1 /*
2 * Copyright (c) 2001, 2014, 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/tenuredGeneration.hpp"
57 #include "oops/oop.inline.hpp"
58 #include "prims/jvmtiExport.hpp"
59 #include "runtime/atomic.inline.hpp"
60 #include "runtime/globals_extension.hpp"
61 #include "runtime/handles.inline.hpp"
62 #include "runtime/java.hpp"
63 #include "runtime/orderAccess.inline.hpp"
64 #include "runtime/vmThread.hpp"
65 #include "services/memoryService.hpp"
66 #include "services/runtimeService.hpp"
67
68 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
69
70 // statics
71 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
72 bool CMSCollector::_full_gc_requested = false;
73 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
74
75 //////////////////////////////////////////////////////////////////
76 // In support of CMS/VM thread synchronization
77 //////////////////////////////////////////////////////////////////
78 // We split use of the CGC_lock into 2 "levels".
79 // The low-level locking is of the usual CGC_lock monitor. We introduce
80 // a higher level "token" (hereafter "CMS token") built on top of the
81 // low level monitor (hereafter "CGC lock").
82 // The token-passing protocol gives priority to the VM thread. The
83 // CMS-lock doesn't provide any fairness guarantees, but clients
84 // should ensure that it is only held for very short, bounded
85 // durations.
86 //
87 // When either of the CMS thread or the VM thread is involved in
88 // collection operations during which it does not want the other
89 // thread to interfere, it obtains the CMS token.
90 //
91 // If either thread tries to get the token while the other has
92 // it, that thread waits. However, if the VM thread and CMS thread
93 // both want the token, then the VM thread gets priority while the
94 // CMS thread waits. This ensures, for instance, that the "concurrent"
95 // phases of the CMS thread's work do not block out the VM thread
96 // for long periods of time as the CMS thread continues to hog
97 // the token. (See bug 4616232).
98 //
99 // The baton-passing functions are, however, controlled by the
100 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
101 // and here the low-level CMS lock, not the high level token,
102 // ensures mutual exclusion.
103 //
104 // Two important conditions that we have to satisfy:
105 // 1. if a thread does a low-level wait on the CMS lock, then it
106 // relinquishes the CMS token if it were holding that token
107 // when it acquired the low-level CMS lock.
108 // 2. any low-level notifications on the low-level lock
109 // should only be sent when a thread has relinquished the token.
110 //
111 // In the absence of either property, we'd have potential deadlock.
112 //
113 // We protect each of the CMS (concurrent and sequential) phases
114 // with the CMS _token_, not the CMS _lock_.
115 //
116 // The only code protected by CMS lock is the token acquisition code
117 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
118 // baton-passing code.
119 //
120 // Unfortunately, i couldn't come up with a good abstraction to factor and
121 // hide the naked CGC_lock manipulation in the baton-passing code
122 // further below. That's something we should try to do. Also, the proof
123 // of correctness of this 2-level locking scheme is far from obvious,
124 // and potentially quite slippery. We have an uneasy suspicion, for instance,
125 // that there may be a theoretical possibility of delay/starvation in the
126 // low-level lock/wait/notify scheme used for the baton-passing because of
127 // potential interference with the priority scheme embodied in the
128 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
129 // invocation further below and marked with "XXX 20011219YSR".
130 // Indeed, as we note elsewhere, this may become yet more slippery
131 // in the presence of multiple CMS and/or multiple VM threads. XXX
132
133 class CMSTokenSync: public StackObj {
134 private:
135 bool _is_cms_thread;
136 public:
137 CMSTokenSync(bool is_cms_thread):
138 _is_cms_thread(is_cms_thread) {
139 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
140 "Incorrect argument to constructor");
141 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
142 }
143
144 ~CMSTokenSync() {
145 assert(_is_cms_thread ?
146 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
147 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
148 "Incorrect state");
149 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
150 }
151 };
152
153 // Convenience class that does a CMSTokenSync, and then acquires
154 // upto three locks.
155 class CMSTokenSyncWithLocks: public CMSTokenSync {
156 private:
157 // Note: locks are acquired in textual declaration order
158 // and released in the opposite order
159 MutexLockerEx _locker1, _locker2, _locker3;
160 public:
161 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
162 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
163 CMSTokenSync(is_cms_thread),
164 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
165 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
166 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
167 { }
168 };
169
170
171 //////////////////////////////////////////////////////////////////
172 // Concurrent Mark-Sweep Generation /////////////////////////////
173 //////////////////////////////////////////////////////////////////
174
175 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
176
177 // This struct contains per-thread things necessary to support parallel
178 // young-gen collection.
179 class CMSParGCThreadState: public CHeapObj<mtGC> {
180 public:
181 CFLS_LAB lab;
182 PromotionInfo promo;
183
184 // Constructor.
185 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
186 promo.setSpace(cfls);
187 }
188 };
189
190 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
191 ReservedSpace rs, size_t initial_byte_size, int level,
192 CardTableRS* ct, bool use_adaptive_freelists,
193 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) :
194 CardGeneration(rs, initial_byte_size, level, ct),
195 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
196 _did_compact(false)
197 {
198 HeapWord* bottom = (HeapWord*) _virtual_space.low();
199 HeapWord* end = (HeapWord*) _virtual_space.high();
200
201 _direct_allocated_words = 0;
202 NOT_PRODUCT(
203 _numObjectsPromoted = 0;
204 _numWordsPromoted = 0;
205 _numObjectsAllocated = 0;
206 _numWordsAllocated = 0;
207 )
208
209 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
210 use_adaptive_freelists,
211 dictionaryChoice);
212 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
213 if (_cmsSpace == NULL) {
214 vm_exit_during_initialization(
215 "CompactibleFreeListSpace allocation failure");
216 }
217 _cmsSpace->_gen = this;
218
219 _gc_stats = new CMSGCStats();
220
221 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
222 // offsets match. The ability to tell free chunks from objects
223 // depends on this property.
224 debug_only(
225 FreeChunk* junk = NULL;
226 assert(UseCompressedClassPointers ||
227 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
228 "Offset of FreeChunk::_prev within FreeChunk must match"
229 " that of OopDesc::_klass within OopDesc");
230 )
231 if (CollectedHeap::use_parallel_gc_threads()) {
232 typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
233 _par_gc_thread_states =
234 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC);
235 if (_par_gc_thread_states == NULL) {
236 vm_exit_during_initialization("Could not allocate par gc structs");
237 }
238 for (uint i = 0; i < ParallelGCThreads; i++) {
239 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
240 if (_par_gc_thread_states[i] == NULL) {
241 vm_exit_during_initialization("Could not allocate par gc structs");
242 }
243 }
244 } else {
245 _par_gc_thread_states = NULL;
246 }
247 _incremental_collection_failed = false;
248 // The "dilatation_factor" is the expansion that can occur on
249 // account of the fact that the minimum object size in the CMS
250 // generation may be larger than that in, say, a contiguous young
251 // generation.
252 // Ideally, in the calculation below, we'd compute the dilatation
253 // factor as: MinChunkSize/(promoting_gen's min object size)
254 // Since we do not have such a general query interface for the
255 // promoting generation, we'll instead just use the minimum
256 // object size (which today is a header's worth of space);
257 // note that all arithmetic is in units of HeapWords.
258 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
259 assert(_dilatation_factor >= 1.0, "from previous assert");
260 }
261
262
263 // The field "_initiating_occupancy" represents the occupancy percentage
264 // at which we trigger a new collection cycle. Unless explicitly specified
265 // via CMSInitiatingOccupancyFraction (argument "io" below), it
266 // is calculated by:
267 //
268 // Let "f" be MinHeapFreeRatio in
269 //
270 // _initiating_occupancy = 100-f +
271 // f * (CMSTriggerRatio/100)
272 // where CMSTriggerRatio is the argument "tr" below.
273 //
274 // That is, if we assume the heap is at its desired maximum occupancy at the
275 // end of a collection, we let CMSTriggerRatio of the (purported) free
276 // space be allocated before initiating a new collection cycle.
277 //
278 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
279 assert(io <= 100 && tr <= 100, "Check the arguments");
280 if (io >= 0) {
281 _initiating_occupancy = (double)io / 100.0;
282 } else {
283 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
284 (double)(tr * MinHeapFreeRatio) / 100.0)
285 / 100.0;
286 }
287 }
288
289 void ConcurrentMarkSweepGeneration::ref_processor_init() {
290 assert(collector() != NULL, "no collector");
291 collector()->ref_processor_init();
292 }
293
294 void CMSCollector::ref_processor_init() {
295 if (_ref_processor == NULL) {
296 // Allocate and initialize a reference processor
297 _ref_processor =
298 new ReferenceProcessor(_span, // span
299 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
300 (int) ParallelGCThreads, // mt processing degree
301 _cmsGen->refs_discovery_is_mt(), // mt discovery
302 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
303 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
304 &_is_alive_closure); // closure for liveness info
305 // Initialize the _ref_processor field of CMSGen
306 _cmsGen->set_ref_processor(_ref_processor);
307
308 }
309 }
310
311 AdaptiveSizePolicy* CMSCollector::size_policy() {
312 GenCollectedHeap* gch = GenCollectedHeap::heap();
313 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
314 "Wrong type of heap");
315 return gch->gen_policy()->size_policy();
316 }
317
318 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
319
320 const char* gen_name = "old";
321 GenCollectorPolicy* gcp = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy();
322
323 // Generation Counters - generation 1, 1 subspace
324 _gen_counters = new GenerationCounters(gen_name, 1, 1,
325 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
326
327 _space_counters = new GSpaceCounters(gen_name, 0,
328 _virtual_space.reserved_size(),
329 this, _gen_counters);
330 }
331
332 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
333 _cms_gen(cms_gen)
334 {
335 assert(alpha <= 100, "bad value");
336 _saved_alpha = alpha;
337
338 // Initialize the alphas to the bootstrap value of 100.
339 _gc0_alpha = _cms_alpha = 100;
340
341 _cms_begin_time.update();
342 _cms_end_time.update();
343
344 _gc0_duration = 0.0;
345 _gc0_period = 0.0;
346 _gc0_promoted = 0;
347
348 _cms_duration = 0.0;
349 _cms_period = 0.0;
350 _cms_allocated = 0;
351
352 _cms_used_at_gc0_begin = 0;
353 _cms_used_at_gc0_end = 0;
354 _allow_duty_cycle_reduction = false;
355 _valid_bits = 0;
356 }
357
358 double CMSStats::cms_free_adjustment_factor(size_t free) const {
359 // TBD: CR 6909490
360 return 1.0;
361 }
362
363 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
364 }
365
366 // If promotion failure handling is on use
367 // the padded average size of the promotion for each
368 // young generation collection.
369 double CMSStats::time_until_cms_gen_full() const {
370 size_t cms_free = _cms_gen->cmsSpace()->free();
371 GenCollectedHeap* gch = GenCollectedHeap::heap();
372 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(),
373 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
374 if (cms_free > expected_promotion) {
375 // Start a cms collection if there isn't enough space to promote
376 // for the next minor collection. Use the padded average as
377 // a safety factor.
378 cms_free -= expected_promotion;
379
380 // Adjust by the safety factor.
381 double cms_free_dbl = (double)cms_free;
382 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
383 // Apply a further correction factor which tries to adjust
384 // for recent occurance of concurrent mode failures.
385 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
386 cms_free_dbl = cms_free_dbl * cms_adjustment;
387
388 if (PrintGCDetails && Verbose) {
389 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
390 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
391 cms_free, expected_promotion);
392 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
393 cms_free_dbl, cms_consumption_rate() + 1.0);
394 }
395 // Add 1 in case the consumption rate goes to zero.
396 return cms_free_dbl / (cms_consumption_rate() + 1.0);
397 }
398 return 0.0;
399 }
400
401 // Compare the duration of the cms collection to the
402 // time remaining before the cms generation is empty.
403 // Note that the time from the start of the cms collection
404 // to the start of the cms sweep (less than the total
405 // duration of the cms collection) can be used. This
406 // has been tried and some applications experienced
407 // promotion failures early in execution. This was
408 // possibly because the averages were not accurate
409 // enough at the beginning.
410 double CMSStats::time_until_cms_start() const {
411 // We add "gc0_period" to the "work" calculation
412 // below because this query is done (mostly) at the
413 // end of a scavenge, so we need to conservatively
414 // account for that much possible delay
415 // in the query so as to avoid concurrent mode failures
416 // due to starting the collection just a wee bit too
417 // late.
418 double work = cms_duration() + gc0_period();
419 double deadline = time_until_cms_gen_full();
420 // If a concurrent mode failure occurred recently, we want to be
421 // more conservative and halve our expected time_until_cms_gen_full()
422 if (work > deadline) {
423 if (Verbose && PrintGCDetails) {
424 gclog_or_tty->print(
425 " CMSCollector: collect because of anticipated promotion "
426 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
427 gc0_period(), time_until_cms_gen_full());
428 }
429 return 0.0;
430 }
431 return work - deadline;
432 }
433
434 #ifndef PRODUCT
435 void CMSStats::print_on(outputStream *st) const {
436 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
437 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
438 gc0_duration(), gc0_period(), gc0_promoted());
439 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
440 cms_duration(), cms_period(), cms_allocated());
441 st->print(",cms_since_beg=%g,cms_since_end=%g",
442 cms_time_since_begin(), cms_time_since_end());
443 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
444 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
445
446 if (valid()) {
447 st->print(",promo_rate=%g,cms_alloc_rate=%g",
448 promotion_rate(), cms_allocation_rate());
449 st->print(",cms_consumption_rate=%g,time_until_full=%g",
450 cms_consumption_rate(), time_until_cms_gen_full());
451 }
452 st->print(" ");
453 }
454 #endif // #ifndef PRODUCT
455
456 CMSCollector::CollectorState CMSCollector::_collectorState =
457 CMSCollector::Idling;
458 bool CMSCollector::_foregroundGCIsActive = false;
459 bool CMSCollector::_foregroundGCShouldWait = false;
460
461 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
462 CardTableRS* ct,
463 ConcurrentMarkSweepPolicy* cp):
464 _cmsGen(cmsGen),
465 _ct(ct),
466 _ref_processor(NULL), // will be set later
467 _conc_workers(NULL), // may be set later
468 _abort_preclean(false),
469 _start_sampling(false),
470 _between_prologue_and_epilogue(false),
471 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
472 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
473 -1 /* lock-free */, "No_lock" /* dummy */),
474 _modUnionClosure(&_modUnionTable),
475 _modUnionClosurePar(&_modUnionTable),
476 // Adjust my span to cover old (cms) gen
477 _span(cmsGen->reserved()),
478 // Construct the is_alive_closure with _span & markBitMap
479 _is_alive_closure(_span, &_markBitMap),
480 _restart_addr(NULL),
481 _overflow_list(NULL),
482 _stats(cmsGen),
483 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
484 //verify that this lock should be acquired with safepoint check.
485 Monitor::_safepoint_check_sometimes)),
486 _eden_chunk_array(NULL), // may be set in ctor body
487 _eden_chunk_capacity(0), // -- ditto --
488 _eden_chunk_index(0), // -- ditto --
489 _survivor_plab_array(NULL), // -- ditto --
490 _survivor_chunk_array(NULL), // -- ditto --
491 _survivor_chunk_capacity(0), // -- ditto --
492 _survivor_chunk_index(0), // -- ditto --
493 _ser_pmc_preclean_ovflw(0),
494 _ser_kac_preclean_ovflw(0),
495 _ser_pmc_remark_ovflw(0),
496 _par_pmc_remark_ovflw(0),
497 _ser_kac_ovflw(0),
498 _par_kac_ovflw(0),
499 #ifndef PRODUCT
500 _num_par_pushes(0),
501 #endif
502 _collection_count_start(0),
503 _verifying(false),
504 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
505 _completed_initialization(false),
506 _collector_policy(cp),
507 _should_unload_classes(CMSClassUnloadingEnabled),
508 _concurrent_cycles_since_last_unload(0),
509 _roots_scanning_options(SharedHeap::SO_None),
510 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
511 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
512 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
513 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
514 _cms_start_registered(false)
515 {
516 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
517 ExplicitGCInvokesConcurrent = true;
518 }
519 // Now expand the span and allocate the collection support structures
520 // (MUT, marking bit map etc.) to cover both generations subject to
521 // collection.
522
523 // For use by dirty card to oop closures.
524 _cmsGen->cmsSpace()->set_collector(this);
525
526 // Allocate MUT and marking bit map
527 {
528 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
529 if (!_markBitMap.allocate(_span)) {
530 warning("Failed to allocate CMS Bit Map");
531 return;
532 }
533 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
534 }
535 {
536 _modUnionTable.allocate(_span);
537 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
538 }
539
540 if (!_markStack.allocate(MarkStackSize)) {
541 warning("Failed to allocate CMS Marking Stack");
542 return;
543 }
544
545 // Support for multi-threaded concurrent phases
546 if (CMSConcurrentMTEnabled) {
547 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
548 // just for now
549 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
550 }
551 if (ConcGCThreads > 1) {
552 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
553 ConcGCThreads, true);
554 if (_conc_workers == NULL) {
555 warning("GC/CMS: _conc_workers allocation failure: "
556 "forcing -CMSConcurrentMTEnabled");
557 CMSConcurrentMTEnabled = false;
558 } else {
559 _conc_workers->initialize_workers();
560 }
561 } else {
562 CMSConcurrentMTEnabled = false;
563 }
564 }
565 if (!CMSConcurrentMTEnabled) {
566 ConcGCThreads = 0;
567 } else {
568 // Turn off CMSCleanOnEnter optimization temporarily for
569 // the MT case where it's not fixed yet; see 6178663.
570 CMSCleanOnEnter = false;
571 }
572 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
573 "Inconsistency");
574
575 // Parallel task queues; these are shared for the
576 // concurrent and stop-world phases of CMS, but
577 // are not shared with parallel scavenge (ParNew).
578 {
579 uint i;
580 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
581
582 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
583 || ParallelRefProcEnabled)
584 && num_queues > 0) {
585 _task_queues = new OopTaskQueueSet(num_queues);
586 if (_task_queues == NULL) {
587 warning("task_queues allocation failure.");
588 return;
589 }
590 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
591 if (_hash_seed == NULL) {
592 warning("_hash_seed array allocation failure");
593 return;
594 }
595
596 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
597 for (i = 0; i < num_queues; i++) {
598 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
599 if (q == NULL) {
600 warning("work_queue allocation failure.");
601 return;
602 }
603 _task_queues->register_queue(i, q);
604 }
605 for (i = 0; i < num_queues; i++) {
606 _task_queues->queue(i)->initialize();
607 _hash_seed[i] = 17; // copied from ParNew
608 }
609 }
610 }
611
612 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
613
614 // Clip CMSBootstrapOccupancy between 0 and 100.
615 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100;
616
617 // Now tell CMS generations the identity of their collector
618 ConcurrentMarkSweepGeneration::set_collector(this);
619
620 // Create & start a CMS thread for this CMS collector
621 _cmsThread = ConcurrentMarkSweepThread::start(this);
622 assert(cmsThread() != NULL, "CMS Thread should have been created");
623 assert(cmsThread()->collector() == this,
624 "CMS Thread should refer to this gen");
625 assert(CGC_lock != NULL, "Where's the CGC_lock?");
626
627 // Support for parallelizing young gen rescan
628 GenCollectedHeap* gch = GenCollectedHeap::heap();
629 assert(gch->prev_gen(_cmsGen)->kind() == Generation::ParNew, "CMS can only be used with ParNew");
630 _young_gen = (ParNewGeneration*)gch->prev_gen(_cmsGen);
631 if (gch->supports_inline_contig_alloc()) {
632 _top_addr = gch->top_addr();
633 _end_addr = gch->end_addr();
634 assert(_young_gen != NULL, "no _young_gen");
635 _eden_chunk_index = 0;
636 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
637 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
638 if (_eden_chunk_array == NULL) {
639 _eden_chunk_capacity = 0;
640 warning("GC/CMS: _eden_chunk_array allocation failure");
641 }
642 }
643 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
644
645 // Support for parallelizing survivor space rescan
646 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
647 const size_t max_plab_samples =
648 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize;
649
650 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
651 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC);
652 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
653 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
654 || _cursor == NULL) {
655 warning("Failed to allocate survivor plab/chunk array");
656 if (_survivor_plab_array != NULL) {
657 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
658 _survivor_plab_array = NULL;
659 }
660 if (_survivor_chunk_array != NULL) {
661 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
662 _survivor_chunk_array = NULL;
663 }
664 if (_cursor != NULL) {
665 FREE_C_HEAP_ARRAY(size_t, _cursor);
666 _cursor = NULL;
667 }
668 } else {
669 _survivor_chunk_capacity = 2*max_plab_samples;
670 for (uint i = 0; i < ParallelGCThreads; i++) {
671 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
672 if (vec == NULL) {
673 warning("Failed to allocate survivor plab array");
674 for (int j = i; j > 0; j--) {
675 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array());
676 }
677 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
678 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
679 _survivor_plab_array = NULL;
680 _survivor_chunk_array = NULL;
681 _survivor_chunk_capacity = 0;
682 break;
683 } else {
684 ChunkArray* cur =
685 ::new (&_survivor_plab_array[i]) ChunkArray(vec,
686 max_plab_samples);
687 assert(cur->end() == 0, "Should be 0");
688 assert(cur->array() == vec, "Should be vec");
689 assert(cur->capacity() == max_plab_samples, "Error");
690 }
691 }
692 }
693 }
694 assert( ( _survivor_plab_array != NULL
695 && _survivor_chunk_array != NULL)
696 || ( _survivor_chunk_capacity == 0
697 && _survivor_chunk_index == 0),
698 "Error");
699
700 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
701 _gc_counters = new CollectorCounters("CMS", 1);
702 _completed_initialization = true;
703 _inter_sweep_timer.start(); // start of time
704 }
705
706 const char* ConcurrentMarkSweepGeneration::name() const {
707 return "concurrent mark-sweep generation";
708 }
709 void ConcurrentMarkSweepGeneration::update_counters() {
710 if (UsePerfData) {
711 _space_counters->update_all();
712 _gen_counters->update_all();
713 }
714 }
715
716 // this is an optimized version of update_counters(). it takes the
717 // used value as a parameter rather than computing it.
718 //
719 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
720 if (UsePerfData) {
721 _space_counters->update_used(used);
722 _space_counters->update_capacity();
723 _gen_counters->update_all();
724 }
725 }
726
727 void ConcurrentMarkSweepGeneration::print() const {
728 Generation::print();
729 cmsSpace()->print();
730 }
731
732 #ifndef PRODUCT
733 void ConcurrentMarkSweepGeneration::print_statistics() {
734 cmsSpace()->printFLCensus(0);
735 }
736 #endif
737
738 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
739 GenCollectedHeap* gch = GenCollectedHeap::heap();
740 if (PrintGCDetails) {
741 if (Verbose) {
742 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
743 level(), short_name(), s, used(), capacity());
744 } else {
745 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
746 level(), short_name(), s, used() / K, capacity() / K);
747 }
748 }
749 if (Verbose) {
750 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
751 gch->used(), gch->capacity());
752 } else {
753 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
754 gch->used() / K, gch->capacity() / K);
755 }
756 }
757
758 size_t
759 ConcurrentMarkSweepGeneration::contiguous_available() const {
760 // dld proposes an improvement in precision here. If the committed
761 // part of the space ends in a free block we should add that to
762 // uncommitted size in the calculation below. Will make this
763 // change later, staying with the approximation below for the
764 // time being. -- ysr.
765 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
766 }
767
768 size_t
769 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
770 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
771 }
772
773 size_t ConcurrentMarkSweepGeneration::max_available() const {
774 return free() + _virtual_space.uncommitted_size();
775 }
776
777 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
778 size_t available = max_available();
779 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
780 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
781 if (Verbose && PrintGCDetails) {
782 gclog_or_tty->print_cr(
783 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
784 "max_promo("SIZE_FORMAT")",
785 res? "":" not", available, res? ">=":"<",
786 av_promo, max_promotion_in_bytes);
787 }
788 return res;
789 }
790
791 // At a promotion failure dump information on block layout in heap
792 // (cms old generation).
793 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
794 if (CMSDumpAtPromotionFailure) {
795 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
796 }
797 }
798
799 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
800 // Clear the promotion information. These pointers can be adjusted
801 // along with all the other pointers into the heap but
802 // compaction is expected to be a rare event with
803 // a heap using cms so don't do it without seeing the need.
804 for (uint i = 0; i < ParallelGCThreads; i++) {
805 _par_gc_thread_states[i]->promo.reset();
806 }
807 }
808
809 void ConcurrentMarkSweepGeneration::compute_new_size() {
810 assert_locked_or_safepoint(Heap_lock);
811
812 // If incremental collection failed, we just want to expand
813 // to the limit.
814 if (incremental_collection_failed()) {
815 clear_incremental_collection_failed();
816 grow_to_reserved();
817 return;
818 }
819
820 // The heap has been compacted but not reset yet.
821 // Any metric such as free() or used() will be incorrect.
822
823 CardGeneration::compute_new_size();
824
825 // Reset again after a possible resizing
826 if (did_compact()) {
827 cmsSpace()->reset_after_compaction();
828 }
829 }
830
831 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
832 assert_locked_or_safepoint(Heap_lock);
833
834 // If incremental collection failed, we just want to expand
835 // to the limit.
836 if (incremental_collection_failed()) {
837 clear_incremental_collection_failed();
838 grow_to_reserved();
839 return;
840 }
841
842 double free_percentage = ((double) free()) / capacity();
843 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
844 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
845
846 // compute expansion delta needed for reaching desired free percentage
847 if (free_percentage < desired_free_percentage) {
848 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
849 assert(desired_capacity >= capacity(), "invalid expansion size");
850 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
851 if (PrintGCDetails && Verbose) {
852 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
853 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
854 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
855 gclog_or_tty->print_cr(" Desired free fraction %f",
856 desired_free_percentage);
857 gclog_or_tty->print_cr(" Maximum free fraction %f",
858 maximum_free_percentage);
859 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000);
860 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
861 desired_capacity/1000);
862 int prev_level = level() - 1;
863 if (prev_level >= 0) {
864 size_t prev_size = 0;
865 GenCollectedHeap* gch = GenCollectedHeap::heap();
866 Generation* prev_gen = gch->get_gen(prev_level);
867 prev_size = prev_gen->capacity();
868 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
869 prev_size/1000);
870 }
871 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
872 unsafe_max_alloc_nogc()/1000);
873 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
874 contiguous_available()/1000);
875 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
876 expand_bytes);
877 }
878 // safe if expansion fails
879 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
880 if (PrintGCDetails && Verbose) {
881 gclog_or_tty->print_cr(" Expanded free fraction %f",
882 ((double) free()) / capacity());
883 }
884 } else {
885 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
886 assert(desired_capacity <= capacity(), "invalid expansion size");
887 size_t shrink_bytes = capacity() - desired_capacity;
888 // Don't shrink unless the delta is greater than the minimum shrink we want
889 if (shrink_bytes >= MinHeapDeltaBytes) {
890 shrink_free_list_by(shrink_bytes);
891 }
892 }
893 }
894
895 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
896 return cmsSpace()->freelistLock();
897 }
898
899 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
900 bool tlab) {
901 CMSSynchronousYieldRequest yr;
902 MutexLockerEx x(freelistLock(),
903 Mutex::_no_safepoint_check_flag);
904 return have_lock_and_allocate(size, tlab);
905 }
906
907 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
908 bool tlab /* ignored */) {
909 assert_lock_strong(freelistLock());
910 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
911 HeapWord* res = cmsSpace()->allocate(adjustedSize);
912 // Allocate the object live (grey) if the background collector has
913 // started marking. This is necessary because the marker may
914 // have passed this address and consequently this object will
915 // not otherwise be greyed and would be incorrectly swept up.
916 // Note that if this object contains references, the writing
917 // of those references will dirty the card containing this object
918 // allowing the object to be blackened (and its references scanned)
919 // either during a preclean phase or at the final checkpoint.
920 if (res != NULL) {
921 // We may block here with an uninitialized object with
922 // its mark-bit or P-bits not yet set. Such objects need
923 // to be safely navigable by block_start().
924 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
925 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
926 collector()->direct_allocated(res, adjustedSize);
927 _direct_allocated_words += adjustedSize;
928 // allocation counters
929 NOT_PRODUCT(
930 _numObjectsAllocated++;
931 _numWordsAllocated += (int)adjustedSize;
932 )
933 }
934 return res;
935 }
936
937 // In the case of direct allocation by mutators in a generation that
938 // is being concurrently collected, the object must be allocated
939 // live (grey) if the background collector has started marking.
940 // This is necessary because the marker may
941 // have passed this address and consequently this object will
942 // not otherwise be greyed and would be incorrectly swept up.
943 // Note that if this object contains references, the writing
944 // of those references will dirty the card containing this object
945 // allowing the object to be blackened (and its references scanned)
946 // either during a preclean phase or at the final checkpoint.
947 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
948 assert(_markBitMap.covers(start, size), "Out of bounds");
949 if (_collectorState >= Marking) {
950 MutexLockerEx y(_markBitMap.lock(),
951 Mutex::_no_safepoint_check_flag);
952 // [see comments preceding SweepClosure::do_blk() below for details]
953 //
954 // Can the P-bits be deleted now? JJJ
955 //
956 // 1. need to mark the object as live so it isn't collected
957 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
958 // 3. need to mark the end of the object so marking, precleaning or sweeping
959 // can skip over uninitialized or unparsable objects. An allocated
960 // object is considered uninitialized for our purposes as long as
961 // its klass word is NULL. All old gen objects are parsable
962 // as soon as they are initialized.)
963 _markBitMap.mark(start); // object is live
964 _markBitMap.mark(start + 1); // object is potentially uninitialized?
965 _markBitMap.mark(start + size - 1);
966 // mark end of object
967 }
968 // check that oop looks uninitialized
969 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
970 }
971
972 void CMSCollector::promoted(bool par, HeapWord* start,
973 bool is_obj_array, size_t obj_size) {
974 assert(_markBitMap.covers(start), "Out of bounds");
975 // See comment in direct_allocated() about when objects should
976 // be allocated live.
977 if (_collectorState >= Marking) {
978 // we already hold the marking bit map lock, taken in
979 // the prologue
980 if (par) {
981 _markBitMap.par_mark(start);
982 } else {
983 _markBitMap.mark(start);
984 }
985 // We don't need to mark the object as uninitialized (as
986 // in direct_allocated above) because this is being done with the
987 // world stopped and the object will be initialized by the
988 // time the marking, precleaning or sweeping get to look at it.
989 // But see the code for copying objects into the CMS generation,
990 // where we need to ensure that concurrent readers of the
991 // block offset table are able to safely navigate a block that
992 // is in flux from being free to being allocated (and in
993 // transition while being copied into) and subsequently
994 // becoming a bona-fide object when the copy/promotion is complete.
995 assert(SafepointSynchronize::is_at_safepoint(),
996 "expect promotion only at safepoints");
997
998 if (_collectorState < Sweeping) {
999 // Mark the appropriate cards in the modUnionTable, so that
1000 // this object gets scanned before the sweep. If this is
1001 // not done, CMS generation references in the object might
1002 // not get marked.
1003 // For the case of arrays, which are otherwise precisely
1004 // marked, we need to dirty the entire array, not just its head.
1005 if (is_obj_array) {
1006 // The [par_]mark_range() method expects mr.end() below to
1007 // be aligned to the granularity of a bit's representation
1008 // in the heap. In the case of the MUT below, that's a
1009 // card size.
1010 MemRegion mr(start,
1011 (HeapWord*)round_to((intptr_t)(start + obj_size),
1012 CardTableModRefBS::card_size /* bytes */));
1013 if (par) {
1014 _modUnionTable.par_mark_range(mr);
1015 } else {
1016 _modUnionTable.mark_range(mr);
1017 }
1018 } else { // not an obj array; we can just mark the head
1019 if (par) {
1020 _modUnionTable.par_mark(start);
1021 } else {
1022 _modUnionTable.mark(start);
1023 }
1024 }
1025 }
1026 }
1027 }
1028
1029 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
1030 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1031 // allocate, copy and if necessary update promoinfo --
1032 // delegate to underlying space.
1033 assert_lock_strong(freelistLock());
1034
1035 #ifndef PRODUCT
1036 if (Universe::heap()->promotion_should_fail()) {
1037 return NULL;
1038 }
1039 #endif // #ifndef PRODUCT
1040
1041 oop res = _cmsSpace->promote(obj, obj_size);
1042 if (res == NULL) {
1043 // expand and retry
1044 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
1045 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
1046 // Since there's currently no next generation, we don't try to promote
1047 // into a more senior generation.
1048 assert(next_gen() == NULL, "assumption, based upon which no attempt "
1049 "is made to pass on a possibly failing "
1050 "promotion to next generation");
1051 res = _cmsSpace->promote(obj, obj_size);
1052 }
1053 if (res != NULL) {
1054 // See comment in allocate() about when objects should
1055 // be allocated live.
1056 assert(obj->is_oop(), "Will dereference klass pointer below");
1057 collector()->promoted(false, // Not parallel
1058 (HeapWord*)res, obj->is_objArray(), obj_size);
1059 // promotion counters
1060 NOT_PRODUCT(
1061 _numObjectsPromoted++;
1062 _numWordsPromoted +=
1063 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1064 )
1065 }
1066 return res;
1067 }
1068
1069
1070 // IMPORTANT: Notes on object size recognition in CMS.
1071 // ---------------------------------------------------
1072 // A block of storage in the CMS generation is always in
1073 // one of three states. A free block (FREE), an allocated
1074 // object (OBJECT) whose size() method reports the correct size,
1075 // and an intermediate state (TRANSIENT) in which its size cannot
1076 // be accurately determined.
1077 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1078 // -----------------------------------------------------
1079 // FREE: klass_word & 1 == 1; mark_word holds block size
1080 //
1081 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1082 // obj->size() computes correct size
1083 //
1084 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1085 //
1086 // STATE IDENTIFICATION: (64 bit+COOPS)
1087 // ------------------------------------
1088 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1089 //
1090 // OBJECT: klass_word installed; klass_word != 0;
1091 // obj->size() computes correct size
1092 //
1093 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1094 //
1095 //
1096 // STATE TRANSITION DIAGRAM
1097 //
1098 // mut / parnew mut / parnew
1099 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1100 // ^ |
1101 // |------------------------ DEAD <------------------------------------|
1102 // sweep mut
1103 //
1104 // While a block is in TRANSIENT state its size cannot be determined
1105 // so readers will either need to come back later or stall until
1106 // the size can be determined. Note that for the case of direct
1107 // allocation, P-bits, when available, may be used to determine the
1108 // size of an object that may not yet have been initialized.
1109
1110 // Things to support parallel young-gen collection.
1111 oop
1112 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1113 oop old, markOop m,
1114 size_t word_sz) {
1115 #ifndef PRODUCT
1116 if (Universe::heap()->promotion_should_fail()) {
1117 return NULL;
1118 }
1119 #endif // #ifndef PRODUCT
1120
1121 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1122 PromotionInfo* promoInfo = &ps->promo;
1123 // if we are tracking promotions, then first ensure space for
1124 // promotion (including spooling space for saving header if necessary).
1125 // then allocate and copy, then track promoted info if needed.
1126 // When tracking (see PromotionInfo::track()), the mark word may
1127 // be displaced and in this case restoration of the mark word
1128 // occurs in the (oop_since_save_marks_)iterate phase.
1129 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1130 // Out of space for allocating spooling buffers;
1131 // try expanding and allocating spooling buffers.
1132 if (!expand_and_ensure_spooling_space(promoInfo)) {
1133 return NULL;
1134 }
1135 }
1136 assert(promoInfo->has_spooling_space(), "Control point invariant");
1137 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1138 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1139 if (obj_ptr == NULL) {
1140 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1141 if (obj_ptr == NULL) {
1142 return NULL;
1143 }
1144 }
1145 oop obj = oop(obj_ptr);
1146 OrderAccess::storestore();
1147 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1148 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1149 // IMPORTANT: See note on object initialization for CMS above.
1150 // Otherwise, copy the object. Here we must be careful to insert the
1151 // klass pointer last, since this marks the block as an allocated object.
1152 // Except with compressed oops it's the mark word.
1153 HeapWord* old_ptr = (HeapWord*)old;
1154 // Restore the mark word copied above.
1155 obj->set_mark(m);
1156 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1157 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1158 OrderAccess::storestore();
1159
1160 if (UseCompressedClassPointers) {
1161 // Copy gap missed by (aligned) header size calculation below
1162 obj->set_klass_gap(old->klass_gap());
1163 }
1164 if (word_sz > (size_t)oopDesc::header_size()) {
1165 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1166 obj_ptr + oopDesc::header_size(),
1167 word_sz - oopDesc::header_size());
1168 }
1169
1170 // Now we can track the promoted object, if necessary. We take care
1171 // to delay the transition from uninitialized to full object
1172 // (i.e., insertion of klass pointer) until after, so that it
1173 // atomically becomes a promoted object.
1174 if (promoInfo->tracking()) {
1175 promoInfo->track((PromotedObject*)obj, old->klass());
1176 }
1177 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1178 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1179 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1180
1181 // Finally, install the klass pointer (this should be volatile).
1182 OrderAccess::storestore();
1183 obj->set_klass(old->klass());
1184 // We should now be able to calculate the right size for this object
1185 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1186
1187 collector()->promoted(true, // parallel
1188 obj_ptr, old->is_objArray(), word_sz);
1189
1190 NOT_PRODUCT(
1191 Atomic::inc_ptr(&_numObjectsPromoted);
1192 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1193 )
1194
1195 return obj;
1196 }
1197
1198 void
1199 ConcurrentMarkSweepGeneration::
1200 par_promote_alloc_done(int thread_num) {
1201 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1202 ps->lab.retire(thread_num);
1203 }
1204
1205 void
1206 ConcurrentMarkSweepGeneration::
1207 par_oop_since_save_marks_iterate_done(int thread_num) {
1208 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1209 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1210 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1211 }
1212
1213 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1214 size_t size,
1215 bool tlab)
1216 {
1217 // We allow a STW collection only if a full
1218 // collection was requested.
1219 return full || should_allocate(size, tlab); // FIX ME !!!
1220 // This and promotion failure handling are connected at the
1221 // hip and should be fixed by untying them.
1222 }
1223
1224 bool CMSCollector::shouldConcurrentCollect() {
1225 if (_full_gc_requested) {
1226 if (Verbose && PrintGCDetails) {
1227 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1228 " gc request (or gc_locker)");
1229 }
1230 return true;
1231 }
1232
1233 FreelistLocker x(this);
1234 // ------------------------------------------------------------------
1235 // Print out lots of information which affects the initiation of
1236 // a collection.
1237 if (PrintCMSInitiationStatistics && stats().valid()) {
1238 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1239 gclog_or_tty->stamp();
1240 gclog_or_tty->cr();
1241 stats().print_on(gclog_or_tty);
1242 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1243 stats().time_until_cms_gen_full());
1244 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1245 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1246 _cmsGen->contiguous_available());
1247 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1248 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1249 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1250 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1251 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1252 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1253 gclog_or_tty->print_cr("metadata initialized %d",
1254 MetaspaceGC::should_concurrent_collect());
1255 }
1256 // ------------------------------------------------------------------
1257
1258 // If the estimated time to complete a cms collection (cms_duration())
1259 // is less than the estimated time remaining until the cms generation
1260 // is full, start a collection.
1261 if (!UseCMSInitiatingOccupancyOnly) {
1262 if (stats().valid()) {
1263 if (stats().time_until_cms_start() == 0.0) {
1264 return true;
1265 }
1266 } else {
1267 // We want to conservatively collect somewhat early in order
1268 // to try and "bootstrap" our CMS/promotion statistics;
1269 // this branch will not fire after the first successful CMS
1270 // collection because the stats should then be valid.
1271 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1272 if (Verbose && PrintGCDetails) {
1273 gclog_or_tty->print_cr(
1274 " CMSCollector: collect for bootstrapping statistics:"
1275 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1276 _bootstrap_occupancy);
1277 }
1278 return true;
1279 }
1280 }
1281 }
1282
1283 // Otherwise, we start a collection cycle if
1284 // old gen want a collection cycle started. Each may use
1285 // an appropriate criterion for making this decision.
1286 // XXX We need to make sure that the gen expansion
1287 // criterion dovetails well with this. XXX NEED TO FIX THIS
1288 if (_cmsGen->should_concurrent_collect()) {
1289 if (Verbose && PrintGCDetails) {
1290 gclog_or_tty->print_cr("CMS old gen initiated");
1291 }
1292 return true;
1293 }
1294
1295 // We start a collection if we believe an incremental collection may fail;
1296 // this is not likely to be productive in practice because it's probably too
1297 // late anyway.
1298 GenCollectedHeap* gch = GenCollectedHeap::heap();
1299 assert(gch->collector_policy()->is_generation_policy(),
1300 "You may want to check the correctness of the following");
1301 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1302 if (Verbose && PrintGCDetails) {
1303 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1304 }
1305 return true;
1306 }
1307
1308 if (MetaspaceGC::should_concurrent_collect()) {
1309 if (Verbose && PrintGCDetails) {
1310 gclog_or_tty->print("CMSCollector: collect for metadata allocation ");
1311 }
1312 return true;
1313 }
1314
1315 // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1316 if (CMSTriggerInterval >= 0) {
1317 if (CMSTriggerInterval == 0) {
1318 // Trigger always
1319 return true;
1320 }
1321
1322 // Check the CMS time since begin (we do not check the stats validity
1323 // as we want to be able to trigger the first CMS cycle as well)
1324 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1325 if (Verbose && PrintGCDetails) {
1326 if (stats().valid()) {
1327 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1328 stats().cms_time_since_begin());
1329 } else {
1330 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)");
1331 }
1332 }
1333 return true;
1334 }
1335 }
1336
1337 return false;
1338 }
1339
1340 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1341
1342 // Clear _expansion_cause fields of constituent generations
1343 void CMSCollector::clear_expansion_cause() {
1344 _cmsGen->clear_expansion_cause();
1345 }
1346
1347 // We should be conservative in starting a collection cycle. To
1348 // start too eagerly runs the risk of collecting too often in the
1349 // extreme. To collect too rarely falls back on full collections,
1350 // which works, even if not optimum in terms of concurrent work.
1351 // As a work around for too eagerly collecting, use the flag
1352 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1353 // giving the user an easily understandable way of controlling the
1354 // collections.
1355 // We want to start a new collection cycle if any of the following
1356 // conditions hold:
1357 // . our current occupancy exceeds the configured initiating occupancy
1358 // for this generation, or
1359 // . we recently needed to expand this space and have not, since that
1360 // expansion, done a collection of this generation, or
1361 // . the underlying space believes that it may be a good idea to initiate
1362 // a concurrent collection (this may be based on criteria such as the
1363 // following: the space uses linear allocation and linear allocation is
1364 // going to fail, or there is believed to be excessive fragmentation in
1365 // the generation, etc... or ...
1366 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1367 // the case of the old generation; see CR 6543076):
1368 // we may be approaching a point at which allocation requests may fail because
1369 // we will be out of sufficient free space given allocation rate estimates.]
1370 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1371
1372 assert_lock_strong(freelistLock());
1373 if (occupancy() > initiating_occupancy()) {
1374 if (PrintGCDetails && Verbose) {
1375 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1376 short_name(), occupancy(), initiating_occupancy());
1377 }
1378 return true;
1379 }
1380 if (UseCMSInitiatingOccupancyOnly) {
1381 return false;
1382 }
1383 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1384 if (PrintGCDetails && Verbose) {
1385 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1386 short_name());
1387 }
1388 return true;
1389 }
1390 if (_cmsSpace->should_concurrent_collect()) {
1391 if (PrintGCDetails && Verbose) {
1392 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1393 short_name());
1394 }
1395 return true;
1396 }
1397 return false;
1398 }
1399
1400 void ConcurrentMarkSweepGeneration::collect(bool full,
1401 bool clear_all_soft_refs,
1402 size_t size,
1403 bool tlab)
1404 {
1405 collector()->collect(full, clear_all_soft_refs, size, tlab);
1406 }
1407
1408 void CMSCollector::collect(bool full,
1409 bool clear_all_soft_refs,
1410 size_t size,
1411 bool tlab)
1412 {
1413 // The following "if" branch is present for defensive reasons.
1414 // In the current uses of this interface, it can be replaced with:
1415 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1416 // But I am not placing that assert here to allow future
1417 // generality in invoking this interface.
1418 if (GC_locker::is_active()) {
1419 // A consistency test for GC_locker
1420 assert(GC_locker::needs_gc(), "Should have been set already");
1421 // Skip this foreground collection, instead
1422 // expanding the heap if necessary.
1423 // Need the free list locks for the call to free() in compute_new_size()
1424 compute_new_size();
1425 return;
1426 }
1427 acquire_control_and_collect(full, clear_all_soft_refs);
1428 }
1429
1430 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1431 GenCollectedHeap* gch = GenCollectedHeap::heap();
1432 unsigned int gc_count = gch->total_full_collections();
1433 if (gc_count == full_gc_count) {
1434 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1435 _full_gc_requested = true;
1436 _full_gc_cause = cause;
1437 CGC_lock->notify(); // nudge CMS thread
1438 } else {
1439 assert(gc_count > full_gc_count, "Error: causal loop");
1440 }
1441 }
1442
1443 bool CMSCollector::is_external_interruption() {
1444 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1445 return GCCause::is_user_requested_gc(cause) ||
1446 GCCause::is_serviceability_requested_gc(cause);
1447 }
1448
1449 void CMSCollector::report_concurrent_mode_interruption() {
1450 if (is_external_interruption()) {
1451 if (PrintGCDetails) {
1452 gclog_or_tty->print(" (concurrent mode interrupted)");
1453 }
1454 } else {
1455 if (PrintGCDetails) {
1456 gclog_or_tty->print(" (concurrent mode failure)");
1457 }
1458 _gc_tracer_cm->report_concurrent_mode_failure();
1459 }
1460 }
1461
1462
1463 // The foreground and background collectors need to coordinate in order
1464 // to make sure that they do not mutually interfere with CMS collections.
1465 // When a background collection is active,
1466 // the foreground collector may need to take over (preempt) and
1467 // synchronously complete an ongoing collection. Depending on the
1468 // frequency of the background collections and the heap usage
1469 // of the application, this preemption can be seldom or frequent.
1470 // There are only certain
1471 // points in the background collection that the "collection-baton"
1472 // can be passed to the foreground collector.
1473 //
1474 // The foreground collector will wait for the baton before
1475 // starting any part of the collection. The foreground collector
1476 // will only wait at one location.
1477 //
1478 // The background collector will yield the baton before starting a new
1479 // phase of the collection (e.g., before initial marking, marking from roots,
1480 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1481 // of the loop which switches the phases. The background collector does some
1482 // of the phases (initial mark, final re-mark) with the world stopped.
1483 // Because of locking involved in stopping the world,
1484 // the foreground collector should not block waiting for the background
1485 // collector when it is doing a stop-the-world phase. The background
1486 // collector will yield the baton at an additional point just before
1487 // it enters a stop-the-world phase. Once the world is stopped, the
1488 // background collector checks the phase of the collection. If the
1489 // phase has not changed, it proceeds with the collection. If the
1490 // phase has changed, it skips that phase of the collection. See
1491 // the comments on the use of the Heap_lock in collect_in_background().
1492 //
1493 // Variable used in baton passing.
1494 // _foregroundGCIsActive - Set to true by the foreground collector when
1495 // it wants the baton. The foreground clears it when it has finished
1496 // the collection.
1497 // _foregroundGCShouldWait - Set to true by the background collector
1498 // when it is running. The foreground collector waits while
1499 // _foregroundGCShouldWait is true.
1500 // CGC_lock - monitor used to protect access to the above variables
1501 // and to notify the foreground and background collectors.
1502 // _collectorState - current state of the CMS collection.
1503 //
1504 // The foreground collector
1505 // acquires the CGC_lock
1506 // sets _foregroundGCIsActive
1507 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1508 // various locks acquired in preparation for the collection
1509 // are released so as not to block the background collector
1510 // that is in the midst of a collection
1511 // proceeds with the collection
1512 // clears _foregroundGCIsActive
1513 // returns
1514 //
1515 // The background collector in a loop iterating on the phases of the
1516 // collection
1517 // acquires the CGC_lock
1518 // sets _foregroundGCShouldWait
1519 // if _foregroundGCIsActive is set
1520 // clears _foregroundGCShouldWait, notifies _CGC_lock
1521 // waits on _CGC_lock for _foregroundGCIsActive to become false
1522 // and exits the loop.
1523 // otherwise
1524 // proceed with that phase of the collection
1525 // if the phase is a stop-the-world phase,
1526 // yield the baton once more just before enqueueing
1527 // the stop-world CMS operation (executed by the VM thread).
1528 // returns after all phases of the collection are done
1529 //
1530
1531 void CMSCollector::acquire_control_and_collect(bool full,
1532 bool clear_all_soft_refs) {
1533 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1534 assert(!Thread::current()->is_ConcurrentGC_thread(),
1535 "shouldn't try to acquire control from self!");
1536
1537 // Start the protocol for acquiring control of the
1538 // collection from the background collector (aka CMS thread).
1539 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1540 "VM thread should have CMS token");
1541 // Remember the possibly interrupted state of an ongoing
1542 // concurrent collection
1543 CollectorState first_state = _collectorState;
1544
1545 // Signal to a possibly ongoing concurrent collection that
1546 // we want to do a foreground collection.
1547 _foregroundGCIsActive = true;
1548
1549 // release locks and wait for a notify from the background collector
1550 // releasing the locks in only necessary for phases which
1551 // do yields to improve the granularity of the collection.
1552 assert_lock_strong(bitMapLock());
1553 // We need to lock the Free list lock for the space that we are
1554 // currently collecting.
1555 assert(haveFreelistLocks(), "Must be holding free list locks");
1556 bitMapLock()->unlock();
1557 releaseFreelistLocks();
1558 {
1559 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1560 if (_foregroundGCShouldWait) {
1561 // We are going to be waiting for action for the CMS thread;
1562 // it had better not be gone (for instance at shutdown)!
1563 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1564 "CMS thread must be running");
1565 // Wait here until the background collector gives us the go-ahead
1566 ConcurrentMarkSweepThread::clear_CMS_flag(
1567 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1568 // Get a possibly blocked CMS thread going:
1569 // Note that we set _foregroundGCIsActive true above,
1570 // without protection of the CGC_lock.
1571 CGC_lock->notify();
1572 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1573 "Possible deadlock");
1574 while (_foregroundGCShouldWait) {
1575 // wait for notification
1576 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1577 // Possibility of delay/starvation here, since CMS token does
1578 // not know to give priority to VM thread? Actually, i think
1579 // there wouldn't be any delay/starvation, but the proof of
1580 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1581 }
1582 ConcurrentMarkSweepThread::set_CMS_flag(
1583 ConcurrentMarkSweepThread::CMS_vm_has_token);
1584 }
1585 }
1586 // The CMS_token is already held. Get back the other locks.
1587 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1588 "VM thread should have CMS token");
1589 getFreelistLocks();
1590 bitMapLock()->lock_without_safepoint_check();
1591 if (TraceCMSState) {
1592 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1593 INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1594 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1595 }
1596
1597 // Inform cms gen if this was due to partial collection failing.
1598 // The CMS gen may use this fact to determine its expansion policy.
1599 GenCollectedHeap* gch = GenCollectedHeap::heap();
1600 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1601 assert(!_cmsGen->incremental_collection_failed(),
1602 "Should have been noticed, reacted to and cleared");
1603 _cmsGen->set_incremental_collection_failed();
1604 }
1605
1606 if (first_state > Idling) {
1607 report_concurrent_mode_interruption();
1608 }
1609
1610 set_did_compact(true);
1611
1612 // If the collection is being acquired from the background
1613 // collector, there may be references on the discovered
1614 // references lists. Abandon those references, since some
1615 // of them may have become unreachable after concurrent
1616 // discovery; the STW compacting collector will redo discovery
1617 // more precisely, without being subject to floating garbage.
1618 // Leaving otherwise unreachable references in the discovered
1619 // lists would require special handling.
1620 ref_processor()->disable_discovery();
1621 ref_processor()->abandon_partial_discovery();
1622 ref_processor()->verify_no_references_recorded();
1623
1624 if (first_state > Idling) {
1625 save_heap_summary();
1626 }
1627
1628 do_compaction_work(clear_all_soft_refs);
1629
1630 // Has the GC time limit been exceeded?
1631 size_t max_eden_size = _young_gen->max_capacity() -
1632 _young_gen->to()->capacity() -
1633 _young_gen->from()->capacity();
1634 GCCause::Cause gc_cause = gch->gc_cause();
1635 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1636 _young_gen->eden()->used(),
1637 _cmsGen->max_capacity(),
1638 max_eden_size,
1639 full,
1640 gc_cause,
1641 gch->collector_policy());
1642
1643 // Reset the expansion cause, now that we just completed
1644 // a collection cycle.
1645 clear_expansion_cause();
1646 _foregroundGCIsActive = false;
1647 return;
1648 }
1649
1650 // Resize the tenured generation
1651 // after obtaining the free list locks for the
1652 // two generations.
1653 void CMSCollector::compute_new_size() {
1654 assert_locked_or_safepoint(Heap_lock);
1655 FreelistLocker z(this);
1656 MetaspaceGC::compute_new_size();
1657 _cmsGen->compute_new_size_free_list();
1658 }
1659
1660 // A work method used by the foreground collector to do
1661 // a mark-sweep-compact.
1662 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1663 GenCollectedHeap* gch = GenCollectedHeap::heap();
1664
1665 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1666 gc_timer->register_gc_start();
1667
1668 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1669 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
1670
1671 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id());
1672
1673 // Temporarily widen the span of the weak reference processing to
1674 // the entire heap.
1675 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1676 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1677 // Temporarily, clear the "is_alive_non_header" field of the
1678 // reference processor.
1679 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1680 // Temporarily make reference _processing_ single threaded (non-MT).
1681 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1682 // Temporarily make refs discovery atomic
1683 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1684 // Temporarily make reference _discovery_ single threaded (non-MT)
1685 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1686
1687 ref_processor()->set_enqueuing_is_done(false);
1688 ref_processor()->enable_discovery();
1689 ref_processor()->setup_policy(clear_all_soft_refs);
1690 // If an asynchronous collection finishes, the _modUnionTable is
1691 // all clear. If we are assuming the collection from an asynchronous
1692 // collection, clear the _modUnionTable.
1693 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1694 "_modUnionTable should be clear if the baton was not passed");
1695 _modUnionTable.clear_all();
1696 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
1697 "mod union for klasses should be clear if the baton was passed");
1698 _ct->klass_rem_set()->clear_mod_union();
1699
1700 // We must adjust the allocation statistics being maintained
1701 // in the free list space. We do so by reading and clearing
1702 // the sweep timer and updating the block flux rate estimates below.
1703 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1704 if (_inter_sweep_timer.is_active()) {
1705 _inter_sweep_timer.stop();
1706 // Note that we do not use this sample to update the _inter_sweep_estimate.
1707 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1708 _inter_sweep_estimate.padded_average(),
1709 _intra_sweep_estimate.padded_average());
1710 }
1711
1712 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
1713 ref_processor(), clear_all_soft_refs);
1714 #ifdef ASSERT
1715 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1716 size_t free_size = cms_space->free();
1717 assert(free_size ==
1718 pointer_delta(cms_space->end(), cms_space->compaction_top())
1719 * HeapWordSize,
1720 "All the free space should be compacted into one chunk at top");
1721 assert(cms_space->dictionary()->total_chunk_size(
1722 debug_only(cms_space->freelistLock())) == 0 ||
1723 cms_space->totalSizeInIndexedFreeLists() == 0,
1724 "All the free space should be in a single chunk");
1725 size_t num = cms_space->totalCount();
1726 assert((free_size == 0 && num == 0) ||
1727 (free_size > 0 && (num == 1 || num == 2)),
1728 "There should be at most 2 free chunks after compaction");
1729 #endif // ASSERT
1730 _collectorState = Resetting;
1731 assert(_restart_addr == NULL,
1732 "Should have been NULL'd before baton was passed");
1733 reset(false /* == !concurrent */);
1734 _cmsGen->reset_after_compaction();
1735 _concurrent_cycles_since_last_unload = 0;
1736
1737 // Clear any data recorded in the PLAB chunk arrays.
1738 if (_survivor_plab_array != NULL) {
1739 reset_survivor_plab_arrays();
1740 }
1741
1742 // Adjust the per-size allocation stats for the next epoch.
1743 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1744 // Restart the "inter sweep timer" for the next epoch.
1745 _inter_sweep_timer.reset();
1746 _inter_sweep_timer.start();
1747
1748 gc_timer->register_gc_end();
1749
1750 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1751
1752 // For a mark-sweep-compact, compute_new_size() will be called
1753 // in the heap's do_collection() method.
1754 }
1755
1756 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1757 ContiguousSpace* eden_space = _young_gen->eden();
1758 ContiguousSpace* from_space = _young_gen->from();
1759 ContiguousSpace* to_space = _young_gen->to();
1760 // Eden
1761 if (_eden_chunk_array != NULL) {
1762 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1763 eden_space->bottom(), eden_space->top(),
1764 eden_space->end(), eden_space->capacity());
1765 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", "
1766 "_eden_chunk_capacity=" SIZE_FORMAT,
1767 _eden_chunk_index, _eden_chunk_capacity);
1768 for (size_t i = 0; i < _eden_chunk_index; i++) {
1769 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1770 i, _eden_chunk_array[i]);
1771 }
1772 }
1773 // Survivor
1774 if (_survivor_chunk_array != NULL) {
1775 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1776 from_space->bottom(), from_space->top(),
1777 from_space->end(), from_space->capacity());
1778 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", "
1779 "_survivor_chunk_capacity=" SIZE_FORMAT,
1780 _survivor_chunk_index, _survivor_chunk_capacity);
1781 for (size_t i = 0; i < _survivor_chunk_index; i++) {
1782 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1783 i, _survivor_chunk_array[i]);
1784 }
1785 }
1786 }
1787
1788 void CMSCollector::getFreelistLocks() const {
1789 // Get locks for all free lists in all generations that this
1790 // collector is responsible for
1791 _cmsGen->freelistLock()->lock_without_safepoint_check();
1792 }
1793
1794 void CMSCollector::releaseFreelistLocks() const {
1795 // Release locks for all free lists in all generations that this
1796 // collector is responsible for
1797 _cmsGen->freelistLock()->unlock();
1798 }
1799
1800 bool CMSCollector::haveFreelistLocks() const {
1801 // Check locks for all free lists in all generations that this
1802 // collector is responsible for
1803 assert_lock_strong(_cmsGen->freelistLock());
1804 PRODUCT_ONLY(ShouldNotReachHere());
1805 return true;
1806 }
1807
1808 // A utility class that is used by the CMS collector to
1809 // temporarily "release" the foreground collector from its
1810 // usual obligation to wait for the background collector to
1811 // complete an ongoing phase before proceeding.
1812 class ReleaseForegroundGC: public StackObj {
1813 private:
1814 CMSCollector* _c;
1815 public:
1816 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1817 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1818 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1819 // allow a potentially blocked foreground collector to proceed
1820 _c->_foregroundGCShouldWait = false;
1821 if (_c->_foregroundGCIsActive) {
1822 CGC_lock->notify();
1823 }
1824 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1825 "Possible deadlock");
1826 }
1827
1828 ~ReleaseForegroundGC() {
1829 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1830 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1831 _c->_foregroundGCShouldWait = true;
1832 }
1833 };
1834
1835 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1836 assert(Thread::current()->is_ConcurrentGC_thread(),
1837 "A CMS asynchronous collection is only allowed on a CMS thread.");
1838
1839 GenCollectedHeap* gch = GenCollectedHeap::heap();
1840 {
1841 bool safepoint_check = Mutex::_no_safepoint_check_flag;
1842 MutexLockerEx hl(Heap_lock, safepoint_check);
1843 FreelistLocker fll(this);
1844 MutexLockerEx x(CGC_lock, safepoint_check);
1845 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
1846 // The foreground collector is active or we're
1847 // not using asynchronous collections. Skip this
1848 // background collection.
1849 assert(!_foregroundGCShouldWait, "Should be clear");
1850 return;
1851 } else {
1852 assert(_collectorState == Idling, "Should be idling before start.");
1853 _collectorState = InitialMarking;
1854 register_gc_start(cause);
1855 // Reset the expansion cause, now that we are about to begin
1856 // a new cycle.
1857 clear_expansion_cause();
1858
1859 // Clear the MetaspaceGC flag since a concurrent collection
1860 // is starting but also clear it after the collection.
1861 MetaspaceGC::set_should_concurrent_collect(false);
1862 }
1863 // Decide if we want to enable class unloading as part of the
1864 // ensuing concurrent GC cycle.
1865 update_should_unload_classes();
1866 _full_gc_requested = false; // acks all outstanding full gc requests
1867 _full_gc_cause = GCCause::_no_gc;
1868 // Signal that we are about to start a collection
1869 gch->increment_total_full_collections(); // ... starting a collection cycle
1870 _collection_count_start = gch->total_full_collections();
1871 }
1872
1873 // Used for PrintGC
1874 size_t prev_used;
1875 if (PrintGC && Verbose) {
1876 prev_used = _cmsGen->used();
1877 }
1878
1879 // The change of the collection state is normally done at this level;
1880 // the exceptions are phases that are executed while the world is
1881 // stopped. For those phases the change of state is done while the
1882 // world is stopped. For baton passing purposes this allows the
1883 // background collector to finish the phase and change state atomically.
1884 // The foreground collector cannot wait on a phase that is done
1885 // while the world is stopped because the foreground collector already
1886 // has the world stopped and would deadlock.
1887 while (_collectorState != Idling) {
1888 if (TraceCMSState) {
1889 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
1890 Thread::current(), _collectorState);
1891 }
1892 // The foreground collector
1893 // holds the Heap_lock throughout its collection.
1894 // holds the CMS token (but not the lock)
1895 // except while it is waiting for the background collector to yield.
1896 //
1897 // The foreground collector should be blocked (not for long)
1898 // if the background collector is about to start a phase
1899 // executed with world stopped. If the background
1900 // collector has already started such a phase, the
1901 // foreground collector is blocked waiting for the
1902 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
1903 // are executed in the VM thread.
1904 //
1905 // The locking order is
1906 // PendingListLock (PLL) -- if applicable (FinalMarking)
1907 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
1908 // CMS token (claimed in
1909 // stop_world_and_do() -->
1910 // safepoint_synchronize() -->
1911 // CMSThread::synchronize())
1912
1913 {
1914 // Check if the FG collector wants us to yield.
1915 CMSTokenSync x(true); // is cms thread
1916 if (waitForForegroundGC()) {
1917 // We yielded to a foreground GC, nothing more to be
1918 // done this round.
1919 assert(_foregroundGCShouldWait == false, "We set it to false in "
1920 "waitForForegroundGC()");
1921 if (TraceCMSState) {
1922 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1923 " exiting collection CMS state %d",
1924 Thread::current(), _collectorState);
1925 }
1926 return;
1927 } else {
1928 // The background collector can run but check to see if the
1929 // foreground collector has done a collection while the
1930 // background collector was waiting to get the CGC_lock
1931 // above. If yes, break so that _foregroundGCShouldWait
1932 // is cleared before returning.
1933 if (_collectorState == Idling) {
1934 break;
1935 }
1936 }
1937 }
1938
1939 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1940 "should be waiting");
1941
1942 switch (_collectorState) {
1943 case InitialMarking:
1944 {
1945 ReleaseForegroundGC x(this);
1946 stats().record_cms_begin();
1947 VM_CMS_Initial_Mark initial_mark_op(this);
1948 VMThread::execute(&initial_mark_op);
1949 }
1950 // The collector state may be any legal state at this point
1951 // since the background collector may have yielded to the
1952 // foreground collector.
1953 break;
1954 case Marking:
1955 // initial marking in checkpointRootsInitialWork has been completed
1956 if (markFromRoots()) { // we were successful
1957 assert(_collectorState == Precleaning, "Collector state should "
1958 "have changed");
1959 } else {
1960 assert(_foregroundGCIsActive, "Internal state inconsistency");
1961 }
1962 break;
1963 case Precleaning:
1964 // marking from roots in markFromRoots has been completed
1965 preclean();
1966 assert(_collectorState == AbortablePreclean ||
1967 _collectorState == FinalMarking,
1968 "Collector state should have changed");
1969 break;
1970 case AbortablePreclean:
1971 abortable_preclean();
1972 assert(_collectorState == FinalMarking, "Collector state should "
1973 "have changed");
1974 break;
1975 case FinalMarking:
1976 {
1977 ReleaseForegroundGC x(this);
1978
1979 VM_CMS_Final_Remark final_remark_op(this);
1980 VMThread::execute(&final_remark_op);
1981 }
1982 assert(_foregroundGCShouldWait, "block post-condition");
1983 break;
1984 case Sweeping:
1985 // final marking in checkpointRootsFinal has been completed
1986 sweep();
1987 assert(_collectorState == Resizing, "Collector state change "
1988 "to Resizing must be done under the free_list_lock");
1989
1990 case Resizing: {
1991 // Sweeping has been completed...
1992 // At this point the background collection has completed.
1993 // Don't move the call to compute_new_size() down
1994 // into code that might be executed if the background
1995 // collection was preempted.
1996 {
1997 ReleaseForegroundGC x(this); // unblock FG collection
1998 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
1999 CMSTokenSync z(true); // not strictly needed.
2000 if (_collectorState == Resizing) {
2001 compute_new_size();
2002 save_heap_summary();
2003 _collectorState = Resetting;
2004 } else {
2005 assert(_collectorState == Idling, "The state should only change"
2006 " because the foreground collector has finished the collection");
2007 }
2008 }
2009 break;
2010 }
2011 case Resetting:
2012 // CMS heap resizing has been completed
2013 reset(true);
2014 assert(_collectorState == Idling, "Collector state should "
2015 "have changed");
2016
2017 MetaspaceGC::set_should_concurrent_collect(false);
2018
2019 stats().record_cms_end();
2020 // Don't move the concurrent_phases_end() and compute_new_size()
2021 // calls to here because a preempted background collection
2022 // has it's state set to "Resetting".
2023 break;
2024 case Idling:
2025 default:
2026 ShouldNotReachHere();
2027 break;
2028 }
2029 if (TraceCMSState) {
2030 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2031 Thread::current(), _collectorState);
2032 }
2033 assert(_foregroundGCShouldWait, "block post-condition");
2034 }
2035
2036 // Should this be in gc_epilogue?
2037 collector_policy()->counters()->update_counters();
2038
2039 {
2040 // Clear _foregroundGCShouldWait and, in the event that the
2041 // foreground collector is waiting, notify it, before
2042 // returning.
2043 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2044 _foregroundGCShouldWait = false;
2045 if (_foregroundGCIsActive) {
2046 CGC_lock->notify();
2047 }
2048 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2049 "Possible deadlock");
2050 }
2051 if (TraceCMSState) {
2052 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2053 " exiting collection CMS state %d",
2054 Thread::current(), _collectorState);
2055 }
2056 if (PrintGC && Verbose) {
2057 _cmsGen->print_heap_change(prev_used);
2058 }
2059 }
2060
2061 void CMSCollector::register_gc_start(GCCause::Cause cause) {
2062 _cms_start_registered = true;
2063 _gc_timer_cm->register_gc_start();
2064 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
2065 }
2066
2067 void CMSCollector::register_gc_end() {
2068 if (_cms_start_registered) {
2069 report_heap_summary(GCWhen::AfterGC);
2070
2071 _gc_timer_cm->register_gc_end();
2072 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
2073 _cms_start_registered = false;
2074 }
2075 }
2076
2077 void CMSCollector::save_heap_summary() {
2078 GenCollectedHeap* gch = GenCollectedHeap::heap();
2079 _last_heap_summary = gch->create_heap_summary();
2080 _last_metaspace_summary = gch->create_metaspace_summary();
2081 }
2082
2083 void CMSCollector::report_heap_summary(GCWhen::Type when) {
2084 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
2085 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
2086 }
2087
2088 bool CMSCollector::waitForForegroundGC() {
2089 bool res = false;
2090 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2091 "CMS thread should have CMS token");
2092 // Block the foreground collector until the
2093 // background collectors decides whether to
2094 // yield.
2095 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2096 _foregroundGCShouldWait = true;
2097 if (_foregroundGCIsActive) {
2098 // The background collector yields to the
2099 // foreground collector and returns a value
2100 // indicating that it has yielded. The foreground
2101 // collector can proceed.
2102 res = true;
2103 _foregroundGCShouldWait = false;
2104 ConcurrentMarkSweepThread::clear_CMS_flag(
2105 ConcurrentMarkSweepThread::CMS_cms_has_token);
2106 ConcurrentMarkSweepThread::set_CMS_flag(
2107 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2108 // Get a possibly blocked foreground thread going
2109 CGC_lock->notify();
2110 if (TraceCMSState) {
2111 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2112 Thread::current(), _collectorState);
2113 }
2114 while (_foregroundGCIsActive) {
2115 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2116 }
2117 ConcurrentMarkSweepThread::set_CMS_flag(
2118 ConcurrentMarkSweepThread::CMS_cms_has_token);
2119 ConcurrentMarkSweepThread::clear_CMS_flag(
2120 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2121 }
2122 if (TraceCMSState) {
2123 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2124 Thread::current(), _collectorState);
2125 }
2126 return res;
2127 }
2128
2129 // Because of the need to lock the free lists and other structures in
2130 // the collector, common to all the generations that the collector is
2131 // collecting, we need the gc_prologues of individual CMS generations
2132 // delegate to their collector. It may have been simpler had the
2133 // current infrastructure allowed one to call a prologue on a
2134 // collector. In the absence of that we have the generation's
2135 // prologue delegate to the collector, which delegates back
2136 // some "local" work to a worker method in the individual generations
2137 // that it's responsible for collecting, while itself doing any
2138 // work common to all generations it's responsible for. A similar
2139 // comment applies to the gc_epilogue()'s.
2140 // The role of the variable _between_prologue_and_epilogue is to
2141 // enforce the invocation protocol.
2142 void CMSCollector::gc_prologue(bool full) {
2143 // Call gc_prologue_work() for the CMSGen
2144 // we are responsible for.
2145
2146 // The following locking discipline assumes that we are only called
2147 // when the world is stopped.
2148 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2149
2150 // The CMSCollector prologue must call the gc_prologues for the
2151 // "generations" that it's responsible
2152 // for.
2153
2154 assert( Thread::current()->is_VM_thread()
2155 || ( CMSScavengeBeforeRemark
2156 && Thread::current()->is_ConcurrentGC_thread()),
2157 "Incorrect thread type for prologue execution");
2158
2159 if (_between_prologue_and_epilogue) {
2160 // We have already been invoked; this is a gc_prologue delegation
2161 // from yet another CMS generation that we are responsible for, just
2162 // ignore it since all relevant work has already been done.
2163 return;
2164 }
2165
2166 // set a bit saying prologue has been called; cleared in epilogue
2167 _between_prologue_and_epilogue = true;
2168 // Claim locks for common data structures, then call gc_prologue_work()
2169 // for each CMSGen.
2170
2171 getFreelistLocks(); // gets free list locks on constituent spaces
2172 bitMapLock()->lock_without_safepoint_check();
2173
2174 // Should call gc_prologue_work() for all cms gens we are responsible for
2175 bool duringMarking = _collectorState >= Marking
2176 && _collectorState < Sweeping;
2177
2178 // The young collections clear the modified oops state, which tells if
2179 // there are any modified oops in the class. The remark phase also needs
2180 // that information. Tell the young collection to save the union of all
2181 // modified klasses.
2182 if (duringMarking) {
2183 _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2184 }
2185
2186 bool registerClosure = duringMarking;
2187
2188 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ?
2189 &_modUnionClosurePar
2190 : &_modUnionClosure;
2191 _cmsGen->gc_prologue_work(full, registerClosure, muc);
2192
2193 if (!full) {
2194 stats().record_gc0_begin();
2195 }
2196 }
2197
2198 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2199
2200 _capacity_at_prologue = capacity();
2201 _used_at_prologue = used();
2202
2203 // Delegate to CMScollector which knows how to coordinate between
2204 // this and any other CMS generations that it is responsible for
2205 // collecting.
2206 collector()->gc_prologue(full);
2207 }
2208
2209 // This is a "private" interface for use by this generation's CMSCollector.
2210 // Not to be called directly by any other entity (for instance,
2211 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2212 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2213 bool registerClosure, ModUnionClosure* modUnionClosure) {
2214 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2215 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2216 "Should be NULL");
2217 if (registerClosure) {
2218 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2219 }
2220 cmsSpace()->gc_prologue();
2221 // Clear stat counters
2222 NOT_PRODUCT(
2223 assert(_numObjectsPromoted == 0, "check");
2224 assert(_numWordsPromoted == 0, "check");
2225 if (Verbose && PrintGC) {
2226 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2227 SIZE_FORMAT" bytes concurrently",
2228 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2229 }
2230 _numObjectsAllocated = 0;
2231 _numWordsAllocated = 0;
2232 )
2233 }
2234
2235 void CMSCollector::gc_epilogue(bool full) {
2236 // The following locking discipline assumes that we are only called
2237 // when the world is stopped.
2238 assert(SafepointSynchronize::is_at_safepoint(),
2239 "world is stopped assumption");
2240
2241 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2242 // if linear allocation blocks need to be appropriately marked to allow the
2243 // the blocks to be parsable. We also check here whether we need to nudge the
2244 // CMS collector thread to start a new cycle (if it's not already active).
2245 assert( Thread::current()->is_VM_thread()
2246 || ( CMSScavengeBeforeRemark
2247 && Thread::current()->is_ConcurrentGC_thread()),
2248 "Incorrect thread type for epilogue execution");
2249
2250 if (!_between_prologue_and_epilogue) {
2251 // We have already been invoked; this is a gc_epilogue delegation
2252 // from yet another CMS generation that we are responsible for, just
2253 // ignore it since all relevant work has already been done.
2254 return;
2255 }
2256 assert(haveFreelistLocks(), "must have freelist locks");
2257 assert_lock_strong(bitMapLock());
2258
2259 _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2260
2261 _cmsGen->gc_epilogue_work(full);
2262
2263 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2264 // in case sampling was not already enabled, enable it
2265 _start_sampling = true;
2266 }
2267 // reset _eden_chunk_array so sampling starts afresh
2268 _eden_chunk_index = 0;
2269
2270 size_t cms_used = _cmsGen->cmsSpace()->used();
2271
2272 // update performance counters - this uses a special version of
2273 // update_counters() that allows the utilization to be passed as a
2274 // parameter, avoiding multiple calls to used().
2275 //
2276 _cmsGen->update_counters(cms_used);
2277
2278 bitMapLock()->unlock();
2279 releaseFreelistLocks();
2280
2281 if (!CleanChunkPoolAsync) {
2282 Chunk::clean_chunk_pool();
2283 }
2284
2285 set_did_compact(false);
2286 _between_prologue_and_epilogue = false; // ready for next cycle
2287 }
2288
2289 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2290 collector()->gc_epilogue(full);
2291
2292 // Also reset promotion tracking in par gc thread states.
2293 for (uint i = 0; i < ParallelGCThreads; i++) {
2294 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2295 }
2296 }
2297
2298 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2299 assert(!incremental_collection_failed(), "Should have been cleared");
2300 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2301 cmsSpace()->gc_epilogue();
2302 // Print stat counters
2303 NOT_PRODUCT(
2304 assert(_numObjectsAllocated == 0, "check");
2305 assert(_numWordsAllocated == 0, "check");
2306 if (Verbose && PrintGC) {
2307 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2308 SIZE_FORMAT" bytes",
2309 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2310 }
2311 _numObjectsPromoted = 0;
2312 _numWordsPromoted = 0;
2313 )
2314
2315 if (PrintGC && Verbose) {
2316 // Call down the chain in contiguous_available needs the freelistLock
2317 // so print this out before releasing the freeListLock.
2318 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2319 contiguous_available());
2320 }
2321 }
2322
2323 #ifndef PRODUCT
2324 bool CMSCollector::have_cms_token() {
2325 Thread* thr = Thread::current();
2326 if (thr->is_VM_thread()) {
2327 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2328 } else if (thr->is_ConcurrentGC_thread()) {
2329 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2330 } else if (thr->is_GC_task_thread()) {
2331 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2332 ParGCRareEvent_lock->owned_by_self();
2333 }
2334 return false;
2335 }
2336 #endif
2337
2338 // Check reachability of the given heap address in CMS generation,
2339 // treating all other generations as roots.
2340 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2341 // We could "guarantee" below, rather than assert, but I'll
2342 // leave these as "asserts" so that an adventurous debugger
2343 // could try this in the product build provided some subset of
2344 // the conditions were met, provided they were interested in the
2345 // results and knew that the computation below wouldn't interfere
2346 // with other concurrent computations mutating the structures
2347 // being read or written.
2348 assert(SafepointSynchronize::is_at_safepoint(),
2349 "Else mutations in object graph will make answer suspect");
2350 assert(have_cms_token(), "Should hold cms token");
2351 assert(haveFreelistLocks(), "must hold free list locks");
2352 assert_lock_strong(bitMapLock());
2353
2354 // Clear the marking bit map array before starting, but, just
2355 // for kicks, first report if the given address is already marked
2356 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", addr,
2357 _markBitMap.isMarked(addr) ? "" : " not");
2358
2359 if (verify_after_remark()) {
2360 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2361 bool result = verification_mark_bm()->isMarked(addr);
2362 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", addr,
2363 result ? "IS" : "is NOT");
2364 return result;
2365 } else {
2366 gclog_or_tty->print_cr("Could not compute result");
2367 return false;
2368 }
2369 }
2370
2371
2372 void
2373 CMSCollector::print_on_error(outputStream* st) {
2374 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2375 if (collector != NULL) {
2376 CMSBitMap* bitmap = &collector->_markBitMap;
2377 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap);
2378 bitmap->print_on_error(st, " Bits: ");
2379
2380 st->cr();
2381
2382 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2383 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap);
2384 mut_bitmap->print_on_error(st, " Bits: ");
2385 }
2386 }
2387
2388 ////////////////////////////////////////////////////////
2389 // CMS Verification Support
2390 ////////////////////////////////////////////////////////
2391 // Following the remark phase, the following invariant
2392 // should hold -- each object in the CMS heap which is
2393 // marked in markBitMap() should be marked in the verification_mark_bm().
2394
2395 class VerifyMarkedClosure: public BitMapClosure {
2396 CMSBitMap* _marks;
2397 bool _failed;
2398
2399 public:
2400 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2401
2402 bool do_bit(size_t offset) {
2403 HeapWord* addr = _marks->offsetToHeapWord(offset);
2404 if (!_marks->isMarked(addr)) {
2405 oop(addr)->print_on(gclog_or_tty);
2406 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2407 _failed = true;
2408 }
2409 return true;
2410 }
2411
2412 bool failed() { return _failed; }
2413 };
2414
2415 bool CMSCollector::verify_after_remark(bool silent) {
2416 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... ");
2417 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2418 static bool init = false;
2419
2420 assert(SafepointSynchronize::is_at_safepoint(),
2421 "Else mutations in object graph will make answer suspect");
2422 assert(have_cms_token(),
2423 "Else there may be mutual interference in use of "
2424 " verification data structures");
2425 assert(_collectorState > Marking && _collectorState <= Sweeping,
2426 "Else marking info checked here may be obsolete");
2427 assert(haveFreelistLocks(), "must hold free list locks");
2428 assert_lock_strong(bitMapLock());
2429
2430
2431 // Allocate marking bit map if not already allocated
2432 if (!init) { // first time
2433 if (!verification_mark_bm()->allocate(_span)) {
2434 return false;
2435 }
2436 init = true;
2437 }
2438
2439 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2440
2441 // Turn off refs discovery -- so we will be tracing through refs.
2442 // This is as intended, because by this time
2443 // GC must already have cleared any refs that need to be cleared,
2444 // and traced those that need to be marked; moreover,
2445 // the marking done here is not going to interfere in any
2446 // way with the marking information used by GC.
2447 NoRefDiscovery no_discovery(ref_processor());
2448
2449 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2450
2451 // Clear any marks from a previous round
2452 verification_mark_bm()->clear_all();
2453 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2454 verify_work_stacks_empty();
2455
2456 GenCollectedHeap* gch = GenCollectedHeap::heap();
2457 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2458 // Update the saved marks which may affect the root scans.
2459 gch->save_marks();
2460
2461 if (CMSRemarkVerifyVariant == 1) {
2462 // In this first variant of verification, we complete
2463 // all marking, then check if the new marks-vector is
2464 // a subset of the CMS marks-vector.
2465 verify_after_remark_work_1();
2466 } else if (CMSRemarkVerifyVariant == 2) {
2467 // In this second variant of verification, we flag an error
2468 // (i.e. an object reachable in the new marks-vector not reachable
2469 // in the CMS marks-vector) immediately, also indicating the
2470 // identify of an object (A) that references the unmarked object (B) --
2471 // presumably, a mutation to A failed to be picked up by preclean/remark?
2472 verify_after_remark_work_2();
2473 } else {
2474 warning("Unrecognized value %d for CMSRemarkVerifyVariant",
2475 CMSRemarkVerifyVariant);
2476 }
2477 if (!silent) gclog_or_tty->print(" done] ");
2478 return true;
2479 }
2480
2481 void CMSCollector::verify_after_remark_work_1() {
2482 ResourceMark rm;
2483 HandleMark hm;
2484 GenCollectedHeap* gch = GenCollectedHeap::heap();
2485
2486 // Get a clear set of claim bits for the roots processing to work with.
2487 ClassLoaderDataGraph::clear_claimed_marks();
2488
2489 // Mark from roots one level into CMS
2490 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2491 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2492
2493 gch->gen_process_roots(_cmsGen->level(),
2494 true, // younger gens are roots
2495 true, // activate StrongRootsScope
2496 SharedHeap::ScanningOption(roots_scanning_options()),
2497 should_unload_classes(),
2498 ¬Older,
2499 NULL,
2500 NULL); // SSS: Provide correct closure
2501
2502 // Now mark from the roots
2503 MarkFromRootsClosure markFromRootsClosure(this, _span,
2504 verification_mark_bm(), verification_mark_stack(),
2505 false /* don't yield */, true /* verifying */);
2506 assert(_restart_addr == NULL, "Expected pre-condition");
2507 verification_mark_bm()->iterate(&markFromRootsClosure);
2508 while (_restart_addr != NULL) {
2509 // Deal with stack overflow: by restarting at the indicated
2510 // address.
2511 HeapWord* ra = _restart_addr;
2512 markFromRootsClosure.reset(ra);
2513 _restart_addr = NULL;
2514 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2515 }
2516 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2517 verify_work_stacks_empty();
2518
2519 // Marking completed -- now verify that each bit marked in
2520 // verification_mark_bm() is also marked in markBitMap(); flag all
2521 // errors by printing corresponding objects.
2522 VerifyMarkedClosure vcl(markBitMap());
2523 verification_mark_bm()->iterate(&vcl);
2524 if (vcl.failed()) {
2525 gclog_or_tty->print("Verification failed");
2526 Universe::heap()->print_on(gclog_or_tty);
2527 fatal("CMS: failed marking verification after remark");
2528 }
2529 }
2530
2531 class VerifyKlassOopsKlassClosure : public KlassClosure {
2532 class VerifyKlassOopsClosure : public OopClosure {
2533 CMSBitMap* _bitmap;
2534 public:
2535 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2536 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2537 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2538 } _oop_closure;
2539 public:
2540 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2541 void do_klass(Klass* k) {
2542 k->oops_do(&_oop_closure);
2543 }
2544 };
2545
2546 void CMSCollector::verify_after_remark_work_2() {
2547 ResourceMark rm;
2548 HandleMark hm;
2549 GenCollectedHeap* gch = GenCollectedHeap::heap();
2550
2551 // Get a clear set of claim bits for the roots processing to work with.
2552 ClassLoaderDataGraph::clear_claimed_marks();
2553
2554 // Mark from roots one level into CMS
2555 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2556 markBitMap());
2557 CLDToOopClosure cld_closure(¬Older, true);
2558
2559 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2560
2561 gch->gen_process_roots(_cmsGen->level(),
2562 true, // younger gens are roots
2563 true, // activate StrongRootsScope
2564 SharedHeap::ScanningOption(roots_scanning_options()),
2565 should_unload_classes(),
2566 ¬Older,
2567 NULL,
2568 &cld_closure);
2569
2570 // Now mark from the roots
2571 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2572 verification_mark_bm(), markBitMap(), verification_mark_stack());
2573 assert(_restart_addr == NULL, "Expected pre-condition");
2574 verification_mark_bm()->iterate(&markFromRootsClosure);
2575 while (_restart_addr != NULL) {
2576 // Deal with stack overflow: by restarting at the indicated
2577 // address.
2578 HeapWord* ra = _restart_addr;
2579 markFromRootsClosure.reset(ra);
2580 _restart_addr = NULL;
2581 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2582 }
2583 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2584 verify_work_stacks_empty();
2585
2586 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
2587 ClassLoaderDataGraph::classes_do(&verify_klass_oops);
2588
2589 // Marking completed -- now verify that each bit marked in
2590 // verification_mark_bm() is also marked in markBitMap(); flag all
2591 // errors by printing corresponding objects.
2592 VerifyMarkedClosure vcl(markBitMap());
2593 verification_mark_bm()->iterate(&vcl);
2594 assert(!vcl.failed(), "Else verification above should not have succeeded");
2595 }
2596
2597 void ConcurrentMarkSweepGeneration::save_marks() {
2598 // delegate to CMS space
2599 cmsSpace()->save_marks();
2600 for (uint i = 0; i < ParallelGCThreads; i++) {
2601 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2602 }
2603 }
2604
2605 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2606 return cmsSpace()->no_allocs_since_save_marks();
2607 }
2608
2609 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2610 \
2611 void ConcurrentMarkSweepGeneration:: \
2612 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2613 cl->set_generation(this); \
2614 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2615 cl->reset_generation(); \
2616 save_marks(); \
2617 }
2618
2619 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2620
2621 void
2622 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2623 if (freelistLock()->owned_by_self()) {
2624 Generation::oop_iterate(cl);
2625 } else {
2626 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2627 Generation::oop_iterate(cl);
2628 }
2629 }
2630
2631 void
2632 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2633 if (freelistLock()->owned_by_self()) {
2634 Generation::object_iterate(cl);
2635 } else {
2636 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2637 Generation::object_iterate(cl);
2638 }
2639 }
2640
2641 void
2642 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2643 if (freelistLock()->owned_by_self()) {
2644 Generation::safe_object_iterate(cl);
2645 } else {
2646 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2647 Generation::safe_object_iterate(cl);
2648 }
2649 }
2650
2651 void
2652 ConcurrentMarkSweepGeneration::post_compact() {
2653 }
2654
2655 void
2656 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2657 // Fix the linear allocation blocks to look like free blocks.
2658
2659 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2660 // are not called when the heap is verified during universe initialization and
2661 // at vm shutdown.
2662 if (freelistLock()->owned_by_self()) {
2663 cmsSpace()->prepare_for_verify();
2664 } else {
2665 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2666 cmsSpace()->prepare_for_verify();
2667 }
2668 }
2669
2670 void
2671 ConcurrentMarkSweepGeneration::verify() {
2672 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2673 // are not called when the heap is verified during universe initialization and
2674 // at vm shutdown.
2675 if (freelistLock()->owned_by_self()) {
2676 cmsSpace()->verify();
2677 } else {
2678 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2679 cmsSpace()->verify();
2680 }
2681 }
2682
2683 void CMSCollector::verify() {
2684 _cmsGen->verify();
2685 }
2686
2687 #ifndef PRODUCT
2688 bool CMSCollector::overflow_list_is_empty() const {
2689 assert(_num_par_pushes >= 0, "Inconsistency");
2690 if (_overflow_list == NULL) {
2691 assert(_num_par_pushes == 0, "Inconsistency");
2692 }
2693 return _overflow_list == NULL;
2694 }
2695
2696 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2697 // merely consolidate assertion checks that appear to occur together frequently.
2698 void CMSCollector::verify_work_stacks_empty() const {
2699 assert(_markStack.isEmpty(), "Marking stack should be empty");
2700 assert(overflow_list_is_empty(), "Overflow list should be empty");
2701 }
2702
2703 void CMSCollector::verify_overflow_empty() const {
2704 assert(overflow_list_is_empty(), "Overflow list should be empty");
2705 assert(no_preserved_marks(), "No preserved marks");
2706 }
2707 #endif // PRODUCT
2708
2709 // Decide if we want to enable class unloading as part of the
2710 // ensuing concurrent GC cycle. We will collect and
2711 // unload classes if it's the case that:
2712 // (1) an explicit gc request has been made and the flag
2713 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
2714 // (2) (a) class unloading is enabled at the command line, and
2715 // (b) old gen is getting really full
2716 // NOTE: Provided there is no change in the state of the heap between
2717 // calls to this method, it should have idempotent results. Moreover,
2718 // its results should be monotonically increasing (i.e. going from 0 to 1,
2719 // but not 1 to 0) between successive calls between which the heap was
2720 // not collected. For the implementation below, it must thus rely on
2721 // the property that concurrent_cycles_since_last_unload()
2722 // will not decrease unless a collection cycle happened and that
2723 // _cmsGen->is_too_full() are
2724 // themselves also monotonic in that sense. See check_monotonicity()
2725 // below.
2726 void CMSCollector::update_should_unload_classes() {
2727 _should_unload_classes = false;
2728 // Condition 1 above
2729 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
2730 _should_unload_classes = true;
2731 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
2732 // Disjuncts 2.b.(i,ii,iii) above
2733 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2734 CMSClassUnloadingMaxInterval)
2735 || _cmsGen->is_too_full();
2736 }
2737 }
2738
2739 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2740 bool res = should_concurrent_collect();
2741 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2742 return res;
2743 }
2744
2745 void CMSCollector::setup_cms_unloading_and_verification_state() {
2746 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2747 || VerifyBeforeExit;
2748 const int rso = SharedHeap::SO_AllCodeCache;
2749
2750 // We set the proper root for this CMS cycle here.
2751 if (should_unload_classes()) { // Should unload classes this cycle
2752 remove_root_scanning_option(rso); // Shrink the root set appropriately
2753 set_verifying(should_verify); // Set verification state for this cycle
2754 return; // Nothing else needs to be done at this time
2755 }
2756
2757 // Not unloading classes this cycle
2758 assert(!should_unload_classes(), "Inconsistency!");
2759
2760 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2761 // Include symbols, strings and code cache elements to prevent their resurrection.
2762 add_root_scanning_option(rso);
2763 set_verifying(true);
2764 } else if (verifying() && !should_verify) {
2765 // We were verifying, but some verification flags got disabled.
2766 set_verifying(false);
2767 // Exclude symbols, strings and code cache elements from root scanning to
2768 // reduce IM and RM pauses.
2769 remove_root_scanning_option(rso);
2770 }
2771 }
2772
2773
2774 #ifndef PRODUCT
2775 HeapWord* CMSCollector::block_start(const void* p) const {
2776 const HeapWord* addr = (HeapWord*)p;
2777 if (_span.contains(p)) {
2778 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2779 return _cmsGen->cmsSpace()->block_start(p);
2780 }
2781 }
2782 return NULL;
2783 }
2784 #endif
2785
2786 HeapWord*
2787 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2788 bool tlab,
2789 bool parallel) {
2790 CMSSynchronousYieldRequest yr;
2791 assert(!tlab, "Can't deal with TLAB allocation");
2792 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2793 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2794 if (GCExpandToAllocateDelayMillis > 0) {
2795 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2796 }
2797 return have_lock_and_allocate(word_size, tlab);
2798 }
2799
2800 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2801 size_t bytes,
2802 size_t expand_bytes,
2803 CMSExpansionCause::Cause cause)
2804 {
2805
2806 bool success = expand(bytes, expand_bytes);
2807
2808 // remember why we expanded; this information is used
2809 // by shouldConcurrentCollect() when making decisions on whether to start
2810 // a new CMS cycle.
2811 if (success) {
2812 set_expansion_cause(cause);
2813 if (PrintGCDetails && Verbose) {
2814 gclog_or_tty->print_cr("Expanded CMS gen for %s",
2815 CMSExpansionCause::to_string(cause));
2816 }
2817 }
2818 }
2819
2820 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2821 HeapWord* res = NULL;
2822 MutexLocker x(ParGCRareEvent_lock);
2823 while (true) {
2824 // Expansion by some other thread might make alloc OK now:
2825 res = ps->lab.alloc(word_sz);
2826 if (res != NULL) return res;
2827 // If there's not enough expansion space available, give up.
2828 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2829 return NULL;
2830 }
2831 // Otherwise, we try expansion.
2832 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2833 // Now go around the loop and try alloc again;
2834 // A competing par_promote might beat us to the expansion space,
2835 // so we may go around the loop again if promotion fails again.
2836 if (GCExpandToAllocateDelayMillis > 0) {
2837 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2838 }
2839 }
2840 }
2841
2842
2843 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2844 PromotionInfo* promo) {
2845 MutexLocker x(ParGCRareEvent_lock);
2846 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2847 while (true) {
2848 // Expansion by some other thread might make alloc OK now:
2849 if (promo->ensure_spooling_space()) {
2850 assert(promo->has_spooling_space(),
2851 "Post-condition of successful ensure_spooling_space()");
2852 return true;
2853 }
2854 // If there's not enough expansion space available, give up.
2855 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2856 return false;
2857 }
2858 // Otherwise, we try expansion.
2859 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2860 // Now go around the loop and try alloc again;
2861 // A competing allocation might beat us to the expansion space,
2862 // so we may go around the loop again if allocation fails again.
2863 if (GCExpandToAllocateDelayMillis > 0) {
2864 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2865 }
2866 }
2867 }
2868
2869 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2870 // Only shrink if a compaction was done so that all the free space
2871 // in the generation is in a contiguous block at the end.
2872 if (did_compact()) {
2873 CardGeneration::shrink(bytes);
2874 }
2875 }
2876
2877 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2878 assert_locked_or_safepoint(Heap_lock);
2879 }
2880
2881 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2882 assert_locked_or_safepoint(Heap_lock);
2883 assert_lock_strong(freelistLock());
2884 if (PrintGCDetails && Verbose) {
2885 warning("Shrinking of CMS not yet implemented");
2886 }
2887 return;
2888 }
2889
2890
2891 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2892 // phases.
2893 class CMSPhaseAccounting: public StackObj {
2894 public:
2895 CMSPhaseAccounting(CMSCollector *collector,
2896 const char *phase,
2897 const GCId gc_id,
2898 bool print_cr = true);
2899 ~CMSPhaseAccounting();
2900
2901 private:
2902 CMSCollector *_collector;
2903 const char *_phase;
2904 elapsedTimer _wallclock;
2905 bool _print_cr;
2906 const GCId _gc_id;
2907
2908 public:
2909 // Not MT-safe; so do not pass around these StackObj's
2910 // where they may be accessed by other threads.
2911 jlong wallclock_millis() {
2912 assert(_wallclock.is_active(), "Wall clock should not stop");
2913 _wallclock.stop(); // to record time
2914 jlong ret = _wallclock.milliseconds();
2915 _wallclock.start(); // restart
2916 return ret;
2917 }
2918 };
2919
2920 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2921 const char *phase,
2922 const GCId gc_id,
2923 bool print_cr) :
2924 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) {
2925
2926 if (PrintCMSStatistics != 0) {
2927 _collector->resetYields();
2928 }
2929 if (PrintGCDetails) {
2930 gclog_or_tty->gclog_stamp(_gc_id);
2931 gclog_or_tty->print_cr("[%s-concurrent-%s-start]",
2932 _collector->cmsGen()->short_name(), _phase);
2933 }
2934 _collector->resetTimer();
2935 _wallclock.start();
2936 _collector->startTimer();
2937 }
2938
2939 CMSPhaseAccounting::~CMSPhaseAccounting() {
2940 assert(_wallclock.is_active(), "Wall clock should not have stopped");
2941 _collector->stopTimer();
2942 _wallclock.stop();
2943 if (PrintGCDetails) {
2944 gclog_or_tty->gclog_stamp(_gc_id);
2945 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
2946 _collector->cmsGen()->short_name(),
2947 _phase, _collector->timerValue(), _wallclock.seconds());
2948 if (_print_cr) {
2949 gclog_or_tty->cr();
2950 }
2951 if (PrintCMSStatistics != 0) {
2952 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
2953 _collector->yields());
2954 }
2955 }
2956 }
2957
2958 // CMS work
2959
2960 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2961 class CMSParMarkTask : public AbstractGangTask {
2962 protected:
2963 CMSCollector* _collector;
2964 int _n_workers;
2965 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) :
2966 AbstractGangTask(name),
2967 _collector(collector),
2968 _n_workers(n_workers) {}
2969 // Work method in support of parallel rescan ... of young gen spaces
2970 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
2971 ContiguousSpace* space,
2972 HeapWord** chunk_array, size_t chunk_top);
2973 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
2974 };
2975
2976 // Parallel initial mark task
2977 class CMSParInitialMarkTask: public CMSParMarkTask {
2978 public:
2979 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) :
2980 CMSParMarkTask("Scan roots and young gen for initial mark in parallel",
2981 collector, n_workers) {}
2982 void work(uint worker_id);
2983 };
2984
2985 // Checkpoint the roots into this generation from outside
2986 // this generation. [Note this initial checkpoint need only
2987 // be approximate -- we'll do a catch up phase subsequently.]
2988 void CMSCollector::checkpointRootsInitial() {
2989 assert(_collectorState == InitialMarking, "Wrong collector state");
2990 check_correct_thread_executing();
2991 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
2992
2993 save_heap_summary();
2994 report_heap_summary(GCWhen::BeforeGC);
2995
2996 ReferenceProcessor* rp = ref_processor();
2997 SpecializationStats::clear();
2998 assert(_restart_addr == NULL, "Control point invariant");
2999 {
3000 // acquire locks for subsequent manipulations
3001 MutexLockerEx x(bitMapLock(),
3002 Mutex::_no_safepoint_check_flag);
3003 checkpointRootsInitialWork();
3004 // enable ("weak") refs discovery
3005 rp->enable_discovery();
3006 _collectorState = Marking;
3007 }
3008 SpecializationStats::print();
3009 }
3010
3011 void CMSCollector::checkpointRootsInitialWork() {
3012 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3013 assert(_collectorState == InitialMarking, "just checking");
3014
3015 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3016 // precede our marking with a collection of all
3017 // younger generations to keep floating garbage to a minimum.
3018 // XXX: we won't do this for now -- it's an optimization to be done later.
3019
3020 // already have locks
3021 assert_lock_strong(bitMapLock());
3022 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3023
3024 // Setup the verification and class unloading state for this
3025 // CMS collection cycle.
3026 setup_cms_unloading_and_verification_state();
3027
3028 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
3029 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
3030
3031 // Reset all the PLAB chunk arrays if necessary.
3032 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3033 reset_survivor_plab_arrays();
3034 }
3035
3036 ResourceMark rm;
3037 HandleMark hm;
3038
3039 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
3040 GenCollectedHeap* gch = GenCollectedHeap::heap();
3041
3042 verify_work_stacks_empty();
3043 verify_overflow_empty();
3044
3045 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3046 // Update the saved marks which may affect the root scans.
3047 gch->save_marks();
3048
3049 // weak reference processing has not started yet.
3050 ref_processor()->set_enqueuing_is_done(false);
3051
3052 // Need to remember all newly created CLDs,
3053 // so that we can guarantee that the remark finds them.
3054 ClassLoaderDataGraph::remember_new_clds(true);
3055
3056 // Whenever a CLD is found, it will be claimed before proceeding to mark
3057 // the klasses. The claimed marks need to be cleared before marking starts.
3058 ClassLoaderDataGraph::clear_claimed_marks();
3059
3060 if (CMSPrintEdenSurvivorChunks) {
3061 print_eden_and_survivor_chunk_arrays();
3062 }
3063
3064 {
3065 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3066 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
3067 // The parallel version.
3068 FlexibleWorkGang* workers = gch->workers();
3069 assert(workers != NULL, "Need parallel worker threads.");
3070 int n_workers = workers->active_workers();
3071 CMSParInitialMarkTask tsk(this, n_workers);
3072 gch->set_par_threads(n_workers);
3073 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
3074 if (n_workers > 1) {
3075 GenCollectedHeap::StrongRootsScope srs(gch);
3076 workers->run_task(&tsk);
3077 } else {
3078 GenCollectedHeap::StrongRootsScope srs(gch);
3079 tsk.work(0);
3080 }
3081 gch->set_par_threads(0);
3082 } else {
3083 // The serial version.
3084 CLDToOopClosure cld_closure(¬Older, true);
3085 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3086 gch->gen_process_roots(_cmsGen->level(),
3087 true, // younger gens are roots
3088 true, // activate StrongRootsScope
3089 SharedHeap::ScanningOption(roots_scanning_options()),
3090 should_unload_classes(),
3091 ¬Older,
3092 NULL,
3093 &cld_closure);
3094 }
3095 }
3096
3097 // Clear mod-union table; it will be dirtied in the prologue of
3098 // CMS generation per each younger generation collection.
3099
3100 assert(_modUnionTable.isAllClear(),
3101 "Was cleared in most recent final checkpoint phase"
3102 " or no bits are set in the gc_prologue before the start of the next "
3103 "subsequent marking phase.");
3104
3105 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
3106
3107 // Save the end of the used_region of the constituent generations
3108 // to be used to limit the extent of sweep in each generation.
3109 save_sweep_limits();
3110 verify_overflow_empty();
3111 }
3112
3113 bool CMSCollector::markFromRoots() {
3114 // we might be tempted to assert that:
3115 // assert(!SafepointSynchronize::is_at_safepoint(),
3116 // "inconsistent argument?");
3117 // However that wouldn't be right, because it's possible that
3118 // a safepoint is indeed in progress as a younger generation
3119 // stop-the-world GC happens even as we mark in this generation.
3120 assert(_collectorState == Marking, "inconsistent state?");
3121 check_correct_thread_executing();
3122 verify_overflow_empty();
3123
3124 // Weak ref discovery note: We may be discovering weak
3125 // refs in this generation concurrent (but interleaved) with
3126 // weak ref discovery by a younger generation collector.
3127
3128 CMSTokenSyncWithLocks ts(true, bitMapLock());
3129 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3130 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3131 bool res = markFromRootsWork();
3132 if (res) {
3133 _collectorState = Precleaning;
3134 } else { // We failed and a foreground collection wants to take over
3135 assert(_foregroundGCIsActive, "internal state inconsistency");
3136 assert(_restart_addr == NULL, "foreground will restart from scratch");
3137 if (PrintGCDetails) {
3138 gclog_or_tty->print_cr("bailing out to foreground collection");
3139 }
3140 }
3141 verify_overflow_empty();
3142 return res;
3143 }
3144
3145 bool CMSCollector::markFromRootsWork() {
3146 // iterate over marked bits in bit map, doing a full scan and mark
3147 // from these roots using the following algorithm:
3148 // . if oop is to the right of the current scan pointer,
3149 // mark corresponding bit (we'll process it later)
3150 // . else (oop is to left of current scan pointer)
3151 // push oop on marking stack
3152 // . drain the marking stack
3153
3154 // Note that when we do a marking step we need to hold the
3155 // bit map lock -- recall that direct allocation (by mutators)
3156 // and promotion (by younger generation collectors) is also
3157 // marking the bit map. [the so-called allocate live policy.]
3158 // Because the implementation of bit map marking is not
3159 // robust wrt simultaneous marking of bits in the same word,
3160 // we need to make sure that there is no such interference
3161 // between concurrent such updates.
3162
3163 // already have locks
3164 assert_lock_strong(bitMapLock());
3165
3166 verify_work_stacks_empty();
3167 verify_overflow_empty();
3168 bool result = false;
3169 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3170 result = do_marking_mt();
3171 } else {
3172 result = do_marking_st();
3173 }
3174 return result;
3175 }
3176
3177 // Forward decl
3178 class CMSConcMarkingTask;
3179
3180 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3181 CMSCollector* _collector;
3182 CMSConcMarkingTask* _task;
3183 public:
3184 virtual void yield();
3185
3186 // "n_threads" is the number of threads to be terminated.
3187 // "queue_set" is a set of work queues of other threads.
3188 // "collector" is the CMS collector associated with this task terminator.
3189 // "yield" indicates whether we need the gang as a whole to yield.
3190 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3191 ParallelTaskTerminator(n_threads, queue_set),
3192 _collector(collector) { }
3193
3194 void set_task(CMSConcMarkingTask* task) {
3195 _task = task;
3196 }
3197 };
3198
3199 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3200 CMSConcMarkingTask* _task;
3201 public:
3202 bool should_exit_termination();
3203 void set_task(CMSConcMarkingTask* task) {
3204 _task = task;
3205 }
3206 };
3207
3208 // MT Concurrent Marking Task
3209 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3210 CMSCollector* _collector;
3211 int _n_workers; // requested/desired # workers
3212 bool _result;
3213 CompactibleFreeListSpace* _cms_space;
3214 char _pad_front[64]; // padding to ...
3215 HeapWord* _global_finger; // ... avoid sharing cache line
3216 char _pad_back[64];
3217 HeapWord* _restart_addr;
3218
3219 // Exposed here for yielding support
3220 Mutex* const _bit_map_lock;
3221
3222 // The per thread work queues, available here for stealing
3223 OopTaskQueueSet* _task_queues;
3224
3225 // Termination (and yielding) support
3226 CMSConcMarkingTerminator _term;
3227 CMSConcMarkingTerminatorTerminator _term_term;
3228
3229 public:
3230 CMSConcMarkingTask(CMSCollector* collector,
3231 CompactibleFreeListSpace* cms_space,
3232 YieldingFlexibleWorkGang* workers,
3233 OopTaskQueueSet* task_queues):
3234 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3235 _collector(collector),
3236 _cms_space(cms_space),
3237 _n_workers(0), _result(true),
3238 _task_queues(task_queues),
3239 _term(_n_workers, task_queues, _collector),
3240 _bit_map_lock(collector->bitMapLock())
3241 {
3242 _requested_size = _n_workers;
3243 _term.set_task(this);
3244 _term_term.set_task(this);
3245 _restart_addr = _global_finger = _cms_space->bottom();
3246 }
3247
3248
3249 OopTaskQueueSet* task_queues() { return _task_queues; }
3250
3251 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3252
3253 HeapWord** global_finger_addr() { return &_global_finger; }
3254
3255 CMSConcMarkingTerminator* terminator() { return &_term; }
3256
3257 virtual void set_for_termination(int active_workers) {
3258 terminator()->reset_for_reuse(active_workers);
3259 }
3260
3261 void work(uint worker_id);
3262 bool should_yield() {
3263 return ConcurrentMarkSweepThread::should_yield()
3264 && !_collector->foregroundGCIsActive();
3265 }
3266
3267 virtual void coordinator_yield(); // stuff done by coordinator
3268 bool result() { return _result; }
3269
3270 void reset(HeapWord* ra) {
3271 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3272 _restart_addr = _global_finger = ra;
3273 _term.reset_for_reuse();
3274 }
3275
3276 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3277 OopTaskQueue* work_q);
3278
3279 private:
3280 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3281 void do_work_steal(int i);
3282 void bump_global_finger(HeapWord* f);
3283 };
3284
3285 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3286 assert(_task != NULL, "Error");
3287 return _task->yielding();
3288 // Note that we do not need the disjunct || _task->should_yield() above
3289 // because we want terminating threads to yield only if the task
3290 // is already in the midst of yielding, which happens only after at least one
3291 // thread has yielded.
3292 }
3293
3294 void CMSConcMarkingTerminator::yield() {
3295 if (_task->should_yield()) {
3296 _task->yield();
3297 } else {
3298 ParallelTaskTerminator::yield();
3299 }
3300 }
3301
3302 ////////////////////////////////////////////////////////////////
3303 // Concurrent Marking Algorithm Sketch
3304 ////////////////////////////////////////////////////////////////
3305 // Until all tasks exhausted (both spaces):
3306 // -- claim next available chunk
3307 // -- bump global finger via CAS
3308 // -- find first object that starts in this chunk
3309 // and start scanning bitmap from that position
3310 // -- scan marked objects for oops
3311 // -- CAS-mark target, and if successful:
3312 // . if target oop is above global finger (volatile read)
3313 // nothing to do
3314 // . if target oop is in chunk and above local finger
3315 // then nothing to do
3316 // . else push on work-queue
3317 // -- Deal with possible overflow issues:
3318 // . local work-queue overflow causes stuff to be pushed on
3319 // global (common) overflow queue
3320 // . always first empty local work queue
3321 // . then get a batch of oops from global work queue if any
3322 // . then do work stealing
3323 // -- When all tasks claimed (both spaces)
3324 // and local work queue empty,
3325 // then in a loop do:
3326 // . check global overflow stack; steal a batch of oops and trace
3327 // . try to steal from other threads oif GOS is empty
3328 // . if neither is available, offer termination
3329 // -- Terminate and return result
3330 //
3331 void CMSConcMarkingTask::work(uint worker_id) {
3332 elapsedTimer _timer;
3333 ResourceMark rm;
3334 HandleMark hm;
3335
3336 DEBUG_ONLY(_collector->verify_overflow_empty();)
3337
3338 // Before we begin work, our work queue should be empty
3339 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3340 // Scan the bitmap covering _cms_space, tracing through grey objects.
3341 _timer.start();
3342 do_scan_and_mark(worker_id, _cms_space);
3343 _timer.stop();
3344 if (PrintCMSStatistics != 0) {
3345 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3346 worker_id, _timer.seconds());
3347 // XXX: need xxx/xxx type of notation, two timers
3348 }
3349
3350 // ... do work stealing
3351 _timer.reset();
3352 _timer.start();
3353 do_work_steal(worker_id);
3354 _timer.stop();
3355 if (PrintCMSStatistics != 0) {
3356 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3357 worker_id, _timer.seconds());
3358 // XXX: need xxx/xxx type of notation, two timers
3359 }
3360 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3361 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3362 // Note that under the current task protocol, the
3363 // following assertion is true even of the spaces
3364 // expanded since the completion of the concurrent
3365 // marking. XXX This will likely change under a strict
3366 // ABORT semantics.
3367 // After perm removal the comparison was changed to
3368 // greater than or equal to from strictly greater than.
3369 // Before perm removal the highest address sweep would
3370 // have been at the end of perm gen but now is at the
3371 // end of the tenured gen.
3372 assert(_global_finger >= _cms_space->end(),
3373 "All tasks have been completed");
3374 DEBUG_ONLY(_collector->verify_overflow_empty();)
3375 }
3376
3377 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3378 HeapWord* read = _global_finger;
3379 HeapWord* cur = read;
3380 while (f > read) {
3381 cur = read;
3382 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3383 if (cur == read) {
3384 // our cas succeeded
3385 assert(_global_finger >= f, "protocol consistency");
3386 break;
3387 }
3388 }
3389 }
3390
3391 // This is really inefficient, and should be redone by
3392 // using (not yet available) block-read and -write interfaces to the
3393 // stack and the work_queue. XXX FIX ME !!!
3394 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3395 OopTaskQueue* work_q) {
3396 // Fast lock-free check
3397 if (ovflw_stk->length() == 0) {
3398 return false;
3399 }
3400 assert(work_q->size() == 0, "Shouldn't steal");
3401 MutexLockerEx ml(ovflw_stk->par_lock(),
3402 Mutex::_no_safepoint_check_flag);
3403 // Grab up to 1/4 the size of the work queue
3404 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3405 (size_t)ParGCDesiredObjsFromOverflowList);
3406 num = MIN2(num, ovflw_stk->length());
3407 for (int i = (int) num; i > 0; i--) {
3408 oop cur = ovflw_stk->pop();
3409 assert(cur != NULL, "Counted wrong?");
3410 work_q->push(cur);
3411 }
3412 return num > 0;
3413 }
3414
3415 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3416 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3417 int n_tasks = pst->n_tasks();
3418 // We allow that there may be no tasks to do here because
3419 // we are restarting after a stack overflow.
3420 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3421 uint nth_task = 0;
3422
3423 HeapWord* aligned_start = sp->bottom();
3424 if (sp->used_region().contains(_restart_addr)) {
3425 // Align down to a card boundary for the start of 0th task
3426 // for this space.
3427 aligned_start =
3428 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3429 CardTableModRefBS::card_size);
3430 }
3431
3432 size_t chunk_size = sp->marking_task_size();
3433 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3434 // Having claimed the nth task in this space,
3435 // compute the chunk that it corresponds to:
3436 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3437 aligned_start + (nth_task+1)*chunk_size);
3438 // Try and bump the global finger via a CAS;
3439 // note that we need to do the global finger bump
3440 // _before_ taking the intersection below, because
3441 // the task corresponding to that region will be
3442 // deemed done even if the used_region() expands
3443 // because of allocation -- as it almost certainly will
3444 // during start-up while the threads yield in the
3445 // closure below.
3446 HeapWord* finger = span.end();
3447 bump_global_finger(finger); // atomically
3448 // There are null tasks here corresponding to chunks
3449 // beyond the "top" address of the space.
3450 span = span.intersection(sp->used_region());
3451 if (!span.is_empty()) { // Non-null task
3452 HeapWord* prev_obj;
3453 assert(!span.contains(_restart_addr) || nth_task == 0,
3454 "Inconsistency");
3455 if (nth_task == 0) {
3456 // For the 0th task, we'll not need to compute a block_start.
3457 if (span.contains(_restart_addr)) {
3458 // In the case of a restart because of stack overflow,
3459 // we might additionally skip a chunk prefix.
3460 prev_obj = _restart_addr;
3461 } else {
3462 prev_obj = span.start();
3463 }
3464 } else {
3465 // We want to skip the first object because
3466 // the protocol is to scan any object in its entirety
3467 // that _starts_ in this span; a fortiori, any
3468 // object starting in an earlier span is scanned
3469 // as part of an earlier claimed task.
3470 // Below we use the "careful" version of block_start
3471 // so we do not try to navigate uninitialized objects.
3472 prev_obj = sp->block_start_careful(span.start());
3473 // Below we use a variant of block_size that uses the
3474 // Printezis bits to avoid waiting for allocated
3475 // objects to become initialized/parsable.
3476 while (prev_obj < span.start()) {
3477 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3478 if (sz > 0) {
3479 prev_obj += sz;
3480 } else {
3481 // In this case we may end up doing a bit of redundant
3482 // scanning, but that appears unavoidable, short of
3483 // locking the free list locks; see bug 6324141.
3484 break;
3485 }
3486 }
3487 }
3488 if (prev_obj < span.end()) {
3489 MemRegion my_span = MemRegion(prev_obj, span.end());
3490 // Do the marking work within a non-empty span --
3491 // the last argument to the constructor indicates whether the
3492 // iteration should be incremental with periodic yields.
3493 Par_MarkFromRootsClosure cl(this, _collector, my_span,
3494 &_collector->_markBitMap,
3495 work_queue(i),
3496 &_collector->_markStack);
3497 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3498 } // else nothing to do for this task
3499 } // else nothing to do for this task
3500 }
3501 // We'd be tempted to assert here that since there are no
3502 // more tasks left to claim in this space, the global_finger
3503 // must exceed space->top() and a fortiori space->end(). However,
3504 // that would not quite be correct because the bumping of
3505 // global_finger occurs strictly after the claiming of a task,
3506 // so by the time we reach here the global finger may not yet
3507 // have been bumped up by the thread that claimed the last
3508 // task.
3509 pst->all_tasks_completed();
3510 }
3511
3512 class Par_ConcMarkingClosure: public MetadataAwareOopClosure {
3513 private:
3514 CMSCollector* _collector;
3515 CMSConcMarkingTask* _task;
3516 MemRegion _span;
3517 CMSBitMap* _bit_map;
3518 CMSMarkStack* _overflow_stack;
3519 OopTaskQueue* _work_queue;
3520 protected:
3521 DO_OOP_WORK_DEFN
3522 public:
3523 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3524 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3525 MetadataAwareOopClosure(collector->ref_processor()),
3526 _collector(collector),
3527 _task(task),
3528 _span(collector->_span),
3529 _work_queue(work_queue),
3530 _bit_map(bit_map),
3531 _overflow_stack(overflow_stack)
3532 { }
3533 virtual void do_oop(oop* p);
3534 virtual void do_oop(narrowOop* p);
3535
3536 void trim_queue(size_t max);
3537 void handle_stack_overflow(HeapWord* lost);
3538 void do_yield_check() {
3539 if (_task->should_yield()) {
3540 _task->yield();
3541 }
3542 }
3543 };
3544
3545 // Grey object scanning during work stealing phase --
3546 // the salient assumption here is that any references
3547 // that are in these stolen objects being scanned must
3548 // already have been initialized (else they would not have
3549 // been published), so we do not need to check for
3550 // uninitialized objects before pushing here.
3551 void Par_ConcMarkingClosure::do_oop(oop obj) {
3552 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
3553 HeapWord* addr = (HeapWord*)obj;
3554 // Check if oop points into the CMS generation
3555 // and is not marked
3556 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3557 // a white object ...
3558 // If we manage to "claim" the object, by being the
3559 // first thread to mark it, then we push it on our
3560 // marking stack
3561 if (_bit_map->par_mark(addr)) { // ... now grey
3562 // push on work queue (grey set)
3563 bool simulate_overflow = false;
3564 NOT_PRODUCT(
3565 if (CMSMarkStackOverflowALot &&
3566 _collector->simulate_overflow()) {
3567 // simulate a stack overflow
3568 simulate_overflow = true;
3569 }
3570 )
3571 if (simulate_overflow ||
3572 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3573 // stack overflow
3574 if (PrintCMSStatistics != 0) {
3575 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
3576 SIZE_FORMAT, _overflow_stack->capacity());
3577 }
3578 // We cannot assert that the overflow stack is full because
3579 // it may have been emptied since.
3580 assert(simulate_overflow ||
3581 _work_queue->size() == _work_queue->max_elems(),
3582 "Else push should have succeeded");
3583 handle_stack_overflow(addr);
3584 }
3585 } // Else, some other thread got there first
3586 do_yield_check();
3587 }
3588 }
3589
3590 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3591 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3592
3593 void Par_ConcMarkingClosure::trim_queue(size_t max) {
3594 while (_work_queue->size() > max) {
3595 oop new_oop;
3596 if (_work_queue->pop_local(new_oop)) {
3597 assert(new_oop->is_oop(), "Should be an oop");
3598 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3599 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3600 new_oop->oop_iterate(this); // do_oop() above
3601 do_yield_check();
3602 }
3603 }
3604 }
3605
3606 // Upon stack overflow, we discard (part of) the stack,
3607 // remembering the least address amongst those discarded
3608 // in CMSCollector's _restart_address.
3609 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3610 // We need to do this under a mutex to prevent other
3611 // workers from interfering with the work done below.
3612 MutexLockerEx ml(_overflow_stack->par_lock(),
3613 Mutex::_no_safepoint_check_flag);
3614 // Remember the least grey address discarded
3615 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3616 _collector->lower_restart_addr(ra);
3617 _overflow_stack->reset(); // discard stack contents
3618 _overflow_stack->expand(); // expand the stack if possible
3619 }
3620
3621
3622 void CMSConcMarkingTask::do_work_steal(int i) {
3623 OopTaskQueue* work_q = work_queue(i);
3624 oop obj_to_scan;
3625 CMSBitMap* bm = &(_collector->_markBitMap);
3626 CMSMarkStack* ovflw = &(_collector->_markStack);
3627 int* seed = _collector->hash_seed(i);
3628 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3629 while (true) {
3630 cl.trim_queue(0);
3631 assert(work_q->size() == 0, "Should have been emptied above");
3632 if (get_work_from_overflow_stack(ovflw, work_q)) {
3633 // Can't assert below because the work obtained from the
3634 // overflow stack may already have been stolen from us.
3635 // assert(work_q->size() > 0, "Work from overflow stack");
3636 continue;
3637 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3638 assert(obj_to_scan->is_oop(), "Should be an oop");
3639 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3640 obj_to_scan->oop_iterate(&cl);
3641 } else if (terminator()->offer_termination(&_term_term)) {
3642 assert(work_q->size() == 0, "Impossible!");
3643 break;
3644 } else if (yielding() || should_yield()) {
3645 yield();
3646 }
3647 }
3648 }
3649
3650 // This is run by the CMS (coordinator) thread.
3651 void CMSConcMarkingTask::coordinator_yield() {
3652 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3653 "CMS thread should hold CMS token");
3654 // First give up the locks, then yield, then re-lock
3655 // We should probably use a constructor/destructor idiom to
3656 // do this unlock/lock or modify the MutexUnlocker class to
3657 // serve our purpose. XXX
3658 assert_lock_strong(_bit_map_lock);
3659 _bit_map_lock->unlock();
3660 ConcurrentMarkSweepThread::desynchronize(true);
3661 _collector->stopTimer();
3662 if (PrintCMSStatistics != 0) {
3663 _collector->incrementYields();
3664 }
3665
3666 // It is possible for whichever thread initiated the yield request
3667 // not to get a chance to wake up and take the bitmap lock between
3668 // this thread releasing it and reacquiring it. So, while the
3669 // should_yield() flag is on, let's sleep for a bit to give the
3670 // other thread a chance to wake up. The limit imposed on the number
3671 // of iterations is defensive, to avoid any unforseen circumstances
3672 // putting us into an infinite loop. Since it's always been this
3673 // (coordinator_yield()) method that was observed to cause the
3674 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3675 // which is by default non-zero. For the other seven methods that
3676 // also perform the yield operation, as are using a different
3677 // parameter (CMSYieldSleepCount) which is by default zero. This way we
3678 // can enable the sleeping for those methods too, if necessary.
3679 // See 6442774.
3680 //
3681 // We really need to reconsider the synchronization between the GC
3682 // thread and the yield-requesting threads in the future and we
3683 // should really use wait/notify, which is the recommended
3684 // way of doing this type of interaction. Additionally, we should
3685 // consolidate the eight methods that do the yield operation and they
3686 // are almost identical into one for better maintainability and
3687 // readability. See 6445193.
3688 //
3689 // Tony 2006.06.29
3690 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3691 ConcurrentMarkSweepThread::should_yield() &&
3692 !CMSCollector::foregroundGCIsActive(); ++i) {
3693 os::sleep(Thread::current(), 1, false);
3694 }
3695
3696 ConcurrentMarkSweepThread::synchronize(true);
3697 _bit_map_lock->lock_without_safepoint_check();
3698 _collector->startTimer();
3699 }
3700
3701 bool CMSCollector::do_marking_mt() {
3702 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3703 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers(
3704 conc_workers()->total_workers(),
3705 conc_workers()->active_workers(),
3706 Threads::number_of_non_daemon_threads());
3707 conc_workers()->set_active_workers(num_workers);
3708
3709 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
3710
3711 CMSConcMarkingTask tsk(this,
3712 cms_space,
3713 conc_workers(),
3714 task_queues());
3715
3716 // Since the actual number of workers we get may be different
3717 // from the number we requested above, do we need to do anything different
3718 // below? In particular, may be we need to subclass the SequantialSubTasksDone
3719 // class?? XXX
3720 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3721
3722 // Refs discovery is already non-atomic.
3723 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3724 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3725 conc_workers()->start_task(&tsk);
3726 while (tsk.yielded()) {
3727 tsk.coordinator_yield();
3728 conc_workers()->continue_task(&tsk);
3729 }
3730 // If the task was aborted, _restart_addr will be non-NULL
3731 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3732 while (_restart_addr != NULL) {
3733 // XXX For now we do not make use of ABORTED state and have not
3734 // yet implemented the right abort semantics (even in the original
3735 // single-threaded CMS case). That needs some more investigation
3736 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3737 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3738 // If _restart_addr is non-NULL, a marking stack overflow
3739 // occurred; we need to do a fresh marking iteration from the
3740 // indicated restart address.
3741 if (_foregroundGCIsActive) {
3742 // We may be running into repeated stack overflows, having
3743 // reached the limit of the stack size, while making very
3744 // slow forward progress. It may be best to bail out and
3745 // let the foreground collector do its job.
3746 // Clear _restart_addr, so that foreground GC
3747 // works from scratch. This avoids the headache of
3748 // a "rescan" which would otherwise be needed because
3749 // of the dirty mod union table & card table.
3750 _restart_addr = NULL;
3751 return false;
3752 }
3753 // Adjust the task to restart from _restart_addr
3754 tsk.reset(_restart_addr);
3755 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3756 _restart_addr);
3757 _restart_addr = NULL;
3758 // Get the workers going again
3759 conc_workers()->start_task(&tsk);
3760 while (tsk.yielded()) {
3761 tsk.coordinator_yield();
3762 conc_workers()->continue_task(&tsk);
3763 }
3764 }
3765 assert(tsk.completed(), "Inconsistency");
3766 assert(tsk.result() == true, "Inconsistency");
3767 return true;
3768 }
3769
3770 bool CMSCollector::do_marking_st() {
3771 ResourceMark rm;
3772 HandleMark hm;
3773
3774 // Temporarily make refs discovery single threaded (non-MT)
3775 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3776 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3777 &_markStack, CMSYield);
3778 // the last argument to iterate indicates whether the iteration
3779 // should be incremental with periodic yields.
3780 _markBitMap.iterate(&markFromRootsClosure);
3781 // If _restart_addr is non-NULL, a marking stack overflow
3782 // occurred; we need to do a fresh iteration from the
3783 // indicated restart address.
3784 while (_restart_addr != NULL) {
3785 if (_foregroundGCIsActive) {
3786 // We may be running into repeated stack overflows, having
3787 // reached the limit of the stack size, while making very
3788 // slow forward progress. It may be best to bail out and
3789 // let the foreground collector do its job.
3790 // Clear _restart_addr, so that foreground GC
3791 // works from scratch. This avoids the headache of
3792 // a "rescan" which would otherwise be needed because
3793 // of the dirty mod union table & card table.
3794 _restart_addr = NULL;
3795 return false; // indicating failure to complete marking
3796 }
3797 // Deal with stack overflow:
3798 // we restart marking from _restart_addr
3799 HeapWord* ra = _restart_addr;
3800 markFromRootsClosure.reset(ra);
3801 _restart_addr = NULL;
3802 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3803 }
3804 return true;
3805 }
3806
3807 void CMSCollector::preclean() {
3808 check_correct_thread_executing();
3809 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3810 verify_work_stacks_empty();
3811 verify_overflow_empty();
3812 _abort_preclean = false;
3813 if (CMSPrecleaningEnabled) {
3814 if (!CMSEdenChunksRecordAlways) {
3815 _eden_chunk_index = 0;
3816 }
3817 size_t used = get_eden_used();
3818 size_t capacity = get_eden_capacity();
3819 // Don't start sampling unless we will get sufficiently
3820 // many samples.
3821 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
3822 * CMSScheduleRemarkEdenPenetration)) {
3823 _start_sampling = true;
3824 } else {
3825 _start_sampling = false;
3826 }
3827 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3828 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3829 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3830 }
3831 CMSTokenSync x(true); // is cms thread
3832 if (CMSPrecleaningEnabled) {
3833 sample_eden();
3834 _collectorState = AbortablePreclean;
3835 } else {
3836 _collectorState = FinalMarking;
3837 }
3838 verify_work_stacks_empty();
3839 verify_overflow_empty();
3840 }
3841
3842 // Try and schedule the remark such that young gen
3843 // occupancy is CMSScheduleRemarkEdenPenetration %.
3844 void CMSCollector::abortable_preclean() {
3845 check_correct_thread_executing();
3846 assert(CMSPrecleaningEnabled, "Inconsistent control state");
3847 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3848
3849 // If Eden's current occupancy is below this threshold,
3850 // immediately schedule the remark; else preclean
3851 // past the next scavenge in an effort to
3852 // schedule the pause as described above. By choosing
3853 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3854 // we will never do an actual abortable preclean cycle.
3855 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3856 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3857 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3858 // We need more smarts in the abortable preclean
3859 // loop below to deal with cases where allocation
3860 // in young gen is very very slow, and our precleaning
3861 // is running a losing race against a horde of
3862 // mutators intent on flooding us with CMS updates
3863 // (dirty cards).
3864 // One, admittedly dumb, strategy is to give up
3865 // after a certain number of abortable precleaning loops
3866 // or after a certain maximum time. We want to make
3867 // this smarter in the next iteration.
3868 // XXX FIX ME!!! YSR
3869 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3870 while (!(should_abort_preclean() ||
3871 ConcurrentMarkSweepThread::should_terminate())) {
3872 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3873 cumworkdone += workdone;
3874 loops++;
3875 // Voluntarily terminate abortable preclean phase if we have
3876 // been at it for too long.
3877 if ((CMSMaxAbortablePrecleanLoops != 0) &&
3878 loops >= CMSMaxAbortablePrecleanLoops) {
3879 if (PrintGCDetails) {
3880 gclog_or_tty->print(" CMS: abort preclean due to loops ");
3881 }
3882 break;
3883 }
3884 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3885 if (PrintGCDetails) {
3886 gclog_or_tty->print(" CMS: abort preclean due to time ");
3887 }
3888 break;
3889 }
3890 // If we are doing little work each iteration, we should
3891 // take a short break.
3892 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3893 // Sleep for some time, waiting for work to accumulate
3894 stopTimer();
3895 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3896 startTimer();
3897 waited++;
3898 }
3899 }
3900 if (PrintCMSStatistics > 0) {
3901 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3902 loops, waited, cumworkdone);
3903 }
3904 }
3905 CMSTokenSync x(true); // is cms thread
3906 if (_collectorState != Idling) {
3907 assert(_collectorState == AbortablePreclean,
3908 "Spontaneous state transition?");
3909 _collectorState = FinalMarking;
3910 } // Else, a foreground collection completed this CMS cycle.
3911 return;
3912 }
3913
3914 // Respond to an Eden sampling opportunity
3915 void CMSCollector::sample_eden() {
3916 // Make sure a young gc cannot sneak in between our
3917 // reading and recording of a sample.
3918 assert(Thread::current()->is_ConcurrentGC_thread(),
3919 "Only the cms thread may collect Eden samples");
3920 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3921 "Should collect samples while holding CMS token");
3922 if (!_start_sampling) {
3923 return;
3924 }
3925 // When CMSEdenChunksRecordAlways is true, the eden chunk array
3926 // is populated by the young generation.
3927 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3928 if (_eden_chunk_index < _eden_chunk_capacity) {
3929 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
3930 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3931 "Unexpected state of Eden");
3932 // We'd like to check that what we just sampled is an oop-start address;
3933 // however, we cannot do that here since the object may not yet have been
3934 // initialized. So we'll instead do the check when we _use_ this sample
3935 // later.
3936 if (_eden_chunk_index == 0 ||
3937 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3938 _eden_chunk_array[_eden_chunk_index-1])
3939 >= CMSSamplingGrain)) {
3940 _eden_chunk_index++; // commit sample
3941 }
3942 }
3943 }
3944 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3945 size_t used = get_eden_used();
3946 size_t capacity = get_eden_capacity();
3947 assert(used <= capacity, "Unexpected state of Eden");
3948 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3949 _abort_preclean = true;
3950 }
3951 }
3952 }
3953
3954
3955 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3956 assert(_collectorState == Precleaning ||
3957 _collectorState == AbortablePreclean, "incorrect state");
3958 ResourceMark rm;
3959 HandleMark hm;
3960
3961 // Precleaning is currently not MT but the reference processor
3962 // may be set for MT. Disable it temporarily here.
3963 ReferenceProcessor* rp = ref_processor();
3964 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3965
3966 // Do one pass of scrubbing the discovered reference lists
3967 // to remove any reference objects with strongly-reachable
3968 // referents.
3969 if (clean_refs) {
3970 CMSPrecleanRefsYieldClosure yield_cl(this);
3971 assert(rp->span().equals(_span), "Spans should be equal");
3972 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3973 &_markStack, true /* preclean */);
3974 CMSDrainMarkingStackClosure complete_trace(this,
3975 _span, &_markBitMap, &_markStack,
3976 &keep_alive, true /* preclean */);
3977
3978 // We don't want this step to interfere with a young
3979 // collection because we don't want to take CPU
3980 // or memory bandwidth away from the young GC threads
3981 // (which may be as many as there are CPUs).
3982 // Note that we don't need to protect ourselves from
3983 // interference with mutators because they can't
3984 // manipulate the discovered reference lists nor affect
3985 // the computed reachability of the referents, the
3986 // only properties manipulated by the precleaning
3987 // of these reference lists.
3988 stopTimer();
3989 CMSTokenSyncWithLocks x(true /* is cms thread */,
3990 bitMapLock());
3991 startTimer();
3992 sample_eden();
3993
3994 // The following will yield to allow foreground
3995 // collection to proceed promptly. XXX YSR:
3996 // The code in this method may need further
3997 // tweaking for better performance and some restructuring
3998 // for cleaner interfaces.
3999 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
4000 rp->preclean_discovered_references(
4001 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
4002 gc_timer, _gc_tracer_cm->gc_id());
4003 }
4004
4005 if (clean_survivor) { // preclean the active survivor space(s)
4006 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4007 &_markBitMap, &_modUnionTable,
4008 &_markStack, true /* precleaning phase */);
4009 stopTimer();
4010 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4011 bitMapLock());
4012 startTimer();
4013 unsigned int before_count =
4014 GenCollectedHeap::heap()->total_collections();
4015 SurvivorSpacePrecleanClosure
4016 sss_cl(this, _span, &_markBitMap, &_markStack,
4017 &pam_cl, before_count, CMSYield);
4018 _young_gen->from()->object_iterate_careful(&sss_cl);
4019 _young_gen->to()->object_iterate_careful(&sss_cl);
4020 }
4021 MarkRefsIntoAndScanClosure
4022 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4023 &_markStack, this, CMSYield,
4024 true /* precleaning phase */);
4025 // CAUTION: The following closure has persistent state that may need to
4026 // be reset upon a decrease in the sequence of addresses it
4027 // processes.
4028 ScanMarkedObjectsAgainCarefullyClosure
4029 smoac_cl(this, _span,
4030 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
4031
4032 // Preclean dirty cards in ModUnionTable and CardTable using
4033 // appropriate convergence criterion;
4034 // repeat CMSPrecleanIter times unless we find that
4035 // we are losing.
4036 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4037 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4038 "Bad convergence multiplier");
4039 assert(CMSPrecleanThreshold >= 100,
4040 "Unreasonably low CMSPrecleanThreshold");
4041
4042 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4043 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4044 numIter < CMSPrecleanIter;
4045 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4046 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4047 if (Verbose && PrintGCDetails) {
4048 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
4049 }
4050 // Either there are very few dirty cards, so re-mark
4051 // pause will be small anyway, or our pre-cleaning isn't
4052 // that much faster than the rate at which cards are being
4053 // dirtied, so we might as well stop and re-mark since
4054 // precleaning won't improve our re-mark time by much.
4055 if (curNumCards <= CMSPrecleanThreshold ||
4056 (numIter > 0 &&
4057 (curNumCards * CMSPrecleanDenominator >
4058 lastNumCards * CMSPrecleanNumerator))) {
4059 numIter++;
4060 cumNumCards += curNumCards;
4061 break;
4062 }
4063 }
4064
4065 preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
4066
4067 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4068 cumNumCards += curNumCards;
4069 if (PrintGCDetails && PrintCMSStatistics != 0) {
4070 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
4071 curNumCards, cumNumCards, numIter);
4072 }
4073 return cumNumCards; // as a measure of useful work done
4074 }
4075
4076 // PRECLEANING NOTES:
4077 // Precleaning involves:
4078 // . reading the bits of the modUnionTable and clearing the set bits.
4079 // . For the cards corresponding to the set bits, we scan the
4080 // objects on those cards. This means we need the free_list_lock
4081 // so that we can safely iterate over the CMS space when scanning
4082 // for oops.
4083 // . When we scan the objects, we'll be both reading and setting
4084 // marks in the marking bit map, so we'll need the marking bit map.
4085 // . For protecting _collector_state transitions, we take the CGC_lock.
4086 // Note that any races in the reading of of card table entries by the
4087 // CMS thread on the one hand and the clearing of those entries by the
4088 // VM thread or the setting of those entries by the mutator threads on the
4089 // other are quite benign. However, for efficiency it makes sense to keep
4090 // the VM thread from racing with the CMS thread while the latter is
4091 // dirty card info to the modUnionTable. We therefore also use the
4092 // CGC_lock to protect the reading of the card table and the mod union
4093 // table by the CM thread.
4094 // . We run concurrently with mutator updates, so scanning
4095 // needs to be done carefully -- we should not try to scan
4096 // potentially uninitialized objects.
4097 //
4098 // Locking strategy: While holding the CGC_lock, we scan over and
4099 // reset a maximal dirty range of the mod union / card tables, then lock
4100 // the free_list_lock and bitmap lock to do a full marking, then
4101 // release these locks; and repeat the cycle. This allows for a
4102 // certain amount of fairness in the sharing of these locks between
4103 // the CMS collector on the one hand, and the VM thread and the
4104 // mutators on the other.
4105
4106 // NOTE: preclean_mod_union_table() and preclean_card_table()
4107 // further below are largely identical; if you need to modify
4108 // one of these methods, please check the other method too.
4109
4110 size_t CMSCollector::preclean_mod_union_table(
4111 ConcurrentMarkSweepGeneration* gen,
4112 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4113 verify_work_stacks_empty();
4114 verify_overflow_empty();
4115
4116 // strategy: starting with the first card, accumulate contiguous
4117 // ranges of dirty cards; clear these cards, then scan the region
4118 // covered by these cards.
4119
4120 // Since all of the MUT is committed ahead, we can just use
4121 // that, in case the generations expand while we are precleaning.
4122 // It might also be fine to just use the committed part of the
4123 // generation, but we might potentially miss cards when the
4124 // generation is rapidly expanding while we are in the midst
4125 // of precleaning.
4126 HeapWord* startAddr = gen->reserved().start();
4127 HeapWord* endAddr = gen->reserved().end();
4128
4129 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4130
4131 size_t numDirtyCards, cumNumDirtyCards;
4132 HeapWord *nextAddr, *lastAddr;
4133 for (cumNumDirtyCards = numDirtyCards = 0,
4134 nextAddr = lastAddr = startAddr;
4135 nextAddr < endAddr;
4136 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4137
4138 ResourceMark rm;
4139 HandleMark hm;
4140
4141 MemRegion dirtyRegion;
4142 {
4143 stopTimer();
4144 // Potential yield point
4145 CMSTokenSync ts(true);
4146 startTimer();
4147 sample_eden();
4148 // Get dirty region starting at nextOffset (inclusive),
4149 // simultaneously clearing it.
4150 dirtyRegion =
4151 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4152 assert(dirtyRegion.start() >= nextAddr,
4153 "returned region inconsistent?");
4154 }
4155 // Remember where the next search should begin.
4156 // The returned region (if non-empty) is a right open interval,
4157 // so lastOffset is obtained from the right end of that
4158 // interval.
4159 lastAddr = dirtyRegion.end();
4160 // Should do something more transparent and less hacky XXX
4161 numDirtyCards =
4162 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4163
4164 // We'll scan the cards in the dirty region (with periodic
4165 // yields for foreground GC as needed).
4166 if (!dirtyRegion.is_empty()) {
4167 assert(numDirtyCards > 0, "consistency check");
4168 HeapWord* stop_point = NULL;
4169 stopTimer();
4170 // Potential yield point
4171 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4172 bitMapLock());
4173 startTimer();
4174 {
4175 verify_work_stacks_empty();
4176 verify_overflow_empty();
4177 sample_eden();
4178 stop_point =
4179 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4180 }
4181 if (stop_point != NULL) {
4182 // The careful iteration stopped early either because it found an
4183 // uninitialized object, or because we were in the midst of an
4184 // "abortable preclean", which should now be aborted. Redirty
4185 // the bits corresponding to the partially-scanned or unscanned
4186 // cards. We'll either restart at the next block boundary or
4187 // abort the preclean.
4188 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4189 "Should only be AbortablePreclean.");
4190 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4191 if (should_abort_preclean()) {
4192 break; // out of preclean loop
4193 } else {
4194 // Compute the next address at which preclean should pick up;
4195 // might need bitMapLock in order to read P-bits.
4196 lastAddr = next_card_start_after_block(stop_point);
4197 }
4198 }
4199 } else {
4200 assert(lastAddr == endAddr, "consistency check");
4201 assert(numDirtyCards == 0, "consistency check");
4202 break;
4203 }
4204 }
4205 verify_work_stacks_empty();
4206 verify_overflow_empty();
4207 return cumNumDirtyCards;
4208 }
4209
4210 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4211 // below are largely identical; if you need to modify
4212 // one of these methods, please check the other method too.
4213
4214 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4215 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4216 // strategy: it's similar to precleamModUnionTable above, in that
4217 // we accumulate contiguous ranges of dirty cards, mark these cards
4218 // precleaned, then scan the region covered by these cards.
4219 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4220 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4221
4222 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4223
4224 size_t numDirtyCards, cumNumDirtyCards;
4225 HeapWord *lastAddr, *nextAddr;
4226
4227 for (cumNumDirtyCards = numDirtyCards = 0,
4228 nextAddr = lastAddr = startAddr;
4229 nextAddr < endAddr;
4230 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4231
4232 ResourceMark rm;
4233 HandleMark hm;
4234
4235 MemRegion dirtyRegion;
4236 {
4237 // See comments in "Precleaning notes" above on why we
4238 // do this locking. XXX Could the locking overheads be
4239 // too high when dirty cards are sparse? [I don't think so.]
4240 stopTimer();
4241 CMSTokenSync x(true); // is cms thread
4242 startTimer();
4243 sample_eden();
4244 // Get and clear dirty region from card table
4245 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4246 MemRegion(nextAddr, endAddr),
4247 true,
4248 CardTableModRefBS::precleaned_card_val());
4249
4250 assert(dirtyRegion.start() >= nextAddr,
4251 "returned region inconsistent?");
4252 }
4253 lastAddr = dirtyRegion.end();
4254 numDirtyCards =
4255 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4256
4257 if (!dirtyRegion.is_empty()) {
4258 stopTimer();
4259 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4260 startTimer();
4261 sample_eden();
4262 verify_work_stacks_empty();
4263 verify_overflow_empty();
4264 HeapWord* stop_point =
4265 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4266 if (stop_point != NULL) {
4267 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4268 "Should only be AbortablePreclean.");
4269 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4270 if (should_abort_preclean()) {
4271 break; // out of preclean loop
4272 } else {
4273 // Compute the next address at which preclean should pick up.
4274 lastAddr = next_card_start_after_block(stop_point);
4275 }
4276 }
4277 } else {
4278 break;
4279 }
4280 }
4281 verify_work_stacks_empty();
4282 verify_overflow_empty();
4283 return cumNumDirtyCards;
4284 }
4285
4286 class PrecleanKlassClosure : public KlassClosure {
4287 KlassToOopClosure _cm_klass_closure;
4288 public:
4289 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4290 void do_klass(Klass* k) {
4291 if (k->has_accumulated_modified_oops()) {
4292 k->clear_accumulated_modified_oops();
4293
4294 _cm_klass_closure.do_klass(k);
4295 }
4296 }
4297 };
4298
4299 // The freelist lock is needed to prevent asserts, is it really needed?
4300 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4301
4302 cl->set_freelistLock(freelistLock);
4303
4304 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4305
4306 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4307 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4308 PrecleanKlassClosure preclean_klass_closure(cl);
4309 ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
4310
4311 verify_work_stacks_empty();
4312 verify_overflow_empty();
4313 }
4314
4315 void CMSCollector::checkpointRootsFinal() {
4316 assert(_collectorState == FinalMarking, "incorrect state transition?");
4317 check_correct_thread_executing();
4318 // world is stopped at this checkpoint
4319 assert(SafepointSynchronize::is_at_safepoint(),
4320 "world should be stopped");
4321 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4322
4323 verify_work_stacks_empty();
4324 verify_overflow_empty();
4325
4326 SpecializationStats::clear();
4327 if (PrintGCDetails) {
4328 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4329 _young_gen->used() / K,
4330 _young_gen->capacity() / K);
4331 }
4332 {
4333 if (CMSScavengeBeforeRemark) {
4334 GenCollectedHeap* gch = GenCollectedHeap::heap();
4335 // Temporarily set flag to false, GCH->do_collection will
4336 // expect it to be false and set to true
4337 FlagSetting fl(gch->_is_gc_active, false);
4338 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
4339 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4340 int level = _cmsGen->level() - 1;
4341 if (level >= 0) {
4342 gch->do_collection(true, // full (i.e. force, see below)
4343 false, // !clear_all_soft_refs
4344 0, // size
4345 false, // is_tlab
4346 level // max_level
4347 );
4348 }
4349 }
4350 FreelistLocker x(this);
4351 MutexLockerEx y(bitMapLock(),
4352 Mutex::_no_safepoint_check_flag);
4353 checkpointRootsFinalWork();
4354 }
4355 verify_work_stacks_empty();
4356 verify_overflow_empty();
4357 SpecializationStats::print();
4358 }
4359
4360 void CMSCollector::checkpointRootsFinalWork() {
4361 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4362
4363 assert(haveFreelistLocks(), "must have free list locks");
4364 assert_lock_strong(bitMapLock());
4365
4366 ResourceMark rm;
4367 HandleMark hm;
4368
4369 GenCollectedHeap* gch = GenCollectedHeap::heap();
4370
4371 if (should_unload_classes()) {
4372 CodeCache::gc_prologue();
4373 }
4374 assert(haveFreelistLocks(), "must have free list locks");
4375 assert_lock_strong(bitMapLock());
4376
4377 // We might assume that we need not fill TLAB's when
4378 // CMSScavengeBeforeRemark is set, because we may have just done
4379 // a scavenge which would have filled all TLAB's -- and besides
4380 // Eden would be empty. This however may not always be the case --
4381 // for instance although we asked for a scavenge, it may not have
4382 // happened because of a JNI critical section. We probably need
4383 // a policy for deciding whether we can in that case wait until
4384 // the critical section releases and then do the remark following
4385 // the scavenge, and skip it here. In the absence of that policy,
4386 // or of an indication of whether the scavenge did indeed occur,
4387 // we cannot rely on TLAB's having been filled and must do
4388 // so here just in case a scavenge did not happen.
4389 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4390 // Update the saved marks which may affect the root scans.
4391 gch->save_marks();
4392
4393 if (CMSPrintEdenSurvivorChunks) {
4394 print_eden_and_survivor_chunk_arrays();
4395 }
4396
4397 {
4398 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4399
4400 // Note on the role of the mod union table:
4401 // Since the marker in "markFromRoots" marks concurrently with
4402 // mutators, it is possible for some reachable objects not to have been
4403 // scanned. For instance, an only reference to an object A was
4404 // placed in object B after the marker scanned B. Unless B is rescanned,
4405 // A would be collected. Such updates to references in marked objects
4406 // are detected via the mod union table which is the set of all cards
4407 // dirtied since the first checkpoint in this GC cycle and prior to
4408 // the most recent young generation GC, minus those cleaned up by the
4409 // concurrent precleaning.
4410 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
4411 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
4412 do_remark_parallel();
4413 } else {
4414 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4415 _gc_timer_cm, _gc_tracer_cm->gc_id());
4416 do_remark_non_parallel();
4417 }
4418 }
4419 verify_work_stacks_empty();
4420 verify_overflow_empty();
4421
4422 {
4423 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4424 refProcessingWork();
4425 }
4426 verify_work_stacks_empty();
4427 verify_overflow_empty();
4428
4429 if (should_unload_classes()) {
4430 CodeCache::gc_epilogue();
4431 }
4432 JvmtiExport::gc_epilogue();
4433
4434 // If we encountered any (marking stack / work queue) overflow
4435 // events during the current CMS cycle, take appropriate
4436 // remedial measures, where possible, so as to try and avoid
4437 // recurrence of that condition.
4438 assert(_markStack.isEmpty(), "No grey objects");
4439 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4440 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4441 if (ser_ovflw > 0) {
4442 if (PrintCMSStatistics != 0) {
4443 gclog_or_tty->print_cr("Marking stack overflow (benign) "
4444 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
4445 ", kac_preclean="SIZE_FORMAT")",
4446 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4447 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4448 }
4449 _markStack.expand();
4450 _ser_pmc_remark_ovflw = 0;
4451 _ser_pmc_preclean_ovflw = 0;
4452 _ser_kac_preclean_ovflw = 0;
4453 _ser_kac_ovflw = 0;
4454 }
4455 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4456 if (PrintCMSStatistics != 0) {
4457 gclog_or_tty->print_cr("Work queue overflow (benign) "
4458 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4459 _par_pmc_remark_ovflw, _par_kac_ovflw);
4460 }
4461 _par_pmc_remark_ovflw = 0;
4462 _par_kac_ovflw = 0;
4463 }
4464 if (PrintCMSStatistics != 0) {
4465 if (_markStack._hit_limit > 0) {
4466 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
4467 _markStack._hit_limit);
4468 }
4469 if (_markStack._failed_double > 0) {
4470 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
4471 " current capacity "SIZE_FORMAT,
4472 _markStack._failed_double,
4473 _markStack.capacity());
4474 }
4475 }
4476 _markStack._hit_limit = 0;
4477 _markStack._failed_double = 0;
4478
4479 if ((VerifyAfterGC || VerifyDuringGC) &&
4480 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4481 verify_after_remark();
4482 }
4483
4484 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4485
4486 // Change under the freelistLocks.
4487 _collectorState = Sweeping;
4488 // Call isAllClear() under bitMapLock
4489 assert(_modUnionTable.isAllClear(),
4490 "Should be clear by end of the final marking");
4491 assert(_ct->klass_rem_set()->mod_union_is_clear(),
4492 "Should be clear by end of the final marking");
4493 }
4494
4495 void CMSParInitialMarkTask::work(uint worker_id) {
4496 elapsedTimer _timer;
4497 ResourceMark rm;
4498 HandleMark hm;
4499
4500 // ---------- scan from roots --------------
4501 _timer.start();
4502 GenCollectedHeap* gch = GenCollectedHeap::heap();
4503 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4504
4505 // ---------- young gen roots --------------
4506 {
4507 work_on_young_gen_roots(worker_id, &par_mri_cl);
4508 _timer.stop();
4509 if (PrintCMSStatistics != 0) {
4510 gclog_or_tty->print_cr(
4511 "Finished young gen initial mark scan work in %dth thread: %3.3f sec",
4512 worker_id, _timer.seconds());
4513 }
4514 }
4515
4516 // ---------- remaining roots --------------
4517 _timer.reset();
4518 _timer.start();
4519
4520 CLDToOopClosure cld_closure(&par_mri_cl, true);
4521
4522 gch->gen_process_roots(_collector->_cmsGen->level(),
4523 false, // yg was scanned above
4524 false, // this is parallel code
4525 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4526 _collector->should_unload_classes(),
4527 &par_mri_cl,
4528 NULL,
4529 &cld_closure);
4530 assert(_collector->should_unload_classes()
4531 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache),
4532 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4533 _timer.stop();
4534 if (PrintCMSStatistics != 0) {
4535 gclog_or_tty->print_cr(
4536 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec",
4537 worker_id, _timer.seconds());
4538 }
4539 }
4540
4541 // Parallel remark task
4542 class CMSParRemarkTask: public CMSParMarkTask {
4543 CompactibleFreeListSpace* _cms_space;
4544
4545 // The per-thread work queues, available here for stealing.
4546 OopTaskQueueSet* _task_queues;
4547 ParallelTaskTerminator _term;
4548
4549 public:
4550 // A value of 0 passed to n_workers will cause the number of
4551 // workers to be taken from the active workers in the work gang.
4552 CMSParRemarkTask(CMSCollector* collector,
4553 CompactibleFreeListSpace* cms_space,
4554 int n_workers, FlexibleWorkGang* workers,
4555 OopTaskQueueSet* task_queues):
4556 CMSParMarkTask("Rescan roots and grey objects in parallel",
4557 collector, n_workers),
4558 _cms_space(cms_space),
4559 _task_queues(task_queues),
4560 _term(n_workers, task_queues) { }
4561
4562 OopTaskQueueSet* task_queues() { return _task_queues; }
4563
4564 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4565
4566 ParallelTaskTerminator* terminator() { return &_term; }
4567 int n_workers() { return _n_workers; }
4568
4569 void work(uint worker_id);
4570
4571 private:
4572 // ... of dirty cards in old space
4573 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4574 Par_MarkRefsIntoAndScanClosure* cl);
4575
4576 // ... work stealing for the above
4577 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4578 };
4579
4580 class RemarkKlassClosure : public KlassClosure {
4581 KlassToOopClosure _cm_klass_closure;
4582 public:
4583 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4584 void do_klass(Klass* k) {
4585 // Check if we have modified any oops in the Klass during the concurrent marking.
4586 if (k->has_accumulated_modified_oops()) {
4587 k->clear_accumulated_modified_oops();
4588
4589 // We could have transfered the current modified marks to the accumulated marks,
4590 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4591 } else if (k->has_modified_oops()) {
4592 // Don't clear anything, this info is needed by the next young collection.
4593 } else {
4594 // No modified oops in the Klass.
4595 return;
4596 }
4597
4598 // The klass has modified fields, need to scan the klass.
4599 _cm_klass_closure.do_klass(k);
4600 }
4601 };
4602
4603 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
4604 ParNewGeneration* young_gen = _collector->_young_gen;
4605 ContiguousSpace* eden_space = young_gen->eden();
4606 ContiguousSpace* from_space = young_gen->from();
4607 ContiguousSpace* to_space = young_gen->to();
4608
4609 HeapWord** eca = _collector->_eden_chunk_array;
4610 size_t ect = _collector->_eden_chunk_index;
4611 HeapWord** sca = _collector->_survivor_chunk_array;
4612 size_t sct = _collector->_survivor_chunk_index;
4613
4614 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4615 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4616
4617 do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
4618 do_young_space_rescan(worker_id, cl, from_space, sca, sct);
4619 do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
4620 }
4621
4622 // work_queue(i) is passed to the closure
4623 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
4624 // also is passed to do_dirty_card_rescan_tasks() and to
4625 // do_work_steal() to select the i-th task_queue.
4626
4627 void CMSParRemarkTask::work(uint worker_id) {
4628 elapsedTimer _timer;
4629 ResourceMark rm;
4630 HandleMark hm;
4631
4632 // ---------- rescan from roots --------------
4633 _timer.start();
4634 GenCollectedHeap* gch = GenCollectedHeap::heap();
4635 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4636 _collector->_span, _collector->ref_processor(),
4637 &(_collector->_markBitMap),
4638 work_queue(worker_id));
4639
4640 // Rescan young gen roots first since these are likely
4641 // coarsely partitioned and may, on that account, constitute
4642 // the critical path; thus, it's best to start off that
4643 // work first.
4644 // ---------- young gen roots --------------
4645 {
4646 work_on_young_gen_roots(worker_id, &par_mrias_cl);
4647 _timer.stop();
4648 if (PrintCMSStatistics != 0) {
4649 gclog_or_tty->print_cr(
4650 "Finished young gen rescan work in %dth thread: %3.3f sec",
4651 worker_id, _timer.seconds());
4652 }
4653 }
4654
4655 // ---------- remaining roots --------------
4656 _timer.reset();
4657 _timer.start();
4658 gch->gen_process_roots(_collector->_cmsGen->level(),
4659 false, // yg was scanned above
4660 false, // this is parallel code
4661 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4662 _collector->should_unload_classes(),
4663 &par_mrias_cl,
4664 NULL,
4665 NULL); // The dirty klasses will be handled below
4666
4667 assert(_collector->should_unload_classes()
4668 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache),
4669 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4670 _timer.stop();
4671 if (PrintCMSStatistics != 0) {
4672 gclog_or_tty->print_cr(
4673 "Finished remaining root rescan work in %dth thread: %3.3f sec",
4674 worker_id, _timer.seconds());
4675 }
4676
4677 // ---------- unhandled CLD scanning ----------
4678 if (worker_id == 0) { // Single threaded at the moment.
4679 _timer.reset();
4680 _timer.start();
4681
4682 // Scan all new class loader data objects and new dependencies that were
4683 // introduced during concurrent marking.
4684 ResourceMark rm;
4685 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4686 for (int i = 0; i < array->length(); i++) {
4687 par_mrias_cl.do_class_loader_data(array->at(i));
4688 }
4689
4690 // We don't need to keep track of new CLDs anymore.
4691 ClassLoaderDataGraph::remember_new_clds(false);
4692
4693 _timer.stop();
4694 if (PrintCMSStatistics != 0) {
4695 gclog_or_tty->print_cr(
4696 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec",
4697 worker_id, _timer.seconds());
4698 }
4699 }
4700
4701 // ---------- dirty klass scanning ----------
4702 if (worker_id == 0) { // Single threaded at the moment.
4703 _timer.reset();
4704 _timer.start();
4705
4706 // Scan all classes that was dirtied during the concurrent marking phase.
4707 RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
4708 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
4709
4710 _timer.stop();
4711 if (PrintCMSStatistics != 0) {
4712 gclog_or_tty->print_cr(
4713 "Finished dirty klass scanning work in %dth thread: %3.3f sec",
4714 worker_id, _timer.seconds());
4715 }
4716 }
4717
4718 // We might have added oops to ClassLoaderData::_handles during the
4719 // concurrent marking phase. These oops point to newly allocated objects
4720 // that are guaranteed to be kept alive. Either by the direct allocation
4721 // code, or when the young collector processes the roots. Hence,
4722 // we don't have to revisit the _handles block during the remark phase.
4723
4724 // ---------- rescan dirty cards ------------
4725 _timer.reset();
4726 _timer.start();
4727
4728 // Do the rescan tasks for each of the two spaces
4729 // (cms_space) in turn.
4730 // "worker_id" is passed to select the task_queue for "worker_id"
4731 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4732 _timer.stop();
4733 if (PrintCMSStatistics != 0) {
4734 gclog_or_tty->print_cr(
4735 "Finished dirty card rescan work in %dth thread: %3.3f sec",
4736 worker_id, _timer.seconds());
4737 }
4738
4739 // ---------- steal work from other threads ...
4740 // ---------- ... and drain overflow list.
4741 _timer.reset();
4742 _timer.start();
4743 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4744 _timer.stop();
4745 if (PrintCMSStatistics != 0) {
4746 gclog_or_tty->print_cr(
4747 "Finished work stealing in %dth thread: %3.3f sec",
4748 worker_id, _timer.seconds());
4749 }
4750 }
4751
4752 // Note that parameter "i" is not used.
4753 void
4754 CMSParMarkTask::do_young_space_rescan(uint worker_id,
4755 OopsInGenClosure* cl, ContiguousSpace* space,
4756 HeapWord** chunk_array, size_t chunk_top) {
4757 // Until all tasks completed:
4758 // . claim an unclaimed task
4759 // . compute region boundaries corresponding to task claimed
4760 // using chunk_array
4761 // . par_oop_iterate(cl) over that region
4762
4763 ResourceMark rm;
4764 HandleMark hm;
4765
4766 SequentialSubTasksDone* pst = space->par_seq_tasks();
4767
4768 uint nth_task = 0;
4769 uint n_tasks = pst->n_tasks();
4770
4771 if (n_tasks > 0) {
4772 assert(pst->valid(), "Uninitialized use?");
4773 HeapWord *start, *end;
4774 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4775 // We claimed task # nth_task; compute its boundaries.
4776 if (chunk_top == 0) { // no samples were taken
4777 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4778 start = space->bottom();
4779 end = space->top();
4780 } else if (nth_task == 0) {
4781 start = space->bottom();
4782 end = chunk_array[nth_task];
4783 } else if (nth_task < (uint)chunk_top) {
4784 assert(nth_task >= 1, "Control point invariant");
4785 start = chunk_array[nth_task - 1];
4786 end = chunk_array[nth_task];
4787 } else {
4788 assert(nth_task == (uint)chunk_top, "Control point invariant");
4789 start = chunk_array[chunk_top - 1];
4790 end = space->top();
4791 }
4792 MemRegion mr(start, end);
4793 // Verify that mr is in space
4794 assert(mr.is_empty() || space->used_region().contains(mr),
4795 "Should be in space");
4796 // Verify that "start" is an object boundary
4797 assert(mr.is_empty() || oop(mr.start())->is_oop(),
4798 "Should be an oop");
4799 space->par_oop_iterate(mr, cl);
4800 }
4801 pst->all_tasks_completed();
4802 }
4803 }
4804
4805 void
4806 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4807 CompactibleFreeListSpace* sp, int i,
4808 Par_MarkRefsIntoAndScanClosure* cl) {
4809 // Until all tasks completed:
4810 // . claim an unclaimed task
4811 // . compute region boundaries corresponding to task claimed
4812 // . transfer dirty bits ct->mut for that region
4813 // . apply rescanclosure to dirty mut bits for that region
4814
4815 ResourceMark rm;
4816 HandleMark hm;
4817
4818 OopTaskQueue* work_q = work_queue(i);
4819 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4820 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4821 // CAUTION: This closure has state that persists across calls to
4822 // the work method dirty_range_iterate_clear() in that it has
4823 // embedded in it a (subtype of) UpwardsObjectClosure. The
4824 // use of that state in the embedded UpwardsObjectClosure instance
4825 // assumes that the cards are always iterated (even if in parallel
4826 // by several threads) in monotonically increasing order per each
4827 // thread. This is true of the implementation below which picks
4828 // card ranges (chunks) in monotonically increasing order globally
4829 // and, a-fortiori, in monotonically increasing order per thread
4830 // (the latter order being a subsequence of the former).
4831 // If the work code below is ever reorganized into a more chaotic
4832 // work-partitioning form than the current "sequential tasks"
4833 // paradigm, the use of that persistent state will have to be
4834 // revisited and modified appropriately. See also related
4835 // bug 4756801 work on which should examine this code to make
4836 // sure that the changes there do not run counter to the
4837 // assumptions made here and necessary for correctness and
4838 // efficiency. Note also that this code might yield inefficient
4839 // behavior in the case of very large objects that span one or
4840 // more work chunks. Such objects would potentially be scanned
4841 // several times redundantly. Work on 4756801 should try and
4842 // address that performance anomaly if at all possible. XXX
4843 MemRegion full_span = _collector->_span;
4844 CMSBitMap* bm = &(_collector->_markBitMap); // shared
4845 MarkFromDirtyCardsClosure
4846 greyRescanClosure(_collector, full_span, // entire span of interest
4847 sp, bm, work_q, cl);
4848
4849 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4850 assert(pst->valid(), "Uninitialized use?");
4851 uint nth_task = 0;
4852 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
4853 MemRegion span = sp->used_region();
4854 HeapWord* start_addr = span.start();
4855 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
4856 alignment);
4857 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4858 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
4859 start_addr, "Check alignment");
4860 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
4861 chunk_size, "Check alignment");
4862
4863 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4864 // Having claimed the nth_task, compute corresponding mem-region,
4865 // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4866 // The alignment restriction ensures that we do not need any
4867 // synchronization with other gang-workers while setting or
4868 // clearing bits in thus chunk of the MUT.
4869 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4870 start_addr + (nth_task+1)*chunk_size);
4871 // The last chunk's end might be way beyond end of the
4872 // used region. In that case pull back appropriately.
4873 if (this_span.end() > end_addr) {
4874 this_span.set_end(end_addr);
4875 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4876 }
4877 // Iterate over the dirty cards covering this chunk, marking them
4878 // precleaned, and setting the corresponding bits in the mod union
4879 // table. Since we have been careful to partition at Card and MUT-word
4880 // boundaries no synchronization is needed between parallel threads.
4881 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
4882 &modUnionClosure);
4883
4884 // Having transferred these marks into the modUnionTable,
4885 // rescan the marked objects on the dirty cards in the modUnionTable.
4886 // Even if this is at a synchronous collection, the initial marking
4887 // may have been done during an asynchronous collection so there
4888 // may be dirty bits in the mod-union table.
4889 _collector->_modUnionTable.dirty_range_iterate_clear(
4890 this_span, &greyRescanClosure);
4891 _collector->_modUnionTable.verifyNoOneBitsInRange(
4892 this_span.start(),
4893 this_span.end());
4894 }
4895 pst->all_tasks_completed(); // declare that i am done
4896 }
4897
4898 // . see if we can share work_queues with ParNew? XXX
4899 void
4900 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
4901 int* seed) {
4902 OopTaskQueue* work_q = work_queue(i);
4903 NOT_PRODUCT(int num_steals = 0;)
4904 oop obj_to_scan;
4905 CMSBitMap* bm = &(_collector->_markBitMap);
4906
4907 while (true) {
4908 // Completely finish any left over work from (an) earlier round(s)
4909 cl->trim_queue(0);
4910 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4911 (size_t)ParGCDesiredObjsFromOverflowList);
4912 // Now check if there's any work in the overflow list
4913 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4914 // only affects the number of attempts made to get work from the
4915 // overflow list and does not affect the number of workers. Just
4916 // pass ParallelGCThreads so this behavior is unchanged.
4917 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4918 work_q,
4919 ParallelGCThreads)) {
4920 // found something in global overflow list;
4921 // not yet ready to go stealing work from others.
4922 // We'd like to assert(work_q->size() != 0, ...)
4923 // because we just took work from the overflow list,
4924 // but of course we can't since all of that could have
4925 // been already stolen from us.
4926 // "He giveth and He taketh away."
4927 continue;
4928 }
4929 // Verify that we have no work before we resort to stealing
4930 assert(work_q->size() == 0, "Have work, shouldn't steal");
4931 // Try to steal from other queues that have work
4932 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4933 NOT_PRODUCT(num_steals++;)
4934 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
4935 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4936 // Do scanning work
4937 obj_to_scan->oop_iterate(cl);
4938 // Loop around, finish this work, and try to steal some more
4939 } else if (terminator()->offer_termination()) {
4940 break; // nirvana from the infinite cycle
4941 }
4942 }
4943 NOT_PRODUCT(
4944 if (PrintCMSStatistics != 0) {
4945 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
4946 }
4947 )
4948 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4949 "Else our work is not yet done");
4950 }
4951
4952 // Record object boundaries in _eden_chunk_array by sampling the eden
4953 // top in the slow-path eden object allocation code path and record
4954 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4955 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4956 // sampling in sample_eden() that activates during the part of the
4957 // preclean phase.
4958 void CMSCollector::sample_eden_chunk() {
4959 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4960 if (_eden_chunk_lock->try_lock()) {
4961 // Record a sample. This is the critical section. The contents
4962 // of the _eden_chunk_array have to be non-decreasing in the
4963 // address order.
4964 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4965 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4966 "Unexpected state of Eden");
4967 if (_eden_chunk_index == 0 ||
4968 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4969 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4970 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4971 _eden_chunk_index++; // commit sample
4972 }
4973 _eden_chunk_lock->unlock();
4974 }
4975 }
4976 }
4977
4978 // Return a thread-local PLAB recording array, as appropriate.
4979 void* CMSCollector::get_data_recorder(int thr_num) {
4980 if (_survivor_plab_array != NULL &&
4981 (CMSPLABRecordAlways ||
4982 (_collectorState > Marking && _collectorState < FinalMarking))) {
4983 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4984 ChunkArray* ca = &_survivor_plab_array[thr_num];
4985 ca->reset(); // clear it so that fresh data is recorded
4986 return (void*) ca;
4987 } else {
4988 return NULL;
4989 }
4990 }
4991
4992 // Reset all the thread-local PLAB recording arrays
4993 void CMSCollector::reset_survivor_plab_arrays() {
4994 for (uint i = 0; i < ParallelGCThreads; i++) {
4995 _survivor_plab_array[i].reset();
4996 }
4997 }
4998
4999 // Merge the per-thread plab arrays into the global survivor chunk
5000 // array which will provide the partitioning of the survivor space
5001 // for CMS initial scan and rescan.
5002 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
5003 int no_of_gc_threads) {
5004 assert(_survivor_plab_array != NULL, "Error");
5005 assert(_survivor_chunk_array != NULL, "Error");
5006 assert(_collectorState == FinalMarking ||
5007 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
5008 for (int j = 0; j < no_of_gc_threads; j++) {
5009 _cursor[j] = 0;
5010 }
5011 HeapWord* top = surv->top();
5012 size_t i;
5013 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
5014 HeapWord* min_val = top; // Higher than any PLAB address
5015 uint min_tid = 0; // position of min_val this round
5016 for (int j = 0; j < no_of_gc_threads; j++) {
5017 ChunkArray* cur_sca = &_survivor_plab_array[j];
5018 if (_cursor[j] == cur_sca->end()) {
5019 continue;
5020 }
5021 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5022 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5023 assert(surv->used_region().contains(cur_val), "Out of bounds value");
5024 if (cur_val < min_val) {
5025 min_tid = j;
5026 min_val = cur_val;
5027 } else {
5028 assert(cur_val < top, "All recorded addresses should be less");
5029 }
5030 }
5031 // At this point min_val and min_tid are respectively
5032 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5033 // and the thread (j) that witnesses that address.
5034 // We record this address in the _survivor_chunk_array[i]
5035 // and increment _cursor[min_tid] prior to the next round i.
5036 if (min_val == top) {
5037 break;
5038 }
5039 _survivor_chunk_array[i] = min_val;
5040 _cursor[min_tid]++;
5041 }
5042 // We are all done; record the size of the _survivor_chunk_array
5043 _survivor_chunk_index = i; // exclusive: [0, i)
5044 if (PrintCMSStatistics > 0) {
5045 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5046 }
5047 // Verify that we used up all the recorded entries
5048 #ifdef ASSERT
5049 size_t total = 0;
5050 for (int j = 0; j < no_of_gc_threads; j++) {
5051 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5052 total += _cursor[j];
5053 }
5054 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5055 // Check that the merged array is in sorted order
5056 if (total > 0) {
5057 for (size_t i = 0; i < total - 1; i++) {
5058 if (PrintCMSStatistics > 0) {
5059 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5060 i, _survivor_chunk_array[i]);
5061 }
5062 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5063 "Not sorted");
5064 }
5065 }
5066 #endif // ASSERT
5067 }
5068
5069 // Set up the space's par_seq_tasks structure for work claiming
5070 // for parallel initial scan and rescan of young gen.
5071 // See ParRescanTask where this is currently used.
5072 void
5073 CMSCollector::
5074 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5075 assert(n_threads > 0, "Unexpected n_threads argument");
5076
5077 // Eden space
5078 if (!_young_gen->eden()->is_empty()) {
5079 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
5080 assert(!pst->valid(), "Clobbering existing data?");
5081 // Each valid entry in [0, _eden_chunk_index) represents a task.
5082 size_t n_tasks = _eden_chunk_index + 1;
5083 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5084 // Sets the condition for completion of the subtask (how many threads
5085 // need to finish in order to be done).
5086 pst->set_n_threads(n_threads);
5087 pst->set_n_tasks((int)n_tasks);
5088 }
5089
5090 // Merge the survivor plab arrays into _survivor_chunk_array
5091 if (_survivor_plab_array != NULL) {
5092 merge_survivor_plab_arrays(_young_gen->from(), n_threads);
5093 } else {
5094 assert(_survivor_chunk_index == 0, "Error");
5095 }
5096
5097 // To space
5098 {
5099 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
5100 assert(!pst->valid(), "Clobbering existing data?");
5101 // Sets the condition for completion of the subtask (how many threads
5102 // need to finish in order to be done).
5103 pst->set_n_threads(n_threads);
5104 pst->set_n_tasks(1);
5105 assert(pst->valid(), "Error");
5106 }
5107
5108 // From space
5109 {
5110 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
5111 assert(!pst->valid(), "Clobbering existing data?");
5112 size_t n_tasks = _survivor_chunk_index + 1;
5113 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5114 // Sets the condition for completion of the subtask (how many threads
5115 // need to finish in order to be done).
5116 pst->set_n_threads(n_threads);
5117 pst->set_n_tasks((int)n_tasks);
5118 assert(pst->valid(), "Error");
5119 }
5120 }
5121
5122 // Parallel version of remark
5123 void CMSCollector::do_remark_parallel() {
5124 GenCollectedHeap* gch = GenCollectedHeap::heap();
5125 FlexibleWorkGang* workers = gch->workers();
5126 assert(workers != NULL, "Need parallel worker threads.");
5127 // Choose to use the number of GC workers most recently set
5128 // into "active_workers". If active_workers is not set, set it
5129 // to ParallelGCThreads.
5130 int n_workers = workers->active_workers();
5131 if (n_workers == 0) {
5132 assert(n_workers > 0, "Should have been set during scavenge");
5133 n_workers = ParallelGCThreads;
5134 workers->set_active_workers(n_workers);
5135 }
5136 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5137
5138 CMSParRemarkTask tsk(this,
5139 cms_space,
5140 n_workers, workers, task_queues());
5141
5142 // Set up for parallel process_roots work.
5143 gch->set_par_threads(n_workers);
5144 // We won't be iterating over the cards in the card table updating
5145 // the younger_gen cards, so we shouldn't call the following else
5146 // the verification code as well as subsequent younger_refs_iterate
5147 // code would get confused. XXX
5148 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5149
5150 // The young gen rescan work will not be done as part of
5151 // process_roots (which currently doesn't know how to
5152 // parallelize such a scan), but rather will be broken up into
5153 // a set of parallel tasks (via the sampling that the [abortable]
5154 // preclean phase did of eden, plus the [two] tasks of
5155 // scanning the [two] survivor spaces. Further fine-grain
5156 // parallelization of the scanning of the survivor spaces
5157 // themselves, and of precleaning of the younger gen itself
5158 // is deferred to the future.
5159 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5160
5161 // The dirty card rescan work is broken up into a "sequence"
5162 // of parallel tasks (per constituent space) that are dynamically
5163 // claimed by the parallel threads.
5164 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5165
5166 // It turns out that even when we're using 1 thread, doing the work in a
5167 // separate thread causes wide variance in run times. We can't help this
5168 // in the multi-threaded case, but we special-case n=1 here to get
5169 // repeatable measurements of the 1-thread overhead of the parallel code.
5170 if (n_workers > 1) {
5171 // Make refs discovery MT-safe, if it isn't already: it may not
5172 // necessarily be so, since it's possible that we are doing
5173 // ST marking.
5174 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5175 GenCollectedHeap::StrongRootsScope srs(gch);
5176 workers->run_task(&tsk);
5177 } else {
5178 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5179 GenCollectedHeap::StrongRootsScope srs(gch);
5180 tsk.work(0);
5181 }
5182
5183 gch->set_par_threads(0); // 0 ==> non-parallel.
5184 // restore, single-threaded for now, any preserved marks
5185 // as a result of work_q overflow
5186 restore_preserved_marks_if_any();
5187 }
5188
5189 // Non-parallel version of remark
5190 void CMSCollector::do_remark_non_parallel() {
5191 ResourceMark rm;
5192 HandleMark hm;
5193 GenCollectedHeap* gch = GenCollectedHeap::heap();
5194 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5195
5196 MarkRefsIntoAndScanClosure
5197 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
5198 &_markStack, this,
5199 false /* should_yield */, false /* not precleaning */);
5200 MarkFromDirtyCardsClosure
5201 markFromDirtyCardsClosure(this, _span,
5202 NULL, // space is set further below
5203 &_markBitMap, &_markStack, &mrias_cl);
5204 {
5205 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5206 // Iterate over the dirty cards, setting the corresponding bits in the
5207 // mod union table.
5208 {
5209 ModUnionClosure modUnionClosure(&_modUnionTable);
5210 _ct->ct_bs()->dirty_card_iterate(
5211 _cmsGen->used_region(),
5212 &modUnionClosure);
5213 }
5214 // Having transferred these marks into the modUnionTable, we just need
5215 // to rescan the marked objects on the dirty cards in the modUnionTable.
5216 // The initial marking may have been done during an asynchronous
5217 // collection so there may be dirty bits in the mod-union table.
5218 const int alignment =
5219 CardTableModRefBS::card_size * BitsPerWord;
5220 {
5221 // ... First handle dirty cards in CMS gen
5222 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5223 MemRegion ur = _cmsGen->used_region();
5224 HeapWord* lb = ur.start();
5225 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5226 MemRegion cms_span(lb, ub);
5227 _modUnionTable.dirty_range_iterate_clear(cms_span,
5228 &markFromDirtyCardsClosure);
5229 verify_work_stacks_empty();
5230 if (PrintCMSStatistics != 0) {
5231 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5232 markFromDirtyCardsClosure.num_dirty_cards());
5233 }
5234 }
5235 }
5236 if (VerifyDuringGC &&
5237 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5238 HandleMark hm; // Discard invalid handles created during verification
5239 Universe::verify();
5240 }
5241 {
5242 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5243
5244 verify_work_stacks_empty();
5245
5246 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5247 GenCollectedHeap::StrongRootsScope srs(gch);
5248
5249 gch->gen_process_roots(_cmsGen->level(),
5250 true, // younger gens as roots
5251 false, // use the local StrongRootsScope
5252 SharedHeap::ScanningOption(roots_scanning_options()),
5253 should_unload_classes(),
5254 &mrias_cl,
5255 NULL,
5256 NULL); // The dirty klasses will be handled below
5257
5258 assert(should_unload_classes()
5259 || (roots_scanning_options() & SharedHeap::SO_AllCodeCache),
5260 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5261 }
5262
5263 {
5264 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5265
5266 verify_work_stacks_empty();
5267
5268 // Scan all class loader data objects that might have been introduced
5269 // during concurrent marking.
5270 ResourceMark rm;
5271 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5272 for (int i = 0; i < array->length(); i++) {
5273 mrias_cl.do_class_loader_data(array->at(i));
5274 }
5275
5276 // We don't need to keep track of new CLDs anymore.
5277 ClassLoaderDataGraph::remember_new_clds(false);
5278
5279 verify_work_stacks_empty();
5280 }
5281
5282 {
5283 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5284
5285 verify_work_stacks_empty();
5286
5287 RemarkKlassClosure remark_klass_closure(&mrias_cl);
5288 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5289
5290 verify_work_stacks_empty();
5291 }
5292
5293 // We might have added oops to ClassLoaderData::_handles during the
5294 // concurrent marking phase. These oops point to newly allocated objects
5295 // that are guaranteed to be kept alive. Either by the direct allocation
5296 // code, or when the young collector processes the roots. Hence,
5297 // we don't have to revisit the _handles block during the remark phase.
5298
5299 verify_work_stacks_empty();
5300 // Restore evacuated mark words, if any, used for overflow list links
5301 if (!CMSOverflowEarlyRestoration) {
5302 restore_preserved_marks_if_any();
5303 }
5304 verify_overflow_empty();
5305 }
5306
5307 ////////////////////////////////////////////////////////
5308 // Parallel Reference Processing Task Proxy Class
5309 ////////////////////////////////////////////////////////
5310 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5311 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5312 CMSCollector* _collector;
5313 CMSBitMap* _mark_bit_map;
5314 const MemRegion _span;
5315 ProcessTask& _task;
5316
5317 public:
5318 CMSRefProcTaskProxy(ProcessTask& task,
5319 CMSCollector* collector,
5320 const MemRegion& span,
5321 CMSBitMap* mark_bit_map,
5322 AbstractWorkGang* workers,
5323 OopTaskQueueSet* task_queues):
5324 // XXX Should superclass AGTWOQ also know about AWG since it knows
5325 // about the task_queues used by the AWG? Then it could initialize
5326 // the terminator() object. See 6984287. The set_for_termination()
5327 // below is a temporary band-aid for the regression in 6984287.
5328 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5329 task_queues),
5330 _task(task),
5331 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5332 {
5333 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5334 "Inconsistency in _span");
5335 set_for_termination(workers->active_workers());
5336 }
5337
5338 OopTaskQueueSet* task_queues() { return queues(); }
5339
5340 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5341
5342 void do_work_steal(int i,
5343 CMSParDrainMarkingStackClosure* drain,
5344 CMSParKeepAliveClosure* keep_alive,
5345 int* seed);
5346
5347 virtual void work(uint worker_id);
5348 };
5349
5350 void CMSRefProcTaskProxy::work(uint worker_id) {
5351 ResourceMark rm;
5352 HandleMark hm;
5353 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5354 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5355 _mark_bit_map,
5356 work_queue(worker_id));
5357 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5358 _mark_bit_map,
5359 work_queue(worker_id));
5360 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5361 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5362 if (_task.marks_oops_alive()) {
5363 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5364 _collector->hash_seed(worker_id));
5365 }
5366 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5367 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5368 }
5369
5370 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5371 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5372 EnqueueTask& _task;
5373
5374 public:
5375 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5376 : AbstractGangTask("Enqueue reference objects in parallel"),
5377 _task(task)
5378 { }
5379
5380 virtual void work(uint worker_id)
5381 {
5382 _task.work(worker_id);
5383 }
5384 };
5385
5386 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5387 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5388 _span(span),
5389 _bit_map(bit_map),
5390 _work_queue(work_queue),
5391 _mark_and_push(collector, span, bit_map, work_queue),
5392 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5393 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5394 { }
5395
5396 // . see if we can share work_queues with ParNew? XXX
5397 void CMSRefProcTaskProxy::do_work_steal(int i,
5398 CMSParDrainMarkingStackClosure* drain,
5399 CMSParKeepAliveClosure* keep_alive,
5400 int* seed) {
5401 OopTaskQueue* work_q = work_queue(i);
5402 NOT_PRODUCT(int num_steals = 0;)
5403 oop obj_to_scan;
5404
5405 while (true) {
5406 // Completely finish any left over work from (an) earlier round(s)
5407 drain->trim_queue(0);
5408 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5409 (size_t)ParGCDesiredObjsFromOverflowList);
5410 // Now check if there's any work in the overflow list
5411 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5412 // only affects the number of attempts made to get work from the
5413 // overflow list and does not affect the number of workers. Just
5414 // pass ParallelGCThreads so this behavior is unchanged.
5415 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5416 work_q,
5417 ParallelGCThreads)) {
5418 // Found something in global overflow list;
5419 // not yet ready to go stealing work from others.
5420 // We'd like to assert(work_q->size() != 0, ...)
5421 // because we just took work from the overflow list,
5422 // but of course we can't, since all of that might have
5423 // been already stolen from us.
5424 continue;
5425 }
5426 // Verify that we have no work before we resort to stealing
5427 assert(work_q->size() == 0, "Have work, shouldn't steal");
5428 // Try to steal from other queues that have work
5429 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5430 NOT_PRODUCT(num_steals++;)
5431 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5432 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5433 // Do scanning work
5434 obj_to_scan->oop_iterate(keep_alive);
5435 // Loop around, finish this work, and try to steal some more
5436 } else if (terminator()->offer_termination()) {
5437 break; // nirvana from the infinite cycle
5438 }
5439 }
5440 NOT_PRODUCT(
5441 if (PrintCMSStatistics != 0) {
5442 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5443 }
5444 )
5445 }
5446
5447 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5448 {
5449 GenCollectedHeap* gch = GenCollectedHeap::heap();
5450 FlexibleWorkGang* workers = gch->workers();
5451 assert(workers != NULL, "Need parallel worker threads.");
5452 CMSRefProcTaskProxy rp_task(task, &_collector,
5453 _collector.ref_processor()->span(),
5454 _collector.markBitMap(),
5455 workers, _collector.task_queues());
5456 workers->run_task(&rp_task);
5457 }
5458
5459 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5460 {
5461
5462 GenCollectedHeap* gch = GenCollectedHeap::heap();
5463 FlexibleWorkGang* workers = gch->workers();
5464 assert(workers != NULL, "Need parallel worker threads.");
5465 CMSRefEnqueueTaskProxy enq_task(task);
5466 workers->run_task(&enq_task);
5467 }
5468
5469 void CMSCollector::refProcessingWork() {
5470 ResourceMark rm;
5471 HandleMark hm;
5472
5473 ReferenceProcessor* rp = ref_processor();
5474 assert(rp->span().equals(_span), "Spans should be equal");
5475 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5476 // Process weak references.
5477 rp->setup_policy(false);
5478 verify_work_stacks_empty();
5479
5480 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5481 &_markStack, false /* !preclean */);
5482 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5483 _span, &_markBitMap, &_markStack,
5484 &cmsKeepAliveClosure, false /* !preclean */);
5485 {
5486 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5487
5488 ReferenceProcessorStats stats;
5489 if (rp->processing_is_mt()) {
5490 // Set the degree of MT here. If the discovery is done MT, there
5491 // may have been a different number of threads doing the discovery
5492 // and a different number of discovered lists may have Ref objects.
5493 // That is OK as long as the Reference lists are balanced (see
5494 // balance_all_queues() and balance_queues()).
5495 GenCollectedHeap* gch = GenCollectedHeap::heap();
5496 int active_workers = ParallelGCThreads;
5497 FlexibleWorkGang* workers = gch->workers();
5498 if (workers != NULL) {
5499 active_workers = workers->active_workers();
5500 // The expectation is that active_workers will have already
5501 // been set to a reasonable value. If it has not been set,
5502 // investigate.
5503 assert(active_workers > 0, "Should have been set during scavenge");
5504 }
5505 rp->set_active_mt_degree(active_workers);
5506 CMSRefProcTaskExecutor task_executor(*this);
5507 stats = rp->process_discovered_references(&_is_alive_closure,
5508 &cmsKeepAliveClosure,
5509 &cmsDrainMarkingStackClosure,
5510 &task_executor,
5511 _gc_timer_cm,
5512 _gc_tracer_cm->gc_id());
5513 } else {
5514 stats = rp->process_discovered_references(&_is_alive_closure,
5515 &cmsKeepAliveClosure,
5516 &cmsDrainMarkingStackClosure,
5517 NULL,
5518 _gc_timer_cm,
5519 _gc_tracer_cm->gc_id());
5520 }
5521 _gc_tracer_cm->report_gc_reference_stats(stats);
5522
5523 }
5524
5525 // This is the point where the entire marking should have completed.
5526 verify_work_stacks_empty();
5527
5528 if (should_unload_classes()) {
5529 {
5530 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5531
5532 // Unload classes and purge the SystemDictionary.
5533 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5534
5535 // Unload nmethods.
5536 CodeCache::do_unloading(&_is_alive_closure, purged_class);
5537
5538 // Prune dead klasses from subklass/sibling/implementor lists.
5539 Klass::clean_weak_klass_links(&_is_alive_closure);
5540 }
5541
5542 {
5543 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5544 // Clean up unreferenced symbols in symbol table.
5545 SymbolTable::unlink();
5546 }
5547
5548 {
5549 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5550 // Delete entries for dead interned strings.
5551 StringTable::unlink(&_is_alive_closure);
5552 }
5553 }
5554
5555
5556 // Restore any preserved marks as a result of mark stack or
5557 // work queue overflow
5558 restore_preserved_marks_if_any(); // done single-threaded for now
5559
5560 rp->set_enqueuing_is_done(true);
5561 if (rp->processing_is_mt()) {
5562 rp->balance_all_queues();
5563 CMSRefProcTaskExecutor task_executor(*this);
5564 rp->enqueue_discovered_references(&task_executor);
5565 } else {
5566 rp->enqueue_discovered_references(NULL);
5567 }
5568 rp->verify_no_references_recorded();
5569 assert(!rp->discovery_enabled(), "should have been disabled");
5570 }
5571
5572 #ifndef PRODUCT
5573 void CMSCollector::check_correct_thread_executing() {
5574 Thread* t = Thread::current();
5575 // Only the VM thread or the CMS thread should be here.
5576 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5577 "Unexpected thread type");
5578 // If this is the vm thread, the foreground process
5579 // should not be waiting. Note that _foregroundGCIsActive is
5580 // true while the foreground collector is waiting.
5581 if (_foregroundGCShouldWait) {
5582 // We cannot be the VM thread
5583 assert(t->is_ConcurrentGC_thread(),
5584 "Should be CMS thread");
5585 } else {
5586 // We can be the CMS thread only if we are in a stop-world
5587 // phase of CMS collection.
5588 if (t->is_ConcurrentGC_thread()) {
5589 assert(_collectorState == InitialMarking ||
5590 _collectorState == FinalMarking,
5591 "Should be a stop-world phase");
5592 // The CMS thread should be holding the CMS_token.
5593 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5594 "Potential interference with concurrently "
5595 "executing VM thread");
5596 }
5597 }
5598 }
5599 #endif
5600
5601 void CMSCollector::sweep() {
5602 assert(_collectorState == Sweeping, "just checking");
5603 check_correct_thread_executing();
5604 verify_work_stacks_empty();
5605 verify_overflow_empty();
5606 increment_sweep_count();
5607 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5608
5609 _inter_sweep_timer.stop();
5610 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5611
5612 assert(!_intra_sweep_timer.is_active(), "Should not be active");
5613 _intra_sweep_timer.reset();
5614 _intra_sweep_timer.start();
5615 {
5616 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5617 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails);
5618 // First sweep the old gen
5619 {
5620 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5621 bitMapLock());
5622 sweepWork(_cmsGen);
5623 }
5624
5625 // Update Universe::_heap_*_at_gc figures.
5626 // We need all the free list locks to make the abstract state
5627 // transition from Sweeping to Resetting. See detailed note
5628 // further below.
5629 {
5630 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5631 // Update heap occupancy information which is used as
5632 // input to soft ref clearing policy at the next gc.
5633 Universe::update_heap_info_at_gc();
5634 _collectorState = Resizing;
5635 }
5636 }
5637 verify_work_stacks_empty();
5638 verify_overflow_empty();
5639
5640 if (should_unload_classes()) {
5641 // Delay purge to the beginning of the next safepoint. Metaspace::contains
5642 // requires that the virtual spaces are stable and not deleted.
5643 ClassLoaderDataGraph::set_should_purge(true);
5644 }
5645
5646 _intra_sweep_timer.stop();
5647 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5648
5649 _inter_sweep_timer.reset();
5650 _inter_sweep_timer.start();
5651
5652 // We need to use a monotonically non-decreasing time in ms
5653 // or we will see time-warp warnings and os::javaTimeMillis()
5654 // does not guarantee monotonicity.
5655 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5656 update_time_of_last_gc(now);
5657
5658 // NOTE on abstract state transitions:
5659 // Mutators allocate-live and/or mark the mod-union table dirty
5660 // based on the state of the collection. The former is done in
5661 // the interval [Marking, Sweeping] and the latter in the interval
5662 // [Marking, Sweeping). Thus the transitions into the Marking state
5663 // and out of the Sweeping state must be synchronously visible
5664 // globally to the mutators.
5665 // The transition into the Marking state happens with the world
5666 // stopped so the mutators will globally see it. Sweeping is
5667 // done asynchronously by the background collector so the transition
5668 // from the Sweeping state to the Resizing state must be done
5669 // under the freelistLock (as is the check for whether to
5670 // allocate-live and whether to dirty the mod-union table).
5671 assert(_collectorState == Resizing, "Change of collector state to"
5672 " Resizing must be done under the freelistLocks (plural)");
5673
5674 // Now that sweeping has been completed, we clear
5675 // the incremental_collection_failed flag,
5676 // thus inviting a younger gen collection to promote into
5677 // this generation. If such a promotion may still fail,
5678 // the flag will be set again when a young collection is
5679 // attempted.
5680 GenCollectedHeap* gch = GenCollectedHeap::heap();
5681 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
5682 gch->update_full_collections_completed(_collection_count_start);
5683 }
5684
5685 // FIX ME!!! Looks like this belongs in CFLSpace, with
5686 // CMSGen merely delegating to it.
5687 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5688 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5689 HeapWord* minAddr = _cmsSpace->bottom();
5690 HeapWord* largestAddr =
5691 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5692 if (largestAddr == NULL) {
5693 // The dictionary appears to be empty. In this case
5694 // try to coalesce at the end of the heap.
5695 largestAddr = _cmsSpace->end();
5696 }
5697 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5698 size_t nearLargestOffset =
5699 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5700 if (PrintFLSStatistics != 0) {
5701 gclog_or_tty->print_cr(
5702 "CMS: Large Block: " PTR_FORMAT ";"
5703 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5704 largestAddr,
5705 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset);
5706 }
5707 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5708 }
5709
5710 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5711 return addr >= _cmsSpace->nearLargestChunk();
5712 }
5713
5714 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5715 return _cmsSpace->find_chunk_at_end();
5716 }
5717
5718 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
5719 bool full) {
5720 // The next lower level has been collected. Gather any statistics
5721 // that are of interest at this point.
5722 if (!full && (current_level + 1) == level()) {
5723 // Gather statistics on the young generation collection.
5724 collector()->stats().record_gc0_end(used());
5725 }
5726 }
5727
5728 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen) {
5729 // We iterate over the space(s) underlying this generation,
5730 // checking the mark bit map to see if the bits corresponding
5731 // to specific blocks are marked or not. Blocks that are
5732 // marked are live and are not swept up. All remaining blocks
5733 // are swept up, with coalescing on-the-fly as we sweep up
5734 // contiguous free and/or garbage blocks:
5735 // We need to ensure that the sweeper synchronizes with allocators
5736 // and stop-the-world collectors. In particular, the following
5737 // locks are used:
5738 // . CMS token: if this is held, a stop the world collection cannot occur
5739 // . freelistLock: if this is held no allocation can occur from this
5740 // generation by another thread
5741 // . bitMapLock: if this is held, no other thread can access or update
5742 //
5743
5744 // Note that we need to hold the freelistLock if we use
5745 // block iterate below; else the iterator might go awry if
5746 // a mutator (or promotion) causes block contents to change
5747 // (for instance if the allocator divvies up a block).
5748 // If we hold the free list lock, for all practical purposes
5749 // young generation GC's can't occur (they'll usually need to
5750 // promote), so we might as well prevent all young generation
5751 // GC's while we do a sweeping step. For the same reason, we might
5752 // as well take the bit map lock for the entire duration
5753
5754 // check that we hold the requisite locks
5755 assert(have_cms_token(), "Should hold cms token");
5756 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5757 assert_lock_strong(gen->freelistLock());
5758 assert_lock_strong(bitMapLock());
5759
5760 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5761 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
5762 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5763 _inter_sweep_estimate.padded_average(),
5764 _intra_sweep_estimate.padded_average());
5765 gen->setNearLargestChunk();
5766
5767 {
5768 SweepClosure sweepClosure(this, gen, &_markBitMap, CMSYield);
5769 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5770 // We need to free-up/coalesce garbage/blocks from a
5771 // co-terminal free run. This is done in the SweepClosure
5772 // destructor; so, do not remove this scope, else the
5773 // end-of-sweep-census below will be off by a little bit.
5774 }
5775 gen->cmsSpace()->sweep_completed();
5776 gen->cmsSpace()->endSweepFLCensus(sweep_count());
5777 if (should_unload_classes()) { // unloaded classes this cycle,
5778 _concurrent_cycles_since_last_unload = 0; // ... reset count
5779 } else { // did not unload classes,
5780 _concurrent_cycles_since_last_unload++; // ... increment count
5781 }
5782 }
5783
5784 // Reset CMS data structures (for now just the marking bit map)
5785 // preparatory for the next cycle.
5786 void CMSCollector::reset(bool concurrent) {
5787 if (concurrent) {
5788 CMSTokenSyncWithLocks ts(true, bitMapLock());
5789
5790 // If the state is not "Resetting", the foreground thread
5791 // has done a collection and the resetting.
5792 if (_collectorState != Resetting) {
5793 assert(_collectorState == Idling, "The state should only change"
5794 " because the foreground collector has finished the collection");
5795 return;
5796 }
5797
5798 // Clear the mark bitmap (no grey objects to start with)
5799 // for the next cycle.
5800 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5801 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails);
5802
5803 HeapWord* curAddr = _markBitMap.startWord();
5804 while (curAddr < _markBitMap.endWord()) {
5805 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
5806 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5807 _markBitMap.clear_large_range(chunk);
5808 if (ConcurrentMarkSweepThread::should_yield() &&
5809 !foregroundGCIsActive() &&
5810 CMSYield) {
5811 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5812 "CMS thread should hold CMS token");
5813 assert_lock_strong(bitMapLock());
5814 bitMapLock()->unlock();
5815 ConcurrentMarkSweepThread::desynchronize(true);
5816 stopTimer();
5817 if (PrintCMSStatistics != 0) {
5818 incrementYields();
5819 }
5820
5821 // See the comment in coordinator_yield()
5822 for (unsigned i = 0; i < CMSYieldSleepCount &&
5823 ConcurrentMarkSweepThread::should_yield() &&
5824 !CMSCollector::foregroundGCIsActive(); ++i) {
5825 os::sleep(Thread::current(), 1, false);
5826 }
5827
5828 ConcurrentMarkSweepThread::synchronize(true);
5829 bitMapLock()->lock_without_safepoint_check();
5830 startTimer();
5831 }
5832 curAddr = chunk.end();
5833 }
5834 // A successful mostly concurrent collection has been done.
5835 // Because only the full (i.e., concurrent mode failure) collections
5836 // are being measured for gc overhead limits, clean the "near" flag
5837 // and count.
5838 size_policy()->reset_gc_overhead_limit_count();
5839 _collectorState = Idling;
5840 } else {
5841 // already have the lock
5842 assert(_collectorState == Resetting, "just checking");
5843 assert_lock_strong(bitMapLock());
5844 _markBitMap.clear_all();
5845 _collectorState = Idling;
5846 }
5847
5848 register_gc_end();
5849 }
5850
5851 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5852 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5853 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id());
5854 TraceCollectorStats tcs(counters());
5855
5856 switch (op) {
5857 case CMS_op_checkpointRootsInitial: {
5858 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5859 checkpointRootsInitial();
5860 if (PrintGC) {
5861 _cmsGen->printOccupancy("initial-mark");
5862 }
5863 break;
5864 }
5865 case CMS_op_checkpointRootsFinal: {
5866 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5867 checkpointRootsFinal();
5868 if (PrintGC) {
5869 _cmsGen->printOccupancy("remark");
5870 }
5871 break;
5872 }
5873 default:
5874 fatal("No such CMS_op");
5875 }
5876 }
5877
5878 #ifndef PRODUCT
5879 size_t const CMSCollector::skip_header_HeapWords() {
5880 return FreeChunk::header_size();
5881 }
5882
5883 // Try and collect here conditions that should hold when
5884 // CMS thread is exiting. The idea is that the foreground GC
5885 // thread should not be blocked if it wants to terminate
5886 // the CMS thread and yet continue to run the VM for a while
5887 // after that.
5888 void CMSCollector::verify_ok_to_terminate() const {
5889 assert(Thread::current()->is_ConcurrentGC_thread(),
5890 "should be called by CMS thread");
5891 assert(!_foregroundGCShouldWait, "should be false");
5892 // We could check here that all the various low-level locks
5893 // are not held by the CMS thread, but that is overkill; see
5894 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5895 // is checked.
5896 }
5897 #endif
5898
5899 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5900 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5901 "missing Printezis mark?");
5902 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5903 size_t size = pointer_delta(nextOneAddr + 1, addr);
5904 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5905 "alignment problem");
5906 assert(size >= 3, "Necessary for Printezis marks to work");
5907 return size;
5908 }
5909
5910 // A variant of the above (block_size_using_printezis_bits()) except
5911 // that we return 0 if the P-bits are not yet set.
5912 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5913 if (_markBitMap.isMarked(addr + 1)) {
5914 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5915 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5916 size_t size = pointer_delta(nextOneAddr + 1, addr);
5917 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5918 "alignment problem");
5919 assert(size >= 3, "Necessary for Printezis marks to work");
5920 return size;
5921 }
5922 return 0;
5923 }
5924
5925 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5926 size_t sz = 0;
5927 oop p = (oop)addr;
5928 if (p->klass_or_null() != NULL) {
5929 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5930 } else {
5931 sz = block_size_using_printezis_bits(addr);
5932 }
5933 assert(sz > 0, "size must be nonzero");
5934 HeapWord* next_block = addr + sz;
5935 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
5936 CardTableModRefBS::card_size);
5937 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
5938 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
5939 "must be different cards");
5940 return next_card;
5941 }
5942
5943
5944 // CMS Bit Map Wrapper /////////////////////////////////////////
5945
5946 // Construct a CMS bit map infrastructure, but don't create the
5947 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5948 // further below.
5949 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5950 _bm(),
5951 _shifter(shifter),
5952 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5953 Monitor::_safepoint_check_sometimes) : NULL)
5954 {
5955 _bmStartWord = 0;
5956 _bmWordSize = 0;
5957 }
5958
5959 bool CMSBitMap::allocate(MemRegion mr) {
5960 _bmStartWord = mr.start();
5961 _bmWordSize = mr.word_size();
5962 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5963 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5964 if (!brs.is_reserved()) {
5965 warning("CMS bit map allocation failure");
5966 return false;
5967 }
5968 // For now we'll just commit all of the bit map up front.
5969 // Later on we'll try to be more parsimonious with swap.
5970 if (!_virtual_space.initialize(brs, brs.size())) {
5971 warning("CMS bit map backing store failure");
5972 return false;
5973 }
5974 assert(_virtual_space.committed_size() == brs.size(),
5975 "didn't reserve backing store for all of CMS bit map?");
5976 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
5977 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5978 _bmWordSize, "inconsistency in bit map sizing");
5979 _bm.set_size(_bmWordSize >> _shifter);
5980
5981 // bm.clear(); // can we rely on getting zero'd memory? verify below
5982 assert(isAllClear(),
5983 "Expected zero'd memory from ReservedSpace constructor");
5984 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5985 "consistency check");
5986 return true;
5987 }
5988
5989 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5990 HeapWord *next_addr, *end_addr, *last_addr;
5991 assert_locked();
5992 assert(covers(mr), "out-of-range error");
5993 // XXX assert that start and end are appropriately aligned
5994 for (next_addr = mr.start(), end_addr = mr.end();
5995 next_addr < end_addr; next_addr = last_addr) {
5996 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5997 last_addr = dirty_region.end();
5998 if (!dirty_region.is_empty()) {
5999 cl->do_MemRegion(dirty_region);
6000 } else {
6001 assert(last_addr == end_addr, "program logic");
6002 return;
6003 }
6004 }
6005 }
6006
6007 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
6008 _bm.print_on_error(st, prefix);
6009 }
6010
6011 #ifndef PRODUCT
6012 void CMSBitMap::assert_locked() const {
6013 CMSLockVerifier::assert_locked(lock());
6014 }
6015
6016 bool CMSBitMap::covers(MemRegion mr) const {
6017 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6018 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6019 "size inconsistency");
6020 return (mr.start() >= _bmStartWord) &&
6021 (mr.end() <= endWord());
6022 }
6023
6024 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6025 return (start >= _bmStartWord && (start + size) <= endWord());
6026 }
6027
6028 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6029 // verify that there are no 1 bits in the interval [left, right)
6030 FalseBitMapClosure falseBitMapClosure;
6031 iterate(&falseBitMapClosure, left, right);
6032 }
6033
6034 void CMSBitMap::region_invariant(MemRegion mr)
6035 {
6036 assert_locked();
6037 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6038 assert(!mr.is_empty(), "unexpected empty region");
6039 assert(covers(mr), "mr should be covered by bit map");
6040 // convert address range into offset range
6041 size_t start_ofs = heapWordToOffset(mr.start());
6042 // Make sure that end() is appropriately aligned
6043 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6044 (1 << (_shifter+LogHeapWordSize))),
6045 "Misaligned mr.end()");
6046 size_t end_ofs = heapWordToOffset(mr.end());
6047 assert(end_ofs > start_ofs, "Should mark at least one bit");
6048 }
6049
6050 #endif
6051
6052 bool CMSMarkStack::allocate(size_t size) {
6053 // allocate a stack of the requisite depth
6054 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6055 size * sizeof(oop)));
6056 if (!rs.is_reserved()) {
6057 warning("CMSMarkStack allocation failure");
6058 return false;
6059 }
6060 if (!_virtual_space.initialize(rs, rs.size())) {
6061 warning("CMSMarkStack backing store failure");
6062 return false;
6063 }
6064 assert(_virtual_space.committed_size() == rs.size(),
6065 "didn't reserve backing store for all of CMS stack?");
6066 _base = (oop*)(_virtual_space.low());
6067 _index = 0;
6068 _capacity = size;
6069 NOT_PRODUCT(_max_depth = 0);
6070 return true;
6071 }
6072
6073 // XXX FIX ME !!! In the MT case we come in here holding a
6074 // leaf lock. For printing we need to take a further lock
6075 // which has lower rank. We need to recalibrate the two
6076 // lock-ranks involved in order to be able to print the
6077 // messages below. (Or defer the printing to the caller.
6078 // For now we take the expedient path of just disabling the
6079 // messages for the problematic case.)
6080 void CMSMarkStack::expand() {
6081 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6082 if (_capacity == MarkStackSizeMax) {
6083 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6084 // We print a warning message only once per CMS cycle.
6085 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6086 }
6087 return;
6088 }
6089 // Double capacity if possible
6090 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6091 // Do not give up existing stack until we have managed to
6092 // get the double capacity that we desired.
6093 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6094 new_capacity * sizeof(oop)));
6095 if (rs.is_reserved()) {
6096 // Release the backing store associated with old stack
6097 _virtual_space.release();
6098 // Reinitialize virtual space for new stack
6099 if (!_virtual_space.initialize(rs, rs.size())) {
6100 fatal("Not enough swap for expanded marking stack");
6101 }
6102 _base = (oop*)(_virtual_space.low());
6103 _index = 0;
6104 _capacity = new_capacity;
6105 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6106 // Failed to double capacity, continue;
6107 // we print a detail message only once per CMS cycle.
6108 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6109 SIZE_FORMAT"K",
6110 _capacity / K, new_capacity / K);
6111 }
6112 }
6113
6114
6115 // Closures
6116 // XXX: there seems to be a lot of code duplication here;
6117 // should refactor and consolidate common code.
6118
6119 // This closure is used to mark refs into the CMS generation in
6120 // the CMS bit map. Called at the first checkpoint. This closure
6121 // assumes that we do not need to re-mark dirty cards; if the CMS
6122 // generation on which this is used is not an oldest
6123 // generation then this will lose younger_gen cards!
6124
6125 MarkRefsIntoClosure::MarkRefsIntoClosure(
6126 MemRegion span, CMSBitMap* bitMap):
6127 _span(span),
6128 _bitMap(bitMap)
6129 {
6130 assert(_ref_processor == NULL, "deliberately left NULL");
6131 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6132 }
6133
6134 void MarkRefsIntoClosure::do_oop(oop obj) {
6135 // if p points into _span, then mark corresponding bit in _markBitMap
6136 assert(obj->is_oop(), "expected an oop");
6137 HeapWord* addr = (HeapWord*)obj;
6138 if (_span.contains(addr)) {
6139 // this should be made more efficient
6140 _bitMap->mark(addr);
6141 }
6142 }
6143
6144 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6145 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6146
6147 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
6148 MemRegion span, CMSBitMap* bitMap):
6149 _span(span),
6150 _bitMap(bitMap)
6151 {
6152 assert(_ref_processor == NULL, "deliberately left NULL");
6153 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6154 }
6155
6156 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
6157 // if p points into _span, then mark corresponding bit in _markBitMap
6158 assert(obj->is_oop(), "expected an oop");
6159 HeapWord* addr = (HeapWord*)obj;
6160 if (_span.contains(addr)) {
6161 // this should be made more efficient
6162 _bitMap->par_mark(addr);
6163 }
6164 }
6165
6166 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6167 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6168
6169 // A variant of the above, used for CMS marking verification.
6170 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6171 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6172 _span(span),
6173 _verification_bm(verification_bm),
6174 _cms_bm(cms_bm)
6175 {
6176 assert(_ref_processor == NULL, "deliberately left NULL");
6177 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6178 }
6179
6180 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6181 // if p points into _span, then mark corresponding bit in _markBitMap
6182 assert(obj->is_oop(), "expected an oop");
6183 HeapWord* addr = (HeapWord*)obj;
6184 if (_span.contains(addr)) {
6185 _verification_bm->mark(addr);
6186 if (!_cms_bm->isMarked(addr)) {
6187 oop(addr)->print();
6188 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6189 fatal("... aborting");
6190 }
6191 }
6192 }
6193
6194 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6195 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6196
6197 //////////////////////////////////////////////////
6198 // MarkRefsIntoAndScanClosure
6199 //////////////////////////////////////////////////
6200
6201 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6202 ReferenceProcessor* rp,
6203 CMSBitMap* bit_map,
6204 CMSBitMap* mod_union_table,
6205 CMSMarkStack* mark_stack,
6206 CMSCollector* collector,
6207 bool should_yield,
6208 bool concurrent_precleaning):
6209 _collector(collector),
6210 _span(span),
6211 _bit_map(bit_map),
6212 _mark_stack(mark_stack),
6213 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6214 mark_stack, concurrent_precleaning),
6215 _yield(should_yield),
6216 _concurrent_precleaning(concurrent_precleaning),
6217 _freelistLock(NULL)
6218 {
6219 _ref_processor = rp;
6220 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6221 }
6222
6223 // This closure is used to mark refs into the CMS generation at the
6224 // second (final) checkpoint, and to scan and transitively follow
6225 // the unmarked oops. It is also used during the concurrent precleaning
6226 // phase while scanning objects on dirty cards in the CMS generation.
6227 // The marks are made in the marking bit map and the marking stack is
6228 // used for keeping the (newly) grey objects during the scan.
6229 // The parallel version (Par_...) appears further below.
6230 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6231 if (obj != NULL) {
6232 assert(obj->is_oop(), "expected an oop");
6233 HeapWord* addr = (HeapWord*)obj;
6234 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6235 assert(_collector->overflow_list_is_empty(),
6236 "overflow list should be empty");
6237 if (_span.contains(addr) &&
6238 !_bit_map->isMarked(addr)) {
6239 // mark bit map (object is now grey)
6240 _bit_map->mark(addr);
6241 // push on marking stack (stack should be empty), and drain the
6242 // stack by applying this closure to the oops in the oops popped
6243 // from the stack (i.e. blacken the grey objects)
6244 bool res = _mark_stack->push(obj);
6245 assert(res, "Should have space to push on empty stack");
6246 do {
6247 oop new_oop = _mark_stack->pop();
6248 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6249 assert(_bit_map->isMarked((HeapWord*)new_oop),
6250 "only grey objects on this stack");
6251 // iterate over the oops in this oop, marking and pushing
6252 // the ones in CMS heap (i.e. in _span).
6253 new_oop->oop_iterate(&_pushAndMarkClosure);
6254 // check if it's time to yield
6255 do_yield_check();
6256 } while (!_mark_stack->isEmpty() ||
6257 (!_concurrent_precleaning && take_from_overflow_list()));
6258 // if marking stack is empty, and we are not doing this
6259 // during precleaning, then check the overflow list
6260 }
6261 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6262 assert(_collector->overflow_list_is_empty(),
6263 "overflow list was drained above");
6264 // We could restore evacuated mark words, if any, used for
6265 // overflow list links here because the overflow list is
6266 // provably empty here. That would reduce the maximum
6267 // size requirements for preserved_{oop,mark}_stack.
6268 // But we'll just postpone it until we are all done
6269 // so we can just stream through.
6270 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6271 _collector->restore_preserved_marks_if_any();
6272 assert(_collector->no_preserved_marks(), "No preserved marks");
6273 }
6274 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6275 "All preserved marks should have been restored above");
6276 }
6277 }
6278
6279 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6280 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6281
6282 void MarkRefsIntoAndScanClosure::do_yield_work() {
6283 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6284 "CMS thread should hold CMS token");
6285 assert_lock_strong(_freelistLock);
6286 assert_lock_strong(_bit_map->lock());
6287 // relinquish the free_list_lock and bitMaplock()
6288 _bit_map->lock()->unlock();
6289 _freelistLock->unlock();
6290 ConcurrentMarkSweepThread::desynchronize(true);
6291 _collector->stopTimer();
6292 if (PrintCMSStatistics != 0) {
6293 _collector->incrementYields();
6294 }
6295
6296 // See the comment in coordinator_yield()
6297 for (unsigned i = 0;
6298 i < CMSYieldSleepCount &&
6299 ConcurrentMarkSweepThread::should_yield() &&
6300 !CMSCollector::foregroundGCIsActive();
6301 ++i) {
6302 os::sleep(Thread::current(), 1, false);
6303 }
6304
6305 ConcurrentMarkSweepThread::synchronize(true);
6306 _freelistLock->lock_without_safepoint_check();
6307 _bit_map->lock()->lock_without_safepoint_check();
6308 _collector->startTimer();
6309 }
6310
6311 ///////////////////////////////////////////////////////////
6312 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6313 // MarkRefsIntoAndScanClosure
6314 ///////////////////////////////////////////////////////////
6315 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6316 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6317 CMSBitMap* bit_map, OopTaskQueue* work_queue):
6318 _span(span),
6319 _bit_map(bit_map),
6320 _work_queue(work_queue),
6321 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6322 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6323 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
6324 {
6325 _ref_processor = rp;
6326 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6327 }
6328
6329 // This closure is used to mark refs into the CMS generation at the
6330 // second (final) checkpoint, and to scan and transitively follow
6331 // the unmarked oops. The marks are made in the marking bit map and
6332 // the work_queue is used for keeping the (newly) grey objects during
6333 // the scan phase whence they are also available for stealing by parallel
6334 // threads. Since the marking bit map is shared, updates are
6335 // synchronized (via CAS).
6336 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6337 if (obj != NULL) {
6338 // Ignore mark word because this could be an already marked oop
6339 // that may be chained at the end of the overflow list.
6340 assert(obj->is_oop(true), "expected an oop");
6341 HeapWord* addr = (HeapWord*)obj;
6342 if (_span.contains(addr) &&
6343 !_bit_map->isMarked(addr)) {
6344 // mark bit map (object will become grey):
6345 // It is possible for several threads to be
6346 // trying to "claim" this object concurrently;
6347 // the unique thread that succeeds in marking the
6348 // object first will do the subsequent push on
6349 // to the work queue (or overflow list).
6350 if (_bit_map->par_mark(addr)) {
6351 // push on work_queue (which may not be empty), and trim the
6352 // queue to an appropriate length by applying this closure to
6353 // the oops in the oops popped from the stack (i.e. blacken the
6354 // grey objects)
6355 bool res = _work_queue->push(obj);
6356 assert(res, "Low water mark should be less than capacity?");
6357 trim_queue(_low_water_mark);
6358 } // Else, another thread claimed the object
6359 }
6360 }
6361 }
6362
6363 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6364 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6365
6366 // This closure is used to rescan the marked objects on the dirty cards
6367 // in the mod union table and the card table proper.
6368 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6369 oop p, MemRegion mr) {
6370
6371 size_t size = 0;
6372 HeapWord* addr = (HeapWord*)p;
6373 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6374 assert(_span.contains(addr), "we are scanning the CMS generation");
6375 // check if it's time to yield
6376 if (do_yield_check()) {
6377 // We yielded for some foreground stop-world work,
6378 // and we have been asked to abort this ongoing preclean cycle.
6379 return 0;
6380 }
6381 if (_bitMap->isMarked(addr)) {
6382 // it's marked; is it potentially uninitialized?
6383 if (p->klass_or_null() != NULL) {
6384 // an initialized object; ignore mark word in verification below
6385 // since we are running concurrent with mutators
6386 assert(p->is_oop(true), "should be an oop");
6387 if (p->is_objArray()) {
6388 // objArrays are precisely marked; restrict scanning
6389 // to dirty cards only.
6390 size = CompactibleFreeListSpace::adjustObjectSize(
6391 p->oop_iterate(_scanningClosure, mr));
6392 } else {
6393 // A non-array may have been imprecisely marked; we need
6394 // to scan object in its entirety.
6395 size = CompactibleFreeListSpace::adjustObjectSize(
6396 p->oop_iterate(_scanningClosure));
6397 }
6398 #ifdef ASSERT
6399 size_t direct_size =
6400 CompactibleFreeListSpace::adjustObjectSize(p->size());
6401 assert(size == direct_size, "Inconsistency in size");
6402 assert(size >= 3, "Necessary for Printezis marks to work");
6403 if (!_bitMap->isMarked(addr+1)) {
6404 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6405 } else {
6406 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6407 assert(_bitMap->isMarked(addr+size-1),
6408 "inconsistent Printezis mark");
6409 }
6410 #endif // ASSERT
6411 } else {
6412 // An uninitialized object.
6413 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6414 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6415 size = pointer_delta(nextOneAddr + 1, addr);
6416 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6417 "alignment problem");
6418 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6419 // will dirty the card when the klass pointer is installed in the
6420 // object (signaling the completion of initialization).
6421 }
6422 } else {
6423 // Either a not yet marked object or an uninitialized object
6424 if (p->klass_or_null() == NULL) {
6425 // An uninitialized object, skip to the next card, since
6426 // we may not be able to read its P-bits yet.
6427 assert(size == 0, "Initial value");
6428 } else {
6429 // An object not (yet) reached by marking: we merely need to
6430 // compute its size so as to go look at the next block.
6431 assert(p->is_oop(true), "should be an oop");
6432 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6433 }
6434 }
6435 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6436 return size;
6437 }
6438
6439 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6440 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6441 "CMS thread should hold CMS token");
6442 assert_lock_strong(_freelistLock);
6443 assert_lock_strong(_bitMap->lock());
6444 // relinquish the free_list_lock and bitMaplock()
6445 _bitMap->lock()->unlock();
6446 _freelistLock->unlock();
6447 ConcurrentMarkSweepThread::desynchronize(true);
6448 _collector->stopTimer();
6449 if (PrintCMSStatistics != 0) {
6450 _collector->incrementYields();
6451 }
6452
6453 // See the comment in coordinator_yield()
6454 for (unsigned i = 0; i < CMSYieldSleepCount &&
6455 ConcurrentMarkSweepThread::should_yield() &&
6456 !CMSCollector::foregroundGCIsActive(); ++i) {
6457 os::sleep(Thread::current(), 1, false);
6458 }
6459
6460 ConcurrentMarkSweepThread::synchronize(true);
6461 _freelistLock->lock_without_safepoint_check();
6462 _bitMap->lock()->lock_without_safepoint_check();
6463 _collector->startTimer();
6464 }
6465
6466
6467 //////////////////////////////////////////////////////////////////
6468 // SurvivorSpacePrecleanClosure
6469 //////////////////////////////////////////////////////////////////
6470 // This (single-threaded) closure is used to preclean the oops in
6471 // the survivor spaces.
6472 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6473
6474 HeapWord* addr = (HeapWord*)p;
6475 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6476 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6477 assert(p->klass_or_null() != NULL, "object should be initialized");
6478 // an initialized object; ignore mark word in verification below
6479 // since we are running concurrent with mutators
6480 assert(p->is_oop(true), "should be an oop");
6481 // Note that we do not yield while we iterate over
6482 // the interior oops of p, pushing the relevant ones
6483 // on our marking stack.
6484 size_t size = p->oop_iterate(_scanning_closure);
6485 do_yield_check();
6486 // Observe that below, we do not abandon the preclean
6487 // phase as soon as we should; rather we empty the
6488 // marking stack before returning. This is to satisfy
6489 // some existing assertions. In general, it may be a
6490 // good idea to abort immediately and complete the marking
6491 // from the grey objects at a later time.
6492 while (!_mark_stack->isEmpty()) {
6493 oop new_oop = _mark_stack->pop();
6494 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6495 assert(_bit_map->isMarked((HeapWord*)new_oop),
6496 "only grey objects on this stack");
6497 // iterate over the oops in this oop, marking and pushing
6498 // the ones in CMS heap (i.e. in _span).
6499 new_oop->oop_iterate(_scanning_closure);
6500 // check if it's time to yield
6501 do_yield_check();
6502 }
6503 unsigned int after_count =
6504 GenCollectedHeap::heap()->total_collections();
6505 bool abort = (_before_count != after_count) ||
6506 _collector->should_abort_preclean();
6507 return abort ? 0 : size;
6508 }
6509
6510 void SurvivorSpacePrecleanClosure::do_yield_work() {
6511 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6512 "CMS thread should hold CMS token");
6513 assert_lock_strong(_bit_map->lock());
6514 // Relinquish the bit map lock
6515 _bit_map->lock()->unlock();
6516 ConcurrentMarkSweepThread::desynchronize(true);
6517 _collector->stopTimer();
6518 if (PrintCMSStatistics != 0) {
6519 _collector->incrementYields();
6520 }
6521
6522 // See the comment in coordinator_yield()
6523 for (unsigned i = 0; i < CMSYieldSleepCount &&
6524 ConcurrentMarkSweepThread::should_yield() &&
6525 !CMSCollector::foregroundGCIsActive(); ++i) {
6526 os::sleep(Thread::current(), 1, false);
6527 }
6528
6529 ConcurrentMarkSweepThread::synchronize(true);
6530 _bit_map->lock()->lock_without_safepoint_check();
6531 _collector->startTimer();
6532 }
6533
6534 // This closure is used to rescan the marked objects on the dirty cards
6535 // in the mod union table and the card table proper. In the parallel
6536 // case, although the bitMap is shared, we do a single read so the
6537 // isMarked() query is "safe".
6538 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6539 // Ignore mark word because we are running concurrent with mutators
6540 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p)));
6541 HeapWord* addr = (HeapWord*)p;
6542 assert(_span.contains(addr), "we are scanning the CMS generation");
6543 bool is_obj_array = false;
6544 #ifdef ASSERT
6545 if (!_parallel) {
6546 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6547 assert(_collector->overflow_list_is_empty(),
6548 "overflow list should be empty");
6549
6550 }
6551 #endif // ASSERT
6552 if (_bit_map->isMarked(addr)) {
6553 // Obj arrays are precisely marked, non-arrays are not;
6554 // so we scan objArrays precisely and non-arrays in their
6555 // entirety.
6556 if (p->is_objArray()) {
6557 is_obj_array = true;
6558 if (_parallel) {
6559 p->oop_iterate(_par_scan_closure, mr);
6560 } else {
6561 p->oop_iterate(_scan_closure, mr);
6562 }
6563 } else {
6564 if (_parallel) {
6565 p->oop_iterate(_par_scan_closure);
6566 } else {
6567 p->oop_iterate(_scan_closure);
6568 }
6569 }
6570 }
6571 #ifdef ASSERT
6572 if (!_parallel) {
6573 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6574 assert(_collector->overflow_list_is_empty(),
6575 "overflow list should be empty");
6576
6577 }
6578 #endif // ASSERT
6579 return is_obj_array;
6580 }
6581
6582 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6583 MemRegion span,
6584 CMSBitMap* bitMap, CMSMarkStack* markStack,
6585 bool should_yield, bool verifying):
6586 _collector(collector),
6587 _span(span),
6588 _bitMap(bitMap),
6589 _mut(&collector->_modUnionTable),
6590 _markStack(markStack),
6591 _yield(should_yield),
6592 _skipBits(0)
6593 {
6594 assert(_markStack->isEmpty(), "stack should be empty");
6595 _finger = _bitMap->startWord();
6596 _threshold = _finger;
6597 assert(_collector->_restart_addr == NULL, "Sanity check");
6598 assert(_span.contains(_finger), "Out of bounds _finger?");
6599 DEBUG_ONLY(_verifying = verifying;)
6600 }
6601
6602 void MarkFromRootsClosure::reset(HeapWord* addr) {
6603 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6604 assert(_span.contains(addr), "Out of bounds _finger?");
6605 _finger = addr;
6606 _threshold = (HeapWord*)round_to(
6607 (intptr_t)_finger, CardTableModRefBS::card_size);
6608 }
6609
6610 // Should revisit to see if this should be restructured for
6611 // greater efficiency.
6612 bool MarkFromRootsClosure::do_bit(size_t offset) {
6613 if (_skipBits > 0) {
6614 _skipBits--;
6615 return true;
6616 }
6617 // convert offset into a HeapWord*
6618 HeapWord* addr = _bitMap->startWord() + offset;
6619 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6620 "address out of range");
6621 assert(_bitMap->isMarked(addr), "tautology");
6622 if (_bitMap->isMarked(addr+1)) {
6623 // this is an allocated but not yet initialized object
6624 assert(_skipBits == 0, "tautology");
6625 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6626 oop p = oop(addr);
6627 if (p->klass_or_null() == NULL) {
6628 DEBUG_ONLY(if (!_verifying) {)
6629 // We re-dirty the cards on which this object lies and increase
6630 // the _threshold so that we'll come back to scan this object
6631 // during the preclean or remark phase. (CMSCleanOnEnter)
6632 if (CMSCleanOnEnter) {
6633 size_t sz = _collector->block_size_using_printezis_bits(addr);
6634 HeapWord* end_card_addr = (HeapWord*)round_to(
6635 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6636 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6637 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6638 // Bump _threshold to end_card_addr; note that
6639 // _threshold cannot possibly exceed end_card_addr, anyhow.
6640 // This prevents future clearing of the card as the scan proceeds
6641 // to the right.
6642 assert(_threshold <= end_card_addr,
6643 "Because we are just scanning into this object");
6644 if (_threshold < end_card_addr) {
6645 _threshold = end_card_addr;
6646 }
6647 if (p->klass_or_null() != NULL) {
6648 // Redirty the range of cards...
6649 _mut->mark_range(redirty_range);
6650 } // ...else the setting of klass will dirty the card anyway.
6651 }
6652 DEBUG_ONLY(})
6653 return true;
6654 }
6655 }
6656 scanOopsInOop(addr);
6657 return true;
6658 }
6659
6660 // We take a break if we've been at this for a while,
6661 // so as to avoid monopolizing the locks involved.
6662 void MarkFromRootsClosure::do_yield_work() {
6663 // First give up the locks, then yield, then re-lock
6664 // We should probably use a constructor/destructor idiom to
6665 // do this unlock/lock or modify the MutexUnlocker class to
6666 // serve our purpose. XXX
6667 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6668 "CMS thread should hold CMS token");
6669 assert_lock_strong(_bitMap->lock());
6670 _bitMap->lock()->unlock();
6671 ConcurrentMarkSweepThread::desynchronize(true);
6672 _collector->stopTimer();
6673 if (PrintCMSStatistics != 0) {
6674 _collector->incrementYields();
6675 }
6676
6677 // See the comment in coordinator_yield()
6678 for (unsigned i = 0; i < CMSYieldSleepCount &&
6679 ConcurrentMarkSweepThread::should_yield() &&
6680 !CMSCollector::foregroundGCIsActive(); ++i) {
6681 os::sleep(Thread::current(), 1, false);
6682 }
6683
6684 ConcurrentMarkSweepThread::synchronize(true);
6685 _bitMap->lock()->lock_without_safepoint_check();
6686 _collector->startTimer();
6687 }
6688
6689 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6690 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6691 assert(_markStack->isEmpty(),
6692 "should drain stack to limit stack usage");
6693 // convert ptr to an oop preparatory to scanning
6694 oop obj = oop(ptr);
6695 // Ignore mark word in verification below, since we
6696 // may be running concurrent with mutators.
6697 assert(obj->is_oop(true), "should be an oop");
6698 assert(_finger <= ptr, "_finger runneth ahead");
6699 // advance the finger to right end of this object
6700 _finger = ptr + obj->size();
6701 assert(_finger > ptr, "we just incremented it above");
6702 // On large heaps, it may take us some time to get through
6703 // the marking phase. During
6704 // this time it's possible that a lot of mutations have
6705 // accumulated in the card table and the mod union table --
6706 // these mutation records are redundant until we have
6707 // actually traced into the corresponding card.
6708 // Here, we check whether advancing the finger would make
6709 // us cross into a new card, and if so clear corresponding
6710 // cards in the MUT (preclean them in the card-table in the
6711 // future).
6712
6713 DEBUG_ONLY(if (!_verifying) {)
6714 // The clean-on-enter optimization is disabled by default,
6715 // until we fix 6178663.
6716 if (CMSCleanOnEnter && (_finger > _threshold)) {
6717 // [_threshold, _finger) represents the interval
6718 // of cards to be cleared in MUT (or precleaned in card table).
6719 // The set of cards to be cleared is all those that overlap
6720 // with the interval [_threshold, _finger); note that
6721 // _threshold is always kept card-aligned but _finger isn't
6722 // always card-aligned.
6723 HeapWord* old_threshold = _threshold;
6724 assert(old_threshold == (HeapWord*)round_to(
6725 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6726 "_threshold should always be card-aligned");
6727 _threshold = (HeapWord*)round_to(
6728 (intptr_t)_finger, CardTableModRefBS::card_size);
6729 MemRegion mr(old_threshold, _threshold);
6730 assert(!mr.is_empty(), "Control point invariant");
6731 assert(_span.contains(mr), "Should clear within span");
6732 _mut->clear_range(mr);
6733 }
6734 DEBUG_ONLY(})
6735 // Note: the finger doesn't advance while we drain
6736 // the stack below.
6737 PushOrMarkClosure pushOrMarkClosure(_collector,
6738 _span, _bitMap, _markStack,
6739 _finger, this);
6740 bool res = _markStack->push(obj);
6741 assert(res, "Empty non-zero size stack should have space for single push");
6742 while (!_markStack->isEmpty()) {
6743 oop new_oop = _markStack->pop();
6744 // Skip verifying header mark word below because we are
6745 // running concurrent with mutators.
6746 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6747 // now scan this oop's oops
6748 new_oop->oop_iterate(&pushOrMarkClosure);
6749 do_yield_check();
6750 }
6751 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6752 }
6753
6754 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
6755 CMSCollector* collector, MemRegion span,
6756 CMSBitMap* bit_map,
6757 OopTaskQueue* work_queue,
6758 CMSMarkStack* overflow_stack):
6759 _collector(collector),
6760 _whole_span(collector->_span),
6761 _span(span),
6762 _bit_map(bit_map),
6763 _mut(&collector->_modUnionTable),
6764 _work_queue(work_queue),
6765 _overflow_stack(overflow_stack),
6766 _skip_bits(0),
6767 _task(task)
6768 {
6769 assert(_work_queue->size() == 0, "work_queue should be empty");
6770 _finger = span.start();
6771 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
6772 assert(_span.contains(_finger), "Out of bounds _finger?");
6773 }
6774
6775 // Should revisit to see if this should be restructured for
6776 // greater efficiency.
6777 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
6778 if (_skip_bits > 0) {
6779 _skip_bits--;
6780 return true;
6781 }
6782 // convert offset into a HeapWord*
6783 HeapWord* addr = _bit_map->startWord() + offset;
6784 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6785 "address out of range");
6786 assert(_bit_map->isMarked(addr), "tautology");
6787 if (_bit_map->isMarked(addr+1)) {
6788 // this is an allocated object that might not yet be initialized
6789 assert(_skip_bits == 0, "tautology");
6790 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
6791 oop p = oop(addr);
6792 if (p->klass_or_null() == NULL) {
6793 // in the case of Clean-on-Enter optimization, redirty card
6794 // and avoid clearing card by increasing the threshold.
6795 return true;
6796 }
6797 }
6798 scan_oops_in_oop(addr);
6799 return true;
6800 }
6801
6802 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6803 assert(_bit_map->isMarked(ptr), "expected bit to be set");
6804 // Should we assert that our work queue is empty or
6805 // below some drain limit?
6806 assert(_work_queue->size() == 0,
6807 "should drain stack to limit stack usage");
6808 // convert ptr to an oop preparatory to scanning
6809 oop obj = oop(ptr);
6810 // Ignore mark word in verification below, since we
6811 // may be running concurrent with mutators.
6812 assert(obj->is_oop(true), "should be an oop");
6813 assert(_finger <= ptr, "_finger runneth ahead");
6814 // advance the finger to right end of this object
6815 _finger = ptr + obj->size();
6816 assert(_finger > ptr, "we just incremented it above");
6817 // On large heaps, it may take us some time to get through
6818 // the marking phase. During
6819 // this time it's possible that a lot of mutations have
6820 // accumulated in the card table and the mod union table --
6821 // these mutation records are redundant until we have
6822 // actually traced into the corresponding card.
6823 // Here, we check whether advancing the finger would make
6824 // us cross into a new card, and if so clear corresponding
6825 // cards in the MUT (preclean them in the card-table in the
6826 // future).
6827
6828 // The clean-on-enter optimization is disabled by default,
6829 // until we fix 6178663.
6830 if (CMSCleanOnEnter && (_finger > _threshold)) {
6831 // [_threshold, _finger) represents the interval
6832 // of cards to be cleared in MUT (or precleaned in card table).
6833 // The set of cards to be cleared is all those that overlap
6834 // with the interval [_threshold, _finger); note that
6835 // _threshold is always kept card-aligned but _finger isn't
6836 // always card-aligned.
6837 HeapWord* old_threshold = _threshold;
6838 assert(old_threshold == (HeapWord*)round_to(
6839 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6840 "_threshold should always be card-aligned");
6841 _threshold = (HeapWord*)round_to(
6842 (intptr_t)_finger, CardTableModRefBS::card_size);
6843 MemRegion mr(old_threshold, _threshold);
6844 assert(!mr.is_empty(), "Control point invariant");
6845 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6846 _mut->clear_range(mr);
6847 }
6848
6849 // Note: the local finger doesn't advance while we drain
6850 // the stack below, but the global finger sure can and will.
6851 HeapWord** gfa = _task->global_finger_addr();
6852 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
6853 _span, _bit_map,
6854 _work_queue,
6855 _overflow_stack,
6856 _finger,
6857 gfa, this);
6858 bool res = _work_queue->push(obj); // overflow could occur here
6859 assert(res, "Will hold once we use workqueues");
6860 while (true) {
6861 oop new_oop;
6862 if (!_work_queue->pop_local(new_oop)) {
6863 // We emptied our work_queue; check if there's stuff that can
6864 // be gotten from the overflow stack.
6865 if (CMSConcMarkingTask::get_work_from_overflow_stack(
6866 _overflow_stack, _work_queue)) {
6867 do_yield_check();
6868 continue;
6869 } else { // done
6870 break;
6871 }
6872 }
6873 // Skip verifying header mark word below because we are
6874 // running concurrent with mutators.
6875 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6876 // now scan this oop's oops
6877 new_oop->oop_iterate(&pushOrMarkClosure);
6878 do_yield_check();
6879 }
6880 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6881 }
6882
6883 // Yield in response to a request from VM Thread or
6884 // from mutators.
6885 void Par_MarkFromRootsClosure::do_yield_work() {
6886 assert(_task != NULL, "sanity");
6887 _task->yield();
6888 }
6889
6890 // A variant of the above used for verifying CMS marking work.
6891 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6892 MemRegion span,
6893 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6894 CMSMarkStack* mark_stack):
6895 _collector(collector),
6896 _span(span),
6897 _verification_bm(verification_bm),
6898 _cms_bm(cms_bm),
6899 _mark_stack(mark_stack),
6900 _pam_verify_closure(collector, span, verification_bm, cms_bm,
6901 mark_stack)
6902 {
6903 assert(_mark_stack->isEmpty(), "stack should be empty");
6904 _finger = _verification_bm->startWord();
6905 assert(_collector->_restart_addr == NULL, "Sanity check");
6906 assert(_span.contains(_finger), "Out of bounds _finger?");
6907 }
6908
6909 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6910 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6911 assert(_span.contains(addr), "Out of bounds _finger?");
6912 _finger = addr;
6913 }
6914
6915 // Should revisit to see if this should be restructured for
6916 // greater efficiency.
6917 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6918 // convert offset into a HeapWord*
6919 HeapWord* addr = _verification_bm->startWord() + offset;
6920 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6921 "address out of range");
6922 assert(_verification_bm->isMarked(addr), "tautology");
6923 assert(_cms_bm->isMarked(addr), "tautology");
6924
6925 assert(_mark_stack->isEmpty(),
6926 "should drain stack to limit stack usage");
6927 // convert addr to an oop preparatory to scanning
6928 oop obj = oop(addr);
6929 assert(obj->is_oop(), "should be an oop");
6930 assert(_finger <= addr, "_finger runneth ahead");
6931 // advance the finger to right end of this object
6932 _finger = addr + obj->size();
6933 assert(_finger > addr, "we just incremented it above");
6934 // Note: the finger doesn't advance while we drain
6935 // the stack below.
6936 bool res = _mark_stack->push(obj);
6937 assert(res, "Empty non-zero size stack should have space for single push");
6938 while (!_mark_stack->isEmpty()) {
6939 oop new_oop = _mark_stack->pop();
6940 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
6941 // now scan this oop's oops
6942 new_oop->oop_iterate(&_pam_verify_closure);
6943 }
6944 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6945 return true;
6946 }
6947
6948 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6949 CMSCollector* collector, MemRegion span,
6950 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6951 CMSMarkStack* mark_stack):
6952 MetadataAwareOopClosure(collector->ref_processor()),
6953 _collector(collector),
6954 _span(span),
6955 _verification_bm(verification_bm),
6956 _cms_bm(cms_bm),
6957 _mark_stack(mark_stack)
6958 { }
6959
6960 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6961 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6962
6963 // Upon stack overflow, we discard (part of) the stack,
6964 // remembering the least address amongst those discarded
6965 // in CMSCollector's _restart_address.
6966 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6967 // Remember the least grey address discarded
6968 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6969 _collector->lower_restart_addr(ra);
6970 _mark_stack->reset(); // discard stack contents
6971 _mark_stack->expand(); // expand the stack if possible
6972 }
6973
6974 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6975 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
6976 HeapWord* addr = (HeapWord*)obj;
6977 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6978 // Oop lies in _span and isn't yet grey or black
6979 _verification_bm->mark(addr); // now grey
6980 if (!_cms_bm->isMarked(addr)) {
6981 oop(addr)->print();
6982 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
6983 addr);
6984 fatal("... aborting");
6985 }
6986
6987 if (!_mark_stack->push(obj)) { // stack overflow
6988 if (PrintCMSStatistics != 0) {
6989 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
6990 SIZE_FORMAT, _mark_stack->capacity());
6991 }
6992 assert(_mark_stack->isFull(), "Else push should have succeeded");
6993 handle_stack_overflow(addr);
6994 }
6995 // anything including and to the right of _finger
6996 // will be scanned as we iterate over the remainder of the
6997 // bit map
6998 }
6999 }
7000
7001 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7002 MemRegion span,
7003 CMSBitMap* bitMap, CMSMarkStack* markStack,
7004 HeapWord* finger, MarkFromRootsClosure* parent) :
7005 MetadataAwareOopClosure(collector->ref_processor()),
7006 _collector(collector),
7007 _span(span),
7008 _bitMap(bitMap),
7009 _markStack(markStack),
7010 _finger(finger),
7011 _parent(parent)
7012 { }
7013
7014 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7015 MemRegion span,
7016 CMSBitMap* bit_map,
7017 OopTaskQueue* work_queue,
7018 CMSMarkStack* overflow_stack,
7019 HeapWord* finger,
7020 HeapWord** global_finger_addr,
7021 Par_MarkFromRootsClosure* parent) :
7022 MetadataAwareOopClosure(collector->ref_processor()),
7023 _collector(collector),
7024 _whole_span(collector->_span),
7025 _span(span),
7026 _bit_map(bit_map),
7027 _work_queue(work_queue),
7028 _overflow_stack(overflow_stack),
7029 _finger(finger),
7030 _global_finger_addr(global_finger_addr),
7031 _parent(parent)
7032 { }
7033
7034 // Assumes thread-safe access by callers, who are
7035 // responsible for mutual exclusion.
7036 void CMSCollector::lower_restart_addr(HeapWord* low) {
7037 assert(_span.contains(low), "Out of bounds addr");
7038 if (_restart_addr == NULL) {
7039 _restart_addr = low;
7040 } else {
7041 _restart_addr = MIN2(_restart_addr, low);
7042 }
7043 }
7044
7045 // Upon stack overflow, we discard (part of) the stack,
7046 // remembering the least address amongst those discarded
7047 // in CMSCollector's _restart_address.
7048 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7049 // Remember the least grey address discarded
7050 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7051 _collector->lower_restart_addr(ra);
7052 _markStack->reset(); // discard stack contents
7053 _markStack->expand(); // expand the stack if possible
7054 }
7055
7056 // Upon stack overflow, we discard (part of) the stack,
7057 // remembering the least address amongst those discarded
7058 // in CMSCollector's _restart_address.
7059 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7060 // We need to do this under a mutex to prevent other
7061 // workers from interfering with the work done below.
7062 MutexLockerEx ml(_overflow_stack->par_lock(),
7063 Mutex::_no_safepoint_check_flag);
7064 // Remember the least grey address discarded
7065 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7066 _collector->lower_restart_addr(ra);
7067 _overflow_stack->reset(); // discard stack contents
7068 _overflow_stack->expand(); // expand the stack if possible
7069 }
7070
7071 void PushOrMarkClosure::do_oop(oop obj) {
7072 // Ignore mark word because we are running concurrent with mutators.
7073 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7074 HeapWord* addr = (HeapWord*)obj;
7075 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7076 // Oop lies in _span and isn't yet grey or black
7077 _bitMap->mark(addr); // now grey
7078 if (addr < _finger) {
7079 // the bit map iteration has already either passed, or
7080 // sampled, this bit in the bit map; we'll need to
7081 // use the marking stack to scan this oop's oops.
7082 bool simulate_overflow = false;
7083 NOT_PRODUCT(
7084 if (CMSMarkStackOverflowALot &&
7085 _collector->simulate_overflow()) {
7086 // simulate a stack overflow
7087 simulate_overflow = true;
7088 }
7089 )
7090 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7091 if (PrintCMSStatistics != 0) {
7092 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7093 SIZE_FORMAT, _markStack->capacity());
7094 }
7095 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7096 handle_stack_overflow(addr);
7097 }
7098 }
7099 // anything including and to the right of _finger
7100 // will be scanned as we iterate over the remainder of the
7101 // bit map
7102 do_yield_check();
7103 }
7104 }
7105
7106 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7107 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7108
7109 void Par_PushOrMarkClosure::do_oop(oop obj) {
7110 // Ignore mark word because we are running concurrent with mutators.
7111 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7112 HeapWord* addr = (HeapWord*)obj;
7113 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7114 // Oop lies in _span and isn't yet grey or black
7115 // We read the global_finger (volatile read) strictly after marking oop
7116 bool res = _bit_map->par_mark(addr); // now grey
7117 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7118 // Should we push this marked oop on our stack?
7119 // -- if someone else marked it, nothing to do
7120 // -- if target oop is above global finger nothing to do
7121 // -- if target oop is in chunk and above local finger
7122 // then nothing to do
7123 // -- else push on work queue
7124 if ( !res // someone else marked it, they will deal with it
7125 || (addr >= *gfa) // will be scanned in a later task
7126 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7127 return;
7128 }
7129 // the bit map iteration has already either passed, or
7130 // sampled, this bit in the bit map; we'll need to
7131 // use the marking stack to scan this oop's oops.
7132 bool simulate_overflow = false;
7133 NOT_PRODUCT(
7134 if (CMSMarkStackOverflowALot &&
7135 _collector->simulate_overflow()) {
7136 // simulate a stack overflow
7137 simulate_overflow = true;
7138 }
7139 )
7140 if (simulate_overflow ||
7141 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7142 // stack overflow
7143 if (PrintCMSStatistics != 0) {
7144 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7145 SIZE_FORMAT, _overflow_stack->capacity());
7146 }
7147 // We cannot assert that the overflow stack is full because
7148 // it may have been emptied since.
7149 assert(simulate_overflow ||
7150 _work_queue->size() == _work_queue->max_elems(),
7151 "Else push should have succeeded");
7152 handle_stack_overflow(addr);
7153 }
7154 do_yield_check();
7155 }
7156 }
7157
7158 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7159 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7160
7161 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7162 MemRegion span,
7163 ReferenceProcessor* rp,
7164 CMSBitMap* bit_map,
7165 CMSBitMap* mod_union_table,
7166 CMSMarkStack* mark_stack,
7167 bool concurrent_precleaning):
7168 MetadataAwareOopClosure(rp),
7169 _collector(collector),
7170 _span(span),
7171 _bit_map(bit_map),
7172 _mod_union_table(mod_union_table),
7173 _mark_stack(mark_stack),
7174 _concurrent_precleaning(concurrent_precleaning)
7175 {
7176 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7177 }
7178
7179 // Grey object rescan during pre-cleaning and second checkpoint phases --
7180 // the non-parallel version (the parallel version appears further below.)
7181 void PushAndMarkClosure::do_oop(oop obj) {
7182 // Ignore mark word verification. If during concurrent precleaning,
7183 // the object monitor may be locked. If during the checkpoint
7184 // phases, the object may already have been reached by a different
7185 // path and may be at the end of the global overflow list (so
7186 // the mark word may be NULL).
7187 assert(obj->is_oop_or_null(true /* ignore mark word */),
7188 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7189 HeapWord* addr = (HeapWord*)obj;
7190 // Check if oop points into the CMS generation
7191 // and is not marked
7192 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7193 // a white object ...
7194 _bit_map->mark(addr); // ... now grey
7195 // push on the marking stack (grey set)
7196 bool simulate_overflow = false;
7197 NOT_PRODUCT(
7198 if (CMSMarkStackOverflowALot &&
7199 _collector->simulate_overflow()) {
7200 // simulate a stack overflow
7201 simulate_overflow = true;
7202 }
7203 )
7204 if (simulate_overflow || !_mark_stack->push(obj)) {
7205 if (_concurrent_precleaning) {
7206 // During precleaning we can just dirty the appropriate card(s)
7207 // in the mod union table, thus ensuring that the object remains
7208 // in the grey set and continue. In the case of object arrays
7209 // we need to dirty all of the cards that the object spans,
7210 // since the rescan of object arrays will be limited to the
7211 // dirty cards.
7212 // Note that no one can be interfering with us in this action
7213 // of dirtying the mod union table, so no locking or atomics
7214 // are required.
7215 if (obj->is_objArray()) {
7216 size_t sz = obj->size();
7217 HeapWord* end_card_addr = (HeapWord*)round_to(
7218 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7219 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7220 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7221 _mod_union_table->mark_range(redirty_range);
7222 } else {
7223 _mod_union_table->mark(addr);
7224 }
7225 _collector->_ser_pmc_preclean_ovflw++;
7226 } else {
7227 // During the remark phase, we need to remember this oop
7228 // in the overflow list.
7229 _collector->push_on_overflow_list(obj);
7230 _collector->_ser_pmc_remark_ovflw++;
7231 }
7232 }
7233 }
7234 }
7235
7236 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7237 MemRegion span,
7238 ReferenceProcessor* rp,
7239 CMSBitMap* bit_map,
7240 OopTaskQueue* work_queue):
7241 MetadataAwareOopClosure(rp),
7242 _collector(collector),
7243 _span(span),
7244 _bit_map(bit_map),
7245 _work_queue(work_queue)
7246 {
7247 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7248 }
7249
7250 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7251 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7252
7253 // Grey object rescan during second checkpoint phase --
7254 // the parallel version.
7255 void Par_PushAndMarkClosure::do_oop(oop obj) {
7256 // In the assert below, we ignore the mark word because
7257 // this oop may point to an already visited object that is
7258 // on the overflow stack (in which case the mark word has
7259 // been hijacked for chaining into the overflow stack --
7260 // if this is the last object in the overflow stack then
7261 // its mark word will be NULL). Because this object may
7262 // have been subsequently popped off the global overflow
7263 // stack, and the mark word possibly restored to the prototypical
7264 // value, by the time we get to examined this failing assert in
7265 // the debugger, is_oop_or_null(false) may subsequently start
7266 // to hold.
7267 assert(obj->is_oop_or_null(true),
7268 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7269 HeapWord* addr = (HeapWord*)obj;
7270 // Check if oop points into the CMS generation
7271 // and is not marked
7272 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7273 // a white object ...
7274 // If we manage to "claim" the object, by being the
7275 // first thread to mark it, then we push it on our
7276 // marking stack
7277 if (_bit_map->par_mark(addr)) { // ... now grey
7278 // push on work queue (grey set)
7279 bool simulate_overflow = false;
7280 NOT_PRODUCT(
7281 if (CMSMarkStackOverflowALot &&
7282 _collector->par_simulate_overflow()) {
7283 // simulate a stack overflow
7284 simulate_overflow = true;
7285 }
7286 )
7287 if (simulate_overflow || !_work_queue->push(obj)) {
7288 _collector->par_push_on_overflow_list(obj);
7289 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7290 }
7291 } // Else, some other thread got there first
7292 }
7293 }
7294
7295 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7296 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7297
7298 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7299 Mutex* bml = _collector->bitMapLock();
7300 assert_lock_strong(bml);
7301 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7302 "CMS thread should hold CMS token");
7303
7304 bml->unlock();
7305 ConcurrentMarkSweepThread::desynchronize(true);
7306
7307 _collector->stopTimer();
7308 if (PrintCMSStatistics != 0) {
7309 _collector->incrementYields();
7310 }
7311
7312 // See the comment in coordinator_yield()
7313 for (unsigned i = 0; i < CMSYieldSleepCount &&
7314 ConcurrentMarkSweepThread::should_yield() &&
7315 !CMSCollector::foregroundGCIsActive(); ++i) {
7316 os::sleep(Thread::current(), 1, false);
7317 }
7318
7319 ConcurrentMarkSweepThread::synchronize(true);
7320 bml->lock();
7321
7322 _collector->startTimer();
7323 }
7324
7325 bool CMSPrecleanRefsYieldClosure::should_return() {
7326 if (ConcurrentMarkSweepThread::should_yield()) {
7327 do_yield_work();
7328 }
7329 return _collector->foregroundGCIsActive();
7330 }
7331
7332 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7333 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7334 "mr should be aligned to start at a card boundary");
7335 // We'd like to assert:
7336 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7337 // "mr should be a range of cards");
7338 // However, that would be too strong in one case -- the last
7339 // partition ends at _unallocated_block which, in general, can be
7340 // an arbitrary boundary, not necessarily card aligned.
7341 if (PrintCMSStatistics != 0) {
7342 _num_dirty_cards +=
7343 mr.word_size()/CardTableModRefBS::card_size_in_words;
7344 }
7345 _space->object_iterate_mem(mr, &_scan_cl);
7346 }
7347
7348 SweepClosure::SweepClosure(CMSCollector* collector,
7349 ConcurrentMarkSweepGeneration* g,
7350 CMSBitMap* bitMap, bool should_yield) :
7351 _collector(collector),
7352 _g(g),
7353 _sp(g->cmsSpace()),
7354 _limit(_sp->sweep_limit()),
7355 _freelistLock(_sp->freelistLock()),
7356 _bitMap(bitMap),
7357 _yield(should_yield),
7358 _inFreeRange(false), // No free range at beginning of sweep
7359 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7360 _lastFreeRangeCoalesced(false),
7361 _freeFinger(g->used_region().start())
7362 {
7363 NOT_PRODUCT(
7364 _numObjectsFreed = 0;
7365 _numWordsFreed = 0;
7366 _numObjectsLive = 0;
7367 _numWordsLive = 0;
7368 _numObjectsAlreadyFree = 0;
7369 _numWordsAlreadyFree = 0;
7370 _last_fc = NULL;
7371
7372 _sp->initializeIndexedFreeListArrayReturnedBytes();
7373 _sp->dictionary()->initialize_dict_returned_bytes();
7374 )
7375 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7376 "sweep _limit out of bounds");
7377 if (CMSTraceSweeper) {
7378 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
7379 _limit);
7380 }
7381 }
7382
7383 void SweepClosure::print_on(outputStream* st) const {
7384 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7385 _sp->bottom(), _sp->end());
7386 tty->print_cr("_limit = " PTR_FORMAT, _limit);
7387 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger);
7388 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);)
7389 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7390 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7391 }
7392
7393 #ifndef PRODUCT
7394 // Assertion checking only: no useful work in product mode --
7395 // however, if any of the flags below become product flags,
7396 // you may need to review this code to see if it needs to be
7397 // enabled in product mode.
7398 SweepClosure::~SweepClosure() {
7399 assert_lock_strong(_freelistLock);
7400 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7401 "sweep _limit out of bounds");
7402 if (inFreeRange()) {
7403 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
7404 print();
7405 ShouldNotReachHere();
7406 }
7407 if (Verbose && PrintGC) {
7408 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes",
7409 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7410 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
7411 SIZE_FORMAT" bytes "
7412 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
7413 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7414 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7415 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
7416 * sizeof(HeapWord);
7417 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
7418
7419 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7420 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7421 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7422 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7423 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes);
7424 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
7425 indexListReturnedBytes);
7426 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
7427 dict_returned_bytes);
7428 }
7429 }
7430 if (CMSTraceSweeper) {
7431 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
7432 _limit);
7433 }
7434 }
7435 #endif // PRODUCT
7436
7437 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7438 bool freeRangeInFreeLists) {
7439 if (CMSTraceSweeper) {
7440 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n",
7441 freeFinger, freeRangeInFreeLists);
7442 }
7443 assert(!inFreeRange(), "Trampling existing free range");
7444 set_inFreeRange(true);
7445 set_lastFreeRangeCoalesced(false);
7446
7447 set_freeFinger(freeFinger);
7448 set_freeRangeInFreeLists(freeRangeInFreeLists);
7449 if (CMSTestInFreeList) {
7450 if (freeRangeInFreeLists) {
7451 FreeChunk* fc = (FreeChunk*) freeFinger;
7452 assert(fc->is_free(), "A chunk on the free list should be free.");
7453 assert(fc->size() > 0, "Free range should have a size");
7454 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7455 }
7456 }
7457 }
7458
7459 // Note that the sweeper runs concurrently with mutators. Thus,
7460 // it is possible for direct allocation in this generation to happen
7461 // in the middle of the sweep. Note that the sweeper also coalesces
7462 // contiguous free blocks. Thus, unless the sweeper and the allocator
7463 // synchronize appropriately freshly allocated blocks may get swept up.
7464 // This is accomplished by the sweeper locking the free lists while
7465 // it is sweeping. Thus blocks that are determined to be free are
7466 // indeed free. There is however one additional complication:
7467 // blocks that have been allocated since the final checkpoint and
7468 // mark, will not have been marked and so would be treated as
7469 // unreachable and swept up. To prevent this, the allocator marks
7470 // the bit map when allocating during the sweep phase. This leads,
7471 // however, to a further complication -- objects may have been allocated
7472 // but not yet initialized -- in the sense that the header isn't yet
7473 // installed. The sweeper can not then determine the size of the block
7474 // in order to skip over it. To deal with this case, we use a technique
7475 // (due to Printezis) to encode such uninitialized block sizes in the
7476 // bit map. Since the bit map uses a bit per every HeapWord, but the
7477 // CMS generation has a minimum object size of 3 HeapWords, it follows
7478 // that "normal marks" won't be adjacent in the bit map (there will
7479 // always be at least two 0 bits between successive 1 bits). We make use
7480 // of these "unused" bits to represent uninitialized blocks -- the bit
7481 // corresponding to the start of the uninitialized object and the next
7482 // bit are both set. Finally, a 1 bit marks the end of the object that
7483 // started with the two consecutive 1 bits to indicate its potentially
7484 // uninitialized state.
7485
7486 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7487 FreeChunk* fc = (FreeChunk*)addr;
7488 size_t res;
7489
7490 // Check if we are done sweeping. Below we check "addr >= _limit" rather
7491 // than "addr == _limit" because although _limit was a block boundary when
7492 // we started the sweep, it may no longer be one because heap expansion
7493 // may have caused us to coalesce the block ending at the address _limit
7494 // with a newly expanded chunk (this happens when _limit was set to the
7495 // previous _end of the space), so we may have stepped past _limit:
7496 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7497 if (addr >= _limit) { // we have swept up to or past the limit: finish up
7498 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7499 "sweep _limit out of bounds");
7500 assert(addr < _sp->end(), "addr out of bounds");
7501 // Flush any free range we might be holding as a single
7502 // coalesced chunk to the appropriate free list.
7503 if (inFreeRange()) {
7504 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7505 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger()));
7506 flush_cur_free_chunk(freeFinger(),
7507 pointer_delta(addr, freeFinger()));
7508 if (CMSTraceSweeper) {
7509 gclog_or_tty->print("Sweep: last chunk: ");
7510 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") "
7511 "[coalesced:%d]\n",
7512 freeFinger(), pointer_delta(addr, freeFinger()),
7513 lastFreeRangeCoalesced() ? 1 : 0);
7514 }
7515 }
7516
7517 // help the iterator loop finish
7518 return pointer_delta(_sp->end(), addr);
7519 }
7520
7521 assert(addr < _limit, "sweep invariant");
7522 // check if we should yield
7523 do_yield_check(addr);
7524 if (fc->is_free()) {
7525 // Chunk that is already free
7526 res = fc->size();
7527 do_already_free_chunk(fc);
7528 debug_only(_sp->verifyFreeLists());
7529 // If we flush the chunk at hand in lookahead_and_flush()
7530 // and it's coalesced with a preceding chunk, then the
7531 // process of "mangling" the payload of the coalesced block
7532 // will cause erasure of the size information from the
7533 // (erstwhile) header of all the coalesced blocks but the
7534 // first, so the first disjunct in the assert will not hold
7535 // in that specific case (in which case the second disjunct
7536 // will hold).
7537 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7538 "Otherwise the size info doesn't change at this step");
7539 NOT_PRODUCT(
7540 _numObjectsAlreadyFree++;
7541 _numWordsAlreadyFree += res;
7542 )
7543 NOT_PRODUCT(_last_fc = fc;)
7544 } else if (!_bitMap->isMarked(addr)) {
7545 // Chunk is fresh garbage
7546 res = do_garbage_chunk(fc);
7547 debug_only(_sp->verifyFreeLists());
7548 NOT_PRODUCT(
7549 _numObjectsFreed++;
7550 _numWordsFreed += res;
7551 )
7552 } else {
7553 // Chunk that is alive.
7554 res = do_live_chunk(fc);
7555 debug_only(_sp->verifyFreeLists());
7556 NOT_PRODUCT(
7557 _numObjectsLive++;
7558 _numWordsLive += res;
7559 )
7560 }
7561 return res;
7562 }
7563
7564 // For the smart allocation, record following
7565 // split deaths - a free chunk is removed from its free list because
7566 // it is being split into two or more chunks.
7567 // split birth - a free chunk is being added to its free list because
7568 // a larger free chunk has been split and resulted in this free chunk.
7569 // coal death - a free chunk is being removed from its free list because
7570 // it is being coalesced into a large free chunk.
7571 // coal birth - a free chunk is being added to its free list because
7572 // it was created when two or more free chunks where coalesced into
7573 // this free chunk.
7574 //
7575 // These statistics are used to determine the desired number of free
7576 // chunks of a given size. The desired number is chosen to be relative
7577 // to the end of a CMS sweep. The desired number at the end of a sweep
7578 // is the
7579 // count-at-end-of-previous-sweep (an amount that was enough)
7580 // - count-at-beginning-of-current-sweep (the excess)
7581 // + split-births (gains in this size during interval)
7582 // - split-deaths (demands on this size during interval)
7583 // where the interval is from the end of one sweep to the end of the
7584 // next.
7585 //
7586 // When sweeping the sweeper maintains an accumulated chunk which is
7587 // the chunk that is made up of chunks that have been coalesced. That
7588 // will be termed the left-hand chunk. A new chunk of garbage that
7589 // is being considered for coalescing will be referred to as the
7590 // right-hand chunk.
7591 //
7592 // When making a decision on whether to coalesce a right-hand chunk with
7593 // the current left-hand chunk, the current count vs. the desired count
7594 // of the left-hand chunk is considered. Also if the right-hand chunk
7595 // is near the large chunk at the end of the heap (see
7596 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7597 // left-hand chunk is coalesced.
7598 //
7599 // When making a decision about whether to split a chunk, the desired count
7600 // vs. the current count of the candidate to be split is also considered.
7601 // If the candidate is underpopulated (currently fewer chunks than desired)
7602 // a chunk of an overpopulated (currently more chunks than desired) size may
7603 // be chosen. The "hint" associated with a free list, if non-null, points
7604 // to a free list which may be overpopulated.
7605 //
7606
7607 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7608 const size_t size = fc->size();
7609 // Chunks that cannot be coalesced are not in the
7610 // free lists.
7611 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7612 assert(_sp->verify_chunk_in_free_list(fc),
7613 "free chunk should be in free lists");
7614 }
7615 // a chunk that is already free, should not have been
7616 // marked in the bit map
7617 HeapWord* const addr = (HeapWord*) fc;
7618 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7619 // Verify that the bit map has no bits marked between
7620 // addr and purported end of this block.
7621 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7622
7623 // Some chunks cannot be coalesced under any circumstances.
7624 // See the definition of cantCoalesce().
7625 if (!fc->cantCoalesce()) {
7626 // This chunk can potentially be coalesced.
7627 if (_sp->adaptive_freelists()) {
7628 // All the work is done in
7629 do_post_free_or_garbage_chunk(fc, size);
7630 } else { // Not adaptive free lists
7631 // this is a free chunk that can potentially be coalesced by the sweeper;
7632 if (!inFreeRange()) {
7633 // if the next chunk is a free block that can't be coalesced
7634 // it doesn't make sense to remove this chunk from the free lists
7635 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
7636 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
7637 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
7638 nextChunk->is_free() && // ... which is free...
7639 nextChunk->cantCoalesce()) { // ... but can't be coalesced
7640 // nothing to do
7641 } else {
7642 // Potentially the start of a new free range:
7643 // Don't eagerly remove it from the free lists.
7644 // No need to remove it if it will just be put
7645 // back again. (Also from a pragmatic point of view
7646 // if it is a free block in a region that is beyond
7647 // any allocated blocks, an assertion will fail)
7648 // Remember the start of a free run.
7649 initialize_free_range(addr, true);
7650 // end - can coalesce with next chunk
7651 }
7652 } else {
7653 // the midst of a free range, we are coalescing
7654 print_free_block_coalesced(fc);
7655 if (CMSTraceSweeper) {
7656 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", fc, size);
7657 }
7658 // remove it from the free lists
7659 _sp->removeFreeChunkFromFreeLists(fc);
7660 set_lastFreeRangeCoalesced(true);
7661 // If the chunk is being coalesced and the current free range is
7662 // in the free lists, remove the current free range so that it
7663 // will be returned to the free lists in its entirety - all
7664 // the coalesced pieces included.
7665 if (freeRangeInFreeLists()) {
7666 FreeChunk* ffc = (FreeChunk*) freeFinger();
7667 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7668 "Size of free range is inconsistent with chunk size.");
7669 if (CMSTestInFreeList) {
7670 assert(_sp->verify_chunk_in_free_list(ffc),
7671 "free range is not in free lists");
7672 }
7673 _sp->removeFreeChunkFromFreeLists(ffc);
7674 set_freeRangeInFreeLists(false);
7675 }
7676 }
7677 }
7678 // Note that if the chunk is not coalescable (the else arm
7679 // below), we unconditionally flush, without needing to do
7680 // a "lookahead," as we do below.
7681 if (inFreeRange()) lookahead_and_flush(fc, size);
7682 } else {
7683 // Code path common to both original and adaptive free lists.
7684
7685 // cant coalesce with previous block; this should be treated
7686 // as the end of a free run if any
7687 if (inFreeRange()) {
7688 // we kicked some butt; time to pick up the garbage
7689 assert(freeFinger() < addr, "freeFinger points too high");
7690 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7691 }
7692 // else, nothing to do, just continue
7693 }
7694 }
7695
7696 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7697 // This is a chunk of garbage. It is not in any free list.
7698 // Add it to a free list or let it possibly be coalesced into
7699 // a larger chunk.
7700 HeapWord* const addr = (HeapWord*) fc;
7701 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7702
7703 if (_sp->adaptive_freelists()) {
7704 // Verify that the bit map has no bits marked between
7705 // addr and purported end of just dead object.
7706 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7707
7708 do_post_free_or_garbage_chunk(fc, size);
7709 } else {
7710 if (!inFreeRange()) {
7711 // start of a new free range
7712 assert(size > 0, "A free range should have a size");
7713 initialize_free_range(addr, false);
7714 } else {
7715 // this will be swept up when we hit the end of the
7716 // free range
7717 if (CMSTraceSweeper) {
7718 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", fc, size);
7719 }
7720 // If the chunk is being coalesced and the current free range is
7721 // in the free lists, remove the current free range so that it
7722 // will be returned to the free lists in its entirety - all
7723 // the coalesced pieces included.
7724 if (freeRangeInFreeLists()) {
7725 FreeChunk* ffc = (FreeChunk*)freeFinger();
7726 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7727 "Size of free range is inconsistent with chunk size.");
7728 if (CMSTestInFreeList) {
7729 assert(_sp->verify_chunk_in_free_list(ffc),
7730 "free range is not in free lists");
7731 }
7732 _sp->removeFreeChunkFromFreeLists(ffc);
7733 set_freeRangeInFreeLists(false);
7734 }
7735 set_lastFreeRangeCoalesced(true);
7736 }
7737 // this will be swept up when we hit the end of the free range
7738
7739 // Verify that the bit map has no bits marked between
7740 // addr and purported end of just dead object.
7741 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7742 }
7743 assert(_limit >= addr + size,
7744 "A freshly garbage chunk can't possibly straddle over _limit");
7745 if (inFreeRange()) lookahead_and_flush(fc, size);
7746 return size;
7747 }
7748
7749 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7750 HeapWord* addr = (HeapWord*) fc;
7751 // The sweeper has just found a live object. Return any accumulated
7752 // left hand chunk to the free lists.
7753 if (inFreeRange()) {
7754 assert(freeFinger() < addr, "freeFinger points too high");
7755 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7756 }
7757
7758 // This object is live: we'd normally expect this to be
7759 // an oop, and like to assert the following:
7760 // assert(oop(addr)->is_oop(), "live block should be an oop");
7761 // However, as we commented above, this may be an object whose
7762 // header hasn't yet been initialized.
7763 size_t size;
7764 assert(_bitMap->isMarked(addr), "Tautology for this control point");
7765 if (_bitMap->isMarked(addr + 1)) {
7766 // Determine the size from the bit map, rather than trying to
7767 // compute it from the object header.
7768 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7769 size = pointer_delta(nextOneAddr + 1, addr);
7770 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7771 "alignment problem");
7772
7773 #ifdef ASSERT
7774 if (oop(addr)->klass_or_null() != NULL) {
7775 // Ignore mark word because we are running concurrent with mutators
7776 assert(oop(addr)->is_oop(true), "live block should be an oop");
7777 assert(size ==
7778 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7779 "P-mark and computed size do not agree");
7780 }
7781 #endif
7782
7783 } else {
7784 // This should be an initialized object that's alive.
7785 assert(oop(addr)->klass_or_null() != NULL,
7786 "Should be an initialized object");
7787 // Ignore mark word because we are running concurrent with mutators
7788 assert(oop(addr)->is_oop(true), "live block should be an oop");
7789 // Verify that the bit map has no bits marked between
7790 // addr and purported end of this block.
7791 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7792 assert(size >= 3, "Necessary for Printezis marks to work");
7793 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7794 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7795 }
7796 return size;
7797 }
7798
7799 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7800 size_t chunkSize) {
7801 // do_post_free_or_garbage_chunk() should only be called in the case
7802 // of the adaptive free list allocator.
7803 const bool fcInFreeLists = fc->is_free();
7804 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
7805 assert((HeapWord*)fc <= _limit, "sweep invariant");
7806 if (CMSTestInFreeList && fcInFreeLists) {
7807 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7808 }
7809
7810 if (CMSTraceSweeper) {
7811 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", fc, chunkSize);
7812 }
7813
7814 HeapWord* const fc_addr = (HeapWord*) fc;
7815
7816 bool coalesce;
7817 const size_t left = pointer_delta(fc_addr, freeFinger());
7818 const size_t right = chunkSize;
7819 switch (FLSCoalescePolicy) {
7820 // numeric value forms a coalition aggressiveness metric
7821 case 0: { // never coalesce
7822 coalesce = false;
7823 break;
7824 }
7825 case 1: { // coalesce if left & right chunks on overpopulated lists
7826 coalesce = _sp->coalOverPopulated(left) &&
7827 _sp->coalOverPopulated(right);
7828 break;
7829 }
7830 case 2: { // coalesce if left chunk on overpopulated list (default)
7831 coalesce = _sp->coalOverPopulated(left);
7832 break;
7833 }
7834 case 3: { // coalesce if left OR right chunk on overpopulated list
7835 coalesce = _sp->coalOverPopulated(left) ||
7836 _sp->coalOverPopulated(right);
7837 break;
7838 }
7839 case 4: { // always coalesce
7840 coalesce = true;
7841 break;
7842 }
7843 default:
7844 ShouldNotReachHere();
7845 }
7846
7847 // Should the current free range be coalesced?
7848 // If the chunk is in a free range and either we decided to coalesce above
7849 // or the chunk is near the large block at the end of the heap
7850 // (isNearLargestChunk() returns true), then coalesce this chunk.
7851 const bool doCoalesce = inFreeRange()
7852 && (coalesce || _g->isNearLargestChunk(fc_addr));
7853 if (doCoalesce) {
7854 // Coalesce the current free range on the left with the new
7855 // chunk on the right. If either is on a free list,
7856 // it must be removed from the list and stashed in the closure.
7857 if (freeRangeInFreeLists()) {
7858 FreeChunk* const ffc = (FreeChunk*)freeFinger();
7859 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7860 "Size of free range is inconsistent with chunk size.");
7861 if (CMSTestInFreeList) {
7862 assert(_sp->verify_chunk_in_free_list(ffc),
7863 "Chunk is not in free lists");
7864 }
7865 _sp->coalDeath(ffc->size());
7866 _sp->removeFreeChunkFromFreeLists(ffc);
7867 set_freeRangeInFreeLists(false);
7868 }
7869 if (fcInFreeLists) {
7870 _sp->coalDeath(chunkSize);
7871 assert(fc->size() == chunkSize,
7872 "The chunk has the wrong size or is not in the free lists");
7873 _sp->removeFreeChunkFromFreeLists(fc);
7874 }
7875 set_lastFreeRangeCoalesced(true);
7876 print_free_block_coalesced(fc);
7877 } else { // not in a free range and/or should not coalesce
7878 // Return the current free range and start a new one.
7879 if (inFreeRange()) {
7880 // In a free range but cannot coalesce with the right hand chunk.
7881 // Put the current free range into the free lists.
7882 flush_cur_free_chunk(freeFinger(),
7883 pointer_delta(fc_addr, freeFinger()));
7884 }
7885 // Set up for new free range. Pass along whether the right hand
7886 // chunk is in the free lists.
7887 initialize_free_range((HeapWord*)fc, fcInFreeLists);
7888 }
7889 }
7890
7891 // Lookahead flush:
7892 // If we are tracking a free range, and this is the last chunk that
7893 // we'll look at because its end crosses past _limit, we'll preemptively
7894 // flush it along with any free range we may be holding on to. Note that
7895 // this can be the case only for an already free or freshly garbage
7896 // chunk. If this block is an object, it can never straddle
7897 // over _limit. The "straddling" occurs when _limit is set at
7898 // the previous end of the space when this cycle started, and
7899 // a subsequent heap expansion caused the previously co-terminal
7900 // free block to be coalesced with the newly expanded portion,
7901 // thus rendering _limit a non-block-boundary making it dangerous
7902 // for the sweeper to step over and examine.
7903 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7904 assert(inFreeRange(), "Should only be called if currently in a free range.");
7905 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7906 assert(_sp->used_region().contains(eob - 1),
7907 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7908 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7909 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7910 eob, eob-1, _limit, _sp->bottom(), _sp->end(), fc, chunk_size));
7911 if (eob >= _limit) {
7912 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7913 if (CMSTraceSweeper) {
7914 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
7915 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7916 "[" PTR_FORMAT "," PTR_FORMAT ")",
7917 _limit, fc, eob, _sp->bottom(), _sp->end());
7918 }
7919 // Return the storage we are tracking back into the free lists.
7920 if (CMSTraceSweeper) {
7921 gclog_or_tty->print_cr("Flushing ... ");
7922 }
7923 assert(freeFinger() < eob, "Error");
7924 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7925 }
7926 }
7927
7928 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7929 assert(inFreeRange(), "Should only be called if currently in a free range.");
7930 assert(size > 0,
7931 "A zero sized chunk cannot be added to the free lists.");
7932 if (!freeRangeInFreeLists()) {
7933 if (CMSTestInFreeList) {
7934 FreeChunk* fc = (FreeChunk*) chunk;
7935 fc->set_size(size);
7936 assert(!_sp->verify_chunk_in_free_list(fc),
7937 "chunk should not be in free lists yet");
7938 }
7939 if (CMSTraceSweeper) {
7940 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists",
7941 chunk, size);
7942 }
7943 // A new free range is going to be starting. The current
7944 // free range has not been added to the free lists yet or
7945 // was removed so add it back.
7946 // If the current free range was coalesced, then the death
7947 // of the free range was recorded. Record a birth now.
7948 if (lastFreeRangeCoalesced()) {
7949 _sp->coalBirth(size);
7950 }
7951 _sp->addChunkAndRepairOffsetTable(chunk, size,
7952 lastFreeRangeCoalesced());
7953 } else if (CMSTraceSweeper) {
7954 gclog_or_tty->print_cr("Already in free list: nothing to flush");
7955 }
7956 set_inFreeRange(false);
7957 set_freeRangeInFreeLists(false);
7958 }
7959
7960 // We take a break if we've been at this for a while,
7961 // so as to avoid monopolizing the locks involved.
7962 void SweepClosure::do_yield_work(HeapWord* addr) {
7963 // Return current free chunk being used for coalescing (if any)
7964 // to the appropriate freelist. After yielding, the next
7965 // free block encountered will start a coalescing range of
7966 // free blocks. If the next free block is adjacent to the
7967 // chunk just flushed, they will need to wait for the next
7968 // sweep to be coalesced.
7969 if (inFreeRange()) {
7970 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7971 }
7972
7973 // First give up the locks, then yield, then re-lock.
7974 // We should probably use a constructor/destructor idiom to
7975 // do this unlock/lock or modify the MutexUnlocker class to
7976 // serve our purpose. XXX
7977 assert_lock_strong(_bitMap->lock());
7978 assert_lock_strong(_freelistLock);
7979 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7980 "CMS thread should hold CMS token");
7981 _bitMap->lock()->unlock();
7982 _freelistLock->unlock();
7983 ConcurrentMarkSweepThread::desynchronize(true);
7984 _collector->stopTimer();
7985 if (PrintCMSStatistics != 0) {
7986 _collector->incrementYields();
7987 }
7988
7989 // See the comment in coordinator_yield()
7990 for (unsigned i = 0; i < CMSYieldSleepCount &&
7991 ConcurrentMarkSweepThread::should_yield() &&
7992 !CMSCollector::foregroundGCIsActive(); ++i) {
7993 os::sleep(Thread::current(), 1, false);
7994 }
7995
7996 ConcurrentMarkSweepThread::synchronize(true);
7997 _freelistLock->lock();
7998 _bitMap->lock()->lock_without_safepoint_check();
7999 _collector->startTimer();
8000 }
8001
8002 #ifndef PRODUCT
8003 // This is actually very useful in a product build if it can
8004 // be called from the debugger. Compile it into the product
8005 // as needed.
8006 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
8007 return debug_cms_space->verify_chunk_in_free_list(fc);
8008 }
8009 #endif
8010
8011 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
8012 if (CMSTraceSweeper) {
8013 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
8014 fc, fc->size());
8015 }
8016 }
8017
8018 // CMSIsAliveClosure
8019 bool CMSIsAliveClosure::do_object_b(oop obj) {
8020 HeapWord* addr = (HeapWord*)obj;
8021 return addr != NULL &&
8022 (!_span.contains(addr) || _bit_map->isMarked(addr));
8023 }
8024
8025
8026 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
8027 MemRegion span,
8028 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
8029 bool cpc):
8030 _collector(collector),
8031 _span(span),
8032 _bit_map(bit_map),
8033 _mark_stack(mark_stack),
8034 _concurrent_precleaning(cpc) {
8035 assert(!_span.is_empty(), "Empty span could spell trouble");
8036 }
8037
8038
8039 // CMSKeepAliveClosure: the serial version
8040 void CMSKeepAliveClosure::do_oop(oop obj) {
8041 HeapWord* addr = (HeapWord*)obj;
8042 if (_span.contains(addr) &&
8043 !_bit_map->isMarked(addr)) {
8044 _bit_map->mark(addr);
8045 bool simulate_overflow = false;
8046 NOT_PRODUCT(
8047 if (CMSMarkStackOverflowALot &&
8048 _collector->simulate_overflow()) {
8049 // simulate a stack overflow
8050 simulate_overflow = true;
8051 }
8052 )
8053 if (simulate_overflow || !_mark_stack->push(obj)) {
8054 if (_concurrent_precleaning) {
8055 // We dirty the overflown object and let the remark
8056 // phase deal with it.
8057 assert(_collector->overflow_list_is_empty(), "Error");
8058 // In the case of object arrays, we need to dirty all of
8059 // the cards that the object spans. No locking or atomics
8060 // are needed since no one else can be mutating the mod union
8061 // table.
8062 if (obj->is_objArray()) {
8063 size_t sz = obj->size();
8064 HeapWord* end_card_addr =
8065 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
8066 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8067 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8068 _collector->_modUnionTable.mark_range(redirty_range);
8069 } else {
8070 _collector->_modUnionTable.mark(addr);
8071 }
8072 _collector->_ser_kac_preclean_ovflw++;
8073 } else {
8074 _collector->push_on_overflow_list(obj);
8075 _collector->_ser_kac_ovflw++;
8076 }
8077 }
8078 }
8079 }
8080
8081 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8082 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8083
8084 // CMSParKeepAliveClosure: a parallel version of the above.
8085 // The work queues are private to each closure (thread),
8086 // but (may be) available for stealing by other threads.
8087 void CMSParKeepAliveClosure::do_oop(oop obj) {
8088 HeapWord* addr = (HeapWord*)obj;
8089 if (_span.contains(addr) &&
8090 !_bit_map->isMarked(addr)) {
8091 // In general, during recursive tracing, several threads
8092 // may be concurrently getting here; the first one to
8093 // "tag" it, claims it.
8094 if (_bit_map->par_mark(addr)) {
8095 bool res = _work_queue->push(obj);
8096 assert(res, "Low water mark should be much less than capacity");
8097 // Do a recursive trim in the hope that this will keep
8098 // stack usage lower, but leave some oops for potential stealers
8099 trim_queue(_low_water_mark);
8100 } // Else, another thread got there first
8101 }
8102 }
8103
8104 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8105 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8106
8107 void CMSParKeepAliveClosure::trim_queue(uint max) {
8108 while (_work_queue->size() > max) {
8109 oop new_oop;
8110 if (_work_queue->pop_local(new_oop)) {
8111 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8112 assert(_bit_map->isMarked((HeapWord*)new_oop),
8113 "no white objects on this stack!");
8114 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8115 // iterate over the oops in this oop, marking and pushing
8116 // the ones in CMS heap (i.e. in _span).
8117 new_oop->oop_iterate(&_mark_and_push);
8118 }
8119 }
8120 }
8121
8122 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8123 CMSCollector* collector,
8124 MemRegion span, CMSBitMap* bit_map,
8125 OopTaskQueue* work_queue):
8126 _collector(collector),
8127 _span(span),
8128 _bit_map(bit_map),
8129 _work_queue(work_queue) { }
8130
8131 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8132 HeapWord* addr = (HeapWord*)obj;
8133 if (_span.contains(addr) &&
8134 !_bit_map->isMarked(addr)) {
8135 if (_bit_map->par_mark(addr)) {
8136 bool simulate_overflow = false;
8137 NOT_PRODUCT(
8138 if (CMSMarkStackOverflowALot &&
8139 _collector->par_simulate_overflow()) {
8140 // simulate a stack overflow
8141 simulate_overflow = true;
8142 }
8143 )
8144 if (simulate_overflow || !_work_queue->push(obj)) {
8145 _collector->par_push_on_overflow_list(obj);
8146 _collector->_par_kac_ovflw++;
8147 }
8148 } // Else another thread got there already
8149 }
8150 }
8151
8152 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8153 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8154
8155 //////////////////////////////////////////////////////////////////
8156 // CMSExpansionCause /////////////////////////////
8157 //////////////////////////////////////////////////////////////////
8158 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8159 switch (cause) {
8160 case _no_expansion:
8161 return "No expansion";
8162 case _satisfy_free_ratio:
8163 return "Free ratio";
8164 case _satisfy_promotion:
8165 return "Satisfy promotion";
8166 case _satisfy_allocation:
8167 return "allocation";
8168 case _allocate_par_lab:
8169 return "Par LAB";
8170 case _allocate_par_spooling_space:
8171 return "Par Spooling Space";
8172 case _adaptive_size_policy:
8173 return "Ergonomics";
8174 default:
8175 return "unknown";
8176 }
8177 }
8178
8179 void CMSDrainMarkingStackClosure::do_void() {
8180 // the max number to take from overflow list at a time
8181 const size_t num = _mark_stack->capacity()/4;
8182 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8183 "Overflow list should be NULL during concurrent phases");
8184 while (!_mark_stack->isEmpty() ||
8185 // if stack is empty, check the overflow list
8186 _collector->take_from_overflow_list(num, _mark_stack)) {
8187 oop obj = _mark_stack->pop();
8188 HeapWord* addr = (HeapWord*)obj;
8189 assert(_span.contains(addr), "Should be within span");
8190 assert(_bit_map->isMarked(addr), "Should be marked");
8191 assert(obj->is_oop(), "Should be an oop");
8192 obj->oop_iterate(_keep_alive);
8193 }
8194 }
8195
8196 void CMSParDrainMarkingStackClosure::do_void() {
8197 // drain queue
8198 trim_queue(0);
8199 }
8200
8201 // Trim our work_queue so its length is below max at return
8202 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8203 while (_work_queue->size() > max) {
8204 oop new_oop;
8205 if (_work_queue->pop_local(new_oop)) {
8206 assert(new_oop->is_oop(), "Expected an oop");
8207 assert(_bit_map->isMarked((HeapWord*)new_oop),
8208 "no white objects on this stack!");
8209 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8210 // iterate over the oops in this oop, marking and pushing
8211 // the ones in CMS heap (i.e. in _span).
8212 new_oop->oop_iterate(&_mark_and_push);
8213 }
8214 }
8215 }
8216
8217 ////////////////////////////////////////////////////////////////////
8218 // Support for Marking Stack Overflow list handling and related code
8219 ////////////////////////////////////////////////////////////////////
8220 // Much of the following code is similar in shape and spirit to the
8221 // code used in ParNewGC. We should try and share that code
8222 // as much as possible in the future.
8223
8224 #ifndef PRODUCT
8225 // Debugging support for CMSStackOverflowALot
8226
8227 // It's OK to call this multi-threaded; the worst thing
8228 // that can happen is that we'll get a bunch of closely
8229 // spaced simulated overflows, but that's OK, in fact
8230 // probably good as it would exercise the overflow code
8231 // under contention.
8232 bool CMSCollector::simulate_overflow() {
8233 if (_overflow_counter-- <= 0) { // just being defensive
8234 _overflow_counter = CMSMarkStackOverflowInterval;
8235 return true;
8236 } else {
8237 return false;
8238 }
8239 }
8240
8241 bool CMSCollector::par_simulate_overflow() {
8242 return simulate_overflow();
8243 }
8244 #endif
8245
8246 // Single-threaded
8247 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8248 assert(stack->isEmpty(), "Expected precondition");
8249 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8250 size_t i = num;
8251 oop cur = _overflow_list;
8252 const markOop proto = markOopDesc::prototype();
8253 NOT_PRODUCT(ssize_t n = 0;)
8254 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8255 next = oop(cur->mark());
8256 cur->set_mark(proto); // until proven otherwise
8257 assert(cur->is_oop(), "Should be an oop");
8258 bool res = stack->push(cur);
8259 assert(res, "Bit off more than can chew?");
8260 NOT_PRODUCT(n++;)
8261 }
8262 _overflow_list = cur;
8263 #ifndef PRODUCT
8264 assert(_num_par_pushes >= n, "Too many pops?");
8265 _num_par_pushes -=n;
8266 #endif
8267 return !stack->isEmpty();
8268 }
8269
8270 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
8271 // (MT-safe) Get a prefix of at most "num" from the list.
8272 // The overflow list is chained through the mark word of
8273 // each object in the list. We fetch the entire list,
8274 // break off a prefix of the right size and return the
8275 // remainder. If other threads try to take objects from
8276 // the overflow list at that time, they will wait for
8277 // some time to see if data becomes available. If (and
8278 // only if) another thread places one or more object(s)
8279 // on the global list before we have returned the suffix
8280 // to the global list, we will walk down our local list
8281 // to find its end and append the global list to
8282 // our suffix before returning it. This suffix walk can
8283 // prove to be expensive (quadratic in the amount of traffic)
8284 // when there are many objects in the overflow list and
8285 // there is much producer-consumer contention on the list.
8286 // *NOTE*: The overflow list manipulation code here and
8287 // in ParNewGeneration:: are very similar in shape,
8288 // except that in the ParNew case we use the old (from/eden)
8289 // copy of the object to thread the list via its klass word.
8290 // Because of the common code, if you make any changes in
8291 // the code below, please check the ParNew version to see if
8292 // similar changes might be needed.
8293 // CR 6797058 has been filed to consolidate the common code.
8294 bool CMSCollector::par_take_from_overflow_list(size_t num,
8295 OopTaskQueue* work_q,
8296 int no_of_gc_threads) {
8297 assert(work_q->size() == 0, "First empty local work queue");
8298 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8299 if (_overflow_list == NULL) {
8300 return false;
8301 }
8302 // Grab the entire list; we'll put back a suffix
8303 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8304 Thread* tid = Thread::current();
8305 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
8306 // set to ParallelGCThreads.
8307 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
8308 size_t sleep_time_millis = MAX2((size_t)1, num/100);
8309 // If the list is busy, we spin for a short while,
8310 // sleeping between attempts to get the list.
8311 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
8312 os::sleep(tid, sleep_time_millis, false);
8313 if (_overflow_list == NULL) {
8314 // Nothing left to take
8315 return false;
8316 } else if (_overflow_list != BUSY) {
8317 // Try and grab the prefix
8318 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8319 }
8320 }
8321 // If the list was found to be empty, or we spun long
8322 // enough, we give up and return empty-handed. If we leave
8323 // the list in the BUSY state below, it must be the case that
8324 // some other thread holds the overflow list and will set it
8325 // to a non-BUSY state in the future.
8326 if (prefix == NULL || prefix == BUSY) {
8327 // Nothing to take or waited long enough
8328 if (prefix == NULL) {
8329 // Write back the NULL in case we overwrote it with BUSY above
8330 // and it is still the same value.
8331 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8332 }
8333 return false;
8334 }
8335 assert(prefix != NULL && prefix != BUSY, "Error");
8336 size_t i = num;
8337 oop cur = prefix;
8338 // Walk down the first "num" objects, unless we reach the end.
8339 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8340 if (cur->mark() == NULL) {
8341 // We have "num" or fewer elements in the list, so there
8342 // is nothing to return to the global list.
8343 // Write back the NULL in lieu of the BUSY we wrote
8344 // above, if it is still the same value.
8345 if (_overflow_list == BUSY) {
8346 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8347 }
8348 } else {
8349 // Chop off the suffix and return it to the global list.
8350 assert(cur->mark() != BUSY, "Error");
8351 oop suffix_head = cur->mark(); // suffix will be put back on global list
8352 cur->set_mark(NULL); // break off suffix
8353 // It's possible that the list is still in the empty(busy) state
8354 // we left it in a short while ago; in that case we may be
8355 // able to place back the suffix without incurring the cost
8356 // of a walk down the list.
8357 oop observed_overflow_list = _overflow_list;
8358 oop cur_overflow_list = observed_overflow_list;
8359 bool attached = false;
8360 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
8361 observed_overflow_list =
8362 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8363 if (cur_overflow_list == observed_overflow_list) {
8364 attached = true;
8365 break;
8366 } else cur_overflow_list = observed_overflow_list;
8367 }
8368 if (!attached) {
8369 // Too bad, someone else sneaked in (at least) an element; we'll need
8370 // to do a splice. Find tail of suffix so we can prepend suffix to global
8371 // list.
8372 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8373 oop suffix_tail = cur;
8374 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8375 "Tautology");
8376 observed_overflow_list = _overflow_list;
8377 do {
8378 cur_overflow_list = observed_overflow_list;
8379 if (cur_overflow_list != BUSY) {
8380 // Do the splice ...
8381 suffix_tail->set_mark(markOop(cur_overflow_list));
8382 } else { // cur_overflow_list == BUSY
8383 suffix_tail->set_mark(NULL);
8384 }
8385 // ... and try to place spliced list back on overflow_list ...
8386 observed_overflow_list =
8387 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8388 } while (cur_overflow_list != observed_overflow_list);
8389 // ... until we have succeeded in doing so.
8390 }
8391 }
8392
8393 // Push the prefix elements on work_q
8394 assert(prefix != NULL, "control point invariant");
8395 const markOop proto = markOopDesc::prototype();
8396 oop next;
8397 NOT_PRODUCT(ssize_t n = 0;)
8398 for (cur = prefix; cur != NULL; cur = next) {
8399 next = oop(cur->mark());
8400 cur->set_mark(proto); // until proven otherwise
8401 assert(cur->is_oop(), "Should be an oop");
8402 bool res = work_q->push(cur);
8403 assert(res, "Bit off more than we can chew?");
8404 NOT_PRODUCT(n++;)
8405 }
8406 #ifndef PRODUCT
8407 assert(_num_par_pushes >= n, "Too many pops?");
8408 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8409 #endif
8410 return true;
8411 }
8412
8413 // Single-threaded
8414 void CMSCollector::push_on_overflow_list(oop p) {
8415 NOT_PRODUCT(_num_par_pushes++;)
8416 assert(p->is_oop(), "Not an oop");
8417 preserve_mark_if_necessary(p);
8418 p->set_mark((markOop)_overflow_list);
8419 _overflow_list = p;
8420 }
8421
8422 // Multi-threaded; use CAS to prepend to overflow list
8423 void CMSCollector::par_push_on_overflow_list(oop p) {
8424 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8425 assert(p->is_oop(), "Not an oop");
8426 par_preserve_mark_if_necessary(p);
8427 oop observed_overflow_list = _overflow_list;
8428 oop cur_overflow_list;
8429 do {
8430 cur_overflow_list = observed_overflow_list;
8431 if (cur_overflow_list != BUSY) {
8432 p->set_mark(markOop(cur_overflow_list));
8433 } else {
8434 p->set_mark(NULL);
8435 }
8436 observed_overflow_list =
8437 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8438 } while (cur_overflow_list != observed_overflow_list);
8439 }
8440 #undef BUSY
8441
8442 // Single threaded
8443 // General Note on GrowableArray: pushes may silently fail
8444 // because we are (temporarily) out of C-heap for expanding
8445 // the stack. The problem is quite ubiquitous and affects
8446 // a lot of code in the JVM. The prudent thing for GrowableArray
8447 // to do (for now) is to exit with an error. However, that may
8448 // be too draconian in some cases because the caller may be
8449 // able to recover without much harm. For such cases, we
8450 // should probably introduce a "soft_push" method which returns
8451 // an indication of success or failure with the assumption that
8452 // the caller may be able to recover from a failure; code in
8453 // the VM can then be changed, incrementally, to deal with such
8454 // failures where possible, thus, incrementally hardening the VM
8455 // in such low resource situations.
8456 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8457 _preserved_oop_stack.push(p);
8458 _preserved_mark_stack.push(m);
8459 assert(m == p->mark(), "Mark word changed");
8460 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8461 "bijection");
8462 }
8463
8464 // Single threaded
8465 void CMSCollector::preserve_mark_if_necessary(oop p) {
8466 markOop m = p->mark();
8467 if (m->must_be_preserved(p)) {
8468 preserve_mark_work(p, m);
8469 }
8470 }
8471
8472 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8473 markOop m = p->mark();
8474 if (m->must_be_preserved(p)) {
8475 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8476 // Even though we read the mark word without holding
8477 // the lock, we are assured that it will not change
8478 // because we "own" this oop, so no other thread can
8479 // be trying to push it on the overflow list; see
8480 // the assertion in preserve_mark_work() that checks
8481 // that m == p->mark().
8482 preserve_mark_work(p, m);
8483 }
8484 }
8485
8486 // We should be able to do this multi-threaded,
8487 // a chunk of stack being a task (this is
8488 // correct because each oop only ever appears
8489 // once in the overflow list. However, it's
8490 // not very easy to completely overlap this with
8491 // other operations, so will generally not be done
8492 // until all work's been completed. Because we
8493 // expect the preserved oop stack (set) to be small,
8494 // it's probably fine to do this single-threaded.
8495 // We can explore cleverer concurrent/overlapped/parallel
8496 // processing of preserved marks if we feel the
8497 // need for this in the future. Stack overflow should
8498 // be so rare in practice and, when it happens, its
8499 // effect on performance so great that this will
8500 // likely just be in the noise anyway.
8501 void CMSCollector::restore_preserved_marks_if_any() {
8502 assert(SafepointSynchronize::is_at_safepoint(),
8503 "world should be stopped");
8504 assert(Thread::current()->is_ConcurrentGC_thread() ||
8505 Thread::current()->is_VM_thread(),
8506 "should be single-threaded");
8507 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8508 "bijection");
8509
8510 while (!_preserved_oop_stack.is_empty()) {
8511 oop p = _preserved_oop_stack.pop();
8512 assert(p->is_oop(), "Should be an oop");
8513 assert(_span.contains(p), "oop should be in _span");
8514 assert(p->mark() == markOopDesc::prototype(),
8515 "Set when taken from overflow list");
8516 markOop m = _preserved_mark_stack.pop();
8517 p->set_mark(m);
8518 }
8519 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8520 "stacks were cleared above");
8521 }
8522
8523 #ifndef PRODUCT
8524 bool CMSCollector::no_preserved_marks() const {
8525 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8526 }
8527 #endif
8528
8529 // Transfer some number of overflown objects to usual marking
8530 // stack. Return true if some objects were transferred.
8531 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8532 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8533 (size_t)ParGCDesiredObjsFromOverflowList);
8534
8535 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8536 assert(_collector->overflow_list_is_empty() || res,
8537 "If list is not empty, we should have taken something");
8538 assert(!res || !_mark_stack->isEmpty(),
8539 "If we took something, it should now be on our stack");
8540 return res;
8541 }
8542
8543 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8544 size_t res = _sp->block_size_no_stall(addr, _collector);
8545 if (_sp->block_is_obj(addr)) {
8546 if (_live_bit_map->isMarked(addr)) {
8547 // It can't have been dead in a previous cycle
8548 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8549 } else {
8550 _dead_bit_map->mark(addr); // mark the dead object
8551 }
8552 }
8553 // Could be 0, if the block size could not be computed without stalling.
8554 return res;
8555 }
8556
8557 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8558
8559 switch (phase) {
8560 case CMSCollector::InitialMarking:
8561 initialize(true /* fullGC */ ,
8562 cause /* cause of the GC */,
8563 true /* recordGCBeginTime */,
8564 true /* recordPreGCUsage */,
8565 false /* recordPeakUsage */,
8566 false /* recordPostGCusage */,
8567 true /* recordAccumulatedGCTime */,
8568 false /* recordGCEndTime */,
8569 false /* countCollection */ );
8570 break;
8571
8572 case CMSCollector::FinalMarking:
8573 initialize(true /* fullGC */ ,
8574 cause /* cause of the GC */,
8575 false /* recordGCBeginTime */,
8576 false /* recordPreGCUsage */,
8577 false /* recordPeakUsage */,
8578 false /* recordPostGCusage */,
8579 true /* recordAccumulatedGCTime */,
8580 false /* recordGCEndTime */,
8581 false /* countCollection */ );
8582 break;
8583
8584 case CMSCollector::Sweeping:
8585 initialize(true /* fullGC */ ,
8586 cause /* cause of the GC */,
8587 false /* recordGCBeginTime */,
8588 false /* recordPreGCUsage */,
8589 true /* recordPeakUsage */,
8590 true /* recordPostGCusage */,
8591 false /* recordAccumulatedGCTime */,
8592 true /* recordGCEndTime */,
8593 true /* countCollection */ );
8594 break;
8595
8596 default:
8597 ShouldNotReachHere();
8598 }
8599 }