1 /*
2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/classLoaderData.hpp"
27 #include "classfile/stringTable.hpp"
28 #include "classfile/symbolTable.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/codeCache.hpp"
31 #include "gc/cms/cmsCollectorPolicy.hpp"
32 #include "gc/cms/cmsHeap.hpp"
33 #include "gc/cms/cmsOopClosures.inline.hpp"
34 #include "gc/cms/compactibleFreeListSpace.hpp"
35 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp"
36 #include "gc/cms/concurrentMarkSweepThread.hpp"
37 #include "gc/cms/parNewGeneration.hpp"
38 #include "gc/cms/vmCMSOperations.hpp"
39 #include "gc/serial/genMarkSweep.hpp"
40 #include "gc/serial/tenuredGeneration.hpp"
41 #include "gc/shared/adaptiveSizePolicy.hpp"
42 #include "gc/shared/cardGeneration.inline.hpp"
43 #include "gc/shared/cardTableRS.hpp"
44 #include "gc/shared/collectedHeap.inline.hpp"
45 #include "gc/shared/collectorCounters.hpp"
46 #include "gc/shared/collectorPolicy.hpp"
47 #include "gc/shared/gcLocker.inline.hpp"
48 #include "gc/shared/gcPolicyCounters.hpp"
49 #include "gc/shared/gcTimer.hpp"
50 #include "gc/shared/gcTrace.hpp"
51 #include "gc/shared/gcTraceTime.inline.hpp"
52 #include "gc/shared/genCollectedHeap.hpp"
53 #include "gc/shared/genOopClosures.inline.hpp"
54 #include "gc/shared/isGCActiveMark.hpp"
55 #include "gc/shared/referencePolicy.hpp"
56 #include "gc/shared/strongRootsScope.hpp"
57 #include "gc/shared/taskqueue.inline.hpp"
58 #include "gc/shared/weakProcessor.hpp"
59 #include "logging/log.hpp"
60 #include "logging/logStream.hpp"
61 #include "memory/allocation.hpp"
62 #include "memory/iterator.inline.hpp"
63 #include "memory/padded.hpp"
64 #include "memory/resourceArea.hpp"
65 #include "oops/oop.inline.hpp"
66 #include "prims/jvmtiExport.hpp"
67 #include "runtime/atomic.hpp"
68 #include "runtime/globals_extension.hpp"
69 #include "runtime/handles.inline.hpp"
70 #include "runtime/java.hpp"
71 #include "runtime/orderAccess.inline.hpp"
72 #include "runtime/timer.hpp"
73 #include "runtime/vmThread.hpp"
74 #include "services/memoryService.hpp"
75 #include "services/runtimeService.hpp"
76 #include "utilities/align.hpp"
77 #include "utilities/stack.inline.hpp"
78
79 // statics
80 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
81 bool CMSCollector::_full_gc_requested = false;
82 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
83
84 //////////////////////////////////////////////////////////////////
85 // In support of CMS/VM thread synchronization
86 //////////////////////////////////////////////////////////////////
87 // We split use of the CGC_lock into 2 "levels".
88 // The low-level locking is of the usual CGC_lock monitor. We introduce
89 // a higher level "token" (hereafter "CMS token") built on top of the
90 // low level monitor (hereafter "CGC lock").
91 // The token-passing protocol gives priority to the VM thread. The
92 // CMS-lock doesn't provide any fairness guarantees, but clients
93 // should ensure that it is only held for very short, bounded
94 // durations.
95 //
96 // When either of the CMS thread or the VM thread is involved in
97 // collection operations during which it does not want the other
98 // thread to interfere, it obtains the CMS token.
99 //
100 // If either thread tries to get the token while the other has
101 // it, that thread waits. However, if the VM thread and CMS thread
102 // both want the token, then the VM thread gets priority while the
103 // CMS thread waits. This ensures, for instance, that the "concurrent"
104 // phases of the CMS thread's work do not block out the VM thread
105 // for long periods of time as the CMS thread continues to hog
106 // the token. (See bug 4616232).
107 //
108 // The baton-passing functions are, however, controlled by the
109 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
110 // and here the low-level CMS lock, not the high level token,
111 // ensures mutual exclusion.
112 //
113 // Two important conditions that we have to satisfy:
114 // 1. if a thread does a low-level wait on the CMS lock, then it
115 // relinquishes the CMS token if it were holding that token
116 // when it acquired the low-level CMS lock.
117 // 2. any low-level notifications on the low-level lock
118 // should only be sent when a thread has relinquished the token.
119 //
120 // In the absence of either property, we'd have potential deadlock.
121 //
122 // We protect each of the CMS (concurrent and sequential) phases
123 // with the CMS _token_, not the CMS _lock_.
124 //
125 // The only code protected by CMS lock is the token acquisition code
126 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
127 // baton-passing code.
128 //
129 // Unfortunately, i couldn't come up with a good abstraction to factor and
130 // hide the naked CGC_lock manipulation in the baton-passing code
131 // further below. That's something we should try to do. Also, the proof
132 // of correctness of this 2-level locking scheme is far from obvious,
133 // and potentially quite slippery. We have an uneasy suspicion, for instance,
134 // that there may be a theoretical possibility of delay/starvation in the
135 // low-level lock/wait/notify scheme used for the baton-passing because of
136 // potential interference with the priority scheme embodied in the
137 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
138 // invocation further below and marked with "XXX 20011219YSR".
139 // Indeed, as we note elsewhere, this may become yet more slippery
140 // in the presence of multiple CMS and/or multiple VM threads. XXX
141
142 class CMSTokenSync: public StackObj {
143 private:
144 bool _is_cms_thread;
145 public:
146 CMSTokenSync(bool is_cms_thread):
147 _is_cms_thread(is_cms_thread) {
148 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
149 "Incorrect argument to constructor");
150 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
151 }
152
153 ~CMSTokenSync() {
154 assert(_is_cms_thread ?
155 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
156 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
157 "Incorrect state");
158 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
159 }
160 };
161
162 // Convenience class that does a CMSTokenSync, and then acquires
163 // upto three locks.
164 class CMSTokenSyncWithLocks: public CMSTokenSync {
165 private:
166 // Note: locks are acquired in textual declaration order
167 // and released in the opposite order
168 MutexLockerEx _locker1, _locker2, _locker3;
169 public:
170 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
171 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
172 CMSTokenSync(is_cms_thread),
173 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
174 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
175 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
176 { }
177 };
178
179
180 //////////////////////////////////////////////////////////////////
181 // Concurrent Mark-Sweep Generation /////////////////////////////
182 //////////////////////////////////////////////////////////////////
183
184 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
185
186 // This struct contains per-thread things necessary to support parallel
187 // young-gen collection.
188 class CMSParGCThreadState: public CHeapObj<mtGC> {
189 public:
190 CompactibleFreeListSpaceLAB lab;
191 PromotionInfo promo;
192
193 // Constructor.
194 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
195 promo.setSpace(cfls);
196 }
197 };
198
199 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
200 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) :
201 CardGeneration(rs, initial_byte_size, ct),
202 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
203 _did_compact(false)
204 {
205 HeapWord* bottom = (HeapWord*) _virtual_space.low();
206 HeapWord* end = (HeapWord*) _virtual_space.high();
207
208 _direct_allocated_words = 0;
209 NOT_PRODUCT(
210 _numObjectsPromoted = 0;
211 _numWordsPromoted = 0;
212 _numObjectsAllocated = 0;
213 _numWordsAllocated = 0;
214 )
215
216 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end));
217 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
218 _cmsSpace->_old_gen = this;
219
220 _gc_stats = new CMSGCStats();
221
222 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
223 // offsets match. The ability to tell free chunks from objects
224 // depends on this property.
225 debug_only(
226 FreeChunk* junk = NULL;
227 assert(UseCompressedClassPointers ||
228 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
229 "Offset of FreeChunk::_prev within FreeChunk must match"
230 " that of OopDesc::_klass within OopDesc");
231 )
232
233 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC);
234 for (uint i = 0; i < ParallelGCThreads; i++) {
235 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
236 }
237
238 _incremental_collection_failed = false;
239 // The "dilatation_factor" is the expansion that can occur on
240 // account of the fact that the minimum object size in the CMS
241 // generation may be larger than that in, say, a contiguous young
242 // generation.
243 // Ideally, in the calculation below, we'd compute the dilatation
244 // factor as: MinChunkSize/(promoting_gen's min object size)
245 // Since we do not have such a general query interface for the
246 // promoting generation, we'll instead just use the minimum
247 // object size (which today is a header's worth of space);
248 // note that all arithmetic is in units of HeapWords.
249 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
250 assert(_dilatation_factor >= 1.0, "from previous assert");
251 }
252
253
254 // The field "_initiating_occupancy" represents the occupancy percentage
255 // at which we trigger a new collection cycle. Unless explicitly specified
256 // via CMSInitiatingOccupancyFraction (argument "io" below), it
257 // is calculated by:
258 //
259 // Let "f" be MinHeapFreeRatio in
260 //
261 // _initiating_occupancy = 100-f +
262 // f * (CMSTriggerRatio/100)
263 // where CMSTriggerRatio is the argument "tr" below.
264 //
265 // That is, if we assume the heap is at its desired maximum occupancy at the
266 // end of a collection, we let CMSTriggerRatio of the (purported) free
267 // space be allocated before initiating a new collection cycle.
268 //
269 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
270 assert(io <= 100 && tr <= 100, "Check the arguments");
271 if (io >= 0) {
272 _initiating_occupancy = (double)io / 100.0;
273 } else {
274 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
275 (double)(tr * MinHeapFreeRatio) / 100.0)
276 / 100.0;
277 }
278 }
279
280 void ConcurrentMarkSweepGeneration::ref_processor_init() {
281 assert(collector() != NULL, "no collector");
282 collector()->ref_processor_init();
283 }
284
285 void CMSCollector::ref_processor_init() {
286 if (_ref_processor == NULL) {
287 // Allocate and initialize a reference processor
288 _ref_processor =
289 new ReferenceProcessor(_span, // span
290 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
291 ParallelGCThreads, // mt processing degree
292 _cmsGen->refs_discovery_is_mt(), // mt discovery
293 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
294 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
295 &_is_alive_closure); // closure for liveness info
296 // Initialize the _ref_processor field of CMSGen
297 _cmsGen->set_ref_processor(_ref_processor);
298
299 }
300 }
301
302 AdaptiveSizePolicy* CMSCollector::size_policy() {
303 CMSHeap* heap = CMSHeap::heap();
304 return heap->gen_policy()->size_policy();
305 }
306
307 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
308
309 const char* gen_name = "old";
310 GenCollectorPolicy* gcp = CMSHeap::heap()->gen_policy();
311 // Generation Counters - generation 1, 1 subspace
312 _gen_counters = new GenerationCounters(gen_name, 1, 1,
313 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
314
315 _space_counters = new GSpaceCounters(gen_name, 0,
316 _virtual_space.reserved_size(),
317 this, _gen_counters);
318 }
319
320 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
321 _cms_gen(cms_gen)
322 {
323 assert(alpha <= 100, "bad value");
324 _saved_alpha = alpha;
325
326 // Initialize the alphas to the bootstrap value of 100.
327 _gc0_alpha = _cms_alpha = 100;
328
329 _cms_begin_time.update();
330 _cms_end_time.update();
331
332 _gc0_duration = 0.0;
333 _gc0_period = 0.0;
334 _gc0_promoted = 0;
335
336 _cms_duration = 0.0;
337 _cms_period = 0.0;
338 _cms_allocated = 0;
339
340 _cms_used_at_gc0_begin = 0;
341 _cms_used_at_gc0_end = 0;
342 _allow_duty_cycle_reduction = false;
343 _valid_bits = 0;
344 }
345
346 double CMSStats::cms_free_adjustment_factor(size_t free) const {
347 // TBD: CR 6909490
348 return 1.0;
349 }
350
351 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
352 }
353
354 // If promotion failure handling is on use
355 // the padded average size of the promotion for each
356 // young generation collection.
357 double CMSStats::time_until_cms_gen_full() const {
358 size_t cms_free = _cms_gen->cmsSpace()->free();
359 CMSHeap* heap = CMSHeap::heap();
360 size_t expected_promotion = MIN2(heap->young_gen()->capacity(),
361 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
362 if (cms_free > expected_promotion) {
363 // Start a cms collection if there isn't enough space to promote
364 // for the next young collection. Use the padded average as
365 // a safety factor.
366 cms_free -= expected_promotion;
367
368 // Adjust by the safety factor.
369 double cms_free_dbl = (double)cms_free;
370 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0;
371 // Apply a further correction factor which tries to adjust
372 // for recent occurance of concurrent mode failures.
373 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
374 cms_free_dbl = cms_free_dbl * cms_adjustment;
375
376 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
377 cms_free, expected_promotion);
378 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0);
379 // Add 1 in case the consumption rate goes to zero.
380 return cms_free_dbl / (cms_consumption_rate() + 1.0);
381 }
382 return 0.0;
383 }
384
385 // Compare the duration of the cms collection to the
386 // time remaining before the cms generation is empty.
387 // Note that the time from the start of the cms collection
388 // to the start of the cms sweep (less than the total
389 // duration of the cms collection) can be used. This
390 // has been tried and some applications experienced
391 // promotion failures early in execution. This was
392 // possibly because the averages were not accurate
393 // enough at the beginning.
394 double CMSStats::time_until_cms_start() const {
395 // We add "gc0_period" to the "work" calculation
396 // below because this query is done (mostly) at the
397 // end of a scavenge, so we need to conservatively
398 // account for that much possible delay
399 // in the query so as to avoid concurrent mode failures
400 // due to starting the collection just a wee bit too
401 // late.
402 double work = cms_duration() + gc0_period();
403 double deadline = time_until_cms_gen_full();
404 // If a concurrent mode failure occurred recently, we want to be
405 // more conservative and halve our expected time_until_cms_gen_full()
406 if (work > deadline) {
407 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ",
408 cms_duration(), gc0_period(), time_until_cms_gen_full());
409 return 0.0;
410 }
411 return work - deadline;
412 }
413
414 #ifndef PRODUCT
415 void CMSStats::print_on(outputStream *st) const {
416 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
417 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
418 gc0_duration(), gc0_period(), gc0_promoted());
419 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
420 cms_duration(), cms_period(), cms_allocated());
421 st->print(",cms_since_beg=%g,cms_since_end=%g",
422 cms_time_since_begin(), cms_time_since_end());
423 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
424 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
425
426 if (valid()) {
427 st->print(",promo_rate=%g,cms_alloc_rate=%g",
428 promotion_rate(), cms_allocation_rate());
429 st->print(",cms_consumption_rate=%g,time_until_full=%g",
430 cms_consumption_rate(), time_until_cms_gen_full());
431 }
432 st->cr();
433 }
434 #endif // #ifndef PRODUCT
435
436 CMSCollector::CollectorState CMSCollector::_collectorState =
437 CMSCollector::Idling;
438 bool CMSCollector::_foregroundGCIsActive = false;
439 bool CMSCollector::_foregroundGCShouldWait = false;
440
441 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
442 CardTableRS* ct,
443 ConcurrentMarkSweepPolicy* cp):
444 _cmsGen(cmsGen),
445 _ct(ct),
446 _ref_processor(NULL), // will be set later
447 _conc_workers(NULL), // may be set later
448 _abort_preclean(false),
449 _start_sampling(false),
450 _between_prologue_and_epilogue(false),
451 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
452 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
453 -1 /* lock-free */, "No_lock" /* dummy */),
454 _modUnionClosurePar(&_modUnionTable),
455 // Adjust my span to cover old (cms) gen
456 _span(cmsGen->reserved()),
457 // Construct the is_alive_closure with _span & markBitMap
458 _is_alive_closure(_span, &_markBitMap),
459 _restart_addr(NULL),
460 _overflow_list(NULL),
461 _stats(cmsGen),
462 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
463 //verify that this lock should be acquired with safepoint check.
464 Monitor::_safepoint_check_sometimes)),
465 _eden_chunk_array(NULL), // may be set in ctor body
466 _eden_chunk_capacity(0), // -- ditto --
467 _eden_chunk_index(0), // -- ditto --
468 _survivor_plab_array(NULL), // -- ditto --
469 _survivor_chunk_array(NULL), // -- ditto --
470 _survivor_chunk_capacity(0), // -- ditto --
471 _survivor_chunk_index(0), // -- ditto --
472 _ser_pmc_preclean_ovflw(0),
473 _ser_kac_preclean_ovflw(0),
474 _ser_pmc_remark_ovflw(0),
475 _par_pmc_remark_ovflw(0),
476 _ser_kac_ovflw(0),
477 _par_kac_ovflw(0),
478 #ifndef PRODUCT
479 _num_par_pushes(0),
480 #endif
481 _collection_count_start(0),
482 _verifying(false),
483 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
484 _completed_initialization(false),
485 _collector_policy(cp),
486 _should_unload_classes(CMSClassUnloadingEnabled),
487 _concurrent_cycles_since_last_unload(0),
488 _roots_scanning_options(GenCollectedHeap::SO_None),
489 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
490 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
491 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
492 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
493 _cms_start_registered(false)
494 {
495 // Now expand the span and allocate the collection support structures
496 // (MUT, marking bit map etc.) to cover both generations subject to
497 // collection.
498
499 // For use by dirty card to oop closures.
500 _cmsGen->cmsSpace()->set_collector(this);
501
502 // Allocate MUT and marking bit map
503 {
504 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
505 if (!_markBitMap.allocate(_span)) {
506 log_warning(gc)("Failed to allocate CMS Bit Map");
507 return;
508 }
509 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
510 }
511 {
512 _modUnionTable.allocate(_span);
513 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
514 }
515
516 if (!_markStack.allocate(MarkStackSize)) {
517 log_warning(gc)("Failed to allocate CMS Marking Stack");
518 return;
519 }
520
521 // Support for multi-threaded concurrent phases
522 if (CMSConcurrentMTEnabled) {
523 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
524 // just for now
525 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4);
526 }
527 if (ConcGCThreads > 1) {
528 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
529 ConcGCThreads, true);
530 if (_conc_workers == NULL) {
531 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled");
532 CMSConcurrentMTEnabled = false;
533 } else {
534 _conc_workers->initialize_workers();
535 }
536 } else {
537 CMSConcurrentMTEnabled = false;
538 }
539 }
540 if (!CMSConcurrentMTEnabled) {
541 ConcGCThreads = 0;
542 } else {
543 // Turn off CMSCleanOnEnter optimization temporarily for
544 // the MT case where it's not fixed yet; see 6178663.
545 CMSCleanOnEnter = false;
546 }
547 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
548 "Inconsistency");
549 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads);
550 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
551
552 // Parallel task queues; these are shared for the
553 // concurrent and stop-world phases of CMS, but
554 // are not shared with parallel scavenge (ParNew).
555 {
556 uint i;
557 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads);
558
559 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
560 || ParallelRefProcEnabled)
561 && num_queues > 0) {
562 _task_queues = new OopTaskQueueSet(num_queues);
563 if (_task_queues == NULL) {
564 log_warning(gc)("task_queues allocation failure.");
565 return;
566 }
567 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
568 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
569 for (i = 0; i < num_queues; i++) {
570 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
571 if (q == NULL) {
572 log_warning(gc)("work_queue allocation failure.");
573 return;
574 }
575 _task_queues->register_queue(i, q);
576 }
577 for (i = 0; i < num_queues; i++) {
578 _task_queues->queue(i)->initialize();
579 _hash_seed[i] = 17; // copied from ParNew
580 }
581 }
582 }
583
584 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
585
586 // Clip CMSBootstrapOccupancy between 0 and 100.
587 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0;
588
589 // Now tell CMS generations the identity of their collector
590 ConcurrentMarkSweepGeneration::set_collector(this);
591
592 // Create & start a CMS thread for this CMS collector
593 _cmsThread = ConcurrentMarkSweepThread::start(this);
594 assert(cmsThread() != NULL, "CMS Thread should have been created");
595 assert(cmsThread()->collector() == this,
596 "CMS Thread should refer to this gen");
597 assert(CGC_lock != NULL, "Where's the CGC_lock?");
598
599 // Support for parallelizing young gen rescan
600 CMSHeap* heap = CMSHeap::heap();
601 assert(heap->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
602 _young_gen = (ParNewGeneration*)heap->young_gen();
603 if (heap->supports_inline_contig_alloc()) {
604 _top_addr = heap->top_addr();
605 _end_addr = heap->end_addr();
606 assert(_young_gen != NULL, "no _young_gen");
607 _eden_chunk_index = 0;
608 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain;
609 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
610 }
611
612 // Support for parallelizing survivor space rescan
613 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
614 const size_t max_plab_samples =
615 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize);
616
617 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
618 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
619 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
620 _survivor_chunk_capacity = max_plab_samples;
621 for (uint i = 0; i < ParallelGCThreads; i++) {
622 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
623 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
624 assert(cur->end() == 0, "Should be 0");
625 assert(cur->array() == vec, "Should be vec");
626 assert(cur->capacity() == max_plab_samples, "Error");
627 }
628 }
629
630 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
631 _gc_counters = new CollectorCounters("CMS", 1);
632 _cgc_counters = new CollectorCounters("CMS stop-the-world phases", 2);
633 _completed_initialization = true;
634 _inter_sweep_timer.start(); // start of time
635 }
636
637 const char* ConcurrentMarkSweepGeneration::name() const {
638 return "concurrent mark-sweep generation";
639 }
640 void ConcurrentMarkSweepGeneration::update_counters() {
641 if (UsePerfData) {
642 _space_counters->update_all();
643 _gen_counters->update_all();
644 }
645 }
646
647 // this is an optimized version of update_counters(). it takes the
648 // used value as a parameter rather than computing it.
649 //
650 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
651 if (UsePerfData) {
652 _space_counters->update_used(used);
653 _space_counters->update_capacity();
654 _gen_counters->update_all();
655 }
656 }
657
658 void ConcurrentMarkSweepGeneration::print() const {
659 Generation::print();
660 cmsSpace()->print();
661 }
662
663 #ifndef PRODUCT
664 void ConcurrentMarkSweepGeneration::print_statistics() {
665 cmsSpace()->printFLCensus(0);
666 }
667 #endif
668
669 size_t
670 ConcurrentMarkSweepGeneration::contiguous_available() const {
671 // dld proposes an improvement in precision here. If the committed
672 // part of the space ends in a free block we should add that to
673 // uncommitted size in the calculation below. Will make this
674 // change later, staying with the approximation below for the
675 // time being. -- ysr.
676 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
677 }
678
679 size_t
680 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
681 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
682 }
683
684 size_t ConcurrentMarkSweepGeneration::max_available() const {
685 return free() + _virtual_space.uncommitted_size();
686 }
687
688 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
689 size_t available = max_available();
690 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
691 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
692 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")",
693 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes);
694 return res;
695 }
696
697 // At a promotion failure dump information on block layout in heap
698 // (cms old generation).
699 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
700 Log(gc, promotion) log;
701 if (log.is_trace()) {
702 LogStream ls(log.trace());
703 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls);
704 }
705 }
706
707 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
708 // Clear the promotion information. These pointers can be adjusted
709 // along with all the other pointers into the heap but
710 // compaction is expected to be a rare event with
711 // a heap using cms so don't do it without seeing the need.
712 for (uint i = 0; i < ParallelGCThreads; i++) {
713 _par_gc_thread_states[i]->promo.reset();
714 }
715 }
716
717 void ConcurrentMarkSweepGeneration::compute_new_size() {
718 assert_locked_or_safepoint(Heap_lock);
719
720 // If incremental collection failed, we just want to expand
721 // to the limit.
722 if (incremental_collection_failed()) {
723 clear_incremental_collection_failed();
724 grow_to_reserved();
725 return;
726 }
727
728 // The heap has been compacted but not reset yet.
729 // Any metric such as free() or used() will be incorrect.
730
731 CardGeneration::compute_new_size();
732
733 // Reset again after a possible resizing
734 if (did_compact()) {
735 cmsSpace()->reset_after_compaction();
736 }
737 }
738
739 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
740 assert_locked_or_safepoint(Heap_lock);
741
742 // If incremental collection failed, we just want to expand
743 // to the limit.
744 if (incremental_collection_failed()) {
745 clear_incremental_collection_failed();
746 grow_to_reserved();
747 return;
748 }
749
750 double free_percentage = ((double) free()) / capacity();
751 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
752 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
753
754 // compute expansion delta needed for reaching desired free percentage
755 if (free_percentage < desired_free_percentage) {
756 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
757 assert(desired_capacity >= capacity(), "invalid expansion size");
758 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
759 Log(gc) log;
760 if (log.is_trace()) {
761 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
762 log.trace("From compute_new_size: ");
763 log.trace(" Free fraction %f", free_percentage);
764 log.trace(" Desired free fraction %f", desired_free_percentage);
765 log.trace(" Maximum free fraction %f", maximum_free_percentage);
766 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000);
767 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000);
768 CMSHeap* heap = CMSHeap::heap();
769 assert(heap->is_old_gen(this), "The CMS generation should always be the old generation");
770 size_t young_size = heap->young_gen()->capacity();
771 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000);
772 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000);
773 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000);
774 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes);
775 }
776 // safe if expansion fails
777 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
778 log.trace(" Expanded free fraction %f", ((double) free()) / capacity());
779 } else {
780 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
781 assert(desired_capacity <= capacity(), "invalid expansion size");
782 size_t shrink_bytes = capacity() - desired_capacity;
783 // Don't shrink unless the delta is greater than the minimum shrink we want
784 if (shrink_bytes >= MinHeapDeltaBytes) {
785 shrink_free_list_by(shrink_bytes);
786 }
787 }
788 }
789
790 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
791 return cmsSpace()->freelistLock();
792 }
793
794 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) {
795 CMSSynchronousYieldRequest yr;
796 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
797 return have_lock_and_allocate(size, tlab);
798 }
799
800 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
801 bool tlab /* ignored */) {
802 assert_lock_strong(freelistLock());
803 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
804 HeapWord* res = cmsSpace()->allocate(adjustedSize);
805 // Allocate the object live (grey) if the background collector has
806 // started marking. This is necessary because the marker may
807 // have passed this address and consequently this object will
808 // not otherwise be greyed and would be incorrectly swept up.
809 // Note that if this object contains references, the writing
810 // of those references will dirty the card containing this object
811 // allowing the object to be blackened (and its references scanned)
812 // either during a preclean phase or at the final checkpoint.
813 if (res != NULL) {
814 // We may block here with an uninitialized object with
815 // its mark-bit or P-bits not yet set. Such objects need
816 // to be safely navigable by block_start().
817 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
818 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
819 collector()->direct_allocated(res, adjustedSize);
820 _direct_allocated_words += adjustedSize;
821 // allocation counters
822 NOT_PRODUCT(
823 _numObjectsAllocated++;
824 _numWordsAllocated += (int)adjustedSize;
825 )
826 }
827 return res;
828 }
829
830 // In the case of direct allocation by mutators in a generation that
831 // is being concurrently collected, the object must be allocated
832 // live (grey) if the background collector has started marking.
833 // This is necessary because the marker may
834 // have passed this address and consequently this object will
835 // not otherwise be greyed and would be incorrectly swept up.
836 // Note that if this object contains references, the writing
837 // of those references will dirty the card containing this object
838 // allowing the object to be blackened (and its references scanned)
839 // either during a preclean phase or at the final checkpoint.
840 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
841 assert(_markBitMap.covers(start, size), "Out of bounds");
842 if (_collectorState >= Marking) {
843 MutexLockerEx y(_markBitMap.lock(),
844 Mutex::_no_safepoint_check_flag);
845 // [see comments preceding SweepClosure::do_blk() below for details]
846 //
847 // Can the P-bits be deleted now? JJJ
848 //
849 // 1. need to mark the object as live so it isn't collected
850 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
851 // 3. need to mark the end of the object so marking, precleaning or sweeping
852 // can skip over uninitialized or unparsable objects. An allocated
853 // object is considered uninitialized for our purposes as long as
854 // its klass word is NULL. All old gen objects are parsable
855 // as soon as they are initialized.)
856 _markBitMap.mark(start); // object is live
857 _markBitMap.mark(start + 1); // object is potentially uninitialized?
858 _markBitMap.mark(start + size - 1);
859 // mark end of object
860 }
861 // check that oop looks uninitialized
862 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
863 }
864
865 void CMSCollector::promoted(bool par, HeapWord* start,
866 bool is_obj_array, size_t obj_size) {
867 assert(_markBitMap.covers(start), "Out of bounds");
868 // See comment in direct_allocated() about when objects should
869 // be allocated live.
870 if (_collectorState >= Marking) {
871 // we already hold the marking bit map lock, taken in
872 // the prologue
873 if (par) {
874 _markBitMap.par_mark(start);
875 } else {
876 _markBitMap.mark(start);
877 }
878 // We don't need to mark the object as uninitialized (as
879 // in direct_allocated above) because this is being done with the
880 // world stopped and the object will be initialized by the
881 // time the marking, precleaning or sweeping get to look at it.
882 // But see the code for copying objects into the CMS generation,
883 // where we need to ensure that concurrent readers of the
884 // block offset table are able to safely navigate a block that
885 // is in flux from being free to being allocated (and in
886 // transition while being copied into) and subsequently
887 // becoming a bona-fide object when the copy/promotion is complete.
888 assert(SafepointSynchronize::is_at_safepoint(),
889 "expect promotion only at safepoints");
890
891 if (_collectorState < Sweeping) {
892 // Mark the appropriate cards in the modUnionTable, so that
893 // this object gets scanned before the sweep. If this is
894 // not done, CMS generation references in the object might
895 // not get marked.
896 // For the case of arrays, which are otherwise precisely
897 // marked, we need to dirty the entire array, not just its head.
898 if (is_obj_array) {
899 // The [par_]mark_range() method expects mr.end() below to
900 // be aligned to the granularity of a bit's representation
901 // in the heap. In the case of the MUT below, that's a
902 // card size.
903 MemRegion mr(start,
904 align_up(start + obj_size,
905 CardTableModRefBS::card_size /* bytes */));
906 if (par) {
907 _modUnionTable.par_mark_range(mr);
908 } else {
909 _modUnionTable.mark_range(mr);
910 }
911 } else { // not an obj array; we can just mark the head
912 if (par) {
913 _modUnionTable.par_mark(start);
914 } else {
915 _modUnionTable.mark(start);
916 }
917 }
918 }
919 }
920 }
921
922 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
923 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
924 // allocate, copy and if necessary update promoinfo --
925 // delegate to underlying space.
926 assert_lock_strong(freelistLock());
927
928 #ifndef PRODUCT
929 if (CMSHeap::heap()->promotion_should_fail()) {
930 return NULL;
931 }
932 #endif // #ifndef PRODUCT
933
934 oop res = _cmsSpace->promote(obj, obj_size);
935 if (res == NULL) {
936 // expand and retry
937 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
938 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
939 // Since this is the old generation, we don't try to promote
940 // into a more senior generation.
941 res = _cmsSpace->promote(obj, obj_size);
942 }
943 if (res != NULL) {
944 // See comment in allocate() about when objects should
945 // be allocated live.
946 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below");
947 collector()->promoted(false, // Not parallel
948 (HeapWord*)res, obj->is_objArray(), obj_size);
949 // promotion counters
950 NOT_PRODUCT(
951 _numObjectsPromoted++;
952 _numWordsPromoted +=
953 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
954 )
955 }
956 return res;
957 }
958
959
960 // IMPORTANT: Notes on object size recognition in CMS.
961 // ---------------------------------------------------
962 // A block of storage in the CMS generation is always in
963 // one of three states. A free block (FREE), an allocated
964 // object (OBJECT) whose size() method reports the correct size,
965 // and an intermediate state (TRANSIENT) in which its size cannot
966 // be accurately determined.
967 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
968 // -----------------------------------------------------
969 // FREE: klass_word & 1 == 1; mark_word holds block size
970 //
971 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
972 // obj->size() computes correct size
973 //
974 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
975 //
976 // STATE IDENTIFICATION: (64 bit+COOPS)
977 // ------------------------------------
978 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
979 //
980 // OBJECT: klass_word installed; klass_word != 0;
981 // obj->size() computes correct size
982 //
983 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
984 //
985 //
986 // STATE TRANSITION DIAGRAM
987 //
988 // mut / parnew mut / parnew
989 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
990 // ^ |
991 // |------------------------ DEAD <------------------------------------|
992 // sweep mut
993 //
994 // While a block is in TRANSIENT state its size cannot be determined
995 // so readers will either need to come back later or stall until
996 // the size can be determined. Note that for the case of direct
997 // allocation, P-bits, when available, may be used to determine the
998 // size of an object that may not yet have been initialized.
999
1000 // Things to support parallel young-gen collection.
1001 oop
1002 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1003 oop old, markOop m,
1004 size_t word_sz) {
1005 #ifndef PRODUCT
1006 if (CMSHeap::heap()->promotion_should_fail()) {
1007 return NULL;
1008 }
1009 #endif // #ifndef PRODUCT
1010
1011 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1012 PromotionInfo* promoInfo = &ps->promo;
1013 // if we are tracking promotions, then first ensure space for
1014 // promotion (including spooling space for saving header if necessary).
1015 // then allocate and copy, then track promoted info if needed.
1016 // When tracking (see PromotionInfo::track()), the mark word may
1017 // be displaced and in this case restoration of the mark word
1018 // occurs in the (oop_since_save_marks_)iterate phase.
1019 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1020 // Out of space for allocating spooling buffers;
1021 // try expanding and allocating spooling buffers.
1022 if (!expand_and_ensure_spooling_space(promoInfo)) {
1023 return NULL;
1024 }
1025 }
1026 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant");
1027 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1028 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1029 if (obj_ptr == NULL) {
1030 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1031 if (obj_ptr == NULL) {
1032 return NULL;
1033 }
1034 }
1035 oop obj = oop(obj_ptr);
1036 OrderAccess::storestore();
1037 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1038 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1039 // IMPORTANT: See note on object initialization for CMS above.
1040 // Otherwise, copy the object. Here we must be careful to insert the
1041 // klass pointer last, since this marks the block as an allocated object.
1042 // Except with compressed oops it's the mark word.
1043 HeapWord* old_ptr = (HeapWord*)old;
1044 // Restore the mark word copied above.
1045 obj->set_mark(m);
1046 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1047 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1048 OrderAccess::storestore();
1049
1050 if (UseCompressedClassPointers) {
1051 // Copy gap missed by (aligned) header size calculation below
1052 obj->set_klass_gap(old->klass_gap());
1053 }
1054 if (word_sz > (size_t)oopDesc::header_size()) {
1055 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1056 obj_ptr + oopDesc::header_size(),
1057 word_sz - oopDesc::header_size());
1058 }
1059
1060 // Now we can track the promoted object, if necessary. We take care
1061 // to delay the transition from uninitialized to full object
1062 // (i.e., insertion of klass pointer) until after, so that it
1063 // atomically becomes a promoted object.
1064 if (promoInfo->tracking()) {
1065 promoInfo->track((PromotedObject*)obj, old->klass());
1066 }
1067 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1068 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1069 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below");
1070
1071 // Finally, install the klass pointer (this should be volatile).
1072 OrderAccess::storestore();
1073 obj->set_klass(old->klass());
1074 // We should now be able to calculate the right size for this object
1075 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1076
1077 collector()->promoted(true, // parallel
1078 obj_ptr, old->is_objArray(), word_sz);
1079
1080 NOT_PRODUCT(
1081 Atomic::inc(&_numObjectsPromoted);
1082 Atomic::add(alloc_sz, &_numWordsPromoted);
1083 )
1084
1085 return obj;
1086 }
1087
1088 void
1089 ConcurrentMarkSweepGeneration::
1090 par_promote_alloc_done(int thread_num) {
1091 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1092 ps->lab.retire(thread_num);
1093 }
1094
1095 void
1096 ConcurrentMarkSweepGeneration::
1097 par_oop_since_save_marks_iterate_done(int thread_num) {
1098 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1099 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1100 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1101
1102 // Because card-scanning has been completed, subsequent phases
1103 // (e.g., reference processing) will not need to recognize which
1104 // objects have been promoted during this GC. So, we can now disable
1105 // promotion tracking.
1106 ps->promo.stopTrackingPromotions();
1107 }
1108
1109 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1110 size_t size,
1111 bool tlab)
1112 {
1113 // We allow a STW collection only if a full
1114 // collection was requested.
1115 return full || should_allocate(size, tlab); // FIX ME !!!
1116 // This and promotion failure handling are connected at the
1117 // hip and should be fixed by untying them.
1118 }
1119
1120 bool CMSCollector::shouldConcurrentCollect() {
1121 LogTarget(Trace, gc) log;
1122
1123 if (_full_gc_requested) {
1124 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)");
1125 return true;
1126 }
1127
1128 FreelistLocker x(this);
1129 // ------------------------------------------------------------------
1130 // Print out lots of information which affects the initiation of
1131 // a collection.
1132 if (log.is_enabled() && stats().valid()) {
1133 log.print("CMSCollector shouldConcurrentCollect: ");
1134
1135 LogStream out(log);
1136 stats().print_on(&out);
1137
1138 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full());
1139 log.print("free=" SIZE_FORMAT, _cmsGen->free());
1140 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available());
1141 log.print("promotion_rate=%g", stats().promotion_rate());
1142 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate());
1143 log.print("occupancy=%3.7f", _cmsGen->occupancy());
1144 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1145 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1146 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1147 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect());
1148 }
1149 // ------------------------------------------------------------------
1150
1151 // If the estimated time to complete a cms collection (cms_duration())
1152 // is less than the estimated time remaining until the cms generation
1153 // is full, start a collection.
1154 if (!UseCMSInitiatingOccupancyOnly) {
1155 if (stats().valid()) {
1156 if (stats().time_until_cms_start() == 0.0) {
1157 return true;
1158 }
1159 } else {
1160 // We want to conservatively collect somewhat early in order
1161 // to try and "bootstrap" our CMS/promotion statistics;
1162 // this branch will not fire after the first successful CMS
1163 // collection because the stats should then be valid.
1164 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1165 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f",
1166 _cmsGen->occupancy(), _bootstrap_occupancy);
1167 return true;
1168 }
1169 }
1170 }
1171
1172 // Otherwise, we start a collection cycle if
1173 // old gen want a collection cycle started. Each may use
1174 // an appropriate criterion for making this decision.
1175 // XXX We need to make sure that the gen expansion
1176 // criterion dovetails well with this. XXX NEED TO FIX THIS
1177 if (_cmsGen->should_concurrent_collect()) {
1178 log.print("CMS old gen initiated");
1179 return true;
1180 }
1181
1182 // We start a collection if we believe an incremental collection may fail;
1183 // this is not likely to be productive in practice because it's probably too
1184 // late anyway.
1185 CMSHeap* heap = CMSHeap::heap();
1186 assert(heap->collector_policy()->is_generation_policy(),
1187 "You may want to check the correctness of the following");
1188 if (heap->incremental_collection_will_fail(true /* consult_young */)) {
1189 log.print("CMSCollector: collect because incremental collection will fail ");
1190 return true;
1191 }
1192
1193 if (MetaspaceGC::should_concurrent_collect()) {
1194 log.print("CMSCollector: collect for metadata allocation ");
1195 return true;
1196 }
1197
1198 // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1199 if (CMSTriggerInterval >= 0) {
1200 if (CMSTriggerInterval == 0) {
1201 // Trigger always
1202 return true;
1203 }
1204
1205 // Check the CMS time since begin (we do not check the stats validity
1206 // as we want to be able to trigger the first CMS cycle as well)
1207 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1208 if (stats().valid()) {
1209 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1210 stats().cms_time_since_begin());
1211 } else {
1212 log.print("CMSCollector: collect because of trigger interval (first collection)");
1213 }
1214 return true;
1215 }
1216 }
1217
1218 return false;
1219 }
1220
1221 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1222
1223 // Clear _expansion_cause fields of constituent generations
1224 void CMSCollector::clear_expansion_cause() {
1225 _cmsGen->clear_expansion_cause();
1226 }
1227
1228 // We should be conservative in starting a collection cycle. To
1229 // start too eagerly runs the risk of collecting too often in the
1230 // extreme. To collect too rarely falls back on full collections,
1231 // which works, even if not optimum in terms of concurrent work.
1232 // As a work around for too eagerly collecting, use the flag
1233 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1234 // giving the user an easily understandable way of controlling the
1235 // collections.
1236 // We want to start a new collection cycle if any of the following
1237 // conditions hold:
1238 // . our current occupancy exceeds the configured initiating occupancy
1239 // for this generation, or
1240 // . we recently needed to expand this space and have not, since that
1241 // expansion, done a collection of this generation, or
1242 // . the underlying space believes that it may be a good idea to initiate
1243 // a concurrent collection (this may be based on criteria such as the
1244 // following: the space uses linear allocation and linear allocation is
1245 // going to fail, or there is believed to be excessive fragmentation in
1246 // the generation, etc... or ...
1247 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1248 // the case of the old generation; see CR 6543076):
1249 // we may be approaching a point at which allocation requests may fail because
1250 // we will be out of sufficient free space given allocation rate estimates.]
1251 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1252
1253 assert_lock_strong(freelistLock());
1254 if (occupancy() > initiating_occupancy()) {
1255 log_trace(gc)(" %s: collect because of occupancy %f / %f ",
1256 short_name(), occupancy(), initiating_occupancy());
1257 return true;
1258 }
1259 if (UseCMSInitiatingOccupancyOnly) {
1260 return false;
1261 }
1262 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1263 log_trace(gc)(" %s: collect because expanded for allocation ", short_name());
1264 return true;
1265 }
1266 return false;
1267 }
1268
1269 void ConcurrentMarkSweepGeneration::collect(bool full,
1270 bool clear_all_soft_refs,
1271 size_t size,
1272 bool tlab)
1273 {
1274 collector()->collect(full, clear_all_soft_refs, size, tlab);
1275 }
1276
1277 void CMSCollector::collect(bool full,
1278 bool clear_all_soft_refs,
1279 size_t size,
1280 bool tlab)
1281 {
1282 // The following "if" branch is present for defensive reasons.
1283 // In the current uses of this interface, it can be replaced with:
1284 // assert(!GCLocker.is_active(), "Can't be called otherwise");
1285 // But I am not placing that assert here to allow future
1286 // generality in invoking this interface.
1287 if (GCLocker::is_active()) {
1288 // A consistency test for GCLocker
1289 assert(GCLocker::needs_gc(), "Should have been set already");
1290 // Skip this foreground collection, instead
1291 // expanding the heap if necessary.
1292 // Need the free list locks for the call to free() in compute_new_size()
1293 compute_new_size();
1294 return;
1295 }
1296 acquire_control_and_collect(full, clear_all_soft_refs);
1297 }
1298
1299 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1300 CMSHeap* heap = CMSHeap::heap();
1301 unsigned int gc_count = heap->total_full_collections();
1302 if (gc_count == full_gc_count) {
1303 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1304 _full_gc_requested = true;
1305 _full_gc_cause = cause;
1306 CGC_lock->notify(); // nudge CMS thread
1307 } else {
1308 assert(gc_count > full_gc_count, "Error: causal loop");
1309 }
1310 }
1311
1312 bool CMSCollector::is_external_interruption() {
1313 GCCause::Cause cause = CMSHeap::heap()->gc_cause();
1314 return GCCause::is_user_requested_gc(cause) ||
1315 GCCause::is_serviceability_requested_gc(cause);
1316 }
1317
1318 void CMSCollector::report_concurrent_mode_interruption() {
1319 if (is_external_interruption()) {
1320 log_debug(gc)("Concurrent mode interrupted");
1321 } else {
1322 log_debug(gc)("Concurrent mode failure");
1323 _gc_tracer_cm->report_concurrent_mode_failure();
1324 }
1325 }
1326
1327
1328 // The foreground and background collectors need to coordinate in order
1329 // to make sure that they do not mutually interfere with CMS collections.
1330 // When a background collection is active,
1331 // the foreground collector may need to take over (preempt) and
1332 // synchronously complete an ongoing collection. Depending on the
1333 // frequency of the background collections and the heap usage
1334 // of the application, this preemption can be seldom or frequent.
1335 // There are only certain
1336 // points in the background collection that the "collection-baton"
1337 // can be passed to the foreground collector.
1338 //
1339 // The foreground collector will wait for the baton before
1340 // starting any part of the collection. The foreground collector
1341 // will only wait at one location.
1342 //
1343 // The background collector will yield the baton before starting a new
1344 // phase of the collection (e.g., before initial marking, marking from roots,
1345 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1346 // of the loop which switches the phases. The background collector does some
1347 // of the phases (initial mark, final re-mark) with the world stopped.
1348 // Because of locking involved in stopping the world,
1349 // the foreground collector should not block waiting for the background
1350 // collector when it is doing a stop-the-world phase. The background
1351 // collector will yield the baton at an additional point just before
1352 // it enters a stop-the-world phase. Once the world is stopped, the
1353 // background collector checks the phase of the collection. If the
1354 // phase has not changed, it proceeds with the collection. If the
1355 // phase has changed, it skips that phase of the collection. See
1356 // the comments on the use of the Heap_lock in collect_in_background().
1357 //
1358 // Variable used in baton passing.
1359 // _foregroundGCIsActive - Set to true by the foreground collector when
1360 // it wants the baton. The foreground clears it when it has finished
1361 // the collection.
1362 // _foregroundGCShouldWait - Set to true by the background collector
1363 // when it is running. The foreground collector waits while
1364 // _foregroundGCShouldWait is true.
1365 // CGC_lock - monitor used to protect access to the above variables
1366 // and to notify the foreground and background collectors.
1367 // _collectorState - current state of the CMS collection.
1368 //
1369 // The foreground collector
1370 // acquires the CGC_lock
1371 // sets _foregroundGCIsActive
1372 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1373 // various locks acquired in preparation for the collection
1374 // are released so as not to block the background collector
1375 // that is in the midst of a collection
1376 // proceeds with the collection
1377 // clears _foregroundGCIsActive
1378 // returns
1379 //
1380 // The background collector in a loop iterating on the phases of the
1381 // collection
1382 // acquires the CGC_lock
1383 // sets _foregroundGCShouldWait
1384 // if _foregroundGCIsActive is set
1385 // clears _foregroundGCShouldWait, notifies _CGC_lock
1386 // waits on _CGC_lock for _foregroundGCIsActive to become false
1387 // and exits the loop.
1388 // otherwise
1389 // proceed with that phase of the collection
1390 // if the phase is a stop-the-world phase,
1391 // yield the baton once more just before enqueueing
1392 // the stop-world CMS operation (executed by the VM thread).
1393 // returns after all phases of the collection are done
1394 //
1395
1396 void CMSCollector::acquire_control_and_collect(bool full,
1397 bool clear_all_soft_refs) {
1398 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1399 assert(!Thread::current()->is_ConcurrentGC_thread(),
1400 "shouldn't try to acquire control from self!");
1401
1402 // Start the protocol for acquiring control of the
1403 // collection from the background collector (aka CMS thread).
1404 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1405 "VM thread should have CMS token");
1406 // Remember the possibly interrupted state of an ongoing
1407 // concurrent collection
1408 CollectorState first_state = _collectorState;
1409
1410 // Signal to a possibly ongoing concurrent collection that
1411 // we want to do a foreground collection.
1412 _foregroundGCIsActive = true;
1413
1414 // release locks and wait for a notify from the background collector
1415 // releasing the locks in only necessary for phases which
1416 // do yields to improve the granularity of the collection.
1417 assert_lock_strong(bitMapLock());
1418 // We need to lock the Free list lock for the space that we are
1419 // currently collecting.
1420 assert(haveFreelistLocks(), "Must be holding free list locks");
1421 bitMapLock()->unlock();
1422 releaseFreelistLocks();
1423 {
1424 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1425 if (_foregroundGCShouldWait) {
1426 // We are going to be waiting for action for the CMS thread;
1427 // it had better not be gone (for instance at shutdown)!
1428 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(),
1429 "CMS thread must be running");
1430 // Wait here until the background collector gives us the go-ahead
1431 ConcurrentMarkSweepThread::clear_CMS_flag(
1432 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1433 // Get a possibly blocked CMS thread going:
1434 // Note that we set _foregroundGCIsActive true above,
1435 // without protection of the CGC_lock.
1436 CGC_lock->notify();
1437 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1438 "Possible deadlock");
1439 while (_foregroundGCShouldWait) {
1440 // wait for notification
1441 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1442 // Possibility of delay/starvation here, since CMS token does
1443 // not know to give priority to VM thread? Actually, i think
1444 // there wouldn't be any delay/starvation, but the proof of
1445 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1446 }
1447 ConcurrentMarkSweepThread::set_CMS_flag(
1448 ConcurrentMarkSweepThread::CMS_vm_has_token);
1449 }
1450 }
1451 // The CMS_token is already held. Get back the other locks.
1452 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1453 "VM thread should have CMS token");
1454 getFreelistLocks();
1455 bitMapLock()->lock_without_safepoint_check();
1456 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d",
1457 p2i(Thread::current()), first_state);
1458 log_debug(gc, state)(" gets control with state %d", _collectorState);
1459
1460 // Inform cms gen if this was due to partial collection failing.
1461 // The CMS gen may use this fact to determine its expansion policy.
1462 CMSHeap* heap = CMSHeap::heap();
1463 if (heap->incremental_collection_will_fail(false /* don't consult_young */)) {
1464 assert(!_cmsGen->incremental_collection_failed(),
1465 "Should have been noticed, reacted to and cleared");
1466 _cmsGen->set_incremental_collection_failed();
1467 }
1468
1469 if (first_state > Idling) {
1470 report_concurrent_mode_interruption();
1471 }
1472
1473 set_did_compact(true);
1474
1475 // If the collection is being acquired from the background
1476 // collector, there may be references on the discovered
1477 // references lists. Abandon those references, since some
1478 // of them may have become unreachable after concurrent
1479 // discovery; the STW compacting collector will redo discovery
1480 // more precisely, without being subject to floating garbage.
1481 // Leaving otherwise unreachable references in the discovered
1482 // lists would require special handling.
1483 ref_processor()->disable_discovery();
1484 ref_processor()->abandon_partial_discovery();
1485 ref_processor()->verify_no_references_recorded();
1486
1487 if (first_state > Idling) {
1488 save_heap_summary();
1489 }
1490
1491 do_compaction_work(clear_all_soft_refs);
1492
1493 // Has the GC time limit been exceeded?
1494 size_t max_eden_size = _young_gen->max_eden_size();
1495 GCCause::Cause gc_cause = heap->gc_cause();
1496 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1497 _young_gen->eden()->used(),
1498 _cmsGen->max_capacity(),
1499 max_eden_size,
1500 full,
1501 gc_cause,
1502 heap->collector_policy());
1503
1504 // Reset the expansion cause, now that we just completed
1505 // a collection cycle.
1506 clear_expansion_cause();
1507 _foregroundGCIsActive = false;
1508 return;
1509 }
1510
1511 // Resize the tenured generation
1512 // after obtaining the free list locks for the
1513 // two generations.
1514 void CMSCollector::compute_new_size() {
1515 assert_locked_or_safepoint(Heap_lock);
1516 FreelistLocker z(this);
1517 MetaspaceGC::compute_new_size();
1518 _cmsGen->compute_new_size_free_list();
1519 }
1520
1521 // A work method used by the foreground collector to do
1522 // a mark-sweep-compact.
1523 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1524 CMSHeap* heap = CMSHeap::heap();
1525
1526 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1527 gc_timer->register_gc_start();
1528
1529 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1530 gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start());
1531
1532 heap->pre_full_gc_dump(gc_timer);
1533
1534 GCTraceTime(Trace, gc, phases) t("CMS:MSC");
1535
1536 // Temporarily widen the span of the weak reference processing to
1537 // the entire heap.
1538 MemRegion new_span(CMSHeap::heap()->reserved_region());
1539 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1540 // Temporarily, clear the "is_alive_non_header" field of the
1541 // reference processor.
1542 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1543 // Temporarily make reference _processing_ single threaded (non-MT).
1544 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1545 // Temporarily make refs discovery atomic
1546 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1547 // Temporarily make reference _discovery_ single threaded (non-MT)
1548 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1549
1550 ref_processor()->set_enqueuing_is_done(false);
1551 ref_processor()->enable_discovery();
1552 ref_processor()->setup_policy(clear_all_soft_refs);
1553 // If an asynchronous collection finishes, the _modUnionTable is
1554 // all clear. If we are assuming the collection from an asynchronous
1555 // collection, clear the _modUnionTable.
1556 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1557 "_modUnionTable should be clear if the baton was not passed");
1558 _modUnionTable.clear_all();
1559 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(),
1560 "mod union for klasses should be clear if the baton was passed");
1561 _ct->cld_rem_set()->clear_mod_union();
1562
1563
1564 // We must adjust the allocation statistics being maintained
1565 // in the free list space. We do so by reading and clearing
1566 // the sweep timer and updating the block flux rate estimates below.
1567 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1568 if (_inter_sweep_timer.is_active()) {
1569 _inter_sweep_timer.stop();
1570 // Note that we do not use this sample to update the _inter_sweep_estimate.
1571 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1572 _inter_sweep_estimate.padded_average(),
1573 _intra_sweep_estimate.padded_average());
1574 }
1575
1576 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs);
1577 #ifdef ASSERT
1578 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1579 size_t free_size = cms_space->free();
1580 assert(free_size ==
1581 pointer_delta(cms_space->end(), cms_space->compaction_top())
1582 * HeapWordSize,
1583 "All the free space should be compacted into one chunk at top");
1584 assert(cms_space->dictionary()->total_chunk_size(
1585 debug_only(cms_space->freelistLock())) == 0 ||
1586 cms_space->totalSizeInIndexedFreeLists() == 0,
1587 "All the free space should be in a single chunk");
1588 size_t num = cms_space->totalCount();
1589 assert((free_size == 0 && num == 0) ||
1590 (free_size > 0 && (num == 1 || num == 2)),
1591 "There should be at most 2 free chunks after compaction");
1592 #endif // ASSERT
1593 _collectorState = Resetting;
1594 assert(_restart_addr == NULL,
1595 "Should have been NULL'd before baton was passed");
1596 reset_stw();
1597 _cmsGen->reset_after_compaction();
1598 _concurrent_cycles_since_last_unload = 0;
1599
1600 // Clear any data recorded in the PLAB chunk arrays.
1601 if (_survivor_plab_array != NULL) {
1602 reset_survivor_plab_arrays();
1603 }
1604
1605 // Adjust the per-size allocation stats for the next epoch.
1606 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1607 // Restart the "inter sweep timer" for the next epoch.
1608 _inter_sweep_timer.reset();
1609 _inter_sweep_timer.start();
1610
1611 // No longer a need to do a concurrent collection for Metaspace.
1612 MetaspaceGC::set_should_concurrent_collect(false);
1613
1614 heap->post_full_gc_dump(gc_timer);
1615
1616 gc_timer->register_gc_end();
1617
1618 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1619
1620 // For a mark-sweep-compact, compute_new_size() will be called
1621 // in the heap's do_collection() method.
1622 }
1623
1624 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1625 Log(gc, heap) log;
1626 if (!log.is_trace()) {
1627 return;
1628 }
1629
1630 ContiguousSpace* eden_space = _young_gen->eden();
1631 ContiguousSpace* from_space = _young_gen->from();
1632 ContiguousSpace* to_space = _young_gen->to();
1633 // Eden
1634 if (_eden_chunk_array != NULL) {
1635 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1636 p2i(eden_space->bottom()), p2i(eden_space->top()),
1637 p2i(eden_space->end()), eden_space->capacity());
1638 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT,
1639 _eden_chunk_index, _eden_chunk_capacity);
1640 for (size_t i = 0; i < _eden_chunk_index; i++) {
1641 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i]));
1642 }
1643 }
1644 // Survivor
1645 if (_survivor_chunk_array != NULL) {
1646 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1647 p2i(from_space->bottom()), p2i(from_space->top()),
1648 p2i(from_space->end()), from_space->capacity());
1649 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT,
1650 _survivor_chunk_index, _survivor_chunk_capacity);
1651 for (size_t i = 0; i < _survivor_chunk_index; i++) {
1652 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i]));
1653 }
1654 }
1655 }
1656
1657 void CMSCollector::getFreelistLocks() const {
1658 // Get locks for all free lists in all generations that this
1659 // collector is responsible for
1660 _cmsGen->freelistLock()->lock_without_safepoint_check();
1661 }
1662
1663 void CMSCollector::releaseFreelistLocks() const {
1664 // Release locks for all free lists in all generations that this
1665 // collector is responsible for
1666 _cmsGen->freelistLock()->unlock();
1667 }
1668
1669 bool CMSCollector::haveFreelistLocks() const {
1670 // Check locks for all free lists in all generations that this
1671 // collector is responsible for
1672 assert_lock_strong(_cmsGen->freelistLock());
1673 PRODUCT_ONLY(ShouldNotReachHere());
1674 return true;
1675 }
1676
1677 // A utility class that is used by the CMS collector to
1678 // temporarily "release" the foreground collector from its
1679 // usual obligation to wait for the background collector to
1680 // complete an ongoing phase before proceeding.
1681 class ReleaseForegroundGC: public StackObj {
1682 private:
1683 CMSCollector* _c;
1684 public:
1685 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1686 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1687 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1688 // allow a potentially blocked foreground collector to proceed
1689 _c->_foregroundGCShouldWait = false;
1690 if (_c->_foregroundGCIsActive) {
1691 CGC_lock->notify();
1692 }
1693 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1694 "Possible deadlock");
1695 }
1696
1697 ~ReleaseForegroundGC() {
1698 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1699 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1700 _c->_foregroundGCShouldWait = true;
1701 }
1702 };
1703
1704 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1705 assert(Thread::current()->is_ConcurrentGC_thread(),
1706 "A CMS asynchronous collection is only allowed on a CMS thread.");
1707
1708 CMSHeap* heap = CMSHeap::heap();
1709 {
1710 bool safepoint_check = Mutex::_no_safepoint_check_flag;
1711 MutexLockerEx hl(Heap_lock, safepoint_check);
1712 FreelistLocker fll(this);
1713 MutexLockerEx x(CGC_lock, safepoint_check);
1714 if (_foregroundGCIsActive) {
1715 // The foreground collector is. Skip this
1716 // background collection.
1717 assert(!_foregroundGCShouldWait, "Should be clear");
1718 return;
1719 } else {
1720 assert(_collectorState == Idling, "Should be idling before start.");
1721 _collectorState = InitialMarking;
1722 register_gc_start(cause);
1723 // Reset the expansion cause, now that we are about to begin
1724 // a new cycle.
1725 clear_expansion_cause();
1726
1727 // Clear the MetaspaceGC flag since a concurrent collection
1728 // is starting but also clear it after the collection.
1729 MetaspaceGC::set_should_concurrent_collect(false);
1730 }
1731 // Decide if we want to enable class unloading as part of the
1732 // ensuing concurrent GC cycle.
1733 update_should_unload_classes();
1734 _full_gc_requested = false; // acks all outstanding full gc requests
1735 _full_gc_cause = GCCause::_no_gc;
1736 // Signal that we are about to start a collection
1737 heap->increment_total_full_collections(); // ... starting a collection cycle
1738 _collection_count_start = heap->total_full_collections();
1739 }
1740
1741 size_t prev_used = _cmsGen->used();
1742
1743 // The change of the collection state is normally done at this level;
1744 // the exceptions are phases that are executed while the world is
1745 // stopped. For those phases the change of state is done while the
1746 // world is stopped. For baton passing purposes this allows the
1747 // background collector to finish the phase and change state atomically.
1748 // The foreground collector cannot wait on a phase that is done
1749 // while the world is stopped because the foreground collector already
1750 // has the world stopped and would deadlock.
1751 while (_collectorState != Idling) {
1752 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d",
1753 p2i(Thread::current()), _collectorState);
1754 // The foreground collector
1755 // holds the Heap_lock throughout its collection.
1756 // holds the CMS token (but not the lock)
1757 // except while it is waiting for the background collector to yield.
1758 //
1759 // The foreground collector should be blocked (not for long)
1760 // if the background collector is about to start a phase
1761 // executed with world stopped. If the background
1762 // collector has already started such a phase, the
1763 // foreground collector is blocked waiting for the
1764 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
1765 // are executed in the VM thread.
1766 //
1767 // The locking order is
1768 // PendingListLock (PLL) -- if applicable (FinalMarking)
1769 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
1770 // CMS token (claimed in
1771 // stop_world_and_do() -->
1772 // safepoint_synchronize() -->
1773 // CMSThread::synchronize())
1774
1775 {
1776 // Check if the FG collector wants us to yield.
1777 CMSTokenSync x(true); // is cms thread
1778 if (waitForForegroundGC()) {
1779 // We yielded to a foreground GC, nothing more to be
1780 // done this round.
1781 assert(_foregroundGCShouldWait == false, "We set it to false in "
1782 "waitForForegroundGC()");
1783 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1784 p2i(Thread::current()), _collectorState);
1785 return;
1786 } else {
1787 // The background collector can run but check to see if the
1788 // foreground collector has done a collection while the
1789 // background collector was waiting to get the CGC_lock
1790 // above. If yes, break so that _foregroundGCShouldWait
1791 // is cleared before returning.
1792 if (_collectorState == Idling) {
1793 break;
1794 }
1795 }
1796 }
1797
1798 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1799 "should be waiting");
1800
1801 switch (_collectorState) {
1802 case InitialMarking:
1803 {
1804 ReleaseForegroundGC x(this);
1805 stats().record_cms_begin();
1806 VM_CMS_Initial_Mark initial_mark_op(this);
1807 VMThread::execute(&initial_mark_op);
1808 }
1809 // The collector state may be any legal state at this point
1810 // since the background collector may have yielded to the
1811 // foreground collector.
1812 break;
1813 case Marking:
1814 // initial marking in checkpointRootsInitialWork has been completed
1815 if (markFromRoots()) { // we were successful
1816 assert(_collectorState == Precleaning, "Collector state should "
1817 "have changed");
1818 } else {
1819 assert(_foregroundGCIsActive, "Internal state inconsistency");
1820 }
1821 break;
1822 case Precleaning:
1823 // marking from roots in markFromRoots has been completed
1824 preclean();
1825 assert(_collectorState == AbortablePreclean ||
1826 _collectorState == FinalMarking,
1827 "Collector state should have changed");
1828 break;
1829 case AbortablePreclean:
1830 abortable_preclean();
1831 assert(_collectorState == FinalMarking, "Collector state should "
1832 "have changed");
1833 break;
1834 case FinalMarking:
1835 {
1836 ReleaseForegroundGC x(this);
1837
1838 VM_CMS_Final_Remark final_remark_op(this);
1839 VMThread::execute(&final_remark_op);
1840 }
1841 assert(_foregroundGCShouldWait, "block post-condition");
1842 break;
1843 case Sweeping:
1844 // final marking in checkpointRootsFinal has been completed
1845 sweep();
1846 assert(_collectorState == Resizing, "Collector state change "
1847 "to Resizing must be done under the free_list_lock");
1848
1849 case Resizing: {
1850 // Sweeping has been completed...
1851 // At this point the background collection has completed.
1852 // Don't move the call to compute_new_size() down
1853 // into code that might be executed if the background
1854 // collection was preempted.
1855 {
1856 ReleaseForegroundGC x(this); // unblock FG collection
1857 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
1858 CMSTokenSync z(true); // not strictly needed.
1859 if (_collectorState == Resizing) {
1860 compute_new_size();
1861 save_heap_summary();
1862 _collectorState = Resetting;
1863 } else {
1864 assert(_collectorState == Idling, "The state should only change"
1865 " because the foreground collector has finished the collection");
1866 }
1867 }
1868 break;
1869 }
1870 case Resetting:
1871 // CMS heap resizing has been completed
1872 reset_concurrent();
1873 assert(_collectorState == Idling, "Collector state should "
1874 "have changed");
1875
1876 MetaspaceGC::set_should_concurrent_collect(false);
1877
1878 stats().record_cms_end();
1879 // Don't move the concurrent_phases_end() and compute_new_size()
1880 // calls to here because a preempted background collection
1881 // has it's state set to "Resetting".
1882 break;
1883 case Idling:
1884 default:
1885 ShouldNotReachHere();
1886 break;
1887 }
1888 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d",
1889 p2i(Thread::current()), _collectorState);
1890 assert(_foregroundGCShouldWait, "block post-condition");
1891 }
1892
1893 // Should this be in gc_epilogue?
1894 collector_policy()->counters()->update_counters();
1895
1896 {
1897 // Clear _foregroundGCShouldWait and, in the event that the
1898 // foreground collector is waiting, notify it, before
1899 // returning.
1900 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1901 _foregroundGCShouldWait = false;
1902 if (_foregroundGCIsActive) {
1903 CGC_lock->notify();
1904 }
1905 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1906 "Possible deadlock");
1907 }
1908 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1909 p2i(Thread::current()), _collectorState);
1910 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)",
1911 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K);
1912 }
1913
1914 void CMSCollector::register_gc_start(GCCause::Cause cause) {
1915 _cms_start_registered = true;
1916 _gc_timer_cm->register_gc_start();
1917 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
1918 }
1919
1920 void CMSCollector::register_gc_end() {
1921 if (_cms_start_registered) {
1922 report_heap_summary(GCWhen::AfterGC);
1923
1924 _gc_timer_cm->register_gc_end();
1925 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1926 _cms_start_registered = false;
1927 }
1928 }
1929
1930 void CMSCollector::save_heap_summary() {
1931 CMSHeap* heap = CMSHeap::heap();
1932 _last_heap_summary = heap->create_heap_summary();
1933 _last_metaspace_summary = heap->create_metaspace_summary();
1934 }
1935
1936 void CMSCollector::report_heap_summary(GCWhen::Type when) {
1937 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
1938 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
1939 }
1940
1941 bool CMSCollector::waitForForegroundGC() {
1942 bool res = false;
1943 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1944 "CMS thread should have CMS token");
1945 // Block the foreground collector until the
1946 // background collectors decides whether to
1947 // yield.
1948 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1949 _foregroundGCShouldWait = true;
1950 if (_foregroundGCIsActive) {
1951 // The background collector yields to the
1952 // foreground collector and returns a value
1953 // indicating that it has yielded. The foreground
1954 // collector can proceed.
1955 res = true;
1956 _foregroundGCShouldWait = false;
1957 ConcurrentMarkSweepThread::clear_CMS_flag(
1958 ConcurrentMarkSweepThread::CMS_cms_has_token);
1959 ConcurrentMarkSweepThread::set_CMS_flag(
1960 ConcurrentMarkSweepThread::CMS_cms_wants_token);
1961 // Get a possibly blocked foreground thread going
1962 CGC_lock->notify();
1963 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
1964 p2i(Thread::current()), _collectorState);
1965 while (_foregroundGCIsActive) {
1966 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1967 }
1968 ConcurrentMarkSweepThread::set_CMS_flag(
1969 ConcurrentMarkSweepThread::CMS_cms_has_token);
1970 ConcurrentMarkSweepThread::clear_CMS_flag(
1971 ConcurrentMarkSweepThread::CMS_cms_wants_token);
1972 }
1973 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
1974 p2i(Thread::current()), _collectorState);
1975 return res;
1976 }
1977
1978 // Because of the need to lock the free lists and other structures in
1979 // the collector, common to all the generations that the collector is
1980 // collecting, we need the gc_prologues of individual CMS generations
1981 // delegate to their collector. It may have been simpler had the
1982 // current infrastructure allowed one to call a prologue on a
1983 // collector. In the absence of that we have the generation's
1984 // prologue delegate to the collector, which delegates back
1985 // some "local" work to a worker method in the individual generations
1986 // that it's responsible for collecting, while itself doing any
1987 // work common to all generations it's responsible for. A similar
1988 // comment applies to the gc_epilogue()'s.
1989 // The role of the variable _between_prologue_and_epilogue is to
1990 // enforce the invocation protocol.
1991 void CMSCollector::gc_prologue(bool full) {
1992 // Call gc_prologue_work() for the CMSGen
1993 // we are responsible for.
1994
1995 // The following locking discipline assumes that we are only called
1996 // when the world is stopped.
1997 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
1998
1999 // The CMSCollector prologue must call the gc_prologues for the
2000 // "generations" that it's responsible
2001 // for.
2002
2003 assert( Thread::current()->is_VM_thread()
2004 || ( CMSScavengeBeforeRemark
2005 && Thread::current()->is_ConcurrentGC_thread()),
2006 "Incorrect thread type for prologue execution");
2007
2008 if (_between_prologue_and_epilogue) {
2009 // We have already been invoked; this is a gc_prologue delegation
2010 // from yet another CMS generation that we are responsible for, just
2011 // ignore it since all relevant work has already been done.
2012 return;
2013 }
2014
2015 // set a bit saying prologue has been called; cleared in epilogue
2016 _between_prologue_and_epilogue = true;
2017 // Claim locks for common data structures, then call gc_prologue_work()
2018 // for each CMSGen.
2019
2020 getFreelistLocks(); // gets free list locks on constituent spaces
2021 bitMapLock()->lock_without_safepoint_check();
2022
2023 // Should call gc_prologue_work() for all cms gens we are responsible for
2024 bool duringMarking = _collectorState >= Marking
2025 && _collectorState < Sweeping;
2026
2027 // The young collections clear the modified oops state, which tells if
2028 // there are any modified oops in the class. The remark phase also needs
2029 // that information. Tell the young collection to save the union of all
2030 // modified klasses.
2031 if (duringMarking) {
2032 _ct->cld_rem_set()->set_accumulate_modified_oops(true);
2033 }
2034
2035 bool registerClosure = duringMarking;
2036
2037 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2038
2039 if (!full) {
2040 stats().record_gc0_begin();
2041 }
2042 }
2043
2044 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2045
2046 _capacity_at_prologue = capacity();
2047 _used_at_prologue = used();
2048
2049 // We enable promotion tracking so that card-scanning can recognize
2050 // which objects have been promoted during this GC and skip them.
2051 for (uint i = 0; i < ParallelGCThreads; i++) {
2052 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2053 }
2054
2055 // Delegate to CMScollector which knows how to coordinate between
2056 // this and any other CMS generations that it is responsible for
2057 // collecting.
2058 collector()->gc_prologue(full);
2059 }
2060
2061 // This is a "private" interface for use by this generation's CMSCollector.
2062 // Not to be called directly by any other entity (for instance,
2063 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2064 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2065 bool registerClosure, ModUnionClosure* modUnionClosure) {
2066 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2067 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2068 "Should be NULL");
2069 if (registerClosure) {
2070 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2071 }
2072 cmsSpace()->gc_prologue();
2073 // Clear stat counters
2074 NOT_PRODUCT(
2075 assert(_numObjectsPromoted == 0, "check");
2076 assert(_numWordsPromoted == 0, "check");
2077 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently",
2078 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2079 _numObjectsAllocated = 0;
2080 _numWordsAllocated = 0;
2081 )
2082 }
2083
2084 void CMSCollector::gc_epilogue(bool full) {
2085 // The following locking discipline assumes that we are only called
2086 // when the world is stopped.
2087 assert(SafepointSynchronize::is_at_safepoint(),
2088 "world is stopped assumption");
2089
2090 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2091 // if linear allocation blocks need to be appropriately marked to allow the
2092 // the blocks to be parsable. We also check here whether we need to nudge the
2093 // CMS collector thread to start a new cycle (if it's not already active).
2094 assert( Thread::current()->is_VM_thread()
2095 || ( CMSScavengeBeforeRemark
2096 && Thread::current()->is_ConcurrentGC_thread()),
2097 "Incorrect thread type for epilogue execution");
2098
2099 if (!_between_prologue_and_epilogue) {
2100 // We have already been invoked; this is a gc_epilogue delegation
2101 // from yet another CMS generation that we are responsible for, just
2102 // ignore it since all relevant work has already been done.
2103 return;
2104 }
2105 assert(haveFreelistLocks(), "must have freelist locks");
2106 assert_lock_strong(bitMapLock());
2107
2108 _ct->cld_rem_set()->set_accumulate_modified_oops(false);
2109
2110 _cmsGen->gc_epilogue_work(full);
2111
2112 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2113 // in case sampling was not already enabled, enable it
2114 _start_sampling = true;
2115 }
2116 // reset _eden_chunk_array so sampling starts afresh
2117 _eden_chunk_index = 0;
2118
2119 size_t cms_used = _cmsGen->cmsSpace()->used();
2120
2121 // update performance counters - this uses a special version of
2122 // update_counters() that allows the utilization to be passed as a
2123 // parameter, avoiding multiple calls to used().
2124 //
2125 _cmsGen->update_counters(cms_used);
2126
2127 bitMapLock()->unlock();
2128 releaseFreelistLocks();
2129
2130 if (!CleanChunkPoolAsync) {
2131 Chunk::clean_chunk_pool();
2132 }
2133
2134 set_did_compact(false);
2135 _between_prologue_and_epilogue = false; // ready for next cycle
2136 }
2137
2138 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2139 collector()->gc_epilogue(full);
2140
2141 // When using ParNew, promotion tracking should have already been
2142 // disabled. However, the prologue (which enables promotion
2143 // tracking) and epilogue are called irrespective of the type of
2144 // GC. So they will also be called before and after Full GCs, during
2145 // which promotion tracking will not be explicitly disabled. So,
2146 // it's safer to also disable it here too (to be symmetric with
2147 // enabling it in the prologue).
2148 for (uint i = 0; i < ParallelGCThreads; i++) {
2149 _par_gc_thread_states[i]->promo.stopTrackingPromotions();
2150 }
2151 }
2152
2153 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2154 assert(!incremental_collection_failed(), "Should have been cleared");
2155 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2156 cmsSpace()->gc_epilogue();
2157 // Print stat counters
2158 NOT_PRODUCT(
2159 assert(_numObjectsAllocated == 0, "check");
2160 assert(_numWordsAllocated == 0, "check");
2161 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
2162 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2163 _numObjectsPromoted = 0;
2164 _numWordsPromoted = 0;
2165 )
2166
2167 // Call down the chain in contiguous_available needs the freelistLock
2168 // so print this out before releasing the freeListLock.
2169 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available());
2170 }
2171
2172 #ifndef PRODUCT
2173 bool CMSCollector::have_cms_token() {
2174 Thread* thr = Thread::current();
2175 if (thr->is_VM_thread()) {
2176 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2177 } else if (thr->is_ConcurrentGC_thread()) {
2178 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2179 } else if (thr->is_GC_task_thread()) {
2180 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2181 ParGCRareEvent_lock->owned_by_self();
2182 }
2183 return false;
2184 }
2185
2186 // Check reachability of the given heap address in CMS generation,
2187 // treating all other generations as roots.
2188 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2189 // We could "guarantee" below, rather than assert, but I'll
2190 // leave these as "asserts" so that an adventurous debugger
2191 // could try this in the product build provided some subset of
2192 // the conditions were met, provided they were interested in the
2193 // results and knew that the computation below wouldn't interfere
2194 // with other concurrent computations mutating the structures
2195 // being read or written.
2196 assert(SafepointSynchronize::is_at_safepoint(),
2197 "Else mutations in object graph will make answer suspect");
2198 assert(have_cms_token(), "Should hold cms token");
2199 assert(haveFreelistLocks(), "must hold free list locks");
2200 assert_lock_strong(bitMapLock());
2201
2202 // Clear the marking bit map array before starting, but, just
2203 // for kicks, first report if the given address is already marked
2204 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr),
2205 _markBitMap.isMarked(addr) ? "" : " not");
2206
2207 if (verify_after_remark()) {
2208 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2209 bool result = verification_mark_bm()->isMarked(addr);
2210 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr),
2211 result ? "IS" : "is NOT");
2212 return result;
2213 } else {
2214 tty->print_cr("Could not compute result");
2215 return false;
2216 }
2217 }
2218 #endif
2219
2220 void
2221 CMSCollector::print_on_error(outputStream* st) {
2222 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2223 if (collector != NULL) {
2224 CMSBitMap* bitmap = &collector->_markBitMap;
2225 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2226 bitmap->print_on_error(st, " Bits: ");
2227
2228 st->cr();
2229
2230 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2231 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2232 mut_bitmap->print_on_error(st, " Bits: ");
2233 }
2234 }
2235
2236 ////////////////////////////////////////////////////////
2237 // CMS Verification Support
2238 ////////////////////////////////////////////////////////
2239 // Following the remark phase, the following invariant
2240 // should hold -- each object in the CMS heap which is
2241 // marked in markBitMap() should be marked in the verification_mark_bm().
2242
2243 class VerifyMarkedClosure: public BitMapClosure {
2244 CMSBitMap* _marks;
2245 bool _failed;
2246
2247 public:
2248 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2249
2250 bool do_bit(size_t offset) {
2251 HeapWord* addr = _marks->offsetToHeapWord(offset);
2252 if (!_marks->isMarked(addr)) {
2253 Log(gc, verify) log;
2254 ResourceMark rm;
2255 LogStream ls(log.error());
2256 oop(addr)->print_on(&ls);
2257 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
2258 _failed = true;
2259 }
2260 return true;
2261 }
2262
2263 bool failed() { return _failed; }
2264 };
2265
2266 bool CMSCollector::verify_after_remark() {
2267 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking.");
2268 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2269 static bool init = false;
2270
2271 assert(SafepointSynchronize::is_at_safepoint(),
2272 "Else mutations in object graph will make answer suspect");
2273 assert(have_cms_token(),
2274 "Else there may be mutual interference in use of "
2275 " verification data structures");
2276 assert(_collectorState > Marking && _collectorState <= Sweeping,
2277 "Else marking info checked here may be obsolete");
2278 assert(haveFreelistLocks(), "must hold free list locks");
2279 assert_lock_strong(bitMapLock());
2280
2281
2282 // Allocate marking bit map if not already allocated
2283 if (!init) { // first time
2284 if (!verification_mark_bm()->allocate(_span)) {
2285 return false;
2286 }
2287 init = true;
2288 }
2289
2290 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2291
2292 // Turn off refs discovery -- so we will be tracing through refs.
2293 // This is as intended, because by this time
2294 // GC must already have cleared any refs that need to be cleared,
2295 // and traced those that need to be marked; moreover,
2296 // the marking done here is not going to interfere in any
2297 // way with the marking information used by GC.
2298 NoRefDiscovery no_discovery(ref_processor());
2299
2300 #if COMPILER2_OR_JVMCI
2301 DerivedPointerTableDeactivate dpt_deact;
2302 #endif
2303
2304 // Clear any marks from a previous round
2305 verification_mark_bm()->clear_all();
2306 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2307 verify_work_stacks_empty();
2308
2309 CMSHeap* heap = CMSHeap::heap();
2310 heap->ensure_parsability(false); // fill TLABs, but no need to retire them
2311 // Update the saved marks which may affect the root scans.
2312 heap->save_marks();
2313
2314 if (CMSRemarkVerifyVariant == 1) {
2315 // In this first variant of verification, we complete
2316 // all marking, then check if the new marks-vector is
2317 // a subset of the CMS marks-vector.
2318 verify_after_remark_work_1();
2319 } else {
2320 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2");
2321 // In this second variant of verification, we flag an error
2322 // (i.e. an object reachable in the new marks-vector not reachable
2323 // in the CMS marks-vector) immediately, also indicating the
2324 // identify of an object (A) that references the unmarked object (B) --
2325 // presumably, a mutation to A failed to be picked up by preclean/remark?
2326 verify_after_remark_work_2();
2327 }
2328
2329 return true;
2330 }
2331
2332 void CMSCollector::verify_after_remark_work_1() {
2333 ResourceMark rm;
2334 HandleMark hm;
2335 CMSHeap* heap = CMSHeap::heap();
2336
2337 // Get a clear set of claim bits for the roots processing to work with.
2338 ClassLoaderDataGraph::clear_claimed_marks();
2339
2340 // Mark from roots one level into CMS
2341 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2342 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2343
2344 {
2345 StrongRootsScope srs(1);
2346
2347 heap->cms_process_roots(&srs,
2348 true, // young gen as roots
2349 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2350 should_unload_classes(),
2351 ¬Older,
2352 NULL);
2353 }
2354
2355 // Now mark from the roots
2356 MarkFromRootsClosure markFromRootsClosure(this, _span,
2357 verification_mark_bm(), verification_mark_stack(),
2358 false /* don't yield */, true /* verifying */);
2359 assert(_restart_addr == NULL, "Expected pre-condition");
2360 verification_mark_bm()->iterate(&markFromRootsClosure);
2361 while (_restart_addr != NULL) {
2362 // Deal with stack overflow: by restarting at the indicated
2363 // address.
2364 HeapWord* ra = _restart_addr;
2365 markFromRootsClosure.reset(ra);
2366 _restart_addr = NULL;
2367 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2368 }
2369 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2370 verify_work_stacks_empty();
2371
2372 // Marking completed -- now verify that each bit marked in
2373 // verification_mark_bm() is also marked in markBitMap(); flag all
2374 // errors by printing corresponding objects.
2375 VerifyMarkedClosure vcl(markBitMap());
2376 verification_mark_bm()->iterate(&vcl);
2377 if (vcl.failed()) {
2378 Log(gc, verify) log;
2379 log.error("Failed marking verification after remark");
2380 ResourceMark rm;
2381 LogStream ls(log.error());
2382 heap->print_on(&ls);
2383 fatal("CMS: failed marking verification after remark");
2384 }
2385 }
2386
2387 class VerifyCLDOopsCLDClosure : public CLDClosure {
2388 class VerifyCLDOopsClosure : public OopClosure {
2389 CMSBitMap* _bitmap;
2390 public:
2391 VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2392 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2393 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2394 } _oop_closure;
2395 public:
2396 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2397 void do_cld(ClassLoaderData* cld) {
2398 cld->oops_do(&_oop_closure, false, false);
2399 }
2400 };
2401
2402 void CMSCollector::verify_after_remark_work_2() {
2403 ResourceMark rm;
2404 HandleMark hm;
2405 CMSHeap* heap = CMSHeap::heap();
2406
2407 // Get a clear set of claim bits for the roots processing to work with.
2408 ClassLoaderDataGraph::clear_claimed_marks();
2409
2410 // Mark from roots one level into CMS
2411 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2412 markBitMap());
2413 CLDToOopClosure cld_closure(¬Older, true);
2414
2415 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2416
2417 {
2418 StrongRootsScope srs(1);
2419
2420 heap->cms_process_roots(&srs,
2421 true, // young gen as roots
2422 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2423 should_unload_classes(),
2424 ¬Older,
2425 &cld_closure);
2426 }
2427
2428 // Now mark from the roots
2429 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2430 verification_mark_bm(), markBitMap(), verification_mark_stack());
2431 assert(_restart_addr == NULL, "Expected pre-condition");
2432 verification_mark_bm()->iterate(&markFromRootsClosure);
2433 while (_restart_addr != NULL) {
2434 // Deal with stack overflow: by restarting at the indicated
2435 // address.
2436 HeapWord* ra = _restart_addr;
2437 markFromRootsClosure.reset(ra);
2438 _restart_addr = NULL;
2439 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2440 }
2441 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2442 verify_work_stacks_empty();
2443
2444 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm());
2445 ClassLoaderDataGraph::cld_do(&verify_cld_oops);
2446
2447 // Marking completed -- now verify that each bit marked in
2448 // verification_mark_bm() is also marked in markBitMap(); flag all
2449 // errors by printing corresponding objects.
2450 VerifyMarkedClosure vcl(markBitMap());
2451 verification_mark_bm()->iterate(&vcl);
2452 assert(!vcl.failed(), "Else verification above should not have succeeded");
2453 }
2454
2455 void ConcurrentMarkSweepGeneration::save_marks() {
2456 // delegate to CMS space
2457 cmsSpace()->save_marks();
2458 }
2459
2460 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2461 return cmsSpace()->no_allocs_since_save_marks();
2462 }
2463
2464 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2465 \
2466 void ConcurrentMarkSweepGeneration:: \
2467 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2468 cl->set_generation(this); \
2469 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2470 cl->reset_generation(); \
2471 save_marks(); \
2472 }
2473
2474 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2475
2476 void
2477 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2478 if (freelistLock()->owned_by_self()) {
2479 Generation::oop_iterate(cl);
2480 } else {
2481 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2482 Generation::oop_iterate(cl);
2483 }
2484 }
2485
2486 void
2487 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2488 if (freelistLock()->owned_by_self()) {
2489 Generation::object_iterate(cl);
2490 } else {
2491 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2492 Generation::object_iterate(cl);
2493 }
2494 }
2495
2496 void
2497 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2498 if (freelistLock()->owned_by_self()) {
2499 Generation::safe_object_iterate(cl);
2500 } else {
2501 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2502 Generation::safe_object_iterate(cl);
2503 }
2504 }
2505
2506 void
2507 ConcurrentMarkSweepGeneration::post_compact() {
2508 }
2509
2510 void
2511 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2512 // Fix the linear allocation blocks to look like free blocks.
2513
2514 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2515 // are not called when the heap is verified during universe initialization and
2516 // at vm shutdown.
2517 if (freelistLock()->owned_by_self()) {
2518 cmsSpace()->prepare_for_verify();
2519 } else {
2520 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2521 cmsSpace()->prepare_for_verify();
2522 }
2523 }
2524
2525 void
2526 ConcurrentMarkSweepGeneration::verify() {
2527 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2528 // are not called when the heap is verified during universe initialization and
2529 // at vm shutdown.
2530 if (freelistLock()->owned_by_self()) {
2531 cmsSpace()->verify();
2532 } else {
2533 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2534 cmsSpace()->verify();
2535 }
2536 }
2537
2538 void CMSCollector::verify() {
2539 _cmsGen->verify();
2540 }
2541
2542 #ifndef PRODUCT
2543 bool CMSCollector::overflow_list_is_empty() const {
2544 assert(_num_par_pushes >= 0, "Inconsistency");
2545 if (_overflow_list == NULL) {
2546 assert(_num_par_pushes == 0, "Inconsistency");
2547 }
2548 return _overflow_list == NULL;
2549 }
2550
2551 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2552 // merely consolidate assertion checks that appear to occur together frequently.
2553 void CMSCollector::verify_work_stacks_empty() const {
2554 assert(_markStack.isEmpty(), "Marking stack should be empty");
2555 assert(overflow_list_is_empty(), "Overflow list should be empty");
2556 }
2557
2558 void CMSCollector::verify_overflow_empty() const {
2559 assert(overflow_list_is_empty(), "Overflow list should be empty");
2560 assert(no_preserved_marks(), "No preserved marks");
2561 }
2562 #endif // PRODUCT
2563
2564 // Decide if we want to enable class unloading as part of the
2565 // ensuing concurrent GC cycle. We will collect and
2566 // unload classes if it's the case that:
2567 // (a) class unloading is enabled at the command line, and
2568 // (b) old gen is getting really full
2569 // NOTE: Provided there is no change in the state of the heap between
2570 // calls to this method, it should have idempotent results. Moreover,
2571 // its results should be monotonically increasing (i.e. going from 0 to 1,
2572 // but not 1 to 0) between successive calls between which the heap was
2573 // not collected. For the implementation below, it must thus rely on
2574 // the property that concurrent_cycles_since_last_unload()
2575 // will not decrease unless a collection cycle happened and that
2576 // _cmsGen->is_too_full() are
2577 // themselves also monotonic in that sense. See check_monotonicity()
2578 // below.
2579 void CMSCollector::update_should_unload_classes() {
2580 _should_unload_classes = false;
2581 if (CMSClassUnloadingEnabled) {
2582 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2583 CMSClassUnloadingMaxInterval)
2584 || _cmsGen->is_too_full();
2585 }
2586 }
2587
2588 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2589 bool res = should_concurrent_collect();
2590 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2591 return res;
2592 }
2593
2594 void CMSCollector::setup_cms_unloading_and_verification_state() {
2595 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2596 || VerifyBeforeExit;
2597 const int rso = GenCollectedHeap::SO_AllCodeCache;
2598
2599 // We set the proper root for this CMS cycle here.
2600 if (should_unload_classes()) { // Should unload classes this cycle
2601 remove_root_scanning_option(rso); // Shrink the root set appropriately
2602 set_verifying(should_verify); // Set verification state for this cycle
2603 return; // Nothing else needs to be done at this time
2604 }
2605
2606 // Not unloading classes this cycle
2607 assert(!should_unload_classes(), "Inconsistency!");
2608
2609 // If we are not unloading classes then add SO_AllCodeCache to root
2610 // scanning options.
2611 add_root_scanning_option(rso);
2612
2613 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2614 set_verifying(true);
2615 } else if (verifying() && !should_verify) {
2616 // We were verifying, but some verification flags got disabled.
2617 set_verifying(false);
2618 // Exclude symbols, strings and code cache elements from root scanning to
2619 // reduce IM and RM pauses.
2620 remove_root_scanning_option(rso);
2621 }
2622 }
2623
2624
2625 #ifndef PRODUCT
2626 HeapWord* CMSCollector::block_start(const void* p) const {
2627 const HeapWord* addr = (HeapWord*)p;
2628 if (_span.contains(p)) {
2629 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2630 return _cmsGen->cmsSpace()->block_start(p);
2631 }
2632 }
2633 return NULL;
2634 }
2635 #endif
2636
2637 HeapWord*
2638 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2639 bool tlab,
2640 bool parallel) {
2641 CMSSynchronousYieldRequest yr;
2642 assert(!tlab, "Can't deal with TLAB allocation");
2643 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2644 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2645 if (GCExpandToAllocateDelayMillis > 0) {
2646 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2647 }
2648 return have_lock_and_allocate(word_size, tlab);
2649 }
2650
2651 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2652 size_t bytes,
2653 size_t expand_bytes,
2654 CMSExpansionCause::Cause cause)
2655 {
2656
2657 bool success = expand(bytes, expand_bytes);
2658
2659 // remember why we expanded; this information is used
2660 // by shouldConcurrentCollect() when making decisions on whether to start
2661 // a new CMS cycle.
2662 if (success) {
2663 set_expansion_cause(cause);
2664 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause));
2665 }
2666 }
2667
2668 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2669 HeapWord* res = NULL;
2670 MutexLocker x(ParGCRareEvent_lock);
2671 while (true) {
2672 // Expansion by some other thread might make alloc OK now:
2673 res = ps->lab.alloc(word_sz);
2674 if (res != NULL) return res;
2675 // If there's not enough expansion space available, give up.
2676 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2677 return NULL;
2678 }
2679 // Otherwise, we try expansion.
2680 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2681 // Now go around the loop and try alloc again;
2682 // A competing par_promote might beat us to the expansion space,
2683 // so we may go around the loop again if promotion fails again.
2684 if (GCExpandToAllocateDelayMillis > 0) {
2685 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2686 }
2687 }
2688 }
2689
2690
2691 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2692 PromotionInfo* promo) {
2693 MutexLocker x(ParGCRareEvent_lock);
2694 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2695 while (true) {
2696 // Expansion by some other thread might make alloc OK now:
2697 if (promo->ensure_spooling_space()) {
2698 assert(promo->has_spooling_space(),
2699 "Post-condition of successful ensure_spooling_space()");
2700 return true;
2701 }
2702 // If there's not enough expansion space available, give up.
2703 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2704 return false;
2705 }
2706 // Otherwise, we try expansion.
2707 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2708 // Now go around the loop and try alloc again;
2709 // A competing allocation might beat us to the expansion space,
2710 // so we may go around the loop again if allocation fails again.
2711 if (GCExpandToAllocateDelayMillis > 0) {
2712 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2713 }
2714 }
2715 }
2716
2717 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2718 // Only shrink if a compaction was done so that all the free space
2719 // in the generation is in a contiguous block at the end.
2720 if (did_compact()) {
2721 CardGeneration::shrink(bytes);
2722 }
2723 }
2724
2725 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2726 assert_locked_or_safepoint(Heap_lock);
2727 }
2728
2729 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2730 assert_locked_or_safepoint(Heap_lock);
2731 assert_lock_strong(freelistLock());
2732 log_trace(gc)("Shrinking of CMS not yet implemented");
2733 return;
2734 }
2735
2736
2737 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2738 // phases.
2739 class CMSPhaseAccounting: public StackObj {
2740 public:
2741 CMSPhaseAccounting(CMSCollector *collector,
2742 const char *title);
2743 ~CMSPhaseAccounting();
2744
2745 private:
2746 CMSCollector *_collector;
2747 const char *_title;
2748 GCTraceConcTime(Info, gc) _trace_time;
2749
2750 public:
2751 // Not MT-safe; so do not pass around these StackObj's
2752 // where they may be accessed by other threads.
2753 double wallclock_millis() {
2754 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time());
2755 }
2756 };
2757
2758 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2759 const char *title) :
2760 _collector(collector), _title(title), _trace_time(title) {
2761
2762 _collector->resetYields();
2763 _collector->resetTimer();
2764 _collector->startTimer();
2765 _collector->gc_timer_cm()->register_gc_concurrent_start(title);
2766 }
2767
2768 CMSPhaseAccounting::~CMSPhaseAccounting() {
2769 _collector->gc_timer_cm()->register_gc_concurrent_end();
2770 _collector->stopTimer();
2771 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks()));
2772 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields());
2773 }
2774
2775 // CMS work
2776
2777 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2778 class CMSParMarkTask : public AbstractGangTask {
2779 protected:
2780 CMSCollector* _collector;
2781 uint _n_workers;
2782 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2783 AbstractGangTask(name),
2784 _collector(collector),
2785 _n_workers(n_workers) {}
2786 // Work method in support of parallel rescan ... of young gen spaces
2787 void do_young_space_rescan(OopsInGenClosure* cl,
2788 ContiguousSpace* space,
2789 HeapWord** chunk_array, size_t chunk_top);
2790 void work_on_young_gen_roots(OopsInGenClosure* cl);
2791 };
2792
2793 // Parallel initial mark task
2794 class CMSParInitialMarkTask: public CMSParMarkTask {
2795 StrongRootsScope* _strong_roots_scope;
2796 public:
2797 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) :
2798 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers),
2799 _strong_roots_scope(strong_roots_scope) {}
2800 void work(uint worker_id);
2801 };
2802
2803 // Checkpoint the roots into this generation from outside
2804 // this generation. [Note this initial checkpoint need only
2805 // be approximate -- we'll do a catch up phase subsequently.]
2806 void CMSCollector::checkpointRootsInitial() {
2807 assert(_collectorState == InitialMarking, "Wrong collector state");
2808 check_correct_thread_executing();
2809 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
2810
2811 save_heap_summary();
2812 report_heap_summary(GCWhen::BeforeGC);
2813
2814 ReferenceProcessor* rp = ref_processor();
2815 assert(_restart_addr == NULL, "Control point invariant");
2816 {
2817 // acquire locks for subsequent manipulations
2818 MutexLockerEx x(bitMapLock(),
2819 Mutex::_no_safepoint_check_flag);
2820 checkpointRootsInitialWork();
2821 // enable ("weak") refs discovery
2822 rp->enable_discovery();
2823 _collectorState = Marking;
2824 }
2825 }
2826
2827 void CMSCollector::checkpointRootsInitialWork() {
2828 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2829 assert(_collectorState == InitialMarking, "just checking");
2830
2831 // Already have locks.
2832 assert_lock_strong(bitMapLock());
2833 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2834
2835 // Setup the verification and class unloading state for this
2836 // CMS collection cycle.
2837 setup_cms_unloading_and_verification_state();
2838
2839 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm);
2840
2841 // Reset all the PLAB chunk arrays if necessary.
2842 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2843 reset_survivor_plab_arrays();
2844 }
2845
2846 ResourceMark rm;
2847 HandleMark hm;
2848
2849 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2850 CMSHeap* heap = CMSHeap::heap();
2851
2852 verify_work_stacks_empty();
2853 verify_overflow_empty();
2854
2855 heap->ensure_parsability(false); // fill TLABs, but no need to retire them
2856 // Update the saved marks which may affect the root scans.
2857 heap->save_marks();
2858
2859 // weak reference processing has not started yet.
2860 ref_processor()->set_enqueuing_is_done(false);
2861
2862 // Need to remember all newly created CLDs,
2863 // so that we can guarantee that the remark finds them.
2864 ClassLoaderDataGraph::remember_new_clds(true);
2865
2866 // Whenever a CLD is found, it will be claimed before proceeding to mark
2867 // the klasses. The claimed marks need to be cleared before marking starts.
2868 ClassLoaderDataGraph::clear_claimed_marks();
2869
2870 print_eden_and_survivor_chunk_arrays();
2871
2872 {
2873 #if COMPILER2_OR_JVMCI
2874 DerivedPointerTableDeactivate dpt_deact;
2875 #endif
2876 if (CMSParallelInitialMarkEnabled) {
2877 // The parallel version.
2878 WorkGang* workers = heap->workers();
2879 assert(workers != NULL, "Need parallel worker threads.");
2880 uint n_workers = workers->active_workers();
2881
2882 StrongRootsScope srs(n_workers);
2883
2884 CMSParInitialMarkTask tsk(this, &srs, n_workers);
2885 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
2886 // If the total workers is greater than 1, then multiple workers
2887 // may be used at some time and the initialization has been set
2888 // such that the single threaded path cannot be used.
2889 if (workers->total_workers() > 1) {
2890 workers->run_task(&tsk);
2891 } else {
2892 tsk.work(0);
2893 }
2894 } else {
2895 // The serial version.
2896 CLDToOopClosure cld_closure(¬Older, true);
2897 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2898
2899 StrongRootsScope srs(1);
2900
2901 heap->cms_process_roots(&srs,
2902 true, // young gen as roots
2903 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2904 should_unload_classes(),
2905 ¬Older,
2906 &cld_closure);
2907 }
2908 }
2909
2910 // Clear mod-union table; it will be dirtied in the prologue of
2911 // CMS generation per each young generation collection.
2912
2913 assert(_modUnionTable.isAllClear(),
2914 "Was cleared in most recent final checkpoint phase"
2915 " or no bits are set in the gc_prologue before the start of the next "
2916 "subsequent marking phase.");
2917
2918 assert(_ct->cld_rem_set()->mod_union_is_clear(), "Must be");
2919
2920 // Save the end of the used_region of the constituent generations
2921 // to be used to limit the extent of sweep in each generation.
2922 save_sweep_limits();
2923 verify_overflow_empty();
2924 }
2925
2926 bool CMSCollector::markFromRoots() {
2927 // we might be tempted to assert that:
2928 // assert(!SafepointSynchronize::is_at_safepoint(),
2929 // "inconsistent argument?");
2930 // However that wouldn't be right, because it's possible that
2931 // a safepoint is indeed in progress as a young generation
2932 // stop-the-world GC happens even as we mark in this generation.
2933 assert(_collectorState == Marking, "inconsistent state?");
2934 check_correct_thread_executing();
2935 verify_overflow_empty();
2936
2937 // Weak ref discovery note: We may be discovering weak
2938 // refs in this generation concurrent (but interleaved) with
2939 // weak ref discovery by the young generation collector.
2940
2941 CMSTokenSyncWithLocks ts(true, bitMapLock());
2942 GCTraceCPUTime tcpu;
2943 CMSPhaseAccounting pa(this, "Concurrent Mark");
2944 bool res = markFromRootsWork();
2945 if (res) {
2946 _collectorState = Precleaning;
2947 } else { // We failed and a foreground collection wants to take over
2948 assert(_foregroundGCIsActive, "internal state inconsistency");
2949 assert(_restart_addr == NULL, "foreground will restart from scratch");
2950 log_debug(gc)("bailing out to foreground collection");
2951 }
2952 verify_overflow_empty();
2953 return res;
2954 }
2955
2956 bool CMSCollector::markFromRootsWork() {
2957 // iterate over marked bits in bit map, doing a full scan and mark
2958 // from these roots using the following algorithm:
2959 // . if oop is to the right of the current scan pointer,
2960 // mark corresponding bit (we'll process it later)
2961 // . else (oop is to left of current scan pointer)
2962 // push oop on marking stack
2963 // . drain the marking stack
2964
2965 // Note that when we do a marking step we need to hold the
2966 // bit map lock -- recall that direct allocation (by mutators)
2967 // and promotion (by the young generation collector) is also
2968 // marking the bit map. [the so-called allocate live policy.]
2969 // Because the implementation of bit map marking is not
2970 // robust wrt simultaneous marking of bits in the same word,
2971 // we need to make sure that there is no such interference
2972 // between concurrent such updates.
2973
2974 // already have locks
2975 assert_lock_strong(bitMapLock());
2976
2977 verify_work_stacks_empty();
2978 verify_overflow_empty();
2979 bool result = false;
2980 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
2981 result = do_marking_mt();
2982 } else {
2983 result = do_marking_st();
2984 }
2985 return result;
2986 }
2987
2988 // Forward decl
2989 class CMSConcMarkingTask;
2990
2991 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
2992 CMSCollector* _collector;
2993 CMSConcMarkingTask* _task;
2994 public:
2995 virtual void yield();
2996
2997 // "n_threads" is the number of threads to be terminated.
2998 // "queue_set" is a set of work queues of other threads.
2999 // "collector" is the CMS collector associated with this task terminator.
3000 // "yield" indicates whether we need the gang as a whole to yield.
3001 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3002 ParallelTaskTerminator(n_threads, queue_set),
3003 _collector(collector) { }
3004
3005 void set_task(CMSConcMarkingTask* task) {
3006 _task = task;
3007 }
3008 };
3009
3010 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3011 CMSConcMarkingTask* _task;
3012 public:
3013 bool should_exit_termination();
3014 void set_task(CMSConcMarkingTask* task) {
3015 _task = task;
3016 }
3017 };
3018
3019 // MT Concurrent Marking Task
3020 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3021 CMSCollector* _collector;
3022 uint _n_workers; // requested/desired # workers
3023 bool _result;
3024 CompactibleFreeListSpace* _cms_space;
3025 char _pad_front[64]; // padding to ...
3026 HeapWord* volatile _global_finger; // ... avoid sharing cache line
3027 char _pad_back[64];
3028 HeapWord* _restart_addr;
3029
3030 // Exposed here for yielding support
3031 Mutex* const _bit_map_lock;
3032
3033 // The per thread work queues, available here for stealing
3034 OopTaskQueueSet* _task_queues;
3035
3036 // Termination (and yielding) support
3037 CMSConcMarkingTerminator _term;
3038 CMSConcMarkingTerminatorTerminator _term_term;
3039
3040 public:
3041 CMSConcMarkingTask(CMSCollector* collector,
3042 CompactibleFreeListSpace* cms_space,
3043 YieldingFlexibleWorkGang* workers,
3044 OopTaskQueueSet* task_queues):
3045 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3046 _collector(collector),
3047 _cms_space(cms_space),
3048 _n_workers(0), _result(true),
3049 _task_queues(task_queues),
3050 _term(_n_workers, task_queues, _collector),
3051 _bit_map_lock(collector->bitMapLock())
3052 {
3053 _requested_size = _n_workers;
3054 _term.set_task(this);
3055 _term_term.set_task(this);
3056 _restart_addr = _global_finger = _cms_space->bottom();
3057 }
3058
3059
3060 OopTaskQueueSet* task_queues() { return _task_queues; }
3061
3062 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3063
3064 HeapWord* volatile* global_finger_addr() { return &_global_finger; }
3065
3066 CMSConcMarkingTerminator* terminator() { return &_term; }
3067
3068 virtual void set_for_termination(uint active_workers) {
3069 terminator()->reset_for_reuse(active_workers);
3070 }
3071
3072 void work(uint worker_id);
3073 bool should_yield() {
3074 return ConcurrentMarkSweepThread::should_yield()
3075 && !_collector->foregroundGCIsActive();
3076 }
3077
3078 virtual void coordinator_yield(); // stuff done by coordinator
3079 bool result() { return _result; }
3080
3081 void reset(HeapWord* ra) {
3082 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3083 _restart_addr = _global_finger = ra;
3084 _term.reset_for_reuse();
3085 }
3086
3087 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3088 OopTaskQueue* work_q);
3089
3090 private:
3091 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3092 void do_work_steal(int i);
3093 void bump_global_finger(HeapWord* f);
3094 };
3095
3096 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3097 assert(_task != NULL, "Error");
3098 return _task->yielding();
3099 // Note that we do not need the disjunct || _task->should_yield() above
3100 // because we want terminating threads to yield only if the task
3101 // is already in the midst of yielding, which happens only after at least one
3102 // thread has yielded.
3103 }
3104
3105 void CMSConcMarkingTerminator::yield() {
3106 if (_task->should_yield()) {
3107 _task->yield();
3108 } else {
3109 ParallelTaskTerminator::yield();
3110 }
3111 }
3112
3113 ////////////////////////////////////////////////////////////////
3114 // Concurrent Marking Algorithm Sketch
3115 ////////////////////////////////////////////////////////////////
3116 // Until all tasks exhausted (both spaces):
3117 // -- claim next available chunk
3118 // -- bump global finger via CAS
3119 // -- find first object that starts in this chunk
3120 // and start scanning bitmap from that position
3121 // -- scan marked objects for oops
3122 // -- CAS-mark target, and if successful:
3123 // . if target oop is above global finger (volatile read)
3124 // nothing to do
3125 // . if target oop is in chunk and above local finger
3126 // then nothing to do
3127 // . else push on work-queue
3128 // -- Deal with possible overflow issues:
3129 // . local work-queue overflow causes stuff to be pushed on
3130 // global (common) overflow queue
3131 // . always first empty local work queue
3132 // . then get a batch of oops from global work queue if any
3133 // . then do work stealing
3134 // -- When all tasks claimed (both spaces)
3135 // and local work queue empty,
3136 // then in a loop do:
3137 // . check global overflow stack; steal a batch of oops and trace
3138 // . try to steal from other threads oif GOS is empty
3139 // . if neither is available, offer termination
3140 // -- Terminate and return result
3141 //
3142 void CMSConcMarkingTask::work(uint worker_id) {
3143 elapsedTimer _timer;
3144 ResourceMark rm;
3145 HandleMark hm;
3146
3147 DEBUG_ONLY(_collector->verify_overflow_empty();)
3148
3149 // Before we begin work, our work queue should be empty
3150 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3151 // Scan the bitmap covering _cms_space, tracing through grey objects.
3152 _timer.start();
3153 do_scan_and_mark(worker_id, _cms_space);
3154 _timer.stop();
3155 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3156
3157 // ... do work stealing
3158 _timer.reset();
3159 _timer.start();
3160 do_work_steal(worker_id);
3161 _timer.stop();
3162 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3163 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3164 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3165 // Note that under the current task protocol, the
3166 // following assertion is true even of the spaces
3167 // expanded since the completion of the concurrent
3168 // marking. XXX This will likely change under a strict
3169 // ABORT semantics.
3170 // After perm removal the comparison was changed to
3171 // greater than or equal to from strictly greater than.
3172 // Before perm removal the highest address sweep would
3173 // have been at the end of perm gen but now is at the
3174 // end of the tenured gen.
3175 assert(_global_finger >= _cms_space->end(),
3176 "All tasks have been completed");
3177 DEBUG_ONLY(_collector->verify_overflow_empty();)
3178 }
3179
3180 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3181 HeapWord* read = _global_finger;
3182 HeapWord* cur = read;
3183 while (f > read) {
3184 cur = read;
3185 read = Atomic::cmpxchg(f, &_global_finger, cur);
3186 if (cur == read) {
3187 // our cas succeeded
3188 assert(_global_finger >= f, "protocol consistency");
3189 break;
3190 }
3191 }
3192 }
3193
3194 // This is really inefficient, and should be redone by
3195 // using (not yet available) block-read and -write interfaces to the
3196 // stack and the work_queue. XXX FIX ME !!!
3197 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3198 OopTaskQueue* work_q) {
3199 // Fast lock-free check
3200 if (ovflw_stk->length() == 0) {
3201 return false;
3202 }
3203 assert(work_q->size() == 0, "Shouldn't steal");
3204 MutexLockerEx ml(ovflw_stk->par_lock(),
3205 Mutex::_no_safepoint_check_flag);
3206 // Grab up to 1/4 the size of the work queue
3207 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3208 (size_t)ParGCDesiredObjsFromOverflowList);
3209 num = MIN2(num, ovflw_stk->length());
3210 for (int i = (int) num; i > 0; i--) {
3211 oop cur = ovflw_stk->pop();
3212 assert(cur != NULL, "Counted wrong?");
3213 work_q->push(cur);
3214 }
3215 return num > 0;
3216 }
3217
3218 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3219 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3220 int n_tasks = pst->n_tasks();
3221 // We allow that there may be no tasks to do here because
3222 // we are restarting after a stack overflow.
3223 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3224 uint nth_task = 0;
3225
3226 HeapWord* aligned_start = sp->bottom();
3227 if (sp->used_region().contains(_restart_addr)) {
3228 // Align down to a card boundary for the start of 0th task
3229 // for this space.
3230 aligned_start = align_down(_restart_addr, CardTableModRefBS::card_size);
3231 }
3232
3233 size_t chunk_size = sp->marking_task_size();
3234 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3235 // Having claimed the nth task in this space,
3236 // compute the chunk that it corresponds to:
3237 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3238 aligned_start + (nth_task+1)*chunk_size);
3239 // Try and bump the global finger via a CAS;
3240 // note that we need to do the global finger bump
3241 // _before_ taking the intersection below, because
3242 // the task corresponding to that region will be
3243 // deemed done even if the used_region() expands
3244 // because of allocation -- as it almost certainly will
3245 // during start-up while the threads yield in the
3246 // closure below.
3247 HeapWord* finger = span.end();
3248 bump_global_finger(finger); // atomically
3249 // There are null tasks here corresponding to chunks
3250 // beyond the "top" address of the space.
3251 span = span.intersection(sp->used_region());
3252 if (!span.is_empty()) { // Non-null task
3253 HeapWord* prev_obj;
3254 assert(!span.contains(_restart_addr) || nth_task == 0,
3255 "Inconsistency");
3256 if (nth_task == 0) {
3257 // For the 0th task, we'll not need to compute a block_start.
3258 if (span.contains(_restart_addr)) {
3259 // In the case of a restart because of stack overflow,
3260 // we might additionally skip a chunk prefix.
3261 prev_obj = _restart_addr;
3262 } else {
3263 prev_obj = span.start();
3264 }
3265 } else {
3266 // We want to skip the first object because
3267 // the protocol is to scan any object in its entirety
3268 // that _starts_ in this span; a fortiori, any
3269 // object starting in an earlier span is scanned
3270 // as part of an earlier claimed task.
3271 // Below we use the "careful" version of block_start
3272 // so we do not try to navigate uninitialized objects.
3273 prev_obj = sp->block_start_careful(span.start());
3274 // Below we use a variant of block_size that uses the
3275 // Printezis bits to avoid waiting for allocated
3276 // objects to become initialized/parsable.
3277 while (prev_obj < span.start()) {
3278 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3279 if (sz > 0) {
3280 prev_obj += sz;
3281 } else {
3282 // In this case we may end up doing a bit of redundant
3283 // scanning, but that appears unavoidable, short of
3284 // locking the free list locks; see bug 6324141.
3285 break;
3286 }
3287 }
3288 }
3289 if (prev_obj < span.end()) {
3290 MemRegion my_span = MemRegion(prev_obj, span.end());
3291 // Do the marking work within a non-empty span --
3292 // the last argument to the constructor indicates whether the
3293 // iteration should be incremental with periodic yields.
3294 ParMarkFromRootsClosure cl(this, _collector, my_span,
3295 &_collector->_markBitMap,
3296 work_queue(i),
3297 &_collector->_markStack);
3298 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3299 } // else nothing to do for this task
3300 } // else nothing to do for this task
3301 }
3302 // We'd be tempted to assert here that since there are no
3303 // more tasks left to claim in this space, the global_finger
3304 // must exceed space->top() and a fortiori space->end(). However,
3305 // that would not quite be correct because the bumping of
3306 // global_finger occurs strictly after the claiming of a task,
3307 // so by the time we reach here the global finger may not yet
3308 // have been bumped up by the thread that claimed the last
3309 // task.
3310 pst->all_tasks_completed();
3311 }
3312
3313 class ParConcMarkingClosure: public MetadataAwareOopClosure {
3314 private:
3315 CMSCollector* _collector;
3316 CMSConcMarkingTask* _task;
3317 MemRegion _span;
3318 CMSBitMap* _bit_map;
3319 CMSMarkStack* _overflow_stack;
3320 OopTaskQueue* _work_queue;
3321 protected:
3322 DO_OOP_WORK_DEFN
3323 public:
3324 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3325 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3326 MetadataAwareOopClosure(collector->ref_processor()),
3327 _collector(collector),
3328 _task(task),
3329 _span(collector->_span),
3330 _work_queue(work_queue),
3331 _bit_map(bit_map),
3332 _overflow_stack(overflow_stack)
3333 { }
3334 virtual void do_oop(oop* p);
3335 virtual void do_oop(narrowOop* p);
3336
3337 void trim_queue(size_t max);
3338 void handle_stack_overflow(HeapWord* lost);
3339 void do_yield_check() {
3340 if (_task->should_yield()) {
3341 _task->yield();
3342 }
3343 }
3344 };
3345
3346 DO_OOP_WORK_IMPL(ParConcMarkingClosure)
3347
3348 // Grey object scanning during work stealing phase --
3349 // the salient assumption here is that any references
3350 // that are in these stolen objects being scanned must
3351 // already have been initialized (else they would not have
3352 // been published), so we do not need to check for
3353 // uninitialized objects before pushing here.
3354 void ParConcMarkingClosure::do_oop(oop obj) {
3355 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
3356 HeapWord* addr = (HeapWord*)obj;
3357 // Check if oop points into the CMS generation
3358 // and is not marked
3359 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3360 // a white object ...
3361 // If we manage to "claim" the object, by being the
3362 // first thread to mark it, then we push it on our
3363 // marking stack
3364 if (_bit_map->par_mark(addr)) { // ... now grey
3365 // push on work queue (grey set)
3366 bool simulate_overflow = false;
3367 NOT_PRODUCT(
3368 if (CMSMarkStackOverflowALot &&
3369 _collector->simulate_overflow()) {
3370 // simulate a stack overflow
3371 simulate_overflow = true;
3372 }
3373 )
3374 if (simulate_overflow ||
3375 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3376 // stack overflow
3377 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
3378 // We cannot assert that the overflow stack is full because
3379 // it may have been emptied since.
3380 assert(simulate_overflow ||
3381 _work_queue->size() == _work_queue->max_elems(),
3382 "Else push should have succeeded");
3383 handle_stack_overflow(addr);
3384 }
3385 } // Else, some other thread got there first
3386 do_yield_check();
3387 }
3388 }
3389
3390 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); }
3391 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); }
3392
3393 void ParConcMarkingClosure::trim_queue(size_t max) {
3394 while (_work_queue->size() > max) {
3395 oop new_oop;
3396 if (_work_queue->pop_local(new_oop)) {
3397 assert(oopDesc::is_oop(new_oop), "Should be an oop");
3398 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3399 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3400 new_oop->oop_iterate(this); // do_oop() above
3401 do_yield_check();
3402 }
3403 }
3404 }
3405
3406 // Upon stack overflow, we discard (part of) the stack,
3407 // remembering the least address amongst those discarded
3408 // in CMSCollector's _restart_address.
3409 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3410 // We need to do this under a mutex to prevent other
3411 // workers from interfering with the work done below.
3412 MutexLockerEx ml(_overflow_stack->par_lock(),
3413 Mutex::_no_safepoint_check_flag);
3414 // Remember the least grey address discarded
3415 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3416 _collector->lower_restart_addr(ra);
3417 _overflow_stack->reset(); // discard stack contents
3418 _overflow_stack->expand(); // expand the stack if possible
3419 }
3420
3421
3422 void CMSConcMarkingTask::do_work_steal(int i) {
3423 OopTaskQueue* work_q = work_queue(i);
3424 oop obj_to_scan;
3425 CMSBitMap* bm = &(_collector->_markBitMap);
3426 CMSMarkStack* ovflw = &(_collector->_markStack);
3427 int* seed = _collector->hash_seed(i);
3428 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3429 while (true) {
3430 cl.trim_queue(0);
3431 assert(work_q->size() == 0, "Should have been emptied above");
3432 if (get_work_from_overflow_stack(ovflw, work_q)) {
3433 // Can't assert below because the work obtained from the
3434 // overflow stack may already have been stolen from us.
3435 // assert(work_q->size() > 0, "Work from overflow stack");
3436 continue;
3437 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3438 assert(oopDesc::is_oop(obj_to_scan), "Should be an oop");
3439 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3440 obj_to_scan->oop_iterate(&cl);
3441 } else if (terminator()->offer_termination(&_term_term)) {
3442 assert(work_q->size() == 0, "Impossible!");
3443 break;
3444 } else if (yielding() || should_yield()) {
3445 yield();
3446 }
3447 }
3448 }
3449
3450 // This is run by the CMS (coordinator) thread.
3451 void CMSConcMarkingTask::coordinator_yield() {
3452 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3453 "CMS thread should hold CMS token");
3454 // First give up the locks, then yield, then re-lock
3455 // We should probably use a constructor/destructor idiom to
3456 // do this unlock/lock or modify the MutexUnlocker class to
3457 // serve our purpose. XXX
3458 assert_lock_strong(_bit_map_lock);
3459 _bit_map_lock->unlock();
3460 ConcurrentMarkSweepThread::desynchronize(true);
3461 _collector->stopTimer();
3462 _collector->incrementYields();
3463
3464 // It is possible for whichever thread initiated the yield request
3465 // not to get a chance to wake up and take the bitmap lock between
3466 // this thread releasing it and reacquiring it. So, while the
3467 // should_yield() flag is on, let's sleep for a bit to give the
3468 // other thread a chance to wake up. The limit imposed on the number
3469 // of iterations is defensive, to avoid any unforseen circumstances
3470 // putting us into an infinite loop. Since it's always been this
3471 // (coordinator_yield()) method that was observed to cause the
3472 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3473 // which is by default non-zero. For the other seven methods that
3474 // also perform the yield operation, as are using a different
3475 // parameter (CMSYieldSleepCount) which is by default zero. This way we
3476 // can enable the sleeping for those methods too, if necessary.
3477 // See 6442774.
3478 //
3479 // We really need to reconsider the synchronization between the GC
3480 // thread and the yield-requesting threads in the future and we
3481 // should really use wait/notify, which is the recommended
3482 // way of doing this type of interaction. Additionally, we should
3483 // consolidate the eight methods that do the yield operation and they
3484 // are almost identical into one for better maintainability and
3485 // readability. See 6445193.
3486 //
3487 // Tony 2006.06.29
3488 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3489 ConcurrentMarkSweepThread::should_yield() &&
3490 !CMSCollector::foregroundGCIsActive(); ++i) {
3491 os::sleep(Thread::current(), 1, false);
3492 }
3493
3494 ConcurrentMarkSweepThread::synchronize(true);
3495 _bit_map_lock->lock_without_safepoint_check();
3496 _collector->startTimer();
3497 }
3498
3499 bool CMSCollector::do_marking_mt() {
3500 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3501 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3502 conc_workers()->active_workers(),
3503 Threads::number_of_non_daemon_threads());
3504 num_workers = conc_workers()->update_active_workers(num_workers);
3505 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers());
3506
3507 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
3508
3509 CMSConcMarkingTask tsk(this,
3510 cms_space,
3511 conc_workers(),
3512 task_queues());
3513
3514 // Since the actual number of workers we get may be different
3515 // from the number we requested above, do we need to do anything different
3516 // below? In particular, may be we need to subclass the SequantialSubTasksDone
3517 // class?? XXX
3518 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3519
3520 // Refs discovery is already non-atomic.
3521 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3522 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3523 conc_workers()->start_task(&tsk);
3524 while (tsk.yielded()) {
3525 tsk.coordinator_yield();
3526 conc_workers()->continue_task(&tsk);
3527 }
3528 // If the task was aborted, _restart_addr will be non-NULL
3529 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3530 while (_restart_addr != NULL) {
3531 // XXX For now we do not make use of ABORTED state and have not
3532 // yet implemented the right abort semantics (even in the original
3533 // single-threaded CMS case). That needs some more investigation
3534 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3535 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3536 // If _restart_addr is non-NULL, a marking stack overflow
3537 // occurred; we need to do a fresh marking iteration from the
3538 // indicated restart address.
3539 if (_foregroundGCIsActive) {
3540 // We may be running into repeated stack overflows, having
3541 // reached the limit of the stack size, while making very
3542 // slow forward progress. It may be best to bail out and
3543 // let the foreground collector do its job.
3544 // Clear _restart_addr, so that foreground GC
3545 // works from scratch. This avoids the headache of
3546 // a "rescan" which would otherwise be needed because
3547 // of the dirty mod union table & card table.
3548 _restart_addr = NULL;
3549 return false;
3550 }
3551 // Adjust the task to restart from _restart_addr
3552 tsk.reset(_restart_addr);
3553 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3554 _restart_addr);
3555 _restart_addr = NULL;
3556 // Get the workers going again
3557 conc_workers()->start_task(&tsk);
3558 while (tsk.yielded()) {
3559 tsk.coordinator_yield();
3560 conc_workers()->continue_task(&tsk);
3561 }
3562 }
3563 assert(tsk.completed(), "Inconsistency");
3564 assert(tsk.result() == true, "Inconsistency");
3565 return true;
3566 }
3567
3568 bool CMSCollector::do_marking_st() {
3569 ResourceMark rm;
3570 HandleMark hm;
3571
3572 // Temporarily make refs discovery single threaded (non-MT)
3573 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3574 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3575 &_markStack, CMSYield);
3576 // the last argument to iterate indicates whether the iteration
3577 // should be incremental with periodic yields.
3578 _markBitMap.iterate(&markFromRootsClosure);
3579 // If _restart_addr is non-NULL, a marking stack overflow
3580 // occurred; we need to do a fresh iteration from the
3581 // indicated restart address.
3582 while (_restart_addr != NULL) {
3583 if (_foregroundGCIsActive) {
3584 // We may be running into repeated stack overflows, having
3585 // reached the limit of the stack size, while making very
3586 // slow forward progress. It may be best to bail out and
3587 // let the foreground collector do its job.
3588 // Clear _restart_addr, so that foreground GC
3589 // works from scratch. This avoids the headache of
3590 // a "rescan" which would otherwise be needed because
3591 // of the dirty mod union table & card table.
3592 _restart_addr = NULL;
3593 return false; // indicating failure to complete marking
3594 }
3595 // Deal with stack overflow:
3596 // we restart marking from _restart_addr
3597 HeapWord* ra = _restart_addr;
3598 markFromRootsClosure.reset(ra);
3599 _restart_addr = NULL;
3600 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3601 }
3602 return true;
3603 }
3604
3605 void CMSCollector::preclean() {
3606 check_correct_thread_executing();
3607 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3608 verify_work_stacks_empty();
3609 verify_overflow_empty();
3610 _abort_preclean = false;
3611 if (CMSPrecleaningEnabled) {
3612 if (!CMSEdenChunksRecordAlways) {
3613 _eden_chunk_index = 0;
3614 }
3615 size_t used = get_eden_used();
3616 size_t capacity = get_eden_capacity();
3617 // Don't start sampling unless we will get sufficiently
3618 // many samples.
3619 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100)
3620 * CMSScheduleRemarkEdenPenetration)) {
3621 _start_sampling = true;
3622 } else {
3623 _start_sampling = false;
3624 }
3625 GCTraceCPUTime tcpu;
3626 CMSPhaseAccounting pa(this, "Concurrent Preclean");
3627 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3628 }
3629 CMSTokenSync x(true); // is cms thread
3630 if (CMSPrecleaningEnabled) {
3631 sample_eden();
3632 _collectorState = AbortablePreclean;
3633 } else {
3634 _collectorState = FinalMarking;
3635 }
3636 verify_work_stacks_empty();
3637 verify_overflow_empty();
3638 }
3639
3640 // Try and schedule the remark such that young gen
3641 // occupancy is CMSScheduleRemarkEdenPenetration %.
3642 void CMSCollector::abortable_preclean() {
3643 check_correct_thread_executing();
3644 assert(CMSPrecleaningEnabled, "Inconsistent control state");
3645 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3646
3647 // If Eden's current occupancy is below this threshold,
3648 // immediately schedule the remark; else preclean
3649 // past the next scavenge in an effort to
3650 // schedule the pause as described above. By choosing
3651 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3652 // we will never do an actual abortable preclean cycle.
3653 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3654 GCTraceCPUTime tcpu;
3655 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean");
3656 // We need more smarts in the abortable preclean
3657 // loop below to deal with cases where allocation
3658 // in young gen is very very slow, and our precleaning
3659 // is running a losing race against a horde of
3660 // mutators intent on flooding us with CMS updates
3661 // (dirty cards).
3662 // One, admittedly dumb, strategy is to give up
3663 // after a certain number of abortable precleaning loops
3664 // or after a certain maximum time. We want to make
3665 // this smarter in the next iteration.
3666 // XXX FIX ME!!! YSR
3667 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3668 while (!(should_abort_preclean() ||
3669 ConcurrentMarkSweepThread::cmst()->should_terminate())) {
3670 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3671 cumworkdone += workdone;
3672 loops++;
3673 // Voluntarily terminate abortable preclean phase if we have
3674 // been at it for too long.
3675 if ((CMSMaxAbortablePrecleanLoops != 0) &&
3676 loops >= CMSMaxAbortablePrecleanLoops) {
3677 log_debug(gc)(" CMS: abort preclean due to loops ");
3678 break;
3679 }
3680 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3681 log_debug(gc)(" CMS: abort preclean due to time ");
3682 break;
3683 }
3684 // If we are doing little work each iteration, we should
3685 // take a short break.
3686 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3687 // Sleep for some time, waiting for work to accumulate
3688 stopTimer();
3689 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3690 startTimer();
3691 waited++;
3692 }
3693 }
3694 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3695 loops, waited, cumworkdone);
3696 }
3697 CMSTokenSync x(true); // is cms thread
3698 if (_collectorState != Idling) {
3699 assert(_collectorState == AbortablePreclean,
3700 "Spontaneous state transition?");
3701 _collectorState = FinalMarking;
3702 } // Else, a foreground collection completed this CMS cycle.
3703 return;
3704 }
3705
3706 // Respond to an Eden sampling opportunity
3707 void CMSCollector::sample_eden() {
3708 // Make sure a young gc cannot sneak in between our
3709 // reading and recording of a sample.
3710 assert(Thread::current()->is_ConcurrentGC_thread(),
3711 "Only the cms thread may collect Eden samples");
3712 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3713 "Should collect samples while holding CMS token");
3714 if (!_start_sampling) {
3715 return;
3716 }
3717 // When CMSEdenChunksRecordAlways is true, the eden chunk array
3718 // is populated by the young generation.
3719 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3720 if (_eden_chunk_index < _eden_chunk_capacity) {
3721 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
3722 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3723 "Unexpected state of Eden");
3724 // We'd like to check that what we just sampled is an oop-start address;
3725 // however, we cannot do that here since the object may not yet have been
3726 // initialized. So we'll instead do the check when we _use_ this sample
3727 // later.
3728 if (_eden_chunk_index == 0 ||
3729 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3730 _eden_chunk_array[_eden_chunk_index-1])
3731 >= CMSSamplingGrain)) {
3732 _eden_chunk_index++; // commit sample
3733 }
3734 }
3735 }
3736 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3737 size_t used = get_eden_used();
3738 size_t capacity = get_eden_capacity();
3739 assert(used <= capacity, "Unexpected state of Eden");
3740 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3741 _abort_preclean = true;
3742 }
3743 }
3744 }
3745
3746
3747 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3748 assert(_collectorState == Precleaning ||
3749 _collectorState == AbortablePreclean, "incorrect state");
3750 ResourceMark rm;
3751 HandleMark hm;
3752
3753 // Precleaning is currently not MT but the reference processor
3754 // may be set for MT. Disable it temporarily here.
3755 ReferenceProcessor* rp = ref_processor();
3756 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3757
3758 // Do one pass of scrubbing the discovered reference lists
3759 // to remove any reference objects with strongly-reachable
3760 // referents.
3761 if (clean_refs) {
3762 CMSPrecleanRefsYieldClosure yield_cl(this);
3763 assert(rp->span().equals(_span), "Spans should be equal");
3764 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3765 &_markStack, true /* preclean */);
3766 CMSDrainMarkingStackClosure complete_trace(this,
3767 _span, &_markBitMap, &_markStack,
3768 &keep_alive, true /* preclean */);
3769
3770 // We don't want this step to interfere with a young
3771 // collection because we don't want to take CPU
3772 // or memory bandwidth away from the young GC threads
3773 // (which may be as many as there are CPUs).
3774 // Note that we don't need to protect ourselves from
3775 // interference with mutators because they can't
3776 // manipulate the discovered reference lists nor affect
3777 // the computed reachability of the referents, the
3778 // only properties manipulated by the precleaning
3779 // of these reference lists.
3780 stopTimer();
3781 CMSTokenSyncWithLocks x(true /* is cms thread */,
3782 bitMapLock());
3783 startTimer();
3784 sample_eden();
3785
3786 // The following will yield to allow foreground
3787 // collection to proceed promptly. XXX YSR:
3788 // The code in this method may need further
3789 // tweaking for better performance and some restructuring
3790 // for cleaner interfaces.
3791 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3792 rp->preclean_discovered_references(
3793 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3794 gc_timer);
3795 }
3796
3797 if (clean_survivor) { // preclean the active survivor space(s)
3798 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3799 &_markBitMap, &_modUnionTable,
3800 &_markStack, true /* precleaning phase */);
3801 stopTimer();
3802 CMSTokenSyncWithLocks ts(true /* is cms thread */,
3803 bitMapLock());
3804 startTimer();
3805 unsigned int before_count =
3806 CMSHeap::heap()->total_collections();
3807 SurvivorSpacePrecleanClosure
3808 sss_cl(this, _span, &_markBitMap, &_markStack,
3809 &pam_cl, before_count, CMSYield);
3810 _young_gen->from()->object_iterate_careful(&sss_cl);
3811 _young_gen->to()->object_iterate_careful(&sss_cl);
3812 }
3813 MarkRefsIntoAndScanClosure
3814 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3815 &_markStack, this, CMSYield,
3816 true /* precleaning phase */);
3817 // CAUTION: The following closure has persistent state that may need to
3818 // be reset upon a decrease in the sequence of addresses it
3819 // processes.
3820 ScanMarkedObjectsAgainCarefullyClosure
3821 smoac_cl(this, _span,
3822 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3823
3824 // Preclean dirty cards in ModUnionTable and CardTable using
3825 // appropriate convergence criterion;
3826 // repeat CMSPrecleanIter times unless we find that
3827 // we are losing.
3828 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3829 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3830 "Bad convergence multiplier");
3831 assert(CMSPrecleanThreshold >= 100,
3832 "Unreasonably low CMSPrecleanThreshold");
3833
3834 size_t numIter, cumNumCards, lastNumCards, curNumCards;
3835 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3836 numIter < CMSPrecleanIter;
3837 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3838 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
3839 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3840 // Either there are very few dirty cards, so re-mark
3841 // pause will be small anyway, or our pre-cleaning isn't
3842 // that much faster than the rate at which cards are being
3843 // dirtied, so we might as well stop and re-mark since
3844 // precleaning won't improve our re-mark time by much.
3845 if (curNumCards <= CMSPrecleanThreshold ||
3846 (numIter > 0 &&
3847 (curNumCards * CMSPrecleanDenominator >
3848 lastNumCards * CMSPrecleanNumerator))) {
3849 numIter++;
3850 cumNumCards += curNumCards;
3851 break;
3852 }
3853 }
3854
3855 preclean_cld(&mrias_cl, _cmsGen->freelistLock());
3856
3857 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
3858 cumNumCards += curNumCards;
3859 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
3860 curNumCards, cumNumCards, numIter);
3861 return cumNumCards; // as a measure of useful work done
3862 }
3863
3864 // PRECLEANING NOTES:
3865 // Precleaning involves:
3866 // . reading the bits of the modUnionTable and clearing the set bits.
3867 // . For the cards corresponding to the set bits, we scan the
3868 // objects on those cards. This means we need the free_list_lock
3869 // so that we can safely iterate over the CMS space when scanning
3870 // for oops.
3871 // . When we scan the objects, we'll be both reading and setting
3872 // marks in the marking bit map, so we'll need the marking bit map.
3873 // . For protecting _collector_state transitions, we take the CGC_lock.
3874 // Note that any races in the reading of of card table entries by the
3875 // CMS thread on the one hand and the clearing of those entries by the
3876 // VM thread or the setting of those entries by the mutator threads on the
3877 // other are quite benign. However, for efficiency it makes sense to keep
3878 // the VM thread from racing with the CMS thread while the latter is
3879 // dirty card info to the modUnionTable. We therefore also use the
3880 // CGC_lock to protect the reading of the card table and the mod union
3881 // table by the CM thread.
3882 // . We run concurrently with mutator updates, so scanning
3883 // needs to be done carefully -- we should not try to scan
3884 // potentially uninitialized objects.
3885 //
3886 // Locking strategy: While holding the CGC_lock, we scan over and
3887 // reset a maximal dirty range of the mod union / card tables, then lock
3888 // the free_list_lock and bitmap lock to do a full marking, then
3889 // release these locks; and repeat the cycle. This allows for a
3890 // certain amount of fairness in the sharing of these locks between
3891 // the CMS collector on the one hand, and the VM thread and the
3892 // mutators on the other.
3893
3894 // NOTE: preclean_mod_union_table() and preclean_card_table()
3895 // further below are largely identical; if you need to modify
3896 // one of these methods, please check the other method too.
3897
3898 size_t CMSCollector::preclean_mod_union_table(
3899 ConcurrentMarkSweepGeneration* old_gen,
3900 ScanMarkedObjectsAgainCarefullyClosure* cl) {
3901 verify_work_stacks_empty();
3902 verify_overflow_empty();
3903
3904 // strategy: starting with the first card, accumulate contiguous
3905 // ranges of dirty cards; clear these cards, then scan the region
3906 // covered by these cards.
3907
3908 // Since all of the MUT is committed ahead, we can just use
3909 // that, in case the generations expand while we are precleaning.
3910 // It might also be fine to just use the committed part of the
3911 // generation, but we might potentially miss cards when the
3912 // generation is rapidly expanding while we are in the midst
3913 // of precleaning.
3914 HeapWord* startAddr = old_gen->reserved().start();
3915 HeapWord* endAddr = old_gen->reserved().end();
3916
3917 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
3918
3919 size_t numDirtyCards, cumNumDirtyCards;
3920 HeapWord *nextAddr, *lastAddr;
3921 for (cumNumDirtyCards = numDirtyCards = 0,
3922 nextAddr = lastAddr = startAddr;
3923 nextAddr < endAddr;
3924 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
3925
3926 ResourceMark rm;
3927 HandleMark hm;
3928
3929 MemRegion dirtyRegion;
3930 {
3931 stopTimer();
3932 // Potential yield point
3933 CMSTokenSync ts(true);
3934 startTimer();
3935 sample_eden();
3936 // Get dirty region starting at nextOffset (inclusive),
3937 // simultaneously clearing it.
3938 dirtyRegion =
3939 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
3940 assert(dirtyRegion.start() >= nextAddr,
3941 "returned region inconsistent?");
3942 }
3943 // Remember where the next search should begin.
3944 // The returned region (if non-empty) is a right open interval,
3945 // so lastOffset is obtained from the right end of that
3946 // interval.
3947 lastAddr = dirtyRegion.end();
3948 // Should do something more transparent and less hacky XXX
3949 numDirtyCards =
3950 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
3951
3952 // We'll scan the cards in the dirty region (with periodic
3953 // yields for foreground GC as needed).
3954 if (!dirtyRegion.is_empty()) {
3955 assert(numDirtyCards > 0, "consistency check");
3956 HeapWord* stop_point = NULL;
3957 stopTimer();
3958 // Potential yield point
3959 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(),
3960 bitMapLock());
3961 startTimer();
3962 {
3963 verify_work_stacks_empty();
3964 verify_overflow_empty();
3965 sample_eden();
3966 stop_point =
3967 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
3968 }
3969 if (stop_point != NULL) {
3970 // The careful iteration stopped early either because it found an
3971 // uninitialized object, or because we were in the midst of an
3972 // "abortable preclean", which should now be aborted. Redirty
3973 // the bits corresponding to the partially-scanned or unscanned
3974 // cards. We'll either restart at the next block boundary or
3975 // abort the preclean.
3976 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
3977 "Should only be AbortablePreclean.");
3978 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
3979 if (should_abort_preclean()) {
3980 break; // out of preclean loop
3981 } else {
3982 // Compute the next address at which preclean should pick up;
3983 // might need bitMapLock in order to read P-bits.
3984 lastAddr = next_card_start_after_block(stop_point);
3985 }
3986 }
3987 } else {
3988 assert(lastAddr == endAddr, "consistency check");
3989 assert(numDirtyCards == 0, "consistency check");
3990 break;
3991 }
3992 }
3993 verify_work_stacks_empty();
3994 verify_overflow_empty();
3995 return cumNumDirtyCards;
3996 }
3997
3998 // NOTE: preclean_mod_union_table() above and preclean_card_table()
3999 // below are largely identical; if you need to modify
4000 // one of these methods, please check the other method too.
4001
4002 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen,
4003 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4004 // strategy: it's similar to precleamModUnionTable above, in that
4005 // we accumulate contiguous ranges of dirty cards, mark these cards
4006 // precleaned, then scan the region covered by these cards.
4007 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high());
4008 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low());
4009
4010 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
4011
4012 size_t numDirtyCards, cumNumDirtyCards;
4013 HeapWord *lastAddr, *nextAddr;
4014
4015 for (cumNumDirtyCards = numDirtyCards = 0,
4016 nextAddr = lastAddr = startAddr;
4017 nextAddr < endAddr;
4018 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4019
4020 ResourceMark rm;
4021 HandleMark hm;
4022
4023 MemRegion dirtyRegion;
4024 {
4025 // See comments in "Precleaning notes" above on why we
4026 // do this locking. XXX Could the locking overheads be
4027 // too high when dirty cards are sparse? [I don't think so.]
4028 stopTimer();
4029 CMSTokenSync x(true); // is cms thread
4030 startTimer();
4031 sample_eden();
4032 // Get and clear dirty region from card table
4033 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4034 MemRegion(nextAddr, endAddr),
4035 true,
4036 CardTableModRefBS::precleaned_card_val());
4037
4038 assert(dirtyRegion.start() >= nextAddr,
4039 "returned region inconsistent?");
4040 }
4041 lastAddr = dirtyRegion.end();
4042 numDirtyCards =
4043 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4044
4045 if (!dirtyRegion.is_empty()) {
4046 stopTimer();
4047 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock());
4048 startTimer();
4049 sample_eden();
4050 verify_work_stacks_empty();
4051 verify_overflow_empty();
4052 HeapWord* stop_point =
4053 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4054 if (stop_point != NULL) {
4055 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4056 "Should only be AbortablePreclean.");
4057 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4058 if (should_abort_preclean()) {
4059 break; // out of preclean loop
4060 } else {
4061 // Compute the next address at which preclean should pick up.
4062 lastAddr = next_card_start_after_block(stop_point);
4063 }
4064 }
4065 } else {
4066 break;
4067 }
4068 }
4069 verify_work_stacks_empty();
4070 verify_overflow_empty();
4071 return cumNumDirtyCards;
4072 }
4073
4074 class PrecleanCLDClosure : public CLDClosure {
4075 MetadataAwareOopsInGenClosure* _cm_closure;
4076 public:
4077 PrecleanCLDClosure(MetadataAwareOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {}
4078 void do_cld(ClassLoaderData* cld) {
4079 if (cld->has_accumulated_modified_oops()) {
4080 cld->clear_accumulated_modified_oops();
4081
4082 _cm_closure->do_cld(cld);
4083 }
4084 }
4085 };
4086
4087 // The freelist lock is needed to prevent asserts, is it really needed?
4088 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4089
4090 cl->set_freelistLock(freelistLock);
4091
4092 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4093
4094 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4095 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4096 PrecleanCLDClosure preclean_closure(cl);
4097 ClassLoaderDataGraph::cld_do(&preclean_closure);
4098
4099 verify_work_stacks_empty();
4100 verify_overflow_empty();
4101 }
4102
4103 void CMSCollector::checkpointRootsFinal() {
4104 assert(_collectorState == FinalMarking, "incorrect state transition?");
4105 check_correct_thread_executing();
4106 // world is stopped at this checkpoint
4107 assert(SafepointSynchronize::is_at_safepoint(),
4108 "world should be stopped");
4109 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
4110
4111 verify_work_stacks_empty();
4112 verify_overflow_empty();
4113
4114 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)",
4115 _young_gen->used() / K, _young_gen->capacity() / K);
4116 {
4117 if (CMSScavengeBeforeRemark) {
4118 CMSHeap* heap = CMSHeap::heap();
4119 // Temporarily set flag to false, GCH->do_collection will
4120 // expect it to be false and set to true
4121 FlagSetting fl(heap->_is_gc_active, false);
4122
4123 heap->do_collection(true, // full (i.e. force, see below)
4124 false, // !clear_all_soft_refs
4125 0, // size
4126 false, // is_tlab
4127 GenCollectedHeap::YoungGen // type
4128 );
4129 }
4130 FreelistLocker x(this);
4131 MutexLockerEx y(bitMapLock(),
4132 Mutex::_no_safepoint_check_flag);
4133 checkpointRootsFinalWork();
4134 }
4135 verify_work_stacks_empty();
4136 verify_overflow_empty();
4137 }
4138
4139 void CMSCollector::checkpointRootsFinalWork() {
4140 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm);
4141
4142 assert(haveFreelistLocks(), "must have free list locks");
4143 assert_lock_strong(bitMapLock());
4144
4145 ResourceMark rm;
4146 HandleMark hm;
4147
4148 CMSHeap* heap = CMSHeap::heap();
4149
4150 if (should_unload_classes()) {
4151 CodeCache::gc_prologue();
4152 }
4153 assert(haveFreelistLocks(), "must have free list locks");
4154 assert_lock_strong(bitMapLock());
4155
4156 // We might assume that we need not fill TLAB's when
4157 // CMSScavengeBeforeRemark is set, because we may have just done
4158 // a scavenge which would have filled all TLAB's -- and besides
4159 // Eden would be empty. This however may not always be the case --
4160 // for instance although we asked for a scavenge, it may not have
4161 // happened because of a JNI critical section. We probably need
4162 // a policy for deciding whether we can in that case wait until
4163 // the critical section releases and then do the remark following
4164 // the scavenge, and skip it here. In the absence of that policy,
4165 // or of an indication of whether the scavenge did indeed occur,
4166 // we cannot rely on TLAB's having been filled and must do
4167 // so here just in case a scavenge did not happen.
4168 heap->ensure_parsability(false); // fill TLAB's, but no need to retire them
4169 // Update the saved marks which may affect the root scans.
4170 heap->save_marks();
4171
4172 print_eden_and_survivor_chunk_arrays();
4173
4174 {
4175 #if COMPILER2_OR_JVMCI
4176 DerivedPointerTableDeactivate dpt_deact;
4177 #endif
4178
4179 // Note on the role of the mod union table:
4180 // Since the marker in "markFromRoots" marks concurrently with
4181 // mutators, it is possible for some reachable objects not to have been
4182 // scanned. For instance, an only reference to an object A was
4183 // placed in object B after the marker scanned B. Unless B is rescanned,
4184 // A would be collected. Such updates to references in marked objects
4185 // are detected via the mod union table which is the set of all cards
4186 // dirtied since the first checkpoint in this GC cycle and prior to
4187 // the most recent young generation GC, minus those cleaned up by the
4188 // concurrent precleaning.
4189 if (CMSParallelRemarkEnabled) {
4190 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm);
4191 do_remark_parallel();
4192 } else {
4193 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm);
4194 do_remark_non_parallel();
4195 }
4196 }
4197 verify_work_stacks_empty();
4198 verify_overflow_empty();
4199
4200 {
4201 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm);
4202 refProcessingWork();
4203 }
4204 verify_work_stacks_empty();
4205 verify_overflow_empty();
4206
4207 if (should_unload_classes()) {
4208 CodeCache::gc_epilogue();
4209 }
4210 JvmtiExport::gc_epilogue();
4211
4212 // If we encountered any (marking stack / work queue) overflow
4213 // events during the current CMS cycle, take appropriate
4214 // remedial measures, where possible, so as to try and avoid
4215 // recurrence of that condition.
4216 assert(_markStack.isEmpty(), "No grey objects");
4217 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4218 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4219 if (ser_ovflw > 0) {
4220 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")",
4221 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4222 _markStack.expand();
4223 _ser_pmc_remark_ovflw = 0;
4224 _ser_pmc_preclean_ovflw = 0;
4225 _ser_kac_preclean_ovflw = 0;
4226 _ser_kac_ovflw = 0;
4227 }
4228 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4229 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")",
4230 _par_pmc_remark_ovflw, _par_kac_ovflw);
4231 _par_pmc_remark_ovflw = 0;
4232 _par_kac_ovflw = 0;
4233 }
4234 if (_markStack._hit_limit > 0) {
4235 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")",
4236 _markStack._hit_limit);
4237 }
4238 if (_markStack._failed_double > 0) {
4239 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT,
4240 _markStack._failed_double, _markStack.capacity());
4241 }
4242 _markStack._hit_limit = 0;
4243 _markStack._failed_double = 0;
4244
4245 if ((VerifyAfterGC || VerifyDuringGC) &&
4246 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) {
4247 verify_after_remark();
4248 }
4249
4250 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4251
4252 // Change under the freelistLocks.
4253 _collectorState = Sweeping;
4254 // Call isAllClear() under bitMapLock
4255 assert(_modUnionTable.isAllClear(),
4256 "Should be clear by end of the final marking");
4257 assert(_ct->cld_rem_set()->mod_union_is_clear(),
4258 "Should be clear by end of the final marking");
4259 }
4260
4261 void CMSParInitialMarkTask::work(uint worker_id) {
4262 elapsedTimer _timer;
4263 ResourceMark rm;
4264 HandleMark hm;
4265
4266 // ---------- scan from roots --------------
4267 _timer.start();
4268 CMSHeap* heap = CMSHeap::heap();
4269 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4270
4271 // ---------- young gen roots --------------
4272 {
4273 work_on_young_gen_roots(&par_mri_cl);
4274 _timer.stop();
4275 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4276 }
4277
4278 // ---------- remaining roots --------------
4279 _timer.reset();
4280 _timer.start();
4281
4282 CLDToOopClosure cld_closure(&par_mri_cl, true);
4283
4284 heap->cms_process_roots(_strong_roots_scope,
4285 false, // yg was scanned above
4286 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4287 _collector->should_unload_classes(),
4288 &par_mri_cl,
4289 &cld_closure);
4290 assert(_collector->should_unload_classes()
4291 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4292 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4293 _timer.stop();
4294 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4295 }
4296
4297 // Parallel remark task
4298 class CMSParRemarkTask: public CMSParMarkTask {
4299 CompactibleFreeListSpace* _cms_space;
4300
4301 // The per-thread work queues, available here for stealing.
4302 OopTaskQueueSet* _task_queues;
4303 ParallelTaskTerminator _term;
4304 StrongRootsScope* _strong_roots_scope;
4305
4306 public:
4307 // A value of 0 passed to n_workers will cause the number of
4308 // workers to be taken from the active workers in the work gang.
4309 CMSParRemarkTask(CMSCollector* collector,
4310 CompactibleFreeListSpace* cms_space,
4311 uint n_workers, WorkGang* workers,
4312 OopTaskQueueSet* task_queues,
4313 StrongRootsScope* strong_roots_scope):
4314 CMSParMarkTask("Rescan roots and grey objects in parallel",
4315 collector, n_workers),
4316 _cms_space(cms_space),
4317 _task_queues(task_queues),
4318 _term(n_workers, task_queues),
4319 _strong_roots_scope(strong_roots_scope) { }
4320
4321 OopTaskQueueSet* task_queues() { return _task_queues; }
4322
4323 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4324
4325 ParallelTaskTerminator* terminator() { return &_term; }
4326 uint n_workers() { return _n_workers; }
4327
4328 void work(uint worker_id);
4329
4330 private:
4331 // ... of dirty cards in old space
4332 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4333 ParMarkRefsIntoAndScanClosure* cl);
4334
4335 // ... work stealing for the above
4336 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed);
4337 };
4338
4339 class RemarkCLDClosure : public CLDClosure {
4340 CLDToOopClosure _cm_closure;
4341 public:
4342 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure) {}
4343 void do_cld(ClassLoaderData* cld) {
4344 // Check if we have modified any oops in the CLD during the concurrent marking.
4345 if (cld->has_accumulated_modified_oops()) {
4346 cld->clear_accumulated_modified_oops();
4347
4348 // We could have transfered the current modified marks to the accumulated marks,
4349 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4350 } else if (cld->has_modified_oops()) {
4351 // Don't clear anything, this info is needed by the next young collection.
4352 } else {
4353 // No modified oops in the ClassLoaderData.
4354 return;
4355 }
4356
4357 // The klass has modified fields, need to scan the klass.
4358 _cm_closure.do_cld(cld);
4359 }
4360 };
4361
4362 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) {
4363 ParNewGeneration* young_gen = _collector->_young_gen;
4364 ContiguousSpace* eden_space = young_gen->eden();
4365 ContiguousSpace* from_space = young_gen->from();
4366 ContiguousSpace* to_space = young_gen->to();
4367
4368 HeapWord** eca = _collector->_eden_chunk_array;
4369 size_t ect = _collector->_eden_chunk_index;
4370 HeapWord** sca = _collector->_survivor_chunk_array;
4371 size_t sct = _collector->_survivor_chunk_index;
4372
4373 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4374 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4375
4376 do_young_space_rescan(cl, to_space, NULL, 0);
4377 do_young_space_rescan(cl, from_space, sca, sct);
4378 do_young_space_rescan(cl, eden_space, eca, ect);
4379 }
4380
4381 // work_queue(i) is passed to the closure
4382 // ParMarkRefsIntoAndScanClosure. The "i" parameter
4383 // also is passed to do_dirty_card_rescan_tasks() and to
4384 // do_work_steal() to select the i-th task_queue.
4385
4386 void CMSParRemarkTask::work(uint worker_id) {
4387 elapsedTimer _timer;
4388 ResourceMark rm;
4389 HandleMark hm;
4390
4391 // ---------- rescan from roots --------------
4392 _timer.start();
4393 CMSHeap* heap = CMSHeap::heap();
4394 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4395 _collector->_span, _collector->ref_processor(),
4396 &(_collector->_markBitMap),
4397 work_queue(worker_id));
4398
4399 // Rescan young gen roots first since these are likely
4400 // coarsely partitioned and may, on that account, constitute
4401 // the critical path; thus, it's best to start off that
4402 // work first.
4403 // ---------- young gen roots --------------
4404 {
4405 work_on_young_gen_roots(&par_mrias_cl);
4406 _timer.stop();
4407 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4408 }
4409
4410 // ---------- remaining roots --------------
4411 _timer.reset();
4412 _timer.start();
4413 heap->cms_process_roots(_strong_roots_scope,
4414 false, // yg was scanned above
4415 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4416 _collector->should_unload_classes(),
4417 &par_mrias_cl,
4418 NULL); // The dirty klasses will be handled below
4419
4420 assert(_collector->should_unload_classes()
4421 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4422 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4423 _timer.stop();
4424 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4425
4426 // ---------- unhandled CLD scanning ----------
4427 if (worker_id == 0) { // Single threaded at the moment.
4428 _timer.reset();
4429 _timer.start();
4430
4431 // Scan all new class loader data objects and new dependencies that were
4432 // introduced during concurrent marking.
4433 ResourceMark rm;
4434 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4435 for (int i = 0; i < array->length(); i++) {
4436 par_mrias_cl.do_cld_nv(array->at(i));
4437 }
4438
4439 // We don't need to keep track of new CLDs anymore.
4440 ClassLoaderDataGraph::remember_new_clds(false);
4441
4442 _timer.stop();
4443 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4444 }
4445
4446 // We might have added oops to ClassLoaderData::_handles during the
4447 // concurrent marking phase. These oops do not always point to newly allocated objects
4448 // that are guaranteed to be kept alive. Hence,
4449 // we do have to revisit the _handles block during the remark phase.
4450
4451 // ---------- dirty CLD scanning ----------
4452 if (worker_id == 0) { // Single threaded at the moment.
4453 _timer.reset();
4454 _timer.start();
4455
4456 // Scan all classes that was dirtied during the concurrent marking phase.
4457 RemarkCLDClosure remark_closure(&par_mrias_cl);
4458 ClassLoaderDataGraph::cld_do(&remark_closure);
4459
4460 _timer.stop();
4461 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4462 }
4463
4464
4465 // ---------- rescan dirty cards ------------
4466 _timer.reset();
4467 _timer.start();
4468
4469 // Do the rescan tasks for each of the two spaces
4470 // (cms_space) in turn.
4471 // "worker_id" is passed to select the task_queue for "worker_id"
4472 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4473 _timer.stop();
4474 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4475
4476 // ---------- steal work from other threads ...
4477 // ---------- ... and drain overflow list.
4478 _timer.reset();
4479 _timer.start();
4480 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4481 _timer.stop();
4482 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4483 }
4484
4485 void
4486 CMSParMarkTask::do_young_space_rescan(
4487 OopsInGenClosure* cl, ContiguousSpace* space,
4488 HeapWord** chunk_array, size_t chunk_top) {
4489 // Until all tasks completed:
4490 // . claim an unclaimed task
4491 // . compute region boundaries corresponding to task claimed
4492 // using chunk_array
4493 // . par_oop_iterate(cl) over that region
4494
4495 ResourceMark rm;
4496 HandleMark hm;
4497
4498 SequentialSubTasksDone* pst = space->par_seq_tasks();
4499
4500 uint nth_task = 0;
4501 uint n_tasks = pst->n_tasks();
4502
4503 if (n_tasks > 0) {
4504 assert(pst->valid(), "Uninitialized use?");
4505 HeapWord *start, *end;
4506 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4507 // We claimed task # nth_task; compute its boundaries.
4508 if (chunk_top == 0) { // no samples were taken
4509 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4510 start = space->bottom();
4511 end = space->top();
4512 } else if (nth_task == 0) {
4513 start = space->bottom();
4514 end = chunk_array[nth_task];
4515 } else if (nth_task < (uint)chunk_top) {
4516 assert(nth_task >= 1, "Control point invariant");
4517 start = chunk_array[nth_task - 1];
4518 end = chunk_array[nth_task];
4519 } else {
4520 assert(nth_task == (uint)chunk_top, "Control point invariant");
4521 start = chunk_array[chunk_top - 1];
4522 end = space->top();
4523 }
4524 MemRegion mr(start, end);
4525 // Verify that mr is in space
4526 assert(mr.is_empty() || space->used_region().contains(mr),
4527 "Should be in space");
4528 // Verify that "start" is an object boundary
4529 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())),
4530 "Should be an oop");
4531 space->par_oop_iterate(mr, cl);
4532 }
4533 pst->all_tasks_completed();
4534 }
4535 }
4536
4537 void
4538 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4539 CompactibleFreeListSpace* sp, int i,
4540 ParMarkRefsIntoAndScanClosure* cl) {
4541 // Until all tasks completed:
4542 // . claim an unclaimed task
4543 // . compute region boundaries corresponding to task claimed
4544 // . transfer dirty bits ct->mut for that region
4545 // . apply rescanclosure to dirty mut bits for that region
4546
4547 ResourceMark rm;
4548 HandleMark hm;
4549
4550 OopTaskQueue* work_q = work_queue(i);
4551 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4552 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4553 // CAUTION: This closure has state that persists across calls to
4554 // the work method dirty_range_iterate_clear() in that it has
4555 // embedded in it a (subtype of) UpwardsObjectClosure. The
4556 // use of that state in the embedded UpwardsObjectClosure instance
4557 // assumes that the cards are always iterated (even if in parallel
4558 // by several threads) in monotonically increasing order per each
4559 // thread. This is true of the implementation below which picks
4560 // card ranges (chunks) in monotonically increasing order globally
4561 // and, a-fortiori, in monotonically increasing order per thread
4562 // (the latter order being a subsequence of the former).
4563 // If the work code below is ever reorganized into a more chaotic
4564 // work-partitioning form than the current "sequential tasks"
4565 // paradigm, the use of that persistent state will have to be
4566 // revisited and modified appropriately. See also related
4567 // bug 4756801 work on which should examine this code to make
4568 // sure that the changes there do not run counter to the
4569 // assumptions made here and necessary for correctness and
4570 // efficiency. Note also that this code might yield inefficient
4571 // behavior in the case of very large objects that span one or
4572 // more work chunks. Such objects would potentially be scanned
4573 // several times redundantly. Work on 4756801 should try and
4574 // address that performance anomaly if at all possible. XXX
4575 MemRegion full_span = _collector->_span;
4576 CMSBitMap* bm = &(_collector->_markBitMap); // shared
4577 MarkFromDirtyCardsClosure
4578 greyRescanClosure(_collector, full_span, // entire span of interest
4579 sp, bm, work_q, cl);
4580
4581 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4582 assert(pst->valid(), "Uninitialized use?");
4583 uint nth_task = 0;
4584 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
4585 MemRegion span = sp->used_region();
4586 HeapWord* start_addr = span.start();
4587 HeapWord* end_addr = align_up(span.end(), alignment);
4588 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4589 assert(is_aligned(start_addr, alignment), "Check alignment");
4590 assert(is_aligned(chunk_size, alignment), "Check alignment");
4591
4592 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4593 // Having claimed the nth_task, compute corresponding mem-region,
4594 // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4595 // The alignment restriction ensures that we do not need any
4596 // synchronization with other gang-workers while setting or
4597 // clearing bits in thus chunk of the MUT.
4598 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4599 start_addr + (nth_task+1)*chunk_size);
4600 // The last chunk's end might be way beyond end of the
4601 // used region. In that case pull back appropriately.
4602 if (this_span.end() > end_addr) {
4603 this_span.set_end(end_addr);
4604 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4605 }
4606 // Iterate over the dirty cards covering this chunk, marking them
4607 // precleaned, and setting the corresponding bits in the mod union
4608 // table. Since we have been careful to partition at Card and MUT-word
4609 // boundaries no synchronization is needed between parallel threads.
4610 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
4611 &modUnionClosure);
4612
4613 // Having transferred these marks into the modUnionTable,
4614 // rescan the marked objects on the dirty cards in the modUnionTable.
4615 // Even if this is at a synchronous collection, the initial marking
4616 // may have been done during an asynchronous collection so there
4617 // may be dirty bits in the mod-union table.
4618 _collector->_modUnionTable.dirty_range_iterate_clear(
4619 this_span, &greyRescanClosure);
4620 _collector->_modUnionTable.verifyNoOneBitsInRange(
4621 this_span.start(),
4622 this_span.end());
4623 }
4624 pst->all_tasks_completed(); // declare that i am done
4625 }
4626
4627 // . see if we can share work_queues with ParNew? XXX
4628 void
4629 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl,
4630 int* seed) {
4631 OopTaskQueue* work_q = work_queue(i);
4632 NOT_PRODUCT(int num_steals = 0;)
4633 oop obj_to_scan;
4634 CMSBitMap* bm = &(_collector->_markBitMap);
4635
4636 while (true) {
4637 // Completely finish any left over work from (an) earlier round(s)
4638 cl->trim_queue(0);
4639 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4640 (size_t)ParGCDesiredObjsFromOverflowList);
4641 // Now check if there's any work in the overflow list
4642 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4643 // only affects the number of attempts made to get work from the
4644 // overflow list and does not affect the number of workers. Just
4645 // pass ParallelGCThreads so this behavior is unchanged.
4646 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4647 work_q,
4648 ParallelGCThreads)) {
4649 // found something in global overflow list;
4650 // not yet ready to go stealing work from others.
4651 // We'd like to assert(work_q->size() != 0, ...)
4652 // because we just took work from the overflow list,
4653 // but of course we can't since all of that could have
4654 // been already stolen from us.
4655 // "He giveth and He taketh away."
4656 continue;
4657 }
4658 // Verify that we have no work before we resort to stealing
4659 assert(work_q->size() == 0, "Have work, shouldn't steal");
4660 // Try to steal from other queues that have work
4661 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4662 NOT_PRODUCT(num_steals++;)
4663 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!");
4664 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4665 // Do scanning work
4666 obj_to_scan->oop_iterate(cl);
4667 // Loop around, finish this work, and try to steal some more
4668 } else if (terminator()->offer_termination()) {
4669 break; // nirvana from the infinite cycle
4670 }
4671 }
4672 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
4673 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4674 "Else our work is not yet done");
4675 }
4676
4677 // Record object boundaries in _eden_chunk_array by sampling the eden
4678 // top in the slow-path eden object allocation code path and record
4679 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4680 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4681 // sampling in sample_eden() that activates during the part of the
4682 // preclean phase.
4683 void CMSCollector::sample_eden_chunk() {
4684 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4685 if (_eden_chunk_lock->try_lock()) {
4686 // Record a sample. This is the critical section. The contents
4687 // of the _eden_chunk_array have to be non-decreasing in the
4688 // address order.
4689 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4690 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4691 "Unexpected state of Eden");
4692 if (_eden_chunk_index == 0 ||
4693 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4694 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4695 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4696 _eden_chunk_index++; // commit sample
4697 }
4698 _eden_chunk_lock->unlock();
4699 }
4700 }
4701 }
4702
4703 // Return a thread-local PLAB recording array, as appropriate.
4704 void* CMSCollector::get_data_recorder(int thr_num) {
4705 if (_survivor_plab_array != NULL &&
4706 (CMSPLABRecordAlways ||
4707 (_collectorState > Marking && _collectorState < FinalMarking))) {
4708 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4709 ChunkArray* ca = &_survivor_plab_array[thr_num];
4710 ca->reset(); // clear it so that fresh data is recorded
4711 return (void*) ca;
4712 } else {
4713 return NULL;
4714 }
4715 }
4716
4717 // Reset all the thread-local PLAB recording arrays
4718 void CMSCollector::reset_survivor_plab_arrays() {
4719 for (uint i = 0; i < ParallelGCThreads; i++) {
4720 _survivor_plab_array[i].reset();
4721 }
4722 }
4723
4724 // Merge the per-thread plab arrays into the global survivor chunk
4725 // array which will provide the partitioning of the survivor space
4726 // for CMS initial scan and rescan.
4727 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4728 int no_of_gc_threads) {
4729 assert(_survivor_plab_array != NULL, "Error");
4730 assert(_survivor_chunk_array != NULL, "Error");
4731 assert(_collectorState == FinalMarking ||
4732 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4733 for (int j = 0; j < no_of_gc_threads; j++) {
4734 _cursor[j] = 0;
4735 }
4736 HeapWord* top = surv->top();
4737 size_t i;
4738 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
4739 HeapWord* min_val = top; // Higher than any PLAB address
4740 uint min_tid = 0; // position of min_val this round
4741 for (int j = 0; j < no_of_gc_threads; j++) {
4742 ChunkArray* cur_sca = &_survivor_plab_array[j];
4743 if (_cursor[j] == cur_sca->end()) {
4744 continue;
4745 }
4746 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4747 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4748 assert(surv->used_region().contains(cur_val), "Out of bounds value");
4749 if (cur_val < min_val) {
4750 min_tid = j;
4751 min_val = cur_val;
4752 } else {
4753 assert(cur_val < top, "All recorded addresses should be less");
4754 }
4755 }
4756 // At this point min_val and min_tid are respectively
4757 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4758 // and the thread (j) that witnesses that address.
4759 // We record this address in the _survivor_chunk_array[i]
4760 // and increment _cursor[min_tid] prior to the next round i.
4761 if (min_val == top) {
4762 break;
4763 }
4764 _survivor_chunk_array[i] = min_val;
4765 _cursor[min_tid]++;
4766 }
4767 // We are all done; record the size of the _survivor_chunk_array
4768 _survivor_chunk_index = i; // exclusive: [0, i)
4769 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4770 // Verify that we used up all the recorded entries
4771 #ifdef ASSERT
4772 size_t total = 0;
4773 for (int j = 0; j < no_of_gc_threads; j++) {
4774 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4775 total += _cursor[j];
4776 }
4777 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
4778 // Check that the merged array is in sorted order
4779 if (total > 0) {
4780 for (size_t i = 0; i < total - 1; i++) {
4781 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
4782 i, p2i(_survivor_chunk_array[i]));
4783 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
4784 "Not sorted");
4785 }
4786 }
4787 #endif // ASSERT
4788 }
4789
4790 // Set up the space's par_seq_tasks structure for work claiming
4791 // for parallel initial scan and rescan of young gen.
4792 // See ParRescanTask where this is currently used.
4793 void
4794 CMSCollector::
4795 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
4796 assert(n_threads > 0, "Unexpected n_threads argument");
4797
4798 // Eden space
4799 if (!_young_gen->eden()->is_empty()) {
4800 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
4801 assert(!pst->valid(), "Clobbering existing data?");
4802 // Each valid entry in [0, _eden_chunk_index) represents a task.
4803 size_t n_tasks = _eden_chunk_index + 1;
4804 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
4805 // Sets the condition for completion of the subtask (how many threads
4806 // need to finish in order to be done).
4807 pst->set_n_threads(n_threads);
4808 pst->set_n_tasks((int)n_tasks);
4809 }
4810
4811 // Merge the survivor plab arrays into _survivor_chunk_array
4812 if (_survivor_plab_array != NULL) {
4813 merge_survivor_plab_arrays(_young_gen->from(), n_threads);
4814 } else {
4815 assert(_survivor_chunk_index == 0, "Error");
4816 }
4817
4818 // To space
4819 {
4820 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
4821 assert(!pst->valid(), "Clobbering existing data?");
4822 // Sets the condition for completion of the subtask (how many threads
4823 // need to finish in order to be done).
4824 pst->set_n_threads(n_threads);
4825 pst->set_n_tasks(1);
4826 assert(pst->valid(), "Error");
4827 }
4828
4829 // From space
4830 {
4831 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
4832 assert(!pst->valid(), "Clobbering existing data?");
4833 size_t n_tasks = _survivor_chunk_index + 1;
4834 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
4835 // Sets the condition for completion of the subtask (how many threads
4836 // need to finish in order to be done).
4837 pst->set_n_threads(n_threads);
4838 pst->set_n_tasks((int)n_tasks);
4839 assert(pst->valid(), "Error");
4840 }
4841 }
4842
4843 // Parallel version of remark
4844 void CMSCollector::do_remark_parallel() {
4845 CMSHeap* heap = CMSHeap::heap();
4846 WorkGang* workers = heap->workers();
4847 assert(workers != NULL, "Need parallel worker threads.");
4848 // Choose to use the number of GC workers most recently set
4849 // into "active_workers".
4850 uint n_workers = workers->active_workers();
4851
4852 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4853
4854 StrongRootsScope srs(n_workers);
4855
4856 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
4857
4858 // We won't be iterating over the cards in the card table updating
4859 // the younger_gen cards, so we shouldn't call the following else
4860 // the verification code as well as subsequent younger_refs_iterate
4861 // code would get confused. XXX
4862 // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
4863
4864 // The young gen rescan work will not be done as part of
4865 // process_roots (which currently doesn't know how to
4866 // parallelize such a scan), but rather will be broken up into
4867 // a set of parallel tasks (via the sampling that the [abortable]
4868 // preclean phase did of eden, plus the [two] tasks of
4869 // scanning the [two] survivor spaces. Further fine-grain
4870 // parallelization of the scanning of the survivor spaces
4871 // themselves, and of precleaning of the young gen itself
4872 // is deferred to the future.
4873 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
4874
4875 // The dirty card rescan work is broken up into a "sequence"
4876 // of parallel tasks (per constituent space) that are dynamically
4877 // claimed by the parallel threads.
4878 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
4879
4880 // It turns out that even when we're using 1 thread, doing the work in a
4881 // separate thread causes wide variance in run times. We can't help this
4882 // in the multi-threaded case, but we special-case n=1 here to get
4883 // repeatable measurements of the 1-thread overhead of the parallel code.
4884 if (n_workers > 1) {
4885 // Make refs discovery MT-safe, if it isn't already: it may not
4886 // necessarily be so, since it's possible that we are doing
4887 // ST marking.
4888 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
4889 workers->run_task(&tsk);
4890 } else {
4891 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4892 tsk.work(0);
4893 }
4894
4895 // restore, single-threaded for now, any preserved marks
4896 // as a result of work_q overflow
4897 restore_preserved_marks_if_any();
4898 }
4899
4900 // Non-parallel version of remark
4901 void CMSCollector::do_remark_non_parallel() {
4902 ResourceMark rm;
4903 HandleMark hm;
4904 CMSHeap* heap = CMSHeap::heap();
4905 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4906
4907 MarkRefsIntoAndScanClosure
4908 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
4909 &_markStack, this,
4910 false /* should_yield */, false /* not precleaning */);
4911 MarkFromDirtyCardsClosure
4912 markFromDirtyCardsClosure(this, _span,
4913 NULL, // space is set further below
4914 &_markBitMap, &_markStack, &mrias_cl);
4915 {
4916 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm);
4917 // Iterate over the dirty cards, setting the corresponding bits in the
4918 // mod union table.
4919 {
4920 ModUnionClosure modUnionClosure(&_modUnionTable);
4921 _ct->ct_bs()->dirty_card_iterate(
4922 _cmsGen->used_region(),
4923 &modUnionClosure);
4924 }
4925 // Having transferred these marks into the modUnionTable, we just need
4926 // to rescan the marked objects on the dirty cards in the modUnionTable.
4927 // The initial marking may have been done during an asynchronous
4928 // collection so there may be dirty bits in the mod-union table.
4929 const int alignment =
4930 CardTableModRefBS::card_size * BitsPerWord;
4931 {
4932 // ... First handle dirty cards in CMS gen
4933 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
4934 MemRegion ur = _cmsGen->used_region();
4935 HeapWord* lb = ur.start();
4936 HeapWord* ub = align_up(ur.end(), alignment);
4937 MemRegion cms_span(lb, ub);
4938 _modUnionTable.dirty_range_iterate_clear(cms_span,
4939 &markFromDirtyCardsClosure);
4940 verify_work_stacks_empty();
4941 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards());
4942 }
4943 }
4944 if (VerifyDuringGC &&
4945 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) {
4946 HandleMark hm; // Discard invalid handles created during verification
4947 Universe::verify();
4948 }
4949 {
4950 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm);
4951
4952 verify_work_stacks_empty();
4953
4954 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
4955 StrongRootsScope srs(1);
4956
4957 heap->cms_process_roots(&srs,
4958 true, // young gen as roots
4959 GenCollectedHeap::ScanningOption(roots_scanning_options()),
4960 should_unload_classes(),
4961 &mrias_cl,
4962 NULL); // The dirty klasses will be handled below
4963
4964 assert(should_unload_classes()
4965 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4966 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4967 }
4968
4969 {
4970 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm);
4971
4972 verify_work_stacks_empty();
4973
4974 // Scan all class loader data objects that might have been introduced
4975 // during concurrent marking.
4976 ResourceMark rm;
4977 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4978 for (int i = 0; i < array->length(); i++) {
4979 mrias_cl.do_cld_nv(array->at(i));
4980 }
4981
4982 // We don't need to keep track of new CLDs anymore.
4983 ClassLoaderDataGraph::remember_new_clds(false);
4984
4985 verify_work_stacks_empty();
4986 }
4987
4988 // We might have added oops to ClassLoaderData::_handles during the
4989 // concurrent marking phase. These oops do not point to newly allocated objects
4990 // that are guaranteed to be kept alive. Hence,
4991 // we do have to revisit the _handles block during the remark phase.
4992 {
4993 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm);
4994
4995 verify_work_stacks_empty();
4996
4997 RemarkCLDClosure remark_closure(&mrias_cl);
4998 ClassLoaderDataGraph::cld_do(&remark_closure);
4999
5000 verify_work_stacks_empty();
5001 }
5002
5003 verify_work_stacks_empty();
5004 // Restore evacuated mark words, if any, used for overflow list links
5005 restore_preserved_marks_if_any();
5006
5007 verify_overflow_empty();
5008 }
5009
5010 ////////////////////////////////////////////////////////
5011 // Parallel Reference Processing Task Proxy Class
5012 ////////////////////////////////////////////////////////
5013 class AbstractGangTaskWOopQueues : public AbstractGangTask {
5014 OopTaskQueueSet* _queues;
5015 ParallelTaskTerminator _terminator;
5016 public:
5017 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) :
5018 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {}
5019 ParallelTaskTerminator* terminator() { return &_terminator; }
5020 OopTaskQueueSet* queues() { return _queues; }
5021 };
5022
5023 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5024 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5025 CMSCollector* _collector;
5026 CMSBitMap* _mark_bit_map;
5027 const MemRegion _span;
5028 ProcessTask& _task;
5029
5030 public:
5031 CMSRefProcTaskProxy(ProcessTask& task,
5032 CMSCollector* collector,
5033 const MemRegion& span,
5034 CMSBitMap* mark_bit_map,
5035 AbstractWorkGang* workers,
5036 OopTaskQueueSet* task_queues):
5037 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5038 task_queues,
5039 workers->active_workers()),
5040 _task(task),
5041 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5042 {
5043 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5044 "Inconsistency in _span");
5045 }
5046
5047 OopTaskQueueSet* task_queues() { return queues(); }
5048
5049 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5050
5051 void do_work_steal(int i,
5052 CMSParDrainMarkingStackClosure* drain,
5053 CMSParKeepAliveClosure* keep_alive,
5054 int* seed);
5055
5056 virtual void work(uint worker_id);
5057 };
5058
5059 void CMSRefProcTaskProxy::work(uint worker_id) {
5060 ResourceMark rm;
5061 HandleMark hm;
5062 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5063 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5064 _mark_bit_map,
5065 work_queue(worker_id));
5066 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5067 _mark_bit_map,
5068 work_queue(worker_id));
5069 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5070 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5071 if (_task.marks_oops_alive()) {
5072 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5073 _collector->hash_seed(worker_id));
5074 }
5075 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5076 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5077 }
5078
5079 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5080 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5081 EnqueueTask& _task;
5082
5083 public:
5084 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5085 : AbstractGangTask("Enqueue reference objects in parallel"),
5086 _task(task)
5087 { }
5088
5089 virtual void work(uint worker_id)
5090 {
5091 _task.work(worker_id);
5092 }
5093 };
5094
5095 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5096 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5097 _span(span),
5098 _bit_map(bit_map),
5099 _work_queue(work_queue),
5100 _mark_and_push(collector, span, bit_map, work_queue),
5101 _low_water_mark(MIN2((work_queue->max_elems()/4),
5102 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5103 { }
5104
5105 // . see if we can share work_queues with ParNew? XXX
5106 void CMSRefProcTaskProxy::do_work_steal(int i,
5107 CMSParDrainMarkingStackClosure* drain,
5108 CMSParKeepAliveClosure* keep_alive,
5109 int* seed) {
5110 OopTaskQueue* work_q = work_queue(i);
5111 NOT_PRODUCT(int num_steals = 0;)
5112 oop obj_to_scan;
5113
5114 while (true) {
5115 // Completely finish any left over work from (an) earlier round(s)
5116 drain->trim_queue(0);
5117 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5118 (size_t)ParGCDesiredObjsFromOverflowList);
5119 // Now check if there's any work in the overflow list
5120 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5121 // only affects the number of attempts made to get work from the
5122 // overflow list and does not affect the number of workers. Just
5123 // pass ParallelGCThreads so this behavior is unchanged.
5124 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5125 work_q,
5126 ParallelGCThreads)) {
5127 // Found something in global overflow list;
5128 // not yet ready to go stealing work from others.
5129 // We'd like to assert(work_q->size() != 0, ...)
5130 // because we just took work from the overflow list,
5131 // but of course we can't, since all of that might have
5132 // been already stolen from us.
5133 continue;
5134 }
5135 // Verify that we have no work before we resort to stealing
5136 assert(work_q->size() == 0, "Have work, shouldn't steal");
5137 // Try to steal from other queues that have work
5138 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5139 NOT_PRODUCT(num_steals++;)
5140 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!");
5141 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5142 // Do scanning work
5143 obj_to_scan->oop_iterate(keep_alive);
5144 // Loop around, finish this work, and try to steal some more
5145 } else if (terminator()->offer_termination()) {
5146 break; // nirvana from the infinite cycle
5147 }
5148 }
5149 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
5150 }
5151
5152 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5153 {
5154 CMSHeap* heap = CMSHeap::heap();
5155 WorkGang* workers = heap->workers();
5156 assert(workers != NULL, "Need parallel worker threads.");
5157 CMSRefProcTaskProxy rp_task(task, &_collector,
5158 _collector.ref_processor()->span(),
5159 _collector.markBitMap(),
5160 workers, _collector.task_queues());
5161 workers->run_task(&rp_task);
5162 }
5163
5164 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5165 {
5166
5167 CMSHeap* heap = CMSHeap::heap();
5168 WorkGang* workers = heap->workers();
5169 assert(workers != NULL, "Need parallel worker threads.");
5170 CMSRefEnqueueTaskProxy enq_task(task);
5171 workers->run_task(&enq_task);
5172 }
5173
5174 void CMSCollector::refProcessingWork() {
5175 ResourceMark rm;
5176 HandleMark hm;
5177
5178 ReferenceProcessor* rp = ref_processor();
5179 assert(rp->span().equals(_span), "Spans should be equal");
5180 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5181 // Process weak references.
5182 rp->setup_policy(false);
5183 verify_work_stacks_empty();
5184
5185 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->num_q());
5186 {
5187 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm);
5188
5189 // Setup keep_alive and complete closures.
5190 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5191 &_markStack, false /* !preclean */);
5192 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5193 _span, &_markBitMap, &_markStack,
5194 &cmsKeepAliveClosure, false /* !preclean */);
5195
5196 ReferenceProcessorStats stats;
5197 if (rp->processing_is_mt()) {
5198 // Set the degree of MT here. If the discovery is done MT, there
5199 // may have been a different number of threads doing the discovery
5200 // and a different number of discovered lists may have Ref objects.
5201 // That is OK as long as the Reference lists are balanced (see
5202 // balance_all_queues() and balance_queues()).
5203 CMSHeap* heap = CMSHeap::heap();
5204 uint active_workers = ParallelGCThreads;
5205 WorkGang* workers = heap->workers();
5206 if (workers != NULL) {
5207 active_workers = workers->active_workers();
5208 // The expectation is that active_workers will have already
5209 // been set to a reasonable value. If it has not been set,
5210 // investigate.
5211 assert(active_workers > 0, "Should have been set during scavenge");
5212 }
5213 rp->set_active_mt_degree(active_workers);
5214 CMSRefProcTaskExecutor task_executor(*this);
5215 stats = rp->process_discovered_references(&_is_alive_closure,
5216 &cmsKeepAliveClosure,
5217 &cmsDrainMarkingStackClosure,
5218 &task_executor,
5219 &pt);
5220 } else {
5221 stats = rp->process_discovered_references(&_is_alive_closure,
5222 &cmsKeepAliveClosure,
5223 &cmsDrainMarkingStackClosure,
5224 NULL,
5225 &pt);
5226 }
5227 _gc_tracer_cm->report_gc_reference_stats(stats);
5228 pt.print_all_references();
5229 }
5230
5231 // This is the point where the entire marking should have completed.
5232 verify_work_stacks_empty();
5233
5234 {
5235 GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm);
5236 WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl);
5237 }
5238
5239 if (should_unload_classes()) {
5240 {
5241 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm);
5242
5243 // Unload classes and purge the SystemDictionary.
5244 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure, _gc_timer_cm);
5245
5246 // Unload nmethods.
5247 CodeCache::do_unloading(&_is_alive_closure, purged_class);
5248
5249 // Prune dead klasses from subklass/sibling/implementor lists.
5250 Klass::clean_weak_klass_links(&_is_alive_closure);
5251 }
5252
5253 {
5254 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm);
5255 // Clean up unreferenced symbols in symbol table.
5256 SymbolTable::unlink();
5257 }
5258
5259 {
5260 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm);
5261 // Delete entries for dead interned strings.
5262 StringTable::unlink(&_is_alive_closure);
5263 }
5264 }
5265
5266 // Restore any preserved marks as a result of mark stack or
5267 // work queue overflow
5268 restore_preserved_marks_if_any(); // done single-threaded for now
5269
5270 rp->set_enqueuing_is_done(true);
5271 if (rp->processing_is_mt()) {
5272 rp->balance_all_queues();
5273 CMSRefProcTaskExecutor task_executor(*this);
5274 rp->enqueue_discovered_references(&task_executor, &pt);
5275 } else {
5276 rp->enqueue_discovered_references(NULL, &pt);
5277 }
5278 rp->verify_no_references_recorded();
5279 pt.print_enqueue_phase();
5280 assert(!rp->discovery_enabled(), "should have been disabled");
5281 }
5282
5283 #ifndef PRODUCT
5284 void CMSCollector::check_correct_thread_executing() {
5285 Thread* t = Thread::current();
5286 // Only the VM thread or the CMS thread should be here.
5287 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5288 "Unexpected thread type");
5289 // If this is the vm thread, the foreground process
5290 // should not be waiting. Note that _foregroundGCIsActive is
5291 // true while the foreground collector is waiting.
5292 if (_foregroundGCShouldWait) {
5293 // We cannot be the VM thread
5294 assert(t->is_ConcurrentGC_thread(),
5295 "Should be CMS thread");
5296 } else {
5297 // We can be the CMS thread only if we are in a stop-world
5298 // phase of CMS collection.
5299 if (t->is_ConcurrentGC_thread()) {
5300 assert(_collectorState == InitialMarking ||
5301 _collectorState == FinalMarking,
5302 "Should be a stop-world phase");
5303 // The CMS thread should be holding the CMS_token.
5304 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5305 "Potential interference with concurrently "
5306 "executing VM thread");
5307 }
5308 }
5309 }
5310 #endif
5311
5312 void CMSCollector::sweep() {
5313 assert(_collectorState == Sweeping, "just checking");
5314 check_correct_thread_executing();
5315 verify_work_stacks_empty();
5316 verify_overflow_empty();
5317 increment_sweep_count();
5318 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
5319
5320 _inter_sweep_timer.stop();
5321 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5322
5323 assert(!_intra_sweep_timer.is_active(), "Should not be active");
5324 _intra_sweep_timer.reset();
5325 _intra_sweep_timer.start();
5326 {
5327 GCTraceCPUTime tcpu;
5328 CMSPhaseAccounting pa(this, "Concurrent Sweep");
5329 // First sweep the old gen
5330 {
5331 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5332 bitMapLock());
5333 sweepWork(_cmsGen);
5334 }
5335
5336 // Update Universe::_heap_*_at_gc figures.
5337 // We need all the free list locks to make the abstract state
5338 // transition from Sweeping to Resetting. See detailed note
5339 // further below.
5340 {
5341 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5342 // Update heap occupancy information which is used as
5343 // input to soft ref clearing policy at the next gc.
5344 Universe::update_heap_info_at_gc();
5345 _collectorState = Resizing;
5346 }
5347 }
5348 verify_work_stacks_empty();
5349 verify_overflow_empty();
5350
5351 if (should_unload_classes()) {
5352 // Delay purge to the beginning of the next safepoint. Metaspace::contains
5353 // requires that the virtual spaces are stable and not deleted.
5354 ClassLoaderDataGraph::set_should_purge(true);
5355 }
5356
5357 _intra_sweep_timer.stop();
5358 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5359
5360 _inter_sweep_timer.reset();
5361 _inter_sweep_timer.start();
5362
5363 // We need to use a monotonically non-decreasing time in ms
5364 // or we will see time-warp warnings and os::javaTimeMillis()
5365 // does not guarantee monotonicity.
5366 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5367 update_time_of_last_gc(now);
5368
5369 // NOTE on abstract state transitions:
5370 // Mutators allocate-live and/or mark the mod-union table dirty
5371 // based on the state of the collection. The former is done in
5372 // the interval [Marking, Sweeping] and the latter in the interval
5373 // [Marking, Sweeping). Thus the transitions into the Marking state
5374 // and out of the Sweeping state must be synchronously visible
5375 // globally to the mutators.
5376 // The transition into the Marking state happens with the world
5377 // stopped so the mutators will globally see it. Sweeping is
5378 // done asynchronously by the background collector so the transition
5379 // from the Sweeping state to the Resizing state must be done
5380 // under the freelistLock (as is the check for whether to
5381 // allocate-live and whether to dirty the mod-union table).
5382 assert(_collectorState == Resizing, "Change of collector state to"
5383 " Resizing must be done under the freelistLocks (plural)");
5384
5385 // Now that sweeping has been completed, we clear
5386 // the incremental_collection_failed flag,
5387 // thus inviting a younger gen collection to promote into
5388 // this generation. If such a promotion may still fail,
5389 // the flag will be set again when a young collection is
5390 // attempted.
5391 CMSHeap* heap = CMSHeap::heap();
5392 heap->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
5393 heap->update_full_collections_completed(_collection_count_start);
5394 }
5395
5396 // FIX ME!!! Looks like this belongs in CFLSpace, with
5397 // CMSGen merely delegating to it.
5398 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5399 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5400 HeapWord* minAddr = _cmsSpace->bottom();
5401 HeapWord* largestAddr =
5402 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5403 if (largestAddr == NULL) {
5404 // The dictionary appears to be empty. In this case
5405 // try to coalesce at the end of the heap.
5406 largestAddr = _cmsSpace->end();
5407 }
5408 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5409 size_t nearLargestOffset =
5410 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5411 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5412 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5413 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5414 }
5415
5416 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5417 return addr >= _cmsSpace->nearLargestChunk();
5418 }
5419
5420 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5421 return _cmsSpace->find_chunk_at_end();
5422 }
5423
5424 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation,
5425 bool full) {
5426 // If the young generation has been collected, gather any statistics
5427 // that are of interest at this point.
5428 bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation);
5429 if (!full && current_is_young) {
5430 // Gather statistics on the young generation collection.
5431 collector()->stats().record_gc0_end(used());
5432 }
5433 }
5434
5435 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) {
5436 // We iterate over the space(s) underlying this generation,
5437 // checking the mark bit map to see if the bits corresponding
5438 // to specific blocks are marked or not. Blocks that are
5439 // marked are live and are not swept up. All remaining blocks
5440 // are swept up, with coalescing on-the-fly as we sweep up
5441 // contiguous free and/or garbage blocks:
5442 // We need to ensure that the sweeper synchronizes with allocators
5443 // and stop-the-world collectors. In particular, the following
5444 // locks are used:
5445 // . CMS token: if this is held, a stop the world collection cannot occur
5446 // . freelistLock: if this is held no allocation can occur from this
5447 // generation by another thread
5448 // . bitMapLock: if this is held, no other thread can access or update
5449 //
5450
5451 // Note that we need to hold the freelistLock if we use
5452 // block iterate below; else the iterator might go awry if
5453 // a mutator (or promotion) causes block contents to change
5454 // (for instance if the allocator divvies up a block).
5455 // If we hold the free list lock, for all practical purposes
5456 // young generation GC's can't occur (they'll usually need to
5457 // promote), so we might as well prevent all young generation
5458 // GC's while we do a sweeping step. For the same reason, we might
5459 // as well take the bit map lock for the entire duration
5460
5461 // check that we hold the requisite locks
5462 assert(have_cms_token(), "Should hold cms token");
5463 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5464 assert_lock_strong(old_gen->freelistLock());
5465 assert_lock_strong(bitMapLock());
5466
5467 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5468 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
5469 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5470 _inter_sweep_estimate.padded_average(),
5471 _intra_sweep_estimate.padded_average());
5472 old_gen->setNearLargestChunk();
5473
5474 {
5475 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield);
5476 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5477 // We need to free-up/coalesce garbage/blocks from a
5478 // co-terminal free run. This is done in the SweepClosure
5479 // destructor; so, do not remove this scope, else the
5480 // end-of-sweep-census below will be off by a little bit.
5481 }
5482 old_gen->cmsSpace()->sweep_completed();
5483 old_gen->cmsSpace()->endSweepFLCensus(sweep_count());
5484 if (should_unload_classes()) { // unloaded classes this cycle,
5485 _concurrent_cycles_since_last_unload = 0; // ... reset count
5486 } else { // did not unload classes,
5487 _concurrent_cycles_since_last_unload++; // ... increment count
5488 }
5489 }
5490
5491 // Reset CMS data structures (for now just the marking bit map)
5492 // preparatory for the next cycle.
5493 void CMSCollector::reset_concurrent() {
5494 CMSTokenSyncWithLocks ts(true, bitMapLock());
5495
5496 // If the state is not "Resetting", the foreground thread
5497 // has done a collection and the resetting.
5498 if (_collectorState != Resetting) {
5499 assert(_collectorState == Idling, "The state should only change"
5500 " because the foreground collector has finished the collection");
5501 return;
5502 }
5503
5504 {
5505 // Clear the mark bitmap (no grey objects to start with)
5506 // for the next cycle.
5507 GCTraceCPUTime tcpu;
5508 CMSPhaseAccounting cmspa(this, "Concurrent Reset");
5509
5510 HeapWord* curAddr = _markBitMap.startWord();
5511 while (curAddr < _markBitMap.endWord()) {
5512 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
5513 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5514 _markBitMap.clear_large_range(chunk);
5515 if (ConcurrentMarkSweepThread::should_yield() &&
5516 !foregroundGCIsActive() &&
5517 CMSYield) {
5518 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5519 "CMS thread should hold CMS token");
5520 assert_lock_strong(bitMapLock());
5521 bitMapLock()->unlock();
5522 ConcurrentMarkSweepThread::desynchronize(true);
5523 stopTimer();
5524 incrementYields();
5525
5526 // See the comment in coordinator_yield()
5527 for (unsigned i = 0; i < CMSYieldSleepCount &&
5528 ConcurrentMarkSweepThread::should_yield() &&
5529 !CMSCollector::foregroundGCIsActive(); ++i) {
5530 os::sleep(Thread::current(), 1, false);
5531 }
5532
5533 ConcurrentMarkSweepThread::synchronize(true);
5534 bitMapLock()->lock_without_safepoint_check();
5535 startTimer();
5536 }
5537 curAddr = chunk.end();
5538 }
5539 // A successful mostly concurrent collection has been done.
5540 // Because only the full (i.e., concurrent mode failure) collections
5541 // are being measured for gc overhead limits, clean the "near" flag
5542 // and count.
5543 size_policy()->reset_gc_overhead_limit_count();
5544 _collectorState = Idling;
5545 }
5546
5547 register_gc_end();
5548 }
5549
5550 // Same as above but for STW paths
5551 void CMSCollector::reset_stw() {
5552 // already have the lock
5553 assert(_collectorState == Resetting, "just checking");
5554 assert_lock_strong(bitMapLock());
5555 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id());
5556 _markBitMap.clear_all();
5557 _collectorState = Idling;
5558 register_gc_end();
5559 }
5560
5561 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5562 GCTraceCPUTime tcpu;
5563 TraceCollectorStats tcs_cgc(cgc_counters());
5564
5565 switch (op) {
5566 case CMS_op_checkpointRootsInitial: {
5567 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true);
5568 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT);
5569 checkpointRootsInitial();
5570 break;
5571 }
5572 case CMS_op_checkpointRootsFinal: {
5573 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true);
5574 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT);
5575 checkpointRootsFinal();
5576 break;
5577 }
5578 default:
5579 fatal("No such CMS_op");
5580 }
5581 }
5582
5583 #ifndef PRODUCT
5584 size_t const CMSCollector::skip_header_HeapWords() {
5585 return FreeChunk::header_size();
5586 }
5587
5588 // Try and collect here conditions that should hold when
5589 // CMS thread is exiting. The idea is that the foreground GC
5590 // thread should not be blocked if it wants to terminate
5591 // the CMS thread and yet continue to run the VM for a while
5592 // after that.
5593 void CMSCollector::verify_ok_to_terminate() const {
5594 assert(Thread::current()->is_ConcurrentGC_thread(),
5595 "should be called by CMS thread");
5596 assert(!_foregroundGCShouldWait, "should be false");
5597 // We could check here that all the various low-level locks
5598 // are not held by the CMS thread, but that is overkill; see
5599 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5600 // is checked.
5601 }
5602 #endif
5603
5604 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5605 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5606 "missing Printezis mark?");
5607 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5608 size_t size = pointer_delta(nextOneAddr + 1, addr);
5609 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5610 "alignment problem");
5611 assert(size >= 3, "Necessary for Printezis marks to work");
5612 return size;
5613 }
5614
5615 // A variant of the above (block_size_using_printezis_bits()) except
5616 // that we return 0 if the P-bits are not yet set.
5617 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5618 if (_markBitMap.isMarked(addr + 1)) {
5619 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5620 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5621 size_t size = pointer_delta(nextOneAddr + 1, addr);
5622 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5623 "alignment problem");
5624 assert(size >= 3, "Necessary for Printezis marks to work");
5625 return size;
5626 }
5627 return 0;
5628 }
5629
5630 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5631 size_t sz = 0;
5632 oop p = (oop)addr;
5633 if (p->klass_or_null_acquire() != NULL) {
5634 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5635 } else {
5636 sz = block_size_using_printezis_bits(addr);
5637 }
5638 assert(sz > 0, "size must be nonzero");
5639 HeapWord* next_block = addr + sz;
5640 HeapWord* next_card = align_up(next_block, CardTableModRefBS::card_size);
5641 assert(align_down((uintptr_t)addr, CardTableModRefBS::card_size) <
5642 align_down((uintptr_t)next_card, CardTableModRefBS::card_size),
5643 "must be different cards");
5644 return next_card;
5645 }
5646
5647
5648 // CMS Bit Map Wrapper /////////////////////////////////////////
5649
5650 // Construct a CMS bit map infrastructure, but don't create the
5651 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5652 // further below.
5653 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5654 _bm(),
5655 _shifter(shifter),
5656 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5657 Monitor::_safepoint_check_sometimes) : NULL)
5658 {
5659 _bmStartWord = 0;
5660 _bmWordSize = 0;
5661 }
5662
5663 bool CMSBitMap::allocate(MemRegion mr) {
5664 _bmStartWord = mr.start();
5665 _bmWordSize = mr.word_size();
5666 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5667 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5668 if (!brs.is_reserved()) {
5669 log_warning(gc)("CMS bit map allocation failure");
5670 return false;
5671 }
5672 // For now we'll just commit all of the bit map up front.
5673 // Later on we'll try to be more parsimonious with swap.
5674 if (!_virtual_space.initialize(brs, brs.size())) {
5675 log_warning(gc)("CMS bit map backing store failure");
5676 return false;
5677 }
5678 assert(_virtual_space.committed_size() == brs.size(),
5679 "didn't reserve backing store for all of CMS bit map?");
5680 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5681 _bmWordSize, "inconsistency in bit map sizing");
5682 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter);
5683
5684 // bm.clear(); // can we rely on getting zero'd memory? verify below
5685 assert(isAllClear(),
5686 "Expected zero'd memory from ReservedSpace constructor");
5687 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5688 "consistency check");
5689 return true;
5690 }
5691
5692 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5693 HeapWord *next_addr, *end_addr, *last_addr;
5694 assert_locked();
5695 assert(covers(mr), "out-of-range error");
5696 // XXX assert that start and end are appropriately aligned
5697 for (next_addr = mr.start(), end_addr = mr.end();
5698 next_addr < end_addr; next_addr = last_addr) {
5699 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5700 last_addr = dirty_region.end();
5701 if (!dirty_region.is_empty()) {
5702 cl->do_MemRegion(dirty_region);
5703 } else {
5704 assert(last_addr == end_addr, "program logic");
5705 return;
5706 }
5707 }
5708 }
5709
5710 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5711 _bm.print_on_error(st, prefix);
5712 }
5713
5714 #ifndef PRODUCT
5715 void CMSBitMap::assert_locked() const {
5716 CMSLockVerifier::assert_locked(lock());
5717 }
5718
5719 bool CMSBitMap::covers(MemRegion mr) const {
5720 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5721 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5722 "size inconsistency");
5723 return (mr.start() >= _bmStartWord) &&
5724 (mr.end() <= endWord());
5725 }
5726
5727 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5728 return (start >= _bmStartWord && (start + size) <= endWord());
5729 }
5730
5731 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5732 // verify that there are no 1 bits in the interval [left, right)
5733 FalseBitMapClosure falseBitMapClosure;
5734 iterate(&falseBitMapClosure, left, right);
5735 }
5736
5737 void CMSBitMap::region_invariant(MemRegion mr)
5738 {
5739 assert_locked();
5740 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5741 assert(!mr.is_empty(), "unexpected empty region");
5742 assert(covers(mr), "mr should be covered by bit map");
5743 // convert address range into offset range
5744 size_t start_ofs = heapWordToOffset(mr.start());
5745 // Make sure that end() is appropriately aligned
5746 assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))),
5747 "Misaligned mr.end()");
5748 size_t end_ofs = heapWordToOffset(mr.end());
5749 assert(end_ofs > start_ofs, "Should mark at least one bit");
5750 }
5751
5752 #endif
5753
5754 bool CMSMarkStack::allocate(size_t size) {
5755 // allocate a stack of the requisite depth
5756 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5757 size * sizeof(oop)));
5758 if (!rs.is_reserved()) {
5759 log_warning(gc)("CMSMarkStack allocation failure");
5760 return false;
5761 }
5762 if (!_virtual_space.initialize(rs, rs.size())) {
5763 log_warning(gc)("CMSMarkStack backing store failure");
5764 return false;
5765 }
5766 assert(_virtual_space.committed_size() == rs.size(),
5767 "didn't reserve backing store for all of CMS stack?");
5768 _base = (oop*)(_virtual_space.low());
5769 _index = 0;
5770 _capacity = size;
5771 NOT_PRODUCT(_max_depth = 0);
5772 return true;
5773 }
5774
5775 // XXX FIX ME !!! In the MT case we come in here holding a
5776 // leaf lock. For printing we need to take a further lock
5777 // which has lower rank. We need to recalibrate the two
5778 // lock-ranks involved in order to be able to print the
5779 // messages below. (Or defer the printing to the caller.
5780 // For now we take the expedient path of just disabling the
5781 // messages for the problematic case.)
5782 void CMSMarkStack::expand() {
5783 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
5784 if (_capacity == MarkStackSizeMax) {
5785 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) {
5786 // We print a warning message only once per CMS cycle.
5787 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit");
5788 }
5789 return;
5790 }
5791 // Double capacity if possible
5792 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
5793 // Do not give up existing stack until we have managed to
5794 // get the double capacity that we desired.
5795 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5796 new_capacity * sizeof(oop)));
5797 if (rs.is_reserved()) {
5798 // Release the backing store associated with old stack
5799 _virtual_space.release();
5800 // Reinitialize virtual space for new stack
5801 if (!_virtual_space.initialize(rs, rs.size())) {
5802 fatal("Not enough swap for expanded marking stack");
5803 }
5804 _base = (oop*)(_virtual_space.low());
5805 _index = 0;
5806 _capacity = new_capacity;
5807 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) {
5808 // Failed to double capacity, continue;
5809 // we print a detail message only once per CMS cycle.
5810 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
5811 _capacity / K, new_capacity / K);
5812 }
5813 }
5814
5815
5816 // Closures
5817 // XXX: there seems to be a lot of code duplication here;
5818 // should refactor and consolidate common code.
5819
5820 // This closure is used to mark refs into the CMS generation in
5821 // the CMS bit map. Called at the first checkpoint. This closure
5822 // assumes that we do not need to re-mark dirty cards; if the CMS
5823 // generation on which this is used is not an oldest
5824 // generation then this will lose younger_gen cards!
5825
5826 MarkRefsIntoClosure::MarkRefsIntoClosure(
5827 MemRegion span, CMSBitMap* bitMap):
5828 _span(span),
5829 _bitMap(bitMap)
5830 {
5831 assert(ref_processor() == NULL, "deliberately left NULL");
5832 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5833 }
5834
5835 void MarkRefsIntoClosure::do_oop(oop obj) {
5836 // if p points into _span, then mark corresponding bit in _markBitMap
5837 assert(oopDesc::is_oop(obj), "expected an oop");
5838 HeapWord* addr = (HeapWord*)obj;
5839 if (_span.contains(addr)) {
5840 // this should be made more efficient
5841 _bitMap->mark(addr);
5842 }
5843 }
5844
5845 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
5846 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
5847
5848 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure(
5849 MemRegion span, CMSBitMap* bitMap):
5850 _span(span),
5851 _bitMap(bitMap)
5852 {
5853 assert(ref_processor() == NULL, "deliberately left NULL");
5854 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5855 }
5856
5857 void ParMarkRefsIntoClosure::do_oop(oop obj) {
5858 // if p points into _span, then mark corresponding bit in _markBitMap
5859 assert(oopDesc::is_oop(obj), "expected an oop");
5860 HeapWord* addr = (HeapWord*)obj;
5861 if (_span.contains(addr)) {
5862 // this should be made more efficient
5863 _bitMap->par_mark(addr);
5864 }
5865 }
5866
5867 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); }
5868 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); }
5869
5870 // A variant of the above, used for CMS marking verification.
5871 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
5872 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
5873 _span(span),
5874 _verification_bm(verification_bm),
5875 _cms_bm(cms_bm)
5876 {
5877 assert(ref_processor() == NULL, "deliberately left NULL");
5878 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
5879 }
5880
5881 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
5882 // if p points into _span, then mark corresponding bit in _markBitMap
5883 assert(oopDesc::is_oop(obj), "expected an oop");
5884 HeapWord* addr = (HeapWord*)obj;
5885 if (_span.contains(addr)) {
5886 _verification_bm->mark(addr);
5887 if (!_cms_bm->isMarked(addr)) {
5888 Log(gc, verify) log;
5889 ResourceMark rm;
5890 LogStream ls(log.error());
5891 oop(addr)->print_on(&ls);
5892 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
5893 fatal("... aborting");
5894 }
5895 }
5896 }
5897
5898 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5899 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5900
5901 //////////////////////////////////////////////////
5902 // MarkRefsIntoAndScanClosure
5903 //////////////////////////////////////////////////
5904
5905 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
5906 ReferenceProcessor* rp,
5907 CMSBitMap* bit_map,
5908 CMSBitMap* mod_union_table,
5909 CMSMarkStack* mark_stack,
5910 CMSCollector* collector,
5911 bool should_yield,
5912 bool concurrent_precleaning):
5913 _collector(collector),
5914 _span(span),
5915 _bit_map(bit_map),
5916 _mark_stack(mark_stack),
5917 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
5918 mark_stack, concurrent_precleaning),
5919 _yield(should_yield),
5920 _concurrent_precleaning(concurrent_precleaning),
5921 _freelistLock(NULL)
5922 {
5923 // FIXME: Should initialize in base class constructor.
5924 assert(rp != NULL, "ref_processor shouldn't be NULL");
5925 set_ref_processor_internal(rp);
5926 }
5927
5928 // This closure is used to mark refs into the CMS generation at the
5929 // second (final) checkpoint, and to scan and transitively follow
5930 // the unmarked oops. It is also used during the concurrent precleaning
5931 // phase while scanning objects on dirty cards in the CMS generation.
5932 // The marks are made in the marking bit map and the marking stack is
5933 // used for keeping the (newly) grey objects during the scan.
5934 // The parallel version (Par_...) appears further below.
5935 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
5936 if (obj != NULL) {
5937 assert(oopDesc::is_oop(obj), "expected an oop");
5938 HeapWord* addr = (HeapWord*)obj;
5939 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
5940 assert(_collector->overflow_list_is_empty(),
5941 "overflow list should be empty");
5942 if (_span.contains(addr) &&
5943 !_bit_map->isMarked(addr)) {
5944 // mark bit map (object is now grey)
5945 _bit_map->mark(addr);
5946 // push on marking stack (stack should be empty), and drain the
5947 // stack by applying this closure to the oops in the oops popped
5948 // from the stack (i.e. blacken the grey objects)
5949 bool res = _mark_stack->push(obj);
5950 assert(res, "Should have space to push on empty stack");
5951 do {
5952 oop new_oop = _mark_stack->pop();
5953 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
5954 assert(_bit_map->isMarked((HeapWord*)new_oop),
5955 "only grey objects on this stack");
5956 // iterate over the oops in this oop, marking and pushing
5957 // the ones in CMS heap (i.e. in _span).
5958 new_oop->oop_iterate(&_pushAndMarkClosure);
5959 // check if it's time to yield
5960 do_yield_check();
5961 } while (!_mark_stack->isEmpty() ||
5962 (!_concurrent_precleaning && take_from_overflow_list()));
5963 // if marking stack is empty, and we are not doing this
5964 // during precleaning, then check the overflow list
5965 }
5966 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
5967 assert(_collector->overflow_list_is_empty(),
5968 "overflow list was drained above");
5969
5970 assert(_collector->no_preserved_marks(),
5971 "All preserved marks should have been restored above");
5972 }
5973 }
5974
5975 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5976 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5977
5978 void MarkRefsIntoAndScanClosure::do_yield_work() {
5979 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5980 "CMS thread should hold CMS token");
5981 assert_lock_strong(_freelistLock);
5982 assert_lock_strong(_bit_map->lock());
5983 // relinquish the free_list_lock and bitMaplock()
5984 _bit_map->lock()->unlock();
5985 _freelistLock->unlock();
5986 ConcurrentMarkSweepThread::desynchronize(true);
5987 _collector->stopTimer();
5988 _collector->incrementYields();
5989
5990 // See the comment in coordinator_yield()
5991 for (unsigned i = 0;
5992 i < CMSYieldSleepCount &&
5993 ConcurrentMarkSweepThread::should_yield() &&
5994 !CMSCollector::foregroundGCIsActive();
5995 ++i) {
5996 os::sleep(Thread::current(), 1, false);
5997 }
5998
5999 ConcurrentMarkSweepThread::synchronize(true);
6000 _freelistLock->lock_without_safepoint_check();
6001 _bit_map->lock()->lock_without_safepoint_check();
6002 _collector->startTimer();
6003 }
6004
6005 ///////////////////////////////////////////////////////////
6006 // ParMarkRefsIntoAndScanClosure: a parallel version of
6007 // MarkRefsIntoAndScanClosure
6008 ///////////////////////////////////////////////////////////
6009 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure(
6010 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6011 CMSBitMap* bit_map, OopTaskQueue* work_queue):
6012 _span(span),
6013 _bit_map(bit_map),
6014 _work_queue(work_queue),
6015 _low_water_mark(MIN2((work_queue->max_elems()/4),
6016 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6017 _parPushAndMarkClosure(collector, span, rp, bit_map, work_queue)
6018 {
6019 // FIXME: Should initialize in base class constructor.
6020 assert(rp != NULL, "ref_processor shouldn't be NULL");
6021 set_ref_processor_internal(rp);
6022 }
6023
6024 // This closure is used to mark refs into the CMS generation at the
6025 // second (final) checkpoint, and to scan and transitively follow
6026 // the unmarked oops. The marks are made in the marking bit map and
6027 // the work_queue is used for keeping the (newly) grey objects during
6028 // the scan phase whence they are also available for stealing by parallel
6029 // threads. Since the marking bit map is shared, updates are
6030 // synchronized (via CAS).
6031 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) {
6032 if (obj != NULL) {
6033 // Ignore mark word because this could be an already marked oop
6034 // that may be chained at the end of the overflow list.
6035 assert(oopDesc::is_oop(obj, true), "expected an oop");
6036 HeapWord* addr = (HeapWord*)obj;
6037 if (_span.contains(addr) &&
6038 !_bit_map->isMarked(addr)) {
6039 // mark bit map (object will become grey):
6040 // It is possible for several threads to be
6041 // trying to "claim" this object concurrently;
6042 // the unique thread that succeeds in marking the
6043 // object first will do the subsequent push on
6044 // to the work queue (or overflow list).
6045 if (_bit_map->par_mark(addr)) {
6046 // push on work_queue (which may not be empty), and trim the
6047 // queue to an appropriate length by applying this closure to
6048 // the oops in the oops popped from the stack (i.e. blacken the
6049 // grey objects)
6050 bool res = _work_queue->push(obj);
6051 assert(res, "Low water mark should be less than capacity?");
6052 trim_queue(_low_water_mark);
6053 } // Else, another thread claimed the object
6054 }
6055 }
6056 }
6057
6058 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6059 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6060
6061 // This closure is used to rescan the marked objects on the dirty cards
6062 // in the mod union table and the card table proper.
6063 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6064 oop p, MemRegion mr) {
6065
6066 size_t size = 0;
6067 HeapWord* addr = (HeapWord*)p;
6068 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6069 assert(_span.contains(addr), "we are scanning the CMS generation");
6070 // check if it's time to yield
6071 if (do_yield_check()) {
6072 // We yielded for some foreground stop-world work,
6073 // and we have been asked to abort this ongoing preclean cycle.
6074 return 0;
6075 }
6076 if (_bitMap->isMarked(addr)) {
6077 // it's marked; is it potentially uninitialized?
6078 if (p->klass_or_null_acquire() != NULL) {
6079 // an initialized object; ignore mark word in verification below
6080 // since we are running concurrent with mutators
6081 assert(oopDesc::is_oop(p, true), "should be an oop");
6082 if (p->is_objArray()) {
6083 // objArrays are precisely marked; restrict scanning
6084 // to dirty cards only.
6085 size = CompactibleFreeListSpace::adjustObjectSize(
6086 p->oop_iterate_size(_scanningClosure, mr));
6087 } else {
6088 // A non-array may have been imprecisely marked; we need
6089 // to scan object in its entirety.
6090 size = CompactibleFreeListSpace::adjustObjectSize(
6091 p->oop_iterate_size(_scanningClosure));
6092 }
6093 #ifdef ASSERT
6094 size_t direct_size =
6095 CompactibleFreeListSpace::adjustObjectSize(p->size());
6096 assert(size == direct_size, "Inconsistency in size");
6097 assert(size >= 3, "Necessary for Printezis marks to work");
6098 HeapWord* start_pbit = addr + 1;
6099 HeapWord* end_pbit = addr + size - 1;
6100 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit),
6101 "inconsistent Printezis mark");
6102 // Verify inner mark bits (between Printezis bits) are clear,
6103 // but don't repeat if there are multiple dirty regions for
6104 // the same object, to avoid potential O(N^2) performance.
6105 if (addr != _last_scanned_object) {
6106 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit);
6107 _last_scanned_object = addr;
6108 }
6109 #endif // ASSERT
6110 } else {
6111 // An uninitialized object.
6112 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6113 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6114 size = pointer_delta(nextOneAddr + 1, addr);
6115 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6116 "alignment problem");
6117 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6118 // will dirty the card when the klass pointer is installed in the
6119 // object (signaling the completion of initialization).
6120 }
6121 } else {
6122 // Either a not yet marked object or an uninitialized object
6123 if (p->klass_or_null_acquire() == NULL) {
6124 // An uninitialized object, skip to the next card, since
6125 // we may not be able to read its P-bits yet.
6126 assert(size == 0, "Initial value");
6127 } else {
6128 // An object not (yet) reached by marking: we merely need to
6129 // compute its size so as to go look at the next block.
6130 assert(oopDesc::is_oop(p, true), "should be an oop");
6131 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6132 }
6133 }
6134 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6135 return size;
6136 }
6137
6138 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6139 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6140 "CMS thread should hold CMS token");
6141 assert_lock_strong(_freelistLock);
6142 assert_lock_strong(_bitMap->lock());
6143 // relinquish the free_list_lock and bitMaplock()
6144 _bitMap->lock()->unlock();
6145 _freelistLock->unlock();
6146 ConcurrentMarkSweepThread::desynchronize(true);
6147 _collector->stopTimer();
6148 _collector->incrementYields();
6149
6150 // See the comment in coordinator_yield()
6151 for (unsigned i = 0; i < CMSYieldSleepCount &&
6152 ConcurrentMarkSweepThread::should_yield() &&
6153 !CMSCollector::foregroundGCIsActive(); ++i) {
6154 os::sleep(Thread::current(), 1, false);
6155 }
6156
6157 ConcurrentMarkSweepThread::synchronize(true);
6158 _freelistLock->lock_without_safepoint_check();
6159 _bitMap->lock()->lock_without_safepoint_check();
6160 _collector->startTimer();
6161 }
6162
6163
6164 //////////////////////////////////////////////////////////////////
6165 // SurvivorSpacePrecleanClosure
6166 //////////////////////////////////////////////////////////////////
6167 // This (single-threaded) closure is used to preclean the oops in
6168 // the survivor spaces.
6169 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6170
6171 HeapWord* addr = (HeapWord*)p;
6172 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6173 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6174 assert(p->klass_or_null() != NULL, "object should be initialized");
6175 // an initialized object; ignore mark word in verification below
6176 // since we are running concurrent with mutators
6177 assert(oopDesc::is_oop(p, true), "should be an oop");
6178 // Note that we do not yield while we iterate over
6179 // the interior oops of p, pushing the relevant ones
6180 // on our marking stack.
6181 size_t size = p->oop_iterate_size(_scanning_closure);
6182 do_yield_check();
6183 // Observe that below, we do not abandon the preclean
6184 // phase as soon as we should; rather we empty the
6185 // marking stack before returning. This is to satisfy
6186 // some existing assertions. In general, it may be a
6187 // good idea to abort immediately and complete the marking
6188 // from the grey objects at a later time.
6189 while (!_mark_stack->isEmpty()) {
6190 oop new_oop = _mark_stack->pop();
6191 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
6192 assert(_bit_map->isMarked((HeapWord*)new_oop),
6193 "only grey objects on this stack");
6194 // iterate over the oops in this oop, marking and pushing
6195 // the ones in CMS heap (i.e. in _span).
6196 new_oop->oop_iterate(_scanning_closure);
6197 // check if it's time to yield
6198 do_yield_check();
6199 }
6200 unsigned int after_count =
6201 CMSHeap::heap()->total_collections();
6202 bool abort = (_before_count != after_count) ||
6203 _collector->should_abort_preclean();
6204 return abort ? 0 : size;
6205 }
6206
6207 void SurvivorSpacePrecleanClosure::do_yield_work() {
6208 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6209 "CMS thread should hold CMS token");
6210 assert_lock_strong(_bit_map->lock());
6211 // Relinquish the bit map lock
6212 _bit_map->lock()->unlock();
6213 ConcurrentMarkSweepThread::desynchronize(true);
6214 _collector->stopTimer();
6215 _collector->incrementYields();
6216
6217 // See the comment in coordinator_yield()
6218 for (unsigned i = 0; i < CMSYieldSleepCount &&
6219 ConcurrentMarkSweepThread::should_yield() &&
6220 !CMSCollector::foregroundGCIsActive(); ++i) {
6221 os::sleep(Thread::current(), 1, false);
6222 }
6223
6224 ConcurrentMarkSweepThread::synchronize(true);
6225 _bit_map->lock()->lock_without_safepoint_check();
6226 _collector->startTimer();
6227 }
6228
6229 // This closure is used to rescan the marked objects on the dirty cards
6230 // in the mod union table and the card table proper. In the parallel
6231 // case, although the bitMap is shared, we do a single read so the
6232 // isMarked() query is "safe".
6233 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6234 // Ignore mark word because we are running concurrent with mutators
6235 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p));
6236 HeapWord* addr = (HeapWord*)p;
6237 assert(_span.contains(addr), "we are scanning the CMS generation");
6238 bool is_obj_array = false;
6239 #ifdef ASSERT
6240 if (!_parallel) {
6241 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6242 assert(_collector->overflow_list_is_empty(),
6243 "overflow list should be empty");
6244
6245 }
6246 #endif // ASSERT
6247 if (_bit_map->isMarked(addr)) {
6248 // Obj arrays are precisely marked, non-arrays are not;
6249 // so we scan objArrays precisely and non-arrays in their
6250 // entirety.
6251 if (p->is_objArray()) {
6252 is_obj_array = true;
6253 if (_parallel) {
6254 p->oop_iterate(_par_scan_closure, mr);
6255 } else {
6256 p->oop_iterate(_scan_closure, mr);
6257 }
6258 } else {
6259 if (_parallel) {
6260 p->oop_iterate(_par_scan_closure);
6261 } else {
6262 p->oop_iterate(_scan_closure);
6263 }
6264 }
6265 }
6266 #ifdef ASSERT
6267 if (!_parallel) {
6268 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6269 assert(_collector->overflow_list_is_empty(),
6270 "overflow list should be empty");
6271
6272 }
6273 #endif // ASSERT
6274 return is_obj_array;
6275 }
6276
6277 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6278 MemRegion span,
6279 CMSBitMap* bitMap, CMSMarkStack* markStack,
6280 bool should_yield, bool verifying):
6281 _collector(collector),
6282 _span(span),
6283 _bitMap(bitMap),
6284 _mut(&collector->_modUnionTable),
6285 _markStack(markStack),
6286 _yield(should_yield),
6287 _skipBits(0)
6288 {
6289 assert(_markStack->isEmpty(), "stack should be empty");
6290 _finger = _bitMap->startWord();
6291 _threshold = _finger;
6292 assert(_collector->_restart_addr == NULL, "Sanity check");
6293 assert(_span.contains(_finger), "Out of bounds _finger?");
6294 DEBUG_ONLY(_verifying = verifying;)
6295 }
6296
6297 void MarkFromRootsClosure::reset(HeapWord* addr) {
6298 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6299 assert(_span.contains(addr), "Out of bounds _finger?");
6300 _finger = addr;
6301 _threshold = align_up(_finger, CardTableModRefBS::card_size);
6302 }
6303
6304 // Should revisit to see if this should be restructured for
6305 // greater efficiency.
6306 bool MarkFromRootsClosure::do_bit(size_t offset) {
6307 if (_skipBits > 0) {
6308 _skipBits--;
6309 return true;
6310 }
6311 // convert offset into a HeapWord*
6312 HeapWord* addr = _bitMap->startWord() + offset;
6313 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6314 "address out of range");
6315 assert(_bitMap->isMarked(addr), "tautology");
6316 if (_bitMap->isMarked(addr+1)) {
6317 // this is an allocated but not yet initialized object
6318 assert(_skipBits == 0, "tautology");
6319 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6320 oop p = oop(addr);
6321 if (p->klass_or_null_acquire() == NULL) {
6322 DEBUG_ONLY(if (!_verifying) {)
6323 // We re-dirty the cards on which this object lies and increase
6324 // the _threshold so that we'll come back to scan this object
6325 // during the preclean or remark phase. (CMSCleanOnEnter)
6326 if (CMSCleanOnEnter) {
6327 size_t sz = _collector->block_size_using_printezis_bits(addr);
6328 HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size);
6329 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6330 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6331 // Bump _threshold to end_card_addr; note that
6332 // _threshold cannot possibly exceed end_card_addr, anyhow.
6333 // This prevents future clearing of the card as the scan proceeds
6334 // to the right.
6335 assert(_threshold <= end_card_addr,
6336 "Because we are just scanning into this object");
6337 if (_threshold < end_card_addr) {
6338 _threshold = end_card_addr;
6339 }
6340 if (p->klass_or_null_acquire() != NULL) {
6341 // Redirty the range of cards...
6342 _mut->mark_range(redirty_range);
6343 } // ...else the setting of klass will dirty the card anyway.
6344 }
6345 DEBUG_ONLY(})
6346 return true;
6347 }
6348 }
6349 scanOopsInOop(addr);
6350 return true;
6351 }
6352
6353 // We take a break if we've been at this for a while,
6354 // so as to avoid monopolizing the locks involved.
6355 void MarkFromRootsClosure::do_yield_work() {
6356 // First give up the locks, then yield, then re-lock
6357 // We should probably use a constructor/destructor idiom to
6358 // do this unlock/lock or modify the MutexUnlocker class to
6359 // serve our purpose. XXX
6360 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6361 "CMS thread should hold CMS token");
6362 assert_lock_strong(_bitMap->lock());
6363 _bitMap->lock()->unlock();
6364 ConcurrentMarkSweepThread::desynchronize(true);
6365 _collector->stopTimer();
6366 _collector->incrementYields();
6367
6368 // See the comment in coordinator_yield()
6369 for (unsigned i = 0; i < CMSYieldSleepCount &&
6370 ConcurrentMarkSweepThread::should_yield() &&
6371 !CMSCollector::foregroundGCIsActive(); ++i) {
6372 os::sleep(Thread::current(), 1, false);
6373 }
6374
6375 ConcurrentMarkSweepThread::synchronize(true);
6376 _bitMap->lock()->lock_without_safepoint_check();
6377 _collector->startTimer();
6378 }
6379
6380 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6381 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6382 assert(_markStack->isEmpty(),
6383 "should drain stack to limit stack usage");
6384 // convert ptr to an oop preparatory to scanning
6385 oop obj = oop(ptr);
6386 // Ignore mark word in verification below, since we
6387 // may be running concurrent with mutators.
6388 assert(oopDesc::is_oop(obj, true), "should be an oop");
6389 assert(_finger <= ptr, "_finger runneth ahead");
6390 // advance the finger to right end of this object
6391 _finger = ptr + obj->size();
6392 assert(_finger > ptr, "we just incremented it above");
6393 // On large heaps, it may take us some time to get through
6394 // the marking phase. During
6395 // this time it's possible that a lot of mutations have
6396 // accumulated in the card table and the mod union table --
6397 // these mutation records are redundant until we have
6398 // actually traced into the corresponding card.
6399 // Here, we check whether advancing the finger would make
6400 // us cross into a new card, and if so clear corresponding
6401 // cards in the MUT (preclean them in the card-table in the
6402 // future).
6403
6404 DEBUG_ONLY(if (!_verifying) {)
6405 // The clean-on-enter optimization is disabled by default,
6406 // until we fix 6178663.
6407 if (CMSCleanOnEnter && (_finger > _threshold)) {
6408 // [_threshold, _finger) represents the interval
6409 // of cards to be cleared in MUT (or precleaned in card table).
6410 // The set of cards to be cleared is all those that overlap
6411 // with the interval [_threshold, _finger); note that
6412 // _threshold is always kept card-aligned but _finger isn't
6413 // always card-aligned.
6414 HeapWord* old_threshold = _threshold;
6415 assert(is_aligned(old_threshold, CardTableModRefBS::card_size),
6416 "_threshold should always be card-aligned");
6417 _threshold = align_up(_finger, CardTableModRefBS::card_size);
6418 MemRegion mr(old_threshold, _threshold);
6419 assert(!mr.is_empty(), "Control point invariant");
6420 assert(_span.contains(mr), "Should clear within span");
6421 _mut->clear_range(mr);
6422 }
6423 DEBUG_ONLY(})
6424 // Note: the finger doesn't advance while we drain
6425 // the stack below.
6426 PushOrMarkClosure pushOrMarkClosure(_collector,
6427 _span, _bitMap, _markStack,
6428 _finger, this);
6429 bool res = _markStack->push(obj);
6430 assert(res, "Empty non-zero size stack should have space for single push");
6431 while (!_markStack->isEmpty()) {
6432 oop new_oop = _markStack->pop();
6433 // Skip verifying header mark word below because we are
6434 // running concurrent with mutators.
6435 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop");
6436 // now scan this oop's oops
6437 new_oop->oop_iterate(&pushOrMarkClosure);
6438 do_yield_check();
6439 }
6440 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6441 }
6442
6443 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task,
6444 CMSCollector* collector, MemRegion span,
6445 CMSBitMap* bit_map,
6446 OopTaskQueue* work_queue,
6447 CMSMarkStack* overflow_stack):
6448 _collector(collector),
6449 _whole_span(collector->_span),
6450 _span(span),
6451 _bit_map(bit_map),
6452 _mut(&collector->_modUnionTable),
6453 _work_queue(work_queue),
6454 _overflow_stack(overflow_stack),
6455 _skip_bits(0),
6456 _task(task)
6457 {
6458 assert(_work_queue->size() == 0, "work_queue should be empty");
6459 _finger = span.start();
6460 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
6461 assert(_span.contains(_finger), "Out of bounds _finger?");
6462 }
6463
6464 // Should revisit to see if this should be restructured for
6465 // greater efficiency.
6466 bool ParMarkFromRootsClosure::do_bit(size_t offset) {
6467 if (_skip_bits > 0) {
6468 _skip_bits--;
6469 return true;
6470 }
6471 // convert offset into a HeapWord*
6472 HeapWord* addr = _bit_map->startWord() + offset;
6473 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6474 "address out of range");
6475 assert(_bit_map->isMarked(addr), "tautology");
6476 if (_bit_map->isMarked(addr+1)) {
6477 // this is an allocated object that might not yet be initialized
6478 assert(_skip_bits == 0, "tautology");
6479 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
6480 oop p = oop(addr);
6481 if (p->klass_or_null_acquire() == NULL) {
6482 // in the case of Clean-on-Enter optimization, redirty card
6483 // and avoid clearing card by increasing the threshold.
6484 return true;
6485 }
6486 }
6487 scan_oops_in_oop(addr);
6488 return true;
6489 }
6490
6491 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6492 assert(_bit_map->isMarked(ptr), "expected bit to be set");
6493 // Should we assert that our work queue is empty or
6494 // below some drain limit?
6495 assert(_work_queue->size() == 0,
6496 "should drain stack to limit stack usage");
6497 // convert ptr to an oop preparatory to scanning
6498 oop obj = oop(ptr);
6499 // Ignore mark word in verification below, since we
6500 // may be running concurrent with mutators.
6501 assert(oopDesc::is_oop(obj, true), "should be an oop");
6502 assert(_finger <= ptr, "_finger runneth ahead");
6503 // advance the finger to right end of this object
6504 _finger = ptr + obj->size();
6505 assert(_finger > ptr, "we just incremented it above");
6506 // On large heaps, it may take us some time to get through
6507 // the marking phase. During
6508 // this time it's possible that a lot of mutations have
6509 // accumulated in the card table and the mod union table --
6510 // these mutation records are redundant until we have
6511 // actually traced into the corresponding card.
6512 // Here, we check whether advancing the finger would make
6513 // us cross into a new card, and if so clear corresponding
6514 // cards in the MUT (preclean them in the card-table in the
6515 // future).
6516
6517 // The clean-on-enter optimization is disabled by default,
6518 // until we fix 6178663.
6519 if (CMSCleanOnEnter && (_finger > _threshold)) {
6520 // [_threshold, _finger) represents the interval
6521 // of cards to be cleared in MUT (or precleaned in card table).
6522 // The set of cards to be cleared is all those that overlap
6523 // with the interval [_threshold, _finger); note that
6524 // _threshold is always kept card-aligned but _finger isn't
6525 // always card-aligned.
6526 HeapWord* old_threshold = _threshold;
6527 assert(is_aligned(old_threshold, CardTableModRefBS::card_size),
6528 "_threshold should always be card-aligned");
6529 _threshold = align_up(_finger, CardTableModRefBS::card_size);
6530 MemRegion mr(old_threshold, _threshold);
6531 assert(!mr.is_empty(), "Control point invariant");
6532 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6533 _mut->clear_range(mr);
6534 }
6535
6536 // Note: the local finger doesn't advance while we drain
6537 // the stack below, but the global finger sure can and will.
6538 HeapWord* volatile* gfa = _task->global_finger_addr();
6539 ParPushOrMarkClosure pushOrMarkClosure(_collector,
6540 _span, _bit_map,
6541 _work_queue,
6542 _overflow_stack,
6543 _finger,
6544 gfa, this);
6545 bool res = _work_queue->push(obj); // overflow could occur here
6546 assert(res, "Will hold once we use workqueues");
6547 while (true) {
6548 oop new_oop;
6549 if (!_work_queue->pop_local(new_oop)) {
6550 // We emptied our work_queue; check if there's stuff that can
6551 // be gotten from the overflow stack.
6552 if (CMSConcMarkingTask::get_work_from_overflow_stack(
6553 _overflow_stack, _work_queue)) {
6554 do_yield_check();
6555 continue;
6556 } else { // done
6557 break;
6558 }
6559 }
6560 // Skip verifying header mark word below because we are
6561 // running concurrent with mutators.
6562 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop");
6563 // now scan this oop's oops
6564 new_oop->oop_iterate(&pushOrMarkClosure);
6565 do_yield_check();
6566 }
6567 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6568 }
6569
6570 // Yield in response to a request from VM Thread or
6571 // from mutators.
6572 void ParMarkFromRootsClosure::do_yield_work() {
6573 assert(_task != NULL, "sanity");
6574 _task->yield();
6575 }
6576
6577 // A variant of the above used for verifying CMS marking work.
6578 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6579 MemRegion span,
6580 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6581 CMSMarkStack* mark_stack):
6582 _collector(collector),
6583 _span(span),
6584 _verification_bm(verification_bm),
6585 _cms_bm(cms_bm),
6586 _mark_stack(mark_stack),
6587 _pam_verify_closure(collector, span, verification_bm, cms_bm,
6588 mark_stack)
6589 {
6590 assert(_mark_stack->isEmpty(), "stack should be empty");
6591 _finger = _verification_bm->startWord();
6592 assert(_collector->_restart_addr == NULL, "Sanity check");
6593 assert(_span.contains(_finger), "Out of bounds _finger?");
6594 }
6595
6596 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6597 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6598 assert(_span.contains(addr), "Out of bounds _finger?");
6599 _finger = addr;
6600 }
6601
6602 // Should revisit to see if this should be restructured for
6603 // greater efficiency.
6604 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6605 // convert offset into a HeapWord*
6606 HeapWord* addr = _verification_bm->startWord() + offset;
6607 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6608 "address out of range");
6609 assert(_verification_bm->isMarked(addr), "tautology");
6610 assert(_cms_bm->isMarked(addr), "tautology");
6611
6612 assert(_mark_stack->isEmpty(),
6613 "should drain stack to limit stack usage");
6614 // convert addr to an oop preparatory to scanning
6615 oop obj = oop(addr);
6616 assert(oopDesc::is_oop(obj), "should be an oop");
6617 assert(_finger <= addr, "_finger runneth ahead");
6618 // advance the finger to right end of this object
6619 _finger = addr + obj->size();
6620 assert(_finger > addr, "we just incremented it above");
6621 // Note: the finger doesn't advance while we drain
6622 // the stack below.
6623 bool res = _mark_stack->push(obj);
6624 assert(res, "Empty non-zero size stack should have space for single push");
6625 while (!_mark_stack->isEmpty()) {
6626 oop new_oop = _mark_stack->pop();
6627 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop");
6628 // now scan this oop's oops
6629 new_oop->oop_iterate(&_pam_verify_closure);
6630 }
6631 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6632 return true;
6633 }
6634
6635 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6636 CMSCollector* collector, MemRegion span,
6637 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6638 CMSMarkStack* mark_stack):
6639 MetadataAwareOopClosure(collector->ref_processor()),
6640 _collector(collector),
6641 _span(span),
6642 _verification_bm(verification_bm),
6643 _cms_bm(cms_bm),
6644 _mark_stack(mark_stack)
6645 { }
6646
6647 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6648 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6649
6650 // Upon stack overflow, we discard (part of) the stack,
6651 // remembering the least address amongst those discarded
6652 // in CMSCollector's _restart_address.
6653 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6654 // Remember the least grey address discarded
6655 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6656 _collector->lower_restart_addr(ra);
6657 _mark_stack->reset(); // discard stack contents
6658 _mark_stack->expand(); // expand the stack if possible
6659 }
6660
6661 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6662 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6663 HeapWord* addr = (HeapWord*)obj;
6664 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6665 // Oop lies in _span and isn't yet grey or black
6666 _verification_bm->mark(addr); // now grey
6667 if (!_cms_bm->isMarked(addr)) {
6668 Log(gc, verify) log;
6669 ResourceMark rm;
6670 LogStream ls(log.error());
6671 oop(addr)->print_on(&ls);
6672 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6673 fatal("... aborting");
6674 }
6675
6676 if (!_mark_stack->push(obj)) { // stack overflow
6677 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity());
6678 assert(_mark_stack->isFull(), "Else push should have succeeded");
6679 handle_stack_overflow(addr);
6680 }
6681 // anything including and to the right of _finger
6682 // will be scanned as we iterate over the remainder of the
6683 // bit map
6684 }
6685 }
6686
6687 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6688 MemRegion span,
6689 CMSBitMap* bitMap, CMSMarkStack* markStack,
6690 HeapWord* finger, MarkFromRootsClosure* parent) :
6691 MetadataAwareOopClosure(collector->ref_processor()),
6692 _collector(collector),
6693 _span(span),
6694 _bitMap(bitMap),
6695 _markStack(markStack),
6696 _finger(finger),
6697 _parent(parent)
6698 { }
6699
6700 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector,
6701 MemRegion span,
6702 CMSBitMap* bit_map,
6703 OopTaskQueue* work_queue,
6704 CMSMarkStack* overflow_stack,
6705 HeapWord* finger,
6706 HeapWord* volatile* global_finger_addr,
6707 ParMarkFromRootsClosure* parent) :
6708 MetadataAwareOopClosure(collector->ref_processor()),
6709 _collector(collector),
6710 _whole_span(collector->_span),
6711 _span(span),
6712 _bit_map(bit_map),
6713 _work_queue(work_queue),
6714 _overflow_stack(overflow_stack),
6715 _finger(finger),
6716 _global_finger_addr(global_finger_addr),
6717 _parent(parent)
6718 { }
6719
6720 // Assumes thread-safe access by callers, who are
6721 // responsible for mutual exclusion.
6722 void CMSCollector::lower_restart_addr(HeapWord* low) {
6723 assert(_span.contains(low), "Out of bounds addr");
6724 if (_restart_addr == NULL) {
6725 _restart_addr = low;
6726 } else {
6727 _restart_addr = MIN2(_restart_addr, low);
6728 }
6729 }
6730
6731 // Upon stack overflow, we discard (part of) the stack,
6732 // remembering the least address amongst those discarded
6733 // in CMSCollector's _restart_address.
6734 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6735 // Remember the least grey address discarded
6736 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6737 _collector->lower_restart_addr(ra);
6738 _markStack->reset(); // discard stack contents
6739 _markStack->expand(); // expand the stack if possible
6740 }
6741
6742 // Upon stack overflow, we discard (part of) the stack,
6743 // remembering the least address amongst those discarded
6744 // in CMSCollector's _restart_address.
6745 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6746 // We need to do this under a mutex to prevent other
6747 // workers from interfering with the work done below.
6748 MutexLockerEx ml(_overflow_stack->par_lock(),
6749 Mutex::_no_safepoint_check_flag);
6750 // Remember the least grey address discarded
6751 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
6752 _collector->lower_restart_addr(ra);
6753 _overflow_stack->reset(); // discard stack contents
6754 _overflow_stack->expand(); // expand the stack if possible
6755 }
6756
6757 void PushOrMarkClosure::do_oop(oop obj) {
6758 // Ignore mark word because we are running concurrent with mutators.
6759 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6760 HeapWord* addr = (HeapWord*)obj;
6761 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
6762 // Oop lies in _span and isn't yet grey or black
6763 _bitMap->mark(addr); // now grey
6764 if (addr < _finger) {
6765 // the bit map iteration has already either passed, or
6766 // sampled, this bit in the bit map; we'll need to
6767 // use the marking stack to scan this oop's oops.
6768 bool simulate_overflow = false;
6769 NOT_PRODUCT(
6770 if (CMSMarkStackOverflowALot &&
6771 _collector->simulate_overflow()) {
6772 // simulate a stack overflow
6773 simulate_overflow = true;
6774 }
6775 )
6776 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
6777 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity());
6778 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
6779 handle_stack_overflow(addr);
6780 }
6781 }
6782 // anything including and to the right of _finger
6783 // will be scanned as we iterate over the remainder of the
6784 // bit map
6785 do_yield_check();
6786 }
6787 }
6788
6789 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
6790 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
6791
6792 void ParPushOrMarkClosure::do_oop(oop obj) {
6793 // Ignore mark word because we are running concurrent with mutators.
6794 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6795 HeapWord* addr = (HeapWord*)obj;
6796 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
6797 // Oop lies in _span and isn't yet grey or black
6798 // We read the global_finger (volatile read) strictly after marking oop
6799 bool res = _bit_map->par_mark(addr); // now grey
6800 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
6801 // Should we push this marked oop on our stack?
6802 // -- if someone else marked it, nothing to do
6803 // -- if target oop is above global finger nothing to do
6804 // -- if target oop is in chunk and above local finger
6805 // then nothing to do
6806 // -- else push on work queue
6807 if ( !res // someone else marked it, they will deal with it
6808 || (addr >= *gfa) // will be scanned in a later task
6809 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
6810 return;
6811 }
6812 // the bit map iteration has already either passed, or
6813 // sampled, this bit in the bit map; we'll need to
6814 // use the marking stack to scan this oop's oops.
6815 bool simulate_overflow = false;
6816 NOT_PRODUCT(
6817 if (CMSMarkStackOverflowALot &&
6818 _collector->simulate_overflow()) {
6819 // simulate a stack overflow
6820 simulate_overflow = true;
6821 }
6822 )
6823 if (simulate_overflow ||
6824 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
6825 // stack overflow
6826 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
6827 // We cannot assert that the overflow stack is full because
6828 // it may have been emptied since.
6829 assert(simulate_overflow ||
6830 _work_queue->size() == _work_queue->max_elems(),
6831 "Else push should have succeeded");
6832 handle_stack_overflow(addr);
6833 }
6834 do_yield_check();
6835 }
6836 }
6837
6838 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); }
6839 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); }
6840
6841 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
6842 MemRegion span,
6843 ReferenceProcessor* rp,
6844 CMSBitMap* bit_map,
6845 CMSBitMap* mod_union_table,
6846 CMSMarkStack* mark_stack,
6847 bool concurrent_precleaning):
6848 MetadataAwareOopClosure(rp),
6849 _collector(collector),
6850 _span(span),
6851 _bit_map(bit_map),
6852 _mod_union_table(mod_union_table),
6853 _mark_stack(mark_stack),
6854 _concurrent_precleaning(concurrent_precleaning)
6855 {
6856 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL");
6857 }
6858
6859 // Grey object rescan during pre-cleaning and second checkpoint phases --
6860 // the non-parallel version (the parallel version appears further below.)
6861 void PushAndMarkClosure::do_oop(oop obj) {
6862 // Ignore mark word verification. If during concurrent precleaning,
6863 // the object monitor may be locked. If during the checkpoint
6864 // phases, the object may already have been reached by a different
6865 // path and may be at the end of the global overflow list (so
6866 // the mark word may be NULL).
6867 assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */),
6868 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6869 HeapWord* addr = (HeapWord*)obj;
6870 // Check if oop points into the CMS generation
6871 // and is not marked
6872 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6873 // a white object ...
6874 _bit_map->mark(addr); // ... now grey
6875 // push on the marking stack (grey set)
6876 bool simulate_overflow = false;
6877 NOT_PRODUCT(
6878 if (CMSMarkStackOverflowALot &&
6879 _collector->simulate_overflow()) {
6880 // simulate a stack overflow
6881 simulate_overflow = true;
6882 }
6883 )
6884 if (simulate_overflow || !_mark_stack->push(obj)) {
6885 if (_concurrent_precleaning) {
6886 // During precleaning we can just dirty the appropriate card(s)
6887 // in the mod union table, thus ensuring that the object remains
6888 // in the grey set and continue. In the case of object arrays
6889 // we need to dirty all of the cards that the object spans,
6890 // since the rescan of object arrays will be limited to the
6891 // dirty cards.
6892 // Note that no one can be interfering with us in this action
6893 // of dirtying the mod union table, so no locking or atomics
6894 // are required.
6895 if (obj->is_objArray()) {
6896 size_t sz = obj->size();
6897 HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size);
6898 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6899 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6900 _mod_union_table->mark_range(redirty_range);
6901 } else {
6902 _mod_union_table->mark(addr);
6903 }
6904 _collector->_ser_pmc_preclean_ovflw++;
6905 } else {
6906 // During the remark phase, we need to remember this oop
6907 // in the overflow list.
6908 _collector->push_on_overflow_list(obj);
6909 _collector->_ser_pmc_remark_ovflw++;
6910 }
6911 }
6912 }
6913 }
6914
6915 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector,
6916 MemRegion span,
6917 ReferenceProcessor* rp,
6918 CMSBitMap* bit_map,
6919 OopTaskQueue* work_queue):
6920 MetadataAwareOopClosure(rp),
6921 _collector(collector),
6922 _span(span),
6923 _bit_map(bit_map),
6924 _work_queue(work_queue)
6925 {
6926 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL");
6927 }
6928
6929 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
6930 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
6931
6932 // Grey object rescan during second checkpoint phase --
6933 // the parallel version.
6934 void ParPushAndMarkClosure::do_oop(oop obj) {
6935 // In the assert below, we ignore the mark word because
6936 // this oop may point to an already visited object that is
6937 // on the overflow stack (in which case the mark word has
6938 // been hijacked for chaining into the overflow stack --
6939 // if this is the last object in the overflow stack then
6940 // its mark word will be NULL). Because this object may
6941 // have been subsequently popped off the global overflow
6942 // stack, and the mark word possibly restored to the prototypical
6943 // value, by the time we get to examined this failing assert in
6944 // the debugger, is_oop_or_null(false) may subsequently start
6945 // to hold.
6946 assert(oopDesc::is_oop_or_null(obj, true),
6947 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6948 HeapWord* addr = (HeapWord*)obj;
6949 // Check if oop points into the CMS generation
6950 // and is not marked
6951 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6952 // a white object ...
6953 // If we manage to "claim" the object, by being the
6954 // first thread to mark it, then we push it on our
6955 // marking stack
6956 if (_bit_map->par_mark(addr)) { // ... now grey
6957 // push on work queue (grey set)
6958 bool simulate_overflow = false;
6959 NOT_PRODUCT(
6960 if (CMSMarkStackOverflowALot &&
6961 _collector->par_simulate_overflow()) {
6962 // simulate a stack overflow
6963 simulate_overflow = true;
6964 }
6965 )
6966 if (simulate_overflow || !_work_queue->push(obj)) {
6967 _collector->par_push_on_overflow_list(obj);
6968 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
6969 }
6970 } // Else, some other thread got there first
6971 }
6972 }
6973
6974 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); }
6975 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); }
6976
6977 void CMSPrecleanRefsYieldClosure::do_yield_work() {
6978 Mutex* bml = _collector->bitMapLock();
6979 assert_lock_strong(bml);
6980 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6981 "CMS thread should hold CMS token");
6982
6983 bml->unlock();
6984 ConcurrentMarkSweepThread::desynchronize(true);
6985
6986 _collector->stopTimer();
6987 _collector->incrementYields();
6988
6989 // See the comment in coordinator_yield()
6990 for (unsigned i = 0; i < CMSYieldSleepCount &&
6991 ConcurrentMarkSweepThread::should_yield() &&
6992 !CMSCollector::foregroundGCIsActive(); ++i) {
6993 os::sleep(Thread::current(), 1, false);
6994 }
6995
6996 ConcurrentMarkSweepThread::synchronize(true);
6997 bml->lock();
6998
6999 _collector->startTimer();
7000 }
7001
7002 bool CMSPrecleanRefsYieldClosure::should_return() {
7003 if (ConcurrentMarkSweepThread::should_yield()) {
7004 do_yield_work();
7005 }
7006 return _collector->foregroundGCIsActive();
7007 }
7008
7009 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7010 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7011 "mr should be aligned to start at a card boundary");
7012 // We'd like to assert:
7013 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7014 // "mr should be a range of cards");
7015 // However, that would be too strong in one case -- the last
7016 // partition ends at _unallocated_block which, in general, can be
7017 // an arbitrary boundary, not necessarily card aligned.
7018 _num_dirty_cards += mr.word_size()/CardTableModRefBS::card_size_in_words;
7019 _space->object_iterate_mem(mr, &_scan_cl);
7020 }
7021
7022 SweepClosure::SweepClosure(CMSCollector* collector,
7023 ConcurrentMarkSweepGeneration* g,
7024 CMSBitMap* bitMap, bool should_yield) :
7025 _collector(collector),
7026 _g(g),
7027 _sp(g->cmsSpace()),
7028 _limit(_sp->sweep_limit()),
7029 _freelistLock(_sp->freelistLock()),
7030 _bitMap(bitMap),
7031 _yield(should_yield),
7032 _inFreeRange(false), // No free range at beginning of sweep
7033 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7034 _lastFreeRangeCoalesced(false),
7035 _freeFinger(g->used_region().start())
7036 {
7037 NOT_PRODUCT(
7038 _numObjectsFreed = 0;
7039 _numWordsFreed = 0;
7040 _numObjectsLive = 0;
7041 _numWordsLive = 0;
7042 _numObjectsAlreadyFree = 0;
7043 _numWordsAlreadyFree = 0;
7044 _last_fc = NULL;
7045
7046 _sp->initializeIndexedFreeListArrayReturnedBytes();
7047 _sp->dictionary()->initialize_dict_returned_bytes();
7048 )
7049 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7050 "sweep _limit out of bounds");
7051 log_develop_trace(gc, sweep)("====================");
7052 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit));
7053 }
7054
7055 void SweepClosure::print_on(outputStream* st) const {
7056 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7057 p2i(_sp->bottom()), p2i(_sp->end()));
7058 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7059 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7060 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7061 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7062 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7063 }
7064
7065 #ifndef PRODUCT
7066 // Assertion checking only: no useful work in product mode --
7067 // however, if any of the flags below become product flags,
7068 // you may need to review this code to see if it needs to be
7069 // enabled in product mode.
7070 SweepClosure::~SweepClosure() {
7071 assert_lock_strong(_freelistLock);
7072 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7073 "sweep _limit out of bounds");
7074 if (inFreeRange()) {
7075 Log(gc, sweep) log;
7076 log.error("inFreeRange() should have been reset; dumping state of SweepClosure");
7077 ResourceMark rm;
7078 LogStream ls(log.error());
7079 print_on(&ls);
7080 ShouldNotReachHere();
7081 }
7082
7083 if (log_is_enabled(Debug, gc, sweep)) {
7084 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7085 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7086 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7087 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7088 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord);
7089 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes);
7090 }
7091
7092 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) {
7093 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7094 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7095 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7096 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes",
7097 returned_bytes, indexListReturnedBytes, dict_returned_bytes);
7098 }
7099 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit));
7100 log_develop_trace(gc, sweep)("================");
7101 }
7102 #endif // PRODUCT
7103
7104 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7105 bool freeRangeInFreeLists) {
7106 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)",
7107 p2i(freeFinger), freeRangeInFreeLists);
7108 assert(!inFreeRange(), "Trampling existing free range");
7109 set_inFreeRange(true);
7110 set_lastFreeRangeCoalesced(false);
7111
7112 set_freeFinger(freeFinger);
7113 set_freeRangeInFreeLists(freeRangeInFreeLists);
7114 if (CMSTestInFreeList) {
7115 if (freeRangeInFreeLists) {
7116 FreeChunk* fc = (FreeChunk*) freeFinger;
7117 assert(fc->is_free(), "A chunk on the free list should be free.");
7118 assert(fc->size() > 0, "Free range should have a size");
7119 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7120 }
7121 }
7122 }
7123
7124 // Note that the sweeper runs concurrently with mutators. Thus,
7125 // it is possible for direct allocation in this generation to happen
7126 // in the middle of the sweep. Note that the sweeper also coalesces
7127 // contiguous free blocks. Thus, unless the sweeper and the allocator
7128 // synchronize appropriately freshly allocated blocks may get swept up.
7129 // This is accomplished by the sweeper locking the free lists while
7130 // it is sweeping. Thus blocks that are determined to be free are
7131 // indeed free. There is however one additional complication:
7132 // blocks that have been allocated since the final checkpoint and
7133 // mark, will not have been marked and so would be treated as
7134 // unreachable and swept up. To prevent this, the allocator marks
7135 // the bit map when allocating during the sweep phase. This leads,
7136 // however, to a further complication -- objects may have been allocated
7137 // but not yet initialized -- in the sense that the header isn't yet
7138 // installed. The sweeper can not then determine the size of the block
7139 // in order to skip over it. To deal with this case, we use a technique
7140 // (due to Printezis) to encode such uninitialized block sizes in the
7141 // bit map. Since the bit map uses a bit per every HeapWord, but the
7142 // CMS generation has a minimum object size of 3 HeapWords, it follows
7143 // that "normal marks" won't be adjacent in the bit map (there will
7144 // always be at least two 0 bits between successive 1 bits). We make use
7145 // of these "unused" bits to represent uninitialized blocks -- the bit
7146 // corresponding to the start of the uninitialized object and the next
7147 // bit are both set. Finally, a 1 bit marks the end of the object that
7148 // started with the two consecutive 1 bits to indicate its potentially
7149 // uninitialized state.
7150
7151 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7152 FreeChunk* fc = (FreeChunk*)addr;
7153 size_t res;
7154
7155 // Check if we are done sweeping. Below we check "addr >= _limit" rather
7156 // than "addr == _limit" because although _limit was a block boundary when
7157 // we started the sweep, it may no longer be one because heap expansion
7158 // may have caused us to coalesce the block ending at the address _limit
7159 // with a newly expanded chunk (this happens when _limit was set to the
7160 // previous _end of the space), so we may have stepped past _limit:
7161 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7162 if (addr >= _limit) { // we have swept up to or past the limit: finish up
7163 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7164 "sweep _limit out of bounds");
7165 assert(addr < _sp->end(), "addr out of bounds");
7166 // Flush any free range we might be holding as a single
7167 // coalesced chunk to the appropriate free list.
7168 if (inFreeRange()) {
7169 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7170 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger()));
7171 flush_cur_free_chunk(freeFinger(),
7172 pointer_delta(addr, freeFinger()));
7173 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]",
7174 p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7175 lastFreeRangeCoalesced() ? 1 : 0);
7176 }
7177
7178 // help the iterator loop finish
7179 return pointer_delta(_sp->end(), addr);
7180 }
7181
7182 assert(addr < _limit, "sweep invariant");
7183 // check if we should yield
7184 do_yield_check(addr);
7185 if (fc->is_free()) {
7186 // Chunk that is already free
7187 res = fc->size();
7188 do_already_free_chunk(fc);
7189 debug_only(_sp->verifyFreeLists());
7190 // If we flush the chunk at hand in lookahead_and_flush()
7191 // and it's coalesced with a preceding chunk, then the
7192 // process of "mangling" the payload of the coalesced block
7193 // will cause erasure of the size information from the
7194 // (erstwhile) header of all the coalesced blocks but the
7195 // first, so the first disjunct in the assert will not hold
7196 // in that specific case (in which case the second disjunct
7197 // will hold).
7198 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7199 "Otherwise the size info doesn't change at this step");
7200 NOT_PRODUCT(
7201 _numObjectsAlreadyFree++;
7202 _numWordsAlreadyFree += res;
7203 )
7204 NOT_PRODUCT(_last_fc = fc;)
7205 } else if (!_bitMap->isMarked(addr)) {
7206 // Chunk is fresh garbage
7207 res = do_garbage_chunk(fc);
7208 debug_only(_sp->verifyFreeLists());
7209 NOT_PRODUCT(
7210 _numObjectsFreed++;
7211 _numWordsFreed += res;
7212 )
7213 } else {
7214 // Chunk that is alive.
7215 res = do_live_chunk(fc);
7216 debug_only(_sp->verifyFreeLists());
7217 NOT_PRODUCT(
7218 _numObjectsLive++;
7219 _numWordsLive += res;
7220 )
7221 }
7222 return res;
7223 }
7224
7225 // For the smart allocation, record following
7226 // split deaths - a free chunk is removed from its free list because
7227 // it is being split into two or more chunks.
7228 // split birth - a free chunk is being added to its free list because
7229 // a larger free chunk has been split and resulted in this free chunk.
7230 // coal death - a free chunk is being removed from its free list because
7231 // it is being coalesced into a large free chunk.
7232 // coal birth - a free chunk is being added to its free list because
7233 // it was created when two or more free chunks where coalesced into
7234 // this free chunk.
7235 //
7236 // These statistics are used to determine the desired number of free
7237 // chunks of a given size. The desired number is chosen to be relative
7238 // to the end of a CMS sweep. The desired number at the end of a sweep
7239 // is the
7240 // count-at-end-of-previous-sweep (an amount that was enough)
7241 // - count-at-beginning-of-current-sweep (the excess)
7242 // + split-births (gains in this size during interval)
7243 // - split-deaths (demands on this size during interval)
7244 // where the interval is from the end of one sweep to the end of the
7245 // next.
7246 //
7247 // When sweeping the sweeper maintains an accumulated chunk which is
7248 // the chunk that is made up of chunks that have been coalesced. That
7249 // will be termed the left-hand chunk. A new chunk of garbage that
7250 // is being considered for coalescing will be referred to as the
7251 // right-hand chunk.
7252 //
7253 // When making a decision on whether to coalesce a right-hand chunk with
7254 // the current left-hand chunk, the current count vs. the desired count
7255 // of the left-hand chunk is considered. Also if the right-hand chunk
7256 // is near the large chunk at the end of the heap (see
7257 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7258 // left-hand chunk is coalesced.
7259 //
7260 // When making a decision about whether to split a chunk, the desired count
7261 // vs. the current count of the candidate to be split is also considered.
7262 // If the candidate is underpopulated (currently fewer chunks than desired)
7263 // a chunk of an overpopulated (currently more chunks than desired) size may
7264 // be chosen. The "hint" associated with a free list, if non-null, points
7265 // to a free list which may be overpopulated.
7266 //
7267
7268 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7269 const size_t size = fc->size();
7270 // Chunks that cannot be coalesced are not in the
7271 // free lists.
7272 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7273 assert(_sp->verify_chunk_in_free_list(fc),
7274 "free chunk should be in free lists");
7275 }
7276 // a chunk that is already free, should not have been
7277 // marked in the bit map
7278 HeapWord* const addr = (HeapWord*) fc;
7279 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7280 // Verify that the bit map has no bits marked between
7281 // addr and purported end of this block.
7282 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7283
7284 // Some chunks cannot be coalesced under any circumstances.
7285 // See the definition of cantCoalesce().
7286 if (!fc->cantCoalesce()) {
7287 // This chunk can potentially be coalesced.
7288 // All the work is done in
7289 do_post_free_or_garbage_chunk(fc, size);
7290 // Note that if the chunk is not coalescable (the else arm
7291 // below), we unconditionally flush, without needing to do
7292 // a "lookahead," as we do below.
7293 if (inFreeRange()) lookahead_and_flush(fc, size);
7294 } else {
7295 // Code path common to both original and adaptive free lists.
7296
7297 // cant coalesce with previous block; this should be treated
7298 // as the end of a free run if any
7299 if (inFreeRange()) {
7300 // we kicked some butt; time to pick up the garbage
7301 assert(freeFinger() < addr, "freeFinger points too high");
7302 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7303 }
7304 // else, nothing to do, just continue
7305 }
7306 }
7307
7308 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7309 // This is a chunk of garbage. It is not in any free list.
7310 // Add it to a free list or let it possibly be coalesced into
7311 // a larger chunk.
7312 HeapWord* const addr = (HeapWord*) fc;
7313 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7314
7315 // Verify that the bit map has no bits marked between
7316 // addr and purported end of just dead object.
7317 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7318 do_post_free_or_garbage_chunk(fc, size);
7319
7320 assert(_limit >= addr + size,
7321 "A freshly garbage chunk can't possibly straddle over _limit");
7322 if (inFreeRange()) lookahead_and_flush(fc, size);
7323 return size;
7324 }
7325
7326 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7327 HeapWord* addr = (HeapWord*) fc;
7328 // The sweeper has just found a live object. Return any accumulated
7329 // left hand chunk to the free lists.
7330 if (inFreeRange()) {
7331 assert(freeFinger() < addr, "freeFinger points too high");
7332 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7333 }
7334
7335 // This object is live: we'd normally expect this to be
7336 // an oop, and like to assert the following:
7337 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop");
7338 // However, as we commented above, this may be an object whose
7339 // header hasn't yet been initialized.
7340 size_t size;
7341 assert(_bitMap->isMarked(addr), "Tautology for this control point");
7342 if (_bitMap->isMarked(addr + 1)) {
7343 // Determine the size from the bit map, rather than trying to
7344 // compute it from the object header.
7345 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7346 size = pointer_delta(nextOneAddr + 1, addr);
7347 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7348 "alignment problem");
7349
7350 #ifdef ASSERT
7351 if (oop(addr)->klass_or_null_acquire() != NULL) {
7352 // Ignore mark word because we are running concurrent with mutators
7353 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop");
7354 assert(size ==
7355 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7356 "P-mark and computed size do not agree");
7357 }
7358 #endif
7359
7360 } else {
7361 // This should be an initialized object that's alive.
7362 assert(oop(addr)->klass_or_null_acquire() != NULL,
7363 "Should be an initialized object");
7364 // Ignore mark word because we are running concurrent with mutators
7365 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop");
7366 // Verify that the bit map has no bits marked between
7367 // addr and purported end of this block.
7368 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7369 assert(size >= 3, "Necessary for Printezis marks to work");
7370 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7371 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7372 }
7373 return size;
7374 }
7375
7376 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7377 size_t chunkSize) {
7378 // do_post_free_or_garbage_chunk() should only be called in the case
7379 // of the adaptive free list allocator.
7380 const bool fcInFreeLists = fc->is_free();
7381 assert((HeapWord*)fc <= _limit, "sweep invariant");
7382 if (CMSTestInFreeList && fcInFreeLists) {
7383 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7384 }
7385
7386 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7387
7388 HeapWord* const fc_addr = (HeapWord*) fc;
7389
7390 bool coalesce = false;
7391 const size_t left = pointer_delta(fc_addr, freeFinger());
7392 const size_t right = chunkSize;
7393 switch (FLSCoalescePolicy) {
7394 // numeric value forms a coalition aggressiveness metric
7395 case 0: { // never coalesce
7396 coalesce = false;
7397 break;
7398 }
7399 case 1: { // coalesce if left & right chunks on overpopulated lists
7400 coalesce = _sp->coalOverPopulated(left) &&
7401 _sp->coalOverPopulated(right);
7402 break;
7403 }
7404 case 2: { // coalesce if left chunk on overpopulated list (default)
7405 coalesce = _sp->coalOverPopulated(left);
7406 break;
7407 }
7408 case 3: { // coalesce if left OR right chunk on overpopulated list
7409 coalesce = _sp->coalOverPopulated(left) ||
7410 _sp->coalOverPopulated(right);
7411 break;
7412 }
7413 case 4: { // always coalesce
7414 coalesce = true;
7415 break;
7416 }
7417 default:
7418 ShouldNotReachHere();
7419 }
7420
7421 // Should the current free range be coalesced?
7422 // If the chunk is in a free range and either we decided to coalesce above
7423 // or the chunk is near the large block at the end of the heap
7424 // (isNearLargestChunk() returns true), then coalesce this chunk.
7425 const bool doCoalesce = inFreeRange()
7426 && (coalesce || _g->isNearLargestChunk(fc_addr));
7427 if (doCoalesce) {
7428 // Coalesce the current free range on the left with the new
7429 // chunk on the right. If either is on a free list,
7430 // it must be removed from the list and stashed in the closure.
7431 if (freeRangeInFreeLists()) {
7432 FreeChunk* const ffc = (FreeChunk*)freeFinger();
7433 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7434 "Size of free range is inconsistent with chunk size.");
7435 if (CMSTestInFreeList) {
7436 assert(_sp->verify_chunk_in_free_list(ffc),
7437 "Chunk is not in free lists");
7438 }
7439 _sp->coalDeath(ffc->size());
7440 _sp->removeFreeChunkFromFreeLists(ffc);
7441 set_freeRangeInFreeLists(false);
7442 }
7443 if (fcInFreeLists) {
7444 _sp->coalDeath(chunkSize);
7445 assert(fc->size() == chunkSize,
7446 "The chunk has the wrong size or is not in the free lists");
7447 _sp->removeFreeChunkFromFreeLists(fc);
7448 }
7449 set_lastFreeRangeCoalesced(true);
7450 print_free_block_coalesced(fc);
7451 } else { // not in a free range and/or should not coalesce
7452 // Return the current free range and start a new one.
7453 if (inFreeRange()) {
7454 // In a free range but cannot coalesce with the right hand chunk.
7455 // Put the current free range into the free lists.
7456 flush_cur_free_chunk(freeFinger(),
7457 pointer_delta(fc_addr, freeFinger()));
7458 }
7459 // Set up for new free range. Pass along whether the right hand
7460 // chunk is in the free lists.
7461 initialize_free_range((HeapWord*)fc, fcInFreeLists);
7462 }
7463 }
7464
7465 // Lookahead flush:
7466 // If we are tracking a free range, and this is the last chunk that
7467 // we'll look at because its end crosses past _limit, we'll preemptively
7468 // flush it along with any free range we may be holding on to. Note that
7469 // this can be the case only for an already free or freshly garbage
7470 // chunk. If this block is an object, it can never straddle
7471 // over _limit. The "straddling" occurs when _limit is set at
7472 // the previous end of the space when this cycle started, and
7473 // a subsequent heap expansion caused the previously co-terminal
7474 // free block to be coalesced with the newly expanded portion,
7475 // thus rendering _limit a non-block-boundary making it dangerous
7476 // for the sweeper to step over and examine.
7477 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7478 assert(inFreeRange(), "Should only be called if currently in a free range.");
7479 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7480 assert(_sp->used_region().contains(eob - 1),
7481 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7482 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7483 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7484 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size);
7485 if (eob >= _limit) {
7486 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7487 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block "
7488 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7489 "[" PTR_FORMAT "," PTR_FORMAT ")",
7490 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7491 // Return the storage we are tracking back into the free lists.
7492 log_develop_trace(gc, sweep)("Flushing ... ");
7493 assert(freeFinger() < eob, "Error");
7494 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7495 }
7496 }
7497
7498 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7499 assert(inFreeRange(), "Should only be called if currently in a free range.");
7500 assert(size > 0,
7501 "A zero sized chunk cannot be added to the free lists.");
7502 if (!freeRangeInFreeLists()) {
7503 if (CMSTestInFreeList) {
7504 FreeChunk* fc = (FreeChunk*) chunk;
7505 fc->set_size(size);
7506 assert(!_sp->verify_chunk_in_free_list(fc),
7507 "chunk should not be in free lists yet");
7508 }
7509 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size);
7510 // A new free range is going to be starting. The current
7511 // free range has not been added to the free lists yet or
7512 // was removed so add it back.
7513 // If the current free range was coalesced, then the death
7514 // of the free range was recorded. Record a birth now.
7515 if (lastFreeRangeCoalesced()) {
7516 _sp->coalBirth(size);
7517 }
7518 _sp->addChunkAndRepairOffsetTable(chunk, size,
7519 lastFreeRangeCoalesced());
7520 } else {
7521 log_develop_trace(gc, sweep)("Already in free list: nothing to flush");
7522 }
7523 set_inFreeRange(false);
7524 set_freeRangeInFreeLists(false);
7525 }
7526
7527 // We take a break if we've been at this for a while,
7528 // so as to avoid monopolizing the locks involved.
7529 void SweepClosure::do_yield_work(HeapWord* addr) {
7530 // Return current free chunk being used for coalescing (if any)
7531 // to the appropriate freelist. After yielding, the next
7532 // free block encountered will start a coalescing range of
7533 // free blocks. If the next free block is adjacent to the
7534 // chunk just flushed, they will need to wait for the next
7535 // sweep to be coalesced.
7536 if (inFreeRange()) {
7537 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7538 }
7539
7540 // First give up the locks, then yield, then re-lock.
7541 // We should probably use a constructor/destructor idiom to
7542 // do this unlock/lock or modify the MutexUnlocker class to
7543 // serve our purpose. XXX
7544 assert_lock_strong(_bitMap->lock());
7545 assert_lock_strong(_freelistLock);
7546 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7547 "CMS thread should hold CMS token");
7548 _bitMap->lock()->unlock();
7549 _freelistLock->unlock();
7550 ConcurrentMarkSweepThread::desynchronize(true);
7551 _collector->stopTimer();
7552 _collector->incrementYields();
7553
7554 // See the comment in coordinator_yield()
7555 for (unsigned i = 0; i < CMSYieldSleepCount &&
7556 ConcurrentMarkSweepThread::should_yield() &&
7557 !CMSCollector::foregroundGCIsActive(); ++i) {
7558 os::sleep(Thread::current(), 1, false);
7559 }
7560
7561 ConcurrentMarkSweepThread::synchronize(true);
7562 _freelistLock->lock();
7563 _bitMap->lock()->lock_without_safepoint_check();
7564 _collector->startTimer();
7565 }
7566
7567 #ifndef PRODUCT
7568 // This is actually very useful in a product build if it can
7569 // be called from the debugger. Compile it into the product
7570 // as needed.
7571 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7572 return debug_cms_space->verify_chunk_in_free_list(fc);
7573 }
7574 #endif
7575
7576 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7577 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7578 p2i(fc), fc->size());
7579 }
7580
7581 // CMSIsAliveClosure
7582 bool CMSIsAliveClosure::do_object_b(oop obj) {
7583 HeapWord* addr = (HeapWord*)obj;
7584 return addr != NULL &&
7585 (!_span.contains(addr) || _bit_map->isMarked(addr));
7586 }
7587
7588
7589 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7590 MemRegion span,
7591 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7592 bool cpc):
7593 _collector(collector),
7594 _span(span),
7595 _bit_map(bit_map),
7596 _mark_stack(mark_stack),
7597 _concurrent_precleaning(cpc) {
7598 assert(!_span.is_empty(), "Empty span could spell trouble");
7599 }
7600
7601
7602 // CMSKeepAliveClosure: the serial version
7603 void CMSKeepAliveClosure::do_oop(oop obj) {
7604 HeapWord* addr = (HeapWord*)obj;
7605 if (_span.contains(addr) &&
7606 !_bit_map->isMarked(addr)) {
7607 _bit_map->mark(addr);
7608 bool simulate_overflow = false;
7609 NOT_PRODUCT(
7610 if (CMSMarkStackOverflowALot &&
7611 _collector->simulate_overflow()) {
7612 // simulate a stack overflow
7613 simulate_overflow = true;
7614 }
7615 )
7616 if (simulate_overflow || !_mark_stack->push(obj)) {
7617 if (_concurrent_precleaning) {
7618 // We dirty the overflown object and let the remark
7619 // phase deal with it.
7620 assert(_collector->overflow_list_is_empty(), "Error");
7621 // In the case of object arrays, we need to dirty all of
7622 // the cards that the object spans. No locking or atomics
7623 // are needed since no one else can be mutating the mod union
7624 // table.
7625 if (obj->is_objArray()) {
7626 size_t sz = obj->size();
7627 HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size);
7628 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7629 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7630 _collector->_modUnionTable.mark_range(redirty_range);
7631 } else {
7632 _collector->_modUnionTable.mark(addr);
7633 }
7634 _collector->_ser_kac_preclean_ovflw++;
7635 } else {
7636 _collector->push_on_overflow_list(obj);
7637 _collector->_ser_kac_ovflw++;
7638 }
7639 }
7640 }
7641 }
7642
7643 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7644 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7645
7646 // CMSParKeepAliveClosure: a parallel version of the above.
7647 // The work queues are private to each closure (thread),
7648 // but (may be) available for stealing by other threads.
7649 void CMSParKeepAliveClosure::do_oop(oop obj) {
7650 HeapWord* addr = (HeapWord*)obj;
7651 if (_span.contains(addr) &&
7652 !_bit_map->isMarked(addr)) {
7653 // In general, during recursive tracing, several threads
7654 // may be concurrently getting here; the first one to
7655 // "tag" it, claims it.
7656 if (_bit_map->par_mark(addr)) {
7657 bool res = _work_queue->push(obj);
7658 assert(res, "Low water mark should be much less than capacity");
7659 // Do a recursive trim in the hope that this will keep
7660 // stack usage lower, but leave some oops for potential stealers
7661 trim_queue(_low_water_mark);
7662 } // Else, another thread got there first
7663 }
7664 }
7665
7666 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7667 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7668
7669 void CMSParKeepAliveClosure::trim_queue(uint max) {
7670 while (_work_queue->size() > max) {
7671 oop new_oop;
7672 if (_work_queue->pop_local(new_oop)) {
7673 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
7674 assert(_bit_map->isMarked((HeapWord*)new_oop),
7675 "no white objects on this stack!");
7676 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7677 // iterate over the oops in this oop, marking and pushing
7678 // the ones in CMS heap (i.e. in _span).
7679 new_oop->oop_iterate(&_mark_and_push);
7680 }
7681 }
7682 }
7683
7684 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
7685 CMSCollector* collector,
7686 MemRegion span, CMSBitMap* bit_map,
7687 OopTaskQueue* work_queue):
7688 _collector(collector),
7689 _span(span),
7690 _bit_map(bit_map),
7691 _work_queue(work_queue) { }
7692
7693 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
7694 HeapWord* addr = (HeapWord*)obj;
7695 if (_span.contains(addr) &&
7696 !_bit_map->isMarked(addr)) {
7697 if (_bit_map->par_mark(addr)) {
7698 bool simulate_overflow = false;
7699 NOT_PRODUCT(
7700 if (CMSMarkStackOverflowALot &&
7701 _collector->par_simulate_overflow()) {
7702 // simulate a stack overflow
7703 simulate_overflow = true;
7704 }
7705 )
7706 if (simulate_overflow || !_work_queue->push(obj)) {
7707 _collector->par_push_on_overflow_list(obj);
7708 _collector->_par_kac_ovflw++;
7709 }
7710 } // Else another thread got there already
7711 }
7712 }
7713
7714 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7715 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7716
7717 //////////////////////////////////////////////////////////////////
7718 // CMSExpansionCause /////////////////////////////
7719 //////////////////////////////////////////////////////////////////
7720 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
7721 switch (cause) {
7722 case _no_expansion:
7723 return "No expansion";
7724 case _satisfy_free_ratio:
7725 return "Free ratio";
7726 case _satisfy_promotion:
7727 return "Satisfy promotion";
7728 case _satisfy_allocation:
7729 return "allocation";
7730 case _allocate_par_lab:
7731 return "Par LAB";
7732 case _allocate_par_spooling_space:
7733 return "Par Spooling Space";
7734 case _adaptive_size_policy:
7735 return "Ergonomics";
7736 default:
7737 return "unknown";
7738 }
7739 }
7740
7741 void CMSDrainMarkingStackClosure::do_void() {
7742 // the max number to take from overflow list at a time
7743 const size_t num = _mark_stack->capacity()/4;
7744 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
7745 "Overflow list should be NULL during concurrent phases");
7746 while (!_mark_stack->isEmpty() ||
7747 // if stack is empty, check the overflow list
7748 _collector->take_from_overflow_list(num, _mark_stack)) {
7749 oop obj = _mark_stack->pop();
7750 HeapWord* addr = (HeapWord*)obj;
7751 assert(_span.contains(addr), "Should be within span");
7752 assert(_bit_map->isMarked(addr), "Should be marked");
7753 assert(oopDesc::is_oop(obj), "Should be an oop");
7754 obj->oop_iterate(_keep_alive);
7755 }
7756 }
7757
7758 void CMSParDrainMarkingStackClosure::do_void() {
7759 // drain queue
7760 trim_queue(0);
7761 }
7762
7763 // Trim our work_queue so its length is below max at return
7764 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
7765 while (_work_queue->size() > max) {
7766 oop new_oop;
7767 if (_work_queue->pop_local(new_oop)) {
7768 assert(oopDesc::is_oop(new_oop), "Expected an oop");
7769 assert(_bit_map->isMarked((HeapWord*)new_oop),
7770 "no white objects on this stack!");
7771 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7772 // iterate over the oops in this oop, marking and pushing
7773 // the ones in CMS heap (i.e. in _span).
7774 new_oop->oop_iterate(&_mark_and_push);
7775 }
7776 }
7777 }
7778
7779 ////////////////////////////////////////////////////////////////////
7780 // Support for Marking Stack Overflow list handling and related code
7781 ////////////////////////////////////////////////////////////////////
7782 // Much of the following code is similar in shape and spirit to the
7783 // code used in ParNewGC. We should try and share that code
7784 // as much as possible in the future.
7785
7786 #ifndef PRODUCT
7787 // Debugging support for CMSStackOverflowALot
7788
7789 // It's OK to call this multi-threaded; the worst thing
7790 // that can happen is that we'll get a bunch of closely
7791 // spaced simulated overflows, but that's OK, in fact
7792 // probably good as it would exercise the overflow code
7793 // under contention.
7794 bool CMSCollector::simulate_overflow() {
7795 if (_overflow_counter-- <= 0) { // just being defensive
7796 _overflow_counter = CMSMarkStackOverflowInterval;
7797 return true;
7798 } else {
7799 return false;
7800 }
7801 }
7802
7803 bool CMSCollector::par_simulate_overflow() {
7804 return simulate_overflow();
7805 }
7806 #endif
7807
7808 // Single-threaded
7809 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
7810 assert(stack->isEmpty(), "Expected precondition");
7811 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
7812 size_t i = num;
7813 oop cur = _overflow_list;
7814 const markOop proto = markOopDesc::prototype();
7815 NOT_PRODUCT(ssize_t n = 0;)
7816 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
7817 next = oop(cur->mark());
7818 cur->set_mark(proto); // until proven otherwise
7819 assert(oopDesc::is_oop(cur), "Should be an oop");
7820 bool res = stack->push(cur);
7821 assert(res, "Bit off more than can chew?");
7822 NOT_PRODUCT(n++;)
7823 }
7824 _overflow_list = cur;
7825 #ifndef PRODUCT
7826 assert(_num_par_pushes >= n, "Too many pops?");
7827 _num_par_pushes -=n;
7828 #endif
7829 return !stack->isEmpty();
7830 }
7831
7832 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
7833 // (MT-safe) Get a prefix of at most "num" from the list.
7834 // The overflow list is chained through the mark word of
7835 // each object in the list. We fetch the entire list,
7836 // break off a prefix of the right size and return the
7837 // remainder. If other threads try to take objects from
7838 // the overflow list at that time, they will wait for
7839 // some time to see if data becomes available. If (and
7840 // only if) another thread places one or more object(s)
7841 // on the global list before we have returned the suffix
7842 // to the global list, we will walk down our local list
7843 // to find its end and append the global list to
7844 // our suffix before returning it. This suffix walk can
7845 // prove to be expensive (quadratic in the amount of traffic)
7846 // when there are many objects in the overflow list and
7847 // there is much producer-consumer contention on the list.
7848 // *NOTE*: The overflow list manipulation code here and
7849 // in ParNewGeneration:: are very similar in shape,
7850 // except that in the ParNew case we use the old (from/eden)
7851 // copy of the object to thread the list via its klass word.
7852 // Because of the common code, if you make any changes in
7853 // the code below, please check the ParNew version to see if
7854 // similar changes might be needed.
7855 // CR 6797058 has been filed to consolidate the common code.
7856 bool CMSCollector::par_take_from_overflow_list(size_t num,
7857 OopTaskQueue* work_q,
7858 int no_of_gc_threads) {
7859 assert(work_q->size() == 0, "First empty local work queue");
7860 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
7861 if (_overflow_list == NULL) {
7862 return false;
7863 }
7864 // Grab the entire list; we'll put back a suffix
7865 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
7866 Thread* tid = Thread::current();
7867 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
7868 // set to ParallelGCThreads.
7869 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
7870 size_t sleep_time_millis = MAX2((size_t)1, num/100);
7871 // If the list is busy, we spin for a short while,
7872 // sleeping between attempts to get the list.
7873 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
7874 os::sleep(tid, sleep_time_millis, false);
7875 if (_overflow_list == NULL) {
7876 // Nothing left to take
7877 return false;
7878 } else if (_overflow_list != BUSY) {
7879 // Try and grab the prefix
7880 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
7881 }
7882 }
7883 // If the list was found to be empty, or we spun long
7884 // enough, we give up and return empty-handed. If we leave
7885 // the list in the BUSY state below, it must be the case that
7886 // some other thread holds the overflow list and will set it
7887 // to a non-BUSY state in the future.
7888 if (prefix == NULL || prefix == BUSY) {
7889 // Nothing to take or waited long enough
7890 if (prefix == NULL) {
7891 // Write back the NULL in case we overwrote it with BUSY above
7892 // and it is still the same value.
7893 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
7894 }
7895 return false;
7896 }
7897 assert(prefix != NULL && prefix != BUSY, "Error");
7898 size_t i = num;
7899 oop cur = prefix;
7900 // Walk down the first "num" objects, unless we reach the end.
7901 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
7902 if (cur->mark() == NULL) {
7903 // We have "num" or fewer elements in the list, so there
7904 // is nothing to return to the global list.
7905 // Write back the NULL in lieu of the BUSY we wrote
7906 // above, if it is still the same value.
7907 if (_overflow_list == BUSY) {
7908 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
7909 }
7910 } else {
7911 // Chop off the suffix and return it to the global list.
7912 assert(cur->mark() != BUSY, "Error");
7913 oop suffix_head = cur->mark(); // suffix will be put back on global list
7914 cur->set_mark(NULL); // break off suffix
7915 // It's possible that the list is still in the empty(busy) state
7916 // we left it in a short while ago; in that case we may be
7917 // able to place back the suffix without incurring the cost
7918 // of a walk down the list.
7919 oop observed_overflow_list = _overflow_list;
7920 oop cur_overflow_list = observed_overflow_list;
7921 bool attached = false;
7922 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
7923 observed_overflow_list =
7924 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list);
7925 if (cur_overflow_list == observed_overflow_list) {
7926 attached = true;
7927 break;
7928 } else cur_overflow_list = observed_overflow_list;
7929 }
7930 if (!attached) {
7931 // Too bad, someone else sneaked in (at least) an element; we'll need
7932 // to do a splice. Find tail of suffix so we can prepend suffix to global
7933 // list.
7934 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
7935 oop suffix_tail = cur;
7936 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
7937 "Tautology");
7938 observed_overflow_list = _overflow_list;
7939 do {
7940 cur_overflow_list = observed_overflow_list;
7941 if (cur_overflow_list != BUSY) {
7942 // Do the splice ...
7943 suffix_tail->set_mark(markOop(cur_overflow_list));
7944 } else { // cur_overflow_list == BUSY
7945 suffix_tail->set_mark(NULL);
7946 }
7947 // ... and try to place spliced list back on overflow_list ...
7948 observed_overflow_list =
7949 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list);
7950 } while (cur_overflow_list != observed_overflow_list);
7951 // ... until we have succeeded in doing so.
7952 }
7953 }
7954
7955 // Push the prefix elements on work_q
7956 assert(prefix != NULL, "control point invariant");
7957 const markOop proto = markOopDesc::prototype();
7958 oop next;
7959 NOT_PRODUCT(ssize_t n = 0;)
7960 for (cur = prefix; cur != NULL; cur = next) {
7961 next = oop(cur->mark());
7962 cur->set_mark(proto); // until proven otherwise
7963 assert(oopDesc::is_oop(cur), "Should be an oop");
7964 bool res = work_q->push(cur);
7965 assert(res, "Bit off more than we can chew?");
7966 NOT_PRODUCT(n++;)
7967 }
7968 #ifndef PRODUCT
7969 assert(_num_par_pushes >= n, "Too many pops?");
7970 Atomic::sub(n, &_num_par_pushes);
7971 #endif
7972 return true;
7973 }
7974
7975 // Single-threaded
7976 void CMSCollector::push_on_overflow_list(oop p) {
7977 NOT_PRODUCT(_num_par_pushes++;)
7978 assert(oopDesc::is_oop(p), "Not an oop");
7979 preserve_mark_if_necessary(p);
7980 p->set_mark((markOop)_overflow_list);
7981 _overflow_list = p;
7982 }
7983
7984 // Multi-threaded; use CAS to prepend to overflow list
7985 void CMSCollector::par_push_on_overflow_list(oop p) {
7986 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);)
7987 assert(oopDesc::is_oop(p), "Not an oop");
7988 par_preserve_mark_if_necessary(p);
7989 oop observed_overflow_list = _overflow_list;
7990 oop cur_overflow_list;
7991 do {
7992 cur_overflow_list = observed_overflow_list;
7993 if (cur_overflow_list != BUSY) {
7994 p->set_mark(markOop(cur_overflow_list));
7995 } else {
7996 p->set_mark(NULL);
7997 }
7998 observed_overflow_list =
7999 Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list);
8000 } while (cur_overflow_list != observed_overflow_list);
8001 }
8002 #undef BUSY
8003
8004 // Single threaded
8005 // General Note on GrowableArray: pushes may silently fail
8006 // because we are (temporarily) out of C-heap for expanding
8007 // the stack. The problem is quite ubiquitous and affects
8008 // a lot of code in the JVM. The prudent thing for GrowableArray
8009 // to do (for now) is to exit with an error. However, that may
8010 // be too draconian in some cases because the caller may be
8011 // able to recover without much harm. For such cases, we
8012 // should probably introduce a "soft_push" method which returns
8013 // an indication of success or failure with the assumption that
8014 // the caller may be able to recover from a failure; code in
8015 // the VM can then be changed, incrementally, to deal with such
8016 // failures where possible, thus, incrementally hardening the VM
8017 // in such low resource situations.
8018 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8019 _preserved_oop_stack.push(p);
8020 _preserved_mark_stack.push(m);
8021 assert(m == p->mark(), "Mark word changed");
8022 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8023 "bijection");
8024 }
8025
8026 // Single threaded
8027 void CMSCollector::preserve_mark_if_necessary(oop p) {
8028 markOop m = p->mark();
8029 if (m->must_be_preserved(p)) {
8030 preserve_mark_work(p, m);
8031 }
8032 }
8033
8034 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8035 markOop m = p->mark();
8036 if (m->must_be_preserved(p)) {
8037 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8038 // Even though we read the mark word without holding
8039 // the lock, we are assured that it will not change
8040 // because we "own" this oop, so no other thread can
8041 // be trying to push it on the overflow list; see
8042 // the assertion in preserve_mark_work() that checks
8043 // that m == p->mark().
8044 preserve_mark_work(p, m);
8045 }
8046 }
8047
8048 // We should be able to do this multi-threaded,
8049 // a chunk of stack being a task (this is
8050 // correct because each oop only ever appears
8051 // once in the overflow list. However, it's
8052 // not very easy to completely overlap this with
8053 // other operations, so will generally not be done
8054 // until all work's been completed. Because we
8055 // expect the preserved oop stack (set) to be small,
8056 // it's probably fine to do this single-threaded.
8057 // We can explore cleverer concurrent/overlapped/parallel
8058 // processing of preserved marks if we feel the
8059 // need for this in the future. Stack overflow should
8060 // be so rare in practice and, when it happens, its
8061 // effect on performance so great that this will
8062 // likely just be in the noise anyway.
8063 void CMSCollector::restore_preserved_marks_if_any() {
8064 assert(SafepointSynchronize::is_at_safepoint(),
8065 "world should be stopped");
8066 assert(Thread::current()->is_ConcurrentGC_thread() ||
8067 Thread::current()->is_VM_thread(),
8068 "should be single-threaded");
8069 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8070 "bijection");
8071
8072 while (!_preserved_oop_stack.is_empty()) {
8073 oop p = _preserved_oop_stack.pop();
8074 assert(oopDesc::is_oop(p), "Should be an oop");
8075 assert(_span.contains(p), "oop should be in _span");
8076 assert(p->mark() == markOopDesc::prototype(),
8077 "Set when taken from overflow list");
8078 markOop m = _preserved_mark_stack.pop();
8079 p->set_mark(m);
8080 }
8081 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8082 "stacks were cleared above");
8083 }
8084
8085 #ifndef PRODUCT
8086 bool CMSCollector::no_preserved_marks() const {
8087 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8088 }
8089 #endif
8090
8091 // Transfer some number of overflown objects to usual marking
8092 // stack. Return true if some objects were transferred.
8093 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8094 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8095 (size_t)ParGCDesiredObjsFromOverflowList);
8096
8097 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8098 assert(_collector->overflow_list_is_empty() || res,
8099 "If list is not empty, we should have taken something");
8100 assert(!res || !_mark_stack->isEmpty(),
8101 "If we took something, it should now be on our stack");
8102 return res;
8103 }
8104
8105 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8106 size_t res = _sp->block_size_no_stall(addr, _collector);
8107 if (_sp->block_is_obj(addr)) {
8108 if (_live_bit_map->isMarked(addr)) {
8109 // It can't have been dead in a previous cycle
8110 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8111 } else {
8112 _dead_bit_map->mark(addr); // mark the dead object
8113 }
8114 }
8115 // Could be 0, if the block size could not be computed without stalling.
8116 return res;
8117 }
8118
8119 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8120 GCMemoryManager* manager = CMSHeap::heap()->old_manager();
8121 switch (phase) {
8122 case CMSCollector::InitialMarking:
8123 initialize(manager /* GC manager */ ,
8124 cause /* cause of the GC */,
8125 true /* recordGCBeginTime */,
8126 true /* recordPreGCUsage */,
8127 false /* recordPeakUsage */,
8128 false /* recordPostGCusage */,
8129 true /* recordAccumulatedGCTime */,
8130 false /* recordGCEndTime */,
8131 false /* countCollection */ );
8132 break;
8133
8134 case CMSCollector::FinalMarking:
8135 initialize(manager /* GC manager */ ,
8136 cause /* cause of the GC */,
8137 false /* recordGCBeginTime */,
8138 false /* recordPreGCUsage */,
8139 false /* recordPeakUsage */,
8140 false /* recordPostGCusage */,
8141 true /* recordAccumulatedGCTime */,
8142 false /* recordGCEndTime */,
8143 false /* countCollection */ );
8144 break;
8145
8146 case CMSCollector::Sweeping:
8147 initialize(manager /* GC manager */ ,
8148 cause /* cause of the GC */,
8149 false /* recordGCBeginTime */,
8150 false /* recordPreGCUsage */,
8151 true /* recordPeakUsage */,
8152 true /* recordPostGCusage */,
8153 false /* recordAccumulatedGCTime */,
8154 true /* recordGCEndTime */,
8155 true /* countCollection */ );
8156 break;
8157
8158 default:
8159 ShouldNotReachHere();
8160 }
8161 }