284 _skipBits(0)
6285 {
6286 assert(_markStack->isEmpty(), "stack should be empty");
6287 _finger = _bitMap->startWord();
6288 _threshold = _finger;
6289 assert(_collector->_restart_addr == NULL, "Sanity check");
6290 assert(_span.contains(_finger), "Out of bounds _finger?");
6291 DEBUG_ONLY(_verifying = verifying;)
6292 }
6293
6294 void MarkFromRootsClosure::reset(HeapWord* addr) {
6295 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6296 assert(_span.contains(addr), "Out of bounds _finger?");
6297 _finger = addr;
6298 _threshold = align_up(_finger, CardTableModRefBS::card_size);
6299 }
6300
6301 // Should revisit to see if this should be restructured for
6302 // greater efficiency.
6303 bool MarkFromRootsClosure::do_bit(size_t offset) {
6304 if (_skipBits > 0) {
6305 _skipBits--;
6306 return true;
6307 }
6308 // convert offset into a HeapWord*
6309 HeapWord* addr = _bitMap->startWord() + offset;
6310 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6311 "address out of range");
6312 assert(_bitMap->isMarked(addr), "tautology");
6313 if (_bitMap->isMarked(addr+1)) {
6314 // this is an allocated but not yet initialized object
6315 assert(_skipBits == 0, "tautology");
6316 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6317 oop p = oop(addr);
6318 if (p->klass_or_null_acquire() == NULL) {
6319 DEBUG_ONLY(if (!_verifying) {)
6320 // We re-dirty the cards on which this object lies and increase
6321 // the _threshold so that we'll come back to scan this object
6322 // during the preclean or remark phase. (CMSCleanOnEnter)
6323 if (CMSCleanOnEnter) {
6324 size_t sz = _collector->block_size_using_printezis_bits(addr);
6325 HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size);
6326 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6327 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6328 // Bump _threshold to end_card_addr; note that
6329 // _threshold cannot possibly exceed end_card_addr, anyhow.
6330 // This prevents future clearing of the card as the scan proceeds
6331 // to the right.
6332 assert(_threshold <= end_card_addr,
6333 "Because we are just scanning into this object");
6334 if (_threshold < end_card_addr) {
6335 _threshold = end_card_addr;
6336 }
6337 if (p->klass_or_null_acquire() != NULL) {
6338 // Redirty the range of cards...
6339 _mut->mark_range(redirty_range);
6340 } // ...else the setting of klass will dirty the card anyway.
6341 }
6342 DEBUG_ONLY(})
6343 return true;
6344 }
6345 }
6346 scanOopsInOop(addr);
6347 return true;
6348 }
6349
6350 // We take a break if we've been at this for a while,
6351 // so as to avoid monopolizing the locks involved.
6352 void MarkFromRootsClosure::do_yield_work() {
6353 // First give up the locks, then yield, then re-lock
6354 // We should probably use a constructor/destructor idiom to
6355 // do this unlock/lock or modify the MutexUnlocker class to
6356 // serve our purpose. XXX
6357 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6358 "CMS thread should hold CMS token");
6359 assert_lock_strong(_bitMap->lock());
6360 _bitMap->lock()->unlock();
6361 ConcurrentMarkSweepThread::desynchronize(true);
6362 _collector->stopTimer();
6363 _collector->incrementYields();
6364
6365 // See the comment in coordinator_yield()
6366 for (unsigned i = 0; i < CMSYieldSleepCount &&
6367 ConcurrentMarkSweepThread::should_yield() &&
6368 !CMSCollector::foregroundGCIsActive(); ++i) {
6369 os::sleep(Thread::current(), 1, false);
6370 }
6371
6372 ConcurrentMarkSweepThread::synchronize(true);
6373 _bitMap->lock()->lock_without_safepoint_check();
6374 _collector->startTimer();
6375 }
6376
6377 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6378 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6379 assert(_markStack->isEmpty(),
6380 "should drain stack to limit stack usage");
6381 // convert ptr to an oop preparatory to scanning
6382 oop obj = oop(ptr);
6383 // Ignore mark word in verification below, since we
6384 // may be running concurrent with mutators.
6385 assert(oopDesc::is_oop(obj, true), "should be an oop");
6386 assert(_finger <= ptr, "_finger runneth ahead");
6387 // advance the finger to right end of this object
6388 _finger = ptr + obj->size();
6389 assert(_finger > ptr, "we just incremented it above");
6390 // On large heaps, it may take us some time to get through
6391 // the marking phase. During
6392 // this time it's possible that a lot of mutations have
6393 // accumulated in the card table and the mod union table --
6394 // these mutation records are redundant until we have
6395 // actually traced into the corresponding card.
6396 // Here, we check whether advancing the finger would make
6397 // us cross into a new card, and if so clear corresponding
6398 // cards in the MUT (preclean them in the card-table in the
6399 // future).
6400
6401 DEBUG_ONLY(if (!_verifying) {)
6402 // The clean-on-enter optimization is disabled by default,
6403 // until we fix 6178663.
6404 if (CMSCleanOnEnter && (_finger > _threshold)) {
6405 // [_threshold, _finger) represents the interval
6406 // of cards to be cleared in MUT (or precleaned in card table).
6407 // The set of cards to be cleared is all those that overlap
6408 // with the interval [_threshold, _finger); note that
6409 // _threshold is always kept card-aligned but _finger isn't
6410 // always card-aligned.
6411 HeapWord* old_threshold = _threshold;
6412 assert(is_aligned(old_threshold, CardTableModRefBS::card_size),
6413 "_threshold should always be card-aligned");
6414 _threshold = align_up(_finger, CardTableModRefBS::card_size);
6415 MemRegion mr(old_threshold, _threshold);
6416 assert(!mr.is_empty(), "Control point invariant");
6417 assert(_span.contains(mr), "Should clear within span");
6418 _mut->clear_range(mr);
6419 }
6420 DEBUG_ONLY(})
6421 // Note: the finger doesn't advance while we drain
6422 // the stack below.
6423 PushOrMarkClosure pushOrMarkClosure(_collector,
6424 _span, _bitMap, _markStack,
6425 _finger, this);
6426 bool res = _markStack->push(obj);
6427 assert(res, "Empty non-zero size stack should have space for single push");
6428 while (!_markStack->isEmpty()) {
6429 oop new_oop = _markStack->pop();
6430 // Skip verifying header mark word below because we are
6431 // running concurrent with mutators.
6432 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop");
6433 // now scan this oop's oops
6434 new_oop->oop_iterate(&pushOrMarkClosure);
6435 do_yield_check();
6436 }
6437 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6438 }
6439
6440 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task,
6441 CMSCollector* collector, MemRegion span,
6442 CMSBitMap* bit_map,
6443 OopTaskQueue* work_queue,
6444 CMSMarkStack* overflow_stack):
6445 _collector(collector),
6446 _whole_span(collector->_span),
6447 _span(span),
6448 _bit_map(bit_map),
6449 _mut(&collector->_modUnionTable),
6450 _work_queue(work_queue),
6451 _overflow_stack(overflow_stack),
6452 _skip_bits(0),
6453 _task(task)
6454 {
6455 assert(_work_queue->size() == 0, "work_queue should be empty");
6456 _finger = span.start();
6457 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
6458 assert(_span.contains(_finger), "Out of bounds _finger?");
6459 }
6460
6461 // Should revisit to see if this should be restructured for
6462 // greater efficiency.
6463 bool ParMarkFromRootsClosure::do_bit(size_t offset) {
6464 if (_skip_bits > 0) {
6465 _skip_bits--;
6466 return true;
6467 }
6468 // convert offset into a HeapWord*
6469 HeapWord* addr = _bit_map->startWord() + offset;
6470 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6471 "address out of range");
6472 assert(_bit_map->isMarked(addr), "tautology");
6473 if (_bit_map->isMarked(addr+1)) {
6474 // this is an allocated object that might not yet be initialized
6475 assert(_skip_bits == 0, "tautology");
6476 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
6477 oop p = oop(addr);
6478 if (p->klass_or_null_acquire() == NULL) {
6479 // in the case of Clean-on-Enter optimization, redirty card
6480 // and avoid clearing card by increasing the threshold.
6481 return true;
6482 }
6483 }
6484 scan_oops_in_oop(addr);
6485 return true;
6486 }
6487
6488 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6489 assert(_bit_map->isMarked(ptr), "expected bit to be set");
6490 // Should we assert that our work queue is empty or
6491 // below some drain limit?
6492 assert(_work_queue->size() == 0,
6493 "should drain stack to limit stack usage");
6494 // convert ptr to an oop preparatory to scanning
6495 oop obj = oop(ptr);
6496 // Ignore mark word in verification below, since we
6497 // may be running concurrent with mutators.
6498 assert(oopDesc::is_oop(obj, true), "should be an oop");
6499 assert(_finger <= ptr, "_finger runneth ahead");
6500 // advance the finger to right end of this object
6501 _finger = ptr + obj->size();
6502 assert(_finger > ptr, "we just incremented it above");
6503 // On large heaps, it may take us some time to get through
6504 // the marking phase. During
6505 // this time it's possible that a lot of mutations have
6506 // accumulated in the card table and the mod union table --
6507 // these mutation records are redundant until we have
6508 // actually traced into the corresponding card.
6509 // Here, we check whether advancing the finger would make
6510 // us cross into a new card, and if so clear corresponding
6511 // cards in the MUT (preclean them in the card-table in the
6512 // future).
6513
6514 // The clean-on-enter optimization is disabled by default,
6515 // until we fix 6178663.
6516 if (CMSCleanOnEnter && (_finger > _threshold)) {
6517 // [_threshold, _finger) represents the interval
6518 // of cards to be cleared in MUT (or precleaned in card table).
6519 // The set of cards to be cleared is all those that overlap
6520 // with the interval [_threshold, _finger); note that
6521 // _threshold is always kept card-aligned but _finger isn't
6522 // always card-aligned.
6523 HeapWord* old_threshold = _threshold;
6524 assert(is_aligned(old_threshold, CardTableModRefBS::card_size),
6525 "_threshold should always be card-aligned");
6526 _threshold = align_up(_finger, CardTableModRefBS::card_size);
6527 MemRegion mr(old_threshold, _threshold);
6528 assert(!mr.is_empty(), "Control point invariant");
6529 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6530 _mut->clear_range(mr);
6531 }
6532
6533 // Note: the local finger doesn't advance while we drain
6534 // the stack below, but the global finger sure can and will.
6535 HeapWord* volatile* gfa = _task->global_finger_addr();
6536 ParPushOrMarkClosure pushOrMarkClosure(_collector,
6537 _span, _bit_map,
6538 _work_queue,
6539 _overflow_stack,
6540 _finger,
6541 gfa, this);
6542 bool res = _work_queue->push(obj); // overflow could occur here
6543 assert(res, "Will hold once we use workqueues");
6544 while (true) {
6545 oop new_oop;
6546 if (!_work_queue->pop_local(new_oop)) {
6547 // We emptied our work_queue; check if there's stuff that can
6548 // be gotten from the overflow stack.
6549 if (CMSConcMarkingTask::get_work_from_overflow_stack(
6550 _overflow_stack, _work_queue)) {
6551 do_yield_check();
6552 continue;
6553 } else { // done
6554 break;
6555 }
6556 }
6557 // Skip verifying header mark word below because we are
6558 // running concurrent with mutators.
6559 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop");
6560 // now scan this oop's oops
6561 new_oop->oop_iterate(&pushOrMarkClosure);
6562 do_yield_check();
6563 }
6564 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6565 }
6566
6567 // Yield in response to a request from VM Thread or
6568 // from mutators.
6569 void ParMarkFromRootsClosure::do_yield_work() {
6570 assert(_task != NULL, "sanity");
6571 _task->yield();
6572 }
6573
6574 // A variant of the above used for verifying CMS marking work.
6575 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6576 MemRegion span,
6577 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6578 CMSMarkStack* mark_stack):
6579 _collector(collector),
6580 _span(span),
6581 _verification_bm(verification_bm),
6582 _cms_bm(cms_bm),
6583 _mark_stack(mark_stack),
6584 _pam_verify_closure(collector, span, verification_bm, cms_bm,
6585 mark_stack)
6586 {
6587 assert(_mark_stack->isEmpty(), "stack should be empty");
6588 _finger = _verification_bm->startWord();
6589 assert(_collector->_restart_addr == NULL, "Sanity check");
6590 assert(_span.contains(_finger), "Out of bounds _finger?");
6591 }
6592
6593 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6594 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6595 assert(_span.contains(addr), "Out of bounds _finger?");
6596 _finger = addr;
6597 }
6598
6599 // Should revisit to see if this should be restructured for
6600 // greater efficiency.
6601 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6602 // convert offset into a HeapWord*
6603 HeapWord* addr = _verification_bm->startWord() + offset;
6604 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6605 "address out of range");
6606 assert(_verification_bm->isMarked(addr), "tautology");
6607 assert(_cms_bm->isMarked(addr), "tautology");
6608
6609 assert(_mark_stack->isEmpty(),
6610 "should drain stack to limit stack usage");
6611 // convert addr to an oop preparatory to scanning
6612 oop obj = oop(addr);
6613 assert(oopDesc::is_oop(obj), "should be an oop");
6614 assert(_finger <= addr, "_finger runneth ahead");
6615 // advance the finger to right end of this object
6616 _finger = addr + obj->size();
6617 assert(_finger > addr, "we just incremented it above");
6618 // Note: the finger doesn't advance while we drain
6619 // the stack below.
6620 bool res = _mark_stack->push(obj);
6621 assert(res, "Empty non-zero size stack should have space for single push");
6622 while (!_mark_stack->isEmpty()) {
6623 oop new_oop = _mark_stack->pop();
6624 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop");
6625 // now scan this oop's oops
6626 new_oop->oop_iterate(&_pam_verify_closure);
6627 }
6628 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6629 return true;
6630 }
6631
6632 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6633 CMSCollector* collector, MemRegion span,
6634 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6635 CMSMarkStack* mark_stack):
6636 MetadataAwareOopClosure(collector->ref_processor()),
6637 _collector(collector),
6638 _span(span),
6639 _verification_bm(verification_bm),
6640 _cms_bm(cms_bm),
6641 _mark_stack(mark_stack)
6642 { }
6643
6644 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6645 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6646
6647 // Upon stack overflow, we discard (part of) the stack,
6648 // remembering the least address amongst those discarded
6649 // in CMSCollector's _restart_address.
6650 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6651 // Remember the least grey address discarded
6652 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6653 _collector->lower_restart_addr(ra);
6654 _mark_stack->reset(); // discard stack contents
6655 _mark_stack->expand(); // expand the stack if possible
6656 }
6657
6658 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6659 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6660 HeapWord* addr = (HeapWord*)obj;
6661 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6662 // Oop lies in _span and isn't yet grey or black
6663 _verification_bm->mark(addr); // now grey
6664 if (!_cms_bm->isMarked(addr)) {
6665 Log(gc, verify) log;
6666 ResourceMark rm;
6667 LogStream ls(log.error());
6668 oop(addr)->print_on(&ls);
6669 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6670 fatal("... aborting");
6671 }
6672
6673 if (!_mark_stack->push(obj)) { // stack overflow
6674 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity());
6675 assert(_mark_stack->isFull(), "Else push should have succeeded");
6676 handle_stack_overflow(addr);
6677 }
6678 // anything including and to the right of _finger
6679 // will be scanned as we iterate over the remainder of the
6680 // bit map
6681 }
6682 }
6683
6684 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6685 MemRegion span,
6686 CMSBitMap* bitMap, CMSMarkStack* markStack,
6687 HeapWord* finger, MarkFromRootsClosure* parent) :
6688 MetadataAwareOopClosure(collector->ref_processor()),
6689 _collector(collector),
6690 _span(span),
6691 _bitMap(bitMap),
6692 _markStack(markStack),
6693 _finger(finger),
6694 _parent(parent)
6695 { }
6696
6697 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector,
6698 MemRegion span,
6699 CMSBitMap* bit_map,
6700 OopTaskQueue* work_queue,
6701 CMSMarkStack* overflow_stack,
6702 HeapWord* finger,
6703 HeapWord* volatile* global_finger_addr,
6704 ParMarkFromRootsClosure* parent) :
6705 MetadataAwareOopClosure(collector->ref_processor()),
6706 _collector(collector),
6707 _whole_span(collector->_span),
6708 _span(span),
6709 _bit_map(bit_map),
6710 _work_queue(work_queue),
6711 _overflow_stack(overflow_stack),
6712 _finger(finger),
6713 _global_finger_addr(global_finger_addr),
6714 _parent(parent)
6715 { }
6716
6717 // Assumes thread-safe access by callers, who are
6718 // responsible for mutual exclusion.
6719 void CMSCollector::lower_restart_addr(HeapWord* low) {
6720 assert(_span.contains(low), "Out of bounds addr");
6721 if (_restart_addr == NULL) {
6722 _restart_addr = low;
6723 } else {
6724 _restart_addr = MIN2(_restart_addr, low);
6725 }
6726 }
6727
6728 // Upon stack overflow, we discard (part of) the stack,
6729 // remembering the least address amongst those discarded
6730 // in CMSCollector's _restart_address.
6731 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6732 // Remember the least grey address discarded
6733 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6734 _collector->lower_restart_addr(ra);
6735 _markStack->reset(); // discard stack contents
6736 _markStack->expand(); // expand the stack if possible
6737 }
6738
6739 // Upon stack overflow, we discard (part of) the stack,
6740 // remembering the least address amongst those discarded
6741 // in CMSCollector's _restart_address.
6742 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6743 // We need to do this under a mutex to prevent other
6744 // workers from interfering with the work done below.
6745 MutexLockerEx ml(_overflow_stack->par_lock(),
6746 Mutex::_no_safepoint_check_flag);
6747 // Remember the least grey address discarded
6748 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
6749 _collector->lower_restart_addr(ra);
6750 _overflow_stack->reset(); // discard stack contents
6751 _overflow_stack->expand(); // expand the stack if possible
6752 }
6753
6754 void PushOrMarkClosure::do_oop(oop obj) {
6755 // Ignore mark word because we are running concurrent with mutators.
6756 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6757 HeapWord* addr = (HeapWord*)obj;
6758 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
6759 // Oop lies in _span and isn't yet grey or black
6760 _bitMap->mark(addr); // now grey
6761 if (addr < _finger) {
6762 // the bit map iteration has already either passed, or
6763 // sampled, this bit in the bit map; we'll need to
6764 // use the marking stack to scan this oop's oops.
6765 bool simulate_overflow = false;
6766 NOT_PRODUCT(
6767 if (CMSMarkStackOverflowALot &&
6768 _collector->simulate_overflow()) {
6769 // simulate a stack overflow
6770 simulate_overflow = true;
6771 }
6772 )
6773 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
6774 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity());
6775 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
6776 handle_stack_overflow(addr);
6777 }
6778 }
6779 // anything including and to the right of _finger
6780 // will be scanned as we iterate over the remainder of the
6781 // bit map
6782 do_yield_check();
6783 }
6784 }
6785
6786 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
6787 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
6788
6789 void ParPushOrMarkClosure::do_oop(oop obj) {
6790 // Ignore mark word because we are running concurrent with mutators.
6791 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6792 HeapWord* addr = (HeapWord*)obj;
6793 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
6794 // Oop lies in _span and isn't yet grey or black
6795 // We read the global_finger (volatile read) strictly after marking oop
6796 bool res = _bit_map->par_mark(addr); // now grey
6797 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
6798 // Should we push this marked oop on our stack?
6799 // -- if someone else marked it, nothing to do
6800 // -- if target oop is above global finger nothing to do
6801 // -- if target oop is in chunk and above local finger
6802 // then nothing to do
6803 // -- else push on work queue
6804 if ( !res // someone else marked it, they will deal with it
6805 || (addr >= *gfa) // will be scanned in a later task
6806 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
6807 return;
6808 }
6809 // the bit map iteration has already either passed, or
6810 // sampled, this bit in the bit map; we'll need to
6811 // use the marking stack to scan this oop's oops.
6812 bool simulate_overflow = false;
6813 NOT_PRODUCT(
6814 if (CMSMarkStackOverflowALot &&
6815 _collector->simulate_overflow()) {
6816 // simulate a stack overflow
6817 simulate_overflow = true;
6818 }
6819 )
6820 if (simulate_overflow ||
6821 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
6822 // stack overflow
6823 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
6824 // We cannot assert that the overflow stack is full because
6825 // it may have been emptied since.
6826 assert(simulate_overflow ||
6827 _work_queue->size() == _work_queue->max_elems(),
6828 "Else push should have succeeded");
6829 handle_stack_overflow(addr);
6830 }
6831 do_yield_check();
6832 }
6833 }
6834
6835 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); }
6836 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); }
6837
6838 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
6839 MemRegion span,
6840 ReferenceProcessor* rp,
6841 CMSBitMap* bit_map,
6842 CMSBitMap* mod_union_table,
6843 CMSMarkStack* mark_stack,
6844 bool concurrent_precleaning):
6845 MetadataAwareOopClosure(rp),
6846 _collector(collector),
6847 _span(span),
6848 _bit_map(bit_map),
6849 _mod_union_table(mod_union_table),
6850 _mark_stack(mark_stack),
6851 _concurrent_precleaning(concurrent_precleaning)
6852 {
6853 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL");
6854 }
6855
6856 // Grey object rescan during pre-cleaning and second checkpoint phases --
6857 // the non-parallel version (the parallel version appears further below.)
6858 void PushAndMarkClosure::do_oop(oop obj) {
6859 // Ignore mark word verification. If during concurrent precleaning,
6860 // the object monitor may be locked. If during the checkpoint
6861 // phases, the object may already have been reached by a different
6862 // path and may be at the end of the global overflow list (so
6863 // the mark word may be NULL).
6864 assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */),
6865 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6866 HeapWord* addr = (HeapWord*)obj;
6867 // Check if oop points into the CMS generation
6868 // and is not marked
6869 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6870 // a white object ...
6871 _bit_map->mark(addr); // ... now grey
6872 // push on the marking stack (grey set)
6873 bool simulate_overflow = false;
6874 NOT_PRODUCT(
6875 if (CMSMarkStackOverflowALot &&
6876 _collector->simulate_overflow()) {
6877 // simulate a stack overflow
6878 simulate_overflow = true;
6879 }
6880 )
6881 if (simulate_overflow || !_mark_stack->push(obj)) {
6882 if (_concurrent_precleaning) {
6883 // During precleaning we can just dirty the appropriate card(s)
6884 // in the mod union table, thus ensuring that the object remains
6885 // in the grey set and continue. In the case of object arrays
6886 // we need to dirty all of the cards that the object spans,
6887 // since the rescan of object arrays will be limited to the
6888 // dirty cards.
6889 // Note that no one can be interfering with us in this action
6890 // of dirtying the mod union table, so no locking or atomics
6891 // are required.
6892 if (obj->is_objArray()) {
6893 size_t sz = obj->size();
6894 HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size);
6895 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6896 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6897 _mod_union_table->mark_range(redirty_range);
6898 } else {
6899 _mod_union_table->mark(addr);
6900 }
6901 _collector->_ser_pmc_preclean_ovflw++;
6902 } else {
6903 // During the remark phase, we need to remember this oop
6904 // in the overflow list.
6905 _collector->push_on_overflow_list(obj);
6906 _collector->_ser_pmc_remark_ovflw++;
6907 }
6908 }
6909 }
6910 }
6911
6912 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector,
6913 MemRegion span,
6914 ReferenceProcessor* rp,
6915 CMSBitMap* bit_map,
6916 OopTaskQueue* work_queue):
6917 MetadataAwareOopClosure(rp),
6918 _collector(collector),
6919 _span(span),
6920 _bit_map(bit_map),
6921 _work_queue(work_queue)
6922 {
6923 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL");
6924 }
6925
6926 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
6927 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
6928
6929 // Grey object rescan during second checkpoint phase --
6930 // the parallel version.
6931 void ParPushAndMarkClosure::do_oop(oop obj) {
6932 // In the assert below, we ignore the mark word because
6933 // this oop may point to an already visited object that is
6934 // on the overflow stack (in which case the mark word has
6935 // been hijacked for chaining into the overflow stack --
6936 // if this is the last object in the overflow stack then
6937 // its mark word will be NULL). Because this object may
6938 // have been subsequently popped off the global overflow
6939 // stack, and the mark word possibly restored to the prototypical
6940 // value, by the time we get to examined this failing assert in
6941 // the debugger, is_oop_or_null(false) may subsequently start
6942 // to hold.
6943 assert(oopDesc::is_oop_or_null(obj, true),
6944 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6945 HeapWord* addr = (HeapWord*)obj;
6946 // Check if oop points into the CMS generation
6947 // and is not marked
6948 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6949 // a white object ...
6950 // If we manage to "claim" the object, by being the
6951 // first thread to mark it, then we push it on our
6952 // marking stack
6953 if (_bit_map->par_mark(addr)) { // ... now grey
6954 // push on work queue (grey set)
6955 bool simulate_overflow = false;
6956 NOT_PRODUCT(
6957 if (CMSMarkStackOverflowALot &&
6958 _collector->par_simulate_overflow()) {
6959 // simulate a stack overflow
6960 simulate_overflow = true;
6961 }
6962 )
6963 if (simulate_overflow || !_work_queue->push(obj)) {
6964 _collector->par_push_on_overflow_list(obj);
6965 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
6966 }
6967 } // Else, some other thread got there first
6968 }
6969 }
6970
6971 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); }
6972 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); }
6973
6974 void CMSPrecleanRefsYieldClosure::do_yield_work() {
6975 Mutex* bml = _collector->bitMapLock();
6976 assert_lock_strong(bml);
6977 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6978 "CMS thread should hold CMS token");
6979
6980 bml->unlock();
6981 ConcurrentMarkSweepThread::desynchronize(true);
6982
6983 _collector->stopTimer();
6984 _collector->incrementYields();
6985
6986 // See the comment in coordinator_yield()
6987 for (unsigned i = 0; i < CMSYieldSleepCount &&
6988 ConcurrentMarkSweepThread::should_yield() &&
6989 !CMSCollector::foregroundGCIsActive(); ++i) {
6990 os::sleep(Thread::current(), 1, false);
6991 }
6992
6993 ConcurrentMarkSweepThread::synchronize(true);
6994 bml->lock();
6995
6996 _collector->startTimer();
6997 }
6998
6999 bool CMSPrecleanRefsYieldClosure::should_return() {
7000 if (ConcurrentMarkSweepThread::should_yield()) {
7001 do_yield_work();
7002 }
7003 return _collector->foregroundGCIsActive();
7004 }
7005
7006 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7007 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7008 "mr should be aligned to start at a card boundary");
7009 // We'd like to assert:
7010 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7011 // "mr should be a range of cards");
7012 // However, that would be too strong in one case -- the last
7013 // partition ends at _unallocated_block which, in general, can be
7014 // an arbitrary boundary, not necessarily card aligned.
7015 _num_dirty_cards += mr.word_size()/CardTableModRefBS::card_size_in_words;
7016 _space->object_iterate_mem(mr, &_scan_cl);
7017 }
7018
7019 SweepClosure::SweepClosure(CMSCollector* collector,
7020 ConcurrentMarkSweepGeneration* g,
7021 CMSBitMap* bitMap, bool should_yield) :
7022 _collector(collector),
7023 _g(g),
7024 _sp(g->cmsSpace()),
7025 _limit(_sp->sweep_limit()),
7026 _freelistLock(_sp->freelistLock()),
7027 _bitMap(bitMap),
7028 _yield(should_yield),
7029 _inFreeRange(false), // No free range at beginning of sweep
7030 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7031 _lastFreeRangeCoalesced(false),
7032 _freeFinger(g->used_region().start())
7033 {
7034 NOT_PRODUCT(
7035 _numObjectsFreed = 0;
7036 _numWordsFreed = 0;
7037 _numObjectsLive = 0;
7038 _numWordsLive = 0;
7039 _numObjectsAlreadyFree = 0;
7040 _numWordsAlreadyFree = 0;
7041 _last_fc = NULL;
7042
7043 _sp->initializeIndexedFreeListArrayReturnedBytes();
7044 _sp->dictionary()->initialize_dict_returned_bytes();
7045 )
7046 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7047 "sweep _limit out of bounds");
7048 log_develop_trace(gc, sweep)("====================");
7049 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit));
7050 }
7051
7052 void SweepClosure::print_on(outputStream* st) const {
7053 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7054 p2i(_sp->bottom()), p2i(_sp->end()));
7055 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7056 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7057 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7058 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7059 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7060 }
7061
7062 #ifndef PRODUCT
7063 // Assertion checking only: no useful work in product mode --
7064 // however, if any of the flags below become product flags,
7065 // you may need to review this code to see if it needs to be
7066 // enabled in product mode.
7067 SweepClosure::~SweepClosure() {
7068 assert_lock_strong(_freelistLock);
7069 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7070 "sweep _limit out of bounds");
7071 if (inFreeRange()) {
7072 Log(gc, sweep) log;
7073 log.error("inFreeRange() should have been reset; dumping state of SweepClosure");
7074 ResourceMark rm;
7075 LogStream ls(log.error());
7076 print_on(&ls);
7077 ShouldNotReachHere();
7078 }
7079
7080 if (log_is_enabled(Debug, gc, sweep)) {
7081 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7082 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7083 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7084 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7085 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord);
7086 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes);
7087 }
7088
7089 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) {
7090 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7091 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7092 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7093 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes",
7094 returned_bytes, indexListReturnedBytes, dict_returned_bytes);
7095 }
7096 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit));
7097 log_develop_trace(gc, sweep)("================");
7098 }
7099 #endif // PRODUCT
7100
7101 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7102 bool freeRangeInFreeLists) {
7103 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)",
7104 p2i(freeFinger), freeRangeInFreeLists);
7105 assert(!inFreeRange(), "Trampling existing free range");
7106 set_inFreeRange(true);
7107 set_lastFreeRangeCoalesced(false);
7108
7109 set_freeFinger(freeFinger);
7110 set_freeRangeInFreeLists(freeRangeInFreeLists);
7111 if (CMSTestInFreeList) {
7112 if (freeRangeInFreeLists) {
7113 FreeChunk* fc = (FreeChunk*) freeFinger;
7114 assert(fc->is_free(), "A chunk on the free list should be free.");
7115 assert(fc->size() > 0, "Free range should have a size");
7116 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7117 }
7118 }
7119 }
7120
7121 // Note that the sweeper runs concurrently with mutators. Thus,
7122 // it is possible for direct allocation in this generation to happen
7123 // in the middle of the sweep. Note that the sweeper also coalesces
7124 // contiguous free blocks. Thus, unless the sweeper and the allocator
7125 // synchronize appropriately freshly allocated blocks may get swept up.
7126 // This is accomplished by the sweeper locking the free lists while
7127 // it is sweeping. Thus blocks that are determined to be free are
7128 // indeed free. There is however one additional complication:
7129 // blocks that have been allocated since the final checkpoint and
7130 // mark, will not have been marked and so would be treated as
7131 // unreachable and swept up. To prevent this, the allocator marks
7132 // the bit map when allocating during the sweep phase. This leads,
7133 // however, to a further complication -- objects may have been allocated
7134 // but not yet initialized -- in the sense that the header isn't yet
7135 // installed. The sweeper can not then determine the size of the block
7136 // in order to skip over it. To deal with this case, we use a technique
7137 // (due to Printezis) to encode such uninitialized block sizes in the
7138 // bit map. Since the bit map uses a bit per every HeapWord, but the
7139 // CMS generation has a minimum object size of 3 HeapWords, it follows
7140 // that "normal marks" won't be adjacent in the bit map (there will
7141 // always be at least two 0 bits between successive 1 bits). We make use
7142 // of these "unused" bits to represent uninitialized blocks -- the bit
7143 // corresponding to the start of the uninitialized object and the next
7144 // bit are both set. Finally, a 1 bit marks the end of the object that
7145 // started with the two consecutive 1 bits to indicate its potentially
7146 // uninitialized state.
7147
7148 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7149 FreeChunk* fc = (FreeChunk*)addr;
7150 size_t res;
7151
7152 // Check if we are done sweeping. Below we check "addr >= _limit" rather
7153 // than "addr == _limit" because although _limit was a block boundary when
7154 // we started the sweep, it may no longer be one because heap expansion
7155 // may have caused us to coalesce the block ending at the address _limit
7156 // with a newly expanded chunk (this happens when _limit was set to the
7157 // previous _end of the space), so we may have stepped past _limit:
7158 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7159 if (addr >= _limit) { // we have swept up to or past the limit: finish up
7160 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7161 "sweep _limit out of bounds");
7162 assert(addr < _sp->end(), "addr out of bounds");
7163 // Flush any free range we might be holding as a single
7164 // coalesced chunk to the appropriate free list.
7165 if (inFreeRange()) {
7166 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7167 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger()));
7168 flush_cur_free_chunk(freeFinger(),
7169 pointer_delta(addr, freeFinger()));
7170 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]",
7171 p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7172 lastFreeRangeCoalesced() ? 1 : 0);
7173 }
7174
7175 // help the iterator loop finish
7176 return pointer_delta(_sp->end(), addr);
7177 }
7178
7179 assert(addr < _limit, "sweep invariant");
7180 // check if we should yield
7181 do_yield_check(addr);
7182 if (fc->is_free()) {
7183 // Chunk that is already free
7184 res = fc->size();
7185 do_already_free_chunk(fc);
7186 debug_only(_sp->verifyFreeLists());
7187 // If we flush the chunk at hand in lookahead_and_flush()
7188 // and it's coalesced with a preceding chunk, then the
7189 // process of "mangling" the payload of the coalesced block
7190 // will cause erasure of the size information from the
7191 // (erstwhile) header of all the coalesced blocks but the
7192 // first, so the first disjunct in the assert will not hold
7193 // in that specific case (in which case the second disjunct
7194 // will hold).
7195 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7196 "Otherwise the size info doesn't change at this step");
7197 NOT_PRODUCT(
7198 _numObjectsAlreadyFree++;
7199 _numWordsAlreadyFree += res;
7200 )
7201 NOT_PRODUCT(_last_fc = fc;)
7202 } else if (!_bitMap->isMarked(addr)) {
7203 // Chunk is fresh garbage
7204 res = do_garbage_chunk(fc);
7205 debug_only(_sp->verifyFreeLists());
7206 NOT_PRODUCT(
7207 _numObjectsFreed++;
7208 _numWordsFreed += res;
7209 )
7210 } else {
7211 // Chunk that is alive.
7212 res = do_live_chunk(fc);
7213 debug_only(_sp->verifyFreeLists());
7214 NOT_PRODUCT(
7215 _numObjectsLive++;
7216 _numWordsLive += res;
7217 )
7218 }
7219 return res;
7220 }
7221
7222 // For the smart allocation, record following
7223 // split deaths - a free chunk is removed from its free list because
7224 // it is being split into two or more chunks.
7225 // split birth - a free chunk is being added to its free list because
7226 // a larger free chunk has been split and resulted in this free chunk.
7227 // coal death - a free chunk is being removed from its free list because
7228 // it is being coalesced into a large free chunk.
7229 // coal birth - a free chunk is being added to its free list because
7230 // it was created when two or more free chunks where coalesced into
7231 // this free chunk.
7232 //
7233 // These statistics are used to determine the desired number of free
7234 // chunks of a given size. The desired number is chosen to be relative
7235 // to the end of a CMS sweep. The desired number at the end of a sweep
7236 // is the
7237 // count-at-end-of-previous-sweep (an amount that was enough)
7238 // - count-at-beginning-of-current-sweep (the excess)
7239 // + split-births (gains in this size during interval)
7240 // - split-deaths (demands on this size during interval)
7241 // where the interval is from the end of one sweep to the end of the
7242 // next.
7243 //
7244 // When sweeping the sweeper maintains an accumulated chunk which is
7245 // the chunk that is made up of chunks that have been coalesced. That
7246 // will be termed the left-hand chunk. A new chunk of garbage that
7247 // is being considered for coalescing will be referred to as the
7248 // right-hand chunk.
7249 //
7250 // When making a decision on whether to coalesce a right-hand chunk with
7251 // the current left-hand chunk, the current count vs. the desired count
7252 // of the left-hand chunk is considered. Also if the right-hand chunk
7253 // is near the large chunk at the end of the heap (see
7254 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7255 // left-hand chunk is coalesced.
7256 //
7257 // When making a decision about whether to split a chunk, the desired count
7258 // vs. the current count of the candidate to be split is also considered.
7259 // If the candidate is underpopulated (currently fewer chunks than desired)
7260 // a chunk of an overpopulated (currently more chunks than desired) size may
7261 // be chosen. The "hint" associated with a free list, if non-null, points
7262 // to a free list which may be overpopulated.
7263 //
7264
7265 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7266 const size_t size = fc->size();
7267 // Chunks that cannot be coalesced are not in the
7268 // free lists.
7269 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7270 assert(_sp->verify_chunk_in_free_list(fc),
7271 "free chunk should be in free lists");
7272 }
7273 // a chunk that is already free, should not have been
7274 // marked in the bit map
7275 HeapWord* const addr = (HeapWord*) fc;
7276 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7277 // Verify that the bit map has no bits marked between
7278 // addr and purported end of this block.
7279 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7280
7281 // Some chunks cannot be coalesced under any circumstances.
7282 // See the definition of cantCoalesce().
7283 if (!fc->cantCoalesce()) {
7284 // This chunk can potentially be coalesced.
7285 // All the work is done in
7286 do_post_free_or_garbage_chunk(fc, size);
7287 // Note that if the chunk is not coalescable (the else arm
7288 // below), we unconditionally flush, without needing to do
7289 // a "lookahead," as we do below.
7290 if (inFreeRange()) lookahead_and_flush(fc, size);
7291 } else {
7292 // Code path common to both original and adaptive free lists.
7293
7294 // cant coalesce with previous block; this should be treated
7295 // as the end of a free run if any
7296 if (inFreeRange()) {
7297 // we kicked some butt; time to pick up the garbage
7298 assert(freeFinger() < addr, "freeFinger points too high");
7299 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7300 }
7301 // else, nothing to do, just continue
7302 }
7303 }
7304
7305 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7306 // This is a chunk of garbage. It is not in any free list.
7307 // Add it to a free list or let it possibly be coalesced into
7308 // a larger chunk.
7309 HeapWord* const addr = (HeapWord*) fc;
7310 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7311
7312 // Verify that the bit map has no bits marked between
7313 // addr and purported end of just dead object.
7314 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7315 do_post_free_or_garbage_chunk(fc, size);
7316
7317 assert(_limit >= addr + size,
7318 "A freshly garbage chunk can't possibly straddle over _limit");
7319 if (inFreeRange()) lookahead_and_flush(fc, size);
7320 return size;
7321 }
7322
7323 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7324 HeapWord* addr = (HeapWord*) fc;
7325 // The sweeper has just found a live object. Return any accumulated
7326 // left hand chunk to the free lists.
7327 if (inFreeRange()) {
7328 assert(freeFinger() < addr, "freeFinger points too high");
7329 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7330 }
7331
7332 // This object is live: we'd normally expect this to be
7333 // an oop, and like to assert the following:
7334 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop");
7335 // However, as we commented above, this may be an object whose
7336 // header hasn't yet been initialized.
7337 size_t size;
7338 assert(_bitMap->isMarked(addr), "Tautology for this control point");
7339 if (_bitMap->isMarked(addr + 1)) {
7340 // Determine the size from the bit map, rather than trying to
7341 // compute it from the object header.
7342 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7343 size = pointer_delta(nextOneAddr + 1, addr);
7344 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7345 "alignment problem");
7346
7347 #ifdef ASSERT
7348 if (oop(addr)->klass_or_null_acquire() != NULL) {
7349 // Ignore mark word because we are running concurrent with mutators
7350 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop");
7351 assert(size ==
7352 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7353 "P-mark and computed size do not agree");
7354 }
7355 #endif
7356
7357 } else {
7358 // This should be an initialized object that's alive.
7359 assert(oop(addr)->klass_or_null_acquire() != NULL,
7360 "Should be an initialized object");
7361 // Ignore mark word because we are running concurrent with mutators
7362 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop");
7363 // Verify that the bit map has no bits marked between
7364 // addr and purported end of this block.
7365 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7366 assert(size >= 3, "Necessary for Printezis marks to work");
7367 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7368 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7369 }
7370 return size;
7371 }
7372
7373 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7374 size_t chunkSize) {
7375 // do_post_free_or_garbage_chunk() should only be called in the case
7376 // of the adaptive free list allocator.
7377 const bool fcInFreeLists = fc->is_free();
7378 assert((HeapWord*)fc <= _limit, "sweep invariant");
7379 if (CMSTestInFreeList && fcInFreeLists) {
7380 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7381 }
7382
7383 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7384
7385 HeapWord* const fc_addr = (HeapWord*) fc;
7386
7387 bool coalesce = false;
7388 const size_t left = pointer_delta(fc_addr, freeFinger());
7389 const size_t right = chunkSize;
7390 switch (FLSCoalescePolicy) {
7391 // numeric value forms a coalition aggressiveness metric
7392 case 0: { // never coalesce
7393 coalesce = false;
7394 break;
7395 }
7396 case 1: { // coalesce if left & right chunks on overpopulated lists
7397 coalesce = _sp->coalOverPopulated(left) &&
7398 _sp->coalOverPopulated(right);
7399 break;
7400 }
7401 case 2: { // coalesce if left chunk on overpopulated list (default)
7402 coalesce = _sp->coalOverPopulated(left);
7403 break;
7404 }
7405 case 3: { // coalesce if left OR right chunk on overpopulated list
7406 coalesce = _sp->coalOverPopulated(left) ||
7407 _sp->coalOverPopulated(right);
7408 break;
7409 }
7410 case 4: { // always coalesce
7411 coalesce = true;
7412 break;
7413 }
7414 default:
7415 ShouldNotReachHere();
7416 }
7417
7418 // Should the current free range be coalesced?
7419 // If the chunk is in a free range and either we decided to coalesce above
7420 // or the chunk is near the large block at the end of the heap
7421 // (isNearLargestChunk() returns true), then coalesce this chunk.
7422 const bool doCoalesce = inFreeRange()
7423 && (coalesce || _g->isNearLargestChunk(fc_addr));
7424 if (doCoalesce) {
7425 // Coalesce the current free range on the left with the new
7426 // chunk on the right. If either is on a free list,
7427 // it must be removed from the list and stashed in the closure.
7428 if (freeRangeInFreeLists()) {
7429 FreeChunk* const ffc = (FreeChunk*)freeFinger();
7430 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7431 "Size of free range is inconsistent with chunk size.");
7432 if (CMSTestInFreeList) {
7433 assert(_sp->verify_chunk_in_free_list(ffc),
7434 "Chunk is not in free lists");
7435 }
7436 _sp->coalDeath(ffc->size());
7437 _sp->removeFreeChunkFromFreeLists(ffc);
7438 set_freeRangeInFreeLists(false);
7439 }
7440 if (fcInFreeLists) {
7441 _sp->coalDeath(chunkSize);
7442 assert(fc->size() == chunkSize,
7443 "The chunk has the wrong size or is not in the free lists");
7444 _sp->removeFreeChunkFromFreeLists(fc);
7445 }
7446 set_lastFreeRangeCoalesced(true);
7447 print_free_block_coalesced(fc);
7448 } else { // not in a free range and/or should not coalesce
7449 // Return the current free range and start a new one.
7450 if (inFreeRange()) {
7451 // In a free range but cannot coalesce with the right hand chunk.
7452 // Put the current free range into the free lists.
7453 flush_cur_free_chunk(freeFinger(),
7454 pointer_delta(fc_addr, freeFinger()));
7455 }
7456 // Set up for new free range. Pass along whether the right hand
7457 // chunk is in the free lists.
7458 initialize_free_range((HeapWord*)fc, fcInFreeLists);
7459 }
7460 }
7461
7462 // Lookahead flush:
7463 // If we are tracking a free range, and this is the last chunk that
7464 // we'll look at because its end crosses past _limit, we'll preemptively
7465 // flush it along with any free range we may be holding on to. Note that
7466 // this can be the case only for an already free or freshly garbage
7467 // chunk. If this block is an object, it can never straddle
7468 // over _limit. The "straddling" occurs when _limit is set at
7469 // the previous end of the space when this cycle started, and
7470 // a subsequent heap expansion caused the previously co-terminal
7471 // free block to be coalesced with the newly expanded portion,
7472 // thus rendering _limit a non-block-boundary making it dangerous
7473 // for the sweeper to step over and examine.
7474 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7475 assert(inFreeRange(), "Should only be called if currently in a free range.");
7476 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7477 assert(_sp->used_region().contains(eob - 1),
7478 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7479 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7480 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7481 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size);
7482 if (eob >= _limit) {
7483 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7484 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block "
7485 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7486 "[" PTR_FORMAT "," PTR_FORMAT ")",
7487 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7488 // Return the storage we are tracking back into the free lists.
7489 log_develop_trace(gc, sweep)("Flushing ... ");
7490 assert(freeFinger() < eob, "Error");
7491 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7492 }
7493 }
7494
7495 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7496 assert(inFreeRange(), "Should only be called if currently in a free range.");
7497 assert(size > 0,
7498 "A zero sized chunk cannot be added to the free lists.");
7499 if (!freeRangeInFreeLists()) {
7500 if (CMSTestInFreeList) {
7501 FreeChunk* fc = (FreeChunk*) chunk;
7502 fc->set_size(size);
7503 assert(!_sp->verify_chunk_in_free_list(fc),
7504 "chunk should not be in free lists yet");
7505 }
7506 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size);
7507 // A new free range is going to be starting. The current
7508 // free range has not been added to the free lists yet or
7509 // was removed so add it back.
7510 // If the current free range was coalesced, then the death
7511 // of the free range was recorded. Record a birth now.
7512 if (lastFreeRangeCoalesced()) {
7513 _sp->coalBirth(size);
7514 }
7515 _sp->addChunkAndRepairOffsetTable(chunk, size,
7516 lastFreeRangeCoalesced());
7517 } else {
7518 log_develop_trace(gc, sweep)("Already in free list: nothing to flush");
7519 }
7520 set_inFreeRange(false);
7521 set_freeRangeInFreeLists(false);
7522 }
7523
7524 // We take a break if we've been at this for a while,
7525 // so as to avoid monopolizing the locks involved.
7526 void SweepClosure::do_yield_work(HeapWord* addr) {
7527 // Return current free chunk being used for coalescing (if any)
7528 // to the appropriate freelist. After yielding, the next
7529 // free block encountered will start a coalescing range of
7530 // free blocks. If the next free block is adjacent to the
7531 // chunk just flushed, they will need to wait for the next
7532 // sweep to be coalesced.
7533 if (inFreeRange()) {
7534 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7535 }
7536
7537 // First give up the locks, then yield, then re-lock.
7538 // We should probably use a constructor/destructor idiom to
7539 // do this unlock/lock or modify the MutexUnlocker class to
7540 // serve our purpose. XXX
7541 assert_lock_strong(_bitMap->lock());
7542 assert_lock_strong(_freelistLock);
7543 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7544 "CMS thread should hold CMS token");
7545 _bitMap->lock()->unlock();
7546 _freelistLock->unlock();
7547 ConcurrentMarkSweepThread::desynchronize(true);
7548 _collector->stopTimer();
7549 _collector->incrementYields();
7550
7551 // See the comment in coordinator_yield()
7552 for (unsigned i = 0; i < CMSYieldSleepCount &&
7553 ConcurrentMarkSweepThread::should_yield() &&
7554 !CMSCollector::foregroundGCIsActive(); ++i) {
7555 os::sleep(Thread::current(), 1, false);
7556 }
7557
7558 ConcurrentMarkSweepThread::synchronize(true);
7559 _freelistLock->lock();
7560 _bitMap->lock()->lock_without_safepoint_check();
7561 _collector->startTimer();
7562 }
7563
7564 #ifndef PRODUCT
7565 // This is actually very useful in a product build if it can
7566 // be called from the debugger. Compile it into the product
7567 // as needed.
7568 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7569 return debug_cms_space->verify_chunk_in_free_list(fc);
7570 }
7571 #endif
7572
7573 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7574 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7575 p2i(fc), fc->size());
7576 }
7577
7578 // CMSIsAliveClosure
7579 bool CMSIsAliveClosure::do_object_b(oop obj) {
7580 HeapWord* addr = (HeapWord*)obj;
7581 return addr != NULL &&
7582 (!_span.contains(addr) || _bit_map->isMarked(addr));
7583 }
7584
7585
7586 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7587 MemRegion span,
7588 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7589 bool cpc):
7590 _collector(collector),
7591 _span(span),
7592 _bit_map(bit_map),
7593 _mark_stack(mark_stack),
7594 _concurrent_precleaning(cpc) {
7595 assert(!_span.is_empty(), "Empty span could spell trouble");
7596 }
7597
7598
7599 // CMSKeepAliveClosure: the serial version
7600 void CMSKeepAliveClosure::do_oop(oop obj) {
7601 HeapWord* addr = (HeapWord*)obj;
7602 if (_span.contains(addr) &&
7603 !_bit_map->isMarked(addr)) {
7604 _bit_map->mark(addr);
7605 bool simulate_overflow = false;
7606 NOT_PRODUCT(
7607 if (CMSMarkStackOverflowALot &&
7608 _collector->simulate_overflow()) {
7609 // simulate a stack overflow
7610 simulate_overflow = true;
7611 }
7612 )
7613 if (simulate_overflow || !_mark_stack->push(obj)) {
7614 if (_concurrent_precleaning) {
7615 // We dirty the overflown object and let the remark
7616 // phase deal with it.
7617 assert(_collector->overflow_list_is_empty(), "Error");
7618 // In the case of object arrays, we need to dirty all of
7619 // the cards that the object spans. No locking or atomics
7620 // are needed since no one else can be mutating the mod union
7621 // table.
7622 if (obj->is_objArray()) {
7623 size_t sz = obj->size();
7624 HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size);
7625 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7626 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7627 _collector->_modUnionTable.mark_range(redirty_range);
7628 } else {
7629 _collector->_modUnionTable.mark(addr);
7630 }
7631 _collector->_ser_kac_preclean_ovflw++;
7632 } else {
7633 _collector->push_on_overflow_list(obj);
7634 _collector->_ser_kac_ovflw++;
7635 }
7636 }
7637 }
7638 }
7639
7640 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7641 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7642
7643 // CMSParKeepAliveClosure: a parallel version of the above.
7644 // The work queues are private to each closure (thread),
7645 // but (may be) available for stealing by other threads.
7646 void CMSParKeepAliveClosure::do_oop(oop obj) {
7647 HeapWord* addr = (HeapWord*)obj;
7648 if (_span.contains(addr) &&
7649 !_bit_map->isMarked(addr)) {
7650 // In general, during recursive tracing, several threads
7651 // may be concurrently getting here; the first one to
7652 // "tag" it, claims it.
7653 if (_bit_map->par_mark(addr)) {
7654 bool res = _work_queue->push(obj);
7655 assert(res, "Low water mark should be much less than capacity");
7656 // Do a recursive trim in the hope that this will keep
7657 // stack usage lower, but leave some oops for potential stealers
7658 trim_queue(_low_water_mark);
7659 } // Else, another thread got there first
7660 }
7661 }
7662
7663 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7664 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7665
7666 void CMSParKeepAliveClosure::trim_queue(uint max) {
7667 while (_work_queue->size() > max) {
7668 oop new_oop;
7669 if (_work_queue->pop_local(new_oop)) {
7670 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
7671 assert(_bit_map->isMarked((HeapWord*)new_oop),
7672 "no white objects on this stack!");
7673 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7674 // iterate over the oops in this oop, marking and pushing
7675 // the ones in CMS heap (i.e. in _span).
7676 new_oop->oop_iterate(&_mark_and_push);
7677 }
7678 }
7679 }
7680
7681 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
7682 CMSCollector* collector,
7683 MemRegion span, CMSBitMap* bit_map,
7684 OopTaskQueue* work_queue):
7685 _collector(collector),
7686 _span(span),
7687 _bit_map(bit_map),
7688 _work_queue(work_queue) { }
7689
7690 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
7691 HeapWord* addr = (HeapWord*)obj;
7692 if (_span.contains(addr) &&
7693 !_bit_map->isMarked(addr)) {
7694 if (_bit_map->par_mark(addr)) {
7695 bool simulate_overflow = false;
7696 NOT_PRODUCT(
7697 if (CMSMarkStackOverflowALot &&
7698 _collector->par_simulate_overflow()) {
7699 // simulate a stack overflow
7700 simulate_overflow = true;
7701 }
7702 )
7703 if (simulate_overflow || !_work_queue->push(obj)) {
7704 _collector->par_push_on_overflow_list(obj);
7705 _collector->_par_kac_ovflw++;
7706 }
7707 } // Else another thread got there already
7708 }
7709 }
7710
7711 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7712 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7713
7714 //////////////////////////////////////////////////////////////////
7715 // CMSExpansionCause /////////////////////////////
7716 //////////////////////////////////////////////////////////////////
7717 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
7718 switch (cause) {
7719 case _no_expansion:
7720 return "No expansion";
7721 case _satisfy_free_ratio:
7722 return "Free ratio";
7723 case _satisfy_promotion:
7724 return "Satisfy promotion";
7725 case _satisfy_allocation:
7726 return "allocation";
7727 case _allocate_par_lab:
7728 return "Par LAB";
7729 case _allocate_par_spooling_space:
7730 return "Par Spooling Space";
7731 case _adaptive_size_policy:
7732 return "Ergonomics";
7733 default:
7734 return "unknown";
7735 }
7736 }
7737
7738 void CMSDrainMarkingStackClosure::do_void() {
7739 // the max number to take from overflow list at a time
7740 const size_t num = _mark_stack->capacity()/4;
7741 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
7742 "Overflow list should be NULL during concurrent phases");
7743 while (!_mark_stack->isEmpty() ||
7744 // if stack is empty, check the overflow list
7745 _collector->take_from_overflow_list(num, _mark_stack)) {
7746 oop obj = _mark_stack->pop();
7747 HeapWord* addr = (HeapWord*)obj;
7748 assert(_span.contains(addr), "Should be within span");
7749 assert(_bit_map->isMarked(addr), "Should be marked");
7750 assert(oopDesc::is_oop(obj), "Should be an oop");
7751 obj->oop_iterate(_keep_alive);
7752 }
7753 }
7754
7755 void CMSParDrainMarkingStackClosure::do_void() {
7756 // drain queue
7757 trim_queue(0);
7758 }
7759
7760 // Trim our work_queue so its length is below max at return
7761 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
7762 while (_work_queue->size() > max) {
7763 oop new_oop;
7764 if (_work_queue->pop_local(new_oop)) {
7765 assert(oopDesc::is_oop(new_oop), "Expected an oop");
7766 assert(_bit_map->isMarked((HeapWord*)new_oop),
7767 "no white objects on this stack!");
7768 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7769 // iterate over the oops in this oop, marking and pushing
7770 // the ones in CMS heap (i.e. in _span).
7771 new_oop->oop_iterate(&_mark_and_push);
7772 }
7773 }
7774 }
7775
7776 ////////////////////////////////////////////////////////////////////
7777 // Support for Marking Stack Overflow list handling and related code
7778 ////////////////////////////////////////////////////////////////////
7779 // Much of the following code is similar in shape and spirit to the
7780 // code used in ParNewGC. We should try and share that code
7781 // as much as possible in the future.
7782
7783 #ifndef PRODUCT
7784 // Debugging support for CMSStackOverflowALot
7785
7786 // It's OK to call this multi-threaded; the worst thing
7787 // that can happen is that we'll get a bunch of closely
7788 // spaced simulated overflows, but that's OK, in fact
7789 // probably good as it would exercise the overflow code
7790 // under contention.
7791 bool CMSCollector::simulate_overflow() {
7792 if (_overflow_counter-- <= 0) { // just being defensive
7793 _overflow_counter = CMSMarkStackOverflowInterval;
7794 return true;
7795 } else {
7796 return false;
7797 }
7798 }
7799
7800 bool CMSCollector::par_simulate_overflow() {
7801 return simulate_overflow();
7802 }
7803 #endif
7804
7805 // Single-threaded
7806 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
7807 assert(stack->isEmpty(), "Expected precondition");
7808 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
7809 size_t i = num;
7810 oop cur = _overflow_list;
7811 const markOop proto = markOopDesc::prototype();
7812 NOT_PRODUCT(ssize_t n = 0;)
7813 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
7814 next = oop(cur->mark());
7815 cur->set_mark(proto); // until proven otherwise
7816 assert(oopDesc::is_oop(cur), "Should be an oop");
7817 bool res = stack->push(cur);
7818 assert(res, "Bit off more than can chew?");
7819 NOT_PRODUCT(n++;)
7820 }
7821 _overflow_list = cur;
7822 #ifndef PRODUCT
7823 assert(_num_par_pushes >= n, "Too many pops?");
7824 _num_par_pushes -=n;
7825 #endif
7826 return !stack->isEmpty();
7827 }
7828
7829 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
7830 // (MT-safe) Get a prefix of at most "num" from the list.
7831 // The overflow list is chained through the mark word of
7832 // each object in the list. We fetch the entire list,
7833 // break off a prefix of the right size and return the
7834 // remainder. If other threads try to take objects from
7835 // the overflow list at that time, they will wait for
7836 // some time to see if data becomes available. If (and
7837 // only if) another thread places one or more object(s)
7838 // on the global list before we have returned the suffix
7839 // to the global list, we will walk down our local list
7840 // to find its end and append the global list to
7841 // our suffix before returning it. This suffix walk can
7842 // prove to be expensive (quadratic in the amount of traffic)
7843 // when there are many objects in the overflow list and
7844 // there is much producer-consumer contention on the list.
7845 // *NOTE*: The overflow list manipulation code here and
7846 // in ParNewGeneration:: are very similar in shape,
7847 // except that in the ParNew case we use the old (from/eden)
7848 // copy of the object to thread the list via its klass word.
7849 // Because of the common code, if you make any changes in
7850 // the code below, please check the ParNew version to see if
7851 // similar changes might be needed.
7852 // CR 6797058 has been filed to consolidate the common code.
7853 bool CMSCollector::par_take_from_overflow_list(size_t num,
7854 OopTaskQueue* work_q,
7855 int no_of_gc_threads) {
7856 assert(work_q->size() == 0, "First empty local work queue");
7857 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
7858 if (_overflow_list == NULL) {
7859 return false;
7860 }
7861 // Grab the entire list; we'll put back a suffix
7862 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
7863 Thread* tid = Thread::current();
7864 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
7865 // set to ParallelGCThreads.
7866 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
7867 size_t sleep_time_millis = MAX2((size_t)1, num/100);
7868 // If the list is busy, we spin for a short while,
7869 // sleeping between attempts to get the list.
7870 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
7871 os::sleep(tid, sleep_time_millis, false);
7872 if (_overflow_list == NULL) {
7873 // Nothing left to take
7874 return false;
7875 } else if (_overflow_list != BUSY) {
7876 // Try and grab the prefix
7877 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
7878 }
7879 }
7880 // If the list was found to be empty, or we spun long
7881 // enough, we give up and return empty-handed. If we leave
7882 // the list in the BUSY state below, it must be the case that
7883 // some other thread holds the overflow list and will set it
7884 // to a non-BUSY state in the future.
7885 if (prefix == NULL || prefix == BUSY) {
7886 // Nothing to take or waited long enough
7887 if (prefix == NULL) {
7888 // Write back the NULL in case we overwrote it with BUSY above
7889 // and it is still the same value.
7890 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
7891 }
7892 return false;
7893 }
7894 assert(prefix != NULL && prefix != BUSY, "Error");
7895 size_t i = num;
7896 oop cur = prefix;
7897 // Walk down the first "num" objects, unless we reach the end.
7898 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
7899 if (cur->mark() == NULL) {
7900 // We have "num" or fewer elements in the list, so there
7901 // is nothing to return to the global list.
7902 // Write back the NULL in lieu of the BUSY we wrote
7903 // above, if it is still the same value.
7904 if (_overflow_list == BUSY) {
7905 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
7906 }
7907 } else {
7908 // Chop off the suffix and return it to the global list.
7909 assert(cur->mark() != BUSY, "Error");
7910 oop suffix_head = cur->mark(); // suffix will be put back on global list
7911 cur->set_mark(NULL); // break off suffix
7912 // It's possible that the list is still in the empty(busy) state
7913 // we left it in a short while ago; in that case we may be
7914 // able to place back the suffix without incurring the cost
7915 // of a walk down the list.
7916 oop observed_overflow_list = _overflow_list;
7917 oop cur_overflow_list = observed_overflow_list;
7918 bool attached = false;
7919 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
7920 observed_overflow_list =
7921 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list);
7922 if (cur_overflow_list == observed_overflow_list) {
7923 attached = true;
7924 break;
7925 } else cur_overflow_list = observed_overflow_list;
7926 }
7927 if (!attached) {
7928 // Too bad, someone else sneaked in (at least) an element; we'll need
7929 // to do a splice. Find tail of suffix so we can prepend suffix to global
7930 // list.
7931 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
7932 oop suffix_tail = cur;
7933 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
7934 "Tautology");
7935 observed_overflow_list = _overflow_list;
7936 do {
7937 cur_overflow_list = observed_overflow_list;
7938 if (cur_overflow_list != BUSY) {
7939 // Do the splice ...
7940 suffix_tail->set_mark(markOop(cur_overflow_list));
7941 } else { // cur_overflow_list == BUSY
7942 suffix_tail->set_mark(NULL);
7943 }
7944 // ... and try to place spliced list back on overflow_list ...
7945 observed_overflow_list =
7946 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list);
7947 } while (cur_overflow_list != observed_overflow_list);
7948 // ... until we have succeeded in doing so.
7949 }
7950 }
7951
7952 // Push the prefix elements on work_q
7953 assert(prefix != NULL, "control point invariant");
7954 const markOop proto = markOopDesc::prototype();
7955 oop next;
7956 NOT_PRODUCT(ssize_t n = 0;)
7957 for (cur = prefix; cur != NULL; cur = next) {
7958 next = oop(cur->mark());
7959 cur->set_mark(proto); // until proven otherwise
7960 assert(oopDesc::is_oop(cur), "Should be an oop");
7961 bool res = work_q->push(cur);
7962 assert(res, "Bit off more than we can chew?");
7963 NOT_PRODUCT(n++;)
7964 }
7965 #ifndef PRODUCT
7966 assert(_num_par_pushes >= n, "Too many pops?");
7967 Atomic::sub(n, &_num_par_pushes);
7968 #endif
7969 return true;
7970 }
7971
7972 // Single-threaded
7973 void CMSCollector::push_on_overflow_list(oop p) {
7974 NOT_PRODUCT(_num_par_pushes++;)
7975 assert(oopDesc::is_oop(p), "Not an oop");
7976 preserve_mark_if_necessary(p);
7977 p->set_mark((markOop)_overflow_list);
7978 _overflow_list = p;
7979 }
7980
7981 // Multi-threaded; use CAS to prepend to overflow list
7982 void CMSCollector::par_push_on_overflow_list(oop p) {
7983 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);)
7984 assert(oopDesc::is_oop(p), "Not an oop");
7985 par_preserve_mark_if_necessary(p);
7986 oop observed_overflow_list = _overflow_list;
7987 oop cur_overflow_list;
7988 do {
7989 cur_overflow_list = observed_overflow_list;
7990 if (cur_overflow_list != BUSY) {
7991 p->set_mark(markOop(cur_overflow_list));
7992 } else {
7993 p->set_mark(NULL);
7994 }
7995 observed_overflow_list =
7996 Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list);
7997 } while (cur_overflow_list != observed_overflow_list);
7998 }
7999 #undef BUSY
8000
8001 // Single threaded
8002 // General Note on GrowableArray: pushes may silently fail
8003 // because we are (temporarily) out of C-heap for expanding
8004 // the stack. The problem is quite ubiquitous and affects
8005 // a lot of code in the JVM. The prudent thing for GrowableArray
8006 // to do (for now) is to exit with an error. However, that may
8007 // be too draconian in some cases because the caller may be
8008 // able to recover without much harm. For such cases, we
8009 // should probably introduce a "soft_push" method which returns
8010 // an indication of success or failure with the assumption that
8011 // the caller may be able to recover from a failure; code in
8012 // the VM can then be changed, incrementally, to deal with such
8013 // failures where possible, thus, incrementally hardening the VM
8014 // in such low resource situations.
8015 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8016 _preserved_oop_stack.push(p);
8017 _preserved_mark_stack.push(m);
8018 assert(m == p->mark(), "Mark word changed");
8019 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8020 "bijection");
8021 }
8022
8023 // Single threaded
8024 void CMSCollector::preserve_mark_if_necessary(oop p) {
8025 markOop m = p->mark();
8026 if (m->must_be_preserved(p)) {
8027 preserve_mark_work(p, m);
8028 }
8029 }
8030
8031 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8032 markOop m = p->mark();
8033 if (m->must_be_preserved(p)) {
8034 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8035 // Even though we read the mark word without holding
8036 // the lock, we are assured that it will not change
8037 // because we "own" this oop, so no other thread can
8038 // be trying to push it on the overflow list; see
8039 // the assertion in preserve_mark_work() that checks
8040 // that m == p->mark().
8041 preserve_mark_work(p, m);
8042 }
8043 }
8044
8045 // We should be able to do this multi-threaded,
8046 // a chunk of stack being a task (this is
8047 // correct because each oop only ever appears
8048 // once in the overflow list. However, it's
8049 // not very easy to completely overlap this with
8050 // other operations, so will generally not be done
8051 // until all work's been completed. Because we
8052 // expect the preserved oop stack (set) to be small,
8053 // it's probably fine to do this single-threaded.
8054 // We can explore cleverer concurrent/overlapped/parallel
8055 // processing of preserved marks if we feel the
8056 // need for this in the future. Stack overflow should
8057 // be so rare in practice and, when it happens, its
8058 // effect on performance so great that this will
8059 // likely just be in the noise anyway.
8060 void CMSCollector::restore_preserved_marks_if_any() {
8061 assert(SafepointSynchronize::is_at_safepoint(),
8062 "world should be stopped");
8063 assert(Thread::current()->is_ConcurrentGC_thread() ||
8064 Thread::current()->is_VM_thread(),
8065 "should be single-threaded");
8066 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8067 "bijection");
8068
8069 while (!_preserved_oop_stack.is_empty()) {
8070 oop p = _preserved_oop_stack.pop();
8071 assert(oopDesc::is_oop(p), "Should be an oop");
8072 assert(_span.contains(p), "oop should be in _span");
8073 assert(p->mark() == markOopDesc::prototype(),
8074 "Set when taken from overflow list");
8075 markOop m = _preserved_mark_stack.pop();
8076 p->set_mark(m);
8077 }
8078 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8079 "stacks were cleared above");
8080 }
8081
8082 #ifndef PRODUCT
8083 bool CMSCollector::no_preserved_marks() const {
8084 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8085 }
8086 #endif
8087
8088 // Transfer some number of overflown objects to usual marking
8089 // stack. Return true if some objects were transferred.
8090 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8091 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8092 (size_t)ParGCDesiredObjsFromOverflowList);
8093
8094 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8095 assert(_collector->overflow_list_is_empty() || res,
8096 "If list is not empty, we should have taken something");
8097 assert(!res || !_mark_stack->isEmpty(),
8098 "If we took something, it should now be on our stack");
8099 return res;
8100 }
8101
8102 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8103 size_t res = _sp->block_size_no_stall(addr, _collector);
8104 if (_sp->block_is_obj(addr)) {
8105 if (_live_bit_map->isMarked(addr)) {
8106 // It can't have been dead in a previous cycle
8107 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8108 } else {
8109 _dead_bit_map->mark(addr); // mark the dead object
8110 }
8111 }
8112 // Could be 0, if the block size could not be computed without stalling.
8113 return res;
8114 }
8115
8116 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8117 GCMemoryManager* manager = CMSHeap::heap()->old_manager();
8118 switch (phase) {
8119 case CMSCollector::InitialMarking:
8120 initialize(manager /* GC manager */ ,
8121 cause /* cause of the GC */,
8122 true /* recordGCBeginTime */,
8123 true /* recordPreGCUsage */,
8124 false /* recordPeakUsage */,
8125 false /* recordPostGCusage */,
8126 true /* recordAccumulatedGCTime */,
8127 false /* recordGCEndTime */,
8128 false /* countCollection */ );
8129 break;
8130
8131 case CMSCollector::FinalMarking:
8132 initialize(manager /* GC manager */ ,
8133 cause /* cause of the GC */,
8134 false /* recordGCBeginTime */,
8135 false /* recordPreGCUsage */,
8136 false /* recordPeakUsage */,
8137 false /* recordPostGCusage */,
8138 true /* recordAccumulatedGCTime */,
8139 false /* recordGCEndTime */,
8140 false /* countCollection */ );
8141 break;
8142
8143 case CMSCollector::Sweeping:
8144 initialize(manager /* GC manager */ ,
8145 cause /* cause of the GC */,
8146 false /* recordGCBeginTime */,
8147 false /* recordPreGCUsage */,
8148 true /* recordPeakUsage */,
8149 true /* recordPostGCusage */,
8150 false /* recordAccumulatedGCTime */,
8151 true /* recordGCEndTime */,
8152 true /* countCollection */ );
8153 break;
8154
8155 default:
8156 ShouldNotReachHere();
8157 }
8158 }