1 /*
2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/classLoaderData.hpp"
27 #include "classfile/stringTable.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "code/codeCache.hpp"
30 #include "gc_implementation/shared/adaptiveSizePolicy.hpp"
31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp"
32 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp"
33 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp"
34 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp"
35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp"
36 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp"
37 #include "gc_implementation/parNew/parNewGeneration.hpp"
38 #include "gc_implementation/shared/collectorCounters.hpp"
39 #include "gc_implementation/shared/gcTimer.hpp"
40 #include "gc_implementation/shared/gcTrace.hpp"
41 #include "gc_implementation/shared/gcTraceTime.hpp"
42 #include "gc_implementation/shared/isGCActiveMark.hpp"
43 #include "gc_interface/collectedHeap.inline.hpp"
44 #include "memory/allocation.hpp"
45 #include "memory/cardGeneration.inline.hpp"
46 #include "memory/cardTableRS.hpp"
47 #include "memory/collectorPolicy.hpp"
48 #include "memory/gcLocker.inline.hpp"
49 #include "memory/genCollectedHeap.hpp"
50 #include "memory/genMarkSweep.hpp"
51 #include "memory/genOopClosures.inline.hpp"
52 #include "memory/iterator.inline.hpp"
53 #include "memory/padded.hpp"
54 #include "memory/referencePolicy.hpp"
55 #include "memory/resourceArea.hpp"
56 #include "memory/strongRootsScope.hpp"
57 #include "memory/tenuredGeneration.hpp"
58 #include "oops/oop.inline.hpp"
59 #include "prims/jvmtiExport.hpp"
60 #include "runtime/atomic.inline.hpp"
61 #include "runtime/globals_extension.hpp"
62 #include "runtime/handles.inline.hpp"
63 #include "runtime/java.hpp"
64 #include "runtime/orderAccess.inline.hpp"
65 #include "runtime/vmThread.hpp"
66 #include "services/memoryService.hpp"
67 #include "services/runtimeService.hpp"
68 #include "utilities/stack.inline.hpp"
69 #include "utilities/taskqueue.inline.hpp"
70
71 // statics
72 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
73 bool CMSCollector::_full_gc_requested = false;
74 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
75
76 //////////////////////////////////////////////////////////////////
77 // In support of CMS/VM thread synchronization
78 //////////////////////////////////////////////////////////////////
79 // We split use of the CGC_lock into 2 "levels".
80 // The low-level locking is of the usual CGC_lock monitor. We introduce
81 // a higher level "token" (hereafter "CMS token") built on top of the
82 // low level monitor (hereafter "CGC lock").
83 // The token-passing protocol gives priority to the VM thread. The
84 // CMS-lock doesn't provide any fairness guarantees, but clients
85 // should ensure that it is only held for very short, bounded
86 // durations.
87 //
88 // When either of the CMS thread or the VM thread is involved in
89 // collection operations during which it does not want the other
90 // thread to interfere, it obtains the CMS token.
91 //
92 // If either thread tries to get the token while the other has
93 // it, that thread waits. However, if the VM thread and CMS thread
94 // both want the token, then the VM thread gets priority while the
95 // CMS thread waits. This ensures, for instance, that the "concurrent"
96 // phases of the CMS thread's work do not block out the VM thread
97 // for long periods of time as the CMS thread continues to hog
98 // the token. (See bug 4616232).
99 //
100 // The baton-passing functions are, however, controlled by the
101 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
102 // and here the low-level CMS lock, not the high level token,
103 // ensures mutual exclusion.
104 //
105 // Two important conditions that we have to satisfy:
106 // 1. if a thread does a low-level wait on the CMS lock, then it
107 // relinquishes the CMS token if it were holding that token
108 // when it acquired the low-level CMS lock.
109 // 2. any low-level notifications on the low-level lock
110 // should only be sent when a thread has relinquished the token.
111 //
112 // In the absence of either property, we'd have potential deadlock.
113 //
114 // We protect each of the CMS (concurrent and sequential) phases
115 // with the CMS _token_, not the CMS _lock_.
116 //
117 // The only code protected by CMS lock is the token acquisition code
118 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
119 // baton-passing code.
120 //
121 // Unfortunately, i couldn't come up with a good abstraction to factor and
122 // hide the naked CGC_lock manipulation in the baton-passing code
123 // further below. That's something we should try to do. Also, the proof
124 // of correctness of this 2-level locking scheme is far from obvious,
125 // and potentially quite slippery. We have an uneasy suspicion, for instance,
126 // that there may be a theoretical possibility of delay/starvation in the
127 // low-level lock/wait/notify scheme used for the baton-passing because of
128 // potential interference with the priority scheme embodied in the
129 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
130 // invocation further below and marked with "XXX 20011219YSR".
131 // Indeed, as we note elsewhere, this may become yet more slippery
132 // in the presence of multiple CMS and/or multiple VM threads. XXX
133
134 class CMSTokenSync: public StackObj {
135 private:
136 bool _is_cms_thread;
137 public:
138 CMSTokenSync(bool is_cms_thread):
139 _is_cms_thread(is_cms_thread) {
140 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
141 "Incorrect argument to constructor");
142 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
143 }
144
145 ~CMSTokenSync() {
146 assert(_is_cms_thread ?
147 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
148 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
149 "Incorrect state");
150 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
151 }
152 };
153
154 // Convenience class that does a CMSTokenSync, and then acquires
155 // upto three locks.
156 class CMSTokenSyncWithLocks: public CMSTokenSync {
157 private:
158 // Note: locks are acquired in textual declaration order
159 // and released in the opposite order
160 MutexLockerEx _locker1, _locker2, _locker3;
161 public:
162 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
163 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
164 CMSTokenSync(is_cms_thread),
165 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
166 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
167 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
168 { }
169 };
170
171
172 //////////////////////////////////////////////////////////////////
173 // Concurrent Mark-Sweep Generation /////////////////////////////
174 //////////////////////////////////////////////////////////////////
175
176 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
177
178 // This struct contains per-thread things necessary to support parallel
179 // young-gen collection.
180 class CMSParGCThreadState: public CHeapObj<mtGC> {
181 public:
182 CFLS_LAB lab;
183 PromotionInfo promo;
184
185 // Constructor.
186 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
187 promo.setSpace(cfls);
188 }
189 };
190
191 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
192 ReservedSpace rs, size_t initial_byte_size, int level,
193 CardTableRS* ct, bool use_adaptive_freelists,
194 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) :
195 CardGeneration(rs, initial_byte_size, level, ct),
196 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
197 _did_compact(false)
198 {
199 HeapWord* bottom = (HeapWord*) _virtual_space.low();
200 HeapWord* end = (HeapWord*) _virtual_space.high();
201
202 _direct_allocated_words = 0;
203 NOT_PRODUCT(
204 _numObjectsPromoted = 0;
205 _numWordsPromoted = 0;
206 _numObjectsAllocated = 0;
207 _numWordsAllocated = 0;
208 )
209
210 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
211 use_adaptive_freelists,
212 dictionaryChoice);
213 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
214 _cmsSpace->_gen = this;
215
216 _gc_stats = new CMSGCStats();
217
218 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
219 // offsets match. The ability to tell free chunks from objects
220 // depends on this property.
221 debug_only(
222 FreeChunk* junk = NULL;
223 assert(UseCompressedClassPointers ||
224 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
225 "Offset of FreeChunk::_prev within FreeChunk must match"
226 " that of OopDesc::_klass within OopDesc");
227 )
228
229 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC);
230 for (uint i = 0; i < ParallelGCThreads; i++) {
231 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
232 }
233
234 _incremental_collection_failed = false;
235 // The "dilatation_factor" is the expansion that can occur on
236 // account of the fact that the minimum object size in the CMS
237 // generation may be larger than that in, say, a contiguous young
238 // generation.
239 // Ideally, in the calculation below, we'd compute the dilatation
240 // factor as: MinChunkSize/(promoting_gen's min object size)
241 // Since we do not have such a general query interface for the
242 // promoting generation, we'll instead just use the minimum
243 // object size (which today is a header's worth of space);
244 // note that all arithmetic is in units of HeapWords.
245 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
246 assert(_dilatation_factor >= 1.0, "from previous assert");
247 }
248
249
250 // The field "_initiating_occupancy" represents the occupancy percentage
251 // at which we trigger a new collection cycle. Unless explicitly specified
252 // via CMSInitiatingOccupancyFraction (argument "io" below), it
253 // is calculated by:
254 //
255 // Let "f" be MinHeapFreeRatio in
256 //
257 // _initiating_occupancy = 100-f +
258 // f * (CMSTriggerRatio/100)
259 // where CMSTriggerRatio is the argument "tr" below.
260 //
261 // That is, if we assume the heap is at its desired maximum occupancy at the
262 // end of a collection, we let CMSTriggerRatio of the (purported) free
263 // space be allocated before initiating a new collection cycle.
264 //
265 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
266 assert(io <= 100 && tr <= 100, "Check the arguments");
267 if (io >= 0) {
268 _initiating_occupancy = (double)io / 100.0;
269 } else {
270 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
271 (double)(tr * MinHeapFreeRatio) / 100.0)
272 / 100.0;
273 }
274 }
275
276 void ConcurrentMarkSweepGeneration::ref_processor_init() {
277 assert(collector() != NULL, "no collector");
278 collector()->ref_processor_init();
279 }
280
281 void CMSCollector::ref_processor_init() {
282 if (_ref_processor == NULL) {
283 // Allocate and initialize a reference processor
284 _ref_processor =
285 new ReferenceProcessor(_span, // span
286 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
287 (int) ParallelGCThreads, // mt processing degree
288 _cmsGen->refs_discovery_is_mt(), // mt discovery
289 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
290 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
291 &_is_alive_closure); // closure for liveness info
292 // Initialize the _ref_processor field of CMSGen
293 _cmsGen->set_ref_processor(_ref_processor);
294
295 }
296 }
297
298 AdaptiveSizePolicy* CMSCollector::size_policy() {
299 GenCollectedHeap* gch = GenCollectedHeap::heap();
300 return gch->gen_policy()->size_policy();
301 }
302
303 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
304
305 const char* gen_name = "old";
306 GenCollectorPolicy* gcp = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy();
307
308 // Generation Counters - generation 1, 1 subspace
309 _gen_counters = new GenerationCounters(gen_name, 1, 1,
310 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
311
312 _space_counters = new GSpaceCounters(gen_name, 0,
313 _virtual_space.reserved_size(),
314 this, _gen_counters);
315 }
316
317 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
318 _cms_gen(cms_gen)
319 {
320 assert(alpha <= 100, "bad value");
321 _saved_alpha = alpha;
322
323 // Initialize the alphas to the bootstrap value of 100.
324 _gc0_alpha = _cms_alpha = 100;
325
326 _cms_begin_time.update();
327 _cms_end_time.update();
328
329 _gc0_duration = 0.0;
330 _gc0_period = 0.0;
331 _gc0_promoted = 0;
332
333 _cms_duration = 0.0;
334 _cms_period = 0.0;
335 _cms_allocated = 0;
336
337 _cms_used_at_gc0_begin = 0;
338 _cms_used_at_gc0_end = 0;
339 _allow_duty_cycle_reduction = false;
340 _valid_bits = 0;
341 }
342
343 double CMSStats::cms_free_adjustment_factor(size_t free) const {
344 // TBD: CR 6909490
345 return 1.0;
346 }
347
348 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
349 }
350
351 // If promotion failure handling is on use
352 // the padded average size of the promotion for each
353 // young generation collection.
354 double CMSStats::time_until_cms_gen_full() const {
355 size_t cms_free = _cms_gen->cmsSpace()->free();
356 GenCollectedHeap* gch = GenCollectedHeap::heap();
357 size_t expected_promotion = MIN2(gch->young_gen()->capacity(),
358 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
359 if (cms_free > expected_promotion) {
360 // Start a cms collection if there isn't enough space to promote
361 // for the next minor collection. Use the padded average as
362 // a safety factor.
363 cms_free -= expected_promotion;
364
365 // Adjust by the safety factor.
366 double cms_free_dbl = (double)cms_free;
367 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
368 // Apply a further correction factor which tries to adjust
369 // for recent occurance of concurrent mode failures.
370 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
371 cms_free_dbl = cms_free_dbl * cms_adjustment;
372
373 if (PrintGCDetails && Verbose) {
374 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
375 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
376 cms_free, expected_promotion);
377 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
378 cms_free_dbl, cms_consumption_rate() + 1.0);
379 }
380 // Add 1 in case the consumption rate goes to zero.
381 return cms_free_dbl / (cms_consumption_rate() + 1.0);
382 }
383 return 0.0;
384 }
385
386 // Compare the duration of the cms collection to the
387 // time remaining before the cms generation is empty.
388 // Note that the time from the start of the cms collection
389 // to the start of the cms sweep (less than the total
390 // duration of the cms collection) can be used. This
391 // has been tried and some applications experienced
392 // promotion failures early in execution. This was
393 // possibly because the averages were not accurate
394 // enough at the beginning.
395 double CMSStats::time_until_cms_start() const {
396 // We add "gc0_period" to the "work" calculation
397 // below because this query is done (mostly) at the
398 // end of a scavenge, so we need to conservatively
399 // account for that much possible delay
400 // in the query so as to avoid concurrent mode failures
401 // due to starting the collection just a wee bit too
402 // late.
403 double work = cms_duration() + gc0_period();
404 double deadline = time_until_cms_gen_full();
405 // If a concurrent mode failure occurred recently, we want to be
406 // more conservative and halve our expected time_until_cms_gen_full()
407 if (work > deadline) {
408 if (Verbose && PrintGCDetails) {
409 gclog_or_tty->print(
410 " CMSCollector: collect because of anticipated promotion "
411 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
412 gc0_period(), time_until_cms_gen_full());
413 }
414 return 0.0;
415 }
416 return work - deadline;
417 }
418
419 #ifndef PRODUCT
420 void CMSStats::print_on(outputStream *st) const {
421 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
422 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
423 gc0_duration(), gc0_period(), gc0_promoted());
424 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
425 cms_duration(), cms_period(), cms_allocated());
426 st->print(",cms_since_beg=%g,cms_since_end=%g",
427 cms_time_since_begin(), cms_time_since_end());
428 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
429 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
430
431 if (valid()) {
432 st->print(",promo_rate=%g,cms_alloc_rate=%g",
433 promotion_rate(), cms_allocation_rate());
434 st->print(",cms_consumption_rate=%g,time_until_full=%g",
435 cms_consumption_rate(), time_until_cms_gen_full());
436 }
437 st->print(" ");
438 }
439 #endif // #ifndef PRODUCT
440
441 CMSCollector::CollectorState CMSCollector::_collectorState =
442 CMSCollector::Idling;
443 bool CMSCollector::_foregroundGCIsActive = false;
444 bool CMSCollector::_foregroundGCShouldWait = false;
445
446 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
447 CardTableRS* ct,
448 ConcurrentMarkSweepPolicy* cp):
449 _cmsGen(cmsGen),
450 _ct(ct),
451 _ref_processor(NULL), // will be set later
452 _conc_workers(NULL), // may be set later
453 _abort_preclean(false),
454 _start_sampling(false),
455 _between_prologue_and_epilogue(false),
456 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
457 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
458 -1 /* lock-free */, "No_lock" /* dummy */),
459 _modUnionClosurePar(&_modUnionTable),
460 // Adjust my span to cover old (cms) gen
461 _span(cmsGen->reserved()),
462 // Construct the is_alive_closure with _span & markBitMap
463 _is_alive_closure(_span, &_markBitMap),
464 _restart_addr(NULL),
465 _overflow_list(NULL),
466 _stats(cmsGen),
467 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
468 //verify that this lock should be acquired with safepoint check.
469 Monitor::_safepoint_check_sometimes)),
470 _eden_chunk_array(NULL), // may be set in ctor body
471 _eden_chunk_capacity(0), // -- ditto --
472 _eden_chunk_index(0), // -- ditto --
473 _survivor_plab_array(NULL), // -- ditto --
474 _survivor_chunk_array(NULL), // -- ditto --
475 _survivor_chunk_capacity(0), // -- ditto --
476 _survivor_chunk_index(0), // -- ditto --
477 _ser_pmc_preclean_ovflw(0),
478 _ser_kac_preclean_ovflw(0),
479 _ser_pmc_remark_ovflw(0),
480 _par_pmc_remark_ovflw(0),
481 _ser_kac_ovflw(0),
482 _par_kac_ovflw(0),
483 #ifndef PRODUCT
484 _num_par_pushes(0),
485 #endif
486 _collection_count_start(0),
487 _verifying(false),
488 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
489 _completed_initialization(false),
490 _collector_policy(cp),
491 _should_unload_classes(CMSClassUnloadingEnabled),
492 _concurrent_cycles_since_last_unload(0),
493 _roots_scanning_options(GenCollectedHeap::SO_None),
494 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
495 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
496 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
497 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
498 _cms_start_registered(false)
499 {
500 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
501 ExplicitGCInvokesConcurrent = true;
502 }
503 // Now expand the span and allocate the collection support structures
504 // (MUT, marking bit map etc.) to cover both generations subject to
505 // collection.
506
507 // For use by dirty card to oop closures.
508 _cmsGen->cmsSpace()->set_collector(this);
509
510 // Allocate MUT and marking bit map
511 {
512 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
513 if (!_markBitMap.allocate(_span)) {
514 warning("Failed to allocate CMS Bit Map");
515 return;
516 }
517 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
518 }
519 {
520 _modUnionTable.allocate(_span);
521 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
522 }
523
524 if (!_markStack.allocate(MarkStackSize)) {
525 warning("Failed to allocate CMS Marking Stack");
526 return;
527 }
528
529 // Support for multi-threaded concurrent phases
530 if (CMSConcurrentMTEnabled) {
531 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
532 // just for now
533 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
534 }
535 if (ConcGCThreads > 1) {
536 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
537 ConcGCThreads, true);
538 if (_conc_workers == NULL) {
539 warning("GC/CMS: _conc_workers allocation failure: "
540 "forcing -CMSConcurrentMTEnabled");
541 CMSConcurrentMTEnabled = false;
542 } else {
543 _conc_workers->initialize_workers();
544 }
545 } else {
546 CMSConcurrentMTEnabled = false;
547 }
548 }
549 if (!CMSConcurrentMTEnabled) {
550 ConcGCThreads = 0;
551 } else {
552 // Turn off CMSCleanOnEnter optimization temporarily for
553 // the MT case where it's not fixed yet; see 6178663.
554 CMSCleanOnEnter = false;
555 }
556 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
557 "Inconsistency");
558
559 // Parallel task queues; these are shared for the
560 // concurrent and stop-world phases of CMS, but
561 // are not shared with parallel scavenge (ParNew).
562 {
563 uint i;
564 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
565
566 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
567 || ParallelRefProcEnabled)
568 && num_queues > 0) {
569 _task_queues = new OopTaskQueueSet(num_queues);
570 if (_task_queues == NULL) {
571 warning("task_queues allocation failure.");
572 return;
573 }
574 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
575 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
576 for (i = 0; i < num_queues; i++) {
577 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
578 if (q == NULL) {
579 warning("work_queue allocation failure.");
580 return;
581 }
582 _task_queues->register_queue(i, q);
583 }
584 for (i = 0; i < num_queues; i++) {
585 _task_queues->queue(i)->initialize();
586 _hash_seed[i] = 17; // copied from ParNew
587 }
588 }
589 }
590
591 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
592
593 // Clip CMSBootstrapOccupancy between 0 and 100.
594 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100;
595
596 // Now tell CMS generations the identity of their collector
597 ConcurrentMarkSweepGeneration::set_collector(this);
598
599 // Create & start a CMS thread for this CMS collector
600 _cmsThread = ConcurrentMarkSweepThread::start(this);
601 assert(cmsThread() != NULL, "CMS Thread should have been created");
602 assert(cmsThread()->collector() == this,
603 "CMS Thread should refer to this gen");
604 assert(CGC_lock != NULL, "Where's the CGC_lock?");
605
606 // Support for parallelizing young gen rescan
607 GenCollectedHeap* gch = GenCollectedHeap::heap();
608 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
609 _young_gen = (ParNewGeneration*)gch->young_gen();
610 if (gch->supports_inline_contig_alloc()) {
611 _top_addr = gch->top_addr();
612 _end_addr = gch->end_addr();
613 assert(_young_gen != NULL, "no _young_gen");
614 _eden_chunk_index = 0;
615 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
616 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
617 }
618
619 // Support for parallelizing survivor space rescan
620 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
621 const size_t max_plab_samples =
622 ((DefNewGeneration*)_young_gen)->max_survivor_size() / plab_sample_minimum_size();
623
624 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
625 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC);
626 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
627 _survivor_chunk_capacity = 2*max_plab_samples;
628 for (uint i = 0; i < ParallelGCThreads; i++) {
629 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
630 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
631 assert(cur->end() == 0, "Should be 0");
632 assert(cur->array() == vec, "Should be vec");
633 assert(cur->capacity() == max_plab_samples, "Error");
634 }
635 }
636
637 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
638 _gc_counters = new CollectorCounters("CMS", 1);
639 _completed_initialization = true;
640 _inter_sweep_timer.start(); // start of time
641 }
642
643 size_t CMSCollector::plab_sample_minimum_size() {
644 // The default value of MinTLABSize is 2k, but there is
645 // no way to get the default value if the flag has been overridden.
646 return MAX2(ThreadLocalAllocBuffer::min_size() * HeapWordSize, 2 * K);
647 }
648
649 const char* ConcurrentMarkSweepGeneration::name() const {
650 return "concurrent mark-sweep generation";
651 }
652 void ConcurrentMarkSweepGeneration::update_counters() {
653 if (UsePerfData) {
654 _space_counters->update_all();
655 _gen_counters->update_all();
656 }
657 }
658
659 // this is an optimized version of update_counters(). it takes the
660 // used value as a parameter rather than computing it.
661 //
662 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
663 if (UsePerfData) {
664 _space_counters->update_used(used);
665 _space_counters->update_capacity();
666 _gen_counters->update_all();
667 }
668 }
669
670 void ConcurrentMarkSweepGeneration::print() const {
671 Generation::print();
672 cmsSpace()->print();
673 }
674
675 #ifndef PRODUCT
676 void ConcurrentMarkSweepGeneration::print_statistics() {
677 cmsSpace()->printFLCensus(0);
678 }
679 #endif
680
681 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
682 GenCollectedHeap* gch = GenCollectedHeap::heap();
683 if (PrintGCDetails) {
684 if (Verbose) {
685 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
686 level(), short_name(), s, used(), capacity());
687 } else {
688 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
689 level(), short_name(), s, used() / K, capacity() / K);
690 }
691 }
692 if (Verbose) {
693 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
694 gch->used(), gch->capacity());
695 } else {
696 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
697 gch->used() / K, gch->capacity() / K);
698 }
699 }
700
701 size_t
702 ConcurrentMarkSweepGeneration::contiguous_available() const {
703 // dld proposes an improvement in precision here. If the committed
704 // part of the space ends in a free block we should add that to
705 // uncommitted size in the calculation below. Will make this
706 // change later, staying with the approximation below for the
707 // time being. -- ysr.
708 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
709 }
710
711 size_t
712 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
713 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
714 }
715
716 size_t ConcurrentMarkSweepGeneration::max_available() const {
717 return free() + _virtual_space.uncommitted_size();
718 }
719
720 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
721 size_t available = max_available();
722 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
723 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
724 if (Verbose && PrintGCDetails) {
725 gclog_or_tty->print_cr(
726 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
727 "max_promo("SIZE_FORMAT")",
728 res? "":" not", available, res? ">=":"<",
729 av_promo, max_promotion_in_bytes);
730 }
731 return res;
732 }
733
734 // At a promotion failure dump information on block layout in heap
735 // (cms old generation).
736 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
737 if (CMSDumpAtPromotionFailure) {
738 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
739 }
740 }
741
742 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
743 // Clear the promotion information. These pointers can be adjusted
744 // along with all the other pointers into the heap but
745 // compaction is expected to be a rare event with
746 // a heap using cms so don't do it without seeing the need.
747 for (uint i = 0; i < ParallelGCThreads; i++) {
748 _par_gc_thread_states[i]->promo.reset();
749 }
750 }
751
752 void ConcurrentMarkSweepGeneration::compute_new_size() {
753 assert_locked_or_safepoint(Heap_lock);
754
755 // If incremental collection failed, we just want to expand
756 // to the limit.
757 if (incremental_collection_failed()) {
758 clear_incremental_collection_failed();
759 grow_to_reserved();
760 return;
761 }
762
763 // The heap has been compacted but not reset yet.
764 // Any metric such as free() or used() will be incorrect.
765
766 CardGeneration::compute_new_size();
767
768 // Reset again after a possible resizing
769 if (did_compact()) {
770 cmsSpace()->reset_after_compaction();
771 }
772 }
773
774 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
775 assert_locked_or_safepoint(Heap_lock);
776
777 // If incremental collection failed, we just want to expand
778 // to the limit.
779 if (incremental_collection_failed()) {
780 clear_incremental_collection_failed();
781 grow_to_reserved();
782 return;
783 }
784
785 double free_percentage = ((double) free()) / capacity();
786 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
787 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
788
789 // compute expansion delta needed for reaching desired free percentage
790 if (free_percentage < desired_free_percentage) {
791 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
792 assert(desired_capacity >= capacity(), "invalid expansion size");
793 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
794 if (PrintGCDetails && Verbose) {
795 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
796 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
797 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
798 gclog_or_tty->print_cr(" Desired free fraction %f",
799 desired_free_percentage);
800 gclog_or_tty->print_cr(" Maximum free fraction %f",
801 maximum_free_percentage);
802 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000);
803 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
804 desired_capacity/1000);
805 int prev_level = level() - 1;
806 if (prev_level >= 0) {
807 size_t prev_size = 0;
808 GenCollectedHeap* gch = GenCollectedHeap::heap();
809 Generation* prev_gen = gch->young_gen();
810 prev_size = prev_gen->capacity();
811 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
812 prev_size/1000);
813 }
814 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
815 unsafe_max_alloc_nogc()/1000);
816 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
817 contiguous_available()/1000);
818 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
819 expand_bytes);
820 }
821 // safe if expansion fails
822 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
823 if (PrintGCDetails && Verbose) {
824 gclog_or_tty->print_cr(" Expanded free fraction %f",
825 ((double) free()) / capacity());
826 }
827 } else {
828 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
829 assert(desired_capacity <= capacity(), "invalid expansion size");
830 size_t shrink_bytes = capacity() - desired_capacity;
831 // Don't shrink unless the delta is greater than the minimum shrink we want
832 if (shrink_bytes >= MinHeapDeltaBytes) {
833 shrink_free_list_by(shrink_bytes);
834 }
835 }
836 }
837
838 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
839 return cmsSpace()->freelistLock();
840 }
841
842 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
843 bool tlab) {
844 CMSSynchronousYieldRequest yr;
845 MutexLockerEx x(freelistLock(),
846 Mutex::_no_safepoint_check_flag);
847 return have_lock_and_allocate(size, tlab);
848 }
849
850 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
851 bool tlab /* ignored */) {
852 assert_lock_strong(freelistLock());
853 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
854 HeapWord* res = cmsSpace()->allocate(adjustedSize);
855 // Allocate the object live (grey) if the background collector has
856 // started marking. This is necessary because the marker may
857 // have passed this address and consequently this object will
858 // not otherwise be greyed and would be incorrectly swept up.
859 // Note that if this object contains references, the writing
860 // of those references will dirty the card containing this object
861 // allowing the object to be blackened (and its references scanned)
862 // either during a preclean phase or at the final checkpoint.
863 if (res != NULL) {
864 // We may block here with an uninitialized object with
865 // its mark-bit or P-bits not yet set. Such objects need
866 // to be safely navigable by block_start().
867 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
868 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
869 collector()->direct_allocated(res, adjustedSize);
870 _direct_allocated_words += adjustedSize;
871 // allocation counters
872 NOT_PRODUCT(
873 _numObjectsAllocated++;
874 _numWordsAllocated += (int)adjustedSize;
875 )
876 }
877 return res;
878 }
879
880 // In the case of direct allocation by mutators in a generation that
881 // is being concurrently collected, the object must be allocated
882 // live (grey) if the background collector has started marking.
883 // This is necessary because the marker may
884 // have passed this address and consequently this object will
885 // not otherwise be greyed and would be incorrectly swept up.
886 // Note that if this object contains references, the writing
887 // of those references will dirty the card containing this object
888 // allowing the object to be blackened (and its references scanned)
889 // either during a preclean phase or at the final checkpoint.
890 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
891 assert(_markBitMap.covers(start, size), "Out of bounds");
892 if (_collectorState >= Marking) {
893 MutexLockerEx y(_markBitMap.lock(),
894 Mutex::_no_safepoint_check_flag);
895 // [see comments preceding SweepClosure::do_blk() below for details]
896 //
897 // Can the P-bits be deleted now? JJJ
898 //
899 // 1. need to mark the object as live so it isn't collected
900 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
901 // 3. need to mark the end of the object so marking, precleaning or sweeping
902 // can skip over uninitialized or unparsable objects. An allocated
903 // object is considered uninitialized for our purposes as long as
904 // its klass word is NULL. All old gen objects are parsable
905 // as soon as they are initialized.)
906 _markBitMap.mark(start); // object is live
907 _markBitMap.mark(start + 1); // object is potentially uninitialized?
908 _markBitMap.mark(start + size - 1);
909 // mark end of object
910 }
911 // check that oop looks uninitialized
912 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
913 }
914
915 void CMSCollector::promoted(bool par, HeapWord* start,
916 bool is_obj_array, size_t obj_size) {
917 assert(_markBitMap.covers(start), "Out of bounds");
918 // See comment in direct_allocated() about when objects should
919 // be allocated live.
920 if (_collectorState >= Marking) {
921 // we already hold the marking bit map lock, taken in
922 // the prologue
923 if (par) {
924 _markBitMap.par_mark(start);
925 } else {
926 _markBitMap.mark(start);
927 }
928 // We don't need to mark the object as uninitialized (as
929 // in direct_allocated above) because this is being done with the
930 // world stopped and the object will be initialized by the
931 // time the marking, precleaning or sweeping get to look at it.
932 // But see the code for copying objects into the CMS generation,
933 // where we need to ensure that concurrent readers of the
934 // block offset table are able to safely navigate a block that
935 // is in flux from being free to being allocated (and in
936 // transition while being copied into) and subsequently
937 // becoming a bona-fide object when the copy/promotion is complete.
938 assert(SafepointSynchronize::is_at_safepoint(),
939 "expect promotion only at safepoints");
940
941 if (_collectorState < Sweeping) {
942 // Mark the appropriate cards in the modUnionTable, so that
943 // this object gets scanned before the sweep. If this is
944 // not done, CMS generation references in the object might
945 // not get marked.
946 // For the case of arrays, which are otherwise precisely
947 // marked, we need to dirty the entire array, not just its head.
948 if (is_obj_array) {
949 // The [par_]mark_range() method expects mr.end() below to
950 // be aligned to the granularity of a bit's representation
951 // in the heap. In the case of the MUT below, that's a
952 // card size.
953 MemRegion mr(start,
954 (HeapWord*)round_to((intptr_t)(start + obj_size),
955 CardTableModRefBS::card_size /* bytes */));
956 if (par) {
957 _modUnionTable.par_mark_range(mr);
958 } else {
959 _modUnionTable.mark_range(mr);
960 }
961 } else { // not an obj array; we can just mark the head
962 if (par) {
963 _modUnionTable.par_mark(start);
964 } else {
965 _modUnionTable.mark(start);
966 }
967 }
968 }
969 }
970 }
971
972 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
973 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
974 // allocate, copy and if necessary update promoinfo --
975 // delegate to underlying space.
976 assert_lock_strong(freelistLock());
977
978 #ifndef PRODUCT
979 if (GenCollectedHeap::heap()->promotion_should_fail()) {
980 return NULL;
981 }
982 #endif // #ifndef PRODUCT
983
984 oop res = _cmsSpace->promote(obj, obj_size);
985 if (res == NULL) {
986 // expand and retry
987 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
988 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
989 // Since this is the old generation, we don't try to promote
990 // into a more senior generation.
991 res = _cmsSpace->promote(obj, obj_size);
992 }
993 if (res != NULL) {
994 // See comment in allocate() about when objects should
995 // be allocated live.
996 assert(obj->is_oop(), "Will dereference klass pointer below");
997 collector()->promoted(false, // Not parallel
998 (HeapWord*)res, obj->is_objArray(), obj_size);
999 // promotion counters
1000 NOT_PRODUCT(
1001 _numObjectsPromoted++;
1002 _numWordsPromoted +=
1003 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1004 )
1005 }
1006 return res;
1007 }
1008
1009
1010 // IMPORTANT: Notes on object size recognition in CMS.
1011 // ---------------------------------------------------
1012 // A block of storage in the CMS generation is always in
1013 // one of three states. A free block (FREE), an allocated
1014 // object (OBJECT) whose size() method reports the correct size,
1015 // and an intermediate state (TRANSIENT) in which its size cannot
1016 // be accurately determined.
1017 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1018 // -----------------------------------------------------
1019 // FREE: klass_word & 1 == 1; mark_word holds block size
1020 //
1021 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1022 // obj->size() computes correct size
1023 //
1024 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1025 //
1026 // STATE IDENTIFICATION: (64 bit+COOPS)
1027 // ------------------------------------
1028 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1029 //
1030 // OBJECT: klass_word installed; klass_word != 0;
1031 // obj->size() computes correct size
1032 //
1033 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1034 //
1035 //
1036 // STATE TRANSITION DIAGRAM
1037 //
1038 // mut / parnew mut / parnew
1039 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1040 // ^ |
1041 // |------------------------ DEAD <------------------------------------|
1042 // sweep mut
1043 //
1044 // While a block is in TRANSIENT state its size cannot be determined
1045 // so readers will either need to come back later or stall until
1046 // the size can be determined. Note that for the case of direct
1047 // allocation, P-bits, when available, may be used to determine the
1048 // size of an object that may not yet have been initialized.
1049
1050 // Things to support parallel young-gen collection.
1051 oop
1052 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1053 oop old, markOop m,
1054 size_t word_sz) {
1055 #ifndef PRODUCT
1056 if (GenCollectedHeap::heap()->promotion_should_fail()) {
1057 return NULL;
1058 }
1059 #endif // #ifndef PRODUCT
1060
1061 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1062 PromotionInfo* promoInfo = &ps->promo;
1063 // if we are tracking promotions, then first ensure space for
1064 // promotion (including spooling space for saving header if necessary).
1065 // then allocate and copy, then track promoted info if needed.
1066 // When tracking (see PromotionInfo::track()), the mark word may
1067 // be displaced and in this case restoration of the mark word
1068 // occurs in the (oop_since_save_marks_)iterate phase.
1069 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1070 // Out of space for allocating spooling buffers;
1071 // try expanding and allocating spooling buffers.
1072 if (!expand_and_ensure_spooling_space(promoInfo)) {
1073 return NULL;
1074 }
1075 }
1076 assert(promoInfo->has_spooling_space(), "Control point invariant");
1077 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1078 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1079 if (obj_ptr == NULL) {
1080 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1081 if (obj_ptr == NULL) {
1082 return NULL;
1083 }
1084 }
1085 oop obj = oop(obj_ptr);
1086 OrderAccess::storestore();
1087 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1088 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1089 // IMPORTANT: See note on object initialization for CMS above.
1090 // Otherwise, copy the object. Here we must be careful to insert the
1091 // klass pointer last, since this marks the block as an allocated object.
1092 // Except with compressed oops it's the mark word.
1093 HeapWord* old_ptr = (HeapWord*)old;
1094 // Restore the mark word copied above.
1095 obj->set_mark(m);
1096 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1097 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1098 OrderAccess::storestore();
1099
1100 if (UseCompressedClassPointers) {
1101 // Copy gap missed by (aligned) header size calculation below
1102 obj->set_klass_gap(old->klass_gap());
1103 }
1104 if (word_sz > (size_t)oopDesc::header_size()) {
1105 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1106 obj_ptr + oopDesc::header_size(),
1107 word_sz - oopDesc::header_size());
1108 }
1109
1110 // Now we can track the promoted object, if necessary. We take care
1111 // to delay the transition from uninitialized to full object
1112 // (i.e., insertion of klass pointer) until after, so that it
1113 // atomically becomes a promoted object.
1114 if (promoInfo->tracking()) {
1115 promoInfo->track((PromotedObject*)obj, old->klass());
1116 }
1117 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1118 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1119 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1120
1121 // Finally, install the klass pointer (this should be volatile).
1122 OrderAccess::storestore();
1123 obj->set_klass(old->klass());
1124 // We should now be able to calculate the right size for this object
1125 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1126
1127 collector()->promoted(true, // parallel
1128 obj_ptr, old->is_objArray(), word_sz);
1129
1130 NOT_PRODUCT(
1131 Atomic::inc_ptr(&_numObjectsPromoted);
1132 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1133 )
1134
1135 return obj;
1136 }
1137
1138 void
1139 ConcurrentMarkSweepGeneration::
1140 par_promote_alloc_done(int thread_num) {
1141 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1142 ps->lab.retire(thread_num);
1143 }
1144
1145 void
1146 ConcurrentMarkSweepGeneration::
1147 par_oop_since_save_marks_iterate_done(int thread_num) {
1148 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1149 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1150 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1151 }
1152
1153 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1154 size_t size,
1155 bool tlab)
1156 {
1157 // We allow a STW collection only if a full
1158 // collection was requested.
1159 return full || should_allocate(size, tlab); // FIX ME !!!
1160 // This and promotion failure handling are connected at the
1161 // hip and should be fixed by untying them.
1162 }
1163
1164 bool CMSCollector::shouldConcurrentCollect() {
1165 if (_full_gc_requested) {
1166 if (Verbose && PrintGCDetails) {
1167 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1168 " gc request (or gc_locker)");
1169 }
1170 return true;
1171 }
1172
1173 FreelistLocker x(this);
1174 // ------------------------------------------------------------------
1175 // Print out lots of information which affects the initiation of
1176 // a collection.
1177 if (PrintCMSInitiationStatistics && stats().valid()) {
1178 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1179 gclog_or_tty->stamp();
1180 gclog_or_tty->cr();
1181 stats().print_on(gclog_or_tty);
1182 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1183 stats().time_until_cms_gen_full());
1184 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1185 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1186 _cmsGen->contiguous_available());
1187 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1188 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1189 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1190 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1191 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1192 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1193 gclog_or_tty->print_cr("metadata initialized %d",
1194 MetaspaceGC::should_concurrent_collect());
1195 }
1196 // ------------------------------------------------------------------
1197
1198 // If the estimated time to complete a cms collection (cms_duration())
1199 // is less than the estimated time remaining until the cms generation
1200 // is full, start a collection.
1201 if (!UseCMSInitiatingOccupancyOnly) {
1202 if (stats().valid()) {
1203 if (stats().time_until_cms_start() == 0.0) {
1204 return true;
1205 }
1206 } else {
1207 // We want to conservatively collect somewhat early in order
1208 // to try and "bootstrap" our CMS/promotion statistics;
1209 // this branch will not fire after the first successful CMS
1210 // collection because the stats should then be valid.
1211 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1212 if (Verbose && PrintGCDetails) {
1213 gclog_or_tty->print_cr(
1214 " CMSCollector: collect for bootstrapping statistics:"
1215 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1216 _bootstrap_occupancy);
1217 }
1218 return true;
1219 }
1220 }
1221 }
1222
1223 // Otherwise, we start a collection cycle if
1224 // old gen want a collection cycle started. Each may use
1225 // an appropriate criterion for making this decision.
1226 // XXX We need to make sure that the gen expansion
1227 // criterion dovetails well with this. XXX NEED TO FIX THIS
1228 if (_cmsGen->should_concurrent_collect()) {
1229 if (Verbose && PrintGCDetails) {
1230 gclog_or_tty->print_cr("CMS old gen initiated");
1231 }
1232 return true;
1233 }
1234
1235 // We start a collection if we believe an incremental collection may fail;
1236 // this is not likely to be productive in practice because it's probably too
1237 // late anyway.
1238 GenCollectedHeap* gch = GenCollectedHeap::heap();
1239 assert(gch->collector_policy()->is_generation_policy(),
1240 "You may want to check the correctness of the following");
1241 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1242 if (Verbose && PrintGCDetails) {
1243 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1244 }
1245 return true;
1246 }
1247
1248 if (MetaspaceGC::should_concurrent_collect()) {
1249 if (Verbose && PrintGCDetails) {
1250 gclog_or_tty->print("CMSCollector: collect for metadata allocation ");
1251 }
1252 return true;
1253 }
1254
1255 // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1256 if (CMSTriggerInterval >= 0) {
1257 if (CMSTriggerInterval == 0) {
1258 // Trigger always
1259 return true;
1260 }
1261
1262 // Check the CMS time since begin (we do not check the stats validity
1263 // as we want to be able to trigger the first CMS cycle as well)
1264 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1265 if (Verbose && PrintGCDetails) {
1266 if (stats().valid()) {
1267 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1268 stats().cms_time_since_begin());
1269 } else {
1270 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)");
1271 }
1272 }
1273 return true;
1274 }
1275 }
1276
1277 return false;
1278 }
1279
1280 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1281
1282 // Clear _expansion_cause fields of constituent generations
1283 void CMSCollector::clear_expansion_cause() {
1284 _cmsGen->clear_expansion_cause();
1285 }
1286
1287 // We should be conservative in starting a collection cycle. To
1288 // start too eagerly runs the risk of collecting too often in the
1289 // extreme. To collect too rarely falls back on full collections,
1290 // which works, even if not optimum in terms of concurrent work.
1291 // As a work around for too eagerly collecting, use the flag
1292 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1293 // giving the user an easily understandable way of controlling the
1294 // collections.
1295 // We want to start a new collection cycle if any of the following
1296 // conditions hold:
1297 // . our current occupancy exceeds the configured initiating occupancy
1298 // for this generation, or
1299 // . we recently needed to expand this space and have not, since that
1300 // expansion, done a collection of this generation, or
1301 // . the underlying space believes that it may be a good idea to initiate
1302 // a concurrent collection (this may be based on criteria such as the
1303 // following: the space uses linear allocation and linear allocation is
1304 // going to fail, or there is believed to be excessive fragmentation in
1305 // the generation, etc... or ...
1306 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1307 // the case of the old generation; see CR 6543076):
1308 // we may be approaching a point at which allocation requests may fail because
1309 // we will be out of sufficient free space given allocation rate estimates.]
1310 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1311
1312 assert_lock_strong(freelistLock());
1313 if (occupancy() > initiating_occupancy()) {
1314 if (PrintGCDetails && Verbose) {
1315 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1316 short_name(), occupancy(), initiating_occupancy());
1317 }
1318 return true;
1319 }
1320 if (UseCMSInitiatingOccupancyOnly) {
1321 return false;
1322 }
1323 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1324 if (PrintGCDetails && Verbose) {
1325 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1326 short_name());
1327 }
1328 return true;
1329 }
1330 if (_cmsSpace->should_concurrent_collect()) {
1331 if (PrintGCDetails && Verbose) {
1332 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1333 short_name());
1334 }
1335 return true;
1336 }
1337 return false;
1338 }
1339
1340 void ConcurrentMarkSweepGeneration::collect(bool full,
1341 bool clear_all_soft_refs,
1342 size_t size,
1343 bool tlab)
1344 {
1345 collector()->collect(full, clear_all_soft_refs, size, tlab);
1346 }
1347
1348 void CMSCollector::collect(bool full,
1349 bool clear_all_soft_refs,
1350 size_t size,
1351 bool tlab)
1352 {
1353 // The following "if" branch is present for defensive reasons.
1354 // In the current uses of this interface, it can be replaced with:
1355 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1356 // But I am not placing that assert here to allow future
1357 // generality in invoking this interface.
1358 if (GC_locker::is_active()) {
1359 // A consistency test for GC_locker
1360 assert(GC_locker::needs_gc(), "Should have been set already");
1361 // Skip this foreground collection, instead
1362 // expanding the heap if necessary.
1363 // Need the free list locks for the call to free() in compute_new_size()
1364 compute_new_size();
1365 return;
1366 }
1367 acquire_control_and_collect(full, clear_all_soft_refs);
1368 }
1369
1370 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1371 GenCollectedHeap* gch = GenCollectedHeap::heap();
1372 unsigned int gc_count = gch->total_full_collections();
1373 if (gc_count == full_gc_count) {
1374 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1375 _full_gc_requested = true;
1376 _full_gc_cause = cause;
1377 CGC_lock->notify(); // nudge CMS thread
1378 } else {
1379 assert(gc_count > full_gc_count, "Error: causal loop");
1380 }
1381 }
1382
1383 bool CMSCollector::is_external_interruption() {
1384 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1385 return GCCause::is_user_requested_gc(cause) ||
1386 GCCause::is_serviceability_requested_gc(cause);
1387 }
1388
1389 void CMSCollector::report_concurrent_mode_interruption() {
1390 if (is_external_interruption()) {
1391 if (PrintGCDetails) {
1392 gclog_or_tty->print(" (concurrent mode interrupted)");
1393 }
1394 } else {
1395 if (PrintGCDetails) {
1396 gclog_or_tty->print(" (concurrent mode failure)");
1397 }
1398 _gc_tracer_cm->report_concurrent_mode_failure();
1399 }
1400 }
1401
1402
1403 // The foreground and background collectors need to coordinate in order
1404 // to make sure that they do not mutually interfere with CMS collections.
1405 // When a background collection is active,
1406 // the foreground collector may need to take over (preempt) and
1407 // synchronously complete an ongoing collection. Depending on the
1408 // frequency of the background collections and the heap usage
1409 // of the application, this preemption can be seldom or frequent.
1410 // There are only certain
1411 // points in the background collection that the "collection-baton"
1412 // can be passed to the foreground collector.
1413 //
1414 // The foreground collector will wait for the baton before
1415 // starting any part of the collection. The foreground collector
1416 // will only wait at one location.
1417 //
1418 // The background collector will yield the baton before starting a new
1419 // phase of the collection (e.g., before initial marking, marking from roots,
1420 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1421 // of the loop which switches the phases. The background collector does some
1422 // of the phases (initial mark, final re-mark) with the world stopped.
1423 // Because of locking involved in stopping the world,
1424 // the foreground collector should not block waiting for the background
1425 // collector when it is doing a stop-the-world phase. The background
1426 // collector will yield the baton at an additional point just before
1427 // it enters a stop-the-world phase. Once the world is stopped, the
1428 // background collector checks the phase of the collection. If the
1429 // phase has not changed, it proceeds with the collection. If the
1430 // phase has changed, it skips that phase of the collection. See
1431 // the comments on the use of the Heap_lock in collect_in_background().
1432 //
1433 // Variable used in baton passing.
1434 // _foregroundGCIsActive - Set to true by the foreground collector when
1435 // it wants the baton. The foreground clears it when it has finished
1436 // the collection.
1437 // _foregroundGCShouldWait - Set to true by the background collector
1438 // when it is running. The foreground collector waits while
1439 // _foregroundGCShouldWait is true.
1440 // CGC_lock - monitor used to protect access to the above variables
1441 // and to notify the foreground and background collectors.
1442 // _collectorState - current state of the CMS collection.
1443 //
1444 // The foreground collector
1445 // acquires the CGC_lock
1446 // sets _foregroundGCIsActive
1447 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1448 // various locks acquired in preparation for the collection
1449 // are released so as not to block the background collector
1450 // that is in the midst of a collection
1451 // proceeds with the collection
1452 // clears _foregroundGCIsActive
1453 // returns
1454 //
1455 // The background collector in a loop iterating on the phases of the
1456 // collection
1457 // acquires the CGC_lock
1458 // sets _foregroundGCShouldWait
1459 // if _foregroundGCIsActive is set
1460 // clears _foregroundGCShouldWait, notifies _CGC_lock
1461 // waits on _CGC_lock for _foregroundGCIsActive to become false
1462 // and exits the loop.
1463 // otherwise
1464 // proceed with that phase of the collection
1465 // if the phase is a stop-the-world phase,
1466 // yield the baton once more just before enqueueing
1467 // the stop-world CMS operation (executed by the VM thread).
1468 // returns after all phases of the collection are done
1469 //
1470
1471 void CMSCollector::acquire_control_and_collect(bool full,
1472 bool clear_all_soft_refs) {
1473 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1474 assert(!Thread::current()->is_ConcurrentGC_thread(),
1475 "shouldn't try to acquire control from self!");
1476
1477 // Start the protocol for acquiring control of the
1478 // collection from the background collector (aka CMS thread).
1479 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1480 "VM thread should have CMS token");
1481 // Remember the possibly interrupted state of an ongoing
1482 // concurrent collection
1483 CollectorState first_state = _collectorState;
1484
1485 // Signal to a possibly ongoing concurrent collection that
1486 // we want to do a foreground collection.
1487 _foregroundGCIsActive = true;
1488
1489 // release locks and wait for a notify from the background collector
1490 // releasing the locks in only necessary for phases which
1491 // do yields to improve the granularity of the collection.
1492 assert_lock_strong(bitMapLock());
1493 // We need to lock the Free list lock for the space that we are
1494 // currently collecting.
1495 assert(haveFreelistLocks(), "Must be holding free list locks");
1496 bitMapLock()->unlock();
1497 releaseFreelistLocks();
1498 {
1499 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1500 if (_foregroundGCShouldWait) {
1501 // We are going to be waiting for action for the CMS thread;
1502 // it had better not be gone (for instance at shutdown)!
1503 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1504 "CMS thread must be running");
1505 // Wait here until the background collector gives us the go-ahead
1506 ConcurrentMarkSweepThread::clear_CMS_flag(
1507 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1508 // Get a possibly blocked CMS thread going:
1509 // Note that we set _foregroundGCIsActive true above,
1510 // without protection of the CGC_lock.
1511 CGC_lock->notify();
1512 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1513 "Possible deadlock");
1514 while (_foregroundGCShouldWait) {
1515 // wait for notification
1516 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1517 // Possibility of delay/starvation here, since CMS token does
1518 // not know to give priority to VM thread? Actually, i think
1519 // there wouldn't be any delay/starvation, but the proof of
1520 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1521 }
1522 ConcurrentMarkSweepThread::set_CMS_flag(
1523 ConcurrentMarkSweepThread::CMS_vm_has_token);
1524 }
1525 }
1526 // The CMS_token is already held. Get back the other locks.
1527 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1528 "VM thread should have CMS token");
1529 getFreelistLocks();
1530 bitMapLock()->lock_without_safepoint_check();
1531 if (TraceCMSState) {
1532 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1533 INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state);
1534 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1535 }
1536
1537 // Inform cms gen if this was due to partial collection failing.
1538 // The CMS gen may use this fact to determine its expansion policy.
1539 GenCollectedHeap* gch = GenCollectedHeap::heap();
1540 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1541 assert(!_cmsGen->incremental_collection_failed(),
1542 "Should have been noticed, reacted to and cleared");
1543 _cmsGen->set_incremental_collection_failed();
1544 }
1545
1546 if (first_state > Idling) {
1547 report_concurrent_mode_interruption();
1548 }
1549
1550 set_did_compact(true);
1551
1552 // If the collection is being acquired from the background
1553 // collector, there may be references on the discovered
1554 // references lists. Abandon those references, since some
1555 // of them may have become unreachable after concurrent
1556 // discovery; the STW compacting collector will redo discovery
1557 // more precisely, without being subject to floating garbage.
1558 // Leaving otherwise unreachable references in the discovered
1559 // lists would require special handling.
1560 ref_processor()->disable_discovery();
1561 ref_processor()->abandon_partial_discovery();
1562 ref_processor()->verify_no_references_recorded();
1563
1564 if (first_state > Idling) {
1565 save_heap_summary();
1566 }
1567
1568 do_compaction_work(clear_all_soft_refs);
1569
1570 // Has the GC time limit been exceeded?
1571 size_t max_eden_size = _young_gen->max_capacity() -
1572 _young_gen->to()->capacity() -
1573 _young_gen->from()->capacity();
1574 GCCause::Cause gc_cause = gch->gc_cause();
1575 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1576 _young_gen->eden()->used(),
1577 _cmsGen->max_capacity(),
1578 max_eden_size,
1579 full,
1580 gc_cause,
1581 gch->collector_policy());
1582
1583 // Reset the expansion cause, now that we just completed
1584 // a collection cycle.
1585 clear_expansion_cause();
1586 _foregroundGCIsActive = false;
1587 return;
1588 }
1589
1590 // Resize the tenured generation
1591 // after obtaining the free list locks for the
1592 // two generations.
1593 void CMSCollector::compute_new_size() {
1594 assert_locked_or_safepoint(Heap_lock);
1595 FreelistLocker z(this);
1596 MetaspaceGC::compute_new_size();
1597 _cmsGen->compute_new_size_free_list();
1598 }
1599
1600 // A work method used by the foreground collector to do
1601 // a mark-sweep-compact.
1602 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1603 GenCollectedHeap* gch = GenCollectedHeap::heap();
1604
1605 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1606 gc_timer->register_gc_start();
1607
1608 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1609 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
1610
1611 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id());
1612
1613 // Temporarily widen the span of the weak reference processing to
1614 // the entire heap.
1615 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1616 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1617 // Temporarily, clear the "is_alive_non_header" field of the
1618 // reference processor.
1619 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1620 // Temporarily make reference _processing_ single threaded (non-MT).
1621 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1622 // Temporarily make refs discovery atomic
1623 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1624 // Temporarily make reference _discovery_ single threaded (non-MT)
1625 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1626
1627 ref_processor()->set_enqueuing_is_done(false);
1628 ref_processor()->enable_discovery();
1629 ref_processor()->setup_policy(clear_all_soft_refs);
1630 // If an asynchronous collection finishes, the _modUnionTable is
1631 // all clear. If we are assuming the collection from an asynchronous
1632 // collection, clear the _modUnionTable.
1633 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1634 "_modUnionTable should be clear if the baton was not passed");
1635 _modUnionTable.clear_all();
1636 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
1637 "mod union for klasses should be clear if the baton was passed");
1638 _ct->klass_rem_set()->clear_mod_union();
1639
1640 // We must adjust the allocation statistics being maintained
1641 // in the free list space. We do so by reading and clearing
1642 // the sweep timer and updating the block flux rate estimates below.
1643 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1644 if (_inter_sweep_timer.is_active()) {
1645 _inter_sweep_timer.stop();
1646 // Note that we do not use this sample to update the _inter_sweep_estimate.
1647 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1648 _inter_sweep_estimate.padded_average(),
1649 _intra_sweep_estimate.padded_average());
1650 }
1651
1652 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
1653 ref_processor(), clear_all_soft_refs);
1654 #ifdef ASSERT
1655 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1656 size_t free_size = cms_space->free();
1657 assert(free_size ==
1658 pointer_delta(cms_space->end(), cms_space->compaction_top())
1659 * HeapWordSize,
1660 "All the free space should be compacted into one chunk at top");
1661 assert(cms_space->dictionary()->total_chunk_size(
1662 debug_only(cms_space->freelistLock())) == 0 ||
1663 cms_space->totalSizeInIndexedFreeLists() == 0,
1664 "All the free space should be in a single chunk");
1665 size_t num = cms_space->totalCount();
1666 assert((free_size == 0 && num == 0) ||
1667 (free_size > 0 && (num == 1 || num == 2)),
1668 "There should be at most 2 free chunks after compaction");
1669 #endif // ASSERT
1670 _collectorState = Resetting;
1671 assert(_restart_addr == NULL,
1672 "Should have been NULL'd before baton was passed");
1673 reset(false /* == !concurrent */);
1674 _cmsGen->reset_after_compaction();
1675 _concurrent_cycles_since_last_unload = 0;
1676
1677 // Clear any data recorded in the PLAB chunk arrays.
1678 if (_survivor_plab_array != NULL) {
1679 reset_survivor_plab_arrays();
1680 }
1681
1682 // Adjust the per-size allocation stats for the next epoch.
1683 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1684 // Restart the "inter sweep timer" for the next epoch.
1685 _inter_sweep_timer.reset();
1686 _inter_sweep_timer.start();
1687
1688 gc_timer->register_gc_end();
1689
1690 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1691
1692 // For a mark-sweep-compact, compute_new_size() will be called
1693 // in the heap's do_collection() method.
1694 }
1695
1696 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1697 ContiguousSpace* eden_space = _young_gen->eden();
1698 ContiguousSpace* from_space = _young_gen->from();
1699 ContiguousSpace* to_space = _young_gen->to();
1700 // Eden
1701 if (_eden_chunk_array != NULL) {
1702 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1703 p2i(eden_space->bottom()), p2i(eden_space->top()),
1704 p2i(eden_space->end()), eden_space->capacity());
1705 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", "
1706 "_eden_chunk_capacity=" SIZE_FORMAT,
1707 _eden_chunk_index, _eden_chunk_capacity);
1708 for (size_t i = 0; i < _eden_chunk_index; i++) {
1709 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1710 i, p2i(_eden_chunk_array[i]));
1711 }
1712 }
1713 // Survivor
1714 if (_survivor_chunk_array != NULL) {
1715 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1716 p2i(from_space->bottom()), p2i(from_space->top()),
1717 p2i(from_space->end()), from_space->capacity());
1718 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", "
1719 "_survivor_chunk_capacity=" SIZE_FORMAT,
1720 _survivor_chunk_index, _survivor_chunk_capacity);
1721 for (size_t i = 0; i < _survivor_chunk_index; i++) {
1722 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1723 i, p2i(_survivor_chunk_array[i]));
1724 }
1725 }
1726 }
1727
1728 void CMSCollector::getFreelistLocks() const {
1729 // Get locks for all free lists in all generations that this
1730 // collector is responsible for
1731 _cmsGen->freelistLock()->lock_without_safepoint_check();
1732 }
1733
1734 void CMSCollector::releaseFreelistLocks() const {
1735 // Release locks for all free lists in all generations that this
1736 // collector is responsible for
1737 _cmsGen->freelistLock()->unlock();
1738 }
1739
1740 bool CMSCollector::haveFreelistLocks() const {
1741 // Check locks for all free lists in all generations that this
1742 // collector is responsible for
1743 assert_lock_strong(_cmsGen->freelistLock());
1744 PRODUCT_ONLY(ShouldNotReachHere());
1745 return true;
1746 }
1747
1748 // A utility class that is used by the CMS collector to
1749 // temporarily "release" the foreground collector from its
1750 // usual obligation to wait for the background collector to
1751 // complete an ongoing phase before proceeding.
1752 class ReleaseForegroundGC: public StackObj {
1753 private:
1754 CMSCollector* _c;
1755 public:
1756 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1757 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1758 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1759 // allow a potentially blocked foreground collector to proceed
1760 _c->_foregroundGCShouldWait = false;
1761 if (_c->_foregroundGCIsActive) {
1762 CGC_lock->notify();
1763 }
1764 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1765 "Possible deadlock");
1766 }
1767
1768 ~ReleaseForegroundGC() {
1769 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1770 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1771 _c->_foregroundGCShouldWait = true;
1772 }
1773 };
1774
1775 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1776 assert(Thread::current()->is_ConcurrentGC_thread(),
1777 "A CMS asynchronous collection is only allowed on a CMS thread.");
1778
1779 GenCollectedHeap* gch = GenCollectedHeap::heap();
1780 {
1781 bool safepoint_check = Mutex::_no_safepoint_check_flag;
1782 MutexLockerEx hl(Heap_lock, safepoint_check);
1783 FreelistLocker fll(this);
1784 MutexLockerEx x(CGC_lock, safepoint_check);
1785 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
1786 // The foreground collector is active or we're
1787 // not using asynchronous collections. Skip this
1788 // background collection.
1789 assert(!_foregroundGCShouldWait, "Should be clear");
1790 return;
1791 } else {
1792 assert(_collectorState == Idling, "Should be idling before start.");
1793 _collectorState = InitialMarking;
1794 register_gc_start(cause);
1795 // Reset the expansion cause, now that we are about to begin
1796 // a new cycle.
1797 clear_expansion_cause();
1798
1799 // Clear the MetaspaceGC flag since a concurrent collection
1800 // is starting but also clear it after the collection.
1801 MetaspaceGC::set_should_concurrent_collect(false);
1802 }
1803 // Decide if we want to enable class unloading as part of the
1804 // ensuing concurrent GC cycle.
1805 update_should_unload_classes();
1806 _full_gc_requested = false; // acks all outstanding full gc requests
1807 _full_gc_cause = GCCause::_no_gc;
1808 // Signal that we are about to start a collection
1809 gch->increment_total_full_collections(); // ... starting a collection cycle
1810 _collection_count_start = gch->total_full_collections();
1811 }
1812
1813 // Used for PrintGC
1814 size_t prev_used;
1815 if (PrintGC && Verbose) {
1816 prev_used = _cmsGen->used();
1817 }
1818
1819 // The change of the collection state is normally done at this level;
1820 // the exceptions are phases that are executed while the world is
1821 // stopped. For those phases the change of state is done while the
1822 // world is stopped. For baton passing purposes this allows the
1823 // background collector to finish the phase and change state atomically.
1824 // The foreground collector cannot wait on a phase that is done
1825 // while the world is stopped because the foreground collector already
1826 // has the world stopped and would deadlock.
1827 while (_collectorState != Idling) {
1828 if (TraceCMSState) {
1829 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
1830 p2i(Thread::current()), _collectorState);
1831 }
1832 // The foreground collector
1833 // holds the Heap_lock throughout its collection.
1834 // holds the CMS token (but not the lock)
1835 // except while it is waiting for the background collector to yield.
1836 //
1837 // The foreground collector should be blocked (not for long)
1838 // if the background collector is about to start a phase
1839 // executed with world stopped. If the background
1840 // collector has already started such a phase, the
1841 // foreground collector is blocked waiting for the
1842 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
1843 // are executed in the VM thread.
1844 //
1845 // The locking order is
1846 // PendingListLock (PLL) -- if applicable (FinalMarking)
1847 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
1848 // CMS token (claimed in
1849 // stop_world_and_do() -->
1850 // safepoint_synchronize() -->
1851 // CMSThread::synchronize())
1852
1853 {
1854 // Check if the FG collector wants us to yield.
1855 CMSTokenSync x(true); // is cms thread
1856 if (waitForForegroundGC()) {
1857 // We yielded to a foreground GC, nothing more to be
1858 // done this round.
1859 assert(_foregroundGCShouldWait == false, "We set it to false in "
1860 "waitForForegroundGC()");
1861 if (TraceCMSState) {
1862 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1863 " exiting collection CMS state %d",
1864 p2i(Thread::current()), _collectorState);
1865 }
1866 return;
1867 } else {
1868 // The background collector can run but check to see if the
1869 // foreground collector has done a collection while the
1870 // background collector was waiting to get the CGC_lock
1871 // above. If yes, break so that _foregroundGCShouldWait
1872 // is cleared before returning.
1873 if (_collectorState == Idling) {
1874 break;
1875 }
1876 }
1877 }
1878
1879 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1880 "should be waiting");
1881
1882 switch (_collectorState) {
1883 case InitialMarking:
1884 {
1885 ReleaseForegroundGC x(this);
1886 stats().record_cms_begin();
1887 VM_CMS_Initial_Mark initial_mark_op(this);
1888 VMThread::execute(&initial_mark_op);
1889 }
1890 // The collector state may be any legal state at this point
1891 // since the background collector may have yielded to the
1892 // foreground collector.
1893 break;
1894 case Marking:
1895 // initial marking in checkpointRootsInitialWork has been completed
1896 if (markFromRoots()) { // we were successful
1897 assert(_collectorState == Precleaning, "Collector state should "
1898 "have changed");
1899 } else {
1900 assert(_foregroundGCIsActive, "Internal state inconsistency");
1901 }
1902 break;
1903 case Precleaning:
1904 // marking from roots in markFromRoots has been completed
1905 preclean();
1906 assert(_collectorState == AbortablePreclean ||
1907 _collectorState == FinalMarking,
1908 "Collector state should have changed");
1909 break;
1910 case AbortablePreclean:
1911 abortable_preclean();
1912 assert(_collectorState == FinalMarking, "Collector state should "
1913 "have changed");
1914 break;
1915 case FinalMarking:
1916 {
1917 ReleaseForegroundGC x(this);
1918
1919 VM_CMS_Final_Remark final_remark_op(this);
1920 VMThread::execute(&final_remark_op);
1921 }
1922 assert(_foregroundGCShouldWait, "block post-condition");
1923 break;
1924 case Sweeping:
1925 // final marking in checkpointRootsFinal has been completed
1926 sweep();
1927 assert(_collectorState == Resizing, "Collector state change "
1928 "to Resizing must be done under the free_list_lock");
1929
1930 case Resizing: {
1931 // Sweeping has been completed...
1932 // At this point the background collection has completed.
1933 // Don't move the call to compute_new_size() down
1934 // into code that might be executed if the background
1935 // collection was preempted.
1936 {
1937 ReleaseForegroundGC x(this); // unblock FG collection
1938 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
1939 CMSTokenSync z(true); // not strictly needed.
1940 if (_collectorState == Resizing) {
1941 compute_new_size();
1942 save_heap_summary();
1943 _collectorState = Resetting;
1944 } else {
1945 assert(_collectorState == Idling, "The state should only change"
1946 " because the foreground collector has finished the collection");
1947 }
1948 }
1949 break;
1950 }
1951 case Resetting:
1952 // CMS heap resizing has been completed
1953 reset(true);
1954 assert(_collectorState == Idling, "Collector state should "
1955 "have changed");
1956
1957 MetaspaceGC::set_should_concurrent_collect(false);
1958
1959 stats().record_cms_end();
1960 // Don't move the concurrent_phases_end() and compute_new_size()
1961 // calls to here because a preempted background collection
1962 // has it's state set to "Resetting".
1963 break;
1964 case Idling:
1965 default:
1966 ShouldNotReachHere();
1967 break;
1968 }
1969 if (TraceCMSState) {
1970 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
1971 p2i(Thread::current()), _collectorState);
1972 }
1973 assert(_foregroundGCShouldWait, "block post-condition");
1974 }
1975
1976 // Should this be in gc_epilogue?
1977 collector_policy()->counters()->update_counters();
1978
1979 {
1980 // Clear _foregroundGCShouldWait and, in the event that the
1981 // foreground collector is waiting, notify it, before
1982 // returning.
1983 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1984 _foregroundGCShouldWait = false;
1985 if (_foregroundGCIsActive) {
1986 CGC_lock->notify();
1987 }
1988 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1989 "Possible deadlock");
1990 }
1991 if (TraceCMSState) {
1992 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1993 " exiting collection CMS state %d",
1994 p2i(Thread::current()), _collectorState);
1995 }
1996 if (PrintGC && Verbose) {
1997 _cmsGen->print_heap_change(prev_used);
1998 }
1999 }
2000
2001 void CMSCollector::register_gc_start(GCCause::Cause cause) {
2002 _cms_start_registered = true;
2003 _gc_timer_cm->register_gc_start();
2004 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
2005 }
2006
2007 void CMSCollector::register_gc_end() {
2008 if (_cms_start_registered) {
2009 report_heap_summary(GCWhen::AfterGC);
2010
2011 _gc_timer_cm->register_gc_end();
2012 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
2013 _cms_start_registered = false;
2014 }
2015 }
2016
2017 void CMSCollector::save_heap_summary() {
2018 GenCollectedHeap* gch = GenCollectedHeap::heap();
2019 _last_heap_summary = gch->create_heap_summary();
2020 _last_metaspace_summary = gch->create_metaspace_summary();
2021 }
2022
2023 void CMSCollector::report_heap_summary(GCWhen::Type when) {
2024 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
2025 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
2026 }
2027
2028 bool CMSCollector::waitForForegroundGC() {
2029 bool res = false;
2030 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2031 "CMS thread should have CMS token");
2032 // Block the foreground collector until the
2033 // background collectors decides whether to
2034 // yield.
2035 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2036 _foregroundGCShouldWait = true;
2037 if (_foregroundGCIsActive) {
2038 // The background collector yields to the
2039 // foreground collector and returns a value
2040 // indicating that it has yielded. The foreground
2041 // collector can proceed.
2042 res = true;
2043 _foregroundGCShouldWait = false;
2044 ConcurrentMarkSweepThread::clear_CMS_flag(
2045 ConcurrentMarkSweepThread::CMS_cms_has_token);
2046 ConcurrentMarkSweepThread::set_CMS_flag(
2047 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2048 // Get a possibly blocked foreground thread going
2049 CGC_lock->notify();
2050 if (TraceCMSState) {
2051 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2052 p2i(Thread::current()), _collectorState);
2053 }
2054 while (_foregroundGCIsActive) {
2055 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2056 }
2057 ConcurrentMarkSweepThread::set_CMS_flag(
2058 ConcurrentMarkSweepThread::CMS_cms_has_token);
2059 ConcurrentMarkSweepThread::clear_CMS_flag(
2060 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2061 }
2062 if (TraceCMSState) {
2063 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2064 p2i(Thread::current()), _collectorState);
2065 }
2066 return res;
2067 }
2068
2069 // Because of the need to lock the free lists and other structures in
2070 // the collector, common to all the generations that the collector is
2071 // collecting, we need the gc_prologues of individual CMS generations
2072 // delegate to their collector. It may have been simpler had the
2073 // current infrastructure allowed one to call a prologue on a
2074 // collector. In the absence of that we have the generation's
2075 // prologue delegate to the collector, which delegates back
2076 // some "local" work to a worker method in the individual generations
2077 // that it's responsible for collecting, while itself doing any
2078 // work common to all generations it's responsible for. A similar
2079 // comment applies to the gc_epilogue()'s.
2080 // The role of the variable _between_prologue_and_epilogue is to
2081 // enforce the invocation protocol.
2082 void CMSCollector::gc_prologue(bool full) {
2083 // Call gc_prologue_work() for the CMSGen
2084 // we are responsible for.
2085
2086 // The following locking discipline assumes that we are only called
2087 // when the world is stopped.
2088 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2089
2090 // The CMSCollector prologue must call the gc_prologues for the
2091 // "generations" that it's responsible
2092 // for.
2093
2094 assert( Thread::current()->is_VM_thread()
2095 || ( CMSScavengeBeforeRemark
2096 && Thread::current()->is_ConcurrentGC_thread()),
2097 "Incorrect thread type for prologue execution");
2098
2099 if (_between_prologue_and_epilogue) {
2100 // We have already been invoked; this is a gc_prologue delegation
2101 // from yet another CMS generation that we are responsible for, just
2102 // ignore it since all relevant work has already been done.
2103 return;
2104 }
2105
2106 // set a bit saying prologue has been called; cleared in epilogue
2107 _between_prologue_and_epilogue = true;
2108 // Claim locks for common data structures, then call gc_prologue_work()
2109 // for each CMSGen.
2110
2111 getFreelistLocks(); // gets free list locks on constituent spaces
2112 bitMapLock()->lock_without_safepoint_check();
2113
2114 // Should call gc_prologue_work() for all cms gens we are responsible for
2115 bool duringMarking = _collectorState >= Marking
2116 && _collectorState < Sweeping;
2117
2118 // The young collections clear the modified oops state, which tells if
2119 // there are any modified oops in the class. The remark phase also needs
2120 // that information. Tell the young collection to save the union of all
2121 // modified klasses.
2122 if (duringMarking) {
2123 _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2124 }
2125
2126 bool registerClosure = duringMarking;
2127
2128 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2129
2130 if (!full) {
2131 stats().record_gc0_begin();
2132 }
2133 }
2134
2135 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2136
2137 _capacity_at_prologue = capacity();
2138 _used_at_prologue = used();
2139
2140 // Delegate to CMScollector which knows how to coordinate between
2141 // this and any other CMS generations that it is responsible for
2142 // collecting.
2143 collector()->gc_prologue(full);
2144 }
2145
2146 // This is a "private" interface for use by this generation's CMSCollector.
2147 // Not to be called directly by any other entity (for instance,
2148 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2149 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2150 bool registerClosure, ModUnionClosure* modUnionClosure) {
2151 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2152 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2153 "Should be NULL");
2154 if (registerClosure) {
2155 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2156 }
2157 cmsSpace()->gc_prologue();
2158 // Clear stat counters
2159 NOT_PRODUCT(
2160 assert(_numObjectsPromoted == 0, "check");
2161 assert(_numWordsPromoted == 0, "check");
2162 if (Verbose && PrintGC) {
2163 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2164 SIZE_FORMAT" bytes concurrently",
2165 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2166 }
2167 _numObjectsAllocated = 0;
2168 _numWordsAllocated = 0;
2169 )
2170 }
2171
2172 void CMSCollector::gc_epilogue(bool full) {
2173 // The following locking discipline assumes that we are only called
2174 // when the world is stopped.
2175 assert(SafepointSynchronize::is_at_safepoint(),
2176 "world is stopped assumption");
2177
2178 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2179 // if linear allocation blocks need to be appropriately marked to allow the
2180 // the blocks to be parsable. We also check here whether we need to nudge the
2181 // CMS collector thread to start a new cycle (if it's not already active).
2182 assert( Thread::current()->is_VM_thread()
2183 || ( CMSScavengeBeforeRemark
2184 && Thread::current()->is_ConcurrentGC_thread()),
2185 "Incorrect thread type for epilogue execution");
2186
2187 if (!_between_prologue_and_epilogue) {
2188 // We have already been invoked; this is a gc_epilogue delegation
2189 // from yet another CMS generation that we are responsible for, just
2190 // ignore it since all relevant work has already been done.
2191 return;
2192 }
2193 assert(haveFreelistLocks(), "must have freelist locks");
2194 assert_lock_strong(bitMapLock());
2195
2196 _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2197
2198 _cmsGen->gc_epilogue_work(full);
2199
2200 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2201 // in case sampling was not already enabled, enable it
2202 _start_sampling = true;
2203 }
2204 // reset _eden_chunk_array so sampling starts afresh
2205 _eden_chunk_index = 0;
2206
2207 size_t cms_used = _cmsGen->cmsSpace()->used();
2208
2209 // update performance counters - this uses a special version of
2210 // update_counters() that allows the utilization to be passed as a
2211 // parameter, avoiding multiple calls to used().
2212 //
2213 _cmsGen->update_counters(cms_used);
2214
2215 bitMapLock()->unlock();
2216 releaseFreelistLocks();
2217
2218 if (!CleanChunkPoolAsync) {
2219 Chunk::clean_chunk_pool();
2220 }
2221
2222 set_did_compact(false);
2223 _between_prologue_and_epilogue = false; // ready for next cycle
2224 }
2225
2226 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2227 collector()->gc_epilogue(full);
2228
2229 // Also reset promotion tracking in par gc thread states.
2230 for (uint i = 0; i < ParallelGCThreads; i++) {
2231 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2232 }
2233 }
2234
2235 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2236 assert(!incremental_collection_failed(), "Should have been cleared");
2237 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2238 cmsSpace()->gc_epilogue();
2239 // Print stat counters
2240 NOT_PRODUCT(
2241 assert(_numObjectsAllocated == 0, "check");
2242 assert(_numWordsAllocated == 0, "check");
2243 if (Verbose && PrintGC) {
2244 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2245 SIZE_FORMAT" bytes",
2246 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2247 }
2248 _numObjectsPromoted = 0;
2249 _numWordsPromoted = 0;
2250 )
2251
2252 if (PrintGC && Verbose) {
2253 // Call down the chain in contiguous_available needs the freelistLock
2254 // so print this out before releasing the freeListLock.
2255 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2256 contiguous_available());
2257 }
2258 }
2259
2260 #ifndef PRODUCT
2261 bool CMSCollector::have_cms_token() {
2262 Thread* thr = Thread::current();
2263 if (thr->is_VM_thread()) {
2264 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2265 } else if (thr->is_ConcurrentGC_thread()) {
2266 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2267 } else if (thr->is_GC_task_thread()) {
2268 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2269 ParGCRareEvent_lock->owned_by_self();
2270 }
2271 return false;
2272 }
2273 #endif
2274
2275 // Check reachability of the given heap address in CMS generation,
2276 // treating all other generations as roots.
2277 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2278 // We could "guarantee" below, rather than assert, but I'll
2279 // leave these as "asserts" so that an adventurous debugger
2280 // could try this in the product build provided some subset of
2281 // the conditions were met, provided they were interested in the
2282 // results and knew that the computation below wouldn't interfere
2283 // with other concurrent computations mutating the structures
2284 // being read or written.
2285 assert(SafepointSynchronize::is_at_safepoint(),
2286 "Else mutations in object graph will make answer suspect");
2287 assert(have_cms_token(), "Should hold cms token");
2288 assert(haveFreelistLocks(), "must hold free list locks");
2289 assert_lock_strong(bitMapLock());
2290
2291 // Clear the marking bit map array before starting, but, just
2292 // for kicks, first report if the given address is already marked
2293 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr),
2294 _markBitMap.isMarked(addr) ? "" : " not");
2295
2296 if (verify_after_remark()) {
2297 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2298 bool result = verification_mark_bm()->isMarked(addr);
2299 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr),
2300 result ? "IS" : "is NOT");
2301 return result;
2302 } else {
2303 gclog_or_tty->print_cr("Could not compute result");
2304 return false;
2305 }
2306 }
2307
2308
2309 void
2310 CMSCollector::print_on_error(outputStream* st) {
2311 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2312 if (collector != NULL) {
2313 CMSBitMap* bitmap = &collector->_markBitMap;
2314 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2315 bitmap->print_on_error(st, " Bits: ");
2316
2317 st->cr();
2318
2319 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2320 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2321 mut_bitmap->print_on_error(st, " Bits: ");
2322 }
2323 }
2324
2325 ////////////////////////////////////////////////////////
2326 // CMS Verification Support
2327 ////////////////////////////////////////////////////////
2328 // Following the remark phase, the following invariant
2329 // should hold -- each object in the CMS heap which is
2330 // marked in markBitMap() should be marked in the verification_mark_bm().
2331
2332 class VerifyMarkedClosure: public BitMapClosure {
2333 CMSBitMap* _marks;
2334 bool _failed;
2335
2336 public:
2337 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2338
2339 bool do_bit(size_t offset) {
2340 HeapWord* addr = _marks->offsetToHeapWord(offset);
2341 if (!_marks->isMarked(addr)) {
2342 oop(addr)->print_on(gclog_or_tty);
2343 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", p2i(addr));
2344 _failed = true;
2345 }
2346 return true;
2347 }
2348
2349 bool failed() { return _failed; }
2350 };
2351
2352 bool CMSCollector::verify_after_remark(bool silent) {
2353 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... ");
2354 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2355 static bool init = false;
2356
2357 assert(SafepointSynchronize::is_at_safepoint(),
2358 "Else mutations in object graph will make answer suspect");
2359 assert(have_cms_token(),
2360 "Else there may be mutual interference in use of "
2361 " verification data structures");
2362 assert(_collectorState > Marking && _collectorState <= Sweeping,
2363 "Else marking info checked here may be obsolete");
2364 assert(haveFreelistLocks(), "must hold free list locks");
2365 assert_lock_strong(bitMapLock());
2366
2367
2368 // Allocate marking bit map if not already allocated
2369 if (!init) { // first time
2370 if (!verification_mark_bm()->allocate(_span)) {
2371 return false;
2372 }
2373 init = true;
2374 }
2375
2376 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2377
2378 // Turn off refs discovery -- so we will be tracing through refs.
2379 // This is as intended, because by this time
2380 // GC must already have cleared any refs that need to be cleared,
2381 // and traced those that need to be marked; moreover,
2382 // the marking done here is not going to interfere in any
2383 // way with the marking information used by GC.
2384 NoRefDiscovery no_discovery(ref_processor());
2385
2386 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2387
2388 // Clear any marks from a previous round
2389 verification_mark_bm()->clear_all();
2390 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2391 verify_work_stacks_empty();
2392
2393 GenCollectedHeap* gch = GenCollectedHeap::heap();
2394 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2395 // Update the saved marks which may affect the root scans.
2396 gch->save_marks();
2397
2398 if (CMSRemarkVerifyVariant == 1) {
2399 // In this first variant of verification, we complete
2400 // all marking, then check if the new marks-vector is
2401 // a subset of the CMS marks-vector.
2402 verify_after_remark_work_1();
2403 } else if (CMSRemarkVerifyVariant == 2) {
2404 // In this second variant of verification, we flag an error
2405 // (i.e. an object reachable in the new marks-vector not reachable
2406 // in the CMS marks-vector) immediately, also indicating the
2407 // identify of an object (A) that references the unmarked object (B) --
2408 // presumably, a mutation to A failed to be picked up by preclean/remark?
2409 verify_after_remark_work_2();
2410 } else {
2411 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant",
2412 CMSRemarkVerifyVariant);
2413 }
2414 if (!silent) gclog_or_tty->print(" done] ");
2415 return true;
2416 }
2417
2418 void CMSCollector::verify_after_remark_work_1() {
2419 ResourceMark rm;
2420 HandleMark hm;
2421 GenCollectedHeap* gch = GenCollectedHeap::heap();
2422
2423 // Get a clear set of claim bits for the roots processing to work with.
2424 ClassLoaderDataGraph::clear_claimed_marks();
2425
2426 // Mark from roots one level into CMS
2427 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2428 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2429
2430 gch->gen_process_roots(_cmsGen->level(),
2431 true, // younger gens are roots
2432 true, // activate StrongRootsScope
2433 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2434 should_unload_classes(),
2435 ¬Older,
2436 NULL,
2437 NULL); // SSS: Provide correct closure
2438
2439 // Now mark from the roots
2440 MarkFromRootsClosure markFromRootsClosure(this, _span,
2441 verification_mark_bm(), verification_mark_stack(),
2442 false /* don't yield */, true /* verifying */);
2443 assert(_restart_addr == NULL, "Expected pre-condition");
2444 verification_mark_bm()->iterate(&markFromRootsClosure);
2445 while (_restart_addr != NULL) {
2446 // Deal with stack overflow: by restarting at the indicated
2447 // address.
2448 HeapWord* ra = _restart_addr;
2449 markFromRootsClosure.reset(ra);
2450 _restart_addr = NULL;
2451 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2452 }
2453 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2454 verify_work_stacks_empty();
2455
2456 // Marking completed -- now verify that each bit marked in
2457 // verification_mark_bm() is also marked in markBitMap(); flag all
2458 // errors by printing corresponding objects.
2459 VerifyMarkedClosure vcl(markBitMap());
2460 verification_mark_bm()->iterate(&vcl);
2461 if (vcl.failed()) {
2462 gclog_or_tty->print("Verification failed");
2463 gch->print_on(gclog_or_tty);
2464 fatal("CMS: failed marking verification after remark");
2465 }
2466 }
2467
2468 class VerifyKlassOopsKlassClosure : public KlassClosure {
2469 class VerifyKlassOopsClosure : public OopClosure {
2470 CMSBitMap* _bitmap;
2471 public:
2472 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2473 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2474 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2475 } _oop_closure;
2476 public:
2477 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2478 void do_klass(Klass* k) {
2479 k->oops_do(&_oop_closure);
2480 }
2481 };
2482
2483 void CMSCollector::verify_after_remark_work_2() {
2484 ResourceMark rm;
2485 HandleMark hm;
2486 GenCollectedHeap* gch = GenCollectedHeap::heap();
2487
2488 // Get a clear set of claim bits for the roots processing to work with.
2489 ClassLoaderDataGraph::clear_claimed_marks();
2490
2491 // Mark from roots one level into CMS
2492 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2493 markBitMap());
2494 CLDToOopClosure cld_closure(¬Older, true);
2495
2496 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2497
2498 gch->gen_process_roots(_cmsGen->level(),
2499 true, // younger gens are roots
2500 true, // activate StrongRootsScope
2501 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2502 should_unload_classes(),
2503 ¬Older,
2504 NULL,
2505 &cld_closure);
2506
2507 // Now mark from the roots
2508 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2509 verification_mark_bm(), markBitMap(), verification_mark_stack());
2510 assert(_restart_addr == NULL, "Expected pre-condition");
2511 verification_mark_bm()->iterate(&markFromRootsClosure);
2512 while (_restart_addr != NULL) {
2513 // Deal with stack overflow: by restarting at the indicated
2514 // address.
2515 HeapWord* ra = _restart_addr;
2516 markFromRootsClosure.reset(ra);
2517 _restart_addr = NULL;
2518 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2519 }
2520 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2521 verify_work_stacks_empty();
2522
2523 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
2524 ClassLoaderDataGraph::classes_do(&verify_klass_oops);
2525
2526 // Marking completed -- now verify that each bit marked in
2527 // verification_mark_bm() is also marked in markBitMap(); flag all
2528 // errors by printing corresponding objects.
2529 VerifyMarkedClosure vcl(markBitMap());
2530 verification_mark_bm()->iterate(&vcl);
2531 assert(!vcl.failed(), "Else verification above should not have succeeded");
2532 }
2533
2534 void ConcurrentMarkSweepGeneration::save_marks() {
2535 // delegate to CMS space
2536 cmsSpace()->save_marks();
2537 for (uint i = 0; i < ParallelGCThreads; i++) {
2538 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2539 }
2540 }
2541
2542 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2543 return cmsSpace()->no_allocs_since_save_marks();
2544 }
2545
2546 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2547 \
2548 void ConcurrentMarkSweepGeneration:: \
2549 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2550 cl->set_generation(this); \
2551 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2552 cl->reset_generation(); \
2553 save_marks(); \
2554 }
2555
2556 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2557
2558 void
2559 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2560 if (freelistLock()->owned_by_self()) {
2561 Generation::oop_iterate(cl);
2562 } else {
2563 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2564 Generation::oop_iterate(cl);
2565 }
2566 }
2567
2568 void
2569 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2570 if (freelistLock()->owned_by_self()) {
2571 Generation::object_iterate(cl);
2572 } else {
2573 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2574 Generation::object_iterate(cl);
2575 }
2576 }
2577
2578 void
2579 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2580 if (freelistLock()->owned_by_self()) {
2581 Generation::safe_object_iterate(cl);
2582 } else {
2583 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2584 Generation::safe_object_iterate(cl);
2585 }
2586 }
2587
2588 void
2589 ConcurrentMarkSweepGeneration::post_compact() {
2590 }
2591
2592 void
2593 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2594 // Fix the linear allocation blocks to look like free blocks.
2595
2596 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2597 // are not called when the heap is verified during universe initialization and
2598 // at vm shutdown.
2599 if (freelistLock()->owned_by_self()) {
2600 cmsSpace()->prepare_for_verify();
2601 } else {
2602 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2603 cmsSpace()->prepare_for_verify();
2604 }
2605 }
2606
2607 void
2608 ConcurrentMarkSweepGeneration::verify() {
2609 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2610 // are not called when the heap is verified during universe initialization and
2611 // at vm shutdown.
2612 if (freelistLock()->owned_by_self()) {
2613 cmsSpace()->verify();
2614 } else {
2615 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2616 cmsSpace()->verify();
2617 }
2618 }
2619
2620 void CMSCollector::verify() {
2621 _cmsGen->verify();
2622 }
2623
2624 #ifndef PRODUCT
2625 bool CMSCollector::overflow_list_is_empty() const {
2626 assert(_num_par_pushes >= 0, "Inconsistency");
2627 if (_overflow_list == NULL) {
2628 assert(_num_par_pushes == 0, "Inconsistency");
2629 }
2630 return _overflow_list == NULL;
2631 }
2632
2633 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2634 // merely consolidate assertion checks that appear to occur together frequently.
2635 void CMSCollector::verify_work_stacks_empty() const {
2636 assert(_markStack.isEmpty(), "Marking stack should be empty");
2637 assert(overflow_list_is_empty(), "Overflow list should be empty");
2638 }
2639
2640 void CMSCollector::verify_overflow_empty() const {
2641 assert(overflow_list_is_empty(), "Overflow list should be empty");
2642 assert(no_preserved_marks(), "No preserved marks");
2643 }
2644 #endif // PRODUCT
2645
2646 // Decide if we want to enable class unloading as part of the
2647 // ensuing concurrent GC cycle. We will collect and
2648 // unload classes if it's the case that:
2649 // (1) an explicit gc request has been made and the flag
2650 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
2651 // (2) (a) class unloading is enabled at the command line, and
2652 // (b) old gen is getting really full
2653 // NOTE: Provided there is no change in the state of the heap between
2654 // calls to this method, it should have idempotent results. Moreover,
2655 // its results should be monotonically increasing (i.e. going from 0 to 1,
2656 // but not 1 to 0) between successive calls between which the heap was
2657 // not collected. For the implementation below, it must thus rely on
2658 // the property that concurrent_cycles_since_last_unload()
2659 // will not decrease unless a collection cycle happened and that
2660 // _cmsGen->is_too_full() are
2661 // themselves also monotonic in that sense. See check_monotonicity()
2662 // below.
2663 void CMSCollector::update_should_unload_classes() {
2664 _should_unload_classes = false;
2665 // Condition 1 above
2666 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
2667 _should_unload_classes = true;
2668 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
2669 // Disjuncts 2.b.(i,ii,iii) above
2670 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2671 CMSClassUnloadingMaxInterval)
2672 || _cmsGen->is_too_full();
2673 }
2674 }
2675
2676 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2677 bool res = should_concurrent_collect();
2678 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2679 return res;
2680 }
2681
2682 void CMSCollector::setup_cms_unloading_and_verification_state() {
2683 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2684 || VerifyBeforeExit;
2685 const int rso = GenCollectedHeap::SO_AllCodeCache;
2686
2687 // We set the proper root for this CMS cycle here.
2688 if (should_unload_classes()) { // Should unload classes this cycle
2689 remove_root_scanning_option(rso); // Shrink the root set appropriately
2690 set_verifying(should_verify); // Set verification state for this cycle
2691 return; // Nothing else needs to be done at this time
2692 }
2693
2694 // Not unloading classes this cycle
2695 assert(!should_unload_classes(), "Inconsistency!");
2696
2697 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2698 // Include symbols, strings and code cache elements to prevent their resurrection.
2699 add_root_scanning_option(rso);
2700 set_verifying(true);
2701 } else if (verifying() && !should_verify) {
2702 // We were verifying, but some verification flags got disabled.
2703 set_verifying(false);
2704 // Exclude symbols, strings and code cache elements from root scanning to
2705 // reduce IM and RM pauses.
2706 remove_root_scanning_option(rso);
2707 }
2708 }
2709
2710
2711 #ifndef PRODUCT
2712 HeapWord* CMSCollector::block_start(const void* p) const {
2713 const HeapWord* addr = (HeapWord*)p;
2714 if (_span.contains(p)) {
2715 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2716 return _cmsGen->cmsSpace()->block_start(p);
2717 }
2718 }
2719 return NULL;
2720 }
2721 #endif
2722
2723 HeapWord*
2724 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2725 bool tlab,
2726 bool parallel) {
2727 CMSSynchronousYieldRequest yr;
2728 assert(!tlab, "Can't deal with TLAB allocation");
2729 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2730 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2731 if (GCExpandToAllocateDelayMillis > 0) {
2732 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2733 }
2734 return have_lock_and_allocate(word_size, tlab);
2735 }
2736
2737 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2738 size_t bytes,
2739 size_t expand_bytes,
2740 CMSExpansionCause::Cause cause)
2741 {
2742
2743 bool success = expand(bytes, expand_bytes);
2744
2745 // remember why we expanded; this information is used
2746 // by shouldConcurrentCollect() when making decisions on whether to start
2747 // a new CMS cycle.
2748 if (success) {
2749 set_expansion_cause(cause);
2750 if (PrintGCDetails && Verbose) {
2751 gclog_or_tty->print_cr("Expanded CMS gen for %s",
2752 CMSExpansionCause::to_string(cause));
2753 }
2754 }
2755 }
2756
2757 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2758 HeapWord* res = NULL;
2759 MutexLocker x(ParGCRareEvent_lock);
2760 while (true) {
2761 // Expansion by some other thread might make alloc OK now:
2762 res = ps->lab.alloc(word_sz);
2763 if (res != NULL) return res;
2764 // If there's not enough expansion space available, give up.
2765 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2766 return NULL;
2767 }
2768 // Otherwise, we try expansion.
2769 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2770 // Now go around the loop and try alloc again;
2771 // A competing par_promote might beat us to the expansion space,
2772 // so we may go around the loop again if promotion fails again.
2773 if (GCExpandToAllocateDelayMillis > 0) {
2774 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2775 }
2776 }
2777 }
2778
2779
2780 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2781 PromotionInfo* promo) {
2782 MutexLocker x(ParGCRareEvent_lock);
2783 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2784 while (true) {
2785 // Expansion by some other thread might make alloc OK now:
2786 if (promo->ensure_spooling_space()) {
2787 assert(promo->has_spooling_space(),
2788 "Post-condition of successful ensure_spooling_space()");
2789 return true;
2790 }
2791 // If there's not enough expansion space available, give up.
2792 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2793 return false;
2794 }
2795 // Otherwise, we try expansion.
2796 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2797 // Now go around the loop and try alloc again;
2798 // A competing allocation might beat us to the expansion space,
2799 // so we may go around the loop again if allocation fails again.
2800 if (GCExpandToAllocateDelayMillis > 0) {
2801 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2802 }
2803 }
2804 }
2805
2806 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2807 // Only shrink if a compaction was done so that all the free space
2808 // in the generation is in a contiguous block at the end.
2809 if (did_compact()) {
2810 CardGeneration::shrink(bytes);
2811 }
2812 }
2813
2814 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2815 assert_locked_or_safepoint(Heap_lock);
2816 }
2817
2818 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2819 assert_locked_or_safepoint(Heap_lock);
2820 assert_lock_strong(freelistLock());
2821 if (PrintGCDetails && Verbose) {
2822 warning("Shrinking of CMS not yet implemented");
2823 }
2824 return;
2825 }
2826
2827
2828 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2829 // phases.
2830 class CMSPhaseAccounting: public StackObj {
2831 public:
2832 CMSPhaseAccounting(CMSCollector *collector,
2833 const char *phase,
2834 const GCId gc_id,
2835 bool print_cr = true);
2836 ~CMSPhaseAccounting();
2837
2838 private:
2839 CMSCollector *_collector;
2840 const char *_phase;
2841 elapsedTimer _wallclock;
2842 bool _print_cr;
2843 const GCId _gc_id;
2844
2845 public:
2846 // Not MT-safe; so do not pass around these StackObj's
2847 // where they may be accessed by other threads.
2848 jlong wallclock_millis() {
2849 assert(_wallclock.is_active(), "Wall clock should not stop");
2850 _wallclock.stop(); // to record time
2851 jlong ret = _wallclock.milliseconds();
2852 _wallclock.start(); // restart
2853 return ret;
2854 }
2855 };
2856
2857 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2858 const char *phase,
2859 const GCId gc_id,
2860 bool print_cr) :
2861 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) {
2862
2863 if (PrintCMSStatistics != 0) {
2864 _collector->resetYields();
2865 }
2866 if (PrintGCDetails) {
2867 gclog_or_tty->gclog_stamp(_gc_id);
2868 gclog_or_tty->print_cr("[%s-concurrent-%s-start]",
2869 _collector->cmsGen()->short_name(), _phase);
2870 }
2871 _collector->resetTimer();
2872 _wallclock.start();
2873 _collector->startTimer();
2874 }
2875
2876 CMSPhaseAccounting::~CMSPhaseAccounting() {
2877 assert(_wallclock.is_active(), "Wall clock should not have stopped");
2878 _collector->stopTimer();
2879 _wallclock.stop();
2880 if (PrintGCDetails) {
2881 gclog_or_tty->gclog_stamp(_gc_id);
2882 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
2883 _collector->cmsGen()->short_name(),
2884 _phase, _collector->timerValue(), _wallclock.seconds());
2885 if (_print_cr) {
2886 gclog_or_tty->cr();
2887 }
2888 if (PrintCMSStatistics != 0) {
2889 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
2890 _collector->yields());
2891 }
2892 }
2893 }
2894
2895 // CMS work
2896
2897 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2898 class CMSParMarkTask : public AbstractGangTask {
2899 protected:
2900 CMSCollector* _collector;
2901 uint _n_workers;
2902 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2903 AbstractGangTask(name),
2904 _collector(collector),
2905 _n_workers(n_workers) {}
2906 // Work method in support of parallel rescan ... of young gen spaces
2907 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
2908 ContiguousSpace* space,
2909 HeapWord** chunk_array, size_t chunk_top);
2910 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
2911 };
2912
2913 // Parallel initial mark task
2914 class CMSParInitialMarkTask: public CMSParMarkTask {
2915 public:
2916 CMSParInitialMarkTask(CMSCollector* collector, uint n_workers) :
2917 CMSParMarkTask("Scan roots and young gen for initial mark in parallel",
2918 collector, n_workers) {}
2919 void work(uint worker_id);
2920 };
2921
2922 // Checkpoint the roots into this generation from outside
2923 // this generation. [Note this initial checkpoint need only
2924 // be approximate -- we'll do a catch up phase subsequently.]
2925 void CMSCollector::checkpointRootsInitial() {
2926 assert(_collectorState == InitialMarking, "Wrong collector state");
2927 check_correct_thread_executing();
2928 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
2929
2930 save_heap_summary();
2931 report_heap_summary(GCWhen::BeforeGC);
2932
2933 ReferenceProcessor* rp = ref_processor();
2934 assert(_restart_addr == NULL, "Control point invariant");
2935 {
2936 // acquire locks for subsequent manipulations
2937 MutexLockerEx x(bitMapLock(),
2938 Mutex::_no_safepoint_check_flag);
2939 checkpointRootsInitialWork();
2940 // enable ("weak") refs discovery
2941 rp->enable_discovery();
2942 _collectorState = Marking;
2943 }
2944 }
2945
2946 void CMSCollector::checkpointRootsInitialWork() {
2947 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2948 assert(_collectorState == InitialMarking, "just checking");
2949
2950 // If there has not been a GC[n-1] since last GC[n] cycle completed,
2951 // precede our marking with a collection of all
2952 // younger generations to keep floating garbage to a minimum.
2953 // XXX: we won't do this for now -- it's an optimization to be done later.
2954
2955 // already have locks
2956 assert_lock_strong(bitMapLock());
2957 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2958
2959 // Setup the verification and class unloading state for this
2960 // CMS collection cycle.
2961 setup_cms_unloading_and_verification_state();
2962
2963 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
2964 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
2965
2966 // Reset all the PLAB chunk arrays if necessary.
2967 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2968 reset_survivor_plab_arrays();
2969 }
2970
2971 ResourceMark rm;
2972 HandleMark hm;
2973
2974 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2975 GenCollectedHeap* gch = GenCollectedHeap::heap();
2976
2977 verify_work_stacks_empty();
2978 verify_overflow_empty();
2979
2980 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2981 // Update the saved marks which may affect the root scans.
2982 gch->save_marks();
2983
2984 // weak reference processing has not started yet.
2985 ref_processor()->set_enqueuing_is_done(false);
2986
2987 // Need to remember all newly created CLDs,
2988 // so that we can guarantee that the remark finds them.
2989 ClassLoaderDataGraph::remember_new_clds(true);
2990
2991 // Whenever a CLD is found, it will be claimed before proceeding to mark
2992 // the klasses. The claimed marks need to be cleared before marking starts.
2993 ClassLoaderDataGraph::clear_claimed_marks();
2994
2995 if (CMSPrintEdenSurvivorChunks) {
2996 print_eden_and_survivor_chunk_arrays();
2997 }
2998
2999 {
3000 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3001 if (CMSParallelInitialMarkEnabled) {
3002 // The parallel version.
3003 FlexibleWorkGang* workers = gch->workers();
3004 assert(workers != NULL, "Need parallel worker threads.");
3005 uint n_workers = workers->active_workers();
3006 CMSParInitialMarkTask tsk(this, n_workers);
3007 gch->set_par_threads(n_workers);
3008 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
3009 if (n_workers > 1) {
3010 StrongRootsScope srs;
3011 workers->run_task(&tsk);
3012 } else {
3013 StrongRootsScope srs;
3014 tsk.work(0);
3015 }
3016 gch->set_par_threads(0);
3017 } else {
3018 // The serial version.
3019 CLDToOopClosure cld_closure(¬Older, true);
3020 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3021 gch->gen_process_roots(_cmsGen->level(),
3022 true, // younger gens are roots
3023 true, // activate StrongRootsScope
3024 GenCollectedHeap::ScanningOption(roots_scanning_options()),
3025 should_unload_classes(),
3026 ¬Older,
3027 NULL,
3028 &cld_closure);
3029 }
3030 }
3031
3032 // Clear mod-union table; it will be dirtied in the prologue of
3033 // CMS generation per each younger generation collection.
3034
3035 assert(_modUnionTable.isAllClear(),
3036 "Was cleared in most recent final checkpoint phase"
3037 " or no bits are set in the gc_prologue before the start of the next "
3038 "subsequent marking phase.");
3039
3040 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
3041
3042 // Save the end of the used_region of the constituent generations
3043 // to be used to limit the extent of sweep in each generation.
3044 save_sweep_limits();
3045 verify_overflow_empty();
3046 }
3047
3048 bool CMSCollector::markFromRoots() {
3049 // we might be tempted to assert that:
3050 // assert(!SafepointSynchronize::is_at_safepoint(),
3051 // "inconsistent argument?");
3052 // However that wouldn't be right, because it's possible that
3053 // a safepoint is indeed in progress as a younger generation
3054 // stop-the-world GC happens even as we mark in this generation.
3055 assert(_collectorState == Marking, "inconsistent state?");
3056 check_correct_thread_executing();
3057 verify_overflow_empty();
3058
3059 // Weak ref discovery note: We may be discovering weak
3060 // refs in this generation concurrent (but interleaved) with
3061 // weak ref discovery by a younger generation collector.
3062
3063 CMSTokenSyncWithLocks ts(true, bitMapLock());
3064 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3065 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3066 bool res = markFromRootsWork();
3067 if (res) {
3068 _collectorState = Precleaning;
3069 } else { // We failed and a foreground collection wants to take over
3070 assert(_foregroundGCIsActive, "internal state inconsistency");
3071 assert(_restart_addr == NULL, "foreground will restart from scratch");
3072 if (PrintGCDetails) {
3073 gclog_or_tty->print_cr("bailing out to foreground collection");
3074 }
3075 }
3076 verify_overflow_empty();
3077 return res;
3078 }
3079
3080 bool CMSCollector::markFromRootsWork() {
3081 // iterate over marked bits in bit map, doing a full scan and mark
3082 // from these roots using the following algorithm:
3083 // . if oop is to the right of the current scan pointer,
3084 // mark corresponding bit (we'll process it later)
3085 // . else (oop is to left of current scan pointer)
3086 // push oop on marking stack
3087 // . drain the marking stack
3088
3089 // Note that when we do a marking step we need to hold the
3090 // bit map lock -- recall that direct allocation (by mutators)
3091 // and promotion (by younger generation collectors) is also
3092 // marking the bit map. [the so-called allocate live policy.]
3093 // Because the implementation of bit map marking is not
3094 // robust wrt simultaneous marking of bits in the same word,
3095 // we need to make sure that there is no such interference
3096 // between concurrent such updates.
3097
3098 // already have locks
3099 assert_lock_strong(bitMapLock());
3100
3101 verify_work_stacks_empty();
3102 verify_overflow_empty();
3103 bool result = false;
3104 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3105 result = do_marking_mt();
3106 } else {
3107 result = do_marking_st();
3108 }
3109 return result;
3110 }
3111
3112 // Forward decl
3113 class CMSConcMarkingTask;
3114
3115 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3116 CMSCollector* _collector;
3117 CMSConcMarkingTask* _task;
3118 public:
3119 virtual void yield();
3120
3121 // "n_threads" is the number of threads to be terminated.
3122 // "queue_set" is a set of work queues of other threads.
3123 // "collector" is the CMS collector associated with this task terminator.
3124 // "yield" indicates whether we need the gang as a whole to yield.
3125 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3126 ParallelTaskTerminator(n_threads, queue_set),
3127 _collector(collector) { }
3128
3129 void set_task(CMSConcMarkingTask* task) {
3130 _task = task;
3131 }
3132 };
3133
3134 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3135 CMSConcMarkingTask* _task;
3136 public:
3137 bool should_exit_termination();
3138 void set_task(CMSConcMarkingTask* task) {
3139 _task = task;
3140 }
3141 };
3142
3143 // MT Concurrent Marking Task
3144 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3145 CMSCollector* _collector;
3146 uint _n_workers; // requested/desired # workers
3147 bool _result;
3148 CompactibleFreeListSpace* _cms_space;
3149 char _pad_front[64]; // padding to ...
3150 HeapWord* _global_finger; // ... avoid sharing cache line
3151 char _pad_back[64];
3152 HeapWord* _restart_addr;
3153
3154 // Exposed here for yielding support
3155 Mutex* const _bit_map_lock;
3156
3157 // The per thread work queues, available here for stealing
3158 OopTaskQueueSet* _task_queues;
3159
3160 // Termination (and yielding) support
3161 CMSConcMarkingTerminator _term;
3162 CMSConcMarkingTerminatorTerminator _term_term;
3163
3164 public:
3165 CMSConcMarkingTask(CMSCollector* collector,
3166 CompactibleFreeListSpace* cms_space,
3167 YieldingFlexibleWorkGang* workers,
3168 OopTaskQueueSet* task_queues):
3169 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3170 _collector(collector),
3171 _cms_space(cms_space),
3172 _n_workers(0), _result(true),
3173 _task_queues(task_queues),
3174 _term(_n_workers, task_queues, _collector),
3175 _bit_map_lock(collector->bitMapLock())
3176 {
3177 _requested_size = _n_workers;
3178 _term.set_task(this);
3179 _term_term.set_task(this);
3180 _restart_addr = _global_finger = _cms_space->bottom();
3181 }
3182
3183
3184 OopTaskQueueSet* task_queues() { return _task_queues; }
3185
3186 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3187
3188 HeapWord** global_finger_addr() { return &_global_finger; }
3189
3190 CMSConcMarkingTerminator* terminator() { return &_term; }
3191
3192 virtual void set_for_termination(uint active_workers) {
3193 terminator()->reset_for_reuse(active_workers);
3194 }
3195
3196 void work(uint worker_id);
3197 bool should_yield() {
3198 return ConcurrentMarkSweepThread::should_yield()
3199 && !_collector->foregroundGCIsActive();
3200 }
3201
3202 virtual void coordinator_yield(); // stuff done by coordinator
3203 bool result() { return _result; }
3204
3205 void reset(HeapWord* ra) {
3206 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3207 _restart_addr = _global_finger = ra;
3208 _term.reset_for_reuse();
3209 }
3210
3211 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3212 OopTaskQueue* work_q);
3213
3214 private:
3215 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3216 void do_work_steal(int i);
3217 void bump_global_finger(HeapWord* f);
3218 };
3219
3220 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3221 assert(_task != NULL, "Error");
3222 return _task->yielding();
3223 // Note that we do not need the disjunct || _task->should_yield() above
3224 // because we want terminating threads to yield only if the task
3225 // is already in the midst of yielding, which happens only after at least one
3226 // thread has yielded.
3227 }
3228
3229 void CMSConcMarkingTerminator::yield() {
3230 if (_task->should_yield()) {
3231 _task->yield();
3232 } else {
3233 ParallelTaskTerminator::yield();
3234 }
3235 }
3236
3237 ////////////////////////////////////////////////////////////////
3238 // Concurrent Marking Algorithm Sketch
3239 ////////////////////////////////////////////////////////////////
3240 // Until all tasks exhausted (both spaces):
3241 // -- claim next available chunk
3242 // -- bump global finger via CAS
3243 // -- find first object that starts in this chunk
3244 // and start scanning bitmap from that position
3245 // -- scan marked objects for oops
3246 // -- CAS-mark target, and if successful:
3247 // . if target oop is above global finger (volatile read)
3248 // nothing to do
3249 // . if target oop is in chunk and above local finger
3250 // then nothing to do
3251 // . else push on work-queue
3252 // -- Deal with possible overflow issues:
3253 // . local work-queue overflow causes stuff to be pushed on
3254 // global (common) overflow queue
3255 // . always first empty local work queue
3256 // . then get a batch of oops from global work queue if any
3257 // . then do work stealing
3258 // -- When all tasks claimed (both spaces)
3259 // and local work queue empty,
3260 // then in a loop do:
3261 // . check global overflow stack; steal a batch of oops and trace
3262 // . try to steal from other threads oif GOS is empty
3263 // . if neither is available, offer termination
3264 // -- Terminate and return result
3265 //
3266 void CMSConcMarkingTask::work(uint worker_id) {
3267 elapsedTimer _timer;
3268 ResourceMark rm;
3269 HandleMark hm;
3270
3271 DEBUG_ONLY(_collector->verify_overflow_empty();)
3272
3273 // Before we begin work, our work queue should be empty
3274 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3275 // Scan the bitmap covering _cms_space, tracing through grey objects.
3276 _timer.start();
3277 do_scan_and_mark(worker_id, _cms_space);
3278 _timer.stop();
3279 if (PrintCMSStatistics != 0) {
3280 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3281 worker_id, _timer.seconds());
3282 // XXX: need xxx/xxx type of notation, two timers
3283 }
3284
3285 // ... do work stealing
3286 _timer.reset();
3287 _timer.start();
3288 do_work_steal(worker_id);
3289 _timer.stop();
3290 if (PrintCMSStatistics != 0) {
3291 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3292 worker_id, _timer.seconds());
3293 // XXX: need xxx/xxx type of notation, two timers
3294 }
3295 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3296 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3297 // Note that under the current task protocol, the
3298 // following assertion is true even of the spaces
3299 // expanded since the completion of the concurrent
3300 // marking. XXX This will likely change under a strict
3301 // ABORT semantics.
3302 // After perm removal the comparison was changed to
3303 // greater than or equal to from strictly greater than.
3304 // Before perm removal the highest address sweep would
3305 // have been at the end of perm gen but now is at the
3306 // end of the tenured gen.
3307 assert(_global_finger >= _cms_space->end(),
3308 "All tasks have been completed");
3309 DEBUG_ONLY(_collector->verify_overflow_empty();)
3310 }
3311
3312 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3313 HeapWord* read = _global_finger;
3314 HeapWord* cur = read;
3315 while (f > read) {
3316 cur = read;
3317 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3318 if (cur == read) {
3319 // our cas succeeded
3320 assert(_global_finger >= f, "protocol consistency");
3321 break;
3322 }
3323 }
3324 }
3325
3326 // This is really inefficient, and should be redone by
3327 // using (not yet available) block-read and -write interfaces to the
3328 // stack and the work_queue. XXX FIX ME !!!
3329 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3330 OopTaskQueue* work_q) {
3331 // Fast lock-free check
3332 if (ovflw_stk->length() == 0) {
3333 return false;
3334 }
3335 assert(work_q->size() == 0, "Shouldn't steal");
3336 MutexLockerEx ml(ovflw_stk->par_lock(),
3337 Mutex::_no_safepoint_check_flag);
3338 // Grab up to 1/4 the size of the work queue
3339 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3340 (size_t)ParGCDesiredObjsFromOverflowList);
3341 num = MIN2(num, ovflw_stk->length());
3342 for (int i = (int) num; i > 0; i--) {
3343 oop cur = ovflw_stk->pop();
3344 assert(cur != NULL, "Counted wrong?");
3345 work_q->push(cur);
3346 }
3347 return num > 0;
3348 }
3349
3350 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3351 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3352 int n_tasks = pst->n_tasks();
3353 // We allow that there may be no tasks to do here because
3354 // we are restarting after a stack overflow.
3355 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3356 uint nth_task = 0;
3357
3358 HeapWord* aligned_start = sp->bottom();
3359 if (sp->used_region().contains(_restart_addr)) {
3360 // Align down to a card boundary for the start of 0th task
3361 // for this space.
3362 aligned_start =
3363 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3364 CardTableModRefBS::card_size);
3365 }
3366
3367 size_t chunk_size = sp->marking_task_size();
3368 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3369 // Having claimed the nth task in this space,
3370 // compute the chunk that it corresponds to:
3371 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3372 aligned_start + (nth_task+1)*chunk_size);
3373 // Try and bump the global finger via a CAS;
3374 // note that we need to do the global finger bump
3375 // _before_ taking the intersection below, because
3376 // the task corresponding to that region will be
3377 // deemed done even if the used_region() expands
3378 // because of allocation -- as it almost certainly will
3379 // during start-up while the threads yield in the
3380 // closure below.
3381 HeapWord* finger = span.end();
3382 bump_global_finger(finger); // atomically
3383 // There are null tasks here corresponding to chunks
3384 // beyond the "top" address of the space.
3385 span = span.intersection(sp->used_region());
3386 if (!span.is_empty()) { // Non-null task
3387 HeapWord* prev_obj;
3388 assert(!span.contains(_restart_addr) || nth_task == 0,
3389 "Inconsistency");
3390 if (nth_task == 0) {
3391 // For the 0th task, we'll not need to compute a block_start.
3392 if (span.contains(_restart_addr)) {
3393 // In the case of a restart because of stack overflow,
3394 // we might additionally skip a chunk prefix.
3395 prev_obj = _restart_addr;
3396 } else {
3397 prev_obj = span.start();
3398 }
3399 } else {
3400 // We want to skip the first object because
3401 // the protocol is to scan any object in its entirety
3402 // that _starts_ in this span; a fortiori, any
3403 // object starting in an earlier span is scanned
3404 // as part of an earlier claimed task.
3405 // Below we use the "careful" version of block_start
3406 // so we do not try to navigate uninitialized objects.
3407 prev_obj = sp->block_start_careful(span.start());
3408 // Below we use a variant of block_size that uses the
3409 // Printezis bits to avoid waiting for allocated
3410 // objects to become initialized/parsable.
3411 while (prev_obj < span.start()) {
3412 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3413 if (sz > 0) {
3414 prev_obj += sz;
3415 } else {
3416 // In this case we may end up doing a bit of redundant
3417 // scanning, but that appears unavoidable, short of
3418 // locking the free list locks; see bug 6324141.
3419 break;
3420 }
3421 }
3422 }
3423 if (prev_obj < span.end()) {
3424 MemRegion my_span = MemRegion(prev_obj, span.end());
3425 // Do the marking work within a non-empty span --
3426 // the last argument to the constructor indicates whether the
3427 // iteration should be incremental with periodic yields.
3428 Par_MarkFromRootsClosure cl(this, _collector, my_span,
3429 &_collector->_markBitMap,
3430 work_queue(i),
3431 &_collector->_markStack);
3432 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3433 } // else nothing to do for this task
3434 } // else nothing to do for this task
3435 }
3436 // We'd be tempted to assert here that since there are no
3437 // more tasks left to claim in this space, the global_finger
3438 // must exceed space->top() and a fortiori space->end(). However,
3439 // that would not quite be correct because the bumping of
3440 // global_finger occurs strictly after the claiming of a task,
3441 // so by the time we reach here the global finger may not yet
3442 // have been bumped up by the thread that claimed the last
3443 // task.
3444 pst->all_tasks_completed();
3445 }
3446
3447 class Par_ConcMarkingClosure: public MetadataAwareOopClosure {
3448 private:
3449 CMSCollector* _collector;
3450 CMSConcMarkingTask* _task;
3451 MemRegion _span;
3452 CMSBitMap* _bit_map;
3453 CMSMarkStack* _overflow_stack;
3454 OopTaskQueue* _work_queue;
3455 protected:
3456 DO_OOP_WORK_DEFN
3457 public:
3458 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3459 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3460 MetadataAwareOopClosure(collector->ref_processor()),
3461 _collector(collector),
3462 _task(task),
3463 _span(collector->_span),
3464 _work_queue(work_queue),
3465 _bit_map(bit_map),
3466 _overflow_stack(overflow_stack)
3467 { }
3468 virtual void do_oop(oop* p);
3469 virtual void do_oop(narrowOop* p);
3470
3471 void trim_queue(size_t max);
3472 void handle_stack_overflow(HeapWord* lost);
3473 void do_yield_check() {
3474 if (_task->should_yield()) {
3475 _task->yield();
3476 }
3477 }
3478 };
3479
3480 // Grey object scanning during work stealing phase --
3481 // the salient assumption here is that any references
3482 // that are in these stolen objects being scanned must
3483 // already have been initialized (else they would not have
3484 // been published), so we do not need to check for
3485 // uninitialized objects before pushing here.
3486 void Par_ConcMarkingClosure::do_oop(oop obj) {
3487 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
3488 HeapWord* addr = (HeapWord*)obj;
3489 // Check if oop points into the CMS generation
3490 // and is not marked
3491 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3492 // a white object ...
3493 // If we manage to "claim" the object, by being the
3494 // first thread to mark it, then we push it on our
3495 // marking stack
3496 if (_bit_map->par_mark(addr)) { // ... now grey
3497 // push on work queue (grey set)
3498 bool simulate_overflow = false;
3499 NOT_PRODUCT(
3500 if (CMSMarkStackOverflowALot &&
3501 _collector->simulate_overflow()) {
3502 // simulate a stack overflow
3503 simulate_overflow = true;
3504 }
3505 )
3506 if (simulate_overflow ||
3507 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3508 // stack overflow
3509 if (PrintCMSStatistics != 0) {
3510 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
3511 SIZE_FORMAT, _overflow_stack->capacity());
3512 }
3513 // We cannot assert that the overflow stack is full because
3514 // it may have been emptied since.
3515 assert(simulate_overflow ||
3516 _work_queue->size() == _work_queue->max_elems(),
3517 "Else push should have succeeded");
3518 handle_stack_overflow(addr);
3519 }
3520 } // Else, some other thread got there first
3521 do_yield_check();
3522 }
3523 }
3524
3525 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3526 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3527
3528 void Par_ConcMarkingClosure::trim_queue(size_t max) {
3529 while (_work_queue->size() > max) {
3530 oop new_oop;
3531 if (_work_queue->pop_local(new_oop)) {
3532 assert(new_oop->is_oop(), "Should be an oop");
3533 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3534 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3535 new_oop->oop_iterate(this); // do_oop() above
3536 do_yield_check();
3537 }
3538 }
3539 }
3540
3541 // Upon stack overflow, we discard (part of) the stack,
3542 // remembering the least address amongst those discarded
3543 // in CMSCollector's _restart_address.
3544 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3545 // We need to do this under a mutex to prevent other
3546 // workers from interfering with the work done below.
3547 MutexLockerEx ml(_overflow_stack->par_lock(),
3548 Mutex::_no_safepoint_check_flag);
3549 // Remember the least grey address discarded
3550 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3551 _collector->lower_restart_addr(ra);
3552 _overflow_stack->reset(); // discard stack contents
3553 _overflow_stack->expand(); // expand the stack if possible
3554 }
3555
3556
3557 void CMSConcMarkingTask::do_work_steal(int i) {
3558 OopTaskQueue* work_q = work_queue(i);
3559 oop obj_to_scan;
3560 CMSBitMap* bm = &(_collector->_markBitMap);
3561 CMSMarkStack* ovflw = &(_collector->_markStack);
3562 int* seed = _collector->hash_seed(i);
3563 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3564 while (true) {
3565 cl.trim_queue(0);
3566 assert(work_q->size() == 0, "Should have been emptied above");
3567 if (get_work_from_overflow_stack(ovflw, work_q)) {
3568 // Can't assert below because the work obtained from the
3569 // overflow stack may already have been stolen from us.
3570 // assert(work_q->size() > 0, "Work from overflow stack");
3571 continue;
3572 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3573 assert(obj_to_scan->is_oop(), "Should be an oop");
3574 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3575 obj_to_scan->oop_iterate(&cl);
3576 } else if (terminator()->offer_termination(&_term_term)) {
3577 assert(work_q->size() == 0, "Impossible!");
3578 break;
3579 } else if (yielding() || should_yield()) {
3580 yield();
3581 }
3582 }
3583 }
3584
3585 // This is run by the CMS (coordinator) thread.
3586 void CMSConcMarkingTask::coordinator_yield() {
3587 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3588 "CMS thread should hold CMS token");
3589 // First give up the locks, then yield, then re-lock
3590 // We should probably use a constructor/destructor idiom to
3591 // do this unlock/lock or modify the MutexUnlocker class to
3592 // serve our purpose. XXX
3593 assert_lock_strong(_bit_map_lock);
3594 _bit_map_lock->unlock();
3595 ConcurrentMarkSweepThread::desynchronize(true);
3596 _collector->stopTimer();
3597 if (PrintCMSStatistics != 0) {
3598 _collector->incrementYields();
3599 }
3600
3601 // It is possible for whichever thread initiated the yield request
3602 // not to get a chance to wake up and take the bitmap lock between
3603 // this thread releasing it and reacquiring it. So, while the
3604 // should_yield() flag is on, let's sleep for a bit to give the
3605 // other thread a chance to wake up. The limit imposed on the number
3606 // of iterations is defensive, to avoid any unforseen circumstances
3607 // putting us into an infinite loop. Since it's always been this
3608 // (coordinator_yield()) method that was observed to cause the
3609 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3610 // which is by default non-zero. For the other seven methods that
3611 // also perform the yield operation, as are using a different
3612 // parameter (CMSYieldSleepCount) which is by default zero. This way we
3613 // can enable the sleeping for those methods too, if necessary.
3614 // See 6442774.
3615 //
3616 // We really need to reconsider the synchronization between the GC
3617 // thread and the yield-requesting threads in the future and we
3618 // should really use wait/notify, which is the recommended
3619 // way of doing this type of interaction. Additionally, we should
3620 // consolidate the eight methods that do the yield operation and they
3621 // are almost identical into one for better maintainability and
3622 // readability. See 6445193.
3623 //
3624 // Tony 2006.06.29
3625 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3626 ConcurrentMarkSweepThread::should_yield() &&
3627 !CMSCollector::foregroundGCIsActive(); ++i) {
3628 os::sleep(Thread::current(), 1, false);
3629 }
3630
3631 ConcurrentMarkSweepThread::synchronize(true);
3632 _bit_map_lock->lock_without_safepoint_check();
3633 _collector->startTimer();
3634 }
3635
3636 bool CMSCollector::do_marking_mt() {
3637 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3638 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3639 conc_workers()->active_workers(),
3640 Threads::number_of_non_daemon_threads());
3641 conc_workers()->set_active_workers(num_workers);
3642
3643 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
3644
3645 CMSConcMarkingTask tsk(this,
3646 cms_space,
3647 conc_workers(),
3648 task_queues());
3649
3650 // Since the actual number of workers we get may be different
3651 // from the number we requested above, do we need to do anything different
3652 // below? In particular, may be we need to subclass the SequantialSubTasksDone
3653 // class?? XXX
3654 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3655
3656 // Refs discovery is already non-atomic.
3657 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3658 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3659 conc_workers()->start_task(&tsk);
3660 while (tsk.yielded()) {
3661 tsk.coordinator_yield();
3662 conc_workers()->continue_task(&tsk);
3663 }
3664 // If the task was aborted, _restart_addr will be non-NULL
3665 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3666 while (_restart_addr != NULL) {
3667 // XXX For now we do not make use of ABORTED state and have not
3668 // yet implemented the right abort semantics (even in the original
3669 // single-threaded CMS case). That needs some more investigation
3670 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3671 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3672 // If _restart_addr is non-NULL, a marking stack overflow
3673 // occurred; we need to do a fresh marking iteration from the
3674 // indicated restart address.
3675 if (_foregroundGCIsActive) {
3676 // We may be running into repeated stack overflows, having
3677 // reached the limit of the stack size, while making very
3678 // slow forward progress. It may be best to bail out and
3679 // let the foreground collector do its job.
3680 // Clear _restart_addr, so that foreground GC
3681 // works from scratch. This avoids the headache of
3682 // a "rescan" which would otherwise be needed because
3683 // of the dirty mod union table & card table.
3684 _restart_addr = NULL;
3685 return false;
3686 }
3687 // Adjust the task to restart from _restart_addr
3688 tsk.reset(_restart_addr);
3689 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3690 _restart_addr);
3691 _restart_addr = NULL;
3692 // Get the workers going again
3693 conc_workers()->start_task(&tsk);
3694 while (tsk.yielded()) {
3695 tsk.coordinator_yield();
3696 conc_workers()->continue_task(&tsk);
3697 }
3698 }
3699 assert(tsk.completed(), "Inconsistency");
3700 assert(tsk.result() == true, "Inconsistency");
3701 return true;
3702 }
3703
3704 bool CMSCollector::do_marking_st() {
3705 ResourceMark rm;
3706 HandleMark hm;
3707
3708 // Temporarily make refs discovery single threaded (non-MT)
3709 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3710 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3711 &_markStack, CMSYield);
3712 // the last argument to iterate indicates whether the iteration
3713 // should be incremental with periodic yields.
3714 _markBitMap.iterate(&markFromRootsClosure);
3715 // If _restart_addr is non-NULL, a marking stack overflow
3716 // occurred; we need to do a fresh iteration from the
3717 // indicated restart address.
3718 while (_restart_addr != NULL) {
3719 if (_foregroundGCIsActive) {
3720 // We may be running into repeated stack overflows, having
3721 // reached the limit of the stack size, while making very
3722 // slow forward progress. It may be best to bail out and
3723 // let the foreground collector do its job.
3724 // Clear _restart_addr, so that foreground GC
3725 // works from scratch. This avoids the headache of
3726 // a "rescan" which would otherwise be needed because
3727 // of the dirty mod union table & card table.
3728 _restart_addr = NULL;
3729 return false; // indicating failure to complete marking
3730 }
3731 // Deal with stack overflow:
3732 // we restart marking from _restart_addr
3733 HeapWord* ra = _restart_addr;
3734 markFromRootsClosure.reset(ra);
3735 _restart_addr = NULL;
3736 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3737 }
3738 return true;
3739 }
3740
3741 void CMSCollector::preclean() {
3742 check_correct_thread_executing();
3743 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3744 verify_work_stacks_empty();
3745 verify_overflow_empty();
3746 _abort_preclean = false;
3747 if (CMSPrecleaningEnabled) {
3748 if (!CMSEdenChunksRecordAlways) {
3749 _eden_chunk_index = 0;
3750 }
3751 size_t used = get_eden_used();
3752 size_t capacity = get_eden_capacity();
3753 // Don't start sampling unless we will get sufficiently
3754 // many samples.
3755 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
3756 * CMSScheduleRemarkEdenPenetration)) {
3757 _start_sampling = true;
3758 } else {
3759 _start_sampling = false;
3760 }
3761 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3762 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3763 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3764 }
3765 CMSTokenSync x(true); // is cms thread
3766 if (CMSPrecleaningEnabled) {
3767 sample_eden();
3768 _collectorState = AbortablePreclean;
3769 } else {
3770 _collectorState = FinalMarking;
3771 }
3772 verify_work_stacks_empty();
3773 verify_overflow_empty();
3774 }
3775
3776 // Try and schedule the remark such that young gen
3777 // occupancy is CMSScheduleRemarkEdenPenetration %.
3778 void CMSCollector::abortable_preclean() {
3779 check_correct_thread_executing();
3780 assert(CMSPrecleaningEnabled, "Inconsistent control state");
3781 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3782
3783 // If Eden's current occupancy is below this threshold,
3784 // immediately schedule the remark; else preclean
3785 // past the next scavenge in an effort to
3786 // schedule the pause as described above. By choosing
3787 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3788 // we will never do an actual abortable preclean cycle.
3789 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3790 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3791 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3792 // We need more smarts in the abortable preclean
3793 // loop below to deal with cases where allocation
3794 // in young gen is very very slow, and our precleaning
3795 // is running a losing race against a horde of
3796 // mutators intent on flooding us with CMS updates
3797 // (dirty cards).
3798 // One, admittedly dumb, strategy is to give up
3799 // after a certain number of abortable precleaning loops
3800 // or after a certain maximum time. We want to make
3801 // this smarter in the next iteration.
3802 // XXX FIX ME!!! YSR
3803 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3804 while (!(should_abort_preclean() ||
3805 ConcurrentMarkSweepThread::should_terminate())) {
3806 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3807 cumworkdone += workdone;
3808 loops++;
3809 // Voluntarily terminate abortable preclean phase if we have
3810 // been at it for too long.
3811 if ((CMSMaxAbortablePrecleanLoops != 0) &&
3812 loops >= CMSMaxAbortablePrecleanLoops) {
3813 if (PrintGCDetails) {
3814 gclog_or_tty->print(" CMS: abort preclean due to loops ");
3815 }
3816 break;
3817 }
3818 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3819 if (PrintGCDetails) {
3820 gclog_or_tty->print(" CMS: abort preclean due to time ");
3821 }
3822 break;
3823 }
3824 // If we are doing little work each iteration, we should
3825 // take a short break.
3826 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3827 // Sleep for some time, waiting for work to accumulate
3828 stopTimer();
3829 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3830 startTimer();
3831 waited++;
3832 }
3833 }
3834 if (PrintCMSStatistics > 0) {
3835 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3836 loops, waited, cumworkdone);
3837 }
3838 }
3839 CMSTokenSync x(true); // is cms thread
3840 if (_collectorState != Idling) {
3841 assert(_collectorState == AbortablePreclean,
3842 "Spontaneous state transition?");
3843 _collectorState = FinalMarking;
3844 } // Else, a foreground collection completed this CMS cycle.
3845 return;
3846 }
3847
3848 // Respond to an Eden sampling opportunity
3849 void CMSCollector::sample_eden() {
3850 // Make sure a young gc cannot sneak in between our
3851 // reading and recording of a sample.
3852 assert(Thread::current()->is_ConcurrentGC_thread(),
3853 "Only the cms thread may collect Eden samples");
3854 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3855 "Should collect samples while holding CMS token");
3856 if (!_start_sampling) {
3857 return;
3858 }
3859 // When CMSEdenChunksRecordAlways is true, the eden chunk array
3860 // is populated by the young generation.
3861 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3862 if (_eden_chunk_index < _eden_chunk_capacity) {
3863 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
3864 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3865 "Unexpected state of Eden");
3866 // We'd like to check that what we just sampled is an oop-start address;
3867 // however, we cannot do that here since the object may not yet have been
3868 // initialized. So we'll instead do the check when we _use_ this sample
3869 // later.
3870 if (_eden_chunk_index == 0 ||
3871 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3872 _eden_chunk_array[_eden_chunk_index-1])
3873 >= CMSSamplingGrain)) {
3874 _eden_chunk_index++; // commit sample
3875 }
3876 }
3877 }
3878 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3879 size_t used = get_eden_used();
3880 size_t capacity = get_eden_capacity();
3881 assert(used <= capacity, "Unexpected state of Eden");
3882 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3883 _abort_preclean = true;
3884 }
3885 }
3886 }
3887
3888
3889 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3890 assert(_collectorState == Precleaning ||
3891 _collectorState == AbortablePreclean, "incorrect state");
3892 ResourceMark rm;
3893 HandleMark hm;
3894
3895 // Precleaning is currently not MT but the reference processor
3896 // may be set for MT. Disable it temporarily here.
3897 ReferenceProcessor* rp = ref_processor();
3898 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3899
3900 // Do one pass of scrubbing the discovered reference lists
3901 // to remove any reference objects with strongly-reachable
3902 // referents.
3903 if (clean_refs) {
3904 CMSPrecleanRefsYieldClosure yield_cl(this);
3905 assert(rp->span().equals(_span), "Spans should be equal");
3906 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3907 &_markStack, true /* preclean */);
3908 CMSDrainMarkingStackClosure complete_trace(this,
3909 _span, &_markBitMap, &_markStack,
3910 &keep_alive, true /* preclean */);
3911
3912 // We don't want this step to interfere with a young
3913 // collection because we don't want to take CPU
3914 // or memory bandwidth away from the young GC threads
3915 // (which may be as many as there are CPUs).
3916 // Note that we don't need to protect ourselves from
3917 // interference with mutators because they can't
3918 // manipulate the discovered reference lists nor affect
3919 // the computed reachability of the referents, the
3920 // only properties manipulated by the precleaning
3921 // of these reference lists.
3922 stopTimer();
3923 CMSTokenSyncWithLocks x(true /* is cms thread */,
3924 bitMapLock());
3925 startTimer();
3926 sample_eden();
3927
3928 // The following will yield to allow foreground
3929 // collection to proceed promptly. XXX YSR:
3930 // The code in this method may need further
3931 // tweaking for better performance and some restructuring
3932 // for cleaner interfaces.
3933 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3934 rp->preclean_discovered_references(
3935 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3936 gc_timer, _gc_tracer_cm->gc_id());
3937 }
3938
3939 if (clean_survivor) { // preclean the active survivor space(s)
3940 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3941 &_markBitMap, &_modUnionTable,
3942 &_markStack, true /* precleaning phase */);
3943 stopTimer();
3944 CMSTokenSyncWithLocks ts(true /* is cms thread */,
3945 bitMapLock());
3946 startTimer();
3947 unsigned int before_count =
3948 GenCollectedHeap::heap()->total_collections();
3949 SurvivorSpacePrecleanClosure
3950 sss_cl(this, _span, &_markBitMap, &_markStack,
3951 &pam_cl, before_count, CMSYield);
3952 _young_gen->from()->object_iterate_careful(&sss_cl);
3953 _young_gen->to()->object_iterate_careful(&sss_cl);
3954 }
3955 MarkRefsIntoAndScanClosure
3956 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3957 &_markStack, this, CMSYield,
3958 true /* precleaning phase */);
3959 // CAUTION: The following closure has persistent state that may need to
3960 // be reset upon a decrease in the sequence of addresses it
3961 // processes.
3962 ScanMarkedObjectsAgainCarefullyClosure
3963 smoac_cl(this, _span,
3964 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3965
3966 // Preclean dirty cards in ModUnionTable and CardTable using
3967 // appropriate convergence criterion;
3968 // repeat CMSPrecleanIter times unless we find that
3969 // we are losing.
3970 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3971 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3972 "Bad convergence multiplier");
3973 assert(CMSPrecleanThreshold >= 100,
3974 "Unreasonably low CMSPrecleanThreshold");
3975
3976 size_t numIter, cumNumCards, lastNumCards, curNumCards;
3977 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3978 numIter < CMSPrecleanIter;
3979 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3980 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
3981 if (Verbose && PrintGCDetails) {
3982 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3983 }
3984 // Either there are very few dirty cards, so re-mark
3985 // pause will be small anyway, or our pre-cleaning isn't
3986 // that much faster than the rate at which cards are being
3987 // dirtied, so we might as well stop and re-mark since
3988 // precleaning won't improve our re-mark time by much.
3989 if (curNumCards <= CMSPrecleanThreshold ||
3990 (numIter > 0 &&
3991 (curNumCards * CMSPrecleanDenominator >
3992 lastNumCards * CMSPrecleanNumerator))) {
3993 numIter++;
3994 cumNumCards += curNumCards;
3995 break;
3996 }
3997 }
3998
3999 preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
4000
4001 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4002 cumNumCards += curNumCards;
4003 if (PrintGCDetails && PrintCMSStatistics != 0) {
4004 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
4005 curNumCards, cumNumCards, numIter);
4006 }
4007 return cumNumCards; // as a measure of useful work done
4008 }
4009
4010 // PRECLEANING NOTES:
4011 // Precleaning involves:
4012 // . reading the bits of the modUnionTable and clearing the set bits.
4013 // . For the cards corresponding to the set bits, we scan the
4014 // objects on those cards. This means we need the free_list_lock
4015 // so that we can safely iterate over the CMS space when scanning
4016 // for oops.
4017 // . When we scan the objects, we'll be both reading and setting
4018 // marks in the marking bit map, so we'll need the marking bit map.
4019 // . For protecting _collector_state transitions, we take the CGC_lock.
4020 // Note that any races in the reading of of card table entries by the
4021 // CMS thread on the one hand and the clearing of those entries by the
4022 // VM thread or the setting of those entries by the mutator threads on the
4023 // other are quite benign. However, for efficiency it makes sense to keep
4024 // the VM thread from racing with the CMS thread while the latter is
4025 // dirty card info to the modUnionTable. We therefore also use the
4026 // CGC_lock to protect the reading of the card table and the mod union
4027 // table by the CM thread.
4028 // . We run concurrently with mutator updates, so scanning
4029 // needs to be done carefully -- we should not try to scan
4030 // potentially uninitialized objects.
4031 //
4032 // Locking strategy: While holding the CGC_lock, we scan over and
4033 // reset a maximal dirty range of the mod union / card tables, then lock
4034 // the free_list_lock and bitmap lock to do a full marking, then
4035 // release these locks; and repeat the cycle. This allows for a
4036 // certain amount of fairness in the sharing of these locks between
4037 // the CMS collector on the one hand, and the VM thread and the
4038 // mutators on the other.
4039
4040 // NOTE: preclean_mod_union_table() and preclean_card_table()
4041 // further below are largely identical; if you need to modify
4042 // one of these methods, please check the other method too.
4043
4044 size_t CMSCollector::preclean_mod_union_table(
4045 ConcurrentMarkSweepGeneration* gen,
4046 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4047 verify_work_stacks_empty();
4048 verify_overflow_empty();
4049
4050 // strategy: starting with the first card, accumulate contiguous
4051 // ranges of dirty cards; clear these cards, then scan the region
4052 // covered by these cards.
4053
4054 // Since all of the MUT is committed ahead, we can just use
4055 // that, in case the generations expand while we are precleaning.
4056 // It might also be fine to just use the committed part of the
4057 // generation, but we might potentially miss cards when the
4058 // generation is rapidly expanding while we are in the midst
4059 // of precleaning.
4060 HeapWord* startAddr = gen->reserved().start();
4061 HeapWord* endAddr = gen->reserved().end();
4062
4063 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4064
4065 size_t numDirtyCards, cumNumDirtyCards;
4066 HeapWord *nextAddr, *lastAddr;
4067 for (cumNumDirtyCards = numDirtyCards = 0,
4068 nextAddr = lastAddr = startAddr;
4069 nextAddr < endAddr;
4070 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4071
4072 ResourceMark rm;
4073 HandleMark hm;
4074
4075 MemRegion dirtyRegion;
4076 {
4077 stopTimer();
4078 // Potential yield point
4079 CMSTokenSync ts(true);
4080 startTimer();
4081 sample_eden();
4082 // Get dirty region starting at nextOffset (inclusive),
4083 // simultaneously clearing it.
4084 dirtyRegion =
4085 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4086 assert(dirtyRegion.start() >= nextAddr,
4087 "returned region inconsistent?");
4088 }
4089 // Remember where the next search should begin.
4090 // The returned region (if non-empty) is a right open interval,
4091 // so lastOffset is obtained from the right end of that
4092 // interval.
4093 lastAddr = dirtyRegion.end();
4094 // Should do something more transparent and less hacky XXX
4095 numDirtyCards =
4096 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4097
4098 // We'll scan the cards in the dirty region (with periodic
4099 // yields for foreground GC as needed).
4100 if (!dirtyRegion.is_empty()) {
4101 if (UseCondCardMark) {
4102 // This global fence serializes all stale memory accesses in the process
4103 // allowing fence elision in the mutator write barrier.
4104 // It protects against reading stale references after writing the precleaned card value.
4105 OrderAccess::global_fence();
4106 }
4107
4108 assert(numDirtyCards > 0, "consistency check");
4109 HeapWord* stop_point = NULL;
4110 stopTimer();
4111 // Potential yield point
4112 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4113 bitMapLock());
4114 startTimer();
4115 {
4116 verify_work_stacks_empty();
4117 verify_overflow_empty();
4118 sample_eden();
4119 stop_point =
4120 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4121 }
4122 if (stop_point != NULL) {
4123 // The careful iteration stopped early either because it found an
4124 // uninitialized object, or because we were in the midst of an
4125 // "abortable preclean", which should now be aborted. Redirty
4126 // the bits corresponding to the partially-scanned or unscanned
4127 // cards. We'll either restart at the next block boundary or
4128 // abort the preclean.
4129 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4130 "Should only be AbortablePreclean.");
4131 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4132 if (should_abort_preclean()) {
4133 break; // out of preclean loop
4134 } else {
4135 // Compute the next address at which preclean should pick up;
4136 // might need bitMapLock in order to read P-bits.
4137 lastAddr = next_card_start_after_block(stop_point);
4138 }
4139 }
4140 } else {
4141 assert(lastAddr == endAddr, "consistency check");
4142 assert(numDirtyCards == 0, "consistency check");
4143 break;
4144 }
4145 }
4146 verify_work_stacks_empty();
4147 verify_overflow_empty();
4148 return cumNumDirtyCards;
4149 }
4150
4151 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4152 // below are largely identical; if you need to modify
4153 // one of these methods, please check the other method too.
4154
4155 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4156 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4157 // strategy: it's similar to precleamModUnionTable above, in that
4158 // we accumulate contiguous ranges of dirty cards, mark these cards
4159 // precleaned, then scan the region covered by these cards.
4160 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4161 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4162
4163 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4164
4165 size_t numDirtyCards, cumNumDirtyCards;
4166 HeapWord *lastAddr, *nextAddr;
4167
4168 for (cumNumDirtyCards = numDirtyCards = 0,
4169 nextAddr = lastAddr = startAddr;
4170 nextAddr < endAddr;
4171 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4172
4173 ResourceMark rm;
4174 HandleMark hm;
4175
4176 MemRegion dirtyRegion;
4177 {
4178 // See comments in "Precleaning notes" above on why we
4179 // do this locking. XXX Could the locking overheads be
4180 // too high when dirty cards are sparse? [I don't think so.]
4181 stopTimer();
4182 CMSTokenSync x(true); // is cms thread
4183 startTimer();
4184 sample_eden();
4185 // Get and clear dirty region from card table
4186 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4187 MemRegion(nextAddr, endAddr),
4188 true,
4189 CardTableModRefBS::precleaned_card_val());
4190
4191 assert(dirtyRegion.start() >= nextAddr,
4192 "returned region inconsistent?");
4193 }
4194 lastAddr = dirtyRegion.end();
4195 numDirtyCards =
4196 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4197
4198 if (!dirtyRegion.is_empty()) {
4199 if (UseCondCardMark) {
4200 // This global fence serializes all stale memory accesses in the process
4201 // allowing fence elision in the mutator write barrier.
4202 // It protects against reading stale references after writing the precleaned card value.
4203 OrderAccess::global_fence();
4204 }
4205
4206 stopTimer();
4207 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4208 startTimer();
4209 sample_eden();
4210 verify_work_stacks_empty();
4211 verify_overflow_empty();
4212 HeapWord* stop_point =
4213 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4214 if (stop_point != NULL) {
4215 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4216 "Should only be AbortablePreclean.");
4217 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4218 if (should_abort_preclean()) {
4219 break; // out of preclean loop
4220 } else {
4221 // Compute the next address at which preclean should pick up.
4222 lastAddr = next_card_start_after_block(stop_point);
4223 }
4224 }
4225 } else {
4226 break;
4227 }
4228 }
4229 verify_work_stacks_empty();
4230 verify_overflow_empty();
4231 return cumNumDirtyCards;
4232 }
4233
4234 class PrecleanKlassClosure : public KlassClosure {
4235 KlassToOopClosure _cm_klass_closure;
4236 public:
4237 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4238 void do_klass(Klass* k) {
4239 if (k->has_accumulated_modified_oops()) {
4240 k->clear_accumulated_modified_oops();
4241
4242 _cm_klass_closure.do_klass(k);
4243 }
4244 }
4245 };
4246
4247 // The freelist lock is needed to prevent asserts, is it really needed?
4248 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4249
4250 cl->set_freelistLock(freelistLock);
4251
4252 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4253
4254 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4255 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4256 PrecleanKlassClosure preclean_klass_closure(cl);
4257 ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
4258
4259 verify_work_stacks_empty();
4260 verify_overflow_empty();
4261 }
4262
4263 void CMSCollector::checkpointRootsFinal() {
4264 assert(_collectorState == FinalMarking, "incorrect state transition?");
4265 check_correct_thread_executing();
4266 // world is stopped at this checkpoint
4267 assert(SafepointSynchronize::is_at_safepoint(),
4268 "world should be stopped");
4269 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4270
4271 verify_work_stacks_empty();
4272 verify_overflow_empty();
4273
4274 if (PrintGCDetails) {
4275 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4276 _young_gen->used() / K,
4277 _young_gen->capacity() / K);
4278 }
4279 {
4280 if (CMSScavengeBeforeRemark) {
4281 GenCollectedHeap* gch = GenCollectedHeap::heap();
4282 // Temporarily set flag to false, GCH->do_collection will
4283 // expect it to be false and set to true
4284 FlagSetting fl(gch->_is_gc_active, false);
4285 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
4286 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4287 int level = _cmsGen->level() - 1;
4288 if (level >= 0) {
4289 gch->do_collection(true, // full (i.e. force, see below)
4290 false, // !clear_all_soft_refs
4291 0, // size
4292 false, // is_tlab
4293 level // max_level
4294 );
4295 }
4296 }
4297 FreelistLocker x(this);
4298 MutexLockerEx y(bitMapLock(),
4299 Mutex::_no_safepoint_check_flag);
4300 checkpointRootsFinalWork();
4301 }
4302 verify_work_stacks_empty();
4303 verify_overflow_empty();
4304 }
4305
4306 void CMSCollector::checkpointRootsFinalWork() {
4307 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4308
4309 assert(haveFreelistLocks(), "must have free list locks");
4310 assert_lock_strong(bitMapLock());
4311
4312 ResourceMark rm;
4313 HandleMark hm;
4314
4315 GenCollectedHeap* gch = GenCollectedHeap::heap();
4316
4317 if (should_unload_classes()) {
4318 CodeCache::gc_prologue();
4319 }
4320 assert(haveFreelistLocks(), "must have free list locks");
4321 assert_lock_strong(bitMapLock());
4322
4323 // We might assume that we need not fill TLAB's when
4324 // CMSScavengeBeforeRemark is set, because we may have just done
4325 // a scavenge which would have filled all TLAB's -- and besides
4326 // Eden would be empty. This however may not always be the case --
4327 // for instance although we asked for a scavenge, it may not have
4328 // happened because of a JNI critical section. We probably need
4329 // a policy for deciding whether we can in that case wait until
4330 // the critical section releases and then do the remark following
4331 // the scavenge, and skip it here. In the absence of that policy,
4332 // or of an indication of whether the scavenge did indeed occur,
4333 // we cannot rely on TLAB's having been filled and must do
4334 // so here just in case a scavenge did not happen.
4335 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4336 // Update the saved marks which may affect the root scans.
4337 gch->save_marks();
4338
4339 if (CMSPrintEdenSurvivorChunks) {
4340 print_eden_and_survivor_chunk_arrays();
4341 }
4342
4343 {
4344 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4345
4346 // Note on the role of the mod union table:
4347 // Since the marker in "markFromRoots" marks concurrently with
4348 // mutators, it is possible for some reachable objects not to have been
4349 // scanned. For instance, an only reference to an object A was
4350 // placed in object B after the marker scanned B. Unless B is rescanned,
4351 // A would be collected. Such updates to references in marked objects
4352 // are detected via the mod union table which is the set of all cards
4353 // dirtied since the first checkpoint in this GC cycle and prior to
4354 // the most recent young generation GC, minus those cleaned up by the
4355 // concurrent precleaning.
4356 if (CMSParallelRemarkEnabled) {
4357 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
4358 do_remark_parallel();
4359 } else {
4360 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4361 _gc_timer_cm, _gc_tracer_cm->gc_id());
4362 do_remark_non_parallel();
4363 }
4364 }
4365 verify_work_stacks_empty();
4366 verify_overflow_empty();
4367
4368 {
4369 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4370 refProcessingWork();
4371 }
4372 verify_work_stacks_empty();
4373 verify_overflow_empty();
4374
4375 if (should_unload_classes()) {
4376 CodeCache::gc_epilogue();
4377 }
4378 JvmtiExport::gc_epilogue();
4379
4380 // If we encountered any (marking stack / work queue) overflow
4381 // events during the current CMS cycle, take appropriate
4382 // remedial measures, where possible, so as to try and avoid
4383 // recurrence of that condition.
4384 assert(_markStack.isEmpty(), "No grey objects");
4385 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4386 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4387 if (ser_ovflw > 0) {
4388 if (PrintCMSStatistics != 0) {
4389 gclog_or_tty->print_cr("Marking stack overflow (benign) "
4390 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
4391 ", kac_preclean="SIZE_FORMAT")",
4392 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4393 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4394 }
4395 _markStack.expand();
4396 _ser_pmc_remark_ovflw = 0;
4397 _ser_pmc_preclean_ovflw = 0;
4398 _ser_kac_preclean_ovflw = 0;
4399 _ser_kac_ovflw = 0;
4400 }
4401 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4402 if (PrintCMSStatistics != 0) {
4403 gclog_or_tty->print_cr("Work queue overflow (benign) "
4404 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4405 _par_pmc_remark_ovflw, _par_kac_ovflw);
4406 }
4407 _par_pmc_remark_ovflw = 0;
4408 _par_kac_ovflw = 0;
4409 }
4410 if (PrintCMSStatistics != 0) {
4411 if (_markStack._hit_limit > 0) {
4412 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
4413 _markStack._hit_limit);
4414 }
4415 if (_markStack._failed_double > 0) {
4416 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
4417 " current capacity "SIZE_FORMAT,
4418 _markStack._failed_double,
4419 _markStack.capacity());
4420 }
4421 }
4422 _markStack._hit_limit = 0;
4423 _markStack._failed_double = 0;
4424
4425 if ((VerifyAfterGC || VerifyDuringGC) &&
4426 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4427 verify_after_remark();
4428 }
4429
4430 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4431
4432 // Change under the freelistLocks.
4433 _collectorState = Sweeping;
4434 // Call isAllClear() under bitMapLock
4435 assert(_modUnionTable.isAllClear(),
4436 "Should be clear by end of the final marking");
4437 assert(_ct->klass_rem_set()->mod_union_is_clear(),
4438 "Should be clear by end of the final marking");
4439 }
4440
4441 void CMSParInitialMarkTask::work(uint worker_id) {
4442 elapsedTimer _timer;
4443 ResourceMark rm;
4444 HandleMark hm;
4445
4446 // ---------- scan from roots --------------
4447 _timer.start();
4448 GenCollectedHeap* gch = GenCollectedHeap::heap();
4449 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4450
4451 // ---------- young gen roots --------------
4452 {
4453 work_on_young_gen_roots(worker_id, &par_mri_cl);
4454 _timer.stop();
4455 if (PrintCMSStatistics != 0) {
4456 gclog_or_tty->print_cr(
4457 "Finished young gen initial mark scan work in %dth thread: %3.3f sec",
4458 worker_id, _timer.seconds());
4459 }
4460 }
4461
4462 // ---------- remaining roots --------------
4463 _timer.reset();
4464 _timer.start();
4465
4466 CLDToOopClosure cld_closure(&par_mri_cl, true);
4467
4468 gch->gen_process_roots(_collector->_cmsGen->level(),
4469 false, // yg was scanned above
4470 false, // this is parallel code
4471 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4472 _collector->should_unload_classes(),
4473 &par_mri_cl,
4474 NULL,
4475 &cld_closure);
4476 assert(_collector->should_unload_classes()
4477 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4478 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4479 _timer.stop();
4480 if (PrintCMSStatistics != 0) {
4481 gclog_or_tty->print_cr(
4482 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec",
4483 worker_id, _timer.seconds());
4484 }
4485 }
4486
4487 // Parallel remark task
4488 class CMSParRemarkTask: public CMSParMarkTask {
4489 CompactibleFreeListSpace* _cms_space;
4490
4491 // The per-thread work queues, available here for stealing.
4492 OopTaskQueueSet* _task_queues;
4493 ParallelTaskTerminator _term;
4494
4495 public:
4496 // A value of 0 passed to n_workers will cause the number of
4497 // workers to be taken from the active workers in the work gang.
4498 CMSParRemarkTask(CMSCollector* collector,
4499 CompactibleFreeListSpace* cms_space,
4500 uint n_workers, FlexibleWorkGang* workers,
4501 OopTaskQueueSet* task_queues):
4502 CMSParMarkTask("Rescan roots and grey objects in parallel",
4503 collector, n_workers),
4504 _cms_space(cms_space),
4505 _task_queues(task_queues),
4506 _term(n_workers, task_queues) { }
4507
4508 OopTaskQueueSet* task_queues() { return _task_queues; }
4509
4510 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4511
4512 ParallelTaskTerminator* terminator() { return &_term; }
4513 uint n_workers() { return _n_workers; }
4514
4515 void work(uint worker_id);
4516
4517 private:
4518 // ... of dirty cards in old space
4519 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4520 Par_MarkRefsIntoAndScanClosure* cl);
4521
4522 // ... work stealing for the above
4523 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4524 };
4525
4526 class RemarkKlassClosure : public KlassClosure {
4527 KlassToOopClosure _cm_klass_closure;
4528 public:
4529 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4530 void do_klass(Klass* k) {
4531 // Check if we have modified any oops in the Klass during the concurrent marking.
4532 if (k->has_accumulated_modified_oops()) {
4533 k->clear_accumulated_modified_oops();
4534
4535 // We could have transfered the current modified marks to the accumulated marks,
4536 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4537 } else if (k->has_modified_oops()) {
4538 // Don't clear anything, this info is needed by the next young collection.
4539 } else {
4540 // No modified oops in the Klass.
4541 return;
4542 }
4543
4544 // The klass has modified fields, need to scan the klass.
4545 _cm_klass_closure.do_klass(k);
4546 }
4547 };
4548
4549 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
4550 ParNewGeneration* young_gen = _collector->_young_gen;
4551 ContiguousSpace* eden_space = young_gen->eden();
4552 ContiguousSpace* from_space = young_gen->from();
4553 ContiguousSpace* to_space = young_gen->to();
4554
4555 HeapWord** eca = _collector->_eden_chunk_array;
4556 size_t ect = _collector->_eden_chunk_index;
4557 HeapWord** sca = _collector->_survivor_chunk_array;
4558 size_t sct = _collector->_survivor_chunk_index;
4559
4560 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4561 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4562
4563 do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
4564 do_young_space_rescan(worker_id, cl, from_space, sca, sct);
4565 do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
4566 }
4567
4568 // work_queue(i) is passed to the closure
4569 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
4570 // also is passed to do_dirty_card_rescan_tasks() and to
4571 // do_work_steal() to select the i-th task_queue.
4572
4573 void CMSParRemarkTask::work(uint worker_id) {
4574 elapsedTimer _timer;
4575 ResourceMark rm;
4576 HandleMark hm;
4577
4578 // ---------- rescan from roots --------------
4579 _timer.start();
4580 GenCollectedHeap* gch = GenCollectedHeap::heap();
4581 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4582 _collector->_span, _collector->ref_processor(),
4583 &(_collector->_markBitMap),
4584 work_queue(worker_id));
4585
4586 // Rescan young gen roots first since these are likely
4587 // coarsely partitioned and may, on that account, constitute
4588 // the critical path; thus, it's best to start off that
4589 // work first.
4590 // ---------- young gen roots --------------
4591 {
4592 work_on_young_gen_roots(worker_id, &par_mrias_cl);
4593 _timer.stop();
4594 if (PrintCMSStatistics != 0) {
4595 gclog_or_tty->print_cr(
4596 "Finished young gen rescan work in %dth thread: %3.3f sec",
4597 worker_id, _timer.seconds());
4598 }
4599 }
4600
4601 // ---------- remaining roots --------------
4602 _timer.reset();
4603 _timer.start();
4604 gch->gen_process_roots(_collector->_cmsGen->level(),
4605 false, // yg was scanned above
4606 false, // this is parallel code
4607 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4608 _collector->should_unload_classes(),
4609 &par_mrias_cl,
4610 NULL,
4611 NULL); // The dirty klasses will be handled below
4612
4613 assert(_collector->should_unload_classes()
4614 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4615 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4616 _timer.stop();
4617 if (PrintCMSStatistics != 0) {
4618 gclog_or_tty->print_cr(
4619 "Finished remaining root rescan work in %dth thread: %3.3f sec",
4620 worker_id, _timer.seconds());
4621 }
4622
4623 // ---------- unhandled CLD scanning ----------
4624 if (worker_id == 0) { // Single threaded at the moment.
4625 _timer.reset();
4626 _timer.start();
4627
4628 // Scan all new class loader data objects and new dependencies that were
4629 // introduced during concurrent marking.
4630 ResourceMark rm;
4631 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4632 for (int i = 0; i < array->length(); i++) {
4633 par_mrias_cl.do_class_loader_data(array->at(i));
4634 }
4635
4636 // We don't need to keep track of new CLDs anymore.
4637 ClassLoaderDataGraph::remember_new_clds(false);
4638
4639 _timer.stop();
4640 if (PrintCMSStatistics != 0) {
4641 gclog_or_tty->print_cr(
4642 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec",
4643 worker_id, _timer.seconds());
4644 }
4645 }
4646
4647 // ---------- dirty klass scanning ----------
4648 if (worker_id == 0) { // Single threaded at the moment.
4649 _timer.reset();
4650 _timer.start();
4651
4652 // Scan all classes that was dirtied during the concurrent marking phase.
4653 RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
4654 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
4655
4656 _timer.stop();
4657 if (PrintCMSStatistics != 0) {
4658 gclog_or_tty->print_cr(
4659 "Finished dirty klass scanning work in %dth thread: %3.3f sec",
4660 worker_id, _timer.seconds());
4661 }
4662 }
4663
4664 // We might have added oops to ClassLoaderData::_handles during the
4665 // concurrent marking phase. These oops point to newly allocated objects
4666 // that are guaranteed to be kept alive. Either by the direct allocation
4667 // code, or when the young collector processes the roots. Hence,
4668 // we don't have to revisit the _handles block during the remark phase.
4669
4670 // ---------- rescan dirty cards ------------
4671 _timer.reset();
4672 _timer.start();
4673
4674 // Do the rescan tasks for each of the two spaces
4675 // (cms_space) in turn.
4676 // "worker_id" is passed to select the task_queue for "worker_id"
4677 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4678 _timer.stop();
4679 if (PrintCMSStatistics != 0) {
4680 gclog_or_tty->print_cr(
4681 "Finished dirty card rescan work in %dth thread: %3.3f sec",
4682 worker_id, _timer.seconds());
4683 }
4684
4685 // ---------- steal work from other threads ...
4686 // ---------- ... and drain overflow list.
4687 _timer.reset();
4688 _timer.start();
4689 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4690 _timer.stop();
4691 if (PrintCMSStatistics != 0) {
4692 gclog_or_tty->print_cr(
4693 "Finished work stealing in %dth thread: %3.3f sec",
4694 worker_id, _timer.seconds());
4695 }
4696 }
4697
4698 // Note that parameter "i" is not used.
4699 void
4700 CMSParMarkTask::do_young_space_rescan(uint worker_id,
4701 OopsInGenClosure* cl, ContiguousSpace* space,
4702 HeapWord** chunk_array, size_t chunk_top) {
4703 // Until all tasks completed:
4704 // . claim an unclaimed task
4705 // . compute region boundaries corresponding to task claimed
4706 // using chunk_array
4707 // . par_oop_iterate(cl) over that region
4708
4709 ResourceMark rm;
4710 HandleMark hm;
4711
4712 SequentialSubTasksDone* pst = space->par_seq_tasks();
4713
4714 uint nth_task = 0;
4715 uint n_tasks = pst->n_tasks();
4716
4717 if (n_tasks > 0) {
4718 assert(pst->valid(), "Uninitialized use?");
4719 HeapWord *start, *end;
4720 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4721 // We claimed task # nth_task; compute its boundaries.
4722 if (chunk_top == 0) { // no samples were taken
4723 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4724 start = space->bottom();
4725 end = space->top();
4726 } else if (nth_task == 0) {
4727 start = space->bottom();
4728 end = chunk_array[nth_task];
4729 } else if (nth_task < (uint)chunk_top) {
4730 assert(nth_task >= 1, "Control point invariant");
4731 start = chunk_array[nth_task - 1];
4732 end = chunk_array[nth_task];
4733 } else {
4734 assert(nth_task == (uint)chunk_top, "Control point invariant");
4735 start = chunk_array[chunk_top - 1];
4736 end = space->top();
4737 }
4738 MemRegion mr(start, end);
4739 // Verify that mr is in space
4740 assert(mr.is_empty() || space->used_region().contains(mr),
4741 "Should be in space");
4742 // Verify that "start" is an object boundary
4743 assert(mr.is_empty() || oop(mr.start())->is_oop(),
4744 "Should be an oop");
4745 space->par_oop_iterate(mr, cl);
4746 }
4747 pst->all_tasks_completed();
4748 }
4749 }
4750
4751 void
4752 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4753 CompactibleFreeListSpace* sp, int i,
4754 Par_MarkRefsIntoAndScanClosure* cl) {
4755 // Until all tasks completed:
4756 // . claim an unclaimed task
4757 // . compute region boundaries corresponding to task claimed
4758 // . transfer dirty bits ct->mut for that region
4759 // . apply rescanclosure to dirty mut bits for that region
4760
4761 ResourceMark rm;
4762 HandleMark hm;
4763
4764 OopTaskQueue* work_q = work_queue(i);
4765 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4766 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4767 // CAUTION: This closure has state that persists across calls to
4768 // the work method dirty_range_iterate_clear() in that it has
4769 // embedded in it a (subtype of) UpwardsObjectClosure. The
4770 // use of that state in the embedded UpwardsObjectClosure instance
4771 // assumes that the cards are always iterated (even if in parallel
4772 // by several threads) in monotonically increasing order per each
4773 // thread. This is true of the implementation below which picks
4774 // card ranges (chunks) in monotonically increasing order globally
4775 // and, a-fortiori, in monotonically increasing order per thread
4776 // (the latter order being a subsequence of the former).
4777 // If the work code below is ever reorganized into a more chaotic
4778 // work-partitioning form than the current "sequential tasks"
4779 // paradigm, the use of that persistent state will have to be
4780 // revisited and modified appropriately. See also related
4781 // bug 4756801 work on which should examine this code to make
4782 // sure that the changes there do not run counter to the
4783 // assumptions made here and necessary for correctness and
4784 // efficiency. Note also that this code might yield inefficient
4785 // behavior in the case of very large objects that span one or
4786 // more work chunks. Such objects would potentially be scanned
4787 // several times redundantly. Work on 4756801 should try and
4788 // address that performance anomaly if at all possible. XXX
4789 MemRegion full_span = _collector->_span;
4790 CMSBitMap* bm = &(_collector->_markBitMap); // shared
4791 MarkFromDirtyCardsClosure
4792 greyRescanClosure(_collector, full_span, // entire span of interest
4793 sp, bm, work_q, cl);
4794
4795 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4796 assert(pst->valid(), "Uninitialized use?");
4797 uint nth_task = 0;
4798 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
4799 MemRegion span = sp->used_region();
4800 HeapWord* start_addr = span.start();
4801 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
4802 alignment);
4803 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4804 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
4805 start_addr, "Check alignment");
4806 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
4807 chunk_size, "Check alignment");
4808
4809 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4810 // Having claimed the nth_task, compute corresponding mem-region,
4811 // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4812 // The alignment restriction ensures that we do not need any
4813 // synchronization with other gang-workers while setting or
4814 // clearing bits in thus chunk of the MUT.
4815 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4816 start_addr + (nth_task+1)*chunk_size);
4817 // The last chunk's end might be way beyond end of the
4818 // used region. In that case pull back appropriately.
4819 if (this_span.end() > end_addr) {
4820 this_span.set_end(end_addr);
4821 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4822 }
4823 // Iterate over the dirty cards covering this chunk, marking them
4824 // precleaned, and setting the corresponding bits in the mod union
4825 // table. Since we have been careful to partition at Card and MUT-word
4826 // boundaries no synchronization is needed between parallel threads.
4827 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
4828 &modUnionClosure);
4829
4830 // Having transferred these marks into the modUnionTable,
4831 // rescan the marked objects on the dirty cards in the modUnionTable.
4832 // Even if this is at a synchronous collection, the initial marking
4833 // may have been done during an asynchronous collection so there
4834 // may be dirty bits in the mod-union table.
4835 _collector->_modUnionTable.dirty_range_iterate_clear(
4836 this_span, &greyRescanClosure);
4837 _collector->_modUnionTable.verifyNoOneBitsInRange(
4838 this_span.start(),
4839 this_span.end());
4840 }
4841 pst->all_tasks_completed(); // declare that i am done
4842 }
4843
4844 // . see if we can share work_queues with ParNew? XXX
4845 void
4846 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
4847 int* seed) {
4848 OopTaskQueue* work_q = work_queue(i);
4849 NOT_PRODUCT(int num_steals = 0;)
4850 oop obj_to_scan;
4851 CMSBitMap* bm = &(_collector->_markBitMap);
4852
4853 while (true) {
4854 // Completely finish any left over work from (an) earlier round(s)
4855 cl->trim_queue(0);
4856 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4857 (size_t)ParGCDesiredObjsFromOverflowList);
4858 // Now check if there's any work in the overflow list
4859 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4860 // only affects the number of attempts made to get work from the
4861 // overflow list and does not affect the number of workers. Just
4862 // pass ParallelGCThreads so this behavior is unchanged.
4863 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4864 work_q,
4865 ParallelGCThreads)) {
4866 // found something in global overflow list;
4867 // not yet ready to go stealing work from others.
4868 // We'd like to assert(work_q->size() != 0, ...)
4869 // because we just took work from the overflow list,
4870 // but of course we can't since all of that could have
4871 // been already stolen from us.
4872 // "He giveth and He taketh away."
4873 continue;
4874 }
4875 // Verify that we have no work before we resort to stealing
4876 assert(work_q->size() == 0, "Have work, shouldn't steal");
4877 // Try to steal from other queues that have work
4878 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4879 NOT_PRODUCT(num_steals++;)
4880 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
4881 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4882 // Do scanning work
4883 obj_to_scan->oop_iterate(cl);
4884 // Loop around, finish this work, and try to steal some more
4885 } else if (terminator()->offer_termination()) {
4886 break; // nirvana from the infinite cycle
4887 }
4888 }
4889 NOT_PRODUCT(
4890 if (PrintCMSStatistics != 0) {
4891 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
4892 }
4893 )
4894 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4895 "Else our work is not yet done");
4896 }
4897
4898 // Record object boundaries in _eden_chunk_array by sampling the eden
4899 // top in the slow-path eden object allocation code path and record
4900 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4901 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4902 // sampling in sample_eden() that activates during the part of the
4903 // preclean phase.
4904 void CMSCollector::sample_eden_chunk() {
4905 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4906 if (_eden_chunk_lock->try_lock()) {
4907 // Record a sample. This is the critical section. The contents
4908 // of the _eden_chunk_array have to be non-decreasing in the
4909 // address order.
4910 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4911 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4912 "Unexpected state of Eden");
4913 if (_eden_chunk_index == 0 ||
4914 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4915 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4916 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4917 _eden_chunk_index++; // commit sample
4918 }
4919 _eden_chunk_lock->unlock();
4920 }
4921 }
4922 }
4923
4924 // Return a thread-local PLAB recording array, as appropriate.
4925 void* CMSCollector::get_data_recorder(int thr_num) {
4926 if (_survivor_plab_array != NULL &&
4927 (CMSPLABRecordAlways ||
4928 (_collectorState > Marking && _collectorState < FinalMarking))) {
4929 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4930 ChunkArray* ca = &_survivor_plab_array[thr_num];
4931 ca->reset(); // clear it so that fresh data is recorded
4932 return (void*) ca;
4933 } else {
4934 return NULL;
4935 }
4936 }
4937
4938 // Reset all the thread-local PLAB recording arrays
4939 void CMSCollector::reset_survivor_plab_arrays() {
4940 for (uint i = 0; i < ParallelGCThreads; i++) {
4941 _survivor_plab_array[i].reset();
4942 }
4943 }
4944
4945 // Merge the per-thread plab arrays into the global survivor chunk
4946 // array which will provide the partitioning of the survivor space
4947 // for CMS initial scan and rescan.
4948 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4949 int no_of_gc_threads) {
4950 assert(_survivor_plab_array != NULL, "Error");
4951 assert(_survivor_chunk_array != NULL, "Error");
4952 assert(_collectorState == FinalMarking ||
4953 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4954 for (int j = 0; j < no_of_gc_threads; j++) {
4955 _cursor[j] = 0;
4956 }
4957 HeapWord* top = surv->top();
4958 size_t i;
4959 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
4960 HeapWord* min_val = top; // Higher than any PLAB address
4961 uint min_tid = 0; // position of min_val this round
4962 for (int j = 0; j < no_of_gc_threads; j++) {
4963 ChunkArray* cur_sca = &_survivor_plab_array[j];
4964 if (_cursor[j] == cur_sca->end()) {
4965 continue;
4966 }
4967 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4968 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4969 assert(surv->used_region().contains(cur_val), "Out of bounds value");
4970 if (cur_val < min_val) {
4971 min_tid = j;
4972 min_val = cur_val;
4973 } else {
4974 assert(cur_val < top, "All recorded addresses should be less");
4975 }
4976 }
4977 // At this point min_val and min_tid are respectively
4978 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4979 // and the thread (j) that witnesses that address.
4980 // We record this address in the _survivor_chunk_array[i]
4981 // and increment _cursor[min_tid] prior to the next round i.
4982 if (min_val == top) {
4983 break;
4984 }
4985 _survivor_chunk_array[i] = min_val;
4986 _cursor[min_tid]++;
4987 }
4988 // We are all done; record the size of the _survivor_chunk_array
4989 _survivor_chunk_index = i; // exclusive: [0, i)
4990 if (PrintCMSStatistics > 0) {
4991 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4992 }
4993 // Verify that we used up all the recorded entries
4994 #ifdef ASSERT
4995 size_t total = 0;
4996 for (int j = 0; j < no_of_gc_threads; j++) {
4997 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4998 total += _cursor[j];
4999 }
5000 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5001 // Check that the merged array is in sorted order
5002 if (total > 0) {
5003 for (size_t i = 0; i < total - 1; i++) {
5004 if (PrintCMSStatistics > 0) {
5005 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5006 i, p2i(_survivor_chunk_array[i]));
5007 }
5008 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5009 "Not sorted");
5010 }
5011 }
5012 #endif // ASSERT
5013 }
5014
5015 // Set up the space's par_seq_tasks structure for work claiming
5016 // for parallel initial scan and rescan of young gen.
5017 // See ParRescanTask where this is currently used.
5018 void
5019 CMSCollector::
5020 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5021 assert(n_threads > 0, "Unexpected n_threads argument");
5022
5023 // Eden space
5024 if (!_young_gen->eden()->is_empty()) {
5025 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
5026 assert(!pst->valid(), "Clobbering existing data?");
5027 // Each valid entry in [0, _eden_chunk_index) represents a task.
5028 size_t n_tasks = _eden_chunk_index + 1;
5029 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5030 // Sets the condition for completion of the subtask (how many threads
5031 // need to finish in order to be done).
5032 pst->set_n_threads(n_threads);
5033 pst->set_n_tasks((int)n_tasks);
5034 }
5035
5036 // Merge the survivor plab arrays into _survivor_chunk_array
5037 if (_survivor_plab_array != NULL) {
5038 merge_survivor_plab_arrays(_young_gen->from(), n_threads);
5039 } else {
5040 assert(_survivor_chunk_index == 0, "Error");
5041 }
5042
5043 // To space
5044 {
5045 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
5046 assert(!pst->valid(), "Clobbering existing data?");
5047 // Sets the condition for completion of the subtask (how many threads
5048 // need to finish in order to be done).
5049 pst->set_n_threads(n_threads);
5050 pst->set_n_tasks(1);
5051 assert(pst->valid(), "Error");
5052 }
5053
5054 // From space
5055 {
5056 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
5057 assert(!pst->valid(), "Clobbering existing data?");
5058 size_t n_tasks = _survivor_chunk_index + 1;
5059 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5060 // Sets the condition for completion of the subtask (how many threads
5061 // need to finish in order to be done).
5062 pst->set_n_threads(n_threads);
5063 pst->set_n_tasks((int)n_tasks);
5064 assert(pst->valid(), "Error");
5065 }
5066 }
5067
5068 // Parallel version of remark
5069 void CMSCollector::do_remark_parallel() {
5070 GenCollectedHeap* gch = GenCollectedHeap::heap();
5071 FlexibleWorkGang* workers = gch->workers();
5072 assert(workers != NULL, "Need parallel worker threads.");
5073 // Choose to use the number of GC workers most recently set
5074 // into "active_workers". If active_workers is not set, set it
5075 // to ParallelGCThreads.
5076 uint n_workers = workers->active_workers();
5077 if (n_workers == 0) {
5078 assert(n_workers > 0, "Should have been set during scavenge");
5079 n_workers = ParallelGCThreads;
5080 workers->set_active_workers(n_workers);
5081 }
5082 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5083
5084 CMSParRemarkTask tsk(this,
5085 cms_space,
5086 n_workers, workers, task_queues());
5087
5088 // Set up for parallel process_roots work.
5089 gch->set_par_threads(n_workers);
5090 // We won't be iterating over the cards in the card table updating
5091 // the younger_gen cards, so we shouldn't call the following else
5092 // the verification code as well as subsequent younger_refs_iterate
5093 // code would get confused. XXX
5094 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5095
5096 // The young gen rescan work will not be done as part of
5097 // process_roots (which currently doesn't know how to
5098 // parallelize such a scan), but rather will be broken up into
5099 // a set of parallel tasks (via the sampling that the [abortable]
5100 // preclean phase did of eden, plus the [two] tasks of
5101 // scanning the [two] survivor spaces. Further fine-grain
5102 // parallelization of the scanning of the survivor spaces
5103 // themselves, and of precleaning of the younger gen itself
5104 // is deferred to the future.
5105 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5106
5107 // The dirty card rescan work is broken up into a "sequence"
5108 // of parallel tasks (per constituent space) that are dynamically
5109 // claimed by the parallel threads.
5110 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5111
5112 // It turns out that even when we're using 1 thread, doing the work in a
5113 // separate thread causes wide variance in run times. We can't help this
5114 // in the multi-threaded case, but we special-case n=1 here to get
5115 // repeatable measurements of the 1-thread overhead of the parallel code.
5116 if (n_workers > 1) {
5117 // Make refs discovery MT-safe, if it isn't already: it may not
5118 // necessarily be so, since it's possible that we are doing
5119 // ST marking.
5120 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5121 StrongRootsScope srs;
5122 workers->run_task(&tsk);
5123 } else {
5124 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5125 StrongRootsScope srs;
5126 tsk.work(0);
5127 }
5128
5129 gch->set_par_threads(0); // 0 ==> non-parallel.
5130 // restore, single-threaded for now, any preserved marks
5131 // as a result of work_q overflow
5132 restore_preserved_marks_if_any();
5133 }
5134
5135 // Non-parallel version of remark
5136 void CMSCollector::do_remark_non_parallel() {
5137 ResourceMark rm;
5138 HandleMark hm;
5139 GenCollectedHeap* gch = GenCollectedHeap::heap();
5140 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5141
5142 MarkRefsIntoAndScanClosure
5143 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
5144 &_markStack, this,
5145 false /* should_yield */, false /* not precleaning */);
5146 MarkFromDirtyCardsClosure
5147 markFromDirtyCardsClosure(this, _span,
5148 NULL, // space is set further below
5149 &_markBitMap, &_markStack, &mrias_cl);
5150 {
5151 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5152 // Iterate over the dirty cards, setting the corresponding bits in the
5153 // mod union table.
5154 {
5155 ModUnionClosure modUnionClosure(&_modUnionTable);
5156 _ct->ct_bs()->dirty_card_iterate(
5157 _cmsGen->used_region(),
5158 &modUnionClosure);
5159 }
5160 // Having transferred these marks into the modUnionTable, we just need
5161 // to rescan the marked objects on the dirty cards in the modUnionTable.
5162 // The initial marking may have been done during an asynchronous
5163 // collection so there may be dirty bits in the mod-union table.
5164 const int alignment =
5165 CardTableModRefBS::card_size * BitsPerWord;
5166 {
5167 // ... First handle dirty cards in CMS gen
5168 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5169 MemRegion ur = _cmsGen->used_region();
5170 HeapWord* lb = ur.start();
5171 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5172 MemRegion cms_span(lb, ub);
5173 _modUnionTable.dirty_range_iterate_clear(cms_span,
5174 &markFromDirtyCardsClosure);
5175 verify_work_stacks_empty();
5176 if (PrintCMSStatistics != 0) {
5177 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5178 markFromDirtyCardsClosure.num_dirty_cards());
5179 }
5180 }
5181 }
5182 if (VerifyDuringGC &&
5183 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5184 HandleMark hm; // Discard invalid handles created during verification
5185 Universe::verify();
5186 }
5187 {
5188 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5189
5190 verify_work_stacks_empty();
5191
5192 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5193 StrongRootsScope srs;
5194
5195 gch->gen_process_roots(_cmsGen->level(),
5196 true, // younger gens as roots
5197 false, // use the local StrongRootsScope
5198 GenCollectedHeap::ScanningOption(roots_scanning_options()),
5199 should_unload_classes(),
5200 &mrias_cl,
5201 NULL,
5202 NULL); // The dirty klasses will be handled below
5203
5204 assert(should_unload_classes()
5205 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
5206 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5207 }
5208
5209 {
5210 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5211
5212 verify_work_stacks_empty();
5213
5214 // Scan all class loader data objects that might have been introduced
5215 // during concurrent marking.
5216 ResourceMark rm;
5217 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5218 for (int i = 0; i < array->length(); i++) {
5219 mrias_cl.do_class_loader_data(array->at(i));
5220 }
5221
5222 // We don't need to keep track of new CLDs anymore.
5223 ClassLoaderDataGraph::remember_new_clds(false);
5224
5225 verify_work_stacks_empty();
5226 }
5227
5228 {
5229 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5230
5231 verify_work_stacks_empty();
5232
5233 RemarkKlassClosure remark_klass_closure(&mrias_cl);
5234 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5235
5236 verify_work_stacks_empty();
5237 }
5238
5239 // We might have added oops to ClassLoaderData::_handles during the
5240 // concurrent marking phase. These oops point to newly allocated objects
5241 // that are guaranteed to be kept alive. Either by the direct allocation
5242 // code, or when the young collector processes the roots. Hence,
5243 // we don't have to revisit the _handles block during the remark phase.
5244
5245 verify_work_stacks_empty();
5246 // Restore evacuated mark words, if any, used for overflow list links
5247 if (!CMSOverflowEarlyRestoration) {
5248 restore_preserved_marks_if_any();
5249 }
5250 verify_overflow_empty();
5251 }
5252
5253 ////////////////////////////////////////////////////////
5254 // Parallel Reference Processing Task Proxy Class
5255 ////////////////////////////////////////////////////////
5256 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5257 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5258 CMSCollector* _collector;
5259 CMSBitMap* _mark_bit_map;
5260 const MemRegion _span;
5261 ProcessTask& _task;
5262
5263 public:
5264 CMSRefProcTaskProxy(ProcessTask& task,
5265 CMSCollector* collector,
5266 const MemRegion& span,
5267 CMSBitMap* mark_bit_map,
5268 AbstractWorkGang* workers,
5269 OopTaskQueueSet* task_queues):
5270 // XXX Should superclass AGTWOQ also know about AWG since it knows
5271 // about the task_queues used by the AWG? Then it could initialize
5272 // the terminator() object. See 6984287. The set_for_termination()
5273 // below is a temporary band-aid for the regression in 6984287.
5274 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5275 task_queues),
5276 _task(task),
5277 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5278 {
5279 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5280 "Inconsistency in _span");
5281 set_for_termination(workers->active_workers());
5282 }
5283
5284 OopTaskQueueSet* task_queues() { return queues(); }
5285
5286 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5287
5288 void do_work_steal(int i,
5289 CMSParDrainMarkingStackClosure* drain,
5290 CMSParKeepAliveClosure* keep_alive,
5291 int* seed);
5292
5293 virtual void work(uint worker_id);
5294 };
5295
5296 void CMSRefProcTaskProxy::work(uint worker_id) {
5297 ResourceMark rm;
5298 HandleMark hm;
5299 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5300 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5301 _mark_bit_map,
5302 work_queue(worker_id));
5303 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5304 _mark_bit_map,
5305 work_queue(worker_id));
5306 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5307 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5308 if (_task.marks_oops_alive()) {
5309 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5310 _collector->hash_seed(worker_id));
5311 }
5312 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5313 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5314 }
5315
5316 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5317 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5318 EnqueueTask& _task;
5319
5320 public:
5321 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5322 : AbstractGangTask("Enqueue reference objects in parallel"),
5323 _task(task)
5324 { }
5325
5326 virtual void work(uint worker_id)
5327 {
5328 _task.work(worker_id);
5329 }
5330 };
5331
5332 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5333 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5334 _span(span),
5335 _bit_map(bit_map),
5336 _work_queue(work_queue),
5337 _mark_and_push(collector, span, bit_map, work_queue),
5338 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5339 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5340 { }
5341
5342 // . see if we can share work_queues with ParNew? XXX
5343 void CMSRefProcTaskProxy::do_work_steal(int i,
5344 CMSParDrainMarkingStackClosure* drain,
5345 CMSParKeepAliveClosure* keep_alive,
5346 int* seed) {
5347 OopTaskQueue* work_q = work_queue(i);
5348 NOT_PRODUCT(int num_steals = 0;)
5349 oop obj_to_scan;
5350
5351 while (true) {
5352 // Completely finish any left over work from (an) earlier round(s)
5353 drain->trim_queue(0);
5354 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5355 (size_t)ParGCDesiredObjsFromOverflowList);
5356 // Now check if there's any work in the overflow list
5357 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5358 // only affects the number of attempts made to get work from the
5359 // overflow list and does not affect the number of workers. Just
5360 // pass ParallelGCThreads so this behavior is unchanged.
5361 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5362 work_q,
5363 ParallelGCThreads)) {
5364 // Found something in global overflow list;
5365 // not yet ready to go stealing work from others.
5366 // We'd like to assert(work_q->size() != 0, ...)
5367 // because we just took work from the overflow list,
5368 // but of course we can't, since all of that might have
5369 // been already stolen from us.
5370 continue;
5371 }
5372 // Verify that we have no work before we resort to stealing
5373 assert(work_q->size() == 0, "Have work, shouldn't steal");
5374 // Try to steal from other queues that have work
5375 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5376 NOT_PRODUCT(num_steals++;)
5377 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5378 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5379 // Do scanning work
5380 obj_to_scan->oop_iterate(keep_alive);
5381 // Loop around, finish this work, and try to steal some more
5382 } else if (terminator()->offer_termination()) {
5383 break; // nirvana from the infinite cycle
5384 }
5385 }
5386 NOT_PRODUCT(
5387 if (PrintCMSStatistics != 0) {
5388 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5389 }
5390 )
5391 }
5392
5393 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5394 {
5395 GenCollectedHeap* gch = GenCollectedHeap::heap();
5396 FlexibleWorkGang* workers = gch->workers();
5397 assert(workers != NULL, "Need parallel worker threads.");
5398 CMSRefProcTaskProxy rp_task(task, &_collector,
5399 _collector.ref_processor()->span(),
5400 _collector.markBitMap(),
5401 workers, _collector.task_queues());
5402 workers->run_task(&rp_task);
5403 }
5404
5405 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5406 {
5407
5408 GenCollectedHeap* gch = GenCollectedHeap::heap();
5409 FlexibleWorkGang* workers = gch->workers();
5410 assert(workers != NULL, "Need parallel worker threads.");
5411 CMSRefEnqueueTaskProxy enq_task(task);
5412 workers->run_task(&enq_task);
5413 }
5414
5415 void CMSCollector::refProcessingWork() {
5416 ResourceMark rm;
5417 HandleMark hm;
5418
5419 ReferenceProcessor* rp = ref_processor();
5420 assert(rp->span().equals(_span), "Spans should be equal");
5421 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5422 // Process weak references.
5423 rp->setup_policy(false);
5424 verify_work_stacks_empty();
5425
5426 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5427 &_markStack, false /* !preclean */);
5428 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5429 _span, &_markBitMap, &_markStack,
5430 &cmsKeepAliveClosure, false /* !preclean */);
5431 {
5432 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5433
5434 ReferenceProcessorStats stats;
5435 if (rp->processing_is_mt()) {
5436 // Set the degree of MT here. If the discovery is done MT, there
5437 // may have been a different number of threads doing the discovery
5438 // and a different number of discovered lists may have Ref objects.
5439 // That is OK as long as the Reference lists are balanced (see
5440 // balance_all_queues() and balance_queues()).
5441 GenCollectedHeap* gch = GenCollectedHeap::heap();
5442 uint active_workers = ParallelGCThreads;
5443 FlexibleWorkGang* workers = gch->workers();
5444 if (workers != NULL) {
5445 active_workers = workers->active_workers();
5446 // The expectation is that active_workers will have already
5447 // been set to a reasonable value. If it has not been set,
5448 // investigate.
5449 assert(active_workers > 0, "Should have been set during scavenge");
5450 }
5451 rp->set_active_mt_degree(active_workers);
5452 CMSRefProcTaskExecutor task_executor(*this);
5453 stats = rp->process_discovered_references(&_is_alive_closure,
5454 &cmsKeepAliveClosure,
5455 &cmsDrainMarkingStackClosure,
5456 &task_executor,
5457 _gc_timer_cm,
5458 _gc_tracer_cm->gc_id());
5459 } else {
5460 stats = rp->process_discovered_references(&_is_alive_closure,
5461 &cmsKeepAliveClosure,
5462 &cmsDrainMarkingStackClosure,
5463 NULL,
5464 _gc_timer_cm,
5465 _gc_tracer_cm->gc_id());
5466 }
5467 _gc_tracer_cm->report_gc_reference_stats(stats);
5468
5469 }
5470
5471 // This is the point where the entire marking should have completed.
5472 verify_work_stacks_empty();
5473
5474 if (should_unload_classes()) {
5475 {
5476 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5477
5478 // Unload classes and purge the SystemDictionary.
5479 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5480
5481 // Unload nmethods.
5482 CodeCache::do_unloading(&_is_alive_closure, purged_class);
5483
5484 // Prune dead klasses from subklass/sibling/implementor lists.
5485 Klass::clean_weak_klass_links(&_is_alive_closure);
5486 }
5487
5488 {
5489 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5490 // Clean up unreferenced symbols in symbol table.
5491 SymbolTable::unlink();
5492 }
5493
5494 {
5495 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5496 // Delete entries for dead interned strings.
5497 StringTable::unlink(&_is_alive_closure);
5498 }
5499 }
5500
5501
5502 // Restore any preserved marks as a result of mark stack or
5503 // work queue overflow
5504 restore_preserved_marks_if_any(); // done single-threaded for now
5505
5506 rp->set_enqueuing_is_done(true);
5507 if (rp->processing_is_mt()) {
5508 rp->balance_all_queues();
5509 CMSRefProcTaskExecutor task_executor(*this);
5510 rp->enqueue_discovered_references(&task_executor);
5511 } else {
5512 rp->enqueue_discovered_references(NULL);
5513 }
5514 rp->verify_no_references_recorded();
5515 assert(!rp->discovery_enabled(), "should have been disabled");
5516 }
5517
5518 #ifndef PRODUCT
5519 void CMSCollector::check_correct_thread_executing() {
5520 Thread* t = Thread::current();
5521 // Only the VM thread or the CMS thread should be here.
5522 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5523 "Unexpected thread type");
5524 // If this is the vm thread, the foreground process
5525 // should not be waiting. Note that _foregroundGCIsActive is
5526 // true while the foreground collector is waiting.
5527 if (_foregroundGCShouldWait) {
5528 // We cannot be the VM thread
5529 assert(t->is_ConcurrentGC_thread(),
5530 "Should be CMS thread");
5531 } else {
5532 // We can be the CMS thread only if we are in a stop-world
5533 // phase of CMS collection.
5534 if (t->is_ConcurrentGC_thread()) {
5535 assert(_collectorState == InitialMarking ||
5536 _collectorState == FinalMarking,
5537 "Should be a stop-world phase");
5538 // The CMS thread should be holding the CMS_token.
5539 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5540 "Potential interference with concurrently "
5541 "executing VM thread");
5542 }
5543 }
5544 }
5545 #endif
5546
5547 void CMSCollector::sweep() {
5548 assert(_collectorState == Sweeping, "just checking");
5549 check_correct_thread_executing();
5550 verify_work_stacks_empty();
5551 verify_overflow_empty();
5552 increment_sweep_count();
5553 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5554
5555 _inter_sweep_timer.stop();
5556 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5557
5558 assert(!_intra_sweep_timer.is_active(), "Should not be active");
5559 _intra_sweep_timer.reset();
5560 _intra_sweep_timer.start();
5561 {
5562 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5563 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails);
5564 // First sweep the old gen
5565 {
5566 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5567 bitMapLock());
5568 sweepWork(_cmsGen);
5569 }
5570
5571 // Update Universe::_heap_*_at_gc figures.
5572 // We need all the free list locks to make the abstract state
5573 // transition from Sweeping to Resetting. See detailed note
5574 // further below.
5575 {
5576 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5577 // Update heap occupancy information which is used as
5578 // input to soft ref clearing policy at the next gc.
5579 Universe::update_heap_info_at_gc();
5580 _collectorState = Resizing;
5581 }
5582 }
5583 verify_work_stacks_empty();
5584 verify_overflow_empty();
5585
5586 if (should_unload_classes()) {
5587 // Delay purge to the beginning of the next safepoint. Metaspace::contains
5588 // requires that the virtual spaces are stable and not deleted.
5589 ClassLoaderDataGraph::set_should_purge(true);
5590 }
5591
5592 _intra_sweep_timer.stop();
5593 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5594
5595 _inter_sweep_timer.reset();
5596 _inter_sweep_timer.start();
5597
5598 // We need to use a monotonically non-decreasing time in ms
5599 // or we will see time-warp warnings and os::javaTimeMillis()
5600 // does not guarantee monotonicity.
5601 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5602 update_time_of_last_gc(now);
5603
5604 // NOTE on abstract state transitions:
5605 // Mutators allocate-live and/or mark the mod-union table dirty
5606 // based on the state of the collection. The former is done in
5607 // the interval [Marking, Sweeping] and the latter in the interval
5608 // [Marking, Sweeping). Thus the transitions into the Marking state
5609 // and out of the Sweeping state must be synchronously visible
5610 // globally to the mutators.
5611 // The transition into the Marking state happens with the world
5612 // stopped so the mutators will globally see it. Sweeping is
5613 // done asynchronously by the background collector so the transition
5614 // from the Sweeping state to the Resizing state must be done
5615 // under the freelistLock (as is the check for whether to
5616 // allocate-live and whether to dirty the mod-union table).
5617 assert(_collectorState == Resizing, "Change of collector state to"
5618 " Resizing must be done under the freelistLocks (plural)");
5619
5620 // Now that sweeping has been completed, we clear
5621 // the incremental_collection_failed flag,
5622 // thus inviting a younger gen collection to promote into
5623 // this generation. If such a promotion may still fail,
5624 // the flag will be set again when a young collection is
5625 // attempted.
5626 GenCollectedHeap* gch = GenCollectedHeap::heap();
5627 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
5628 gch->update_full_collections_completed(_collection_count_start);
5629 }
5630
5631 // FIX ME!!! Looks like this belongs in CFLSpace, with
5632 // CMSGen merely delegating to it.
5633 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5634 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5635 HeapWord* minAddr = _cmsSpace->bottom();
5636 HeapWord* largestAddr =
5637 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5638 if (largestAddr == NULL) {
5639 // The dictionary appears to be empty. In this case
5640 // try to coalesce at the end of the heap.
5641 largestAddr = _cmsSpace->end();
5642 }
5643 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5644 size_t nearLargestOffset =
5645 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5646 if (PrintFLSStatistics != 0) {
5647 gclog_or_tty->print_cr(
5648 "CMS: Large Block: " PTR_FORMAT ";"
5649 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5650 p2i(largestAddr),
5651 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5652 }
5653 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5654 }
5655
5656 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5657 return addr >= _cmsSpace->nearLargestChunk();
5658 }
5659
5660 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5661 return _cmsSpace->find_chunk_at_end();
5662 }
5663
5664 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
5665 bool full) {
5666 // The next lower level has been collected. Gather any statistics
5667 // that are of interest at this point.
5668 if (!full && (current_level + 1) == level()) {
5669 // Gather statistics on the young generation collection.
5670 collector()->stats().record_gc0_end(used());
5671 }
5672 }
5673
5674 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen) {
5675 // We iterate over the space(s) underlying this generation,
5676 // checking the mark bit map to see if the bits corresponding
5677 // to specific blocks are marked or not. Blocks that are
5678 // marked are live and are not swept up. All remaining blocks
5679 // are swept up, with coalescing on-the-fly as we sweep up
5680 // contiguous free and/or garbage blocks:
5681 // We need to ensure that the sweeper synchronizes with allocators
5682 // and stop-the-world collectors. In particular, the following
5683 // locks are used:
5684 // . CMS token: if this is held, a stop the world collection cannot occur
5685 // . freelistLock: if this is held no allocation can occur from this
5686 // generation by another thread
5687 // . bitMapLock: if this is held, no other thread can access or update
5688 //
5689
5690 // Note that we need to hold the freelistLock if we use
5691 // block iterate below; else the iterator might go awry if
5692 // a mutator (or promotion) causes block contents to change
5693 // (for instance if the allocator divvies up a block).
5694 // If we hold the free list lock, for all practical purposes
5695 // young generation GC's can't occur (they'll usually need to
5696 // promote), so we might as well prevent all young generation
5697 // GC's while we do a sweeping step. For the same reason, we might
5698 // as well take the bit map lock for the entire duration
5699
5700 // check that we hold the requisite locks
5701 assert(have_cms_token(), "Should hold cms token");
5702 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5703 assert_lock_strong(gen->freelistLock());
5704 assert_lock_strong(bitMapLock());
5705
5706 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5707 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
5708 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5709 _inter_sweep_estimate.padded_average(),
5710 _intra_sweep_estimate.padded_average());
5711 gen->setNearLargestChunk();
5712
5713 {
5714 SweepClosure sweepClosure(this, gen, &_markBitMap, CMSYield);
5715 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5716 // We need to free-up/coalesce garbage/blocks from a
5717 // co-terminal free run. This is done in the SweepClosure
5718 // destructor; so, do not remove this scope, else the
5719 // end-of-sweep-census below will be off by a little bit.
5720 }
5721 gen->cmsSpace()->sweep_completed();
5722 gen->cmsSpace()->endSweepFLCensus(sweep_count());
5723 if (should_unload_classes()) { // unloaded classes this cycle,
5724 _concurrent_cycles_since_last_unload = 0; // ... reset count
5725 } else { // did not unload classes,
5726 _concurrent_cycles_since_last_unload++; // ... increment count
5727 }
5728 }
5729
5730 // Reset CMS data structures (for now just the marking bit map)
5731 // preparatory for the next cycle.
5732 void CMSCollector::reset(bool concurrent) {
5733 if (concurrent) {
5734 CMSTokenSyncWithLocks ts(true, bitMapLock());
5735
5736 // If the state is not "Resetting", the foreground thread
5737 // has done a collection and the resetting.
5738 if (_collectorState != Resetting) {
5739 assert(_collectorState == Idling, "The state should only change"
5740 " because the foreground collector has finished the collection");
5741 return;
5742 }
5743
5744 // Clear the mark bitmap (no grey objects to start with)
5745 // for the next cycle.
5746 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5747 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails);
5748
5749 HeapWord* curAddr = _markBitMap.startWord();
5750 while (curAddr < _markBitMap.endWord()) {
5751 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
5752 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5753 _markBitMap.clear_large_range(chunk);
5754 if (ConcurrentMarkSweepThread::should_yield() &&
5755 !foregroundGCIsActive() &&
5756 CMSYield) {
5757 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5758 "CMS thread should hold CMS token");
5759 assert_lock_strong(bitMapLock());
5760 bitMapLock()->unlock();
5761 ConcurrentMarkSweepThread::desynchronize(true);
5762 stopTimer();
5763 if (PrintCMSStatistics != 0) {
5764 incrementYields();
5765 }
5766
5767 // See the comment in coordinator_yield()
5768 for (unsigned i = 0; i < CMSYieldSleepCount &&
5769 ConcurrentMarkSweepThread::should_yield() &&
5770 !CMSCollector::foregroundGCIsActive(); ++i) {
5771 os::sleep(Thread::current(), 1, false);
5772 }
5773
5774 ConcurrentMarkSweepThread::synchronize(true);
5775 bitMapLock()->lock_without_safepoint_check();
5776 startTimer();
5777 }
5778 curAddr = chunk.end();
5779 }
5780 // A successful mostly concurrent collection has been done.
5781 // Because only the full (i.e., concurrent mode failure) collections
5782 // are being measured for gc overhead limits, clean the "near" flag
5783 // and count.
5784 size_policy()->reset_gc_overhead_limit_count();
5785 _collectorState = Idling;
5786 } else {
5787 // already have the lock
5788 assert(_collectorState == Resetting, "just checking");
5789 assert_lock_strong(bitMapLock());
5790 _markBitMap.clear_all();
5791 _collectorState = Idling;
5792 }
5793
5794 register_gc_end();
5795 }
5796
5797 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5798 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5799 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id());
5800 TraceCollectorStats tcs(counters());
5801
5802 switch (op) {
5803 case CMS_op_checkpointRootsInitial: {
5804 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5805 checkpointRootsInitial();
5806 if (PrintGC) {
5807 _cmsGen->printOccupancy("initial-mark");
5808 }
5809 break;
5810 }
5811 case CMS_op_checkpointRootsFinal: {
5812 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5813 checkpointRootsFinal();
5814 if (PrintGC) {
5815 _cmsGen->printOccupancy("remark");
5816 }
5817 break;
5818 }
5819 default:
5820 fatal("No such CMS_op");
5821 }
5822 }
5823
5824 #ifndef PRODUCT
5825 size_t const CMSCollector::skip_header_HeapWords() {
5826 return FreeChunk::header_size();
5827 }
5828
5829 // Try and collect here conditions that should hold when
5830 // CMS thread is exiting. The idea is that the foreground GC
5831 // thread should not be blocked if it wants to terminate
5832 // the CMS thread and yet continue to run the VM for a while
5833 // after that.
5834 void CMSCollector::verify_ok_to_terminate() const {
5835 assert(Thread::current()->is_ConcurrentGC_thread(),
5836 "should be called by CMS thread");
5837 assert(!_foregroundGCShouldWait, "should be false");
5838 // We could check here that all the various low-level locks
5839 // are not held by the CMS thread, but that is overkill; see
5840 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5841 // is checked.
5842 }
5843 #endif
5844
5845 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5846 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5847 "missing Printezis mark?");
5848 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5849 size_t size = pointer_delta(nextOneAddr + 1, addr);
5850 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5851 "alignment problem");
5852 assert(size >= 3, "Necessary for Printezis marks to work");
5853 return size;
5854 }
5855
5856 // A variant of the above (block_size_using_printezis_bits()) except
5857 // that we return 0 if the P-bits are not yet set.
5858 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5859 if (_markBitMap.isMarked(addr + 1)) {
5860 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5861 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5862 size_t size = pointer_delta(nextOneAddr + 1, addr);
5863 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5864 "alignment problem");
5865 assert(size >= 3, "Necessary for Printezis marks to work");
5866 return size;
5867 }
5868 return 0;
5869 }
5870
5871 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5872 size_t sz = 0;
5873 oop p = (oop)addr;
5874 if (p->klass_or_null() != NULL) {
5875 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5876 } else {
5877 sz = block_size_using_printezis_bits(addr);
5878 }
5879 assert(sz > 0, "size must be nonzero");
5880 HeapWord* next_block = addr + sz;
5881 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
5882 CardTableModRefBS::card_size);
5883 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
5884 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
5885 "must be different cards");
5886 return next_card;
5887 }
5888
5889
5890 // CMS Bit Map Wrapper /////////////////////////////////////////
5891
5892 // Construct a CMS bit map infrastructure, but don't create the
5893 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5894 // further below.
5895 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5896 _bm(),
5897 _shifter(shifter),
5898 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5899 Monitor::_safepoint_check_sometimes) : NULL)
5900 {
5901 _bmStartWord = 0;
5902 _bmWordSize = 0;
5903 }
5904
5905 bool CMSBitMap::allocate(MemRegion mr) {
5906 _bmStartWord = mr.start();
5907 _bmWordSize = mr.word_size();
5908 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5909 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5910 if (!brs.is_reserved()) {
5911 warning("CMS bit map allocation failure");
5912 return false;
5913 }
5914 // For now we'll just commit all of the bit map up front.
5915 // Later on we'll try to be more parsimonious with swap.
5916 if (!_virtual_space.initialize(brs, brs.size())) {
5917 warning("CMS bit map backing store failure");
5918 return false;
5919 }
5920 assert(_virtual_space.committed_size() == brs.size(),
5921 "didn't reserve backing store for all of CMS bit map?");
5922 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
5923 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5924 _bmWordSize, "inconsistency in bit map sizing");
5925 _bm.set_size(_bmWordSize >> _shifter);
5926
5927 // bm.clear(); // can we rely on getting zero'd memory? verify below
5928 assert(isAllClear(),
5929 "Expected zero'd memory from ReservedSpace constructor");
5930 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5931 "consistency check");
5932 return true;
5933 }
5934
5935 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5936 HeapWord *next_addr, *end_addr, *last_addr;
5937 assert_locked();
5938 assert(covers(mr), "out-of-range error");
5939 // XXX assert that start and end are appropriately aligned
5940 for (next_addr = mr.start(), end_addr = mr.end();
5941 next_addr < end_addr; next_addr = last_addr) {
5942 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5943 last_addr = dirty_region.end();
5944 if (!dirty_region.is_empty()) {
5945 cl->do_MemRegion(dirty_region);
5946 } else {
5947 assert(last_addr == end_addr, "program logic");
5948 return;
5949 }
5950 }
5951 }
5952
5953 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5954 _bm.print_on_error(st, prefix);
5955 }
5956
5957 #ifndef PRODUCT
5958 void CMSBitMap::assert_locked() const {
5959 CMSLockVerifier::assert_locked(lock());
5960 }
5961
5962 bool CMSBitMap::covers(MemRegion mr) const {
5963 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5964 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5965 "size inconsistency");
5966 return (mr.start() >= _bmStartWord) &&
5967 (mr.end() <= endWord());
5968 }
5969
5970 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5971 return (start >= _bmStartWord && (start + size) <= endWord());
5972 }
5973
5974 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5975 // verify that there are no 1 bits in the interval [left, right)
5976 FalseBitMapClosure falseBitMapClosure;
5977 iterate(&falseBitMapClosure, left, right);
5978 }
5979
5980 void CMSBitMap::region_invariant(MemRegion mr)
5981 {
5982 assert_locked();
5983 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5984 assert(!mr.is_empty(), "unexpected empty region");
5985 assert(covers(mr), "mr should be covered by bit map");
5986 // convert address range into offset range
5987 size_t start_ofs = heapWordToOffset(mr.start());
5988 // Make sure that end() is appropriately aligned
5989 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
5990 (1 << (_shifter+LogHeapWordSize))),
5991 "Misaligned mr.end()");
5992 size_t end_ofs = heapWordToOffset(mr.end());
5993 assert(end_ofs > start_ofs, "Should mark at least one bit");
5994 }
5995
5996 #endif
5997
5998 bool CMSMarkStack::allocate(size_t size) {
5999 // allocate a stack of the requisite depth
6000 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6001 size * sizeof(oop)));
6002 if (!rs.is_reserved()) {
6003 warning("CMSMarkStack allocation failure");
6004 return false;
6005 }
6006 if (!_virtual_space.initialize(rs, rs.size())) {
6007 warning("CMSMarkStack backing store failure");
6008 return false;
6009 }
6010 assert(_virtual_space.committed_size() == rs.size(),
6011 "didn't reserve backing store for all of CMS stack?");
6012 _base = (oop*)(_virtual_space.low());
6013 _index = 0;
6014 _capacity = size;
6015 NOT_PRODUCT(_max_depth = 0);
6016 return true;
6017 }
6018
6019 // XXX FIX ME !!! In the MT case we come in here holding a
6020 // leaf lock. For printing we need to take a further lock
6021 // which has lower rank. We need to recalibrate the two
6022 // lock-ranks involved in order to be able to print the
6023 // messages below. (Or defer the printing to the caller.
6024 // For now we take the expedient path of just disabling the
6025 // messages for the problematic case.)
6026 void CMSMarkStack::expand() {
6027 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6028 if (_capacity == MarkStackSizeMax) {
6029 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6030 // We print a warning message only once per CMS cycle.
6031 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6032 }
6033 return;
6034 }
6035 // Double capacity if possible
6036 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6037 // Do not give up existing stack until we have managed to
6038 // get the double capacity that we desired.
6039 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6040 new_capacity * sizeof(oop)));
6041 if (rs.is_reserved()) {
6042 // Release the backing store associated with old stack
6043 _virtual_space.release();
6044 // Reinitialize virtual space for new stack
6045 if (!_virtual_space.initialize(rs, rs.size())) {
6046 fatal("Not enough swap for expanded marking stack");
6047 }
6048 _base = (oop*)(_virtual_space.low());
6049 _index = 0;
6050 _capacity = new_capacity;
6051 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6052 // Failed to double capacity, continue;
6053 // we print a detail message only once per CMS cycle.
6054 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6055 SIZE_FORMAT"K",
6056 _capacity / K, new_capacity / K);
6057 }
6058 }
6059
6060
6061 // Closures
6062 // XXX: there seems to be a lot of code duplication here;
6063 // should refactor and consolidate common code.
6064
6065 // This closure is used to mark refs into the CMS generation in
6066 // the CMS bit map. Called at the first checkpoint. This closure
6067 // assumes that we do not need to re-mark dirty cards; if the CMS
6068 // generation on which this is used is not an oldest
6069 // generation then this will lose younger_gen cards!
6070
6071 MarkRefsIntoClosure::MarkRefsIntoClosure(
6072 MemRegion span, CMSBitMap* bitMap):
6073 _span(span),
6074 _bitMap(bitMap)
6075 {
6076 assert(_ref_processor == NULL, "deliberately left NULL");
6077 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6078 }
6079
6080 void MarkRefsIntoClosure::do_oop(oop obj) {
6081 // if p points into _span, then mark corresponding bit in _markBitMap
6082 assert(obj->is_oop(), "expected an oop");
6083 HeapWord* addr = (HeapWord*)obj;
6084 if (_span.contains(addr)) {
6085 // this should be made more efficient
6086 _bitMap->mark(addr);
6087 }
6088 }
6089
6090 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6091 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6092
6093 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
6094 MemRegion span, CMSBitMap* bitMap):
6095 _span(span),
6096 _bitMap(bitMap)
6097 {
6098 assert(_ref_processor == NULL, "deliberately left NULL");
6099 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6100 }
6101
6102 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
6103 // if p points into _span, then mark corresponding bit in _markBitMap
6104 assert(obj->is_oop(), "expected an oop");
6105 HeapWord* addr = (HeapWord*)obj;
6106 if (_span.contains(addr)) {
6107 // this should be made more efficient
6108 _bitMap->par_mark(addr);
6109 }
6110 }
6111
6112 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6113 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6114
6115 // A variant of the above, used for CMS marking verification.
6116 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6117 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6118 _span(span),
6119 _verification_bm(verification_bm),
6120 _cms_bm(cms_bm)
6121 {
6122 assert(_ref_processor == NULL, "deliberately left NULL");
6123 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6124 }
6125
6126 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6127 // if p points into _span, then mark corresponding bit in _markBitMap
6128 assert(obj->is_oop(), "expected an oop");
6129 HeapWord* addr = (HeapWord*)obj;
6130 if (_span.contains(addr)) {
6131 _verification_bm->mark(addr);
6132 if (!_cms_bm->isMarked(addr)) {
6133 oop(addr)->print();
6134 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6135 fatal("... aborting");
6136 }
6137 }
6138 }
6139
6140 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6141 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6142
6143 //////////////////////////////////////////////////
6144 // MarkRefsIntoAndScanClosure
6145 //////////////////////////////////////////////////
6146
6147 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6148 ReferenceProcessor* rp,
6149 CMSBitMap* bit_map,
6150 CMSBitMap* mod_union_table,
6151 CMSMarkStack* mark_stack,
6152 CMSCollector* collector,
6153 bool should_yield,
6154 bool concurrent_precleaning):
6155 _collector(collector),
6156 _span(span),
6157 _bit_map(bit_map),
6158 _mark_stack(mark_stack),
6159 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6160 mark_stack, concurrent_precleaning),
6161 _yield(should_yield),
6162 _concurrent_precleaning(concurrent_precleaning),
6163 _freelistLock(NULL)
6164 {
6165 _ref_processor = rp;
6166 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6167 }
6168
6169 // This closure is used to mark refs into the CMS generation at the
6170 // second (final) checkpoint, and to scan and transitively follow
6171 // the unmarked oops. It is also used during the concurrent precleaning
6172 // phase while scanning objects on dirty cards in the CMS generation.
6173 // The marks are made in the marking bit map and the marking stack is
6174 // used for keeping the (newly) grey objects during the scan.
6175 // The parallel version (Par_...) appears further below.
6176 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6177 if (obj != NULL) {
6178 assert(obj->is_oop(), "expected an oop");
6179 HeapWord* addr = (HeapWord*)obj;
6180 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6181 assert(_collector->overflow_list_is_empty(),
6182 "overflow list should be empty");
6183 if (_span.contains(addr) &&
6184 !_bit_map->isMarked(addr)) {
6185 // mark bit map (object is now grey)
6186 _bit_map->mark(addr);
6187 // push on marking stack (stack should be empty), and drain the
6188 // stack by applying this closure to the oops in the oops popped
6189 // from the stack (i.e. blacken the grey objects)
6190 bool res = _mark_stack->push(obj);
6191 assert(res, "Should have space to push on empty stack");
6192 do {
6193 oop new_oop = _mark_stack->pop();
6194 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6195 assert(_bit_map->isMarked((HeapWord*)new_oop),
6196 "only grey objects on this stack");
6197 // iterate over the oops in this oop, marking and pushing
6198 // the ones in CMS heap (i.e. in _span).
6199 new_oop->oop_iterate(&_pushAndMarkClosure);
6200 // check if it's time to yield
6201 do_yield_check();
6202 } while (!_mark_stack->isEmpty() ||
6203 (!_concurrent_precleaning && take_from_overflow_list()));
6204 // if marking stack is empty, and we are not doing this
6205 // during precleaning, then check the overflow list
6206 }
6207 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6208 assert(_collector->overflow_list_is_empty(),
6209 "overflow list was drained above");
6210 // We could restore evacuated mark words, if any, used for
6211 // overflow list links here because the overflow list is
6212 // provably empty here. That would reduce the maximum
6213 // size requirements for preserved_{oop,mark}_stack.
6214 // But we'll just postpone it until we are all done
6215 // so we can just stream through.
6216 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6217 _collector->restore_preserved_marks_if_any();
6218 assert(_collector->no_preserved_marks(), "No preserved marks");
6219 }
6220 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6221 "All preserved marks should have been restored above");
6222 }
6223 }
6224
6225 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6226 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6227
6228 void MarkRefsIntoAndScanClosure::do_yield_work() {
6229 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6230 "CMS thread should hold CMS token");
6231 assert_lock_strong(_freelistLock);
6232 assert_lock_strong(_bit_map->lock());
6233 // relinquish the free_list_lock and bitMaplock()
6234 _bit_map->lock()->unlock();
6235 _freelistLock->unlock();
6236 ConcurrentMarkSweepThread::desynchronize(true);
6237 _collector->stopTimer();
6238 if (PrintCMSStatistics != 0) {
6239 _collector->incrementYields();
6240 }
6241
6242 // See the comment in coordinator_yield()
6243 for (unsigned i = 0;
6244 i < CMSYieldSleepCount &&
6245 ConcurrentMarkSweepThread::should_yield() &&
6246 !CMSCollector::foregroundGCIsActive();
6247 ++i) {
6248 os::sleep(Thread::current(), 1, false);
6249 }
6250
6251 ConcurrentMarkSweepThread::synchronize(true);
6252 _freelistLock->lock_without_safepoint_check();
6253 _bit_map->lock()->lock_without_safepoint_check();
6254 _collector->startTimer();
6255 }
6256
6257 ///////////////////////////////////////////////////////////
6258 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6259 // MarkRefsIntoAndScanClosure
6260 ///////////////////////////////////////////////////////////
6261 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6262 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6263 CMSBitMap* bit_map, OopTaskQueue* work_queue):
6264 _span(span),
6265 _bit_map(bit_map),
6266 _work_queue(work_queue),
6267 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6268 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6269 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
6270 {
6271 _ref_processor = rp;
6272 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6273 }
6274
6275 // This closure is used to mark refs into the CMS generation at the
6276 // second (final) checkpoint, and to scan and transitively follow
6277 // the unmarked oops. The marks are made in the marking bit map and
6278 // the work_queue is used for keeping the (newly) grey objects during
6279 // the scan phase whence they are also available for stealing by parallel
6280 // threads. Since the marking bit map is shared, updates are
6281 // synchronized (via CAS).
6282 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6283 if (obj != NULL) {
6284 // Ignore mark word because this could be an already marked oop
6285 // that may be chained at the end of the overflow list.
6286 assert(obj->is_oop(true), "expected an oop");
6287 HeapWord* addr = (HeapWord*)obj;
6288 if (_span.contains(addr) &&
6289 !_bit_map->isMarked(addr)) {
6290 // mark bit map (object will become grey):
6291 // It is possible for several threads to be
6292 // trying to "claim" this object concurrently;
6293 // the unique thread that succeeds in marking the
6294 // object first will do the subsequent push on
6295 // to the work queue (or overflow list).
6296 if (_bit_map->par_mark(addr)) {
6297 // push on work_queue (which may not be empty), and trim the
6298 // queue to an appropriate length by applying this closure to
6299 // the oops in the oops popped from the stack (i.e. blacken the
6300 // grey objects)
6301 bool res = _work_queue->push(obj);
6302 assert(res, "Low water mark should be less than capacity?");
6303 trim_queue(_low_water_mark);
6304 } // Else, another thread claimed the object
6305 }
6306 }
6307 }
6308
6309 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6310 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6311
6312 // This closure is used to rescan the marked objects on the dirty cards
6313 // in the mod union table and the card table proper.
6314 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6315 oop p, MemRegion mr) {
6316
6317 size_t size = 0;
6318 HeapWord* addr = (HeapWord*)p;
6319 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6320 assert(_span.contains(addr), "we are scanning the CMS generation");
6321 // check if it's time to yield
6322 if (do_yield_check()) {
6323 // We yielded for some foreground stop-world work,
6324 // and we have been asked to abort this ongoing preclean cycle.
6325 return 0;
6326 }
6327 if (_bitMap->isMarked(addr)) {
6328 // it's marked; is it potentially uninitialized?
6329 if (p->klass_or_null() != NULL) {
6330 // an initialized object; ignore mark word in verification below
6331 // since we are running concurrent with mutators
6332 assert(p->is_oop(true), "should be an oop");
6333 if (p->is_objArray()) {
6334 // objArrays are precisely marked; restrict scanning
6335 // to dirty cards only.
6336 size = CompactibleFreeListSpace::adjustObjectSize(
6337 p->oop_iterate(_scanningClosure, mr));
6338 } else {
6339 // A non-array may have been imprecisely marked; we need
6340 // to scan object in its entirety.
6341 size = CompactibleFreeListSpace::adjustObjectSize(
6342 p->oop_iterate(_scanningClosure));
6343 }
6344 #ifdef ASSERT
6345 size_t direct_size =
6346 CompactibleFreeListSpace::adjustObjectSize(p->size());
6347 assert(size == direct_size, "Inconsistency in size");
6348 assert(size >= 3, "Necessary for Printezis marks to work");
6349 if (!_bitMap->isMarked(addr+1)) {
6350 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6351 } else {
6352 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6353 assert(_bitMap->isMarked(addr+size-1),
6354 "inconsistent Printezis mark");
6355 }
6356 #endif // ASSERT
6357 } else {
6358 // An uninitialized object.
6359 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6360 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6361 size = pointer_delta(nextOneAddr + 1, addr);
6362 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6363 "alignment problem");
6364 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6365 // will dirty the card when the klass pointer is installed in the
6366 // object (signaling the completion of initialization).
6367 }
6368 } else {
6369 // Either a not yet marked object or an uninitialized object
6370 if (p->klass_or_null() == NULL) {
6371 // An uninitialized object, skip to the next card, since
6372 // we may not be able to read its P-bits yet.
6373 assert(size == 0, "Initial value");
6374 } else {
6375 // An object not (yet) reached by marking: we merely need to
6376 // compute its size so as to go look at the next block.
6377 assert(p->is_oop(true), "should be an oop");
6378 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6379 }
6380 }
6381 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6382 return size;
6383 }
6384
6385 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6386 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6387 "CMS thread should hold CMS token");
6388 assert_lock_strong(_freelistLock);
6389 assert_lock_strong(_bitMap->lock());
6390 // relinquish the free_list_lock and bitMaplock()
6391 _bitMap->lock()->unlock();
6392 _freelistLock->unlock();
6393 ConcurrentMarkSweepThread::desynchronize(true);
6394 _collector->stopTimer();
6395 if (PrintCMSStatistics != 0) {
6396 _collector->incrementYields();
6397 }
6398
6399 // See the comment in coordinator_yield()
6400 for (unsigned i = 0; i < CMSYieldSleepCount &&
6401 ConcurrentMarkSweepThread::should_yield() &&
6402 !CMSCollector::foregroundGCIsActive(); ++i) {
6403 os::sleep(Thread::current(), 1, false);
6404 }
6405
6406 ConcurrentMarkSweepThread::synchronize(true);
6407 _freelistLock->lock_without_safepoint_check();
6408 _bitMap->lock()->lock_without_safepoint_check();
6409 _collector->startTimer();
6410 }
6411
6412
6413 //////////////////////////////////////////////////////////////////
6414 // SurvivorSpacePrecleanClosure
6415 //////////////////////////////////////////////////////////////////
6416 // This (single-threaded) closure is used to preclean the oops in
6417 // the survivor spaces.
6418 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6419
6420 HeapWord* addr = (HeapWord*)p;
6421 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6422 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6423 assert(p->klass_or_null() != NULL, "object should be initialized");
6424 // an initialized object; ignore mark word in verification below
6425 // since we are running concurrent with mutators
6426 assert(p->is_oop(true), "should be an oop");
6427 // Note that we do not yield while we iterate over
6428 // the interior oops of p, pushing the relevant ones
6429 // on our marking stack.
6430 size_t size = p->oop_iterate(_scanning_closure);
6431 do_yield_check();
6432 // Observe that below, we do not abandon the preclean
6433 // phase as soon as we should; rather we empty the
6434 // marking stack before returning. This is to satisfy
6435 // some existing assertions. In general, it may be a
6436 // good idea to abort immediately and complete the marking
6437 // from the grey objects at a later time.
6438 while (!_mark_stack->isEmpty()) {
6439 oop new_oop = _mark_stack->pop();
6440 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6441 assert(_bit_map->isMarked((HeapWord*)new_oop),
6442 "only grey objects on this stack");
6443 // iterate over the oops in this oop, marking and pushing
6444 // the ones in CMS heap (i.e. in _span).
6445 new_oop->oop_iterate(_scanning_closure);
6446 // check if it's time to yield
6447 do_yield_check();
6448 }
6449 unsigned int after_count =
6450 GenCollectedHeap::heap()->total_collections();
6451 bool abort = (_before_count != after_count) ||
6452 _collector->should_abort_preclean();
6453 return abort ? 0 : size;
6454 }
6455
6456 void SurvivorSpacePrecleanClosure::do_yield_work() {
6457 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6458 "CMS thread should hold CMS token");
6459 assert_lock_strong(_bit_map->lock());
6460 // Relinquish the bit map lock
6461 _bit_map->lock()->unlock();
6462 ConcurrentMarkSweepThread::desynchronize(true);
6463 _collector->stopTimer();
6464 if (PrintCMSStatistics != 0) {
6465 _collector->incrementYields();
6466 }
6467
6468 // See the comment in coordinator_yield()
6469 for (unsigned i = 0; i < CMSYieldSleepCount &&
6470 ConcurrentMarkSweepThread::should_yield() &&
6471 !CMSCollector::foregroundGCIsActive(); ++i) {
6472 os::sleep(Thread::current(), 1, false);
6473 }
6474
6475 ConcurrentMarkSweepThread::synchronize(true);
6476 _bit_map->lock()->lock_without_safepoint_check();
6477 _collector->startTimer();
6478 }
6479
6480 // This closure is used to rescan the marked objects on the dirty cards
6481 // in the mod union table and the card table proper. In the parallel
6482 // case, although the bitMap is shared, we do a single read so the
6483 // isMarked() query is "safe".
6484 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6485 // Ignore mark word because we are running concurrent with mutators
6486 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p)));
6487 HeapWord* addr = (HeapWord*)p;
6488 assert(_span.contains(addr), "we are scanning the CMS generation");
6489 bool is_obj_array = false;
6490 #ifdef ASSERT
6491 if (!_parallel) {
6492 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6493 assert(_collector->overflow_list_is_empty(),
6494 "overflow list should be empty");
6495
6496 }
6497 #endif // ASSERT
6498 if (_bit_map->isMarked(addr)) {
6499 // Obj arrays are precisely marked, non-arrays are not;
6500 // so we scan objArrays precisely and non-arrays in their
6501 // entirety.
6502 if (p->is_objArray()) {
6503 is_obj_array = true;
6504 if (_parallel) {
6505 p->oop_iterate(_par_scan_closure, mr);
6506 } else {
6507 p->oop_iterate(_scan_closure, mr);
6508 }
6509 } else {
6510 if (_parallel) {
6511 p->oop_iterate(_par_scan_closure);
6512 } else {
6513 p->oop_iterate(_scan_closure);
6514 }
6515 }
6516 }
6517 #ifdef ASSERT
6518 if (!_parallel) {
6519 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6520 assert(_collector->overflow_list_is_empty(),
6521 "overflow list should be empty");
6522
6523 }
6524 #endif // ASSERT
6525 return is_obj_array;
6526 }
6527
6528 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6529 MemRegion span,
6530 CMSBitMap* bitMap, CMSMarkStack* markStack,
6531 bool should_yield, bool verifying):
6532 _collector(collector),
6533 _span(span),
6534 _bitMap(bitMap),
6535 _mut(&collector->_modUnionTable),
6536 _markStack(markStack),
6537 _yield(should_yield),
6538 _skipBits(0)
6539 {
6540 assert(_markStack->isEmpty(), "stack should be empty");
6541 _finger = _bitMap->startWord();
6542 _threshold = _finger;
6543 assert(_collector->_restart_addr == NULL, "Sanity check");
6544 assert(_span.contains(_finger), "Out of bounds _finger?");
6545 DEBUG_ONLY(_verifying = verifying;)
6546 }
6547
6548 void MarkFromRootsClosure::reset(HeapWord* addr) {
6549 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6550 assert(_span.contains(addr), "Out of bounds _finger?");
6551 _finger = addr;
6552 _threshold = (HeapWord*)round_to(
6553 (intptr_t)_finger, CardTableModRefBS::card_size);
6554 }
6555
6556 // Should revisit to see if this should be restructured for
6557 // greater efficiency.
6558 bool MarkFromRootsClosure::do_bit(size_t offset) {
6559 if (_skipBits > 0) {
6560 _skipBits--;
6561 return true;
6562 }
6563 // convert offset into a HeapWord*
6564 HeapWord* addr = _bitMap->startWord() + offset;
6565 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6566 "address out of range");
6567 assert(_bitMap->isMarked(addr), "tautology");
6568 if (_bitMap->isMarked(addr+1)) {
6569 // this is an allocated but not yet initialized object
6570 assert(_skipBits == 0, "tautology");
6571 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6572 oop p = oop(addr);
6573 if (p->klass_or_null() == NULL) {
6574 DEBUG_ONLY(if (!_verifying) {)
6575 // We re-dirty the cards on which this object lies and increase
6576 // the _threshold so that we'll come back to scan this object
6577 // during the preclean or remark phase. (CMSCleanOnEnter)
6578 if (CMSCleanOnEnter) {
6579 size_t sz = _collector->block_size_using_printezis_bits(addr);
6580 HeapWord* end_card_addr = (HeapWord*)round_to(
6581 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6582 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6583 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6584 // Bump _threshold to end_card_addr; note that
6585 // _threshold cannot possibly exceed end_card_addr, anyhow.
6586 // This prevents future clearing of the card as the scan proceeds
6587 // to the right.
6588 assert(_threshold <= end_card_addr,
6589 "Because we are just scanning into this object");
6590 if (_threshold < end_card_addr) {
6591 _threshold = end_card_addr;
6592 }
6593 if (p->klass_or_null() != NULL) {
6594 // Redirty the range of cards...
6595 _mut->mark_range(redirty_range);
6596 } // ...else the setting of klass will dirty the card anyway.
6597 }
6598 DEBUG_ONLY(})
6599 return true;
6600 }
6601 }
6602 scanOopsInOop(addr);
6603 return true;
6604 }
6605
6606 // We take a break if we've been at this for a while,
6607 // so as to avoid monopolizing the locks involved.
6608 void MarkFromRootsClosure::do_yield_work() {
6609 // First give up the locks, then yield, then re-lock
6610 // We should probably use a constructor/destructor idiom to
6611 // do this unlock/lock or modify the MutexUnlocker class to
6612 // serve our purpose. XXX
6613 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6614 "CMS thread should hold CMS token");
6615 assert_lock_strong(_bitMap->lock());
6616 _bitMap->lock()->unlock();
6617 ConcurrentMarkSweepThread::desynchronize(true);
6618 _collector->stopTimer();
6619 if (PrintCMSStatistics != 0) {
6620 _collector->incrementYields();
6621 }
6622
6623 // See the comment in coordinator_yield()
6624 for (unsigned i = 0; i < CMSYieldSleepCount &&
6625 ConcurrentMarkSweepThread::should_yield() &&
6626 !CMSCollector::foregroundGCIsActive(); ++i) {
6627 os::sleep(Thread::current(), 1, false);
6628 }
6629
6630 ConcurrentMarkSweepThread::synchronize(true);
6631 _bitMap->lock()->lock_without_safepoint_check();
6632 _collector->startTimer();
6633 }
6634
6635 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6636 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6637 assert(_markStack->isEmpty(),
6638 "should drain stack to limit stack usage");
6639 // convert ptr to an oop preparatory to scanning
6640 oop obj = oop(ptr);
6641 // Ignore mark word in verification below, since we
6642 // may be running concurrent with mutators.
6643 assert(obj->is_oop(true), "should be an oop");
6644 assert(_finger <= ptr, "_finger runneth ahead");
6645 // advance the finger to right end of this object
6646 _finger = ptr + obj->size();
6647 assert(_finger > ptr, "we just incremented it above");
6648 // On large heaps, it may take us some time to get through
6649 // the marking phase. During
6650 // this time it's possible that a lot of mutations have
6651 // accumulated in the card table and the mod union table --
6652 // these mutation records are redundant until we have
6653 // actually traced into the corresponding card.
6654 // Here, we check whether advancing the finger would make
6655 // us cross into a new card, and if so clear corresponding
6656 // cards in the MUT (preclean them in the card-table in the
6657 // future).
6658
6659 DEBUG_ONLY(if (!_verifying) {)
6660 // The clean-on-enter optimization is disabled by default,
6661 // until we fix 6178663.
6662 if (CMSCleanOnEnter && (_finger > _threshold)) {
6663 // [_threshold, _finger) represents the interval
6664 // of cards to be cleared in MUT (or precleaned in card table).
6665 // The set of cards to be cleared is all those that overlap
6666 // with the interval [_threshold, _finger); note that
6667 // _threshold is always kept card-aligned but _finger isn't
6668 // always card-aligned.
6669 HeapWord* old_threshold = _threshold;
6670 assert(old_threshold == (HeapWord*)round_to(
6671 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6672 "_threshold should always be card-aligned");
6673 _threshold = (HeapWord*)round_to(
6674 (intptr_t)_finger, CardTableModRefBS::card_size);
6675 MemRegion mr(old_threshold, _threshold);
6676 assert(!mr.is_empty(), "Control point invariant");
6677 assert(_span.contains(mr), "Should clear within span");
6678 _mut->clear_range(mr);
6679 }
6680 DEBUG_ONLY(})
6681 // Note: the finger doesn't advance while we drain
6682 // the stack below.
6683 PushOrMarkClosure pushOrMarkClosure(_collector,
6684 _span, _bitMap, _markStack,
6685 _finger, this);
6686 bool res = _markStack->push(obj);
6687 assert(res, "Empty non-zero size stack should have space for single push");
6688 while (!_markStack->isEmpty()) {
6689 oop new_oop = _markStack->pop();
6690 // Skip verifying header mark word below because we are
6691 // running concurrent with mutators.
6692 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6693 // now scan this oop's oops
6694 new_oop->oop_iterate(&pushOrMarkClosure);
6695 do_yield_check();
6696 }
6697 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6698 }
6699
6700 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
6701 CMSCollector* collector, MemRegion span,
6702 CMSBitMap* bit_map,
6703 OopTaskQueue* work_queue,
6704 CMSMarkStack* overflow_stack):
6705 _collector(collector),
6706 _whole_span(collector->_span),
6707 _span(span),
6708 _bit_map(bit_map),
6709 _mut(&collector->_modUnionTable),
6710 _work_queue(work_queue),
6711 _overflow_stack(overflow_stack),
6712 _skip_bits(0),
6713 _task(task)
6714 {
6715 assert(_work_queue->size() == 0, "work_queue should be empty");
6716 _finger = span.start();
6717 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
6718 assert(_span.contains(_finger), "Out of bounds _finger?");
6719 }
6720
6721 // Should revisit to see if this should be restructured for
6722 // greater efficiency.
6723 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
6724 if (_skip_bits > 0) {
6725 _skip_bits--;
6726 return true;
6727 }
6728 // convert offset into a HeapWord*
6729 HeapWord* addr = _bit_map->startWord() + offset;
6730 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6731 "address out of range");
6732 assert(_bit_map->isMarked(addr), "tautology");
6733 if (_bit_map->isMarked(addr+1)) {
6734 // this is an allocated object that might not yet be initialized
6735 assert(_skip_bits == 0, "tautology");
6736 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
6737 oop p = oop(addr);
6738 if (p->klass_or_null() == NULL) {
6739 // in the case of Clean-on-Enter optimization, redirty card
6740 // and avoid clearing card by increasing the threshold.
6741 return true;
6742 }
6743 }
6744 scan_oops_in_oop(addr);
6745 return true;
6746 }
6747
6748 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6749 assert(_bit_map->isMarked(ptr), "expected bit to be set");
6750 // Should we assert that our work queue is empty or
6751 // below some drain limit?
6752 assert(_work_queue->size() == 0,
6753 "should drain stack to limit stack usage");
6754 // convert ptr to an oop preparatory to scanning
6755 oop obj = oop(ptr);
6756 // Ignore mark word in verification below, since we
6757 // may be running concurrent with mutators.
6758 assert(obj->is_oop(true), "should be an oop");
6759 assert(_finger <= ptr, "_finger runneth ahead");
6760 // advance the finger to right end of this object
6761 _finger = ptr + obj->size();
6762 assert(_finger > ptr, "we just incremented it above");
6763 // On large heaps, it may take us some time to get through
6764 // the marking phase. During
6765 // this time it's possible that a lot of mutations have
6766 // accumulated in the card table and the mod union table --
6767 // these mutation records are redundant until we have
6768 // actually traced into the corresponding card.
6769 // Here, we check whether advancing the finger would make
6770 // us cross into a new card, and if so clear corresponding
6771 // cards in the MUT (preclean them in the card-table in the
6772 // future).
6773
6774 // The clean-on-enter optimization is disabled by default,
6775 // until we fix 6178663.
6776 if (CMSCleanOnEnter && (_finger > _threshold)) {
6777 // [_threshold, _finger) represents the interval
6778 // of cards to be cleared in MUT (or precleaned in card table).
6779 // The set of cards to be cleared is all those that overlap
6780 // with the interval [_threshold, _finger); note that
6781 // _threshold is always kept card-aligned but _finger isn't
6782 // always card-aligned.
6783 HeapWord* old_threshold = _threshold;
6784 assert(old_threshold == (HeapWord*)round_to(
6785 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6786 "_threshold should always be card-aligned");
6787 _threshold = (HeapWord*)round_to(
6788 (intptr_t)_finger, CardTableModRefBS::card_size);
6789 MemRegion mr(old_threshold, _threshold);
6790 assert(!mr.is_empty(), "Control point invariant");
6791 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6792 _mut->clear_range(mr);
6793 }
6794
6795 // Note: the local finger doesn't advance while we drain
6796 // the stack below, but the global finger sure can and will.
6797 HeapWord** gfa = _task->global_finger_addr();
6798 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
6799 _span, _bit_map,
6800 _work_queue,
6801 _overflow_stack,
6802 _finger,
6803 gfa, this);
6804 bool res = _work_queue->push(obj); // overflow could occur here
6805 assert(res, "Will hold once we use workqueues");
6806 while (true) {
6807 oop new_oop;
6808 if (!_work_queue->pop_local(new_oop)) {
6809 // We emptied our work_queue; check if there's stuff that can
6810 // be gotten from the overflow stack.
6811 if (CMSConcMarkingTask::get_work_from_overflow_stack(
6812 _overflow_stack, _work_queue)) {
6813 do_yield_check();
6814 continue;
6815 } else { // done
6816 break;
6817 }
6818 }
6819 // Skip verifying header mark word below because we are
6820 // running concurrent with mutators.
6821 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6822 // now scan this oop's oops
6823 new_oop->oop_iterate(&pushOrMarkClosure);
6824 do_yield_check();
6825 }
6826 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6827 }
6828
6829 // Yield in response to a request from VM Thread or
6830 // from mutators.
6831 void Par_MarkFromRootsClosure::do_yield_work() {
6832 assert(_task != NULL, "sanity");
6833 _task->yield();
6834 }
6835
6836 // A variant of the above used for verifying CMS marking work.
6837 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6838 MemRegion span,
6839 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6840 CMSMarkStack* mark_stack):
6841 _collector(collector),
6842 _span(span),
6843 _verification_bm(verification_bm),
6844 _cms_bm(cms_bm),
6845 _mark_stack(mark_stack),
6846 _pam_verify_closure(collector, span, verification_bm, cms_bm,
6847 mark_stack)
6848 {
6849 assert(_mark_stack->isEmpty(), "stack should be empty");
6850 _finger = _verification_bm->startWord();
6851 assert(_collector->_restart_addr == NULL, "Sanity check");
6852 assert(_span.contains(_finger), "Out of bounds _finger?");
6853 }
6854
6855 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6856 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6857 assert(_span.contains(addr), "Out of bounds _finger?");
6858 _finger = addr;
6859 }
6860
6861 // Should revisit to see if this should be restructured for
6862 // greater efficiency.
6863 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6864 // convert offset into a HeapWord*
6865 HeapWord* addr = _verification_bm->startWord() + offset;
6866 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6867 "address out of range");
6868 assert(_verification_bm->isMarked(addr), "tautology");
6869 assert(_cms_bm->isMarked(addr), "tautology");
6870
6871 assert(_mark_stack->isEmpty(),
6872 "should drain stack to limit stack usage");
6873 // convert addr to an oop preparatory to scanning
6874 oop obj = oop(addr);
6875 assert(obj->is_oop(), "should be an oop");
6876 assert(_finger <= addr, "_finger runneth ahead");
6877 // advance the finger to right end of this object
6878 _finger = addr + obj->size();
6879 assert(_finger > addr, "we just incremented it above");
6880 // Note: the finger doesn't advance while we drain
6881 // the stack below.
6882 bool res = _mark_stack->push(obj);
6883 assert(res, "Empty non-zero size stack should have space for single push");
6884 while (!_mark_stack->isEmpty()) {
6885 oop new_oop = _mark_stack->pop();
6886 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
6887 // now scan this oop's oops
6888 new_oop->oop_iterate(&_pam_verify_closure);
6889 }
6890 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6891 return true;
6892 }
6893
6894 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6895 CMSCollector* collector, MemRegion span,
6896 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6897 CMSMarkStack* mark_stack):
6898 MetadataAwareOopClosure(collector->ref_processor()),
6899 _collector(collector),
6900 _span(span),
6901 _verification_bm(verification_bm),
6902 _cms_bm(cms_bm),
6903 _mark_stack(mark_stack)
6904 { }
6905
6906 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6907 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6908
6909 // Upon stack overflow, we discard (part of) the stack,
6910 // remembering the least address amongst those discarded
6911 // in CMSCollector's _restart_address.
6912 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6913 // Remember the least grey address discarded
6914 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6915 _collector->lower_restart_addr(ra);
6916 _mark_stack->reset(); // discard stack contents
6917 _mark_stack->expand(); // expand the stack if possible
6918 }
6919
6920 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6921 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
6922 HeapWord* addr = (HeapWord*)obj;
6923 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6924 // Oop lies in _span and isn't yet grey or black
6925 _verification_bm->mark(addr); // now grey
6926 if (!_cms_bm->isMarked(addr)) {
6927 oop(addr)->print();
6928 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
6929 p2i(addr));
6930 fatal("... aborting");
6931 }
6932
6933 if (!_mark_stack->push(obj)) { // stack overflow
6934 if (PrintCMSStatistics != 0) {
6935 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
6936 SIZE_FORMAT, _mark_stack->capacity());
6937 }
6938 assert(_mark_stack->isFull(), "Else push should have succeeded");
6939 handle_stack_overflow(addr);
6940 }
6941 // anything including and to the right of _finger
6942 // will be scanned as we iterate over the remainder of the
6943 // bit map
6944 }
6945 }
6946
6947 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6948 MemRegion span,
6949 CMSBitMap* bitMap, CMSMarkStack* markStack,
6950 HeapWord* finger, MarkFromRootsClosure* parent) :
6951 MetadataAwareOopClosure(collector->ref_processor()),
6952 _collector(collector),
6953 _span(span),
6954 _bitMap(bitMap),
6955 _markStack(markStack),
6956 _finger(finger),
6957 _parent(parent)
6958 { }
6959
6960 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
6961 MemRegion span,
6962 CMSBitMap* bit_map,
6963 OopTaskQueue* work_queue,
6964 CMSMarkStack* overflow_stack,
6965 HeapWord* finger,
6966 HeapWord** global_finger_addr,
6967 Par_MarkFromRootsClosure* parent) :
6968 MetadataAwareOopClosure(collector->ref_processor()),
6969 _collector(collector),
6970 _whole_span(collector->_span),
6971 _span(span),
6972 _bit_map(bit_map),
6973 _work_queue(work_queue),
6974 _overflow_stack(overflow_stack),
6975 _finger(finger),
6976 _global_finger_addr(global_finger_addr),
6977 _parent(parent)
6978 { }
6979
6980 // Assumes thread-safe access by callers, who are
6981 // responsible for mutual exclusion.
6982 void CMSCollector::lower_restart_addr(HeapWord* low) {
6983 assert(_span.contains(low), "Out of bounds addr");
6984 if (_restart_addr == NULL) {
6985 _restart_addr = low;
6986 } else {
6987 _restart_addr = MIN2(_restart_addr, low);
6988 }
6989 }
6990
6991 // Upon stack overflow, we discard (part of) the stack,
6992 // remembering the least address amongst those discarded
6993 // in CMSCollector's _restart_address.
6994 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6995 // Remember the least grey address discarded
6996 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6997 _collector->lower_restart_addr(ra);
6998 _markStack->reset(); // discard stack contents
6999 _markStack->expand(); // expand the stack if possible
7000 }
7001
7002 // Upon stack overflow, we discard (part of) the stack,
7003 // remembering the least address amongst those discarded
7004 // in CMSCollector's _restart_address.
7005 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7006 // We need to do this under a mutex to prevent other
7007 // workers from interfering with the work done below.
7008 MutexLockerEx ml(_overflow_stack->par_lock(),
7009 Mutex::_no_safepoint_check_flag);
7010 // Remember the least grey address discarded
7011 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7012 _collector->lower_restart_addr(ra);
7013 _overflow_stack->reset(); // discard stack contents
7014 _overflow_stack->expand(); // expand the stack if possible
7015 }
7016
7017 void PushOrMarkClosure::do_oop(oop obj) {
7018 // Ignore mark word because we are running concurrent with mutators.
7019 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7020 HeapWord* addr = (HeapWord*)obj;
7021 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7022 // Oop lies in _span and isn't yet grey or black
7023 _bitMap->mark(addr); // now grey
7024 if (addr < _finger) {
7025 // the bit map iteration has already either passed, or
7026 // sampled, this bit in the bit map; we'll need to
7027 // use the marking stack to scan this oop's oops.
7028 bool simulate_overflow = false;
7029 NOT_PRODUCT(
7030 if (CMSMarkStackOverflowALot &&
7031 _collector->simulate_overflow()) {
7032 // simulate a stack overflow
7033 simulate_overflow = true;
7034 }
7035 )
7036 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7037 if (PrintCMSStatistics != 0) {
7038 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7039 SIZE_FORMAT, _markStack->capacity());
7040 }
7041 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7042 handle_stack_overflow(addr);
7043 }
7044 }
7045 // anything including and to the right of _finger
7046 // will be scanned as we iterate over the remainder of the
7047 // bit map
7048 do_yield_check();
7049 }
7050 }
7051
7052 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7053 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7054
7055 void Par_PushOrMarkClosure::do_oop(oop obj) {
7056 // Ignore mark word because we are running concurrent with mutators.
7057 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7058 HeapWord* addr = (HeapWord*)obj;
7059 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7060 // Oop lies in _span and isn't yet grey or black
7061 // We read the global_finger (volatile read) strictly after marking oop
7062 bool res = _bit_map->par_mark(addr); // now grey
7063 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7064 // Should we push this marked oop on our stack?
7065 // -- if someone else marked it, nothing to do
7066 // -- if target oop is above global finger nothing to do
7067 // -- if target oop is in chunk and above local finger
7068 // then nothing to do
7069 // -- else push on work queue
7070 if ( !res // someone else marked it, they will deal with it
7071 || (addr >= *gfa) // will be scanned in a later task
7072 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7073 return;
7074 }
7075 // the bit map iteration has already either passed, or
7076 // sampled, this bit in the bit map; we'll need to
7077 // use the marking stack to scan this oop's oops.
7078 bool simulate_overflow = false;
7079 NOT_PRODUCT(
7080 if (CMSMarkStackOverflowALot &&
7081 _collector->simulate_overflow()) {
7082 // simulate a stack overflow
7083 simulate_overflow = true;
7084 }
7085 )
7086 if (simulate_overflow ||
7087 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7088 // stack overflow
7089 if (PrintCMSStatistics != 0) {
7090 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7091 SIZE_FORMAT, _overflow_stack->capacity());
7092 }
7093 // We cannot assert that the overflow stack is full because
7094 // it may have been emptied since.
7095 assert(simulate_overflow ||
7096 _work_queue->size() == _work_queue->max_elems(),
7097 "Else push should have succeeded");
7098 handle_stack_overflow(addr);
7099 }
7100 do_yield_check();
7101 }
7102 }
7103
7104 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7105 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7106
7107 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7108 MemRegion span,
7109 ReferenceProcessor* rp,
7110 CMSBitMap* bit_map,
7111 CMSBitMap* mod_union_table,
7112 CMSMarkStack* mark_stack,
7113 bool concurrent_precleaning):
7114 MetadataAwareOopClosure(rp),
7115 _collector(collector),
7116 _span(span),
7117 _bit_map(bit_map),
7118 _mod_union_table(mod_union_table),
7119 _mark_stack(mark_stack),
7120 _concurrent_precleaning(concurrent_precleaning)
7121 {
7122 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7123 }
7124
7125 // Grey object rescan during pre-cleaning and second checkpoint phases --
7126 // the non-parallel version (the parallel version appears further below.)
7127 void PushAndMarkClosure::do_oop(oop obj) {
7128 // Ignore mark word verification. If during concurrent precleaning,
7129 // the object monitor may be locked. If during the checkpoint
7130 // phases, the object may already have been reached by a different
7131 // path and may be at the end of the global overflow list (so
7132 // the mark word may be NULL).
7133 assert(obj->is_oop_or_null(true /* ignore mark word */),
7134 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7135 HeapWord* addr = (HeapWord*)obj;
7136 // Check if oop points into the CMS generation
7137 // and is not marked
7138 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7139 // a white object ...
7140 _bit_map->mark(addr); // ... now grey
7141 // push on the marking stack (grey set)
7142 bool simulate_overflow = false;
7143 NOT_PRODUCT(
7144 if (CMSMarkStackOverflowALot &&
7145 _collector->simulate_overflow()) {
7146 // simulate a stack overflow
7147 simulate_overflow = true;
7148 }
7149 )
7150 if (simulate_overflow || !_mark_stack->push(obj)) {
7151 if (_concurrent_precleaning) {
7152 // During precleaning we can just dirty the appropriate card(s)
7153 // in the mod union table, thus ensuring that the object remains
7154 // in the grey set and continue. In the case of object arrays
7155 // we need to dirty all of the cards that the object spans,
7156 // since the rescan of object arrays will be limited to the
7157 // dirty cards.
7158 // Note that no one can be interfering with us in this action
7159 // of dirtying the mod union table, so no locking or atomics
7160 // are required.
7161 if (obj->is_objArray()) {
7162 size_t sz = obj->size();
7163 HeapWord* end_card_addr = (HeapWord*)round_to(
7164 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7165 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7166 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7167 _mod_union_table->mark_range(redirty_range);
7168 } else {
7169 _mod_union_table->mark(addr);
7170 }
7171 _collector->_ser_pmc_preclean_ovflw++;
7172 } else {
7173 // During the remark phase, we need to remember this oop
7174 // in the overflow list.
7175 _collector->push_on_overflow_list(obj);
7176 _collector->_ser_pmc_remark_ovflw++;
7177 }
7178 }
7179 }
7180 }
7181
7182 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7183 MemRegion span,
7184 ReferenceProcessor* rp,
7185 CMSBitMap* bit_map,
7186 OopTaskQueue* work_queue):
7187 MetadataAwareOopClosure(rp),
7188 _collector(collector),
7189 _span(span),
7190 _bit_map(bit_map),
7191 _work_queue(work_queue)
7192 {
7193 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7194 }
7195
7196 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7197 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7198
7199 // Grey object rescan during second checkpoint phase --
7200 // the parallel version.
7201 void Par_PushAndMarkClosure::do_oop(oop obj) {
7202 // In the assert below, we ignore the mark word because
7203 // this oop may point to an already visited object that is
7204 // on the overflow stack (in which case the mark word has
7205 // been hijacked for chaining into the overflow stack --
7206 // if this is the last object in the overflow stack then
7207 // its mark word will be NULL). Because this object may
7208 // have been subsequently popped off the global overflow
7209 // stack, and the mark word possibly restored to the prototypical
7210 // value, by the time we get to examined this failing assert in
7211 // the debugger, is_oop_or_null(false) may subsequently start
7212 // to hold.
7213 assert(obj->is_oop_or_null(true),
7214 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7215 HeapWord* addr = (HeapWord*)obj;
7216 // Check if oop points into the CMS generation
7217 // and is not marked
7218 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7219 // a white object ...
7220 // If we manage to "claim" the object, by being the
7221 // first thread to mark it, then we push it on our
7222 // marking stack
7223 if (_bit_map->par_mark(addr)) { // ... now grey
7224 // push on work queue (grey set)
7225 bool simulate_overflow = false;
7226 NOT_PRODUCT(
7227 if (CMSMarkStackOverflowALot &&
7228 _collector->par_simulate_overflow()) {
7229 // simulate a stack overflow
7230 simulate_overflow = true;
7231 }
7232 )
7233 if (simulate_overflow || !_work_queue->push(obj)) {
7234 _collector->par_push_on_overflow_list(obj);
7235 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7236 }
7237 } // Else, some other thread got there first
7238 }
7239 }
7240
7241 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7242 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7243
7244 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7245 Mutex* bml = _collector->bitMapLock();
7246 assert_lock_strong(bml);
7247 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7248 "CMS thread should hold CMS token");
7249
7250 bml->unlock();
7251 ConcurrentMarkSweepThread::desynchronize(true);
7252
7253 _collector->stopTimer();
7254 if (PrintCMSStatistics != 0) {
7255 _collector->incrementYields();
7256 }
7257
7258 // See the comment in coordinator_yield()
7259 for (unsigned i = 0; i < CMSYieldSleepCount &&
7260 ConcurrentMarkSweepThread::should_yield() &&
7261 !CMSCollector::foregroundGCIsActive(); ++i) {
7262 os::sleep(Thread::current(), 1, false);
7263 }
7264
7265 ConcurrentMarkSweepThread::synchronize(true);
7266 bml->lock();
7267
7268 _collector->startTimer();
7269 }
7270
7271 bool CMSPrecleanRefsYieldClosure::should_return() {
7272 if (ConcurrentMarkSweepThread::should_yield()) {
7273 do_yield_work();
7274 }
7275 return _collector->foregroundGCIsActive();
7276 }
7277
7278 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7279 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7280 "mr should be aligned to start at a card boundary");
7281 // We'd like to assert:
7282 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7283 // "mr should be a range of cards");
7284 // However, that would be too strong in one case -- the last
7285 // partition ends at _unallocated_block which, in general, can be
7286 // an arbitrary boundary, not necessarily card aligned.
7287 if (PrintCMSStatistics != 0) {
7288 _num_dirty_cards +=
7289 mr.word_size()/CardTableModRefBS::card_size_in_words;
7290 }
7291 _space->object_iterate_mem(mr, &_scan_cl);
7292 }
7293
7294 SweepClosure::SweepClosure(CMSCollector* collector,
7295 ConcurrentMarkSweepGeneration* g,
7296 CMSBitMap* bitMap, bool should_yield) :
7297 _collector(collector),
7298 _g(g),
7299 _sp(g->cmsSpace()),
7300 _limit(_sp->sweep_limit()),
7301 _freelistLock(_sp->freelistLock()),
7302 _bitMap(bitMap),
7303 _yield(should_yield),
7304 _inFreeRange(false), // No free range at beginning of sweep
7305 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7306 _lastFreeRangeCoalesced(false),
7307 _freeFinger(g->used_region().start())
7308 {
7309 NOT_PRODUCT(
7310 _numObjectsFreed = 0;
7311 _numWordsFreed = 0;
7312 _numObjectsLive = 0;
7313 _numWordsLive = 0;
7314 _numObjectsAlreadyFree = 0;
7315 _numWordsAlreadyFree = 0;
7316 _last_fc = NULL;
7317
7318 _sp->initializeIndexedFreeListArrayReturnedBytes();
7319 _sp->dictionary()->initialize_dict_returned_bytes();
7320 )
7321 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7322 "sweep _limit out of bounds");
7323 if (CMSTraceSweeper) {
7324 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
7325 p2i(_limit));
7326 }
7327 }
7328
7329 void SweepClosure::print_on(outputStream* st) const {
7330 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7331 p2i(_sp->bottom()), p2i(_sp->end()));
7332 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7333 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7334 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7335 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7336 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7337 }
7338
7339 #ifndef PRODUCT
7340 // Assertion checking only: no useful work in product mode --
7341 // however, if any of the flags below become product flags,
7342 // you may need to review this code to see if it needs to be
7343 // enabled in product mode.
7344 SweepClosure::~SweepClosure() {
7345 assert_lock_strong(_freelistLock);
7346 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7347 "sweep _limit out of bounds");
7348 if (inFreeRange()) {
7349 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
7350 print();
7351 ShouldNotReachHere();
7352 }
7353 if (Verbose && PrintGC) {
7354 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes",
7355 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7356 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
7357 SIZE_FORMAT" bytes "
7358 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
7359 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7360 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7361 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
7362 * sizeof(HeapWord);
7363 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
7364
7365 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7366 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7367 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7368 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7369 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes);
7370 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
7371 indexListReturnedBytes);
7372 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
7373 dict_returned_bytes);
7374 }
7375 }
7376 if (CMSTraceSweeper) {
7377 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
7378 p2i(_limit));
7379 }
7380 }
7381 #endif // PRODUCT
7382
7383 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7384 bool freeRangeInFreeLists) {
7385 if (CMSTraceSweeper) {
7386 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n",
7387 p2i(freeFinger), freeRangeInFreeLists);
7388 }
7389 assert(!inFreeRange(), "Trampling existing free range");
7390 set_inFreeRange(true);
7391 set_lastFreeRangeCoalesced(false);
7392
7393 set_freeFinger(freeFinger);
7394 set_freeRangeInFreeLists(freeRangeInFreeLists);
7395 if (CMSTestInFreeList) {
7396 if (freeRangeInFreeLists) {
7397 FreeChunk* fc = (FreeChunk*) freeFinger;
7398 assert(fc->is_free(), "A chunk on the free list should be free.");
7399 assert(fc->size() > 0, "Free range should have a size");
7400 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7401 }
7402 }
7403 }
7404
7405 // Note that the sweeper runs concurrently with mutators. Thus,
7406 // it is possible for direct allocation in this generation to happen
7407 // in the middle of the sweep. Note that the sweeper also coalesces
7408 // contiguous free blocks. Thus, unless the sweeper and the allocator
7409 // synchronize appropriately freshly allocated blocks may get swept up.
7410 // This is accomplished by the sweeper locking the free lists while
7411 // it is sweeping. Thus blocks that are determined to be free are
7412 // indeed free. There is however one additional complication:
7413 // blocks that have been allocated since the final checkpoint and
7414 // mark, will not have been marked and so would be treated as
7415 // unreachable and swept up. To prevent this, the allocator marks
7416 // the bit map when allocating during the sweep phase. This leads,
7417 // however, to a further complication -- objects may have been allocated
7418 // but not yet initialized -- in the sense that the header isn't yet
7419 // installed. The sweeper can not then determine the size of the block
7420 // in order to skip over it. To deal with this case, we use a technique
7421 // (due to Printezis) to encode such uninitialized block sizes in the
7422 // bit map. Since the bit map uses a bit per every HeapWord, but the
7423 // CMS generation has a minimum object size of 3 HeapWords, it follows
7424 // that "normal marks" won't be adjacent in the bit map (there will
7425 // always be at least two 0 bits between successive 1 bits). We make use
7426 // of these "unused" bits to represent uninitialized blocks -- the bit
7427 // corresponding to the start of the uninitialized object and the next
7428 // bit are both set. Finally, a 1 bit marks the end of the object that
7429 // started with the two consecutive 1 bits to indicate its potentially
7430 // uninitialized state.
7431
7432 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7433 FreeChunk* fc = (FreeChunk*)addr;
7434 size_t res;
7435
7436 // Check if we are done sweeping. Below we check "addr >= _limit" rather
7437 // than "addr == _limit" because although _limit was a block boundary when
7438 // we started the sweep, it may no longer be one because heap expansion
7439 // may have caused us to coalesce the block ending at the address _limit
7440 // with a newly expanded chunk (this happens when _limit was set to the
7441 // previous _end of the space), so we may have stepped past _limit:
7442 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7443 if (addr >= _limit) { // we have swept up to or past the limit: finish up
7444 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7445 "sweep _limit out of bounds");
7446 assert(addr < _sp->end(), "addr out of bounds");
7447 // Flush any free range we might be holding as a single
7448 // coalesced chunk to the appropriate free list.
7449 if (inFreeRange()) {
7450 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7451 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", p2i(freeFinger())));
7452 flush_cur_free_chunk(freeFinger(),
7453 pointer_delta(addr, freeFinger()));
7454 if (CMSTraceSweeper) {
7455 gclog_or_tty->print("Sweep: last chunk: ");
7456 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") "
7457 "[coalesced:%d]\n",
7458 p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7459 lastFreeRangeCoalesced() ? 1 : 0);
7460 }
7461 }
7462
7463 // help the iterator loop finish
7464 return pointer_delta(_sp->end(), addr);
7465 }
7466
7467 assert(addr < _limit, "sweep invariant");
7468 // check if we should yield
7469 do_yield_check(addr);
7470 if (fc->is_free()) {
7471 // Chunk that is already free
7472 res = fc->size();
7473 do_already_free_chunk(fc);
7474 debug_only(_sp->verifyFreeLists());
7475 // If we flush the chunk at hand in lookahead_and_flush()
7476 // and it's coalesced with a preceding chunk, then the
7477 // process of "mangling" the payload of the coalesced block
7478 // will cause erasure of the size information from the
7479 // (erstwhile) header of all the coalesced blocks but the
7480 // first, so the first disjunct in the assert will not hold
7481 // in that specific case (in which case the second disjunct
7482 // will hold).
7483 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7484 "Otherwise the size info doesn't change at this step");
7485 NOT_PRODUCT(
7486 _numObjectsAlreadyFree++;
7487 _numWordsAlreadyFree += res;
7488 )
7489 NOT_PRODUCT(_last_fc = fc;)
7490 } else if (!_bitMap->isMarked(addr)) {
7491 // Chunk is fresh garbage
7492 res = do_garbage_chunk(fc);
7493 debug_only(_sp->verifyFreeLists());
7494 NOT_PRODUCT(
7495 _numObjectsFreed++;
7496 _numWordsFreed += res;
7497 )
7498 } else {
7499 // Chunk that is alive.
7500 res = do_live_chunk(fc);
7501 debug_only(_sp->verifyFreeLists());
7502 NOT_PRODUCT(
7503 _numObjectsLive++;
7504 _numWordsLive += res;
7505 )
7506 }
7507 return res;
7508 }
7509
7510 // For the smart allocation, record following
7511 // split deaths - a free chunk is removed from its free list because
7512 // it is being split into two or more chunks.
7513 // split birth - a free chunk is being added to its free list because
7514 // a larger free chunk has been split and resulted in this free chunk.
7515 // coal death - a free chunk is being removed from its free list because
7516 // it is being coalesced into a large free chunk.
7517 // coal birth - a free chunk is being added to its free list because
7518 // it was created when two or more free chunks where coalesced into
7519 // this free chunk.
7520 //
7521 // These statistics are used to determine the desired number of free
7522 // chunks of a given size. The desired number is chosen to be relative
7523 // to the end of a CMS sweep. The desired number at the end of a sweep
7524 // is the
7525 // count-at-end-of-previous-sweep (an amount that was enough)
7526 // - count-at-beginning-of-current-sweep (the excess)
7527 // + split-births (gains in this size during interval)
7528 // - split-deaths (demands on this size during interval)
7529 // where the interval is from the end of one sweep to the end of the
7530 // next.
7531 //
7532 // When sweeping the sweeper maintains an accumulated chunk which is
7533 // the chunk that is made up of chunks that have been coalesced. That
7534 // will be termed the left-hand chunk. A new chunk of garbage that
7535 // is being considered for coalescing will be referred to as the
7536 // right-hand chunk.
7537 //
7538 // When making a decision on whether to coalesce a right-hand chunk with
7539 // the current left-hand chunk, the current count vs. the desired count
7540 // of the left-hand chunk is considered. Also if the right-hand chunk
7541 // is near the large chunk at the end of the heap (see
7542 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7543 // left-hand chunk is coalesced.
7544 //
7545 // When making a decision about whether to split a chunk, the desired count
7546 // vs. the current count of the candidate to be split is also considered.
7547 // If the candidate is underpopulated (currently fewer chunks than desired)
7548 // a chunk of an overpopulated (currently more chunks than desired) size may
7549 // be chosen. The "hint" associated with a free list, if non-null, points
7550 // to a free list which may be overpopulated.
7551 //
7552
7553 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7554 const size_t size = fc->size();
7555 // Chunks that cannot be coalesced are not in the
7556 // free lists.
7557 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7558 assert(_sp->verify_chunk_in_free_list(fc),
7559 "free chunk should be in free lists");
7560 }
7561 // a chunk that is already free, should not have been
7562 // marked in the bit map
7563 HeapWord* const addr = (HeapWord*) fc;
7564 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7565 // Verify that the bit map has no bits marked between
7566 // addr and purported end of this block.
7567 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7568
7569 // Some chunks cannot be coalesced under any circumstances.
7570 // See the definition of cantCoalesce().
7571 if (!fc->cantCoalesce()) {
7572 // This chunk can potentially be coalesced.
7573 if (_sp->adaptive_freelists()) {
7574 // All the work is done in
7575 do_post_free_or_garbage_chunk(fc, size);
7576 } else { // Not adaptive free lists
7577 // this is a free chunk that can potentially be coalesced by the sweeper;
7578 if (!inFreeRange()) {
7579 // if the next chunk is a free block that can't be coalesced
7580 // it doesn't make sense to remove this chunk from the free lists
7581 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
7582 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
7583 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
7584 nextChunk->is_free() && // ... which is free...
7585 nextChunk->cantCoalesce()) { // ... but can't be coalesced
7586 // nothing to do
7587 } else {
7588 // Potentially the start of a new free range:
7589 // Don't eagerly remove it from the free lists.
7590 // No need to remove it if it will just be put
7591 // back again. (Also from a pragmatic point of view
7592 // if it is a free block in a region that is beyond
7593 // any allocated blocks, an assertion will fail)
7594 // Remember the start of a free run.
7595 initialize_free_range(addr, true);
7596 // end - can coalesce with next chunk
7597 }
7598 } else {
7599 // the midst of a free range, we are coalescing
7600 print_free_block_coalesced(fc);
7601 if (CMSTraceSweeper) {
7602 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size);
7603 }
7604 // remove it from the free lists
7605 _sp->removeFreeChunkFromFreeLists(fc);
7606 set_lastFreeRangeCoalesced(true);
7607 // If the chunk is being coalesced and the current free range is
7608 // in the free lists, remove the current free range so that it
7609 // will be returned to the free lists in its entirety - all
7610 // the coalesced pieces included.
7611 if (freeRangeInFreeLists()) {
7612 FreeChunk* ffc = (FreeChunk*) freeFinger();
7613 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7614 "Size of free range is inconsistent with chunk size.");
7615 if (CMSTestInFreeList) {
7616 assert(_sp->verify_chunk_in_free_list(ffc),
7617 "free range is not in free lists");
7618 }
7619 _sp->removeFreeChunkFromFreeLists(ffc);
7620 set_freeRangeInFreeLists(false);
7621 }
7622 }
7623 }
7624 // Note that if the chunk is not coalescable (the else arm
7625 // below), we unconditionally flush, without needing to do
7626 // a "lookahead," as we do below.
7627 if (inFreeRange()) lookahead_and_flush(fc, size);
7628 } else {
7629 // Code path common to both original and adaptive free lists.
7630
7631 // cant coalesce with previous block; this should be treated
7632 // as the end of a free run if any
7633 if (inFreeRange()) {
7634 // we kicked some butt; time to pick up the garbage
7635 assert(freeFinger() < addr, "freeFinger points too high");
7636 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7637 }
7638 // else, nothing to do, just continue
7639 }
7640 }
7641
7642 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7643 // This is a chunk of garbage. It is not in any free list.
7644 // Add it to a free list or let it possibly be coalesced into
7645 // a larger chunk.
7646 HeapWord* const addr = (HeapWord*) fc;
7647 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7648
7649 if (_sp->adaptive_freelists()) {
7650 // Verify that the bit map has no bits marked between
7651 // addr and purported end of just dead object.
7652 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7653
7654 do_post_free_or_garbage_chunk(fc, size);
7655 } else {
7656 if (!inFreeRange()) {
7657 // start of a new free range
7658 assert(size > 0, "A free range should have a size");
7659 initialize_free_range(addr, false);
7660 } else {
7661 // this will be swept up when we hit the end of the
7662 // free range
7663 if (CMSTraceSweeper) {
7664 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size);
7665 }
7666 // If the chunk is being coalesced and the current free range is
7667 // in the free lists, remove the current free range so that it
7668 // will be returned to the free lists in its entirety - all
7669 // the coalesced pieces included.
7670 if (freeRangeInFreeLists()) {
7671 FreeChunk* ffc = (FreeChunk*)freeFinger();
7672 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7673 "Size of free range is inconsistent with chunk size.");
7674 if (CMSTestInFreeList) {
7675 assert(_sp->verify_chunk_in_free_list(ffc),
7676 "free range is not in free lists");
7677 }
7678 _sp->removeFreeChunkFromFreeLists(ffc);
7679 set_freeRangeInFreeLists(false);
7680 }
7681 set_lastFreeRangeCoalesced(true);
7682 }
7683 // this will be swept up when we hit the end of the free range
7684
7685 // Verify that the bit map has no bits marked between
7686 // addr and purported end of just dead object.
7687 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7688 }
7689 assert(_limit >= addr + size,
7690 "A freshly garbage chunk can't possibly straddle over _limit");
7691 if (inFreeRange()) lookahead_and_flush(fc, size);
7692 return size;
7693 }
7694
7695 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7696 HeapWord* addr = (HeapWord*) fc;
7697 // The sweeper has just found a live object. Return any accumulated
7698 // left hand chunk to the free lists.
7699 if (inFreeRange()) {
7700 assert(freeFinger() < addr, "freeFinger points too high");
7701 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7702 }
7703
7704 // This object is live: we'd normally expect this to be
7705 // an oop, and like to assert the following:
7706 // assert(oop(addr)->is_oop(), "live block should be an oop");
7707 // However, as we commented above, this may be an object whose
7708 // header hasn't yet been initialized.
7709 size_t size;
7710 assert(_bitMap->isMarked(addr), "Tautology for this control point");
7711 if (_bitMap->isMarked(addr + 1)) {
7712 // Determine the size from the bit map, rather than trying to
7713 // compute it from the object header.
7714 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7715 size = pointer_delta(nextOneAddr + 1, addr);
7716 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7717 "alignment problem");
7718
7719 #ifdef ASSERT
7720 if (oop(addr)->klass_or_null() != NULL) {
7721 // Ignore mark word because we are running concurrent with mutators
7722 assert(oop(addr)->is_oop(true), "live block should be an oop");
7723 assert(size ==
7724 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7725 "P-mark and computed size do not agree");
7726 }
7727 #endif
7728
7729 } else {
7730 // This should be an initialized object that's alive.
7731 assert(oop(addr)->klass_or_null() != NULL,
7732 "Should be an initialized object");
7733 // Ignore mark word because we are running concurrent with mutators
7734 assert(oop(addr)->is_oop(true), "live block should be an oop");
7735 // Verify that the bit map has no bits marked between
7736 // addr and purported end of this block.
7737 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7738 assert(size >= 3, "Necessary for Printezis marks to work");
7739 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7740 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7741 }
7742 return size;
7743 }
7744
7745 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7746 size_t chunkSize) {
7747 // do_post_free_or_garbage_chunk() should only be called in the case
7748 // of the adaptive free list allocator.
7749 const bool fcInFreeLists = fc->is_free();
7750 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
7751 assert((HeapWord*)fc <= _limit, "sweep invariant");
7752 if (CMSTestInFreeList && fcInFreeLists) {
7753 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7754 }
7755
7756 if (CMSTraceSweeper) {
7757 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7758 }
7759
7760 HeapWord* const fc_addr = (HeapWord*) fc;
7761
7762 bool coalesce;
7763 const size_t left = pointer_delta(fc_addr, freeFinger());
7764 const size_t right = chunkSize;
7765 switch (FLSCoalescePolicy) {
7766 // numeric value forms a coalition aggressiveness metric
7767 case 0: { // never coalesce
7768 coalesce = false;
7769 break;
7770 }
7771 case 1: { // coalesce if left & right chunks on overpopulated lists
7772 coalesce = _sp->coalOverPopulated(left) &&
7773 _sp->coalOverPopulated(right);
7774 break;
7775 }
7776 case 2: { // coalesce if left chunk on overpopulated list (default)
7777 coalesce = _sp->coalOverPopulated(left);
7778 break;
7779 }
7780 case 3: { // coalesce if left OR right chunk on overpopulated list
7781 coalesce = _sp->coalOverPopulated(left) ||
7782 _sp->coalOverPopulated(right);
7783 break;
7784 }
7785 case 4: { // always coalesce
7786 coalesce = true;
7787 break;
7788 }
7789 default:
7790 ShouldNotReachHere();
7791 }
7792
7793 // Should the current free range be coalesced?
7794 // If the chunk is in a free range and either we decided to coalesce above
7795 // or the chunk is near the large block at the end of the heap
7796 // (isNearLargestChunk() returns true), then coalesce this chunk.
7797 const bool doCoalesce = inFreeRange()
7798 && (coalesce || _g->isNearLargestChunk(fc_addr));
7799 if (doCoalesce) {
7800 // Coalesce the current free range on the left with the new
7801 // chunk on the right. If either is on a free list,
7802 // it must be removed from the list and stashed in the closure.
7803 if (freeRangeInFreeLists()) {
7804 FreeChunk* const ffc = (FreeChunk*)freeFinger();
7805 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7806 "Size of free range is inconsistent with chunk size.");
7807 if (CMSTestInFreeList) {
7808 assert(_sp->verify_chunk_in_free_list(ffc),
7809 "Chunk is not in free lists");
7810 }
7811 _sp->coalDeath(ffc->size());
7812 _sp->removeFreeChunkFromFreeLists(ffc);
7813 set_freeRangeInFreeLists(false);
7814 }
7815 if (fcInFreeLists) {
7816 _sp->coalDeath(chunkSize);
7817 assert(fc->size() == chunkSize,
7818 "The chunk has the wrong size or is not in the free lists");
7819 _sp->removeFreeChunkFromFreeLists(fc);
7820 }
7821 set_lastFreeRangeCoalesced(true);
7822 print_free_block_coalesced(fc);
7823 } else { // not in a free range and/or should not coalesce
7824 // Return the current free range and start a new one.
7825 if (inFreeRange()) {
7826 // In a free range but cannot coalesce with the right hand chunk.
7827 // Put the current free range into the free lists.
7828 flush_cur_free_chunk(freeFinger(),
7829 pointer_delta(fc_addr, freeFinger()));
7830 }
7831 // Set up for new free range. Pass along whether the right hand
7832 // chunk is in the free lists.
7833 initialize_free_range((HeapWord*)fc, fcInFreeLists);
7834 }
7835 }
7836
7837 // Lookahead flush:
7838 // If we are tracking a free range, and this is the last chunk that
7839 // we'll look at because its end crosses past _limit, we'll preemptively
7840 // flush it along with any free range we may be holding on to. Note that
7841 // this can be the case only for an already free or freshly garbage
7842 // chunk. If this block is an object, it can never straddle
7843 // over _limit. The "straddling" occurs when _limit is set at
7844 // the previous end of the space when this cycle started, and
7845 // a subsequent heap expansion caused the previously co-terminal
7846 // free block to be coalesced with the newly expanded portion,
7847 // thus rendering _limit a non-block-boundary making it dangerous
7848 // for the sweeper to step over and examine.
7849 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7850 assert(inFreeRange(), "Should only be called if currently in a free range.");
7851 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7852 assert(_sp->used_region().contains(eob - 1),
7853 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7854 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7855 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7856 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size));
7857 if (eob >= _limit) {
7858 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7859 if (CMSTraceSweeper) {
7860 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
7861 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7862 "[" PTR_FORMAT "," PTR_FORMAT ")",
7863 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7864 }
7865 // Return the storage we are tracking back into the free lists.
7866 if (CMSTraceSweeper) {
7867 gclog_or_tty->print_cr("Flushing ... ");
7868 }
7869 assert(freeFinger() < eob, "Error");
7870 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7871 }
7872 }
7873
7874 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7875 assert(inFreeRange(), "Should only be called if currently in a free range.");
7876 assert(size > 0,
7877 "A zero sized chunk cannot be added to the free lists.");
7878 if (!freeRangeInFreeLists()) {
7879 if (CMSTestInFreeList) {
7880 FreeChunk* fc = (FreeChunk*) chunk;
7881 fc->set_size(size);
7882 assert(!_sp->verify_chunk_in_free_list(fc),
7883 "chunk should not be in free lists yet");
7884 }
7885 if (CMSTraceSweeper) {
7886 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists",
7887 p2i(chunk), size);
7888 }
7889 // A new free range is going to be starting. The current
7890 // free range has not been added to the free lists yet or
7891 // was removed so add it back.
7892 // If the current free range was coalesced, then the death
7893 // of the free range was recorded. Record a birth now.
7894 if (lastFreeRangeCoalesced()) {
7895 _sp->coalBirth(size);
7896 }
7897 _sp->addChunkAndRepairOffsetTable(chunk, size,
7898 lastFreeRangeCoalesced());
7899 } else if (CMSTraceSweeper) {
7900 gclog_or_tty->print_cr("Already in free list: nothing to flush");
7901 }
7902 set_inFreeRange(false);
7903 set_freeRangeInFreeLists(false);
7904 }
7905
7906 // We take a break if we've been at this for a while,
7907 // so as to avoid monopolizing the locks involved.
7908 void SweepClosure::do_yield_work(HeapWord* addr) {
7909 // Return current free chunk being used for coalescing (if any)
7910 // to the appropriate freelist. After yielding, the next
7911 // free block encountered will start a coalescing range of
7912 // free blocks. If the next free block is adjacent to the
7913 // chunk just flushed, they will need to wait for the next
7914 // sweep to be coalesced.
7915 if (inFreeRange()) {
7916 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7917 }
7918
7919 // First give up the locks, then yield, then re-lock.
7920 // We should probably use a constructor/destructor idiom to
7921 // do this unlock/lock or modify the MutexUnlocker class to
7922 // serve our purpose. XXX
7923 assert_lock_strong(_bitMap->lock());
7924 assert_lock_strong(_freelistLock);
7925 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7926 "CMS thread should hold CMS token");
7927 _bitMap->lock()->unlock();
7928 _freelistLock->unlock();
7929 ConcurrentMarkSweepThread::desynchronize(true);
7930 _collector->stopTimer();
7931 if (PrintCMSStatistics != 0) {
7932 _collector->incrementYields();
7933 }
7934
7935 // See the comment in coordinator_yield()
7936 for (unsigned i = 0; i < CMSYieldSleepCount &&
7937 ConcurrentMarkSweepThread::should_yield() &&
7938 !CMSCollector::foregroundGCIsActive(); ++i) {
7939 os::sleep(Thread::current(), 1, false);
7940 }
7941
7942 ConcurrentMarkSweepThread::synchronize(true);
7943 _freelistLock->lock();
7944 _bitMap->lock()->lock_without_safepoint_check();
7945 _collector->startTimer();
7946 }
7947
7948 #ifndef PRODUCT
7949 // This is actually very useful in a product build if it can
7950 // be called from the debugger. Compile it into the product
7951 // as needed.
7952 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7953 return debug_cms_space->verify_chunk_in_free_list(fc);
7954 }
7955 #endif
7956
7957 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7958 if (CMSTraceSweeper) {
7959 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7960 p2i(fc), fc->size());
7961 }
7962 }
7963
7964 // CMSIsAliveClosure
7965 bool CMSIsAliveClosure::do_object_b(oop obj) {
7966 HeapWord* addr = (HeapWord*)obj;
7967 return addr != NULL &&
7968 (!_span.contains(addr) || _bit_map->isMarked(addr));
7969 }
7970
7971
7972 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7973 MemRegion span,
7974 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7975 bool cpc):
7976 _collector(collector),
7977 _span(span),
7978 _bit_map(bit_map),
7979 _mark_stack(mark_stack),
7980 _concurrent_precleaning(cpc) {
7981 assert(!_span.is_empty(), "Empty span could spell trouble");
7982 }
7983
7984
7985 // CMSKeepAliveClosure: the serial version
7986 void CMSKeepAliveClosure::do_oop(oop obj) {
7987 HeapWord* addr = (HeapWord*)obj;
7988 if (_span.contains(addr) &&
7989 !_bit_map->isMarked(addr)) {
7990 _bit_map->mark(addr);
7991 bool simulate_overflow = false;
7992 NOT_PRODUCT(
7993 if (CMSMarkStackOverflowALot &&
7994 _collector->simulate_overflow()) {
7995 // simulate a stack overflow
7996 simulate_overflow = true;
7997 }
7998 )
7999 if (simulate_overflow || !_mark_stack->push(obj)) {
8000 if (_concurrent_precleaning) {
8001 // We dirty the overflown object and let the remark
8002 // phase deal with it.
8003 assert(_collector->overflow_list_is_empty(), "Error");
8004 // In the case of object arrays, we need to dirty all of
8005 // the cards that the object spans. No locking or atomics
8006 // are needed since no one else can be mutating the mod union
8007 // table.
8008 if (obj->is_objArray()) {
8009 size_t sz = obj->size();
8010 HeapWord* end_card_addr =
8011 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
8012 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8013 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8014 _collector->_modUnionTable.mark_range(redirty_range);
8015 } else {
8016 _collector->_modUnionTable.mark(addr);
8017 }
8018 _collector->_ser_kac_preclean_ovflw++;
8019 } else {
8020 _collector->push_on_overflow_list(obj);
8021 _collector->_ser_kac_ovflw++;
8022 }
8023 }
8024 }
8025 }
8026
8027 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8028 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8029
8030 // CMSParKeepAliveClosure: a parallel version of the above.
8031 // The work queues are private to each closure (thread),
8032 // but (may be) available for stealing by other threads.
8033 void CMSParKeepAliveClosure::do_oop(oop obj) {
8034 HeapWord* addr = (HeapWord*)obj;
8035 if (_span.contains(addr) &&
8036 !_bit_map->isMarked(addr)) {
8037 // In general, during recursive tracing, several threads
8038 // may be concurrently getting here; the first one to
8039 // "tag" it, claims it.
8040 if (_bit_map->par_mark(addr)) {
8041 bool res = _work_queue->push(obj);
8042 assert(res, "Low water mark should be much less than capacity");
8043 // Do a recursive trim in the hope that this will keep
8044 // stack usage lower, but leave some oops for potential stealers
8045 trim_queue(_low_water_mark);
8046 } // Else, another thread got there first
8047 }
8048 }
8049
8050 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8051 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8052
8053 void CMSParKeepAliveClosure::trim_queue(uint max) {
8054 while (_work_queue->size() > max) {
8055 oop new_oop;
8056 if (_work_queue->pop_local(new_oop)) {
8057 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8058 assert(_bit_map->isMarked((HeapWord*)new_oop),
8059 "no white objects on this stack!");
8060 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8061 // iterate over the oops in this oop, marking and pushing
8062 // the ones in CMS heap (i.e. in _span).
8063 new_oop->oop_iterate(&_mark_and_push);
8064 }
8065 }
8066 }
8067
8068 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8069 CMSCollector* collector,
8070 MemRegion span, CMSBitMap* bit_map,
8071 OopTaskQueue* work_queue):
8072 _collector(collector),
8073 _span(span),
8074 _bit_map(bit_map),
8075 _work_queue(work_queue) { }
8076
8077 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8078 HeapWord* addr = (HeapWord*)obj;
8079 if (_span.contains(addr) &&
8080 !_bit_map->isMarked(addr)) {
8081 if (_bit_map->par_mark(addr)) {
8082 bool simulate_overflow = false;
8083 NOT_PRODUCT(
8084 if (CMSMarkStackOverflowALot &&
8085 _collector->par_simulate_overflow()) {
8086 // simulate a stack overflow
8087 simulate_overflow = true;
8088 }
8089 )
8090 if (simulate_overflow || !_work_queue->push(obj)) {
8091 _collector->par_push_on_overflow_list(obj);
8092 _collector->_par_kac_ovflw++;
8093 }
8094 } // Else another thread got there already
8095 }
8096 }
8097
8098 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8099 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8100
8101 //////////////////////////////////////////////////////////////////
8102 // CMSExpansionCause /////////////////////////////
8103 //////////////////////////////////////////////////////////////////
8104 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8105 switch (cause) {
8106 case _no_expansion:
8107 return "No expansion";
8108 case _satisfy_free_ratio:
8109 return "Free ratio";
8110 case _satisfy_promotion:
8111 return "Satisfy promotion";
8112 case _satisfy_allocation:
8113 return "allocation";
8114 case _allocate_par_lab:
8115 return "Par LAB";
8116 case _allocate_par_spooling_space:
8117 return "Par Spooling Space";
8118 case _adaptive_size_policy:
8119 return "Ergonomics";
8120 default:
8121 return "unknown";
8122 }
8123 }
8124
8125 void CMSDrainMarkingStackClosure::do_void() {
8126 // the max number to take from overflow list at a time
8127 const size_t num = _mark_stack->capacity()/4;
8128 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8129 "Overflow list should be NULL during concurrent phases");
8130 while (!_mark_stack->isEmpty() ||
8131 // if stack is empty, check the overflow list
8132 _collector->take_from_overflow_list(num, _mark_stack)) {
8133 oop obj = _mark_stack->pop();
8134 HeapWord* addr = (HeapWord*)obj;
8135 assert(_span.contains(addr), "Should be within span");
8136 assert(_bit_map->isMarked(addr), "Should be marked");
8137 assert(obj->is_oop(), "Should be an oop");
8138 obj->oop_iterate(_keep_alive);
8139 }
8140 }
8141
8142 void CMSParDrainMarkingStackClosure::do_void() {
8143 // drain queue
8144 trim_queue(0);
8145 }
8146
8147 // Trim our work_queue so its length is below max at return
8148 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8149 while (_work_queue->size() > max) {
8150 oop new_oop;
8151 if (_work_queue->pop_local(new_oop)) {
8152 assert(new_oop->is_oop(), "Expected an oop");
8153 assert(_bit_map->isMarked((HeapWord*)new_oop),
8154 "no white objects on this stack!");
8155 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8156 // iterate over the oops in this oop, marking and pushing
8157 // the ones in CMS heap (i.e. in _span).
8158 new_oop->oop_iterate(&_mark_and_push);
8159 }
8160 }
8161 }
8162
8163 ////////////////////////////////////////////////////////////////////
8164 // Support for Marking Stack Overflow list handling and related code
8165 ////////////////////////////////////////////////////////////////////
8166 // Much of the following code is similar in shape and spirit to the
8167 // code used in ParNewGC. We should try and share that code
8168 // as much as possible in the future.
8169
8170 #ifndef PRODUCT
8171 // Debugging support for CMSStackOverflowALot
8172
8173 // It's OK to call this multi-threaded; the worst thing
8174 // that can happen is that we'll get a bunch of closely
8175 // spaced simulated overflows, but that's OK, in fact
8176 // probably good as it would exercise the overflow code
8177 // under contention.
8178 bool CMSCollector::simulate_overflow() {
8179 if (_overflow_counter-- <= 0) { // just being defensive
8180 _overflow_counter = CMSMarkStackOverflowInterval;
8181 return true;
8182 } else {
8183 return false;
8184 }
8185 }
8186
8187 bool CMSCollector::par_simulate_overflow() {
8188 return simulate_overflow();
8189 }
8190 #endif
8191
8192 // Single-threaded
8193 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8194 assert(stack->isEmpty(), "Expected precondition");
8195 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8196 size_t i = num;
8197 oop cur = _overflow_list;
8198 const markOop proto = markOopDesc::prototype();
8199 NOT_PRODUCT(ssize_t n = 0;)
8200 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8201 next = oop(cur->mark());
8202 cur->set_mark(proto); // until proven otherwise
8203 assert(cur->is_oop(), "Should be an oop");
8204 bool res = stack->push(cur);
8205 assert(res, "Bit off more than can chew?");
8206 NOT_PRODUCT(n++;)
8207 }
8208 _overflow_list = cur;
8209 #ifndef PRODUCT
8210 assert(_num_par_pushes >= n, "Too many pops?");
8211 _num_par_pushes -=n;
8212 #endif
8213 return !stack->isEmpty();
8214 }
8215
8216 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
8217 // (MT-safe) Get a prefix of at most "num" from the list.
8218 // The overflow list is chained through the mark word of
8219 // each object in the list. We fetch the entire list,
8220 // break off a prefix of the right size and return the
8221 // remainder. If other threads try to take objects from
8222 // the overflow list at that time, they will wait for
8223 // some time to see if data becomes available. If (and
8224 // only if) another thread places one or more object(s)
8225 // on the global list before we have returned the suffix
8226 // to the global list, we will walk down our local list
8227 // to find its end and append the global list to
8228 // our suffix before returning it. This suffix walk can
8229 // prove to be expensive (quadratic in the amount of traffic)
8230 // when there are many objects in the overflow list and
8231 // there is much producer-consumer contention on the list.
8232 // *NOTE*: The overflow list manipulation code here and
8233 // in ParNewGeneration:: are very similar in shape,
8234 // except that in the ParNew case we use the old (from/eden)
8235 // copy of the object to thread the list via its klass word.
8236 // Because of the common code, if you make any changes in
8237 // the code below, please check the ParNew version to see if
8238 // similar changes might be needed.
8239 // CR 6797058 has been filed to consolidate the common code.
8240 bool CMSCollector::par_take_from_overflow_list(size_t num,
8241 OopTaskQueue* work_q,
8242 int no_of_gc_threads) {
8243 assert(work_q->size() == 0, "First empty local work queue");
8244 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8245 if (_overflow_list == NULL) {
8246 return false;
8247 }
8248 // Grab the entire list; we'll put back a suffix
8249 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8250 Thread* tid = Thread::current();
8251 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
8252 // set to ParallelGCThreads.
8253 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
8254 size_t sleep_time_millis = MAX2((size_t)1, num/100);
8255 // If the list is busy, we spin for a short while,
8256 // sleeping between attempts to get the list.
8257 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
8258 os::sleep(tid, sleep_time_millis, false);
8259 if (_overflow_list == NULL) {
8260 // Nothing left to take
8261 return false;
8262 } else if (_overflow_list != BUSY) {
8263 // Try and grab the prefix
8264 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8265 }
8266 }
8267 // If the list was found to be empty, or we spun long
8268 // enough, we give up and return empty-handed. If we leave
8269 // the list in the BUSY state below, it must be the case that
8270 // some other thread holds the overflow list and will set it
8271 // to a non-BUSY state in the future.
8272 if (prefix == NULL || prefix == BUSY) {
8273 // Nothing to take or waited long enough
8274 if (prefix == NULL) {
8275 // Write back the NULL in case we overwrote it with BUSY above
8276 // and it is still the same value.
8277 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8278 }
8279 return false;
8280 }
8281 assert(prefix != NULL && prefix != BUSY, "Error");
8282 size_t i = num;
8283 oop cur = prefix;
8284 // Walk down the first "num" objects, unless we reach the end.
8285 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8286 if (cur->mark() == NULL) {
8287 // We have "num" or fewer elements in the list, so there
8288 // is nothing to return to the global list.
8289 // Write back the NULL in lieu of the BUSY we wrote
8290 // above, if it is still the same value.
8291 if (_overflow_list == BUSY) {
8292 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8293 }
8294 } else {
8295 // Chop off the suffix and return it to the global list.
8296 assert(cur->mark() != BUSY, "Error");
8297 oop suffix_head = cur->mark(); // suffix will be put back on global list
8298 cur->set_mark(NULL); // break off suffix
8299 // It's possible that the list is still in the empty(busy) state
8300 // we left it in a short while ago; in that case we may be
8301 // able to place back the suffix without incurring the cost
8302 // of a walk down the list.
8303 oop observed_overflow_list = _overflow_list;
8304 oop cur_overflow_list = observed_overflow_list;
8305 bool attached = false;
8306 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
8307 observed_overflow_list =
8308 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8309 if (cur_overflow_list == observed_overflow_list) {
8310 attached = true;
8311 break;
8312 } else cur_overflow_list = observed_overflow_list;
8313 }
8314 if (!attached) {
8315 // Too bad, someone else sneaked in (at least) an element; we'll need
8316 // to do a splice. Find tail of suffix so we can prepend suffix to global
8317 // list.
8318 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8319 oop suffix_tail = cur;
8320 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8321 "Tautology");
8322 observed_overflow_list = _overflow_list;
8323 do {
8324 cur_overflow_list = observed_overflow_list;
8325 if (cur_overflow_list != BUSY) {
8326 // Do the splice ...
8327 suffix_tail->set_mark(markOop(cur_overflow_list));
8328 } else { // cur_overflow_list == BUSY
8329 suffix_tail->set_mark(NULL);
8330 }
8331 // ... and try to place spliced list back on overflow_list ...
8332 observed_overflow_list =
8333 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8334 } while (cur_overflow_list != observed_overflow_list);
8335 // ... until we have succeeded in doing so.
8336 }
8337 }
8338
8339 // Push the prefix elements on work_q
8340 assert(prefix != NULL, "control point invariant");
8341 const markOop proto = markOopDesc::prototype();
8342 oop next;
8343 NOT_PRODUCT(ssize_t n = 0;)
8344 for (cur = prefix; cur != NULL; cur = next) {
8345 next = oop(cur->mark());
8346 cur->set_mark(proto); // until proven otherwise
8347 assert(cur->is_oop(), "Should be an oop");
8348 bool res = work_q->push(cur);
8349 assert(res, "Bit off more than we can chew?");
8350 NOT_PRODUCT(n++;)
8351 }
8352 #ifndef PRODUCT
8353 assert(_num_par_pushes >= n, "Too many pops?");
8354 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8355 #endif
8356 return true;
8357 }
8358
8359 // Single-threaded
8360 void CMSCollector::push_on_overflow_list(oop p) {
8361 NOT_PRODUCT(_num_par_pushes++;)
8362 assert(p->is_oop(), "Not an oop");
8363 preserve_mark_if_necessary(p);
8364 p->set_mark((markOop)_overflow_list);
8365 _overflow_list = p;
8366 }
8367
8368 // Multi-threaded; use CAS to prepend to overflow list
8369 void CMSCollector::par_push_on_overflow_list(oop p) {
8370 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8371 assert(p->is_oop(), "Not an oop");
8372 par_preserve_mark_if_necessary(p);
8373 oop observed_overflow_list = _overflow_list;
8374 oop cur_overflow_list;
8375 do {
8376 cur_overflow_list = observed_overflow_list;
8377 if (cur_overflow_list != BUSY) {
8378 p->set_mark(markOop(cur_overflow_list));
8379 } else {
8380 p->set_mark(NULL);
8381 }
8382 observed_overflow_list =
8383 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8384 } while (cur_overflow_list != observed_overflow_list);
8385 }
8386 #undef BUSY
8387
8388 // Single threaded
8389 // General Note on GrowableArray: pushes may silently fail
8390 // because we are (temporarily) out of C-heap for expanding
8391 // the stack. The problem is quite ubiquitous and affects
8392 // a lot of code in the JVM. The prudent thing for GrowableArray
8393 // to do (for now) is to exit with an error. However, that may
8394 // be too draconian in some cases because the caller may be
8395 // able to recover without much harm. For such cases, we
8396 // should probably introduce a "soft_push" method which returns
8397 // an indication of success or failure with the assumption that
8398 // the caller may be able to recover from a failure; code in
8399 // the VM can then be changed, incrementally, to deal with such
8400 // failures where possible, thus, incrementally hardening the VM
8401 // in such low resource situations.
8402 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8403 _preserved_oop_stack.push(p);
8404 _preserved_mark_stack.push(m);
8405 assert(m == p->mark(), "Mark word changed");
8406 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8407 "bijection");
8408 }
8409
8410 // Single threaded
8411 void CMSCollector::preserve_mark_if_necessary(oop p) {
8412 markOop m = p->mark();
8413 if (m->must_be_preserved(p)) {
8414 preserve_mark_work(p, m);
8415 }
8416 }
8417
8418 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8419 markOop m = p->mark();
8420 if (m->must_be_preserved(p)) {
8421 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8422 // Even though we read the mark word without holding
8423 // the lock, we are assured that it will not change
8424 // because we "own" this oop, so no other thread can
8425 // be trying to push it on the overflow list; see
8426 // the assertion in preserve_mark_work() that checks
8427 // that m == p->mark().
8428 preserve_mark_work(p, m);
8429 }
8430 }
8431
8432 // We should be able to do this multi-threaded,
8433 // a chunk of stack being a task (this is
8434 // correct because each oop only ever appears
8435 // once in the overflow list. However, it's
8436 // not very easy to completely overlap this with
8437 // other operations, so will generally not be done
8438 // until all work's been completed. Because we
8439 // expect the preserved oop stack (set) to be small,
8440 // it's probably fine to do this single-threaded.
8441 // We can explore cleverer concurrent/overlapped/parallel
8442 // processing of preserved marks if we feel the
8443 // need for this in the future. Stack overflow should
8444 // be so rare in practice and, when it happens, its
8445 // effect on performance so great that this will
8446 // likely just be in the noise anyway.
8447 void CMSCollector::restore_preserved_marks_if_any() {
8448 assert(SafepointSynchronize::is_at_safepoint(),
8449 "world should be stopped");
8450 assert(Thread::current()->is_ConcurrentGC_thread() ||
8451 Thread::current()->is_VM_thread(),
8452 "should be single-threaded");
8453 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8454 "bijection");
8455
8456 while (!_preserved_oop_stack.is_empty()) {
8457 oop p = _preserved_oop_stack.pop();
8458 assert(p->is_oop(), "Should be an oop");
8459 assert(_span.contains(p), "oop should be in _span");
8460 assert(p->mark() == markOopDesc::prototype(),
8461 "Set when taken from overflow list");
8462 markOop m = _preserved_mark_stack.pop();
8463 p->set_mark(m);
8464 }
8465 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8466 "stacks were cleared above");
8467 }
8468
8469 #ifndef PRODUCT
8470 bool CMSCollector::no_preserved_marks() const {
8471 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8472 }
8473 #endif
8474
8475 // Transfer some number of overflown objects to usual marking
8476 // stack. Return true if some objects were transferred.
8477 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8478 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8479 (size_t)ParGCDesiredObjsFromOverflowList);
8480
8481 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8482 assert(_collector->overflow_list_is_empty() || res,
8483 "If list is not empty, we should have taken something");
8484 assert(!res || !_mark_stack->isEmpty(),
8485 "If we took something, it should now be on our stack");
8486 return res;
8487 }
8488
8489 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8490 size_t res = _sp->block_size_no_stall(addr, _collector);
8491 if (_sp->block_is_obj(addr)) {
8492 if (_live_bit_map->isMarked(addr)) {
8493 // It can't have been dead in a previous cycle
8494 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8495 } else {
8496 _dead_bit_map->mark(addr); // mark the dead object
8497 }
8498 }
8499 // Could be 0, if the block size could not be computed without stalling.
8500 return res;
8501 }
8502
8503 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8504
8505 switch (phase) {
8506 case CMSCollector::InitialMarking:
8507 initialize(true /* fullGC */ ,
8508 cause /* cause of the GC */,
8509 true /* recordGCBeginTime */,
8510 true /* recordPreGCUsage */,
8511 false /* recordPeakUsage */,
8512 false /* recordPostGCusage */,
8513 true /* recordAccumulatedGCTime */,
8514 false /* recordGCEndTime */,
8515 false /* countCollection */ );
8516 break;
8517
8518 case CMSCollector::FinalMarking:
8519 initialize(true /* fullGC */ ,
8520 cause /* cause of the GC */,
8521 false /* recordGCBeginTime */,
8522 false /* recordPreGCUsage */,
8523 false /* recordPeakUsage */,
8524 false /* recordPostGCusage */,
8525 true /* recordAccumulatedGCTime */,
8526 false /* recordGCEndTime */,
8527 false /* countCollection */ );
8528 break;
8529
8530 case CMSCollector::Sweeping:
8531 initialize(true /* fullGC */ ,
8532 cause /* cause of the GC */,
8533 false /* recordGCBeginTime */,
8534 false /* recordPreGCUsage */,
8535 true /* recordPeakUsage */,
8536 true /* recordPostGCusage */,
8537 false /* recordAccumulatedGCTime */,
8538 true /* recordGCEndTime */,
8539 true /* countCollection */ );
8540 break;
8541
8542 default:
8543 ShouldNotReachHere();
8544 }
8545 }