1 /*
2 * Copyright (c) 2001, 2014, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/classLoaderData.hpp"
27 #include "classfile/stringTable.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "code/codeCache.hpp"
30 #include "gc_implementation/shared/adaptiveSizePolicy.hpp"
31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp"
32 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp"
33 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp"
34 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp"
35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp"
36 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp"
37 #include "gc_implementation/parNew/parNewGeneration.hpp"
38 #include "gc_implementation/shared/collectorCounters.hpp"
39 #include "gc_implementation/shared/gcTimer.hpp"
40 #include "gc_implementation/shared/gcTrace.hpp"
41 #include "gc_implementation/shared/gcTraceTime.hpp"
42 #include "gc_implementation/shared/isGCActiveMark.hpp"
43 #include "gc_interface/collectedHeap.inline.hpp"
44 #include "memory/allocation.hpp"
45 #include "memory/cardTableRS.hpp"
46 #include "memory/collectorPolicy.hpp"
47 #include "memory/gcLocker.inline.hpp"
48 #include "memory/genCollectedHeap.hpp"
49 #include "memory/genMarkSweep.hpp"
50 #include "memory/genOopClosures.inline.hpp"
51 #include "memory/iterator.inline.hpp"
52 #include "memory/padded.hpp"
53 #include "memory/referencePolicy.hpp"
54 #include "memory/resourceArea.hpp"
55 #include "memory/tenuredGeneration.hpp"
56 #include "oops/oop.inline.hpp"
57 #include "prims/jvmtiExport.hpp"
58 #include "runtime/atomic.inline.hpp"
59 #include "runtime/globals_extension.hpp"
60 #include "runtime/handles.inline.hpp"
61 #include "runtime/java.hpp"
62 #include "runtime/orderAccess.inline.hpp"
63 #include "runtime/vmThread.hpp"
64 #include "services/memoryService.hpp"
65 #include "services/runtimeService.hpp"
66
67 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
68
69 // statics
70 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
71 bool CMSCollector::_full_gc_requested = false;
72 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
73
74 //////////////////////////////////////////////////////////////////
75 // In support of CMS/VM thread synchronization
76 //////////////////////////////////////////////////////////////////
77 // We split use of the CGC_lock into 2 "levels".
78 // The low-level locking is of the usual CGC_lock monitor. We introduce
79 // a higher level "token" (hereafter "CMS token") built on top of the
80 // low level monitor (hereafter "CGC lock").
81 // The token-passing protocol gives priority to the VM thread. The
82 // CMS-lock doesn't provide any fairness guarantees, but clients
83 // should ensure that it is only held for very short, bounded
84 // durations.
85 //
86 // When either of the CMS thread or the VM thread is involved in
87 // collection operations during which it does not want the other
88 // thread to interfere, it obtains the CMS token.
89 //
90 // If either thread tries to get the token while the other has
91 // it, that thread waits. However, if the VM thread and CMS thread
92 // both want the token, then the VM thread gets priority while the
93 // CMS thread waits. This ensures, for instance, that the "concurrent"
94 // phases of the CMS thread's work do not block out the VM thread
95 // for long periods of time as the CMS thread continues to hog
96 // the token. (See bug 4616232).
97 //
98 // The baton-passing functions are, however, controlled by the
99 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
100 // and here the low-level CMS lock, not the high level token,
101 // ensures mutual exclusion.
102 //
103 // Two important conditions that we have to satisfy:
104 // 1. if a thread does a low-level wait on the CMS lock, then it
105 // relinquishes the CMS token if it were holding that token
106 // when it acquired the low-level CMS lock.
107 // 2. any low-level notifications on the low-level lock
108 // should only be sent when a thread has relinquished the token.
109 //
110 // In the absence of either property, we'd have potential deadlock.
111 //
112 // We protect each of the CMS (concurrent and sequential) phases
113 // with the CMS _token_, not the CMS _lock_.
114 //
115 // The only code protected by CMS lock is the token acquisition code
116 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
117 // baton-passing code.
118 //
119 // Unfortunately, i couldn't come up with a good abstraction to factor and
120 // hide the naked CGC_lock manipulation in the baton-passing code
121 // further below. That's something we should try to do. Also, the proof
122 // of correctness of this 2-level locking scheme is far from obvious,
123 // and potentially quite slippery. We have an uneasy suspicion, for instance,
124 // that there may be a theoretical possibility of delay/starvation in the
125 // low-level lock/wait/notify scheme used for the baton-passing because of
126 // potential interference with the priority scheme embodied in the
127 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
128 // invocation further below and marked with "XXX 20011219YSR".
129 // Indeed, as we note elsewhere, this may become yet more slippery
130 // in the presence of multiple CMS and/or multiple VM threads. XXX
131
132 class CMSTokenSync: public StackObj {
133 private:
134 bool _is_cms_thread;
135 public:
136 CMSTokenSync(bool is_cms_thread):
137 _is_cms_thread(is_cms_thread) {
138 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
139 "Incorrect argument to constructor");
140 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
141 }
142
143 ~CMSTokenSync() {
144 assert(_is_cms_thread ?
145 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
146 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
147 "Incorrect state");
148 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
149 }
150 };
151
152 // Convenience class that does a CMSTokenSync, and then acquires
153 // upto three locks.
154 class CMSTokenSyncWithLocks: public CMSTokenSync {
155 private:
156 // Note: locks are acquired in textual declaration order
157 // and released in the opposite order
158 MutexLockerEx _locker1, _locker2, _locker3;
159 public:
160 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
161 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
162 CMSTokenSync(is_cms_thread),
163 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
164 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
165 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
166 { }
167 };
168
169
170 // Wrapper class to temporarily disable icms during a foreground cms collection.
171 class ICMSDisabler: public StackObj {
172 public:
173 // The ctor disables icms and wakes up the thread so it notices the change;
174 // the dtor re-enables icms. Note that the CMSCollector methods will check
175 // CMSIncrementalMode.
176 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); }
177 ~ICMSDisabler() { CMSCollector::enable_icms(); }
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 CFLS_LAB 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, int level,
201 CardTableRS* ct, bool use_adaptive_freelists,
202 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) :
203 CardGeneration(rs, initial_byte_size, level, ct),
204 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
205 _debug_collection_type(Concurrent_collection_type),
206 _did_compact(false)
207 {
208 HeapWord* bottom = (HeapWord*) _virtual_space.low();
209 HeapWord* end = (HeapWord*) _virtual_space.high();
210
211 _direct_allocated_words = 0;
212 NOT_PRODUCT(
213 _numObjectsPromoted = 0;
214 _numWordsPromoted = 0;
215 _numObjectsAllocated = 0;
216 _numWordsAllocated = 0;
217 )
218
219 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
220 use_adaptive_freelists,
221 dictionaryChoice);
222 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
223 if (_cmsSpace == NULL) {
224 vm_exit_during_initialization(
225 "CompactibleFreeListSpace allocation failure");
226 }
227 _cmsSpace->_gen = this;
228
229 _gc_stats = new CMSGCStats();
230
231 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
232 // offsets match. The ability to tell free chunks from objects
233 // depends on this property.
234 debug_only(
235 FreeChunk* junk = NULL;
236 assert(UseCompressedClassPointers ||
237 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
238 "Offset of FreeChunk::_prev within FreeChunk must match"
239 " that of OopDesc::_klass within OopDesc");
240 )
241 if (CollectedHeap::use_parallel_gc_threads()) {
242 typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
243 _par_gc_thread_states =
244 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC);
245 if (_par_gc_thread_states == NULL) {
246 vm_exit_during_initialization("Could not allocate par gc structs");
247 }
248 for (uint i = 0; i < ParallelGCThreads; i++) {
249 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
250 if (_par_gc_thread_states[i] == NULL) {
251 vm_exit_during_initialization("Could not allocate par gc structs");
252 }
253 }
254 } else {
255 _par_gc_thread_states = NULL;
256 }
257 _incremental_collection_failed = false;
258 // The "dilatation_factor" is the expansion that can occur on
259 // account of the fact that the minimum object size in the CMS
260 // generation may be larger than that in, say, a contiguous young
261 // generation.
262 // Ideally, in the calculation below, we'd compute the dilatation
263 // factor as: MinChunkSize/(promoting_gen's min object size)
264 // Since we do not have such a general query interface for the
265 // promoting generation, we'll instead just use the minimum
266 // object size (which today is a header's worth of space);
267 // note that all arithmetic is in units of HeapWords.
268 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
269 assert(_dilatation_factor >= 1.0, "from previous assert");
270
271 // Support for CMSFastPromotionFailure
272 reset_promotion_failed();
273 }
274
275
276 // The field "_initiating_occupancy" represents the occupancy percentage
277 // at which we trigger a new collection cycle. Unless explicitly specified
278 // via CMSInitiatingOccupancyFraction (argument "io" below), it
279 // is calculated by:
280 //
281 // Let "f" be MinHeapFreeRatio in
282 //
283 // _initiating_occupancy = 100-f +
284 // f * (CMSTriggerRatio/100)
285 // where CMSTriggerRatio is the argument "tr" below.
286 //
287 // That is, if we assume the heap is at its desired maximum occupancy at the
288 // end of a collection, we let CMSTriggerRatio of the (purported) free
289 // space be allocated before initiating a new collection cycle.
290 //
291 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
292 assert(io <= 100 && tr <= 100, "Check the arguments");
293 if (io >= 0) {
294 _initiating_occupancy = (double)io / 100.0;
295 } else {
296 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
297 (double)(tr * MinHeapFreeRatio) / 100.0)
298 / 100.0;
299 }
300 }
301
302 void ConcurrentMarkSweepGeneration::ref_processor_init() {
303 assert(collector() != NULL, "no collector");
304 collector()->ref_processor_init();
305 }
306
307 void CMSCollector::ref_processor_init() {
308 if (_ref_processor == NULL) {
309 // Allocate and initialize a reference processor
310 _ref_processor =
311 new ReferenceProcessor(_span, // span
312 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
313 (int) ParallelGCThreads, // mt processing degree
314 _cmsGen->refs_discovery_is_mt(), // mt discovery
315 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
316 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
317 &_is_alive_closure); // closure for liveness info
318 // Initialize the _ref_processor field of CMSGen
319 _cmsGen->set_ref_processor(_ref_processor);
320
321 }
322 }
323
324 AdaptiveSizePolicy* CMSCollector::size_policy() {
325 GenCollectedHeap* gch = GenCollectedHeap::heap();
326 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
327 "Wrong type of heap");
328 return gch->gen_policy()->size_policy();
329 }
330
331 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
332
333 const char* gen_name = "old";
334 GenCollectorPolicy* gcp = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy();
335
336 // Generation Counters - generation 1, 1 subspace
337 _gen_counters = new GenerationCounters(gen_name, 1, 1,
338 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
339
340 _space_counters = new GSpaceCounters(gen_name, 0,
341 _virtual_space.reserved_size(),
342 this, _gen_counters);
343 }
344
345 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
346 _cms_gen(cms_gen)
347 {
348 assert(alpha <= 100, "bad value");
349 _saved_alpha = alpha;
350
351 // Initialize the alphas to the bootstrap value of 100.
352 _gc0_alpha = _cms_alpha = 100;
353
354 _cms_begin_time.update();
355 _cms_end_time.update();
356
357 _gc0_duration = 0.0;
358 _gc0_period = 0.0;
359 _gc0_promoted = 0;
360
361 _cms_duration = 0.0;
362 _cms_period = 0.0;
363 _cms_allocated = 0;
364
365 _cms_used_at_gc0_begin = 0;
366 _cms_used_at_gc0_end = 0;
367 _allow_duty_cycle_reduction = false;
368 _valid_bits = 0;
369 _icms_duty_cycle = CMSIncrementalDutyCycle;
370 }
371
372 double CMSStats::cms_free_adjustment_factor(size_t free) const {
373 // TBD: CR 6909490
374 return 1.0;
375 }
376
377 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
378 }
379
380 // If promotion failure handling is on use
381 // the padded average size of the promotion for each
382 // young generation collection.
383 double CMSStats::time_until_cms_gen_full() const {
384 size_t cms_free = _cms_gen->cmsSpace()->free();
385 GenCollectedHeap* gch = GenCollectedHeap::heap();
386 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(),
387 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
388 if (cms_free > expected_promotion) {
389 // Start a cms collection if there isn't enough space to promote
390 // for the next minor collection. Use the padded average as
391 // a safety factor.
392 cms_free -= expected_promotion;
393
394 // Adjust by the safety factor.
395 double cms_free_dbl = (double)cms_free;
396 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
397 // Apply a further correction factor which tries to adjust
398 // for recent occurance of concurrent mode failures.
399 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
400 cms_free_dbl = cms_free_dbl * cms_adjustment;
401
402 if (PrintGCDetails && Verbose) {
403 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
404 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
405 cms_free, expected_promotion);
406 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
407 cms_free_dbl, cms_consumption_rate() + 1.0);
408 }
409 // Add 1 in case the consumption rate goes to zero.
410 return cms_free_dbl / (cms_consumption_rate() + 1.0);
411 }
412 return 0.0;
413 }
414
415 // Compare the duration of the cms collection to the
416 // time remaining before the cms generation is empty.
417 // Note that the time from the start of the cms collection
418 // to the start of the cms sweep (less than the total
419 // duration of the cms collection) can be used. This
420 // has been tried and some applications experienced
421 // promotion failures early in execution. This was
422 // possibly because the averages were not accurate
423 // enough at the beginning.
424 double CMSStats::time_until_cms_start() const {
425 // We add "gc0_period" to the "work" calculation
426 // below because this query is done (mostly) at the
427 // end of a scavenge, so we need to conservatively
428 // account for that much possible delay
429 // in the query so as to avoid concurrent mode failures
430 // due to starting the collection just a wee bit too
431 // late.
432 double work = cms_duration() + gc0_period();
433 double deadline = time_until_cms_gen_full();
434 // If a concurrent mode failure occurred recently, we want to be
435 // more conservative and halve our expected time_until_cms_gen_full()
436 if (work > deadline) {
437 if (Verbose && PrintGCDetails) {
438 gclog_or_tty->print(
439 " CMSCollector: collect because of anticipated promotion "
440 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
441 gc0_period(), time_until_cms_gen_full());
442 }
443 return 0.0;
444 }
445 return work - deadline;
446 }
447
448 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the
449 // amount of change to prevent wild oscillation.
450 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle,
451 unsigned int new_duty_cycle) {
452 assert(old_duty_cycle <= 100, "bad input value");
453 assert(new_duty_cycle <= 100, "bad input value");
454
455 // Note: use subtraction with caution since it may underflow (values are
456 // unsigned). Addition is safe since we're in the range 0-100.
457 unsigned int damped_duty_cycle = new_duty_cycle;
458 if (new_duty_cycle < old_duty_cycle) {
459 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U);
460 if (new_duty_cycle + largest_delta < old_duty_cycle) {
461 damped_duty_cycle = old_duty_cycle - largest_delta;
462 }
463 } else if (new_duty_cycle > old_duty_cycle) {
464 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U);
465 if (new_duty_cycle > old_duty_cycle + largest_delta) {
466 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U);
467 }
468 }
469 assert(damped_duty_cycle <= 100, "invalid duty cycle computed");
470
471 if (CMSTraceIncrementalPacing) {
472 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ",
473 old_duty_cycle, new_duty_cycle, damped_duty_cycle);
474 }
475 return damped_duty_cycle;
476 }
477
478 unsigned int CMSStats::icms_update_duty_cycle_impl() {
479 assert(CMSIncrementalPacing && valid(),
480 "should be handled in icms_update_duty_cycle()");
481
482 double cms_time_so_far = cms_timer().seconds();
483 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M;
484 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far);
485
486 // Avoid division by 0.
487 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01);
488 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full;
489
490 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U);
491 if (new_duty_cycle > _icms_duty_cycle) {
492 // Avoid very small duty cycles (1 or 2); 0 is allowed.
493 if (new_duty_cycle > 2) {
494 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle,
495 new_duty_cycle);
496 }
497 } else if (_allow_duty_cycle_reduction) {
498 // The duty cycle is reduced only once per cms cycle (see record_cms_end()).
499 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle);
500 // Respect the minimum duty cycle.
501 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin;
502 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle);
503 }
504
505 if (PrintGCDetails || CMSTraceIncrementalPacing) {
506 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle);
507 }
508
509 _allow_duty_cycle_reduction = false;
510 return _icms_duty_cycle;
511 }
512
513 #ifndef PRODUCT
514 void CMSStats::print_on(outputStream *st) const {
515 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
516 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
517 gc0_duration(), gc0_period(), gc0_promoted());
518 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
519 cms_duration(), cms_duration_per_mb(),
520 cms_period(), cms_allocated());
521 st->print(",cms_since_beg=%g,cms_since_end=%g",
522 cms_time_since_begin(), cms_time_since_end());
523 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
524 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
525 if (CMSIncrementalMode) {
526 st->print(",dc=%d", icms_duty_cycle());
527 }
528
529 if (valid()) {
530 st->print(",promo_rate=%g,cms_alloc_rate=%g",
531 promotion_rate(), cms_allocation_rate());
532 st->print(",cms_consumption_rate=%g,time_until_full=%g",
533 cms_consumption_rate(), time_until_cms_gen_full());
534 }
535 st->print(" ");
536 }
537 #endif // #ifndef PRODUCT
538
539 CMSCollector::CollectorState CMSCollector::_collectorState =
540 CMSCollector::Idling;
541 bool CMSCollector::_foregroundGCIsActive = false;
542 bool CMSCollector::_foregroundGCShouldWait = false;
543
544 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
545 CardTableRS* ct,
546 ConcurrentMarkSweepPolicy* cp):
547 _cmsGen(cmsGen),
548 _ct(ct),
549 _ref_processor(NULL), // will be set later
550 _conc_workers(NULL), // may be set later
551 _abort_preclean(false),
552 _start_sampling(false),
553 _between_prologue_and_epilogue(false),
554 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
555 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
556 -1 /* lock-free */, "No_lock" /* dummy */),
557 _modUnionClosure(&_modUnionTable),
558 _modUnionClosurePar(&_modUnionTable),
559 // Adjust my span to cover old (cms) gen
560 _span(cmsGen->reserved()),
561 // Construct the is_alive_closure with _span & markBitMap
562 _is_alive_closure(_span, &_markBitMap),
563 _restart_addr(NULL),
564 _overflow_list(NULL),
565 _stats(cmsGen),
566 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true)),
567 _eden_chunk_array(NULL), // may be set in ctor body
568 _eden_chunk_capacity(0), // -- ditto --
569 _eden_chunk_index(0), // -- ditto --
570 _survivor_plab_array(NULL), // -- ditto --
571 _survivor_chunk_array(NULL), // -- ditto --
572 _survivor_chunk_capacity(0), // -- ditto --
573 _survivor_chunk_index(0), // -- ditto --
574 _ser_pmc_preclean_ovflw(0),
575 _ser_kac_preclean_ovflw(0),
576 _ser_pmc_remark_ovflw(0),
577 _par_pmc_remark_ovflw(0),
578 _ser_kac_ovflw(0),
579 _par_kac_ovflw(0),
580 #ifndef PRODUCT
581 _num_par_pushes(0),
582 #endif
583 _collection_count_start(0),
584 _verifying(false),
585 _icms_start_limit(NULL),
586 _icms_stop_limit(NULL),
587 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
588 _completed_initialization(false),
589 _collector_policy(cp),
590 _should_unload_classes(CMSClassUnloadingEnabled),
591 _concurrent_cycles_since_last_unload(0),
592 _roots_scanning_options(SharedHeap::SO_None),
593 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
594 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
595 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
596 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
597 _cms_start_registered(false)
598 {
599 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
600 ExplicitGCInvokesConcurrent = true;
601 }
602 // Now expand the span and allocate the collection support structures
603 // (MUT, marking bit map etc.) to cover both generations subject to
604 // collection.
605
606 // For use by dirty card to oop closures.
607 _cmsGen->cmsSpace()->set_collector(this);
608
609 // Allocate MUT and marking bit map
610 {
611 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
612 if (!_markBitMap.allocate(_span)) {
613 warning("Failed to allocate CMS Bit Map");
614 return;
615 }
616 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
617 }
618 {
619 _modUnionTable.allocate(_span);
620 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
621 }
622
623 if (!_markStack.allocate(MarkStackSize)) {
624 warning("Failed to allocate CMS Marking Stack");
625 return;
626 }
627
628 // Support for multi-threaded concurrent phases
629 if (CMSConcurrentMTEnabled) {
630 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
631 // just for now
632 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
633 }
634 if (ConcGCThreads > 1) {
635 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
636 ConcGCThreads, true);
637 if (_conc_workers == NULL) {
638 warning("GC/CMS: _conc_workers allocation failure: "
639 "forcing -CMSConcurrentMTEnabled");
640 CMSConcurrentMTEnabled = false;
641 } else {
642 _conc_workers->initialize_workers();
643 }
644 } else {
645 CMSConcurrentMTEnabled = false;
646 }
647 }
648 if (!CMSConcurrentMTEnabled) {
649 ConcGCThreads = 0;
650 } else {
651 // Turn off CMSCleanOnEnter optimization temporarily for
652 // the MT case where it's not fixed yet; see 6178663.
653 CMSCleanOnEnter = false;
654 }
655 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
656 "Inconsistency");
657
658 // Parallel task queues; these are shared for the
659 // concurrent and stop-world phases of CMS, but
660 // are not shared with parallel scavenge (ParNew).
661 {
662 uint i;
663 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
664
665 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
666 || ParallelRefProcEnabled)
667 && num_queues > 0) {
668 _task_queues = new OopTaskQueueSet(num_queues);
669 if (_task_queues == NULL) {
670 warning("task_queues allocation failure.");
671 return;
672 }
673 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
674 if (_hash_seed == NULL) {
675 warning("_hash_seed array allocation failure");
676 return;
677 }
678
679 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
680 for (i = 0; i < num_queues; i++) {
681 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
682 if (q == NULL) {
683 warning("work_queue allocation failure.");
684 return;
685 }
686 _task_queues->register_queue(i, q);
687 }
688 for (i = 0; i < num_queues; i++) {
689 _task_queues->queue(i)->initialize();
690 _hash_seed[i] = 17; // copied from ParNew
691 }
692 }
693 }
694
695 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
696
697 // Clip CMSBootstrapOccupancy between 0 and 100.
698 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100;
699
700 _full_gcs_since_conc_gc = 0;
701
702 // Now tell CMS generations the identity of their collector
703 ConcurrentMarkSweepGeneration::set_collector(this);
704
705 // Create & start a CMS thread for this CMS collector
706 _cmsThread = ConcurrentMarkSweepThread::start(this);
707 assert(cmsThread() != NULL, "CMS Thread should have been created");
708 assert(cmsThread()->collector() == this,
709 "CMS Thread should refer to this gen");
710 assert(CGC_lock != NULL, "Where's the CGC_lock?");
711
712 // Support for parallelizing young gen rescan
713 GenCollectedHeap* gch = GenCollectedHeap::heap();
714 _young_gen = gch->prev_gen(_cmsGen);
715 if (gch->supports_inline_contig_alloc()) {
716 _top_addr = gch->top_addr();
717 _end_addr = gch->end_addr();
718 assert(_young_gen != NULL, "no _young_gen");
719 _eden_chunk_index = 0;
720 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
721 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
722 if (_eden_chunk_array == NULL) {
723 _eden_chunk_capacity = 0;
724 warning("GC/CMS: _eden_chunk_array allocation failure");
725 }
726 }
727 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
728
729 // Support for parallelizing survivor space rescan
730 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
731 const size_t max_plab_samples =
732 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize;
733
734 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
735 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC);
736 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
737 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
738 || _cursor == NULL) {
739 warning("Failed to allocate survivor plab/chunk array");
740 if (_survivor_plab_array != NULL) {
741 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC);
742 _survivor_plab_array = NULL;
743 }
744 if (_survivor_chunk_array != NULL) {
745 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC);
746 _survivor_chunk_array = NULL;
747 }
748 if (_cursor != NULL) {
749 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC);
750 _cursor = NULL;
751 }
752 } else {
753 _survivor_chunk_capacity = 2*max_plab_samples;
754 for (uint i = 0; i < ParallelGCThreads; i++) {
755 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
756 if (vec == NULL) {
757 warning("Failed to allocate survivor plab array");
758 for (int j = i; j > 0; j--) {
759 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC);
760 }
761 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC);
762 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC);
763 _survivor_plab_array = NULL;
764 _survivor_chunk_array = NULL;
765 _survivor_chunk_capacity = 0;
766 break;
767 } else {
768 ChunkArray* cur =
769 ::new (&_survivor_plab_array[i]) ChunkArray(vec,
770 max_plab_samples);
771 assert(cur->end() == 0, "Should be 0");
772 assert(cur->array() == vec, "Should be vec");
773 assert(cur->capacity() == max_plab_samples, "Error");
774 }
775 }
776 }
777 }
778 assert( ( _survivor_plab_array != NULL
779 && _survivor_chunk_array != NULL)
780 || ( _survivor_chunk_capacity == 0
781 && _survivor_chunk_index == 0),
782 "Error");
783
784 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
785 _gc_counters = new CollectorCounters("CMS", 1);
786 _completed_initialization = true;
787 _inter_sweep_timer.start(); // start of time
788 }
789
790 const char* ConcurrentMarkSweepGeneration::name() const {
791 return "concurrent mark-sweep generation";
792 }
793 void ConcurrentMarkSweepGeneration::update_counters() {
794 if (UsePerfData) {
795 _space_counters->update_all();
796 _gen_counters->update_all();
797 }
798 }
799
800 // this is an optimized version of update_counters(). it takes the
801 // used value as a parameter rather than computing it.
802 //
803 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
804 if (UsePerfData) {
805 _space_counters->update_used(used);
806 _space_counters->update_capacity();
807 _gen_counters->update_all();
808 }
809 }
810
811 void ConcurrentMarkSweepGeneration::print() const {
812 Generation::print();
813 cmsSpace()->print();
814 }
815
816 #ifndef PRODUCT
817 void ConcurrentMarkSweepGeneration::print_statistics() {
818 cmsSpace()->printFLCensus(0);
819 }
820 #endif
821
822 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
823 GenCollectedHeap* gch = GenCollectedHeap::heap();
824 if (PrintGCDetails) {
825 if (Verbose) {
826 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
827 level(), short_name(), s, used(), capacity());
828 } else {
829 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
830 level(), short_name(), s, used() / K, capacity() / K);
831 }
832 }
833 if (Verbose) {
834 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
835 gch->used(), gch->capacity());
836 } else {
837 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
838 gch->used() / K, gch->capacity() / K);
839 }
840 }
841
842 size_t
843 ConcurrentMarkSweepGeneration::contiguous_available() const {
844 // dld proposes an improvement in precision here. If the committed
845 // part of the space ends in a free block we should add that to
846 // uncommitted size in the calculation below. Will make this
847 // change later, staying with the approximation below for the
848 // time being. -- ysr.
849 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
850 }
851
852 size_t
853 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
854 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
855 }
856
857 size_t ConcurrentMarkSweepGeneration::max_available() const {
858 return free() + _virtual_space.uncommitted_size();
859 }
860
861 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
862 size_t available = max_available();
863 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
864 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
865 if (Verbose && PrintGCDetails) {
866 gclog_or_tty->print_cr(
867 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
868 "max_promo("SIZE_FORMAT")",
869 res? "":" not", available, res? ">=":"<",
870 av_promo, max_promotion_in_bytes);
871 }
872 return res;
873 }
874
875 // At a promotion failure dump information on block layout in heap
876 // (cms old generation).
877 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
878 if (CMSDumpAtPromotionFailure) {
879 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
880 }
881 if (CMSFastPromotionFailure) {
882 reset_promotion_failed();
883 }
884 }
885
886 CompactibleSpace*
887 ConcurrentMarkSweepGeneration::first_compaction_space() const {
888 return _cmsSpace;
889 }
890
891 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
892 // Clear the promotion information. These pointers can be adjusted
893 // along with all the other pointers into the heap but
894 // compaction is expected to be a rare event with
895 // a heap using cms so don't do it without seeing the need.
896 if (CollectedHeap::use_parallel_gc_threads()) {
897 for (uint i = 0; i < ParallelGCThreads; i++) {
898 _par_gc_thread_states[i]->promo.reset();
899 }
900 }
901 }
902
903 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
904 blk->do_space(_cmsSpace);
905 }
906
907 void ConcurrentMarkSweepGeneration::compute_new_size() {
908 assert_locked_or_safepoint(Heap_lock);
909
910 // If incremental collection failed, we just want to expand
911 // to the limit.
912 if (incremental_collection_failed()) {
913 clear_incremental_collection_failed();
914 grow_to_reserved();
915 return;
916 }
917
918 // The heap has been compacted but not reset yet.
919 // Any metric such as free() or used() will be incorrect.
920
921 CardGeneration::compute_new_size();
922
923 // Reset again after a possible resizing
924 if (did_compact()) {
925 cmsSpace()->reset_after_compaction();
926 }
927 }
928
929 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
930 assert_locked_or_safepoint(Heap_lock);
931
932 // If incremental collection failed, we just want to expand
933 // to the limit.
934 if (incremental_collection_failed()) {
935 clear_incremental_collection_failed();
936 grow_to_reserved();
937 return;
938 }
939
940 double free_percentage = ((double) free()) / capacity();
941 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
942 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
943
944 // compute expansion delta needed for reaching desired free percentage
945 if (free_percentage < desired_free_percentage) {
946 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
947 assert(desired_capacity >= capacity(), "invalid expansion size");
948 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
949 if (PrintGCDetails && Verbose) {
950 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
951 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
952 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
953 gclog_or_tty->print_cr(" Desired free fraction %f",
954 desired_free_percentage);
955 gclog_or_tty->print_cr(" Maximum free fraction %f",
956 maximum_free_percentage);
957 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000);
958 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
959 desired_capacity/1000);
960 int prev_level = level() - 1;
961 if (prev_level >= 0) {
962 size_t prev_size = 0;
963 GenCollectedHeap* gch = GenCollectedHeap::heap();
964 Generation* prev_gen = gch->_gens[prev_level];
965 prev_size = prev_gen->capacity();
966 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
967 prev_size/1000);
968 }
969 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
970 unsafe_max_alloc_nogc()/1000);
971 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
972 contiguous_available()/1000);
973 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
974 expand_bytes);
975 }
976 // safe if expansion fails
977 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
978 if (PrintGCDetails && Verbose) {
979 gclog_or_tty->print_cr(" Expanded free fraction %f",
980 ((double) free()) / capacity());
981 }
982 } else {
983 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
984 assert(desired_capacity <= capacity(), "invalid expansion size");
985 size_t shrink_bytes = capacity() - desired_capacity;
986 // Don't shrink unless the delta is greater than the minimum shrink we want
987 if (shrink_bytes >= MinHeapDeltaBytes) {
988 shrink_free_list_by(shrink_bytes);
989 }
990 }
991 }
992
993 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
994 return cmsSpace()->freelistLock();
995 }
996
997 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
998 bool tlab) {
999 CMSSynchronousYieldRequest yr;
1000 MutexLockerEx x(freelistLock(),
1001 Mutex::_no_safepoint_check_flag);
1002 return have_lock_and_allocate(size, tlab);
1003 }
1004
1005 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
1006 bool tlab /* ignored */) {
1007 assert_lock_strong(freelistLock());
1008 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
1009 HeapWord* res = cmsSpace()->allocate(adjustedSize);
1010 // Allocate the object live (grey) if the background collector has
1011 // started marking. This is necessary because the marker may
1012 // have passed this address and consequently this object will
1013 // not otherwise be greyed and would be incorrectly swept up.
1014 // Note that if this object contains references, the writing
1015 // of those references will dirty the card containing this object
1016 // allowing the object to be blackened (and its references scanned)
1017 // either during a preclean phase or at the final checkpoint.
1018 if (res != NULL) {
1019 // We may block here with an uninitialized object with
1020 // its mark-bit or P-bits not yet set. Such objects need
1021 // to be safely navigable by block_start().
1022 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
1023 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
1024 collector()->direct_allocated(res, adjustedSize);
1025 _direct_allocated_words += adjustedSize;
1026 // allocation counters
1027 NOT_PRODUCT(
1028 _numObjectsAllocated++;
1029 _numWordsAllocated += (int)adjustedSize;
1030 )
1031 }
1032 return res;
1033 }
1034
1035 // In the case of direct allocation by mutators in a generation that
1036 // is being concurrently collected, the object must be allocated
1037 // live (grey) if the background collector has started marking.
1038 // This is necessary because the marker may
1039 // have passed this address and consequently this object will
1040 // not otherwise be greyed and would be incorrectly swept up.
1041 // Note that if this object contains references, the writing
1042 // of those references will dirty the card containing this object
1043 // allowing the object to be blackened (and its references scanned)
1044 // either during a preclean phase or at the final checkpoint.
1045 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
1046 assert(_markBitMap.covers(start, size), "Out of bounds");
1047 if (_collectorState >= Marking) {
1048 MutexLockerEx y(_markBitMap.lock(),
1049 Mutex::_no_safepoint_check_flag);
1050 // [see comments preceding SweepClosure::do_blk() below for details]
1051 //
1052 // Can the P-bits be deleted now? JJJ
1053 //
1054 // 1. need to mark the object as live so it isn't collected
1055 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
1056 // 3. need to mark the end of the object so marking, precleaning or sweeping
1057 // can skip over uninitialized or unparsable objects. An allocated
1058 // object is considered uninitialized for our purposes as long as
1059 // its klass word is NULL. All old gen objects are parsable
1060 // as soon as they are initialized.)
1061 _markBitMap.mark(start); // object is live
1062 _markBitMap.mark(start + 1); // object is potentially uninitialized?
1063 _markBitMap.mark(start + size - 1);
1064 // mark end of object
1065 }
1066 // check that oop looks uninitialized
1067 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
1068 }
1069
1070 void CMSCollector::promoted(bool par, HeapWord* start,
1071 bool is_obj_array, size_t obj_size) {
1072 assert(_markBitMap.covers(start), "Out of bounds");
1073 // See comment in direct_allocated() about when objects should
1074 // be allocated live.
1075 if (_collectorState >= Marking) {
1076 // we already hold the marking bit map lock, taken in
1077 // the prologue
1078 if (par) {
1079 _markBitMap.par_mark(start);
1080 } else {
1081 _markBitMap.mark(start);
1082 }
1083 // We don't need to mark the object as uninitialized (as
1084 // in direct_allocated above) because this is being done with the
1085 // world stopped and the object will be initialized by the
1086 // time the marking, precleaning or sweeping get to look at it.
1087 // But see the code for copying objects into the CMS generation,
1088 // where we need to ensure that concurrent readers of the
1089 // block offset table are able to safely navigate a block that
1090 // is in flux from being free to being allocated (and in
1091 // transition while being copied into) and subsequently
1092 // becoming a bona-fide object when the copy/promotion is complete.
1093 assert(SafepointSynchronize::is_at_safepoint(),
1094 "expect promotion only at safepoints");
1095
1096 if (_collectorState < Sweeping) {
1097 // Mark the appropriate cards in the modUnionTable, so that
1098 // this object gets scanned before the sweep. If this is
1099 // not done, CMS generation references in the object might
1100 // not get marked.
1101 // For the case of arrays, which are otherwise precisely
1102 // marked, we need to dirty the entire array, not just its head.
1103 if (is_obj_array) {
1104 // The [par_]mark_range() method expects mr.end() below to
1105 // be aligned to the granularity of a bit's representation
1106 // in the heap. In the case of the MUT below, that's a
1107 // card size.
1108 MemRegion mr(start,
1109 (HeapWord*)round_to((intptr_t)(start + obj_size),
1110 CardTableModRefBS::card_size /* bytes */));
1111 if (par) {
1112 _modUnionTable.par_mark_range(mr);
1113 } else {
1114 _modUnionTable.mark_range(mr);
1115 }
1116 } else { // not an obj array; we can just mark the head
1117 if (par) {
1118 _modUnionTable.par_mark(start);
1119 } else {
1120 _modUnionTable.mark(start);
1121 }
1122 }
1123 }
1124 }
1125 }
1126
1127 static inline size_t percent_of_space(Space* space, HeapWord* addr)
1128 {
1129 size_t delta = pointer_delta(addr, space->bottom());
1130 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
1131 }
1132
1133 void CMSCollector::icms_update_allocation_limits()
1134 {
1135 Generation* young = GenCollectedHeap::heap()->get_gen(0);
1136 EdenSpace* eden = young->as_DefNewGeneration()->eden();
1137
1138 const unsigned int duty_cycle = stats().icms_update_duty_cycle();
1139 if (CMSTraceIncrementalPacing) {
1140 stats().print();
1141 }
1142
1143 assert(duty_cycle <= 100, "invalid duty cycle");
1144 if (duty_cycle != 0) {
1145 // The duty_cycle is a percentage between 0 and 100; convert to words and
1146 // then compute the offset from the endpoints of the space.
1147 size_t free_words = eden->free() / HeapWordSize;
1148 double free_words_dbl = (double)free_words;
1149 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
1150 size_t offset_words = (free_words - duty_cycle_words) / 2;
1151
1152 _icms_start_limit = eden->top() + offset_words;
1153 _icms_stop_limit = eden->end() - offset_words;
1154
1155 // The limits may be adjusted (shifted to the right) by
1156 // CMSIncrementalOffset, to allow the application more mutator time after a
1157 // young gen gc (when all mutators were stopped) and before CMS starts and
1158 // takes away one or more cpus.
1159 if (CMSIncrementalOffset != 0) {
1160 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
1161 size_t adjustment = (size_t)adjustment_dbl;
1162 HeapWord* tmp_stop = _icms_stop_limit + adjustment;
1163 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
1164 _icms_start_limit += adjustment;
1165 _icms_stop_limit = tmp_stop;
1166 }
1167 }
1168 }
1169 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
1170 _icms_start_limit = _icms_stop_limit = eden->end();
1171 }
1172
1173 // Install the new start limit.
1174 eden->set_soft_end(_icms_start_limit);
1175
1176 if (CMSTraceIncrementalMode) {
1177 gclog_or_tty->print(" icms alloc limits: "
1178 PTR_FORMAT "," PTR_FORMAT
1179 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
1180 p2i(_icms_start_limit), p2i(_icms_stop_limit),
1181 percent_of_space(eden, _icms_start_limit),
1182 percent_of_space(eden, _icms_stop_limit));
1183 if (Verbose) {
1184 gclog_or_tty->print("eden: ");
1185 eden->print_on(gclog_or_tty);
1186 }
1187 }
1188 }
1189
1190 // Any changes here should try to maintain the invariant
1191 // that if this method is called with _icms_start_limit
1192 // and _icms_stop_limit both NULL, then it should return NULL
1193 // and not notify the icms thread.
1194 HeapWord*
1195 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
1196 size_t word_size)
1197 {
1198 // A start_limit equal to end() means the duty cycle is 0, so treat that as a
1199 // nop.
1200 if (CMSIncrementalMode && _icms_start_limit != space->end()) {
1201 if (top <= _icms_start_limit) {
1202 if (CMSTraceIncrementalMode) {
1203 space->print_on(gclog_or_tty);
1204 gclog_or_tty->stamp();
1205 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
1206 ", new limit=" PTR_FORMAT
1207 " (" SIZE_FORMAT "%%)",
1208 p2i(top), p2i(_icms_stop_limit),
1209 percent_of_space(space, _icms_stop_limit));
1210 }
1211 ConcurrentMarkSweepThread::start_icms();
1212 assert(top < _icms_stop_limit, "Tautology");
1213 if (word_size < pointer_delta(_icms_stop_limit, top)) {
1214 return _icms_stop_limit;
1215 }
1216
1217 // The allocation will cross both the _start and _stop limits, so do the
1218 // stop notification also and return end().
1219 if (CMSTraceIncrementalMode) {
1220 space->print_on(gclog_or_tty);
1221 gclog_or_tty->stamp();
1222 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
1223 ", new limit=" PTR_FORMAT
1224 " (" SIZE_FORMAT "%%)",
1225 p2i(top), p2i(space->end()),
1226 percent_of_space(space, space->end()));
1227 }
1228 ConcurrentMarkSweepThread::stop_icms();
1229 return space->end();
1230 }
1231
1232 if (top <= _icms_stop_limit) {
1233 if (CMSTraceIncrementalMode) {
1234 space->print_on(gclog_or_tty);
1235 gclog_or_tty->stamp();
1236 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
1237 ", new limit=" PTR_FORMAT
1238 " (" SIZE_FORMAT "%%)",
1239 top, space->end(),
1240 percent_of_space(space, space->end()));
1241 }
1242 ConcurrentMarkSweepThread::stop_icms();
1243 return space->end();
1244 }
1245
1246 if (CMSTraceIncrementalMode) {
1247 space->print_on(gclog_or_tty);
1248 gclog_or_tty->stamp();
1249 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
1250 ", new limit=" PTR_FORMAT,
1251 top, NULL);
1252 }
1253 }
1254
1255 return NULL;
1256 }
1257
1258 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
1259 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1260 // allocate, copy and if necessary update promoinfo --
1261 // delegate to underlying space.
1262 assert_lock_strong(freelistLock());
1263
1264 #ifndef PRODUCT
1265 if (Universe::heap()->promotion_should_fail()) {
1266 return NULL;
1267 }
1268 #endif // #ifndef PRODUCT
1269
1270 oop res = _cmsSpace->promote(obj, obj_size);
1271 if (res == NULL) {
1272 // expand and retry
1273 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
1274 expand(s*HeapWordSize, MinHeapDeltaBytes,
1275 CMSExpansionCause::_satisfy_promotion);
1276 // Since there's currently no next generation, we don't try to promote
1277 // into a more senior generation.
1278 assert(next_gen() == NULL, "assumption, based upon which no attempt "
1279 "is made to pass on a possibly failing "
1280 "promotion to next generation");
1281 res = _cmsSpace->promote(obj, obj_size);
1282 }
1283 if (res != NULL) {
1284 // See comment in allocate() about when objects should
1285 // be allocated live.
1286 assert(obj->is_oop(), "Will dereference klass pointer below");
1287 collector()->promoted(false, // Not parallel
1288 (HeapWord*)res, obj->is_objArray(), obj_size);
1289 // promotion counters
1290 NOT_PRODUCT(
1291 _numObjectsPromoted++;
1292 _numWordsPromoted +=
1293 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1294 )
1295 }
1296 return res;
1297 }
1298
1299
1300 HeapWord*
1301 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
1302 HeapWord* top,
1303 size_t word_sz)
1304 {
1305 return collector()->allocation_limit_reached(space, top, word_sz);
1306 }
1307
1308 // IMPORTANT: Notes on object size recognition in CMS.
1309 // ---------------------------------------------------
1310 // A block of storage in the CMS generation is always in
1311 // one of three states. A free block (FREE), an allocated
1312 // object (OBJECT) whose size() method reports the correct size,
1313 // and an intermediate state (TRANSIENT) in which its size cannot
1314 // be accurately determined.
1315 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1316 // -----------------------------------------------------
1317 // FREE: klass_word & 1 == 1; mark_word holds block size
1318 //
1319 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1320 // obj->size() computes correct size
1321 //
1322 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1323 //
1324 // STATE IDENTIFICATION: (64 bit+COOPS)
1325 // ------------------------------------
1326 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1327 //
1328 // OBJECT: klass_word installed; klass_word != 0;
1329 // obj->size() computes correct size
1330 //
1331 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1332 //
1333 //
1334 // STATE TRANSITION DIAGRAM
1335 //
1336 // mut / parnew mut / parnew
1337 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1338 // ^ |
1339 // |------------------------ DEAD <------------------------------------|
1340 // sweep mut
1341 //
1342 // While a block is in TRANSIENT state its size cannot be determined
1343 // so readers will either need to come back later or stall until
1344 // the size can be determined. Note that for the case of direct
1345 // allocation, P-bits, when available, may be used to determine the
1346 // size of an object that may not yet have been initialized.
1347
1348 // Things to support parallel young-gen collection.
1349 oop
1350 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1351 oop old, markOop m,
1352 size_t word_sz) {
1353 #ifndef PRODUCT
1354 if (Universe::heap()->promotion_should_fail()) {
1355 return NULL;
1356 }
1357 #endif // #ifndef PRODUCT
1358 if (CMSFastPromotionFailure && has_promotion_failed()) {
1359 return NULL;
1360 }
1361
1362 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1363 PromotionInfo* promoInfo = &ps->promo;
1364 // if we are tracking promotions, then first ensure space for
1365 // promotion (including spooling space for saving header if necessary).
1366 // then allocate and copy, then track promoted info if needed.
1367 // When tracking (see PromotionInfo::track()), the mark word may
1368 // be displaced and in this case restoration of the mark word
1369 // occurs in the (oop_since_save_marks_)iterate phase.
1370 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1371 // Out of space for allocating spooling buffers;
1372 // try expanding and allocating spooling buffers.
1373 if (!expand_and_ensure_spooling_space(promoInfo)) {
1374 return NULL;
1375 }
1376 }
1377 assert(promoInfo->has_spooling_space(), "Control point invariant");
1378 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1379 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1380 if (obj_ptr == NULL) {
1381 if (CMSFastPromotionFailure && has_promotion_failed()) {
1382 return NULL;
1383 }
1384 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1385 if (obj_ptr == NULL) {
1386 return NULL;
1387 }
1388 }
1389 oop obj = oop(obj_ptr);
1390 OrderAccess::storestore();
1391 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1392 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1393 // IMPORTANT: See note on object initialization for CMS above.
1394 // Otherwise, copy the object. Here we must be careful to insert the
1395 // klass pointer last, since this marks the block as an allocated object.
1396 // Except with compressed oops it's the mark word.
1397 HeapWord* old_ptr = (HeapWord*)old;
1398 // Restore the mark word copied above.
1399 obj->set_mark(m);
1400 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1401 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1402 OrderAccess::storestore();
1403
1404 if (UseCompressedClassPointers) {
1405 // Copy gap missed by (aligned) header size calculation below
1406 obj->set_klass_gap(old->klass_gap());
1407 }
1408 if (word_sz > (size_t)oopDesc::header_size()) {
1409 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1410 obj_ptr + oopDesc::header_size(),
1411 word_sz - oopDesc::header_size());
1412 }
1413
1414 // Now we can track the promoted object, if necessary. We take care
1415 // to delay the transition from uninitialized to full object
1416 // (i.e., insertion of klass pointer) until after, so that it
1417 // atomically becomes a promoted object.
1418 if (promoInfo->tracking()) {
1419 promoInfo->track((PromotedObject*)obj, old->klass());
1420 }
1421 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1422 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1423 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1424
1425 // Finally, install the klass pointer (this should be volatile).
1426 OrderAccess::storestore();
1427 obj->set_klass(old->klass());
1428 // We should now be able to calculate the right size for this object
1429 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1430
1431 collector()->promoted(true, // parallel
1432 obj_ptr, old->is_objArray(), word_sz);
1433
1434 NOT_PRODUCT(
1435 Atomic::inc_ptr(&_numObjectsPromoted);
1436 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1437 )
1438
1439 return obj;
1440 }
1441
1442 void
1443 ConcurrentMarkSweepGeneration::
1444 par_promote_alloc_undo(int thread_num,
1445 HeapWord* obj, size_t word_sz) {
1446 // CMS does not support promotion undo.
1447 ShouldNotReachHere();
1448 }
1449
1450 void
1451 ConcurrentMarkSweepGeneration::
1452 par_promote_alloc_done(int thread_num) {
1453 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1454 ps->lab.retire(thread_num);
1455 }
1456
1457 void
1458 ConcurrentMarkSweepGeneration::
1459 par_oop_since_save_marks_iterate_done(int thread_num) {
1460 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1461 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1462 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1463 }
1464
1465 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1466 size_t size,
1467 bool tlab)
1468 {
1469 // We allow a STW collection only if a full
1470 // collection was requested.
1471 return full || should_allocate(size, tlab); // FIX ME !!!
1472 // This and promotion failure handling are connected at the
1473 // hip and should be fixed by untying them.
1474 }
1475
1476 bool CMSCollector::shouldConcurrentCollect() {
1477 if (_full_gc_requested) {
1478 if (Verbose && PrintGCDetails) {
1479 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1480 " gc request (or gc_locker)");
1481 }
1482 return true;
1483 }
1484
1485 // For debugging purposes, change the type of collection.
1486 // If the rotation is not on the concurrent collection
1487 // type, don't start a concurrent collection.
1488 NOT_PRODUCT(
1489 if (RotateCMSCollectionTypes &&
1490 (_cmsGen->debug_collection_type() !=
1491 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
1492 assert(_cmsGen->debug_collection_type() !=
1493 ConcurrentMarkSweepGeneration::Unknown_collection_type,
1494 "Bad cms collection type");
1495 return false;
1496 }
1497 )
1498
1499 FreelistLocker x(this);
1500 // ------------------------------------------------------------------
1501 // Print out lots of information which affects the initiation of
1502 // a collection.
1503 if (PrintCMSInitiationStatistics && stats().valid()) {
1504 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1505 gclog_or_tty->stamp();
1506 gclog_or_tty->cr();
1507 stats().print_on(gclog_or_tty);
1508 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1509 stats().time_until_cms_gen_full());
1510 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1511 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1512 _cmsGen->contiguous_available());
1513 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1514 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1515 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1516 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1517 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1518 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1519 gclog_or_tty->print_cr("metadata initialized %d",
1520 MetaspaceGC::should_concurrent_collect());
1521 }
1522 // ------------------------------------------------------------------
1523
1524 // If the estimated time to complete a cms collection (cms_duration())
1525 // is less than the estimated time remaining until the cms generation
1526 // is full, start a collection.
1527 if (!UseCMSInitiatingOccupancyOnly) {
1528 if (stats().valid()) {
1529 if (stats().time_until_cms_start() == 0.0) {
1530 return true;
1531 }
1532 } else {
1533 // We want to conservatively collect somewhat early in order
1534 // to try and "bootstrap" our CMS/promotion statistics;
1535 // this branch will not fire after the first successful CMS
1536 // collection because the stats should then be valid.
1537 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1538 if (Verbose && PrintGCDetails) {
1539 gclog_or_tty->print_cr(
1540 " CMSCollector: collect for bootstrapping statistics:"
1541 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1542 _bootstrap_occupancy);
1543 }
1544 return true;
1545 }
1546 }
1547 }
1548
1549 // Otherwise, we start a collection cycle if
1550 // old gen want a collection cycle started. Each may use
1551 // an appropriate criterion for making this decision.
1552 // XXX We need to make sure that the gen expansion
1553 // criterion dovetails well with this. XXX NEED TO FIX THIS
1554 if (_cmsGen->should_concurrent_collect()) {
1555 if (Verbose && PrintGCDetails) {
1556 gclog_or_tty->print_cr("CMS old gen initiated");
1557 }
1558 return true;
1559 }
1560
1561 // We start a collection if we believe an incremental collection may fail;
1562 // this is not likely to be productive in practice because it's probably too
1563 // late anyway.
1564 GenCollectedHeap* gch = GenCollectedHeap::heap();
1565 assert(gch->collector_policy()->is_generation_policy(),
1566 "You may want to check the correctness of the following");
1567 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1568 if (Verbose && PrintGCDetails) {
1569 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1570 }
1571 return true;
1572 }
1573
1574 if (MetaspaceGC::should_concurrent_collect()) {
1575 if (Verbose && PrintGCDetails) {
1576 gclog_or_tty->print("CMSCollector: collect for metadata allocation ");
1577 }
1578 return true;
1579 }
1580
1581 // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1582 if (CMSTriggerInterval >= 0) {
1583 if (CMSTriggerInterval == 0) {
1584 // Trigger always
1585 return true;
1586 }
1587
1588 // Check the CMS time since begin (we do not check the stats validity
1589 // as we want to be able to trigger the first CMS cycle as well)
1590 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1591 if (Verbose && PrintGCDetails) {
1592 if (stats().valid()) {
1593 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1594 stats().cms_time_since_begin());
1595 } else {
1596 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)");
1597 }
1598 }
1599 return true;
1600 }
1601 }
1602
1603 return false;
1604 }
1605
1606 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1607
1608 // Clear _expansion_cause fields of constituent generations
1609 void CMSCollector::clear_expansion_cause() {
1610 _cmsGen->clear_expansion_cause();
1611 }
1612
1613 // We should be conservative in starting a collection cycle. To
1614 // start too eagerly runs the risk of collecting too often in the
1615 // extreme. To collect too rarely falls back on full collections,
1616 // which works, even if not optimum in terms of concurrent work.
1617 // As a work around for too eagerly collecting, use the flag
1618 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1619 // giving the user an easily understandable way of controlling the
1620 // collections.
1621 // We want to start a new collection cycle if any of the following
1622 // conditions hold:
1623 // . our current occupancy exceeds the configured initiating occupancy
1624 // for this generation, or
1625 // . we recently needed to expand this space and have not, since that
1626 // expansion, done a collection of this generation, or
1627 // . the underlying space believes that it may be a good idea to initiate
1628 // a concurrent collection (this may be based on criteria such as the
1629 // following: the space uses linear allocation and linear allocation is
1630 // going to fail, or there is believed to be excessive fragmentation in
1631 // the generation, etc... or ...
1632 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1633 // the case of the old generation; see CR 6543076):
1634 // we may be approaching a point at which allocation requests may fail because
1635 // we will be out of sufficient free space given allocation rate estimates.]
1636 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1637
1638 assert_lock_strong(freelistLock());
1639 if (occupancy() > initiating_occupancy()) {
1640 if (PrintGCDetails && Verbose) {
1641 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1642 short_name(), occupancy(), initiating_occupancy());
1643 }
1644 return true;
1645 }
1646 if (UseCMSInitiatingOccupancyOnly) {
1647 return false;
1648 }
1649 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1650 if (PrintGCDetails && Verbose) {
1651 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1652 short_name());
1653 }
1654 return true;
1655 }
1656 if (_cmsSpace->should_concurrent_collect()) {
1657 if (PrintGCDetails && Verbose) {
1658 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1659 short_name());
1660 }
1661 return true;
1662 }
1663 return false;
1664 }
1665
1666 void ConcurrentMarkSweepGeneration::collect(bool full,
1667 bool clear_all_soft_refs,
1668 size_t size,
1669 bool tlab)
1670 {
1671 collector()->collect(full, clear_all_soft_refs, size, tlab);
1672 }
1673
1674 void CMSCollector::collect(bool full,
1675 bool clear_all_soft_refs,
1676 size_t size,
1677 bool tlab)
1678 {
1679 if (!UseCMSCollectionPassing && _collectorState > Idling) {
1680 // For debugging purposes skip the collection if the state
1681 // is not currently idle
1682 if (TraceCMSState) {
1683 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
1684 Thread::current(), full, _collectorState);
1685 }
1686 return;
1687 }
1688
1689 // The following "if" branch is present for defensive reasons.
1690 // In the current uses of this interface, it can be replaced with:
1691 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1692 // But I am not placing that assert here to allow future
1693 // generality in invoking this interface.
1694 if (GC_locker::is_active()) {
1695 // A consistency test for GC_locker
1696 assert(GC_locker::needs_gc(), "Should have been set already");
1697 // Skip this foreground collection, instead
1698 // expanding the heap if necessary.
1699 // Need the free list locks for the call to free() in compute_new_size()
1700 compute_new_size();
1701 return;
1702 }
1703 acquire_control_and_collect(full, clear_all_soft_refs);
1704 _full_gcs_since_conc_gc++;
1705 }
1706
1707 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1708 GenCollectedHeap* gch = GenCollectedHeap::heap();
1709 unsigned int gc_count = gch->total_full_collections();
1710 if (gc_count == full_gc_count) {
1711 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1712 _full_gc_requested = true;
1713 _full_gc_cause = cause;
1714 CGC_lock->notify(); // nudge CMS thread
1715 } else {
1716 assert(gc_count > full_gc_count, "Error: causal loop");
1717 }
1718 }
1719
1720 bool CMSCollector::is_external_interruption() {
1721 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1722 return GCCause::is_user_requested_gc(cause) ||
1723 GCCause::is_serviceability_requested_gc(cause);
1724 }
1725
1726 void CMSCollector::report_concurrent_mode_interruption() {
1727 if (is_external_interruption()) {
1728 if (PrintGCDetails) {
1729 gclog_or_tty->print(" (concurrent mode interrupted)");
1730 }
1731 } else {
1732 if (PrintGCDetails) {
1733 gclog_or_tty->print(" (concurrent mode failure)");
1734 }
1735 _gc_tracer_cm->report_concurrent_mode_failure();
1736 }
1737 }
1738
1739
1740 // The foreground and background collectors need to coordinate in order
1741 // to make sure that they do not mutually interfere with CMS collections.
1742 // When a background collection is active,
1743 // the foreground collector may need to take over (preempt) and
1744 // synchronously complete an ongoing collection. Depending on the
1745 // frequency of the background collections and the heap usage
1746 // of the application, this preemption can be seldom or frequent.
1747 // There are only certain
1748 // points in the background collection that the "collection-baton"
1749 // can be passed to the foreground collector.
1750 //
1751 // The foreground collector will wait for the baton before
1752 // starting any part of the collection. The foreground collector
1753 // will only wait at one location.
1754 //
1755 // The background collector will yield the baton before starting a new
1756 // phase of the collection (e.g., before initial marking, marking from roots,
1757 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1758 // of the loop which switches the phases. The background collector does some
1759 // of the phases (initial mark, final re-mark) with the world stopped.
1760 // Because of locking involved in stopping the world,
1761 // the foreground collector should not block waiting for the background
1762 // collector when it is doing a stop-the-world phase. The background
1763 // collector will yield the baton at an additional point just before
1764 // it enters a stop-the-world phase. Once the world is stopped, the
1765 // background collector checks the phase of the collection. If the
1766 // phase has not changed, it proceeds with the collection. If the
1767 // phase has changed, it skips that phase of the collection. See
1768 // the comments on the use of the Heap_lock in collect_in_background().
1769 //
1770 // Variable used in baton passing.
1771 // _foregroundGCIsActive - Set to true by the foreground collector when
1772 // it wants the baton. The foreground clears it when it has finished
1773 // the collection.
1774 // _foregroundGCShouldWait - Set to true by the background collector
1775 // when it is running. The foreground collector waits while
1776 // _foregroundGCShouldWait is true.
1777 // CGC_lock - monitor used to protect access to the above variables
1778 // and to notify the foreground and background collectors.
1779 // _collectorState - current state of the CMS collection.
1780 //
1781 // The foreground collector
1782 // acquires the CGC_lock
1783 // sets _foregroundGCIsActive
1784 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1785 // various locks acquired in preparation for the collection
1786 // are released so as not to block the background collector
1787 // that is in the midst of a collection
1788 // proceeds with the collection
1789 // clears _foregroundGCIsActive
1790 // returns
1791 //
1792 // The background collector in a loop iterating on the phases of the
1793 // collection
1794 // acquires the CGC_lock
1795 // sets _foregroundGCShouldWait
1796 // if _foregroundGCIsActive is set
1797 // clears _foregroundGCShouldWait, notifies _CGC_lock
1798 // waits on _CGC_lock for _foregroundGCIsActive to become false
1799 // and exits the loop.
1800 // otherwise
1801 // proceed with that phase of the collection
1802 // if the phase is a stop-the-world phase,
1803 // yield the baton once more just before enqueueing
1804 // the stop-world CMS operation (executed by the VM thread).
1805 // returns after all phases of the collection are done
1806 //
1807
1808 void CMSCollector::acquire_control_and_collect(bool full,
1809 bool clear_all_soft_refs) {
1810 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1811 assert(!Thread::current()->is_ConcurrentGC_thread(),
1812 "shouldn't try to acquire control from self!");
1813
1814 // Start the protocol for acquiring control of the
1815 // collection from the background collector (aka CMS thread).
1816 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1817 "VM thread should have CMS token");
1818 // Remember the possibly interrupted state of an ongoing
1819 // concurrent collection
1820 CollectorState first_state = _collectorState;
1821
1822 // Signal to a possibly ongoing concurrent collection that
1823 // we want to do a foreground collection.
1824 _foregroundGCIsActive = true;
1825
1826 // Disable incremental mode during a foreground collection.
1827 ICMSDisabler icms_disabler;
1828
1829 // release locks and wait for a notify from the background collector
1830 // releasing the locks in only necessary for phases which
1831 // do yields to improve the granularity of the collection.
1832 assert_lock_strong(bitMapLock());
1833 // We need to lock the Free list lock for the space that we are
1834 // currently collecting.
1835 assert(haveFreelistLocks(), "Must be holding free list locks");
1836 bitMapLock()->unlock();
1837 releaseFreelistLocks();
1838 {
1839 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1840 if (_foregroundGCShouldWait) {
1841 // We are going to be waiting for action for the CMS thread;
1842 // it had better not be gone (for instance at shutdown)!
1843 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1844 "CMS thread must be running");
1845 // Wait here until the background collector gives us the go-ahead
1846 ConcurrentMarkSweepThread::clear_CMS_flag(
1847 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1848 // Get a possibly blocked CMS thread going:
1849 // Note that we set _foregroundGCIsActive true above,
1850 // without protection of the CGC_lock.
1851 CGC_lock->notify();
1852 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1853 "Possible deadlock");
1854 while (_foregroundGCShouldWait) {
1855 // wait for notification
1856 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1857 // Possibility of delay/starvation here, since CMS token does
1858 // not know to give priority to VM thread? Actually, i think
1859 // there wouldn't be any delay/starvation, but the proof of
1860 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1861 }
1862 ConcurrentMarkSweepThread::set_CMS_flag(
1863 ConcurrentMarkSweepThread::CMS_vm_has_token);
1864 }
1865 }
1866 // The CMS_token is already held. Get back the other locks.
1867 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1868 "VM thread should have CMS token");
1869 getFreelistLocks();
1870 bitMapLock()->lock_without_safepoint_check();
1871 if (TraceCMSState) {
1872 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1873 INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1874 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1875 }
1876
1877 // Check if we need to do a compaction, or if not, whether
1878 // we need to start the mark-sweep from scratch.
1879 bool should_compact = false;
1880 bool should_start_over = false;
1881 decide_foreground_collection_type(clear_all_soft_refs,
1882 &should_compact, &should_start_over);
1883
1884 NOT_PRODUCT(
1885 if (RotateCMSCollectionTypes) {
1886 if (_cmsGen->debug_collection_type() ==
1887 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1888 should_compact = true;
1889 } else if (_cmsGen->debug_collection_type() ==
1890 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1891 should_compact = false;
1892 }
1893 }
1894 )
1895
1896 if (first_state > Idling) {
1897 report_concurrent_mode_interruption();
1898 }
1899
1900 set_did_compact(should_compact);
1901 if (should_compact) {
1902 // If the collection is being acquired from the background
1903 // collector, there may be references on the discovered
1904 // references lists that have NULL referents (being those
1905 // that were concurrently cleared by a mutator) or
1906 // that are no longer active (having been enqueued concurrently
1907 // by the mutator).
1908 // Scrub the list of those references because Mark-Sweep-Compact
1909 // code assumes referents are not NULL and that all discovered
1910 // Reference objects are active.
1911 ref_processor()->clean_up_discovered_references();
1912
1913 if (first_state > Idling) {
1914 save_heap_summary();
1915 }
1916
1917 do_compaction_work(clear_all_soft_refs);
1918
1919 // Has the GC time limit been exceeded?
1920 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration();
1921 size_t max_eden_size = young_gen->max_capacity() -
1922 young_gen->to()->capacity() -
1923 young_gen->from()->capacity();
1924 GenCollectedHeap* gch = GenCollectedHeap::heap();
1925 GCCause::Cause gc_cause = gch->gc_cause();
1926 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1927 young_gen->eden()->used(),
1928 _cmsGen->max_capacity(),
1929 max_eden_size,
1930 full,
1931 gc_cause,
1932 gch->collector_policy());
1933 } else {
1934 do_mark_sweep_work(clear_all_soft_refs, first_state,
1935 should_start_over);
1936 }
1937 // Reset the expansion cause, now that we just completed
1938 // a collection cycle.
1939 clear_expansion_cause();
1940 _foregroundGCIsActive = false;
1941 return;
1942 }
1943
1944 // Resize the tenured generation
1945 // after obtaining the free list locks for the
1946 // two generations.
1947 void CMSCollector::compute_new_size() {
1948 assert_locked_or_safepoint(Heap_lock);
1949 FreelistLocker z(this);
1950 MetaspaceGC::compute_new_size();
1951 _cmsGen->compute_new_size_free_list();
1952 }
1953
1954 // A work method used by foreground collection to determine
1955 // what type of collection (compacting or not, continuing or fresh)
1956 // it should do.
1957 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1958 // and CMSCompactWhenClearAllSoftRefs the default in the future
1959 // and do away with the flags after a suitable period.
1960 void CMSCollector::decide_foreground_collection_type(
1961 bool clear_all_soft_refs, bool* should_compact,
1962 bool* should_start_over) {
1963 // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1964 // flag is set, and we have either requested a System.gc() or
1965 // the number of full gc's since the last concurrent cycle
1966 // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1967 // or if an incremental collection has failed
1968 GenCollectedHeap* gch = GenCollectedHeap::heap();
1969 assert(gch->collector_policy()->is_generation_policy(),
1970 "You may want to check the correctness of the following");
1971 // Inform cms gen if this was due to partial collection failing.
1972 // The CMS gen may use this fact to determine its expansion policy.
1973 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1974 assert(!_cmsGen->incremental_collection_failed(),
1975 "Should have been noticed, reacted to and cleared");
1976 _cmsGen->set_incremental_collection_failed();
1977 }
1978 *should_compact =
1979 UseCMSCompactAtFullCollection &&
1980 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1981 GCCause::is_user_requested_gc(gch->gc_cause()) ||
1982 gch->incremental_collection_will_fail(true /* consult_young */));
1983 *should_start_over = false;
1984 if (clear_all_soft_refs && !*should_compact) {
1985 // We are about to do a last ditch collection attempt
1986 // so it would normally make sense to do a compaction
1987 // to reclaim as much space as possible.
1988 if (CMSCompactWhenClearAllSoftRefs) {
1989 // Default: The rationale is that in this case either
1990 // we are past the final marking phase, in which case
1991 // we'd have to start over, or so little has been done
1992 // that there's little point in saving that work. Compaction
1993 // appears to be the sensible choice in either case.
1994 *should_compact = true;
1995 } else {
1996 // We have been asked to clear all soft refs, but not to
1997 // compact. Make sure that we aren't past the final checkpoint
1998 // phase, for that is where we process soft refs. If we are already
1999 // past that phase, we'll need to redo the refs discovery phase and
2000 // if necessary clear soft refs that weren't previously
2001 // cleared. We do so by remembering the phase in which
2002 // we came in, and if we are past the refs processing
2003 // phase, we'll choose to just redo the mark-sweep
2004 // collection from scratch.
2005 if (_collectorState > FinalMarking) {
2006 // We are past the refs processing phase;
2007 // start over and do a fresh synchronous CMS cycle
2008 _collectorState = Resetting; // skip to reset to start new cycle
2009 reset(false /* == !asynch */);
2010 *should_start_over = true;
2011 } // else we can continue a possibly ongoing current cycle
2012 }
2013 }
2014 }
2015
2016 // A work method used by the foreground collector to do
2017 // a mark-sweep-compact.
2018 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
2019 GenCollectedHeap* gch = GenCollectedHeap::heap();
2020
2021 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
2022 gc_timer->register_gc_start();
2023
2024 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
2025 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
2026
2027 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id());
2028 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
2029 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
2030 "collections passed to foreground collector", _full_gcs_since_conc_gc);
2031 }
2032
2033 // Temporarily widen the span of the weak reference processing to
2034 // the entire heap.
2035 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
2036 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
2037 // Temporarily, clear the "is_alive_non_header" field of the
2038 // reference processor.
2039 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
2040 // Temporarily make reference _processing_ single threaded (non-MT).
2041 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
2042 // Temporarily make refs discovery atomic
2043 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
2044 // Temporarily make reference _discovery_ single threaded (non-MT)
2045 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
2046
2047 ref_processor()->set_enqueuing_is_done(false);
2048 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/);
2049 ref_processor()->setup_policy(clear_all_soft_refs);
2050 // If an asynchronous collection finishes, the _modUnionTable is
2051 // all clear. If we are assuming the collection from an asynchronous
2052 // collection, clear the _modUnionTable.
2053 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
2054 "_modUnionTable should be clear if the baton was not passed");
2055 _modUnionTable.clear_all();
2056 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
2057 "mod union for klasses should be clear if the baton was passed");
2058 _ct->klass_rem_set()->clear_mod_union();
2059
2060 // We must adjust the allocation statistics being maintained
2061 // in the free list space. We do so by reading and clearing
2062 // the sweep timer and updating the block flux rate estimates below.
2063 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
2064 if (_inter_sweep_timer.is_active()) {
2065 _inter_sweep_timer.stop();
2066 // Note that we do not use this sample to update the _inter_sweep_estimate.
2067 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
2068 _inter_sweep_estimate.padded_average(),
2069 _intra_sweep_estimate.padded_average());
2070 }
2071
2072 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
2073 ref_processor(), clear_all_soft_refs);
2074 #ifdef ASSERT
2075 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
2076 size_t free_size = cms_space->free();
2077 assert(free_size ==
2078 pointer_delta(cms_space->end(), cms_space->compaction_top())
2079 * HeapWordSize,
2080 "All the free space should be compacted into one chunk at top");
2081 assert(cms_space->dictionary()->total_chunk_size(
2082 debug_only(cms_space->freelistLock())) == 0 ||
2083 cms_space->totalSizeInIndexedFreeLists() == 0,
2084 "All the free space should be in a single chunk");
2085 size_t num = cms_space->totalCount();
2086 assert((free_size == 0 && num == 0) ||
2087 (free_size > 0 && (num == 1 || num == 2)),
2088 "There should be at most 2 free chunks after compaction");
2089 #endif // ASSERT
2090 _collectorState = Resetting;
2091 assert(_restart_addr == NULL,
2092 "Should have been NULL'd before baton was passed");
2093 reset(false /* == !asynch */);
2094 _cmsGen->reset_after_compaction();
2095 _concurrent_cycles_since_last_unload = 0;
2096
2097 // Clear any data recorded in the PLAB chunk arrays.
2098 if (_survivor_plab_array != NULL) {
2099 reset_survivor_plab_arrays();
2100 }
2101
2102 // Adjust the per-size allocation stats for the next epoch.
2103 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
2104 // Restart the "inter sweep timer" for the next epoch.
2105 _inter_sweep_timer.reset();
2106 _inter_sweep_timer.start();
2107
2108 gc_timer->register_gc_end();
2109
2110 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
2111
2112 // For a mark-sweep-compact, compute_new_size() will be called
2113 // in the heap's do_collection() method.
2114 }
2115
2116 // A work method used by the foreground collector to do
2117 // a mark-sweep, after taking over from a possibly on-going
2118 // concurrent mark-sweep collection.
2119 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2120 CollectorState first_state, bool should_start_over) {
2121 if (PrintGC && Verbose) {
2122 gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2123 "collector with count %d",
2124 _full_gcs_since_conc_gc);
2125 }
2126 switch (_collectorState) {
2127 case Idling:
2128 if (first_state == Idling || should_start_over) {
2129 // The background GC was not active, or should
2130 // restarted from scratch; start the cycle.
2131 _collectorState = InitialMarking;
2132 }
2133 // If first_state was not Idling, then a background GC
2134 // was in progress and has now finished. No need to do it
2135 // again. Leave the state as Idling.
2136 break;
2137 case Precleaning:
2138 // In the foreground case don't do the precleaning since
2139 // it is not done concurrently and there is extra work
2140 // required.
2141 _collectorState = FinalMarking;
2142 }
2143 collect_in_foreground(clear_all_soft_refs, GenCollectedHeap::heap()->gc_cause());
2144
2145 // For a mark-sweep, compute_new_size() will be called
2146 // in the heap's do_collection() method.
2147 }
2148
2149
2150 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
2151 DefNewGeneration* dng = _young_gen->as_DefNewGeneration();
2152 EdenSpace* eden_space = dng->eden();
2153 ContiguousSpace* from_space = dng->from();
2154 ContiguousSpace* to_space = dng->to();
2155 // Eden
2156 if (_eden_chunk_array != NULL) {
2157 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
2158 eden_space->bottom(), eden_space->top(),
2159 eden_space->end(), eden_space->capacity());
2160 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", "
2161 "_eden_chunk_capacity=" SIZE_FORMAT,
2162 _eden_chunk_index, _eden_chunk_capacity);
2163 for (size_t i = 0; i < _eden_chunk_index; i++) {
2164 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
2165 i, _eden_chunk_array[i]);
2166 }
2167 }
2168 // Survivor
2169 if (_survivor_chunk_array != NULL) {
2170 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
2171 from_space->bottom(), from_space->top(),
2172 from_space->end(), from_space->capacity());
2173 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", "
2174 "_survivor_chunk_capacity=" SIZE_FORMAT,
2175 _survivor_chunk_index, _survivor_chunk_capacity);
2176 for (size_t i = 0; i < _survivor_chunk_index; i++) {
2177 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
2178 i, _survivor_chunk_array[i]);
2179 }
2180 }
2181 }
2182
2183 void CMSCollector::getFreelistLocks() const {
2184 // Get locks for all free lists in all generations that this
2185 // collector is responsible for
2186 _cmsGen->freelistLock()->lock_without_safepoint_check();
2187 }
2188
2189 void CMSCollector::releaseFreelistLocks() const {
2190 // Release locks for all free lists in all generations that this
2191 // collector is responsible for
2192 _cmsGen->freelistLock()->unlock();
2193 }
2194
2195 bool CMSCollector::haveFreelistLocks() const {
2196 // Check locks for all free lists in all generations that this
2197 // collector is responsible for
2198 assert_lock_strong(_cmsGen->freelistLock());
2199 PRODUCT_ONLY(ShouldNotReachHere());
2200 return true;
2201 }
2202
2203 // A utility class that is used by the CMS collector to
2204 // temporarily "release" the foreground collector from its
2205 // usual obligation to wait for the background collector to
2206 // complete an ongoing phase before proceeding.
2207 class ReleaseForegroundGC: public StackObj {
2208 private:
2209 CMSCollector* _c;
2210 public:
2211 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2212 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2213 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2214 // allow a potentially blocked foreground collector to proceed
2215 _c->_foregroundGCShouldWait = false;
2216 if (_c->_foregroundGCIsActive) {
2217 CGC_lock->notify();
2218 }
2219 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2220 "Possible deadlock");
2221 }
2222
2223 ~ReleaseForegroundGC() {
2224 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2225 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2226 _c->_foregroundGCShouldWait = true;
2227 }
2228 };
2229
2230 // There are separate collect_in_background and collect_in_foreground because of
2231 // the different locking requirements of the background collector and the
2232 // foreground collector. There was originally an attempt to share
2233 // one "collect" method between the background collector and the foreground
2234 // collector but the if-then-else required made it cleaner to have
2235 // separate methods.
2236 void CMSCollector::collect_in_background(bool clear_all_soft_refs, GCCause::Cause cause) {
2237 assert(Thread::current()->is_ConcurrentGC_thread(),
2238 "A CMS asynchronous collection is only allowed on a CMS thread.");
2239
2240 GenCollectedHeap* gch = GenCollectedHeap::heap();
2241 {
2242 bool safepoint_check = Mutex::_no_safepoint_check_flag;
2243 MutexLockerEx hl(Heap_lock, safepoint_check);
2244 FreelistLocker fll(this);
2245 MutexLockerEx x(CGC_lock, safepoint_check);
2246 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2247 // The foreground collector is active or we're
2248 // not using asynchronous collections. Skip this
2249 // background collection.
2250 assert(!_foregroundGCShouldWait, "Should be clear");
2251 return;
2252 } else {
2253 assert(_collectorState == Idling, "Should be idling before start.");
2254 _collectorState = InitialMarking;
2255 register_gc_start(cause);
2256 // Reset the expansion cause, now that we are about to begin
2257 // a new cycle.
2258 clear_expansion_cause();
2259
2260 // Clear the MetaspaceGC flag since a concurrent collection
2261 // is starting but also clear it after the collection.
2262 MetaspaceGC::set_should_concurrent_collect(false);
2263 }
2264 // Decide if we want to enable class unloading as part of the
2265 // ensuing concurrent GC cycle.
2266 update_should_unload_classes();
2267 _full_gc_requested = false; // acks all outstanding full gc requests
2268 _full_gc_cause = GCCause::_no_gc;
2269 // Signal that we are about to start a collection
2270 gch->increment_total_full_collections(); // ... starting a collection cycle
2271 _collection_count_start = gch->total_full_collections();
2272 }
2273
2274 // Used for PrintGC
2275 size_t prev_used;
2276 if (PrintGC && Verbose) {
2277 prev_used = _cmsGen->used(); // XXXPERM
2278 }
2279
2280 // The change of the collection state is normally done at this level;
2281 // the exceptions are phases that are executed while the world is
2282 // stopped. For those phases the change of state is done while the
2283 // world is stopped. For baton passing purposes this allows the
2284 // background collector to finish the phase and change state atomically.
2285 // The foreground collector cannot wait on a phase that is done
2286 // while the world is stopped because the foreground collector already
2287 // has the world stopped and would deadlock.
2288 while (_collectorState != Idling) {
2289 if (TraceCMSState) {
2290 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2291 Thread::current(), _collectorState);
2292 }
2293 // The foreground collector
2294 // holds the Heap_lock throughout its collection.
2295 // holds the CMS token (but not the lock)
2296 // except while it is waiting for the background collector to yield.
2297 //
2298 // The foreground collector should be blocked (not for long)
2299 // if the background collector is about to start a phase
2300 // executed with world stopped. If the background
2301 // collector has already started such a phase, the
2302 // foreground collector is blocked waiting for the
2303 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
2304 // are executed in the VM thread.
2305 //
2306 // The locking order is
2307 // PendingListLock (PLL) -- if applicable (FinalMarking)
2308 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
2309 // CMS token (claimed in
2310 // stop_world_and_do() -->
2311 // safepoint_synchronize() -->
2312 // CMSThread::synchronize())
2313
2314 {
2315 // Check if the FG collector wants us to yield.
2316 CMSTokenSync x(true); // is cms thread
2317 if (waitForForegroundGC()) {
2318 // We yielded to a foreground GC, nothing more to be
2319 // done this round.
2320 assert(_foregroundGCShouldWait == false, "We set it to false in "
2321 "waitForForegroundGC()");
2322 if (TraceCMSState) {
2323 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2324 " exiting collection CMS state %d",
2325 Thread::current(), _collectorState);
2326 }
2327 return;
2328 } else {
2329 // The background collector can run but check to see if the
2330 // foreground collector has done a collection while the
2331 // background collector was waiting to get the CGC_lock
2332 // above. If yes, break so that _foregroundGCShouldWait
2333 // is cleared before returning.
2334 if (_collectorState == Idling) {
2335 break;
2336 }
2337 }
2338 }
2339
2340 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2341 "should be waiting");
2342
2343 switch (_collectorState) {
2344 case InitialMarking:
2345 {
2346 ReleaseForegroundGC x(this);
2347 stats().record_cms_begin();
2348 VM_CMS_Initial_Mark initial_mark_op(this);
2349 VMThread::execute(&initial_mark_op);
2350 }
2351 // The collector state may be any legal state at this point
2352 // since the background collector may have yielded to the
2353 // foreground collector.
2354 break;
2355 case Marking:
2356 // initial marking in checkpointRootsInitialWork has been completed
2357 if (markFromRoots(true)) { // we were successful
2358 assert(_collectorState == Precleaning, "Collector state should "
2359 "have changed");
2360 } else {
2361 assert(_foregroundGCIsActive, "Internal state inconsistency");
2362 }
2363 break;
2364 case Precleaning:
2365 // marking from roots in markFromRoots has been completed
2366 preclean();
2367 assert(_collectorState == AbortablePreclean ||
2368 _collectorState == FinalMarking,
2369 "Collector state should have changed");
2370 break;
2371 case AbortablePreclean:
2372 abortable_preclean();
2373 assert(_collectorState == FinalMarking, "Collector state should "
2374 "have changed");
2375 break;
2376 case FinalMarking:
2377 {
2378 ReleaseForegroundGC x(this);
2379
2380 VM_CMS_Final_Remark final_remark_op(this);
2381 VMThread::execute(&final_remark_op);
2382 }
2383 assert(_foregroundGCShouldWait, "block post-condition");
2384 break;
2385 case Sweeping:
2386 // final marking in checkpointRootsFinal has been completed
2387 sweep(true);
2388 assert(_collectorState == Resizing, "Collector state change "
2389 "to Resizing must be done under the free_list_lock");
2390 _full_gcs_since_conc_gc = 0;
2391
2392 case Resizing: {
2393 // Sweeping has been completed...
2394 // At this point the background collection has completed.
2395 // Don't move the call to compute_new_size() down
2396 // into code that might be executed if the background
2397 // collection was preempted.
2398 {
2399 ReleaseForegroundGC x(this); // unblock FG collection
2400 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
2401 CMSTokenSync z(true); // not strictly needed.
2402 if (_collectorState == Resizing) {
2403 compute_new_size();
2404 save_heap_summary();
2405 _collectorState = Resetting;
2406 } else {
2407 assert(_collectorState == Idling, "The state should only change"
2408 " because the foreground collector has finished the collection");
2409 }
2410 }
2411 break;
2412 }
2413 case Resetting:
2414 // CMS heap resizing has been completed
2415 reset(true);
2416 assert(_collectorState == Idling, "Collector state should "
2417 "have changed");
2418
2419 MetaspaceGC::set_should_concurrent_collect(false);
2420
2421 stats().record_cms_end();
2422 // Don't move the concurrent_phases_end() and compute_new_size()
2423 // calls to here because a preempted background collection
2424 // has it's state set to "Resetting".
2425 break;
2426 case Idling:
2427 default:
2428 ShouldNotReachHere();
2429 break;
2430 }
2431 if (TraceCMSState) {
2432 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2433 Thread::current(), _collectorState);
2434 }
2435 assert(_foregroundGCShouldWait, "block post-condition");
2436 }
2437
2438 // Should this be in gc_epilogue?
2439 collector_policy()->counters()->update_counters();
2440
2441 {
2442 // Clear _foregroundGCShouldWait and, in the event that the
2443 // foreground collector is waiting, notify it, before
2444 // returning.
2445 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2446 _foregroundGCShouldWait = false;
2447 if (_foregroundGCIsActive) {
2448 CGC_lock->notify();
2449 }
2450 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2451 "Possible deadlock");
2452 }
2453 if (TraceCMSState) {
2454 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2455 " exiting collection CMS state %d",
2456 Thread::current(), _collectorState);
2457 }
2458 if (PrintGC && Verbose) {
2459 _cmsGen->print_heap_change(prev_used);
2460 }
2461 }
2462
2463 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) {
2464 if (!_cms_start_registered) {
2465 register_gc_start(cause);
2466 }
2467 }
2468
2469 void CMSCollector::register_gc_start(GCCause::Cause cause) {
2470 _cms_start_registered = true;
2471 _gc_timer_cm->register_gc_start();
2472 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
2473 }
2474
2475 void CMSCollector::register_gc_end() {
2476 if (_cms_start_registered) {
2477 report_heap_summary(GCWhen::AfterGC);
2478
2479 _gc_timer_cm->register_gc_end();
2480 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
2481 _cms_start_registered = false;
2482 }
2483 }
2484
2485 void CMSCollector::save_heap_summary() {
2486 GenCollectedHeap* gch = GenCollectedHeap::heap();
2487 _last_heap_summary = gch->create_heap_summary();
2488 _last_metaspace_summary = gch->create_metaspace_summary();
2489 }
2490
2491 void CMSCollector::report_heap_summary(GCWhen::Type when) {
2492 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
2493 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
2494 }
2495
2496 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) {
2497 assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2498 "Foreground collector should be waiting, not executing");
2499 assert(Thread::current()->is_VM_thread(), "A foreground collection"
2500 "may only be done by the VM Thread with the world stopped");
2501 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2502 "VM thread should have CMS token");
2503
2504 // The gc id is created in register_foreground_gc_start if this collection is synchronous
2505 const GCId gc_id = _collectorState == InitialMarking ? GCId::peek() : _gc_tracer_cm->gc_id();
2506 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2507 true, NULL, gc_id);)
2508 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2509
2510 HandleMark hm; // Discard invalid handles created during verification
2511
2512 if (VerifyBeforeGC &&
2513 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2514 Universe::verify();
2515 }
2516
2517 // Snapshot the soft reference policy to be used in this collection cycle.
2518 ref_processor()->setup_policy(clear_all_soft_refs);
2519
2520 // Decide if class unloading should be done
2521 update_should_unload_classes();
2522
2523 bool init_mark_was_synchronous = false; // until proven otherwise
2524 while (_collectorState != Idling) {
2525 if (TraceCMSState) {
2526 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2527 Thread::current(), _collectorState);
2528 }
2529 switch (_collectorState) {
2530 case InitialMarking:
2531 register_foreground_gc_start(cause);
2532 init_mark_was_synchronous = true; // fact to be exploited in re-mark
2533 checkpointRootsInitial(false);
2534 assert(_collectorState == Marking, "Collector state should have changed"
2535 " within checkpointRootsInitial()");
2536 break;
2537 case Marking:
2538 // initial marking in checkpointRootsInitialWork has been completed
2539 if (VerifyDuringGC &&
2540 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2541 Universe::verify("Verify before initial mark: ");
2542 }
2543 {
2544 bool res = markFromRoots(false);
2545 assert(res && _collectorState == FinalMarking, "Collector state should "
2546 "have changed");
2547 break;
2548 }
2549 case FinalMarking:
2550 if (VerifyDuringGC &&
2551 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2552 Universe::verify("Verify before re-mark: ");
2553 }
2554 checkpointRootsFinal(false, clear_all_soft_refs,
2555 init_mark_was_synchronous);
2556 assert(_collectorState == Sweeping, "Collector state should not "
2557 "have changed within checkpointRootsFinal()");
2558 break;
2559 case Sweeping:
2560 // final marking in checkpointRootsFinal has been completed
2561 if (VerifyDuringGC &&
2562 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2563 Universe::verify("Verify before sweep: ");
2564 }
2565 sweep(false);
2566 assert(_collectorState == Resizing, "Incorrect state");
2567 break;
2568 case Resizing: {
2569 // Sweeping has been completed; the actual resize in this case
2570 // is done separately; nothing to be done in this state.
2571 _collectorState = Resetting;
2572 break;
2573 }
2574 case Resetting:
2575 // The heap has been resized.
2576 if (VerifyDuringGC &&
2577 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2578 Universe::verify("Verify before reset: ");
2579 }
2580 save_heap_summary();
2581 reset(false);
2582 assert(_collectorState == Idling, "Collector state should "
2583 "have changed");
2584 break;
2585 case Precleaning:
2586 case AbortablePreclean:
2587 // Elide the preclean phase
2588 _collectorState = FinalMarking;
2589 break;
2590 default:
2591 ShouldNotReachHere();
2592 }
2593 if (TraceCMSState) {
2594 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2595 Thread::current(), _collectorState);
2596 }
2597 }
2598
2599 if (VerifyAfterGC &&
2600 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2601 Universe::verify();
2602 }
2603 if (TraceCMSState) {
2604 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2605 " exiting collection CMS state %d",
2606 Thread::current(), _collectorState);
2607 }
2608 }
2609
2610 bool CMSCollector::waitForForegroundGC() {
2611 bool res = false;
2612 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2613 "CMS thread should have CMS token");
2614 // Block the foreground collector until the
2615 // background collectors decides whether to
2616 // yield.
2617 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2618 _foregroundGCShouldWait = true;
2619 if (_foregroundGCIsActive) {
2620 // The background collector yields to the
2621 // foreground collector and returns a value
2622 // indicating that it has yielded. The foreground
2623 // collector can proceed.
2624 res = true;
2625 _foregroundGCShouldWait = false;
2626 ConcurrentMarkSweepThread::clear_CMS_flag(
2627 ConcurrentMarkSweepThread::CMS_cms_has_token);
2628 ConcurrentMarkSweepThread::set_CMS_flag(
2629 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2630 // Get a possibly blocked foreground thread going
2631 CGC_lock->notify();
2632 if (TraceCMSState) {
2633 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2634 Thread::current(), _collectorState);
2635 }
2636 while (_foregroundGCIsActive) {
2637 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2638 }
2639 ConcurrentMarkSweepThread::set_CMS_flag(
2640 ConcurrentMarkSweepThread::CMS_cms_has_token);
2641 ConcurrentMarkSweepThread::clear_CMS_flag(
2642 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2643 }
2644 if (TraceCMSState) {
2645 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2646 Thread::current(), _collectorState);
2647 }
2648 return res;
2649 }
2650
2651 // Because of the need to lock the free lists and other structures in
2652 // the collector, common to all the generations that the collector is
2653 // collecting, we need the gc_prologues of individual CMS generations
2654 // delegate to their collector. It may have been simpler had the
2655 // current infrastructure allowed one to call a prologue on a
2656 // collector. In the absence of that we have the generation's
2657 // prologue delegate to the collector, which delegates back
2658 // some "local" work to a worker method in the individual generations
2659 // that it's responsible for collecting, while itself doing any
2660 // work common to all generations it's responsible for. A similar
2661 // comment applies to the gc_epilogue()'s.
2662 // The role of the variable _between_prologue_and_epilogue is to
2663 // enforce the invocation protocol.
2664 void CMSCollector::gc_prologue(bool full) {
2665 // Call gc_prologue_work() for the CMSGen
2666 // we are responsible for.
2667
2668 // The following locking discipline assumes that we are only called
2669 // when the world is stopped.
2670 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2671
2672 // The CMSCollector prologue must call the gc_prologues for the
2673 // "generations" that it's responsible
2674 // for.
2675
2676 assert( Thread::current()->is_VM_thread()
2677 || ( CMSScavengeBeforeRemark
2678 && Thread::current()->is_ConcurrentGC_thread()),
2679 "Incorrect thread type for prologue execution");
2680
2681 if (_between_prologue_and_epilogue) {
2682 // We have already been invoked; this is a gc_prologue delegation
2683 // from yet another CMS generation that we are responsible for, just
2684 // ignore it since all relevant work has already been done.
2685 return;
2686 }
2687
2688 // set a bit saying prologue has been called; cleared in epilogue
2689 _between_prologue_and_epilogue = true;
2690 // Claim locks for common data structures, then call gc_prologue_work()
2691 // for each CMSGen.
2692
2693 getFreelistLocks(); // gets free list locks on constituent spaces
2694 bitMapLock()->lock_without_safepoint_check();
2695
2696 // Should call gc_prologue_work() for all cms gens we are responsible for
2697 bool duringMarking = _collectorState >= Marking
2698 && _collectorState < Sweeping;
2699
2700 // The young collections clear the modified oops state, which tells if
2701 // there are any modified oops in the class. The remark phase also needs
2702 // that information. Tell the young collection to save the union of all
2703 // modified klasses.
2704 if (duringMarking) {
2705 _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2706 }
2707
2708 bool registerClosure = duringMarking;
2709
2710 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ?
2711 &_modUnionClosurePar
2712 : &_modUnionClosure;
2713 _cmsGen->gc_prologue_work(full, registerClosure, muc);
2714
2715 if (!full) {
2716 stats().record_gc0_begin();
2717 }
2718 }
2719
2720 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2721
2722 _capacity_at_prologue = capacity();
2723 _used_at_prologue = used();
2724
2725 // Delegate to CMScollector which knows how to coordinate between
2726 // this and any other CMS generations that it is responsible for
2727 // collecting.
2728 collector()->gc_prologue(full);
2729 }
2730
2731 // This is a "private" interface for use by this generation's CMSCollector.
2732 // Not to be called directly by any other entity (for instance,
2733 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2734 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2735 bool registerClosure, ModUnionClosure* modUnionClosure) {
2736 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2737 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2738 "Should be NULL");
2739 if (registerClosure) {
2740 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2741 }
2742 cmsSpace()->gc_prologue();
2743 // Clear stat counters
2744 NOT_PRODUCT(
2745 assert(_numObjectsPromoted == 0, "check");
2746 assert(_numWordsPromoted == 0, "check");
2747 if (Verbose && PrintGC) {
2748 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2749 SIZE_FORMAT" bytes concurrently",
2750 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2751 }
2752 _numObjectsAllocated = 0;
2753 _numWordsAllocated = 0;
2754 )
2755 }
2756
2757 void CMSCollector::gc_epilogue(bool full) {
2758 // The following locking discipline assumes that we are only called
2759 // when the world is stopped.
2760 assert(SafepointSynchronize::is_at_safepoint(),
2761 "world is stopped assumption");
2762
2763 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2764 // if linear allocation blocks need to be appropriately marked to allow the
2765 // the blocks to be parsable. We also check here whether we need to nudge the
2766 // CMS collector thread to start a new cycle (if it's not already active).
2767 assert( Thread::current()->is_VM_thread()
2768 || ( CMSScavengeBeforeRemark
2769 && Thread::current()->is_ConcurrentGC_thread()),
2770 "Incorrect thread type for epilogue execution");
2771
2772 if (!_between_prologue_and_epilogue) {
2773 // We have already been invoked; this is a gc_epilogue delegation
2774 // from yet another CMS generation that we are responsible for, just
2775 // ignore it since all relevant work has already been done.
2776 return;
2777 }
2778 assert(haveFreelistLocks(), "must have freelist locks");
2779 assert_lock_strong(bitMapLock());
2780
2781 _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2782
2783 _cmsGen->gc_epilogue_work(full);
2784
2785 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2786 // in case sampling was not already enabled, enable it
2787 _start_sampling = true;
2788 }
2789 // reset _eden_chunk_array so sampling starts afresh
2790 _eden_chunk_index = 0;
2791
2792 size_t cms_used = _cmsGen->cmsSpace()->used();
2793
2794 // update performance counters - this uses a special version of
2795 // update_counters() that allows the utilization to be passed as a
2796 // parameter, avoiding multiple calls to used().
2797 //
2798 _cmsGen->update_counters(cms_used);
2799
2800 if (CMSIncrementalMode) {
2801 icms_update_allocation_limits();
2802 }
2803
2804 bitMapLock()->unlock();
2805 releaseFreelistLocks();
2806
2807 if (!CleanChunkPoolAsync) {
2808 Chunk::clean_chunk_pool();
2809 }
2810
2811 set_did_compact(false);
2812 _between_prologue_and_epilogue = false; // ready for next cycle
2813 }
2814
2815 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2816 collector()->gc_epilogue(full);
2817
2818 // Also reset promotion tracking in par gc thread states.
2819 if (CollectedHeap::use_parallel_gc_threads()) {
2820 for (uint i = 0; i < ParallelGCThreads; i++) {
2821 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2822 }
2823 }
2824 }
2825
2826 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2827 assert(!incremental_collection_failed(), "Should have been cleared");
2828 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2829 cmsSpace()->gc_epilogue();
2830 // Print stat counters
2831 NOT_PRODUCT(
2832 assert(_numObjectsAllocated == 0, "check");
2833 assert(_numWordsAllocated == 0, "check");
2834 if (Verbose && PrintGC) {
2835 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2836 SIZE_FORMAT" bytes",
2837 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2838 }
2839 _numObjectsPromoted = 0;
2840 _numWordsPromoted = 0;
2841 )
2842
2843 if (PrintGC && Verbose) {
2844 // Call down the chain in contiguous_available needs the freelistLock
2845 // so print this out before releasing the freeListLock.
2846 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2847 contiguous_available());
2848 }
2849 }
2850
2851 #ifndef PRODUCT
2852 bool CMSCollector::have_cms_token() {
2853 Thread* thr = Thread::current();
2854 if (thr->is_VM_thread()) {
2855 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2856 } else if (thr->is_ConcurrentGC_thread()) {
2857 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2858 } else if (thr->is_GC_task_thread()) {
2859 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2860 ParGCRareEvent_lock->owned_by_self();
2861 }
2862 return false;
2863 }
2864 #endif
2865
2866 // Check reachability of the given heap address in CMS generation,
2867 // treating all other generations as roots.
2868 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2869 // We could "guarantee" below, rather than assert, but I'll
2870 // leave these as "asserts" so that an adventurous debugger
2871 // could try this in the product build provided some subset of
2872 // the conditions were met, provided they were interested in the
2873 // results and knew that the computation below wouldn't interfere
2874 // with other concurrent computations mutating the structures
2875 // being read or written.
2876 assert(SafepointSynchronize::is_at_safepoint(),
2877 "Else mutations in object graph will make answer suspect");
2878 assert(have_cms_token(), "Should hold cms token");
2879 assert(haveFreelistLocks(), "must hold free list locks");
2880 assert_lock_strong(bitMapLock());
2881
2882 // Clear the marking bit map array before starting, but, just
2883 // for kicks, first report if the given address is already marked
2884 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", addr,
2885 _markBitMap.isMarked(addr) ? "" : " not");
2886
2887 if (verify_after_remark()) {
2888 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2889 bool result = verification_mark_bm()->isMarked(addr);
2890 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", addr,
2891 result ? "IS" : "is NOT");
2892 return result;
2893 } else {
2894 gclog_or_tty->print_cr("Could not compute result");
2895 return false;
2896 }
2897 }
2898
2899
2900 void
2901 CMSCollector::print_on_error(outputStream* st) {
2902 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2903 if (collector != NULL) {
2904 CMSBitMap* bitmap = &collector->_markBitMap;
2905 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap);
2906 bitmap->print_on_error(st, " Bits: ");
2907
2908 st->cr();
2909
2910 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2911 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap);
2912 mut_bitmap->print_on_error(st, " Bits: ");
2913 }
2914 }
2915
2916 ////////////////////////////////////////////////////////
2917 // CMS Verification Support
2918 ////////////////////////////////////////////////////////
2919 // Following the remark phase, the following invariant
2920 // should hold -- each object in the CMS heap which is
2921 // marked in markBitMap() should be marked in the verification_mark_bm().
2922
2923 class VerifyMarkedClosure: public BitMapClosure {
2924 CMSBitMap* _marks;
2925 bool _failed;
2926
2927 public:
2928 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2929
2930 bool do_bit(size_t offset) {
2931 HeapWord* addr = _marks->offsetToHeapWord(offset);
2932 if (!_marks->isMarked(addr)) {
2933 oop(addr)->print_on(gclog_or_tty);
2934 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2935 _failed = true;
2936 }
2937 return true;
2938 }
2939
2940 bool failed() { return _failed; }
2941 };
2942
2943 bool CMSCollector::verify_after_remark(bool silent) {
2944 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... ");
2945 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2946 static bool init = false;
2947
2948 assert(SafepointSynchronize::is_at_safepoint(),
2949 "Else mutations in object graph will make answer suspect");
2950 assert(have_cms_token(),
2951 "Else there may be mutual interference in use of "
2952 " verification data structures");
2953 assert(_collectorState > Marking && _collectorState <= Sweeping,
2954 "Else marking info checked here may be obsolete");
2955 assert(haveFreelistLocks(), "must hold free list locks");
2956 assert_lock_strong(bitMapLock());
2957
2958
2959 // Allocate marking bit map if not already allocated
2960 if (!init) { // first time
2961 if (!verification_mark_bm()->allocate(_span)) {
2962 return false;
2963 }
2964 init = true;
2965 }
2966
2967 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2968
2969 // Turn off refs discovery -- so we will be tracing through refs.
2970 // This is as intended, because by this time
2971 // GC must already have cleared any refs that need to be cleared,
2972 // and traced those that need to be marked; moreover,
2973 // the marking done here is not going to interfere in any
2974 // way with the marking information used by GC.
2975 NoRefDiscovery no_discovery(ref_processor());
2976
2977 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2978
2979 // Clear any marks from a previous round
2980 verification_mark_bm()->clear_all();
2981 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2982 verify_work_stacks_empty();
2983
2984 GenCollectedHeap* gch = GenCollectedHeap::heap();
2985 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2986 // Update the saved marks which may affect the root scans.
2987 gch->save_marks();
2988
2989 if (CMSRemarkVerifyVariant == 1) {
2990 // In this first variant of verification, we complete
2991 // all marking, then check if the new marks-vector is
2992 // a subset of the CMS marks-vector.
2993 verify_after_remark_work_1();
2994 } else if (CMSRemarkVerifyVariant == 2) {
2995 // In this second variant of verification, we flag an error
2996 // (i.e. an object reachable in the new marks-vector not reachable
2997 // in the CMS marks-vector) immediately, also indicating the
2998 // identify of an object (A) that references the unmarked object (B) --
2999 // presumably, a mutation to A failed to be picked up by preclean/remark?
3000 verify_after_remark_work_2();
3001 } else {
3002 warning("Unrecognized value %d for CMSRemarkVerifyVariant",
3003 CMSRemarkVerifyVariant);
3004 }
3005 if (!silent) gclog_or_tty->print(" done] ");
3006 return true;
3007 }
3008
3009 void CMSCollector::verify_after_remark_work_1() {
3010 ResourceMark rm;
3011 HandleMark hm;
3012 GenCollectedHeap* gch = GenCollectedHeap::heap();
3013
3014 // Get a clear set of claim bits for the roots processing to work with.
3015 ClassLoaderDataGraph::clear_claimed_marks();
3016
3017 // Mark from roots one level into CMS
3018 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
3019 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3020
3021 gch->gen_process_roots(_cmsGen->level(),
3022 true, // younger gens are roots
3023 true, // activate StrongRootsScope
3024 SharedHeap::ScanningOption(roots_scanning_options()),
3025 should_unload_classes(),
3026 ¬Older,
3027 NULL,
3028 NULL); // SSS: Provide correct closure
3029
3030 // Now mark from the roots
3031 MarkFromRootsClosure markFromRootsClosure(this, _span,
3032 verification_mark_bm(), verification_mark_stack(),
3033 false /* don't yield */, true /* verifying */);
3034 assert(_restart_addr == NULL, "Expected pre-condition");
3035 verification_mark_bm()->iterate(&markFromRootsClosure);
3036 while (_restart_addr != NULL) {
3037 // Deal with stack overflow: by restarting at the indicated
3038 // address.
3039 HeapWord* ra = _restart_addr;
3040 markFromRootsClosure.reset(ra);
3041 _restart_addr = NULL;
3042 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
3043 }
3044 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
3045 verify_work_stacks_empty();
3046
3047 // Marking completed -- now verify that each bit marked in
3048 // verification_mark_bm() is also marked in markBitMap(); flag all
3049 // errors by printing corresponding objects.
3050 VerifyMarkedClosure vcl(markBitMap());
3051 verification_mark_bm()->iterate(&vcl);
3052 if (vcl.failed()) {
3053 gclog_or_tty->print("Verification failed");
3054 Universe::heap()->print_on(gclog_or_tty);
3055 fatal("CMS: failed marking verification after remark");
3056 }
3057 }
3058
3059 class VerifyKlassOopsKlassClosure : public KlassClosure {
3060 class VerifyKlassOopsClosure : public OopClosure {
3061 CMSBitMap* _bitmap;
3062 public:
3063 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
3064 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
3065 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
3066 } _oop_closure;
3067 public:
3068 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
3069 void do_klass(Klass* k) {
3070 k->oops_do(&_oop_closure);
3071 }
3072 };
3073
3074 void CMSCollector::verify_after_remark_work_2() {
3075 ResourceMark rm;
3076 HandleMark hm;
3077 GenCollectedHeap* gch = GenCollectedHeap::heap();
3078
3079 // Get a clear set of claim bits for the roots processing to work with.
3080 ClassLoaderDataGraph::clear_claimed_marks();
3081
3082 // Mark from roots one level into CMS
3083 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
3084 markBitMap());
3085 CLDToOopClosure cld_closure(¬Older, true);
3086
3087 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3088
3089 gch->gen_process_roots(_cmsGen->level(),
3090 true, // younger gens are roots
3091 true, // activate StrongRootsScope
3092 SharedHeap::ScanningOption(roots_scanning_options()),
3093 should_unload_classes(),
3094 ¬Older,
3095 NULL,
3096 &cld_closure);
3097
3098 // Now mark from the roots
3099 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
3100 verification_mark_bm(), markBitMap(), verification_mark_stack());
3101 assert(_restart_addr == NULL, "Expected pre-condition");
3102 verification_mark_bm()->iterate(&markFromRootsClosure);
3103 while (_restart_addr != NULL) {
3104 // Deal with stack overflow: by restarting at the indicated
3105 // address.
3106 HeapWord* ra = _restart_addr;
3107 markFromRootsClosure.reset(ra);
3108 _restart_addr = NULL;
3109 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
3110 }
3111 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
3112 verify_work_stacks_empty();
3113
3114 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
3115 ClassLoaderDataGraph::classes_do(&verify_klass_oops);
3116
3117 // Marking completed -- now verify that each bit marked in
3118 // verification_mark_bm() is also marked in markBitMap(); flag all
3119 // errors by printing corresponding objects.
3120 VerifyMarkedClosure vcl(markBitMap());
3121 verification_mark_bm()->iterate(&vcl);
3122 assert(!vcl.failed(), "Else verification above should not have succeeded");
3123 }
3124
3125 void ConcurrentMarkSweepGeneration::save_marks() {
3126 // delegate to CMS space
3127 cmsSpace()->save_marks();
3128 for (uint i = 0; i < ParallelGCThreads; i++) {
3129 _par_gc_thread_states[i]->promo.startTrackingPromotions();
3130 }
3131 }
3132
3133 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
3134 return cmsSpace()->no_allocs_since_save_marks();
3135 }
3136
3137 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
3138 \
3139 void ConcurrentMarkSweepGeneration:: \
3140 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
3141 cl->set_generation(this); \
3142 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
3143 cl->reset_generation(); \
3144 save_marks(); \
3145 }
3146
3147 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
3148
3149 void
3150 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
3151 cl->set_generation(this);
3152 younger_refs_in_space_iterate(_cmsSpace, cl);
3153 cl->reset_generation();
3154 }
3155
3156 void
3157 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
3158 if (freelistLock()->owned_by_self()) {
3159 Generation::oop_iterate(cl);
3160 } else {
3161 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3162 Generation::oop_iterate(cl);
3163 }
3164 }
3165
3166 void
3167 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
3168 if (freelistLock()->owned_by_self()) {
3169 Generation::object_iterate(cl);
3170 } else {
3171 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3172 Generation::object_iterate(cl);
3173 }
3174 }
3175
3176 void
3177 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
3178 if (freelistLock()->owned_by_self()) {
3179 Generation::safe_object_iterate(cl);
3180 } else {
3181 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3182 Generation::safe_object_iterate(cl);
3183 }
3184 }
3185
3186 void
3187 ConcurrentMarkSweepGeneration::post_compact() {
3188 }
3189
3190 void
3191 ConcurrentMarkSweepGeneration::prepare_for_verify() {
3192 // Fix the linear allocation blocks to look like free blocks.
3193
3194 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3195 // are not called when the heap is verified during universe initialization and
3196 // at vm shutdown.
3197 if (freelistLock()->owned_by_self()) {
3198 cmsSpace()->prepare_for_verify();
3199 } else {
3200 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3201 cmsSpace()->prepare_for_verify();
3202 }
3203 }
3204
3205 void
3206 ConcurrentMarkSweepGeneration::verify() {
3207 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3208 // are not called when the heap is verified during universe initialization and
3209 // at vm shutdown.
3210 if (freelistLock()->owned_by_self()) {
3211 cmsSpace()->verify();
3212 } else {
3213 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3214 cmsSpace()->verify();
3215 }
3216 }
3217
3218 void CMSCollector::verify() {
3219 _cmsGen->verify();
3220 }
3221
3222 #ifndef PRODUCT
3223 bool CMSCollector::overflow_list_is_empty() const {
3224 assert(_num_par_pushes >= 0, "Inconsistency");
3225 if (_overflow_list == NULL) {
3226 assert(_num_par_pushes == 0, "Inconsistency");
3227 }
3228 return _overflow_list == NULL;
3229 }
3230
3231 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3232 // merely consolidate assertion checks that appear to occur together frequently.
3233 void CMSCollector::verify_work_stacks_empty() const {
3234 assert(_markStack.isEmpty(), "Marking stack should be empty");
3235 assert(overflow_list_is_empty(), "Overflow list should be empty");
3236 }
3237
3238 void CMSCollector::verify_overflow_empty() const {
3239 assert(overflow_list_is_empty(), "Overflow list should be empty");
3240 assert(no_preserved_marks(), "No preserved marks");
3241 }
3242 #endif // PRODUCT
3243
3244 // Decide if we want to enable class unloading as part of the
3245 // ensuing concurrent GC cycle. We will collect and
3246 // unload classes if it's the case that:
3247 // (1) an explicit gc request has been made and the flag
3248 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3249 // (2) (a) class unloading is enabled at the command line, and
3250 // (b) old gen is getting really full
3251 // NOTE: Provided there is no change in the state of the heap between
3252 // calls to this method, it should have idempotent results. Moreover,
3253 // its results should be monotonically increasing (i.e. going from 0 to 1,
3254 // but not 1 to 0) between successive calls between which the heap was
3255 // not collected. For the implementation below, it must thus rely on
3256 // the property that concurrent_cycles_since_last_unload()
3257 // will not decrease unless a collection cycle happened and that
3258 // _cmsGen->is_too_full() are
3259 // themselves also monotonic in that sense. See check_monotonicity()
3260 // below.
3261 void CMSCollector::update_should_unload_classes() {
3262 _should_unload_classes = false;
3263 // Condition 1 above
3264 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3265 _should_unload_classes = true;
3266 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3267 // Disjuncts 2.b.(i,ii,iii) above
3268 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3269 CMSClassUnloadingMaxInterval)
3270 || _cmsGen->is_too_full();
3271 }
3272 }
3273
3274 bool ConcurrentMarkSweepGeneration::is_too_full() const {
3275 bool res = should_concurrent_collect();
3276 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
3277 return res;
3278 }
3279
3280 void CMSCollector::setup_cms_unloading_and_verification_state() {
3281 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3282 || VerifyBeforeExit;
3283 const int rso = SharedHeap::SO_AllCodeCache;
3284
3285 // We set the proper root for this CMS cycle here.
3286 if (should_unload_classes()) { // Should unload classes this cycle
3287 remove_root_scanning_option(rso); // Shrink the root set appropriately
3288 set_verifying(should_verify); // Set verification state for this cycle
3289 return; // Nothing else needs to be done at this time
3290 }
3291
3292 // Not unloading classes this cycle
3293 assert(!should_unload_classes(), "Inconsistency!");
3294
3295 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3296 // Include symbols, strings and code cache elements to prevent their resurrection.
3297 add_root_scanning_option(rso);
3298 set_verifying(true);
3299 } else if (verifying() && !should_verify) {
3300 // We were verifying, but some verification flags got disabled.
3301 set_verifying(false);
3302 // Exclude symbols, strings and code cache elements from root scanning to
3303 // reduce IM and RM pauses.
3304 remove_root_scanning_option(rso);
3305 }
3306 }
3307
3308
3309 #ifndef PRODUCT
3310 HeapWord* CMSCollector::block_start(const void* p) const {
3311 const HeapWord* addr = (HeapWord*)p;
3312 if (_span.contains(p)) {
3313 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3314 return _cmsGen->cmsSpace()->block_start(p);
3315 }
3316 }
3317 return NULL;
3318 }
3319 #endif
3320
3321 HeapWord*
3322 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3323 bool tlab,
3324 bool parallel) {
3325 CMSSynchronousYieldRequest yr;
3326 assert(!tlab, "Can't deal with TLAB allocation");
3327 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3328 expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3329 CMSExpansionCause::_satisfy_allocation);
3330 if (GCExpandToAllocateDelayMillis > 0) {
3331 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3332 }
3333 return have_lock_and_allocate(word_size, tlab);
3334 }
3335
3336 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3337 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3338 // to CardGeneration and share it...
3339 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) {
3340 return CardGeneration::expand(bytes, expand_bytes);
3341 }
3342
3343 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3344 CMSExpansionCause::Cause cause)
3345 {
3346
3347 bool success = expand(bytes, expand_bytes);
3348
3349 // remember why we expanded; this information is used
3350 // by shouldConcurrentCollect() when making decisions on whether to start
3351 // a new CMS cycle.
3352 if (success) {
3353 set_expansion_cause(cause);
3354 if (PrintGCDetails && Verbose) {
3355 gclog_or_tty->print_cr("Expanded CMS gen for %s",
3356 CMSExpansionCause::to_string(cause));
3357 }
3358 }
3359 }
3360
3361 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3362 HeapWord* res = NULL;
3363 MutexLocker x(ParGCRareEvent_lock);
3364
3365 // Check again here while holding the lock.
3366 if (CMSFastPromotionFailure && has_promotion_failed()) {
3367 return NULL;
3368 }
3369
3370 while (true) {
3371 // Expansion by some other thread might make alloc OK now:
3372 res = ps->lab.alloc(word_sz);
3373 if (res != NULL) return res;
3374 // If there's not enough expansion space available, give up.
3375 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3376 set_promotion_failed();
3377 return NULL;
3378 }
3379 // Otherwise, we try expansion.
3380 expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3381 CMSExpansionCause::_allocate_par_lab);
3382 // Now go around the loop and try alloc again;
3383 // A competing par_promote might beat us to the expansion space,
3384 // so we may go around the loop again if promotion fails again.
3385 if (GCExpandToAllocateDelayMillis > 0) {
3386 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3387 }
3388 }
3389 }
3390
3391
3392 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3393 PromotionInfo* promo) {
3394 MutexLocker x(ParGCRareEvent_lock);
3395
3396 // Check again here while holding the lock.
3397 if (CMSFastPromotionFailure && has_promotion_failed()) {
3398 return false;
3399 }
3400
3401 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3402 while (true) {
3403 // Expansion by some other thread might make alloc OK now:
3404 if (promo->ensure_spooling_space()) {
3405 assert(promo->has_spooling_space(),
3406 "Post-condition of successful ensure_spooling_space()");
3407 return true;
3408 }
3409 // If there's not enough expansion space available, give up.
3410 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3411 set_promotion_failed();
3412 return false;
3413 }
3414 // Otherwise, we try expansion.
3415 expand(refill_size_bytes, MinHeapDeltaBytes,
3416 CMSExpansionCause::_allocate_par_spooling_space);
3417 // Now go around the loop and try alloc again;
3418 // A competing allocation might beat us to the expansion space,
3419 // so we may go around the loop again if allocation fails again.
3420 if (GCExpandToAllocateDelayMillis > 0) {
3421 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3422 }
3423 }
3424 }
3425
3426
3427 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3428 assert_locked_or_safepoint(ExpandHeap_lock);
3429 // Shrink committed space
3430 _virtual_space.shrink_by(bytes);
3431 // Shrink space; this also shrinks the space's BOT
3432 _cmsSpace->set_end((HeapWord*) _virtual_space.high());
3433 size_t new_word_size = heap_word_size(_cmsSpace->capacity());
3434 // Shrink the shared block offset array
3435 _bts->resize(new_word_size);
3436 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3437 // Shrink the card table
3438 Universe::heap()->barrier_set()->resize_covered_region(mr);
3439
3440 if (Verbose && PrintGC) {
3441 size_t new_mem_size = _virtual_space.committed_size();
3442 size_t old_mem_size = new_mem_size + bytes;
3443 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
3444 name(), old_mem_size/K, new_mem_size/K);
3445 }
3446 }
3447
3448 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3449 assert_locked_or_safepoint(Heap_lock);
3450 size_t size = ReservedSpace::page_align_size_down(bytes);
3451 // Only shrink if a compaction was done so that all the free space
3452 // in the generation is in a contiguous block at the end.
3453 if (size > 0 && did_compact()) {
3454 shrink_by(size);
3455 }
3456 }
3457
3458 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3459 assert_locked_or_safepoint(Heap_lock);
3460 bool result = _virtual_space.expand_by(bytes);
3461 if (result) {
3462 size_t new_word_size =
3463 heap_word_size(_virtual_space.committed_size());
3464 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3465 _bts->resize(new_word_size); // resize the block offset shared array
3466 Universe::heap()->barrier_set()->resize_covered_region(mr);
3467 // Hmmmm... why doesn't CFLS::set_end verify locking?
3468 // This is quite ugly; FIX ME XXX
3469 _cmsSpace->assert_locked(freelistLock());
3470 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3471
3472 // update the space and generation capacity counters
3473 if (UsePerfData) {
3474 _space_counters->update_capacity();
3475 _gen_counters->update_all();
3476 }
3477
3478 if (Verbose && PrintGC) {
3479 size_t new_mem_size = _virtual_space.committed_size();
3480 size_t old_mem_size = new_mem_size - bytes;
3481 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K",
3482 name(), old_mem_size/K, bytes/K, new_mem_size/K);
3483 }
3484 }
3485 return result;
3486 }
3487
3488 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3489 assert_locked_or_safepoint(Heap_lock);
3490 bool success = true;
3491 const size_t remaining_bytes = _virtual_space.uncommitted_size();
3492 if (remaining_bytes > 0) {
3493 success = grow_by(remaining_bytes);
3494 DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3495 }
3496 return success;
3497 }
3498
3499 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
3500 assert_locked_or_safepoint(Heap_lock);
3501 assert_lock_strong(freelistLock());
3502 if (PrintGCDetails && Verbose) {
3503 warning("Shrinking of CMS not yet implemented");
3504 }
3505 return;
3506 }
3507
3508
3509 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3510 // phases.
3511 class CMSPhaseAccounting: public StackObj {
3512 public:
3513 CMSPhaseAccounting(CMSCollector *collector,
3514 const char *phase,
3515 const GCId gc_id,
3516 bool print_cr = true);
3517 ~CMSPhaseAccounting();
3518
3519 private:
3520 CMSCollector *_collector;
3521 const char *_phase;
3522 elapsedTimer _wallclock;
3523 bool _print_cr;
3524 const GCId _gc_id;
3525
3526 public:
3527 // Not MT-safe; so do not pass around these StackObj's
3528 // where they may be accessed by other threads.
3529 jlong wallclock_millis() {
3530 assert(_wallclock.is_active(), "Wall clock should not stop");
3531 _wallclock.stop(); // to record time
3532 jlong ret = _wallclock.milliseconds();
3533 _wallclock.start(); // restart
3534 return ret;
3535 }
3536 };
3537
3538 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3539 const char *phase,
3540 const GCId gc_id,
3541 bool print_cr) :
3542 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) {
3543
3544 if (PrintCMSStatistics != 0) {
3545 _collector->resetYields();
3546 }
3547 if (PrintGCDetails) {
3548 gclog_or_tty->gclog_stamp(_gc_id);
3549 gclog_or_tty->print_cr("[%s-concurrent-%s-start]",
3550 _collector->cmsGen()->short_name(), _phase);
3551 }
3552 _collector->resetTimer();
3553 _wallclock.start();
3554 _collector->startTimer();
3555 }
3556
3557 CMSPhaseAccounting::~CMSPhaseAccounting() {
3558 assert(_wallclock.is_active(), "Wall clock should not have stopped");
3559 _collector->stopTimer();
3560 _wallclock.stop();
3561 if (PrintGCDetails) {
3562 gclog_or_tty->gclog_stamp(_gc_id);
3563 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3564 _collector->cmsGen()->short_name(),
3565 _phase, _collector->timerValue(), _wallclock.seconds());
3566 if (_print_cr) {
3567 gclog_or_tty->cr();
3568 }
3569 if (PrintCMSStatistics != 0) {
3570 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3571 _collector->yields());
3572 }
3573 }
3574 }
3575
3576 // CMS work
3577
3578 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
3579 class CMSParMarkTask : public AbstractGangTask {
3580 protected:
3581 CMSCollector* _collector;
3582 int _n_workers;
3583 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) :
3584 AbstractGangTask(name),
3585 _collector(collector),
3586 _n_workers(n_workers) {}
3587 // Work method in support of parallel rescan ... of young gen spaces
3588 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
3589 ContiguousSpace* space,
3590 HeapWord** chunk_array, size_t chunk_top);
3591 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
3592 };
3593
3594 // Parallel initial mark task
3595 class CMSParInitialMarkTask: public CMSParMarkTask {
3596 public:
3597 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) :
3598 CMSParMarkTask("Scan roots and young gen for initial mark in parallel",
3599 collector, n_workers) {}
3600 void work(uint worker_id);
3601 };
3602
3603 // Checkpoint the roots into this generation from outside
3604 // this generation. [Note this initial checkpoint need only
3605 // be approximate -- we'll do a catch up phase subsequently.]
3606 void CMSCollector::checkpointRootsInitial(bool asynch) {
3607 assert(_collectorState == InitialMarking, "Wrong collector state");
3608 check_correct_thread_executing();
3609 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
3610
3611 save_heap_summary();
3612 report_heap_summary(GCWhen::BeforeGC);
3613
3614 ReferenceProcessor* rp = ref_processor();
3615 SpecializationStats::clear();
3616 assert(_restart_addr == NULL, "Control point invariant");
3617 if (asynch) {
3618 // acquire locks for subsequent manipulations
3619 MutexLockerEx x(bitMapLock(),
3620 Mutex::_no_safepoint_check_flag);
3621 checkpointRootsInitialWork(asynch);
3622 // enable ("weak") refs discovery
3623 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/);
3624 _collectorState = Marking;
3625 } else {
3626 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3627 // which recognizes if we are a CMS generation, and doesn't try to turn on
3628 // discovery; verify that they aren't meddling.
3629 assert(!rp->discovery_is_atomic(),
3630 "incorrect setting of discovery predicate");
3631 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3632 "ref discovery for this generation kind");
3633 // already have locks
3634 checkpointRootsInitialWork(asynch);
3635 // now enable ("weak") refs discovery
3636 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/);
3637 _collectorState = Marking;
3638 }
3639 SpecializationStats::print();
3640 }
3641
3642 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3643 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3644 assert(_collectorState == InitialMarking, "just checking");
3645
3646 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3647 // precede our marking with a collection of all
3648 // younger generations to keep floating garbage to a minimum.
3649 // XXX: we won't do this for now -- it's an optimization to be done later.
3650
3651 // already have locks
3652 assert_lock_strong(bitMapLock());
3653 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3654
3655 // Setup the verification and class unloading state for this
3656 // CMS collection cycle.
3657 setup_cms_unloading_and_verification_state();
3658
3659 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
3660 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
3661
3662 // Reset all the PLAB chunk arrays if necessary.
3663 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3664 reset_survivor_plab_arrays();
3665 }
3666
3667 ResourceMark rm;
3668 HandleMark hm;
3669
3670 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
3671 GenCollectedHeap* gch = GenCollectedHeap::heap();
3672
3673 verify_work_stacks_empty();
3674 verify_overflow_empty();
3675
3676 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3677 // Update the saved marks which may affect the root scans.
3678 gch->save_marks();
3679
3680 // weak reference processing has not started yet.
3681 ref_processor()->set_enqueuing_is_done(false);
3682
3683 // Need to remember all newly created CLDs,
3684 // so that we can guarantee that the remark finds them.
3685 ClassLoaderDataGraph::remember_new_clds(true);
3686
3687 // Whenever a CLD is found, it will be claimed before proceeding to mark
3688 // the klasses. The claimed marks need to be cleared before marking starts.
3689 ClassLoaderDataGraph::clear_claimed_marks();
3690
3691 if (CMSPrintEdenSurvivorChunks) {
3692 print_eden_and_survivor_chunk_arrays();
3693 }
3694
3695 {
3696 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3697 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
3698 // The parallel version.
3699 FlexibleWorkGang* workers = gch->workers();
3700 assert(workers != NULL, "Need parallel worker threads.");
3701 int n_workers = workers->active_workers();
3702 CMSParInitialMarkTask tsk(this, n_workers);
3703 gch->set_par_threads(n_workers);
3704 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
3705 if (n_workers > 1) {
3706 GenCollectedHeap::StrongRootsScope srs(gch);
3707 workers->run_task(&tsk);
3708 } else {
3709 GenCollectedHeap::StrongRootsScope srs(gch);
3710 tsk.work(0);
3711 }
3712 gch->set_par_threads(0);
3713 } else {
3714 // The serial version.
3715 CLDToOopClosure cld_closure(¬Older, true);
3716 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3717 gch->gen_process_roots(_cmsGen->level(),
3718 true, // younger gens are roots
3719 true, // activate StrongRootsScope
3720 SharedHeap::ScanningOption(roots_scanning_options()),
3721 should_unload_classes(),
3722 ¬Older,
3723 NULL,
3724 &cld_closure);
3725 }
3726 }
3727
3728 // Clear mod-union table; it will be dirtied in the prologue of
3729 // CMS generation per each younger generation collection.
3730
3731 assert(_modUnionTable.isAllClear(),
3732 "Was cleared in most recent final checkpoint phase"
3733 " or no bits are set in the gc_prologue before the start of the next "
3734 "subsequent marking phase.");
3735
3736 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
3737
3738 // Save the end of the used_region of the constituent generations
3739 // to be used to limit the extent of sweep in each generation.
3740 save_sweep_limits();
3741 verify_overflow_empty();
3742 }
3743
3744 bool CMSCollector::markFromRoots(bool asynch) {
3745 // we might be tempted to assert that:
3746 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3747 // "inconsistent argument?");
3748 // However that wouldn't be right, because it's possible that
3749 // a safepoint is indeed in progress as a younger generation
3750 // stop-the-world GC happens even as we mark in this generation.
3751 assert(_collectorState == Marking, "inconsistent state?");
3752 check_correct_thread_executing();
3753 verify_overflow_empty();
3754
3755 bool res;
3756 if (asynch) {
3757 // Weak ref discovery note: We may be discovering weak
3758 // refs in this generation concurrent (but interleaved) with
3759 // weak ref discovery by a younger generation collector.
3760
3761 CMSTokenSyncWithLocks ts(true, bitMapLock());
3762 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3763 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3764 res = markFromRootsWork(asynch);
3765 if (res) {
3766 _collectorState = Precleaning;
3767 } else { // We failed and a foreground collection wants to take over
3768 assert(_foregroundGCIsActive, "internal state inconsistency");
3769 assert(_restart_addr == NULL, "foreground will restart from scratch");
3770 if (PrintGCDetails) {
3771 gclog_or_tty->print_cr("bailing out to foreground collection");
3772 }
3773 }
3774 } else {
3775 assert(SafepointSynchronize::is_at_safepoint(),
3776 "inconsistent with asynch == false");
3777 // already have locks
3778 res = markFromRootsWork(asynch);
3779 _collectorState = FinalMarking;
3780 }
3781 verify_overflow_empty();
3782 return res;
3783 }
3784
3785 bool CMSCollector::markFromRootsWork(bool asynch) {
3786 // iterate over marked bits in bit map, doing a full scan and mark
3787 // from these roots using the following algorithm:
3788 // . if oop is to the right of the current scan pointer,
3789 // mark corresponding bit (we'll process it later)
3790 // . else (oop is to left of current scan pointer)
3791 // push oop on marking stack
3792 // . drain the marking stack
3793
3794 // Note that when we do a marking step we need to hold the
3795 // bit map lock -- recall that direct allocation (by mutators)
3796 // and promotion (by younger generation collectors) is also
3797 // marking the bit map. [the so-called allocate live policy.]
3798 // Because the implementation of bit map marking is not
3799 // robust wrt simultaneous marking of bits in the same word,
3800 // we need to make sure that there is no such interference
3801 // between concurrent such updates.
3802
3803 // already have locks
3804 assert_lock_strong(bitMapLock());
3805
3806 verify_work_stacks_empty();
3807 verify_overflow_empty();
3808 bool result = false;
3809 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3810 result = do_marking_mt(asynch);
3811 } else {
3812 result = do_marking_st(asynch);
3813 }
3814 return result;
3815 }
3816
3817 // Forward decl
3818 class CMSConcMarkingTask;
3819
3820 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3821 CMSCollector* _collector;
3822 CMSConcMarkingTask* _task;
3823 public:
3824 virtual void yield();
3825
3826 // "n_threads" is the number of threads to be terminated.
3827 // "queue_set" is a set of work queues of other threads.
3828 // "collector" is the CMS collector associated with this task terminator.
3829 // "yield" indicates whether we need the gang as a whole to yield.
3830 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3831 ParallelTaskTerminator(n_threads, queue_set),
3832 _collector(collector) { }
3833
3834 void set_task(CMSConcMarkingTask* task) {
3835 _task = task;
3836 }
3837 };
3838
3839 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3840 CMSConcMarkingTask* _task;
3841 public:
3842 bool should_exit_termination();
3843 void set_task(CMSConcMarkingTask* task) {
3844 _task = task;
3845 }
3846 };
3847
3848 // MT Concurrent Marking Task
3849 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3850 CMSCollector* _collector;
3851 int _n_workers; // requested/desired # workers
3852 bool _asynch;
3853 bool _result;
3854 CompactibleFreeListSpace* _cms_space;
3855 char _pad_front[64]; // padding to ...
3856 HeapWord* _global_finger; // ... avoid sharing cache line
3857 char _pad_back[64];
3858 HeapWord* _restart_addr;
3859
3860 // Exposed here for yielding support
3861 Mutex* const _bit_map_lock;
3862
3863 // The per thread work queues, available here for stealing
3864 OopTaskQueueSet* _task_queues;
3865
3866 // Termination (and yielding) support
3867 CMSConcMarkingTerminator _term;
3868 CMSConcMarkingTerminatorTerminator _term_term;
3869
3870 public:
3871 CMSConcMarkingTask(CMSCollector* collector,
3872 CompactibleFreeListSpace* cms_space,
3873 bool asynch,
3874 YieldingFlexibleWorkGang* workers,
3875 OopTaskQueueSet* task_queues):
3876 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3877 _collector(collector),
3878 _cms_space(cms_space),
3879 _asynch(asynch), _n_workers(0), _result(true),
3880 _task_queues(task_queues),
3881 _term(_n_workers, task_queues, _collector),
3882 _bit_map_lock(collector->bitMapLock())
3883 {
3884 _requested_size = _n_workers;
3885 _term.set_task(this);
3886 _term_term.set_task(this);
3887 _restart_addr = _global_finger = _cms_space->bottom();
3888 }
3889
3890
3891 OopTaskQueueSet* task_queues() { return _task_queues; }
3892
3893 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3894
3895 HeapWord** global_finger_addr() { return &_global_finger; }
3896
3897 CMSConcMarkingTerminator* terminator() { return &_term; }
3898
3899 virtual void set_for_termination(int active_workers) {
3900 terminator()->reset_for_reuse(active_workers);
3901 }
3902
3903 void work(uint worker_id);
3904 bool should_yield() {
3905 return ConcurrentMarkSweepThread::should_yield()
3906 && !_collector->foregroundGCIsActive()
3907 && _asynch;
3908 }
3909
3910 virtual void coordinator_yield(); // stuff done by coordinator
3911 bool result() { return _result; }
3912
3913 void reset(HeapWord* ra) {
3914 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3915 _restart_addr = _global_finger = ra;
3916 _term.reset_for_reuse();
3917 }
3918
3919 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3920 OopTaskQueue* work_q);
3921
3922 private:
3923 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3924 void do_work_steal(int i);
3925 void bump_global_finger(HeapWord* f);
3926 };
3927
3928 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3929 assert(_task != NULL, "Error");
3930 return _task->yielding();
3931 // Note that we do not need the disjunct || _task->should_yield() above
3932 // because we want terminating threads to yield only if the task
3933 // is already in the midst of yielding, which happens only after at least one
3934 // thread has yielded.
3935 }
3936
3937 void CMSConcMarkingTerminator::yield() {
3938 if (_task->should_yield()) {
3939 _task->yield();
3940 } else {
3941 ParallelTaskTerminator::yield();
3942 }
3943 }
3944
3945 ////////////////////////////////////////////////////////////////
3946 // Concurrent Marking Algorithm Sketch
3947 ////////////////////////////////////////////////////////////////
3948 // Until all tasks exhausted (both spaces):
3949 // -- claim next available chunk
3950 // -- bump global finger via CAS
3951 // -- find first object that starts in this chunk
3952 // and start scanning bitmap from that position
3953 // -- scan marked objects for oops
3954 // -- CAS-mark target, and if successful:
3955 // . if target oop is above global finger (volatile read)
3956 // nothing to do
3957 // . if target oop is in chunk and above local finger
3958 // then nothing to do
3959 // . else push on work-queue
3960 // -- Deal with possible overflow issues:
3961 // . local work-queue overflow causes stuff to be pushed on
3962 // global (common) overflow queue
3963 // . always first empty local work queue
3964 // . then get a batch of oops from global work queue if any
3965 // . then do work stealing
3966 // -- When all tasks claimed (both spaces)
3967 // and local work queue empty,
3968 // then in a loop do:
3969 // . check global overflow stack; steal a batch of oops and trace
3970 // . try to steal from other threads oif GOS is empty
3971 // . if neither is available, offer termination
3972 // -- Terminate and return result
3973 //
3974 void CMSConcMarkingTask::work(uint worker_id) {
3975 elapsedTimer _timer;
3976 ResourceMark rm;
3977 HandleMark hm;
3978
3979 DEBUG_ONLY(_collector->verify_overflow_empty();)
3980
3981 // Before we begin work, our work queue should be empty
3982 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3983 // Scan the bitmap covering _cms_space, tracing through grey objects.
3984 _timer.start();
3985 do_scan_and_mark(worker_id, _cms_space);
3986 _timer.stop();
3987 if (PrintCMSStatistics != 0) {
3988 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3989 worker_id, _timer.seconds());
3990 // XXX: need xxx/xxx type of notation, two timers
3991 }
3992
3993 // ... do work stealing
3994 _timer.reset();
3995 _timer.start();
3996 do_work_steal(worker_id);
3997 _timer.stop();
3998 if (PrintCMSStatistics != 0) {
3999 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
4000 worker_id, _timer.seconds());
4001 // XXX: need xxx/xxx type of notation, two timers
4002 }
4003 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
4004 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
4005 // Note that under the current task protocol, the
4006 // following assertion is true even of the spaces
4007 // expanded since the completion of the concurrent
4008 // marking. XXX This will likely change under a strict
4009 // ABORT semantics.
4010 // After perm removal the comparison was changed to
4011 // greater than or equal to from strictly greater than.
4012 // Before perm removal the highest address sweep would
4013 // have been at the end of perm gen but now is at the
4014 // end of the tenured gen.
4015 assert(_global_finger >= _cms_space->end(),
4016 "All tasks have been completed");
4017 DEBUG_ONLY(_collector->verify_overflow_empty();)
4018 }
4019
4020 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
4021 HeapWord* read = _global_finger;
4022 HeapWord* cur = read;
4023 while (f > read) {
4024 cur = read;
4025 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
4026 if (cur == read) {
4027 // our cas succeeded
4028 assert(_global_finger >= f, "protocol consistency");
4029 break;
4030 }
4031 }
4032 }
4033
4034 // This is really inefficient, and should be redone by
4035 // using (not yet available) block-read and -write interfaces to the
4036 // stack and the work_queue. XXX FIX ME !!!
4037 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
4038 OopTaskQueue* work_q) {
4039 // Fast lock-free check
4040 if (ovflw_stk->length() == 0) {
4041 return false;
4042 }
4043 assert(work_q->size() == 0, "Shouldn't steal");
4044 MutexLockerEx ml(ovflw_stk->par_lock(),
4045 Mutex::_no_safepoint_check_flag);
4046 // Grab up to 1/4 the size of the work queue
4047 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4048 (size_t)ParGCDesiredObjsFromOverflowList);
4049 num = MIN2(num, ovflw_stk->length());
4050 for (int i = (int) num; i > 0; i--) {
4051 oop cur = ovflw_stk->pop();
4052 assert(cur != NULL, "Counted wrong?");
4053 work_q->push(cur);
4054 }
4055 return num > 0;
4056 }
4057
4058 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
4059 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4060 int n_tasks = pst->n_tasks();
4061 // We allow that there may be no tasks to do here because
4062 // we are restarting after a stack overflow.
4063 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
4064 uint nth_task = 0;
4065
4066 HeapWord* aligned_start = sp->bottom();
4067 if (sp->used_region().contains(_restart_addr)) {
4068 // Align down to a card boundary for the start of 0th task
4069 // for this space.
4070 aligned_start =
4071 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
4072 CardTableModRefBS::card_size);
4073 }
4074
4075 size_t chunk_size = sp->marking_task_size();
4076 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4077 // Having claimed the nth task in this space,
4078 // compute the chunk that it corresponds to:
4079 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
4080 aligned_start + (nth_task+1)*chunk_size);
4081 // Try and bump the global finger via a CAS;
4082 // note that we need to do the global finger bump
4083 // _before_ taking the intersection below, because
4084 // the task corresponding to that region will be
4085 // deemed done even if the used_region() expands
4086 // because of allocation -- as it almost certainly will
4087 // during start-up while the threads yield in the
4088 // closure below.
4089 HeapWord* finger = span.end();
4090 bump_global_finger(finger); // atomically
4091 // There are null tasks here corresponding to chunks
4092 // beyond the "top" address of the space.
4093 span = span.intersection(sp->used_region());
4094 if (!span.is_empty()) { // Non-null task
4095 HeapWord* prev_obj;
4096 assert(!span.contains(_restart_addr) || nth_task == 0,
4097 "Inconsistency");
4098 if (nth_task == 0) {
4099 // For the 0th task, we'll not need to compute a block_start.
4100 if (span.contains(_restart_addr)) {
4101 // In the case of a restart because of stack overflow,
4102 // we might additionally skip a chunk prefix.
4103 prev_obj = _restart_addr;
4104 } else {
4105 prev_obj = span.start();
4106 }
4107 } else {
4108 // We want to skip the first object because
4109 // the protocol is to scan any object in its entirety
4110 // that _starts_ in this span; a fortiori, any
4111 // object starting in an earlier span is scanned
4112 // as part of an earlier claimed task.
4113 // Below we use the "careful" version of block_start
4114 // so we do not try to navigate uninitialized objects.
4115 prev_obj = sp->block_start_careful(span.start());
4116 // Below we use a variant of block_size that uses the
4117 // Printezis bits to avoid waiting for allocated
4118 // objects to become initialized/parsable.
4119 while (prev_obj < span.start()) {
4120 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
4121 if (sz > 0) {
4122 prev_obj += sz;
4123 } else {
4124 // In this case we may end up doing a bit of redundant
4125 // scanning, but that appears unavoidable, short of
4126 // locking the free list locks; see bug 6324141.
4127 break;
4128 }
4129 }
4130 }
4131 if (prev_obj < span.end()) {
4132 MemRegion my_span = MemRegion(prev_obj, span.end());
4133 // Do the marking work within a non-empty span --
4134 // the last argument to the constructor indicates whether the
4135 // iteration should be incremental with periodic yields.
4136 Par_MarkFromRootsClosure cl(this, _collector, my_span,
4137 &_collector->_markBitMap,
4138 work_queue(i),
4139 &_collector->_markStack,
4140 _asynch);
4141 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
4142 } // else nothing to do for this task
4143 } // else nothing to do for this task
4144 }
4145 // We'd be tempted to assert here that since there are no
4146 // more tasks left to claim in this space, the global_finger
4147 // must exceed space->top() and a fortiori space->end(). However,
4148 // that would not quite be correct because the bumping of
4149 // global_finger occurs strictly after the claiming of a task,
4150 // so by the time we reach here the global finger may not yet
4151 // have been bumped up by the thread that claimed the last
4152 // task.
4153 pst->all_tasks_completed();
4154 }
4155
4156 class Par_ConcMarkingClosure: public MetadataAwareOopClosure {
4157 private:
4158 CMSCollector* _collector;
4159 CMSConcMarkingTask* _task;
4160 MemRegion _span;
4161 CMSBitMap* _bit_map;
4162 CMSMarkStack* _overflow_stack;
4163 OopTaskQueue* _work_queue;
4164 protected:
4165 DO_OOP_WORK_DEFN
4166 public:
4167 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
4168 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
4169 MetadataAwareOopClosure(collector->ref_processor()),
4170 _collector(collector),
4171 _task(task),
4172 _span(collector->_span),
4173 _work_queue(work_queue),
4174 _bit_map(bit_map),
4175 _overflow_stack(overflow_stack)
4176 { }
4177 virtual void do_oop(oop* p);
4178 virtual void do_oop(narrowOop* p);
4179
4180 void trim_queue(size_t max);
4181 void handle_stack_overflow(HeapWord* lost);
4182 void do_yield_check() {
4183 if (_task->should_yield()) {
4184 _task->yield();
4185 }
4186 }
4187 };
4188
4189 // Grey object scanning during work stealing phase --
4190 // the salient assumption here is that any references
4191 // that are in these stolen objects being scanned must
4192 // already have been initialized (else they would not have
4193 // been published), so we do not need to check for
4194 // uninitialized objects before pushing here.
4195 void Par_ConcMarkingClosure::do_oop(oop obj) {
4196 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
4197 HeapWord* addr = (HeapWord*)obj;
4198 // Check if oop points into the CMS generation
4199 // and is not marked
4200 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
4201 // a white object ...
4202 // If we manage to "claim" the object, by being the
4203 // first thread to mark it, then we push it on our
4204 // marking stack
4205 if (_bit_map->par_mark(addr)) { // ... now grey
4206 // push on work queue (grey set)
4207 bool simulate_overflow = false;
4208 NOT_PRODUCT(
4209 if (CMSMarkStackOverflowALot &&
4210 _collector->simulate_overflow()) {
4211 // simulate a stack overflow
4212 simulate_overflow = true;
4213 }
4214 )
4215 if (simulate_overflow ||
4216 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
4217 // stack overflow
4218 if (PrintCMSStatistics != 0) {
4219 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
4220 SIZE_FORMAT, _overflow_stack->capacity());
4221 }
4222 // We cannot assert that the overflow stack is full because
4223 // it may have been emptied since.
4224 assert(simulate_overflow ||
4225 _work_queue->size() == _work_queue->max_elems(),
4226 "Else push should have succeeded");
4227 handle_stack_overflow(addr);
4228 }
4229 } // Else, some other thread got there first
4230 do_yield_check();
4231 }
4232 }
4233
4234 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4235 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4236
4237 void Par_ConcMarkingClosure::trim_queue(size_t max) {
4238 while (_work_queue->size() > max) {
4239 oop new_oop;
4240 if (_work_queue->pop_local(new_oop)) {
4241 assert(new_oop->is_oop(), "Should be an oop");
4242 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
4243 assert(_span.contains((HeapWord*)new_oop), "Not in span");
4244 new_oop->oop_iterate(this); // do_oop() above
4245 do_yield_check();
4246 }
4247 }
4248 }
4249
4250 // Upon stack overflow, we discard (part of) the stack,
4251 // remembering the least address amongst those discarded
4252 // in CMSCollector's _restart_address.
4253 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
4254 // We need to do this under a mutex to prevent other
4255 // workers from interfering with the work done below.
4256 MutexLockerEx ml(_overflow_stack->par_lock(),
4257 Mutex::_no_safepoint_check_flag);
4258 // Remember the least grey address discarded
4259 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
4260 _collector->lower_restart_addr(ra);
4261 _overflow_stack->reset(); // discard stack contents
4262 _overflow_stack->expand(); // expand the stack if possible
4263 }
4264
4265
4266 void CMSConcMarkingTask::do_work_steal(int i) {
4267 OopTaskQueue* work_q = work_queue(i);
4268 oop obj_to_scan;
4269 CMSBitMap* bm = &(_collector->_markBitMap);
4270 CMSMarkStack* ovflw = &(_collector->_markStack);
4271 int* seed = _collector->hash_seed(i);
4272 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
4273 while (true) {
4274 cl.trim_queue(0);
4275 assert(work_q->size() == 0, "Should have been emptied above");
4276 if (get_work_from_overflow_stack(ovflw, work_q)) {
4277 // Can't assert below because the work obtained from the
4278 // overflow stack may already have been stolen from us.
4279 // assert(work_q->size() > 0, "Work from overflow stack");
4280 continue;
4281 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4282 assert(obj_to_scan->is_oop(), "Should be an oop");
4283 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
4284 obj_to_scan->oop_iterate(&cl);
4285 } else if (terminator()->offer_termination(&_term_term)) {
4286 assert(work_q->size() == 0, "Impossible!");
4287 break;
4288 } else if (yielding() || should_yield()) {
4289 yield();
4290 }
4291 }
4292 }
4293
4294 // This is run by the CMS (coordinator) thread.
4295 void CMSConcMarkingTask::coordinator_yield() {
4296 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4297 "CMS thread should hold CMS token");
4298 // First give up the locks, then yield, then re-lock
4299 // We should probably use a constructor/destructor idiom to
4300 // do this unlock/lock or modify the MutexUnlocker class to
4301 // serve our purpose. XXX
4302 assert_lock_strong(_bit_map_lock);
4303 _bit_map_lock->unlock();
4304 ConcurrentMarkSweepThread::desynchronize(true);
4305 ConcurrentMarkSweepThread::acknowledge_yield_request();
4306 _collector->stopTimer();
4307 if (PrintCMSStatistics != 0) {
4308 _collector->incrementYields();
4309 }
4310 _collector->icms_wait();
4311
4312 // It is possible for whichever thread initiated the yield request
4313 // not to get a chance to wake up and take the bitmap lock between
4314 // this thread releasing it and reacquiring it. So, while the
4315 // should_yield() flag is on, let's sleep for a bit to give the
4316 // other thread a chance to wake up. The limit imposed on the number
4317 // of iterations is defensive, to avoid any unforseen circumstances
4318 // putting us into an infinite loop. Since it's always been this
4319 // (coordinator_yield()) method that was observed to cause the
4320 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4321 // which is by default non-zero. For the other seven methods that
4322 // also perform the yield operation, as are using a different
4323 // parameter (CMSYieldSleepCount) which is by default zero. This way we
4324 // can enable the sleeping for those methods too, if necessary.
4325 // See 6442774.
4326 //
4327 // We really need to reconsider the synchronization between the GC
4328 // thread and the yield-requesting threads in the future and we
4329 // should really use wait/notify, which is the recommended
4330 // way of doing this type of interaction. Additionally, we should
4331 // consolidate the eight methods that do the yield operation and they
4332 // are almost identical into one for better maintainability and
4333 // readability. See 6445193.
4334 //
4335 // Tony 2006.06.29
4336 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4337 ConcurrentMarkSweepThread::should_yield() &&
4338 !CMSCollector::foregroundGCIsActive(); ++i) {
4339 os::sleep(Thread::current(), 1, false);
4340 ConcurrentMarkSweepThread::acknowledge_yield_request();
4341 }
4342
4343 ConcurrentMarkSweepThread::synchronize(true);
4344 _bit_map_lock->lock_without_safepoint_check();
4345 _collector->startTimer();
4346 }
4347
4348 bool CMSCollector::do_marking_mt(bool asynch) {
4349 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
4350 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers(
4351 conc_workers()->total_workers(),
4352 conc_workers()->active_workers(),
4353 Threads::number_of_non_daemon_threads());
4354 conc_workers()->set_active_workers(num_workers);
4355
4356 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4357
4358 CMSConcMarkingTask tsk(this,
4359 cms_space,
4360 asynch,
4361 conc_workers(),
4362 task_queues());
4363
4364 // Since the actual number of workers we get may be different
4365 // from the number we requested above, do we need to do anything different
4366 // below? In particular, may be we need to subclass the SequantialSubTasksDone
4367 // class?? XXX
4368 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4369
4370 // Refs discovery is already non-atomic.
4371 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4372 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
4373 conc_workers()->start_task(&tsk);
4374 while (tsk.yielded()) {
4375 tsk.coordinator_yield();
4376 conc_workers()->continue_task(&tsk);
4377 }
4378 // If the task was aborted, _restart_addr will be non-NULL
4379 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4380 while (_restart_addr != NULL) {
4381 // XXX For now we do not make use of ABORTED state and have not
4382 // yet implemented the right abort semantics (even in the original
4383 // single-threaded CMS case). That needs some more investigation
4384 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4385 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4386 // If _restart_addr is non-NULL, a marking stack overflow
4387 // occurred; we need to do a fresh marking iteration from the
4388 // indicated restart address.
4389 if (_foregroundGCIsActive && asynch) {
4390 // We may be running into repeated stack overflows, having
4391 // reached the limit of the stack size, while making very
4392 // slow forward progress. It may be best to bail out and
4393 // let the foreground collector do its job.
4394 // Clear _restart_addr, so that foreground GC
4395 // works from scratch. This avoids the headache of
4396 // a "rescan" which would otherwise be needed because
4397 // of the dirty mod union table & card table.
4398 _restart_addr = NULL;
4399 return false;
4400 }
4401 // Adjust the task to restart from _restart_addr
4402 tsk.reset(_restart_addr);
4403 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4404 _restart_addr);
4405 _restart_addr = NULL;
4406 // Get the workers going again
4407 conc_workers()->start_task(&tsk);
4408 while (tsk.yielded()) {
4409 tsk.coordinator_yield();
4410 conc_workers()->continue_task(&tsk);
4411 }
4412 }
4413 assert(tsk.completed(), "Inconsistency");
4414 assert(tsk.result() == true, "Inconsistency");
4415 return true;
4416 }
4417
4418 bool CMSCollector::do_marking_st(bool asynch) {
4419 ResourceMark rm;
4420 HandleMark hm;
4421
4422 // Temporarily make refs discovery single threaded (non-MT)
4423 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
4424 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4425 &_markStack, CMSYield && asynch);
4426 // the last argument to iterate indicates whether the iteration
4427 // should be incremental with periodic yields.
4428 _markBitMap.iterate(&markFromRootsClosure);
4429 // If _restart_addr is non-NULL, a marking stack overflow
4430 // occurred; we need to do a fresh iteration from the
4431 // indicated restart address.
4432 while (_restart_addr != NULL) {
4433 if (_foregroundGCIsActive && asynch) {
4434 // We may be running into repeated stack overflows, having
4435 // reached the limit of the stack size, while making very
4436 // slow forward progress. It may be best to bail out and
4437 // let the foreground collector do its job.
4438 // Clear _restart_addr, so that foreground GC
4439 // works from scratch. This avoids the headache of
4440 // a "rescan" which would otherwise be needed because
4441 // of the dirty mod union table & card table.
4442 _restart_addr = NULL;
4443 return false; // indicating failure to complete marking
4444 }
4445 // Deal with stack overflow:
4446 // we restart marking from _restart_addr
4447 HeapWord* ra = _restart_addr;
4448 markFromRootsClosure.reset(ra);
4449 _restart_addr = NULL;
4450 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4451 }
4452 return true;
4453 }
4454
4455 void CMSCollector::preclean() {
4456 check_correct_thread_executing();
4457 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4458 verify_work_stacks_empty();
4459 verify_overflow_empty();
4460 _abort_preclean = false;
4461 if (CMSPrecleaningEnabled) {
4462 if (!CMSEdenChunksRecordAlways) {
4463 _eden_chunk_index = 0;
4464 }
4465 size_t used = get_eden_used();
4466 size_t capacity = get_eden_capacity();
4467 // Don't start sampling unless we will get sufficiently
4468 // many samples.
4469 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4470 * CMSScheduleRemarkEdenPenetration)) {
4471 _start_sampling = true;
4472 } else {
4473 _start_sampling = false;
4474 }
4475 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4476 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
4477 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4478 }
4479 CMSTokenSync x(true); // is cms thread
4480 if (CMSPrecleaningEnabled) {
4481 sample_eden();
4482 _collectorState = AbortablePreclean;
4483 } else {
4484 _collectorState = FinalMarking;
4485 }
4486 verify_work_stacks_empty();
4487 verify_overflow_empty();
4488 }
4489
4490 // Try and schedule the remark such that young gen
4491 // occupancy is CMSScheduleRemarkEdenPenetration %.
4492 void CMSCollector::abortable_preclean() {
4493 check_correct_thread_executing();
4494 assert(CMSPrecleaningEnabled, "Inconsistent control state");
4495 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4496
4497 // If Eden's current occupancy is below this threshold,
4498 // immediately schedule the remark; else preclean
4499 // past the next scavenge in an effort to
4500 // schedule the pause as described above. By choosing
4501 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4502 // we will never do an actual abortable preclean cycle.
4503 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4504 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4505 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
4506 // We need more smarts in the abortable preclean
4507 // loop below to deal with cases where allocation
4508 // in young gen is very very slow, and our precleaning
4509 // is running a losing race against a horde of
4510 // mutators intent on flooding us with CMS updates
4511 // (dirty cards).
4512 // One, admittedly dumb, strategy is to give up
4513 // after a certain number of abortable precleaning loops
4514 // or after a certain maximum time. We want to make
4515 // this smarter in the next iteration.
4516 // XXX FIX ME!!! YSR
4517 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4518 while (!(should_abort_preclean() ||
4519 ConcurrentMarkSweepThread::should_terminate())) {
4520 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4521 cumworkdone += workdone;
4522 loops++;
4523 // Voluntarily terminate abortable preclean phase if we have
4524 // been at it for too long.
4525 if ((CMSMaxAbortablePrecleanLoops != 0) &&
4526 loops >= CMSMaxAbortablePrecleanLoops) {
4527 if (PrintGCDetails) {
4528 gclog_or_tty->print(" CMS: abort preclean due to loops ");
4529 }
4530 break;
4531 }
4532 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4533 if (PrintGCDetails) {
4534 gclog_or_tty->print(" CMS: abort preclean due to time ");
4535 }
4536 break;
4537 }
4538 // If we are doing little work each iteration, we should
4539 // take a short break.
4540 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4541 // Sleep for some time, waiting for work to accumulate
4542 stopTimer();
4543 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4544 startTimer();
4545 waited++;
4546 }
4547 }
4548 if (PrintCMSStatistics > 0) {
4549 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
4550 loops, waited, cumworkdone);
4551 }
4552 }
4553 CMSTokenSync x(true); // is cms thread
4554 if (_collectorState != Idling) {
4555 assert(_collectorState == AbortablePreclean,
4556 "Spontaneous state transition?");
4557 _collectorState = FinalMarking;
4558 } // Else, a foreground collection completed this CMS cycle.
4559 return;
4560 }
4561
4562 // Respond to an Eden sampling opportunity
4563 void CMSCollector::sample_eden() {
4564 // Make sure a young gc cannot sneak in between our
4565 // reading and recording of a sample.
4566 assert(Thread::current()->is_ConcurrentGC_thread(),
4567 "Only the cms thread may collect Eden samples");
4568 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4569 "Should collect samples while holding CMS token");
4570 if (!_start_sampling) {
4571 return;
4572 }
4573 // When CMSEdenChunksRecordAlways is true, the eden chunk array
4574 // is populated by the young generation.
4575 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
4576 if (_eden_chunk_index < _eden_chunk_capacity) {
4577 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
4578 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4579 "Unexpected state of Eden");
4580 // We'd like to check that what we just sampled is an oop-start address;
4581 // however, we cannot do that here since the object may not yet have been
4582 // initialized. So we'll instead do the check when we _use_ this sample
4583 // later.
4584 if (_eden_chunk_index == 0 ||
4585 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4586 _eden_chunk_array[_eden_chunk_index-1])
4587 >= CMSSamplingGrain)) {
4588 _eden_chunk_index++; // commit sample
4589 }
4590 }
4591 }
4592 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4593 size_t used = get_eden_used();
4594 size_t capacity = get_eden_capacity();
4595 assert(used <= capacity, "Unexpected state of Eden");
4596 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4597 _abort_preclean = true;
4598 }
4599 }
4600 }
4601
4602
4603 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4604 assert(_collectorState == Precleaning ||
4605 _collectorState == AbortablePreclean, "incorrect state");
4606 ResourceMark rm;
4607 HandleMark hm;
4608
4609 // Precleaning is currently not MT but the reference processor
4610 // may be set for MT. Disable it temporarily here.
4611 ReferenceProcessor* rp = ref_processor();
4612 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
4613
4614 // Do one pass of scrubbing the discovered reference lists
4615 // to remove any reference objects with strongly-reachable
4616 // referents.
4617 if (clean_refs) {
4618 CMSPrecleanRefsYieldClosure yield_cl(this);
4619 assert(rp->span().equals(_span), "Spans should be equal");
4620 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4621 &_markStack, true /* preclean */);
4622 CMSDrainMarkingStackClosure complete_trace(this,
4623 _span, &_markBitMap, &_markStack,
4624 &keep_alive, true /* preclean */);
4625
4626 // We don't want this step to interfere with a young
4627 // collection because we don't want to take CPU
4628 // or memory bandwidth away from the young GC threads
4629 // (which may be as many as there are CPUs).
4630 // Note that we don't need to protect ourselves from
4631 // interference with mutators because they can't
4632 // manipulate the discovered reference lists nor affect
4633 // the computed reachability of the referents, the
4634 // only properties manipulated by the precleaning
4635 // of these reference lists.
4636 stopTimer();
4637 CMSTokenSyncWithLocks x(true /* is cms thread */,
4638 bitMapLock());
4639 startTimer();
4640 sample_eden();
4641
4642 // The following will yield to allow foreground
4643 // collection to proceed promptly. XXX YSR:
4644 // The code in this method may need further
4645 // tweaking for better performance and some restructuring
4646 // for cleaner interfaces.
4647 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
4648 rp->preclean_discovered_references(
4649 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
4650 gc_timer, _gc_tracer_cm->gc_id());
4651 }
4652
4653 if (clean_survivor) { // preclean the active survivor space(s)
4654 assert(_young_gen->kind() == Generation::DefNew ||
4655 _young_gen->kind() == Generation::ParNew,
4656 "incorrect type for cast");
4657 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4658 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4659 &_markBitMap, &_modUnionTable,
4660 &_markStack, true /* precleaning phase */);
4661 stopTimer();
4662 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4663 bitMapLock());
4664 startTimer();
4665 unsigned int before_count =
4666 GenCollectedHeap::heap()->total_collections();
4667 SurvivorSpacePrecleanClosure
4668 sss_cl(this, _span, &_markBitMap, &_markStack,
4669 &pam_cl, before_count, CMSYield);
4670 dng->from()->object_iterate_careful(&sss_cl);
4671 dng->to()->object_iterate_careful(&sss_cl);
4672 }
4673 MarkRefsIntoAndScanClosure
4674 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4675 &_markStack, this, CMSYield,
4676 true /* precleaning phase */);
4677 // CAUTION: The following closure has persistent state that may need to
4678 // be reset upon a decrease in the sequence of addresses it
4679 // processes.
4680 ScanMarkedObjectsAgainCarefullyClosure
4681 smoac_cl(this, _span,
4682 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
4683
4684 // Preclean dirty cards in ModUnionTable and CardTable using
4685 // appropriate convergence criterion;
4686 // repeat CMSPrecleanIter times unless we find that
4687 // we are losing.
4688 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4689 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4690 "Bad convergence multiplier");
4691 assert(CMSPrecleanThreshold >= 100,
4692 "Unreasonably low CMSPrecleanThreshold");
4693
4694 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4695 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4696 numIter < CMSPrecleanIter;
4697 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4698 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4699 if (Verbose && PrintGCDetails) {
4700 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
4701 }
4702 // Either there are very few dirty cards, so re-mark
4703 // pause will be small anyway, or our pre-cleaning isn't
4704 // that much faster than the rate at which cards are being
4705 // dirtied, so we might as well stop and re-mark since
4706 // precleaning won't improve our re-mark time by much.
4707 if (curNumCards <= CMSPrecleanThreshold ||
4708 (numIter > 0 &&
4709 (curNumCards * CMSPrecleanDenominator >
4710 lastNumCards * CMSPrecleanNumerator))) {
4711 numIter++;
4712 cumNumCards += curNumCards;
4713 break;
4714 }
4715 }
4716
4717 preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
4718
4719 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4720 cumNumCards += curNumCards;
4721 if (PrintGCDetails && PrintCMSStatistics != 0) {
4722 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
4723 curNumCards, cumNumCards, numIter);
4724 }
4725 return cumNumCards; // as a measure of useful work done
4726 }
4727
4728 // PRECLEANING NOTES:
4729 // Precleaning involves:
4730 // . reading the bits of the modUnionTable and clearing the set bits.
4731 // . For the cards corresponding to the set bits, we scan the
4732 // objects on those cards. This means we need the free_list_lock
4733 // so that we can safely iterate over the CMS space when scanning
4734 // for oops.
4735 // . When we scan the objects, we'll be both reading and setting
4736 // marks in the marking bit map, so we'll need the marking bit map.
4737 // . For protecting _collector_state transitions, we take the CGC_lock.
4738 // Note that any races in the reading of of card table entries by the
4739 // CMS thread on the one hand and the clearing of those entries by the
4740 // VM thread or the setting of those entries by the mutator threads on the
4741 // other are quite benign. However, for efficiency it makes sense to keep
4742 // the VM thread from racing with the CMS thread while the latter is
4743 // dirty card info to the modUnionTable. We therefore also use the
4744 // CGC_lock to protect the reading of the card table and the mod union
4745 // table by the CM thread.
4746 // . We run concurrently with mutator updates, so scanning
4747 // needs to be done carefully -- we should not try to scan
4748 // potentially uninitialized objects.
4749 //
4750 // Locking strategy: While holding the CGC_lock, we scan over and
4751 // reset a maximal dirty range of the mod union / card tables, then lock
4752 // the free_list_lock and bitmap lock to do a full marking, then
4753 // release these locks; and repeat the cycle. This allows for a
4754 // certain amount of fairness in the sharing of these locks between
4755 // the CMS collector on the one hand, and the VM thread and the
4756 // mutators on the other.
4757
4758 // NOTE: preclean_mod_union_table() and preclean_card_table()
4759 // further below are largely identical; if you need to modify
4760 // one of these methods, please check the other method too.
4761
4762 size_t CMSCollector::preclean_mod_union_table(
4763 ConcurrentMarkSweepGeneration* gen,
4764 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4765 verify_work_stacks_empty();
4766 verify_overflow_empty();
4767
4768 // strategy: starting with the first card, accumulate contiguous
4769 // ranges of dirty cards; clear these cards, then scan the region
4770 // covered by these cards.
4771
4772 // Since all of the MUT is committed ahead, we can just use
4773 // that, in case the generations expand while we are precleaning.
4774 // It might also be fine to just use the committed part of the
4775 // generation, but we might potentially miss cards when the
4776 // generation is rapidly expanding while we are in the midst
4777 // of precleaning.
4778 HeapWord* startAddr = gen->reserved().start();
4779 HeapWord* endAddr = gen->reserved().end();
4780
4781 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4782
4783 size_t numDirtyCards, cumNumDirtyCards;
4784 HeapWord *nextAddr, *lastAddr;
4785 for (cumNumDirtyCards = numDirtyCards = 0,
4786 nextAddr = lastAddr = startAddr;
4787 nextAddr < endAddr;
4788 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4789
4790 ResourceMark rm;
4791 HandleMark hm;
4792
4793 MemRegion dirtyRegion;
4794 {
4795 stopTimer();
4796 // Potential yield point
4797 CMSTokenSync ts(true);
4798 startTimer();
4799 sample_eden();
4800 // Get dirty region starting at nextOffset (inclusive),
4801 // simultaneously clearing it.
4802 dirtyRegion =
4803 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4804 assert(dirtyRegion.start() >= nextAddr,
4805 "returned region inconsistent?");
4806 }
4807 // Remember where the next search should begin.
4808 // The returned region (if non-empty) is a right open interval,
4809 // so lastOffset is obtained from the right end of that
4810 // interval.
4811 lastAddr = dirtyRegion.end();
4812 // Should do something more transparent and less hacky XXX
4813 numDirtyCards =
4814 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4815
4816 // We'll scan the cards in the dirty region (with periodic
4817 // yields for foreground GC as needed).
4818 if (!dirtyRegion.is_empty()) {
4819 assert(numDirtyCards > 0, "consistency check");
4820 HeapWord* stop_point = NULL;
4821 stopTimer();
4822 // Potential yield point
4823 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4824 bitMapLock());
4825 startTimer();
4826 {
4827 verify_work_stacks_empty();
4828 verify_overflow_empty();
4829 sample_eden();
4830 stop_point =
4831 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4832 }
4833 if (stop_point != NULL) {
4834 // The careful iteration stopped early either because it found an
4835 // uninitialized object, or because we were in the midst of an
4836 // "abortable preclean", which should now be aborted. Redirty
4837 // the bits corresponding to the partially-scanned or unscanned
4838 // cards. We'll either restart at the next block boundary or
4839 // abort the preclean.
4840 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4841 "Should only be AbortablePreclean.");
4842 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4843 if (should_abort_preclean()) {
4844 break; // out of preclean loop
4845 } else {
4846 // Compute the next address at which preclean should pick up;
4847 // might need bitMapLock in order to read P-bits.
4848 lastAddr = next_card_start_after_block(stop_point);
4849 }
4850 }
4851 } else {
4852 assert(lastAddr == endAddr, "consistency check");
4853 assert(numDirtyCards == 0, "consistency check");
4854 break;
4855 }
4856 }
4857 verify_work_stacks_empty();
4858 verify_overflow_empty();
4859 return cumNumDirtyCards;
4860 }
4861
4862 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4863 // below are largely identical; if you need to modify
4864 // one of these methods, please check the other method too.
4865
4866 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4867 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4868 // strategy: it's similar to precleamModUnionTable above, in that
4869 // we accumulate contiguous ranges of dirty cards, mark these cards
4870 // precleaned, then scan the region covered by these cards.
4871 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4872 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4873
4874 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4875
4876 size_t numDirtyCards, cumNumDirtyCards;
4877 HeapWord *lastAddr, *nextAddr;
4878
4879 for (cumNumDirtyCards = numDirtyCards = 0,
4880 nextAddr = lastAddr = startAddr;
4881 nextAddr < endAddr;
4882 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4883
4884 ResourceMark rm;
4885 HandleMark hm;
4886
4887 MemRegion dirtyRegion;
4888 {
4889 // See comments in "Precleaning notes" above on why we
4890 // do this locking. XXX Could the locking overheads be
4891 // too high when dirty cards are sparse? [I don't think so.]
4892 stopTimer();
4893 CMSTokenSync x(true); // is cms thread
4894 startTimer();
4895 sample_eden();
4896 // Get and clear dirty region from card table
4897 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4898 MemRegion(nextAddr, endAddr),
4899 true,
4900 CardTableModRefBS::precleaned_card_val());
4901
4902 assert(dirtyRegion.start() >= nextAddr,
4903 "returned region inconsistent?");
4904 }
4905 lastAddr = dirtyRegion.end();
4906 numDirtyCards =
4907 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4908
4909 if (!dirtyRegion.is_empty()) {
4910 stopTimer();
4911 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4912 startTimer();
4913 sample_eden();
4914 verify_work_stacks_empty();
4915 verify_overflow_empty();
4916 HeapWord* stop_point =
4917 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4918 if (stop_point != NULL) {
4919 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4920 "Should only be AbortablePreclean.");
4921 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4922 if (should_abort_preclean()) {
4923 break; // out of preclean loop
4924 } else {
4925 // Compute the next address at which preclean should pick up.
4926 lastAddr = next_card_start_after_block(stop_point);
4927 }
4928 }
4929 } else {
4930 break;
4931 }
4932 }
4933 verify_work_stacks_empty();
4934 verify_overflow_empty();
4935 return cumNumDirtyCards;
4936 }
4937
4938 class PrecleanKlassClosure : public KlassClosure {
4939 KlassToOopClosure _cm_klass_closure;
4940 public:
4941 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4942 void do_klass(Klass* k) {
4943 if (k->has_accumulated_modified_oops()) {
4944 k->clear_accumulated_modified_oops();
4945
4946 _cm_klass_closure.do_klass(k);
4947 }
4948 }
4949 };
4950
4951 // The freelist lock is needed to prevent asserts, is it really needed?
4952 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4953
4954 cl->set_freelistLock(freelistLock);
4955
4956 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4957
4958 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4959 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4960 PrecleanKlassClosure preclean_klass_closure(cl);
4961 ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
4962
4963 verify_work_stacks_empty();
4964 verify_overflow_empty();
4965 }
4966
4967 void CMSCollector::checkpointRootsFinal(bool asynch,
4968 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4969 assert(_collectorState == FinalMarking, "incorrect state transition?");
4970 check_correct_thread_executing();
4971 // world is stopped at this checkpoint
4972 assert(SafepointSynchronize::is_at_safepoint(),
4973 "world should be stopped");
4974 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4975
4976 verify_work_stacks_empty();
4977 verify_overflow_empty();
4978
4979 SpecializationStats::clear();
4980 if (PrintGCDetails) {
4981 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4982 _young_gen->used() / K,
4983 _young_gen->capacity() / K);
4984 }
4985 if (asynch) {
4986 if (CMSScavengeBeforeRemark) {
4987 GenCollectedHeap* gch = GenCollectedHeap::heap();
4988 // Temporarily set flag to false, GCH->do_collection will
4989 // expect it to be false and set to true
4990 FlagSetting fl(gch->_is_gc_active, false);
4991 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
4992 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4993 int level = _cmsGen->level() - 1;
4994 if (level >= 0) {
4995 gch->do_collection(true, // full (i.e. force, see below)
4996 false, // !clear_all_soft_refs
4997 0, // size
4998 false, // is_tlab
4999 level // max_level
5000 );
5001 }
5002 }
5003 FreelistLocker x(this);
5004 MutexLockerEx y(bitMapLock(),
5005 Mutex::_no_safepoint_check_flag);
5006 assert(!init_mark_was_synchronous, "but that's impossible!");
5007 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
5008 } else {
5009 // already have all the locks
5010 checkpointRootsFinalWork(asynch, clear_all_soft_refs,
5011 init_mark_was_synchronous);
5012 }
5013 verify_work_stacks_empty();
5014 verify_overflow_empty();
5015 SpecializationStats::print();
5016 }
5017
5018 void CMSCollector::checkpointRootsFinalWork(bool asynch,
5019 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
5020
5021 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
5022
5023 assert(haveFreelistLocks(), "must have free list locks");
5024 assert_lock_strong(bitMapLock());
5025
5026 ResourceMark rm;
5027 HandleMark hm;
5028
5029 GenCollectedHeap* gch = GenCollectedHeap::heap();
5030
5031 if (should_unload_classes()) {
5032 CodeCache::gc_prologue();
5033 }
5034 assert(haveFreelistLocks(), "must have free list locks");
5035 assert_lock_strong(bitMapLock());
5036
5037 if (!init_mark_was_synchronous) {
5038 // We might assume that we need not fill TLAB's when
5039 // CMSScavengeBeforeRemark is set, because we may have just done
5040 // a scavenge which would have filled all TLAB's -- and besides
5041 // Eden would be empty. This however may not always be the case --
5042 // for instance although we asked for a scavenge, it may not have
5043 // happened because of a JNI critical section. We probably need
5044 // a policy for deciding whether we can in that case wait until
5045 // the critical section releases and then do the remark following
5046 // the scavenge, and skip it here. In the absence of that policy,
5047 // or of an indication of whether the scavenge did indeed occur,
5048 // we cannot rely on TLAB's having been filled and must do
5049 // so here just in case a scavenge did not happen.
5050 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
5051 // Update the saved marks which may affect the root scans.
5052 gch->save_marks();
5053
5054 if (CMSPrintEdenSurvivorChunks) {
5055 print_eden_and_survivor_chunk_arrays();
5056 }
5057
5058 {
5059 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
5060
5061 // Note on the role of the mod union table:
5062 // Since the marker in "markFromRoots" marks concurrently with
5063 // mutators, it is possible for some reachable objects not to have been
5064 // scanned. For instance, an only reference to an object A was
5065 // placed in object B after the marker scanned B. Unless B is rescanned,
5066 // A would be collected. Such updates to references in marked objects
5067 // are detected via the mod union table which is the set of all cards
5068 // dirtied since the first checkpoint in this GC cycle and prior to
5069 // the most recent young generation GC, minus those cleaned up by the
5070 // concurrent precleaning.
5071 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
5072 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5073 do_remark_parallel();
5074 } else {
5075 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
5076 _gc_timer_cm, _gc_tracer_cm->gc_id());
5077 do_remark_non_parallel();
5078 }
5079 }
5080 } else {
5081 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
5082 // The initial mark was stop-world, so there's no rescanning to
5083 // do; go straight on to the next step below.
5084 }
5085 verify_work_stacks_empty();
5086 verify_overflow_empty();
5087
5088 {
5089 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
5090 refProcessingWork(asynch, clear_all_soft_refs);
5091 }
5092 verify_work_stacks_empty();
5093 verify_overflow_empty();
5094
5095 if (should_unload_classes()) {
5096 CodeCache::gc_epilogue();
5097 }
5098 JvmtiExport::gc_epilogue();
5099
5100 // If we encountered any (marking stack / work queue) overflow
5101 // events during the current CMS cycle, take appropriate
5102 // remedial measures, where possible, so as to try and avoid
5103 // recurrence of that condition.
5104 assert(_markStack.isEmpty(), "No grey objects");
5105 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
5106 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
5107 if (ser_ovflw > 0) {
5108 if (PrintCMSStatistics != 0) {
5109 gclog_or_tty->print_cr("Marking stack overflow (benign) "
5110 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
5111 ", kac_preclean="SIZE_FORMAT")",
5112 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
5113 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
5114 }
5115 _markStack.expand();
5116 _ser_pmc_remark_ovflw = 0;
5117 _ser_pmc_preclean_ovflw = 0;
5118 _ser_kac_preclean_ovflw = 0;
5119 _ser_kac_ovflw = 0;
5120 }
5121 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
5122 if (PrintCMSStatistics != 0) {
5123 gclog_or_tty->print_cr("Work queue overflow (benign) "
5124 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
5125 _par_pmc_remark_ovflw, _par_kac_ovflw);
5126 }
5127 _par_pmc_remark_ovflw = 0;
5128 _par_kac_ovflw = 0;
5129 }
5130 if (PrintCMSStatistics != 0) {
5131 if (_markStack._hit_limit > 0) {
5132 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
5133 _markStack._hit_limit);
5134 }
5135 if (_markStack._failed_double > 0) {
5136 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
5137 " current capacity "SIZE_FORMAT,
5138 _markStack._failed_double,
5139 _markStack.capacity());
5140 }
5141 }
5142 _markStack._hit_limit = 0;
5143 _markStack._failed_double = 0;
5144
5145 if ((VerifyAfterGC || VerifyDuringGC) &&
5146 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5147 verify_after_remark();
5148 }
5149
5150 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
5151
5152 // Change under the freelistLocks.
5153 _collectorState = Sweeping;
5154 // Call isAllClear() under bitMapLock
5155 assert(_modUnionTable.isAllClear(),
5156 "Should be clear by end of the final marking");
5157 assert(_ct->klass_rem_set()->mod_union_is_clear(),
5158 "Should be clear by end of the final marking");
5159 }
5160
5161 void CMSParInitialMarkTask::work(uint worker_id) {
5162 elapsedTimer _timer;
5163 ResourceMark rm;
5164 HandleMark hm;
5165
5166 // ---------- scan from roots --------------
5167 _timer.start();
5168 GenCollectedHeap* gch = GenCollectedHeap::heap();
5169 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
5170
5171 // ---------- young gen roots --------------
5172 {
5173 work_on_young_gen_roots(worker_id, &par_mri_cl);
5174 _timer.stop();
5175 if (PrintCMSStatistics != 0) {
5176 gclog_or_tty->print_cr(
5177 "Finished young gen initial mark scan work in %dth thread: %3.3f sec",
5178 worker_id, _timer.seconds());
5179 }
5180 }
5181
5182 // ---------- remaining roots --------------
5183 _timer.reset();
5184 _timer.start();
5185
5186 CLDToOopClosure cld_closure(&par_mri_cl, true);
5187
5188 gch->gen_process_roots(_collector->_cmsGen->level(),
5189 false, // yg was scanned above
5190 false, // this is parallel code
5191 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5192 _collector->should_unload_classes(),
5193 &par_mri_cl,
5194 NULL,
5195 &cld_closure);
5196 assert(_collector->should_unload_classes()
5197 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache),
5198 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5199 _timer.stop();
5200 if (PrintCMSStatistics != 0) {
5201 gclog_or_tty->print_cr(
5202 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec",
5203 worker_id, _timer.seconds());
5204 }
5205 }
5206
5207 // Parallel remark task
5208 class CMSParRemarkTask: public CMSParMarkTask {
5209 CompactibleFreeListSpace* _cms_space;
5210
5211 // The per-thread work queues, available here for stealing.
5212 OopTaskQueueSet* _task_queues;
5213 ParallelTaskTerminator _term;
5214
5215 public:
5216 // A value of 0 passed to n_workers will cause the number of
5217 // workers to be taken from the active workers in the work gang.
5218 CMSParRemarkTask(CMSCollector* collector,
5219 CompactibleFreeListSpace* cms_space,
5220 int n_workers, FlexibleWorkGang* workers,
5221 OopTaskQueueSet* task_queues):
5222 CMSParMarkTask("Rescan roots and grey objects in parallel",
5223 collector, n_workers),
5224 _cms_space(cms_space),
5225 _task_queues(task_queues),
5226 _term(n_workers, task_queues) { }
5227
5228 OopTaskQueueSet* task_queues() { return _task_queues; }
5229
5230 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5231
5232 ParallelTaskTerminator* terminator() { return &_term; }
5233 int n_workers() { return _n_workers; }
5234
5235 void work(uint worker_id);
5236
5237 private:
5238 // ... of dirty cards in old space
5239 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
5240 Par_MarkRefsIntoAndScanClosure* cl);
5241
5242 // ... work stealing for the above
5243 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
5244 };
5245
5246 class RemarkKlassClosure : public KlassClosure {
5247 KlassToOopClosure _cm_klass_closure;
5248 public:
5249 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
5250 void do_klass(Klass* k) {
5251 // Check if we have modified any oops in the Klass during the concurrent marking.
5252 if (k->has_accumulated_modified_oops()) {
5253 k->clear_accumulated_modified_oops();
5254
5255 // We could have transfered the current modified marks to the accumulated marks,
5256 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
5257 } else if (k->has_modified_oops()) {
5258 // Don't clear anything, this info is needed by the next young collection.
5259 } else {
5260 // No modified oops in the Klass.
5261 return;
5262 }
5263
5264 // The klass has modified fields, need to scan the klass.
5265 _cm_klass_closure.do_klass(k);
5266 }
5267 };
5268
5269 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
5270 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
5271 EdenSpace* eden_space = dng->eden();
5272 ContiguousSpace* from_space = dng->from();
5273 ContiguousSpace* to_space = dng->to();
5274
5275 HeapWord** eca = _collector->_eden_chunk_array;
5276 size_t ect = _collector->_eden_chunk_index;
5277 HeapWord** sca = _collector->_survivor_chunk_array;
5278 size_t sct = _collector->_survivor_chunk_index;
5279
5280 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
5281 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
5282
5283 do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
5284 do_young_space_rescan(worker_id, cl, from_space, sca, sct);
5285 do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
5286 }
5287
5288 // work_queue(i) is passed to the closure
5289 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
5290 // also is passed to do_dirty_card_rescan_tasks() and to
5291 // do_work_steal() to select the i-th task_queue.
5292
5293 void CMSParRemarkTask::work(uint worker_id) {
5294 elapsedTimer _timer;
5295 ResourceMark rm;
5296 HandleMark hm;
5297
5298 // ---------- rescan from roots --------------
5299 _timer.start();
5300 GenCollectedHeap* gch = GenCollectedHeap::heap();
5301 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
5302 _collector->_span, _collector->ref_processor(),
5303 &(_collector->_markBitMap),
5304 work_queue(worker_id));
5305
5306 // Rescan young gen roots first since these are likely
5307 // coarsely partitioned and may, on that account, constitute
5308 // the critical path; thus, it's best to start off that
5309 // work first.
5310 // ---------- young gen roots --------------
5311 {
5312 work_on_young_gen_roots(worker_id, &par_mrias_cl);
5313 _timer.stop();
5314 if (PrintCMSStatistics != 0) {
5315 gclog_or_tty->print_cr(
5316 "Finished young gen rescan work in %dth thread: %3.3f sec",
5317 worker_id, _timer.seconds());
5318 }
5319 }
5320
5321 // ---------- remaining roots --------------
5322 _timer.reset();
5323 _timer.start();
5324 gch->gen_process_roots(_collector->_cmsGen->level(),
5325 false, // yg was scanned above
5326 false, // this is parallel code
5327 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5328 _collector->should_unload_classes(),
5329 &par_mrias_cl,
5330 NULL,
5331 NULL); // The dirty klasses will be handled below
5332
5333 assert(_collector->should_unload_classes()
5334 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache),
5335 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5336 _timer.stop();
5337 if (PrintCMSStatistics != 0) {
5338 gclog_or_tty->print_cr(
5339 "Finished remaining root rescan work in %dth thread: %3.3f sec",
5340 worker_id, _timer.seconds());
5341 }
5342
5343 // ---------- unhandled CLD scanning ----------
5344 if (worker_id == 0) { // Single threaded at the moment.
5345 _timer.reset();
5346 _timer.start();
5347
5348 // Scan all new class loader data objects and new dependencies that were
5349 // introduced during concurrent marking.
5350 ResourceMark rm;
5351 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5352 for (int i = 0; i < array->length(); i++) {
5353 par_mrias_cl.do_class_loader_data(array->at(i));
5354 }
5355
5356 // We don't need to keep track of new CLDs anymore.
5357 ClassLoaderDataGraph::remember_new_clds(false);
5358
5359 _timer.stop();
5360 if (PrintCMSStatistics != 0) {
5361 gclog_or_tty->print_cr(
5362 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec",
5363 worker_id, _timer.seconds());
5364 }
5365 }
5366
5367 // ---------- dirty klass scanning ----------
5368 if (worker_id == 0) { // Single threaded at the moment.
5369 _timer.reset();
5370 _timer.start();
5371
5372 // Scan all classes that was dirtied during the concurrent marking phase.
5373 RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
5374 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5375
5376 _timer.stop();
5377 if (PrintCMSStatistics != 0) {
5378 gclog_or_tty->print_cr(
5379 "Finished dirty klass scanning work in %dth thread: %3.3f sec",
5380 worker_id, _timer.seconds());
5381 }
5382 }
5383
5384 // We might have added oops to ClassLoaderData::_handles during the
5385 // concurrent marking phase. These oops point to newly allocated objects
5386 // that are guaranteed to be kept alive. Either by the direct allocation
5387 // code, or when the young collector processes the roots. Hence,
5388 // we don't have to revisit the _handles block during the remark phase.
5389
5390 // ---------- rescan dirty cards ------------
5391 _timer.reset();
5392 _timer.start();
5393
5394 // Do the rescan tasks for each of the two spaces
5395 // (cms_space) in turn.
5396 // "worker_id" is passed to select the task_queue for "worker_id"
5397 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
5398 _timer.stop();
5399 if (PrintCMSStatistics != 0) {
5400 gclog_or_tty->print_cr(
5401 "Finished dirty card rescan work in %dth thread: %3.3f sec",
5402 worker_id, _timer.seconds());
5403 }
5404
5405 // ---------- steal work from other threads ...
5406 // ---------- ... and drain overflow list.
5407 _timer.reset();
5408 _timer.start();
5409 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
5410 _timer.stop();
5411 if (PrintCMSStatistics != 0) {
5412 gclog_or_tty->print_cr(
5413 "Finished work stealing in %dth thread: %3.3f sec",
5414 worker_id, _timer.seconds());
5415 }
5416 }
5417
5418 // Note that parameter "i" is not used.
5419 void
5420 CMSParMarkTask::do_young_space_rescan(uint worker_id,
5421 OopsInGenClosure* cl, ContiguousSpace* space,
5422 HeapWord** chunk_array, size_t chunk_top) {
5423 // Until all tasks completed:
5424 // . claim an unclaimed task
5425 // . compute region boundaries corresponding to task claimed
5426 // using chunk_array
5427 // . par_oop_iterate(cl) over that region
5428
5429 ResourceMark rm;
5430 HandleMark hm;
5431
5432 SequentialSubTasksDone* pst = space->par_seq_tasks();
5433
5434 uint nth_task = 0;
5435 uint n_tasks = pst->n_tasks();
5436
5437 if (n_tasks > 0) {
5438 assert(pst->valid(), "Uninitialized use?");
5439 HeapWord *start, *end;
5440 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5441 // We claimed task # nth_task; compute its boundaries.
5442 if (chunk_top == 0) { // no samples were taken
5443 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
5444 start = space->bottom();
5445 end = space->top();
5446 } else if (nth_task == 0) {
5447 start = space->bottom();
5448 end = chunk_array[nth_task];
5449 } else if (nth_task < (uint)chunk_top) {
5450 assert(nth_task >= 1, "Control point invariant");
5451 start = chunk_array[nth_task - 1];
5452 end = chunk_array[nth_task];
5453 } else {
5454 assert(nth_task == (uint)chunk_top, "Control point invariant");
5455 start = chunk_array[chunk_top - 1];
5456 end = space->top();
5457 }
5458 MemRegion mr(start, end);
5459 // Verify that mr is in space
5460 assert(mr.is_empty() || space->used_region().contains(mr),
5461 "Should be in space");
5462 // Verify that "start" is an object boundary
5463 assert(mr.is_empty() || oop(mr.start())->is_oop(),
5464 "Should be an oop");
5465 space->par_oop_iterate(mr, cl);
5466 }
5467 pst->all_tasks_completed();
5468 }
5469 }
5470
5471 void
5472 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5473 CompactibleFreeListSpace* sp, int i,
5474 Par_MarkRefsIntoAndScanClosure* cl) {
5475 // Until all tasks completed:
5476 // . claim an unclaimed task
5477 // . compute region boundaries corresponding to task claimed
5478 // . transfer dirty bits ct->mut for that region
5479 // . apply rescanclosure to dirty mut bits for that region
5480
5481 ResourceMark rm;
5482 HandleMark hm;
5483
5484 OopTaskQueue* work_q = work_queue(i);
5485 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5486 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5487 // CAUTION: This closure has state that persists across calls to
5488 // the work method dirty_range_iterate_clear() in that it has
5489 // embedded in it a (subtype of) UpwardsObjectClosure. The
5490 // use of that state in the embedded UpwardsObjectClosure instance
5491 // assumes that the cards are always iterated (even if in parallel
5492 // by several threads) in monotonically increasing order per each
5493 // thread. This is true of the implementation below which picks
5494 // card ranges (chunks) in monotonically increasing order globally
5495 // and, a-fortiori, in monotonically increasing order per thread
5496 // (the latter order being a subsequence of the former).
5497 // If the work code below is ever reorganized into a more chaotic
5498 // work-partitioning form than the current "sequential tasks"
5499 // paradigm, the use of that persistent state will have to be
5500 // revisited and modified appropriately. See also related
5501 // bug 4756801 work on which should examine this code to make
5502 // sure that the changes there do not run counter to the
5503 // assumptions made here and necessary for correctness and
5504 // efficiency. Note also that this code might yield inefficient
5505 // behavior in the case of very large objects that span one or
5506 // more work chunks. Such objects would potentially be scanned
5507 // several times redundantly. Work on 4756801 should try and
5508 // address that performance anomaly if at all possible. XXX
5509 MemRegion full_span = _collector->_span;
5510 CMSBitMap* bm = &(_collector->_markBitMap); // shared
5511 MarkFromDirtyCardsClosure
5512 greyRescanClosure(_collector, full_span, // entire span of interest
5513 sp, bm, work_q, cl);
5514
5515 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5516 assert(pst->valid(), "Uninitialized use?");
5517 uint nth_task = 0;
5518 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5519 MemRegion span = sp->used_region();
5520 HeapWord* start_addr = span.start();
5521 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5522 alignment);
5523 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5524 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5525 start_addr, "Check alignment");
5526 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5527 chunk_size, "Check alignment");
5528
5529 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5530 // Having claimed the nth_task, compute corresponding mem-region,
5531 // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
5532 // The alignment restriction ensures that we do not need any
5533 // synchronization with other gang-workers while setting or
5534 // clearing bits in thus chunk of the MUT.
5535 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5536 start_addr + (nth_task+1)*chunk_size);
5537 // The last chunk's end might be way beyond end of the
5538 // used region. In that case pull back appropriately.
5539 if (this_span.end() > end_addr) {
5540 this_span.set_end(end_addr);
5541 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5542 }
5543 // Iterate over the dirty cards covering this chunk, marking them
5544 // precleaned, and setting the corresponding bits in the mod union
5545 // table. Since we have been careful to partition at Card and MUT-word
5546 // boundaries no synchronization is needed between parallel threads.
5547 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5548 &modUnionClosure);
5549
5550 // Having transferred these marks into the modUnionTable,
5551 // rescan the marked objects on the dirty cards in the modUnionTable.
5552 // Even if this is at a synchronous collection, the initial marking
5553 // may have been done during an asynchronous collection so there
5554 // may be dirty bits in the mod-union table.
5555 _collector->_modUnionTable.dirty_range_iterate_clear(
5556 this_span, &greyRescanClosure);
5557 _collector->_modUnionTable.verifyNoOneBitsInRange(
5558 this_span.start(),
5559 this_span.end());
5560 }
5561 pst->all_tasks_completed(); // declare that i am done
5562 }
5563
5564 // . see if we can share work_queues with ParNew? XXX
5565 void
5566 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5567 int* seed) {
5568 OopTaskQueue* work_q = work_queue(i);
5569 NOT_PRODUCT(int num_steals = 0;)
5570 oop obj_to_scan;
5571 CMSBitMap* bm = &(_collector->_markBitMap);
5572
5573 while (true) {
5574 // Completely finish any left over work from (an) earlier round(s)
5575 cl->trim_queue(0);
5576 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5577 (size_t)ParGCDesiredObjsFromOverflowList);
5578 // Now check if there's any work in the overflow list
5579 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5580 // only affects the number of attempts made to get work from the
5581 // overflow list and does not affect the number of workers. Just
5582 // pass ParallelGCThreads so this behavior is unchanged.
5583 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5584 work_q,
5585 ParallelGCThreads)) {
5586 // found something in global overflow list;
5587 // not yet ready to go stealing work from others.
5588 // We'd like to assert(work_q->size() != 0, ...)
5589 // because we just took work from the overflow list,
5590 // but of course we can't since all of that could have
5591 // been already stolen from us.
5592 // "He giveth and He taketh away."
5593 continue;
5594 }
5595 // Verify that we have no work before we resort to stealing
5596 assert(work_q->size() == 0, "Have work, shouldn't steal");
5597 // Try to steal from other queues that have work
5598 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5599 NOT_PRODUCT(num_steals++;)
5600 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5601 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5602 // Do scanning work
5603 obj_to_scan->oop_iterate(cl);
5604 // Loop around, finish this work, and try to steal some more
5605 } else if (terminator()->offer_termination()) {
5606 break; // nirvana from the infinite cycle
5607 }
5608 }
5609 NOT_PRODUCT(
5610 if (PrintCMSStatistics != 0) {
5611 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5612 }
5613 )
5614 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5615 "Else our work is not yet done");
5616 }
5617
5618 // Record object boundaries in _eden_chunk_array by sampling the eden
5619 // top in the slow-path eden object allocation code path and record
5620 // the boundaries, if CMSEdenChunksRecordAlways is true. If
5621 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
5622 // sampling in sample_eden() that activates during the part of the
5623 // preclean phase.
5624 void CMSCollector::sample_eden_chunk() {
5625 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
5626 if (_eden_chunk_lock->try_lock()) {
5627 // Record a sample. This is the critical section. The contents
5628 // of the _eden_chunk_array have to be non-decreasing in the
5629 // address order.
5630 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
5631 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
5632 "Unexpected state of Eden");
5633 if (_eden_chunk_index == 0 ||
5634 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
5635 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
5636 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
5637 _eden_chunk_index++; // commit sample
5638 }
5639 _eden_chunk_lock->unlock();
5640 }
5641 }
5642 }
5643
5644 // Return a thread-local PLAB recording array, as appropriate.
5645 void* CMSCollector::get_data_recorder(int thr_num) {
5646 if (_survivor_plab_array != NULL &&
5647 (CMSPLABRecordAlways ||
5648 (_collectorState > Marking && _collectorState < FinalMarking))) {
5649 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5650 ChunkArray* ca = &_survivor_plab_array[thr_num];
5651 ca->reset(); // clear it so that fresh data is recorded
5652 return (void*) ca;
5653 } else {
5654 return NULL;
5655 }
5656 }
5657
5658 // Reset all the thread-local PLAB recording arrays
5659 void CMSCollector::reset_survivor_plab_arrays() {
5660 for (uint i = 0; i < ParallelGCThreads; i++) {
5661 _survivor_plab_array[i].reset();
5662 }
5663 }
5664
5665 // Merge the per-thread plab arrays into the global survivor chunk
5666 // array which will provide the partitioning of the survivor space
5667 // for CMS initial scan and rescan.
5668 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
5669 int no_of_gc_threads) {
5670 assert(_survivor_plab_array != NULL, "Error");
5671 assert(_survivor_chunk_array != NULL, "Error");
5672 assert(_collectorState == FinalMarking ||
5673 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
5674 for (int j = 0; j < no_of_gc_threads; j++) {
5675 _cursor[j] = 0;
5676 }
5677 HeapWord* top = surv->top();
5678 size_t i;
5679 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
5680 HeapWord* min_val = top; // Higher than any PLAB address
5681 uint min_tid = 0; // position of min_val this round
5682 for (int j = 0; j < no_of_gc_threads; j++) {
5683 ChunkArray* cur_sca = &_survivor_plab_array[j];
5684 if (_cursor[j] == cur_sca->end()) {
5685 continue;
5686 }
5687 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5688 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5689 assert(surv->used_region().contains(cur_val), "Out of bounds value");
5690 if (cur_val < min_val) {
5691 min_tid = j;
5692 min_val = cur_val;
5693 } else {
5694 assert(cur_val < top, "All recorded addresses should be less");
5695 }
5696 }
5697 // At this point min_val and min_tid are respectively
5698 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5699 // and the thread (j) that witnesses that address.
5700 // We record this address in the _survivor_chunk_array[i]
5701 // and increment _cursor[min_tid] prior to the next round i.
5702 if (min_val == top) {
5703 break;
5704 }
5705 _survivor_chunk_array[i] = min_val;
5706 _cursor[min_tid]++;
5707 }
5708 // We are all done; record the size of the _survivor_chunk_array
5709 _survivor_chunk_index = i; // exclusive: [0, i)
5710 if (PrintCMSStatistics > 0) {
5711 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5712 }
5713 // Verify that we used up all the recorded entries
5714 #ifdef ASSERT
5715 size_t total = 0;
5716 for (int j = 0; j < no_of_gc_threads; j++) {
5717 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5718 total += _cursor[j];
5719 }
5720 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5721 // Check that the merged array is in sorted order
5722 if (total > 0) {
5723 for (size_t i = 0; i < total - 1; i++) {
5724 if (PrintCMSStatistics > 0) {
5725 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5726 i, _survivor_chunk_array[i]);
5727 }
5728 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5729 "Not sorted");
5730 }
5731 }
5732 #endif // ASSERT
5733 }
5734
5735 // Set up the space's par_seq_tasks structure for work claiming
5736 // for parallel initial scan and rescan of young gen.
5737 // See ParRescanTask where this is currently used.
5738 void
5739 CMSCollector::
5740 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5741 assert(n_threads > 0, "Unexpected n_threads argument");
5742 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5743
5744 // Eden space
5745 if (!dng->eden()->is_empty()) {
5746 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5747 assert(!pst->valid(), "Clobbering existing data?");
5748 // Each valid entry in [0, _eden_chunk_index) represents a task.
5749 size_t n_tasks = _eden_chunk_index + 1;
5750 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5751 // Sets the condition for completion of the subtask (how many threads
5752 // need to finish in order to be done).
5753 pst->set_n_threads(n_threads);
5754 pst->set_n_tasks((int)n_tasks);
5755 }
5756
5757 // Merge the survivor plab arrays into _survivor_chunk_array
5758 if (_survivor_plab_array != NULL) {
5759 merge_survivor_plab_arrays(dng->from(), n_threads);
5760 } else {
5761 assert(_survivor_chunk_index == 0, "Error");
5762 }
5763
5764 // To space
5765 {
5766 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5767 assert(!pst->valid(), "Clobbering existing data?");
5768 // Sets the condition for completion of the subtask (how many threads
5769 // need to finish in order to be done).
5770 pst->set_n_threads(n_threads);
5771 pst->set_n_tasks(1);
5772 assert(pst->valid(), "Error");
5773 }
5774
5775 // From space
5776 {
5777 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5778 assert(!pst->valid(), "Clobbering existing data?");
5779 size_t n_tasks = _survivor_chunk_index + 1;
5780 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5781 // Sets the condition for completion of the subtask (how many threads
5782 // need to finish in order to be done).
5783 pst->set_n_threads(n_threads);
5784 pst->set_n_tasks((int)n_tasks);
5785 assert(pst->valid(), "Error");
5786 }
5787 }
5788
5789 // Parallel version of remark
5790 void CMSCollector::do_remark_parallel() {
5791 GenCollectedHeap* gch = GenCollectedHeap::heap();
5792 FlexibleWorkGang* workers = gch->workers();
5793 assert(workers != NULL, "Need parallel worker threads.");
5794 // Choose to use the number of GC workers most recently set
5795 // into "active_workers". If active_workers is not set, set it
5796 // to ParallelGCThreads.
5797 int n_workers = workers->active_workers();
5798 if (n_workers == 0) {
5799 assert(n_workers > 0, "Should have been set during scavenge");
5800 n_workers = ParallelGCThreads;
5801 workers->set_active_workers(n_workers);
5802 }
5803 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5804
5805 CMSParRemarkTask tsk(this,
5806 cms_space,
5807 n_workers, workers, task_queues());
5808
5809 // Set up for parallel process_roots work.
5810 gch->set_par_threads(n_workers);
5811 // We won't be iterating over the cards in the card table updating
5812 // the younger_gen cards, so we shouldn't call the following else
5813 // the verification code as well as subsequent younger_refs_iterate
5814 // code would get confused. XXX
5815 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5816
5817 // The young gen rescan work will not be done as part of
5818 // process_roots (which currently doesn't know how to
5819 // parallelize such a scan), but rather will be broken up into
5820 // a set of parallel tasks (via the sampling that the [abortable]
5821 // preclean phase did of EdenSpace, plus the [two] tasks of
5822 // scanning the [two] survivor spaces. Further fine-grain
5823 // parallelization of the scanning of the survivor spaces
5824 // themselves, and of precleaning of the younger gen itself
5825 // is deferred to the future.
5826 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5827
5828 // The dirty card rescan work is broken up into a "sequence"
5829 // of parallel tasks (per constituent space) that are dynamically
5830 // claimed by the parallel threads.
5831 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5832
5833 // It turns out that even when we're using 1 thread, doing the work in a
5834 // separate thread causes wide variance in run times. We can't help this
5835 // in the multi-threaded case, but we special-case n=1 here to get
5836 // repeatable measurements of the 1-thread overhead of the parallel code.
5837 if (n_workers > 1) {
5838 // Make refs discovery MT-safe, if it isn't already: it may not
5839 // necessarily be so, since it's possible that we are doing
5840 // ST marking.
5841 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5842 GenCollectedHeap::StrongRootsScope srs(gch);
5843 workers->run_task(&tsk);
5844 } else {
5845 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5846 GenCollectedHeap::StrongRootsScope srs(gch);
5847 tsk.work(0);
5848 }
5849
5850 gch->set_par_threads(0); // 0 ==> non-parallel.
5851 // restore, single-threaded for now, any preserved marks
5852 // as a result of work_q overflow
5853 restore_preserved_marks_if_any();
5854 }
5855
5856 // Non-parallel version of remark
5857 void CMSCollector::do_remark_non_parallel() {
5858 ResourceMark rm;
5859 HandleMark hm;
5860 GenCollectedHeap* gch = GenCollectedHeap::heap();
5861 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5862
5863 MarkRefsIntoAndScanClosure
5864 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
5865 &_markStack, this,
5866 false /* should_yield */, false /* not precleaning */);
5867 MarkFromDirtyCardsClosure
5868 markFromDirtyCardsClosure(this, _span,
5869 NULL, // space is set further below
5870 &_markBitMap, &_markStack, &mrias_cl);
5871 {
5872 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5873 // Iterate over the dirty cards, setting the corresponding bits in the
5874 // mod union table.
5875 {
5876 ModUnionClosure modUnionClosure(&_modUnionTable);
5877 _ct->ct_bs()->dirty_card_iterate(
5878 _cmsGen->used_region(),
5879 &modUnionClosure);
5880 }
5881 // Having transferred these marks into the modUnionTable, we just need
5882 // to rescan the marked objects on the dirty cards in the modUnionTable.
5883 // The initial marking may have been done during an asynchronous
5884 // collection so there may be dirty bits in the mod-union table.
5885 const int alignment =
5886 CardTableModRefBS::card_size * BitsPerWord;
5887 {
5888 // ... First handle dirty cards in CMS gen
5889 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5890 MemRegion ur = _cmsGen->used_region();
5891 HeapWord* lb = ur.start();
5892 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5893 MemRegion cms_span(lb, ub);
5894 _modUnionTable.dirty_range_iterate_clear(cms_span,
5895 &markFromDirtyCardsClosure);
5896 verify_work_stacks_empty();
5897 if (PrintCMSStatistics != 0) {
5898 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5899 markFromDirtyCardsClosure.num_dirty_cards());
5900 }
5901 }
5902 }
5903 if (VerifyDuringGC &&
5904 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5905 HandleMark hm; // Discard invalid handles created during verification
5906 Universe::verify();
5907 }
5908 {
5909 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5910
5911 verify_work_stacks_empty();
5912
5913 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5914 GenCollectedHeap::StrongRootsScope srs(gch);
5915
5916 gch->gen_process_roots(_cmsGen->level(),
5917 true, // younger gens as roots
5918 false, // use the local StrongRootsScope
5919 SharedHeap::ScanningOption(roots_scanning_options()),
5920 should_unload_classes(),
5921 &mrias_cl,
5922 NULL,
5923 NULL); // The dirty klasses will be handled below
5924
5925 assert(should_unload_classes()
5926 || (roots_scanning_options() & SharedHeap::SO_AllCodeCache),
5927 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5928 }
5929
5930 {
5931 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5932
5933 verify_work_stacks_empty();
5934
5935 // Scan all class loader data objects that might have been introduced
5936 // during concurrent marking.
5937 ResourceMark rm;
5938 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5939 for (int i = 0; i < array->length(); i++) {
5940 mrias_cl.do_class_loader_data(array->at(i));
5941 }
5942
5943 // We don't need to keep track of new CLDs anymore.
5944 ClassLoaderDataGraph::remember_new_clds(false);
5945
5946 verify_work_stacks_empty();
5947 }
5948
5949 {
5950 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5951
5952 verify_work_stacks_empty();
5953
5954 RemarkKlassClosure remark_klass_closure(&mrias_cl);
5955 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5956
5957 verify_work_stacks_empty();
5958 }
5959
5960 // We might have added oops to ClassLoaderData::_handles during the
5961 // concurrent marking phase. These oops point to newly allocated objects
5962 // that are guaranteed to be kept alive. Either by the direct allocation
5963 // code, or when the young collector processes the roots. Hence,
5964 // we don't have to revisit the _handles block during the remark phase.
5965
5966 verify_work_stacks_empty();
5967 // Restore evacuated mark words, if any, used for overflow list links
5968 if (!CMSOverflowEarlyRestoration) {
5969 restore_preserved_marks_if_any();
5970 }
5971 verify_overflow_empty();
5972 }
5973
5974 ////////////////////////////////////////////////////////
5975 // Parallel Reference Processing Task Proxy Class
5976 ////////////////////////////////////////////////////////
5977 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5978 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5979 CMSCollector* _collector;
5980 CMSBitMap* _mark_bit_map;
5981 const MemRegion _span;
5982 ProcessTask& _task;
5983
5984 public:
5985 CMSRefProcTaskProxy(ProcessTask& task,
5986 CMSCollector* collector,
5987 const MemRegion& span,
5988 CMSBitMap* mark_bit_map,
5989 AbstractWorkGang* workers,
5990 OopTaskQueueSet* task_queues):
5991 // XXX Should superclass AGTWOQ also know about AWG since it knows
5992 // about the task_queues used by the AWG? Then it could initialize
5993 // the terminator() object. See 6984287. The set_for_termination()
5994 // below is a temporary band-aid for the regression in 6984287.
5995 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5996 task_queues),
5997 _task(task),
5998 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5999 {
6000 assert(_collector->_span.equals(_span) && !_span.is_empty(),
6001 "Inconsistency in _span");
6002 set_for_termination(workers->active_workers());
6003 }
6004
6005 OopTaskQueueSet* task_queues() { return queues(); }
6006
6007 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
6008
6009 void do_work_steal(int i,
6010 CMSParDrainMarkingStackClosure* drain,
6011 CMSParKeepAliveClosure* keep_alive,
6012 int* seed);
6013
6014 virtual void work(uint worker_id);
6015 };
6016
6017 void CMSRefProcTaskProxy::work(uint worker_id) {
6018 ResourceMark rm;
6019 HandleMark hm;
6020 assert(_collector->_span.equals(_span), "Inconsistency in _span");
6021 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
6022 _mark_bit_map,
6023 work_queue(worker_id));
6024 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
6025 _mark_bit_map,
6026 work_queue(worker_id));
6027 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
6028 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
6029 if (_task.marks_oops_alive()) {
6030 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
6031 _collector->hash_seed(worker_id));
6032 }
6033 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
6034 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
6035 }
6036
6037 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
6038 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
6039 EnqueueTask& _task;
6040
6041 public:
6042 CMSRefEnqueueTaskProxy(EnqueueTask& task)
6043 : AbstractGangTask("Enqueue reference objects in parallel"),
6044 _task(task)
6045 { }
6046
6047 virtual void work(uint worker_id)
6048 {
6049 _task.work(worker_id);
6050 }
6051 };
6052
6053 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
6054 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
6055 _span(span),
6056 _bit_map(bit_map),
6057 _work_queue(work_queue),
6058 _mark_and_push(collector, span, bit_map, work_queue),
6059 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6060 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
6061 { }
6062
6063 // . see if we can share work_queues with ParNew? XXX
6064 void CMSRefProcTaskProxy::do_work_steal(int i,
6065 CMSParDrainMarkingStackClosure* drain,
6066 CMSParKeepAliveClosure* keep_alive,
6067 int* seed) {
6068 OopTaskQueue* work_q = work_queue(i);
6069 NOT_PRODUCT(int num_steals = 0;)
6070 oop obj_to_scan;
6071
6072 while (true) {
6073 // Completely finish any left over work from (an) earlier round(s)
6074 drain->trim_queue(0);
6075 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
6076 (size_t)ParGCDesiredObjsFromOverflowList);
6077 // Now check if there's any work in the overflow list
6078 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
6079 // only affects the number of attempts made to get work from the
6080 // overflow list and does not affect the number of workers. Just
6081 // pass ParallelGCThreads so this behavior is unchanged.
6082 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
6083 work_q,
6084 ParallelGCThreads)) {
6085 // Found something in global overflow list;
6086 // not yet ready to go stealing work from others.
6087 // We'd like to assert(work_q->size() != 0, ...)
6088 // because we just took work from the overflow list,
6089 // but of course we can't, since all of that might have
6090 // been already stolen from us.
6091 continue;
6092 }
6093 // Verify that we have no work before we resort to stealing
6094 assert(work_q->size() == 0, "Have work, shouldn't steal");
6095 // Try to steal from other queues that have work
6096 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
6097 NOT_PRODUCT(num_steals++;)
6098 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
6099 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
6100 // Do scanning work
6101 obj_to_scan->oop_iterate(keep_alive);
6102 // Loop around, finish this work, and try to steal some more
6103 } else if (terminator()->offer_termination()) {
6104 break; // nirvana from the infinite cycle
6105 }
6106 }
6107 NOT_PRODUCT(
6108 if (PrintCMSStatistics != 0) {
6109 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
6110 }
6111 )
6112 }
6113
6114 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
6115 {
6116 GenCollectedHeap* gch = GenCollectedHeap::heap();
6117 FlexibleWorkGang* workers = gch->workers();
6118 assert(workers != NULL, "Need parallel worker threads.");
6119 CMSRefProcTaskProxy rp_task(task, &_collector,
6120 _collector.ref_processor()->span(),
6121 _collector.markBitMap(),
6122 workers, _collector.task_queues());
6123 workers->run_task(&rp_task);
6124 }
6125
6126 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
6127 {
6128
6129 GenCollectedHeap* gch = GenCollectedHeap::heap();
6130 FlexibleWorkGang* workers = gch->workers();
6131 assert(workers != NULL, "Need parallel worker threads.");
6132 CMSRefEnqueueTaskProxy enq_task(task);
6133 workers->run_task(&enq_task);
6134 }
6135
6136 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
6137
6138 ResourceMark rm;
6139 HandleMark hm;
6140
6141 ReferenceProcessor* rp = ref_processor();
6142 assert(rp->span().equals(_span), "Spans should be equal");
6143 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
6144 // Process weak references.
6145 rp->setup_policy(clear_all_soft_refs);
6146 verify_work_stacks_empty();
6147
6148 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
6149 &_markStack, false /* !preclean */);
6150 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
6151 _span, &_markBitMap, &_markStack,
6152 &cmsKeepAliveClosure, false /* !preclean */);
6153 {
6154 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
6155
6156 ReferenceProcessorStats stats;
6157 if (rp->processing_is_mt()) {
6158 // Set the degree of MT here. If the discovery is done MT, there
6159 // may have been a different number of threads doing the discovery
6160 // and a different number of discovered lists may have Ref objects.
6161 // That is OK as long as the Reference lists are balanced (see
6162 // balance_all_queues() and balance_queues()).
6163 GenCollectedHeap* gch = GenCollectedHeap::heap();
6164 int active_workers = ParallelGCThreads;
6165 FlexibleWorkGang* workers = gch->workers();
6166 if (workers != NULL) {
6167 active_workers = workers->active_workers();
6168 // The expectation is that active_workers will have already
6169 // been set to a reasonable value. If it has not been set,
6170 // investigate.
6171 assert(active_workers > 0, "Should have been set during scavenge");
6172 }
6173 rp->set_active_mt_degree(active_workers);
6174 CMSRefProcTaskExecutor task_executor(*this);
6175 stats = rp->process_discovered_references(&_is_alive_closure,
6176 &cmsKeepAliveClosure,
6177 &cmsDrainMarkingStackClosure,
6178 &task_executor,
6179 _gc_timer_cm,
6180 _gc_tracer_cm->gc_id());
6181 } else {
6182 stats = rp->process_discovered_references(&_is_alive_closure,
6183 &cmsKeepAliveClosure,
6184 &cmsDrainMarkingStackClosure,
6185 NULL,
6186 _gc_timer_cm,
6187 _gc_tracer_cm->gc_id());
6188 }
6189 _gc_tracer_cm->report_gc_reference_stats(stats);
6190
6191 }
6192
6193 // This is the point where the entire marking should have completed.
6194 verify_work_stacks_empty();
6195
6196 if (should_unload_classes()) {
6197 {
6198 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
6199
6200 // Unload classes and purge the SystemDictionary.
6201 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
6202
6203 // Unload nmethods.
6204 CodeCache::do_unloading(&_is_alive_closure, purged_class);
6205
6206 // Prune dead klasses from subklass/sibling/implementor lists.
6207 Klass::clean_weak_klass_links(&_is_alive_closure);
6208 }
6209
6210 {
6211 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
6212 // Clean up unreferenced symbols in symbol table.
6213 SymbolTable::unlink();
6214 }
6215
6216 {
6217 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
6218 // Delete entries for dead interned strings.
6219 StringTable::unlink(&_is_alive_closure);
6220 }
6221 }
6222
6223
6224 // Restore any preserved marks as a result of mark stack or
6225 // work queue overflow
6226 restore_preserved_marks_if_any(); // done single-threaded for now
6227
6228 rp->set_enqueuing_is_done(true);
6229 if (rp->processing_is_mt()) {
6230 rp->balance_all_queues();
6231 CMSRefProcTaskExecutor task_executor(*this);
6232 rp->enqueue_discovered_references(&task_executor);
6233 } else {
6234 rp->enqueue_discovered_references(NULL);
6235 }
6236 rp->verify_no_references_recorded();
6237 assert(!rp->discovery_enabled(), "should have been disabled");
6238 }
6239
6240 #ifndef PRODUCT
6241 void CMSCollector::check_correct_thread_executing() {
6242 Thread* t = Thread::current();
6243 // Only the VM thread or the CMS thread should be here.
6244 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
6245 "Unexpected thread type");
6246 // If this is the vm thread, the foreground process
6247 // should not be waiting. Note that _foregroundGCIsActive is
6248 // true while the foreground collector is waiting.
6249 if (_foregroundGCShouldWait) {
6250 // We cannot be the VM thread
6251 assert(t->is_ConcurrentGC_thread(),
6252 "Should be CMS thread");
6253 } else {
6254 // We can be the CMS thread only if we are in a stop-world
6255 // phase of CMS collection.
6256 if (t->is_ConcurrentGC_thread()) {
6257 assert(_collectorState == InitialMarking ||
6258 _collectorState == FinalMarking,
6259 "Should be a stop-world phase");
6260 // The CMS thread should be holding the CMS_token.
6261 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6262 "Potential interference with concurrently "
6263 "executing VM thread");
6264 }
6265 }
6266 }
6267 #endif
6268
6269 void CMSCollector::sweep(bool asynch) {
6270 assert(_collectorState == Sweeping, "just checking");
6271 check_correct_thread_executing();
6272 verify_work_stacks_empty();
6273 verify_overflow_empty();
6274 increment_sweep_count();
6275 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
6276
6277 _inter_sweep_timer.stop();
6278 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
6279
6280 assert(!_intra_sweep_timer.is_active(), "Should not be active");
6281 _intra_sweep_timer.reset();
6282 _intra_sweep_timer.start();
6283 if (asynch) {
6284 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6285 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails);
6286 // First sweep the old gen
6287 {
6288 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6289 bitMapLock());
6290 sweepWork(_cmsGen, asynch);
6291 }
6292
6293 // Update Universe::_heap_*_at_gc figures.
6294 // We need all the free list locks to make the abstract state
6295 // transition from Sweeping to Resetting. See detailed note
6296 // further below.
6297 {
6298 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
6299 // Update heap occupancy information which is used as
6300 // input to soft ref clearing policy at the next gc.
6301 Universe::update_heap_info_at_gc();
6302 _collectorState = Resizing;
6303 }
6304 } else {
6305 // already have needed locks
6306 sweepWork(_cmsGen, asynch);
6307 // Update heap occupancy information which is used as
6308 // input to soft ref clearing policy at the next gc.
6309 Universe::update_heap_info_at_gc();
6310 _collectorState = Resizing;
6311 }
6312 verify_work_stacks_empty();
6313 verify_overflow_empty();
6314
6315 if (should_unload_classes()) {
6316 // Delay purge to the beginning of the next safepoint. Metaspace::contains
6317 // requires that the virtual spaces are stable and not deleted.
6318 ClassLoaderDataGraph::set_should_purge(true);
6319 }
6320
6321 _intra_sweep_timer.stop();
6322 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
6323
6324 _inter_sweep_timer.reset();
6325 _inter_sweep_timer.start();
6326
6327 // We need to use a monotonically non-decreasing time in ms
6328 // or we will see time-warp warnings and os::javaTimeMillis()
6329 // does not guarantee monotonicity.
6330 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
6331 update_time_of_last_gc(now);
6332
6333 // NOTE on abstract state transitions:
6334 // Mutators allocate-live and/or mark the mod-union table dirty
6335 // based on the state of the collection. The former is done in
6336 // the interval [Marking, Sweeping] and the latter in the interval
6337 // [Marking, Sweeping). Thus the transitions into the Marking state
6338 // and out of the Sweeping state must be synchronously visible
6339 // globally to the mutators.
6340 // The transition into the Marking state happens with the world
6341 // stopped so the mutators will globally see it. Sweeping is
6342 // done asynchronously by the background collector so the transition
6343 // from the Sweeping state to the Resizing state must be done
6344 // under the freelistLock (as is the check for whether to
6345 // allocate-live and whether to dirty the mod-union table).
6346 assert(_collectorState == Resizing, "Change of collector state to"
6347 " Resizing must be done under the freelistLocks (plural)");
6348
6349 // Now that sweeping has been completed, we clear
6350 // the incremental_collection_failed flag,
6351 // thus inviting a younger gen collection to promote into
6352 // this generation. If such a promotion may still fail,
6353 // the flag will be set again when a young collection is
6354 // attempted.
6355 GenCollectedHeap* gch = GenCollectedHeap::heap();
6356 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
6357 gch->update_full_collections_completed(_collection_count_start);
6358 }
6359
6360 // FIX ME!!! Looks like this belongs in CFLSpace, with
6361 // CMSGen merely delegating to it.
6362 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
6363 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
6364 HeapWord* minAddr = _cmsSpace->bottom();
6365 HeapWord* largestAddr =
6366 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
6367 if (largestAddr == NULL) {
6368 // The dictionary appears to be empty. In this case
6369 // try to coalesce at the end of the heap.
6370 largestAddr = _cmsSpace->end();
6371 }
6372 size_t largestOffset = pointer_delta(largestAddr, minAddr);
6373 size_t nearLargestOffset =
6374 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
6375 if (PrintFLSStatistics != 0) {
6376 gclog_or_tty->print_cr(
6377 "CMS: Large Block: " PTR_FORMAT ";"
6378 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
6379 largestAddr,
6380 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset);
6381 }
6382 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
6383 }
6384
6385 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
6386 return addr >= _cmsSpace->nearLargestChunk();
6387 }
6388
6389 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
6390 return _cmsSpace->find_chunk_at_end();
6391 }
6392
6393 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
6394 bool full) {
6395 // The next lower level has been collected. Gather any statistics
6396 // that are of interest at this point.
6397 if (!full && (current_level + 1) == level()) {
6398 // Gather statistics on the young generation collection.
6399 collector()->stats().record_gc0_end(used());
6400 }
6401 }
6402
6403 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
6404 if (PrintGCDetails && Verbose) {
6405 gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
6406 }
6407 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
6408 _debug_collection_type =
6409 (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
6410 if (PrintGCDetails && Verbose) {
6411 gclog_or_tty->print_cr("to %d ", _debug_collection_type);
6412 }
6413 }
6414
6415 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
6416 bool asynch) {
6417 // We iterate over the space(s) underlying this generation,
6418 // checking the mark bit map to see if the bits corresponding
6419 // to specific blocks are marked or not. Blocks that are
6420 // marked are live and are not swept up. All remaining blocks
6421 // are swept up, with coalescing on-the-fly as we sweep up
6422 // contiguous free and/or garbage blocks:
6423 // We need to ensure that the sweeper synchronizes with allocators
6424 // and stop-the-world collectors. In particular, the following
6425 // locks are used:
6426 // . CMS token: if this is held, a stop the world collection cannot occur
6427 // . freelistLock: if this is held no allocation can occur from this
6428 // generation by another thread
6429 // . bitMapLock: if this is held, no other thread can access or update
6430 //
6431
6432 // Note that we need to hold the freelistLock if we use
6433 // block iterate below; else the iterator might go awry if
6434 // a mutator (or promotion) causes block contents to change
6435 // (for instance if the allocator divvies up a block).
6436 // If we hold the free list lock, for all practical purposes
6437 // young generation GC's can't occur (they'll usually need to
6438 // promote), so we might as well prevent all young generation
6439 // GC's while we do a sweeping step. For the same reason, we might
6440 // as well take the bit map lock for the entire duration
6441
6442 // check that we hold the requisite locks
6443 assert(have_cms_token(), "Should hold cms token");
6444 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
6445 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
6446 "Should possess CMS token to sweep");
6447 assert_lock_strong(gen->freelistLock());
6448 assert_lock_strong(bitMapLock());
6449
6450 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
6451 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
6452 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
6453 _inter_sweep_estimate.padded_average(),
6454 _intra_sweep_estimate.padded_average());
6455 gen->setNearLargestChunk();
6456
6457 {
6458 SweepClosure sweepClosure(this, gen, &_markBitMap,
6459 CMSYield && asynch);
6460 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
6461 // We need to free-up/coalesce garbage/blocks from a
6462 // co-terminal free run. This is done in the SweepClosure
6463 // destructor; so, do not remove this scope, else the
6464 // end-of-sweep-census below will be off by a little bit.
6465 }
6466 gen->cmsSpace()->sweep_completed();
6467 gen->cmsSpace()->endSweepFLCensus(sweep_count());
6468 if (should_unload_classes()) { // unloaded classes this cycle,
6469 _concurrent_cycles_since_last_unload = 0; // ... reset count
6470 } else { // did not unload classes,
6471 _concurrent_cycles_since_last_unload++; // ... increment count
6472 }
6473 }
6474
6475 // Reset CMS data structures (for now just the marking bit map)
6476 // preparatory for the next cycle.
6477 void CMSCollector::reset(bool asynch) {
6478 if (asynch) {
6479 CMSTokenSyncWithLocks ts(true, bitMapLock());
6480
6481 // If the state is not "Resetting", the foreground thread
6482 // has done a collection and the resetting.
6483 if (_collectorState != Resetting) {
6484 assert(_collectorState == Idling, "The state should only change"
6485 " because the foreground collector has finished the collection");
6486 return;
6487 }
6488
6489 // Clear the mark bitmap (no grey objects to start with)
6490 // for the next cycle.
6491 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6492 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails);
6493
6494 HeapWord* curAddr = _markBitMap.startWord();
6495 while (curAddr < _markBitMap.endWord()) {
6496 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
6497 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
6498 _markBitMap.clear_large_range(chunk);
6499 if (ConcurrentMarkSweepThread::should_yield() &&
6500 !foregroundGCIsActive() &&
6501 CMSYield) {
6502 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6503 "CMS thread should hold CMS token");
6504 assert_lock_strong(bitMapLock());
6505 bitMapLock()->unlock();
6506 ConcurrentMarkSweepThread::desynchronize(true);
6507 ConcurrentMarkSweepThread::acknowledge_yield_request();
6508 stopTimer();
6509 if (PrintCMSStatistics != 0) {
6510 incrementYields();
6511 }
6512 icms_wait();
6513
6514 // See the comment in coordinator_yield()
6515 for (unsigned i = 0; i < CMSYieldSleepCount &&
6516 ConcurrentMarkSweepThread::should_yield() &&
6517 !CMSCollector::foregroundGCIsActive(); ++i) {
6518 os::sleep(Thread::current(), 1, false);
6519 ConcurrentMarkSweepThread::acknowledge_yield_request();
6520 }
6521
6522 ConcurrentMarkSweepThread::synchronize(true);
6523 bitMapLock()->lock_without_safepoint_check();
6524 startTimer();
6525 }
6526 curAddr = chunk.end();
6527 }
6528 // A successful mostly concurrent collection has been done.
6529 // Because only the full (i.e., concurrent mode failure) collections
6530 // are being measured for gc overhead limits, clean the "near" flag
6531 // and count.
6532 size_policy()->reset_gc_overhead_limit_count();
6533 _collectorState = Idling;
6534 } else {
6535 // already have the lock
6536 assert(_collectorState == Resetting, "just checking");
6537 assert_lock_strong(bitMapLock());
6538 _markBitMap.clear_all();
6539 _collectorState = Idling;
6540 }
6541
6542 // Stop incremental mode after a cycle completes, so that any future cycles
6543 // are triggered by allocation.
6544 stop_icms();
6545
6546 NOT_PRODUCT(
6547 if (RotateCMSCollectionTypes) {
6548 _cmsGen->rotate_debug_collection_type();
6549 }
6550 )
6551
6552 register_gc_end();
6553 }
6554
6555 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
6556 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
6557 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6558 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id());
6559 TraceCollectorStats tcs(counters());
6560
6561 switch (op) {
6562 case CMS_op_checkpointRootsInitial: {
6563 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6564 checkpointRootsInitial(true); // asynch
6565 if (PrintGC) {
6566 _cmsGen->printOccupancy("initial-mark");
6567 }
6568 break;
6569 }
6570 case CMS_op_checkpointRootsFinal: {
6571 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6572 checkpointRootsFinal(true, // asynch
6573 false, // !clear_all_soft_refs
6574 false); // !init_mark_was_synchronous
6575 if (PrintGC) {
6576 _cmsGen->printOccupancy("remark");
6577 }
6578 break;
6579 }
6580 default:
6581 fatal("No such CMS_op");
6582 }
6583 }
6584
6585 #ifndef PRODUCT
6586 size_t const CMSCollector::skip_header_HeapWords() {
6587 return FreeChunk::header_size();
6588 }
6589
6590 // Try and collect here conditions that should hold when
6591 // CMS thread is exiting. The idea is that the foreground GC
6592 // thread should not be blocked if it wants to terminate
6593 // the CMS thread and yet continue to run the VM for a while
6594 // after that.
6595 void CMSCollector::verify_ok_to_terminate() const {
6596 assert(Thread::current()->is_ConcurrentGC_thread(),
6597 "should be called by CMS thread");
6598 assert(!_foregroundGCShouldWait, "should be false");
6599 // We could check here that all the various low-level locks
6600 // are not held by the CMS thread, but that is overkill; see
6601 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6602 // is checked.
6603 }
6604 #endif
6605
6606 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6607 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6608 "missing Printezis mark?");
6609 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6610 size_t size = pointer_delta(nextOneAddr + 1, addr);
6611 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6612 "alignment problem");
6613 assert(size >= 3, "Necessary for Printezis marks to work");
6614 return size;
6615 }
6616
6617 // A variant of the above (block_size_using_printezis_bits()) except
6618 // that we return 0 if the P-bits are not yet set.
6619 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6620 if (_markBitMap.isMarked(addr + 1)) {
6621 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
6622 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6623 size_t size = pointer_delta(nextOneAddr + 1, addr);
6624 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6625 "alignment problem");
6626 assert(size >= 3, "Necessary for Printezis marks to work");
6627 return size;
6628 }
6629 return 0;
6630 }
6631
6632 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6633 size_t sz = 0;
6634 oop p = (oop)addr;
6635 if (p->klass_or_null() != NULL) {
6636 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6637 } else {
6638 sz = block_size_using_printezis_bits(addr);
6639 }
6640 assert(sz > 0, "size must be nonzero");
6641 HeapWord* next_block = addr + sz;
6642 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
6643 CardTableModRefBS::card_size);
6644 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
6645 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6646 "must be different cards");
6647 return next_card;
6648 }
6649
6650
6651 // CMS Bit Map Wrapper /////////////////////////////////////////
6652
6653 // Construct a CMS bit map infrastructure, but don't create the
6654 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6655 // further below.
6656 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6657 _bm(),
6658 _shifter(shifter),
6659 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6660 {
6661 _bmStartWord = 0;
6662 _bmWordSize = 0;
6663 }
6664
6665 bool CMSBitMap::allocate(MemRegion mr) {
6666 _bmStartWord = mr.start();
6667 _bmWordSize = mr.word_size();
6668 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6669 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6670 if (!brs.is_reserved()) {
6671 warning("CMS bit map allocation failure");
6672 return false;
6673 }
6674 // For now we'll just commit all of the bit map up front.
6675 // Later on we'll try to be more parsimonious with swap.
6676 if (!_virtual_space.initialize(brs, brs.size())) {
6677 warning("CMS bit map backing store failure");
6678 return false;
6679 }
6680 assert(_virtual_space.committed_size() == brs.size(),
6681 "didn't reserve backing store for all of CMS bit map?");
6682 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
6683 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6684 _bmWordSize, "inconsistency in bit map sizing");
6685 _bm.set_size(_bmWordSize >> _shifter);
6686
6687 // bm.clear(); // can we rely on getting zero'd memory? verify below
6688 assert(isAllClear(),
6689 "Expected zero'd memory from ReservedSpace constructor");
6690 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6691 "consistency check");
6692 return true;
6693 }
6694
6695 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6696 HeapWord *next_addr, *end_addr, *last_addr;
6697 assert_locked();
6698 assert(covers(mr), "out-of-range error");
6699 // XXX assert that start and end are appropriately aligned
6700 for (next_addr = mr.start(), end_addr = mr.end();
6701 next_addr < end_addr; next_addr = last_addr) {
6702 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6703 last_addr = dirty_region.end();
6704 if (!dirty_region.is_empty()) {
6705 cl->do_MemRegion(dirty_region);
6706 } else {
6707 assert(last_addr == end_addr, "program logic");
6708 return;
6709 }
6710 }
6711 }
6712
6713 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
6714 _bm.print_on_error(st, prefix);
6715 }
6716
6717 #ifndef PRODUCT
6718 void CMSBitMap::assert_locked() const {
6719 CMSLockVerifier::assert_locked(lock());
6720 }
6721
6722 bool CMSBitMap::covers(MemRegion mr) const {
6723 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6724 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6725 "size inconsistency");
6726 return (mr.start() >= _bmStartWord) &&
6727 (mr.end() <= endWord());
6728 }
6729
6730 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6731 return (start >= _bmStartWord && (start + size) <= endWord());
6732 }
6733
6734 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6735 // verify that there are no 1 bits in the interval [left, right)
6736 FalseBitMapClosure falseBitMapClosure;
6737 iterate(&falseBitMapClosure, left, right);
6738 }
6739
6740 void CMSBitMap::region_invariant(MemRegion mr)
6741 {
6742 assert_locked();
6743 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6744 assert(!mr.is_empty(), "unexpected empty region");
6745 assert(covers(mr), "mr should be covered by bit map");
6746 // convert address range into offset range
6747 size_t start_ofs = heapWordToOffset(mr.start());
6748 // Make sure that end() is appropriately aligned
6749 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6750 (1 << (_shifter+LogHeapWordSize))),
6751 "Misaligned mr.end()");
6752 size_t end_ofs = heapWordToOffset(mr.end());
6753 assert(end_ofs > start_ofs, "Should mark at least one bit");
6754 }
6755
6756 #endif
6757
6758 bool CMSMarkStack::allocate(size_t size) {
6759 // allocate a stack of the requisite depth
6760 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6761 size * sizeof(oop)));
6762 if (!rs.is_reserved()) {
6763 warning("CMSMarkStack allocation failure");
6764 return false;
6765 }
6766 if (!_virtual_space.initialize(rs, rs.size())) {
6767 warning("CMSMarkStack backing store failure");
6768 return false;
6769 }
6770 assert(_virtual_space.committed_size() == rs.size(),
6771 "didn't reserve backing store for all of CMS stack?");
6772 _base = (oop*)(_virtual_space.low());
6773 _index = 0;
6774 _capacity = size;
6775 NOT_PRODUCT(_max_depth = 0);
6776 return true;
6777 }
6778
6779 // XXX FIX ME !!! In the MT case we come in here holding a
6780 // leaf lock. For printing we need to take a further lock
6781 // which has lower rank. We need to recalibrate the two
6782 // lock-ranks involved in order to be able to print the
6783 // messages below. (Or defer the printing to the caller.
6784 // For now we take the expedient path of just disabling the
6785 // messages for the problematic case.)
6786 void CMSMarkStack::expand() {
6787 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6788 if (_capacity == MarkStackSizeMax) {
6789 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6790 // We print a warning message only once per CMS cycle.
6791 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6792 }
6793 return;
6794 }
6795 // Double capacity if possible
6796 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6797 // Do not give up existing stack until we have managed to
6798 // get the double capacity that we desired.
6799 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6800 new_capacity * sizeof(oop)));
6801 if (rs.is_reserved()) {
6802 // Release the backing store associated with old stack
6803 _virtual_space.release();
6804 // Reinitialize virtual space for new stack
6805 if (!_virtual_space.initialize(rs, rs.size())) {
6806 fatal("Not enough swap for expanded marking stack");
6807 }
6808 _base = (oop*)(_virtual_space.low());
6809 _index = 0;
6810 _capacity = new_capacity;
6811 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6812 // Failed to double capacity, continue;
6813 // we print a detail message only once per CMS cycle.
6814 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6815 SIZE_FORMAT"K",
6816 _capacity / K, new_capacity / K);
6817 }
6818 }
6819
6820
6821 // Closures
6822 // XXX: there seems to be a lot of code duplication here;
6823 // should refactor and consolidate common code.
6824
6825 // This closure is used to mark refs into the CMS generation in
6826 // the CMS bit map. Called at the first checkpoint. This closure
6827 // assumes that we do not need to re-mark dirty cards; if the CMS
6828 // generation on which this is used is not an oldest
6829 // generation then this will lose younger_gen cards!
6830
6831 MarkRefsIntoClosure::MarkRefsIntoClosure(
6832 MemRegion span, CMSBitMap* bitMap):
6833 _span(span),
6834 _bitMap(bitMap)
6835 {
6836 assert(_ref_processor == NULL, "deliberately left NULL");
6837 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6838 }
6839
6840 void MarkRefsIntoClosure::do_oop(oop obj) {
6841 // if p points into _span, then mark corresponding bit in _markBitMap
6842 assert(obj->is_oop(), "expected an oop");
6843 HeapWord* addr = (HeapWord*)obj;
6844 if (_span.contains(addr)) {
6845 // this should be made more efficient
6846 _bitMap->mark(addr);
6847 }
6848 }
6849
6850 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6851 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6852
6853 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
6854 MemRegion span, CMSBitMap* bitMap):
6855 _span(span),
6856 _bitMap(bitMap)
6857 {
6858 assert(_ref_processor == NULL, "deliberately left NULL");
6859 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6860 }
6861
6862 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
6863 // if p points into _span, then mark corresponding bit in _markBitMap
6864 assert(obj->is_oop(), "expected an oop");
6865 HeapWord* addr = (HeapWord*)obj;
6866 if (_span.contains(addr)) {
6867 // this should be made more efficient
6868 _bitMap->par_mark(addr);
6869 }
6870 }
6871
6872 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6873 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6874
6875 // A variant of the above, used for CMS marking verification.
6876 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6877 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6878 _span(span),
6879 _verification_bm(verification_bm),
6880 _cms_bm(cms_bm)
6881 {
6882 assert(_ref_processor == NULL, "deliberately left NULL");
6883 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6884 }
6885
6886 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6887 // if p points into _span, then mark corresponding bit in _markBitMap
6888 assert(obj->is_oop(), "expected an oop");
6889 HeapWord* addr = (HeapWord*)obj;
6890 if (_span.contains(addr)) {
6891 _verification_bm->mark(addr);
6892 if (!_cms_bm->isMarked(addr)) {
6893 oop(addr)->print();
6894 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6895 fatal("... aborting");
6896 }
6897 }
6898 }
6899
6900 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6901 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6902
6903 //////////////////////////////////////////////////
6904 // MarkRefsIntoAndScanClosure
6905 //////////////////////////////////////////////////
6906
6907 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6908 ReferenceProcessor* rp,
6909 CMSBitMap* bit_map,
6910 CMSBitMap* mod_union_table,
6911 CMSMarkStack* mark_stack,
6912 CMSCollector* collector,
6913 bool should_yield,
6914 bool concurrent_precleaning):
6915 _collector(collector),
6916 _span(span),
6917 _bit_map(bit_map),
6918 _mark_stack(mark_stack),
6919 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6920 mark_stack, concurrent_precleaning),
6921 _yield(should_yield),
6922 _concurrent_precleaning(concurrent_precleaning),
6923 _freelistLock(NULL)
6924 {
6925 _ref_processor = rp;
6926 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6927 }
6928
6929 // This closure is used to mark refs into the CMS generation at the
6930 // second (final) checkpoint, and to scan and transitively follow
6931 // the unmarked oops. It is also used during the concurrent precleaning
6932 // phase while scanning objects on dirty cards in the CMS generation.
6933 // The marks are made in the marking bit map and the marking stack is
6934 // used for keeping the (newly) grey objects during the scan.
6935 // The parallel version (Par_...) appears further below.
6936 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6937 if (obj != NULL) {
6938 assert(obj->is_oop(), "expected an oop");
6939 HeapWord* addr = (HeapWord*)obj;
6940 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6941 assert(_collector->overflow_list_is_empty(),
6942 "overflow list should be empty");
6943 if (_span.contains(addr) &&
6944 !_bit_map->isMarked(addr)) {
6945 // mark bit map (object is now grey)
6946 _bit_map->mark(addr);
6947 // push on marking stack (stack should be empty), and drain the
6948 // stack by applying this closure to the oops in the oops popped
6949 // from the stack (i.e. blacken the grey objects)
6950 bool res = _mark_stack->push(obj);
6951 assert(res, "Should have space to push on empty stack");
6952 do {
6953 oop new_oop = _mark_stack->pop();
6954 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6955 assert(_bit_map->isMarked((HeapWord*)new_oop),
6956 "only grey objects on this stack");
6957 // iterate over the oops in this oop, marking and pushing
6958 // the ones in CMS heap (i.e. in _span).
6959 new_oop->oop_iterate(&_pushAndMarkClosure);
6960 // check if it's time to yield
6961 do_yield_check();
6962 } while (!_mark_stack->isEmpty() ||
6963 (!_concurrent_precleaning && take_from_overflow_list()));
6964 // if marking stack is empty, and we are not doing this
6965 // during precleaning, then check the overflow list
6966 }
6967 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6968 assert(_collector->overflow_list_is_empty(),
6969 "overflow list was drained above");
6970 // We could restore evacuated mark words, if any, used for
6971 // overflow list links here because the overflow list is
6972 // provably empty here. That would reduce the maximum
6973 // size requirements for preserved_{oop,mark}_stack.
6974 // But we'll just postpone it until we are all done
6975 // so we can just stream through.
6976 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6977 _collector->restore_preserved_marks_if_any();
6978 assert(_collector->no_preserved_marks(), "No preserved marks");
6979 }
6980 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6981 "All preserved marks should have been restored above");
6982 }
6983 }
6984
6985 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6986 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6987
6988 void MarkRefsIntoAndScanClosure::do_yield_work() {
6989 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6990 "CMS thread should hold CMS token");
6991 assert_lock_strong(_freelistLock);
6992 assert_lock_strong(_bit_map->lock());
6993 // relinquish the free_list_lock and bitMaplock()
6994 _bit_map->lock()->unlock();
6995 _freelistLock->unlock();
6996 ConcurrentMarkSweepThread::desynchronize(true);
6997 ConcurrentMarkSweepThread::acknowledge_yield_request();
6998 _collector->stopTimer();
6999 if (PrintCMSStatistics != 0) {
7000 _collector->incrementYields();
7001 }
7002 _collector->icms_wait();
7003
7004 // See the comment in coordinator_yield()
7005 for (unsigned i = 0;
7006 i < CMSYieldSleepCount &&
7007 ConcurrentMarkSweepThread::should_yield() &&
7008 !CMSCollector::foregroundGCIsActive();
7009 ++i) {
7010 os::sleep(Thread::current(), 1, false);
7011 ConcurrentMarkSweepThread::acknowledge_yield_request();
7012 }
7013
7014 ConcurrentMarkSweepThread::synchronize(true);
7015 _freelistLock->lock_without_safepoint_check();
7016 _bit_map->lock()->lock_without_safepoint_check();
7017 _collector->startTimer();
7018 }
7019
7020 ///////////////////////////////////////////////////////////
7021 // Par_MarkRefsIntoAndScanClosure: a parallel version of
7022 // MarkRefsIntoAndScanClosure
7023 ///////////////////////////////////////////////////////////
7024 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
7025 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
7026 CMSBitMap* bit_map, OopTaskQueue* work_queue):
7027 _span(span),
7028 _bit_map(bit_map),
7029 _work_queue(work_queue),
7030 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
7031 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
7032 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
7033 {
7034 _ref_processor = rp;
7035 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7036 }
7037
7038 // This closure is used to mark refs into the CMS generation at the
7039 // second (final) checkpoint, and to scan and transitively follow
7040 // the unmarked oops. The marks are made in the marking bit map and
7041 // the work_queue is used for keeping the (newly) grey objects during
7042 // the scan phase whence they are also available for stealing by parallel
7043 // threads. Since the marking bit map is shared, updates are
7044 // synchronized (via CAS).
7045 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
7046 if (obj != NULL) {
7047 // Ignore mark word because this could be an already marked oop
7048 // that may be chained at the end of the overflow list.
7049 assert(obj->is_oop(true), "expected an oop");
7050 HeapWord* addr = (HeapWord*)obj;
7051 if (_span.contains(addr) &&
7052 !_bit_map->isMarked(addr)) {
7053 // mark bit map (object will become grey):
7054 // It is possible for several threads to be
7055 // trying to "claim" this object concurrently;
7056 // the unique thread that succeeds in marking the
7057 // object first will do the subsequent push on
7058 // to the work queue (or overflow list).
7059 if (_bit_map->par_mark(addr)) {
7060 // push on work_queue (which may not be empty), and trim the
7061 // queue to an appropriate length by applying this closure to
7062 // the oops in the oops popped from the stack (i.e. blacken the
7063 // grey objects)
7064 bool res = _work_queue->push(obj);
7065 assert(res, "Low water mark should be less than capacity?");
7066 trim_queue(_low_water_mark);
7067 } // Else, another thread claimed the object
7068 }
7069 }
7070 }
7071
7072 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
7073 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
7074
7075 // This closure is used to rescan the marked objects on the dirty cards
7076 // in the mod union table and the card table proper.
7077 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
7078 oop p, MemRegion mr) {
7079
7080 size_t size = 0;
7081 HeapWord* addr = (HeapWord*)p;
7082 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7083 assert(_span.contains(addr), "we are scanning the CMS generation");
7084 // check if it's time to yield
7085 if (do_yield_check()) {
7086 // We yielded for some foreground stop-world work,
7087 // and we have been asked to abort this ongoing preclean cycle.
7088 return 0;
7089 }
7090 if (_bitMap->isMarked(addr)) {
7091 // it's marked; is it potentially uninitialized?
7092 if (p->klass_or_null() != NULL) {
7093 // an initialized object; ignore mark word in verification below
7094 // since we are running concurrent with mutators
7095 assert(p->is_oop(true), "should be an oop");
7096 if (p->is_objArray()) {
7097 // objArrays are precisely marked; restrict scanning
7098 // to dirty cards only.
7099 size = CompactibleFreeListSpace::adjustObjectSize(
7100 p->oop_iterate(_scanningClosure, mr));
7101 } else {
7102 // A non-array may have been imprecisely marked; we need
7103 // to scan object in its entirety.
7104 size = CompactibleFreeListSpace::adjustObjectSize(
7105 p->oop_iterate(_scanningClosure));
7106 }
7107 #ifdef ASSERT
7108 size_t direct_size =
7109 CompactibleFreeListSpace::adjustObjectSize(p->size());
7110 assert(size == direct_size, "Inconsistency in size");
7111 assert(size >= 3, "Necessary for Printezis marks to work");
7112 if (!_bitMap->isMarked(addr+1)) {
7113 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
7114 } else {
7115 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
7116 assert(_bitMap->isMarked(addr+size-1),
7117 "inconsistent Printezis mark");
7118 }
7119 #endif // ASSERT
7120 } else {
7121 // An uninitialized object.
7122 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
7123 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7124 size = pointer_delta(nextOneAddr + 1, addr);
7125 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7126 "alignment problem");
7127 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
7128 // will dirty the card when the klass pointer is installed in the
7129 // object (signaling the completion of initialization).
7130 }
7131 } else {
7132 // Either a not yet marked object or an uninitialized object
7133 if (p->klass_or_null() == NULL) {
7134 // An uninitialized object, skip to the next card, since
7135 // we may not be able to read its P-bits yet.
7136 assert(size == 0, "Initial value");
7137 } else {
7138 // An object not (yet) reached by marking: we merely need to
7139 // compute its size so as to go look at the next block.
7140 assert(p->is_oop(true), "should be an oop");
7141 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
7142 }
7143 }
7144 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7145 return size;
7146 }
7147
7148 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
7149 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7150 "CMS thread should hold CMS token");
7151 assert_lock_strong(_freelistLock);
7152 assert_lock_strong(_bitMap->lock());
7153 // relinquish the free_list_lock and bitMaplock()
7154 _bitMap->lock()->unlock();
7155 _freelistLock->unlock();
7156 ConcurrentMarkSweepThread::desynchronize(true);
7157 ConcurrentMarkSweepThread::acknowledge_yield_request();
7158 _collector->stopTimer();
7159 if (PrintCMSStatistics != 0) {
7160 _collector->incrementYields();
7161 }
7162 _collector->icms_wait();
7163
7164 // See the comment in coordinator_yield()
7165 for (unsigned i = 0; i < CMSYieldSleepCount &&
7166 ConcurrentMarkSweepThread::should_yield() &&
7167 !CMSCollector::foregroundGCIsActive(); ++i) {
7168 os::sleep(Thread::current(), 1, false);
7169 ConcurrentMarkSweepThread::acknowledge_yield_request();
7170 }
7171
7172 ConcurrentMarkSweepThread::synchronize(true);
7173 _freelistLock->lock_without_safepoint_check();
7174 _bitMap->lock()->lock_without_safepoint_check();
7175 _collector->startTimer();
7176 }
7177
7178
7179 //////////////////////////////////////////////////////////////////
7180 // SurvivorSpacePrecleanClosure
7181 //////////////////////////////////////////////////////////////////
7182 // This (single-threaded) closure is used to preclean the oops in
7183 // the survivor spaces.
7184 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
7185
7186 HeapWord* addr = (HeapWord*)p;
7187 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7188 assert(!_span.contains(addr), "we are scanning the survivor spaces");
7189 assert(p->klass_or_null() != NULL, "object should be initialized");
7190 // an initialized object; ignore mark word in verification below
7191 // since we are running concurrent with mutators
7192 assert(p->is_oop(true), "should be an oop");
7193 // Note that we do not yield while we iterate over
7194 // the interior oops of p, pushing the relevant ones
7195 // on our marking stack.
7196 size_t size = p->oop_iterate(_scanning_closure);
7197 do_yield_check();
7198 // Observe that below, we do not abandon the preclean
7199 // phase as soon as we should; rather we empty the
7200 // marking stack before returning. This is to satisfy
7201 // some existing assertions. In general, it may be a
7202 // good idea to abort immediately and complete the marking
7203 // from the grey objects at a later time.
7204 while (!_mark_stack->isEmpty()) {
7205 oop new_oop = _mark_stack->pop();
7206 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
7207 assert(_bit_map->isMarked((HeapWord*)new_oop),
7208 "only grey objects on this stack");
7209 // iterate over the oops in this oop, marking and pushing
7210 // the ones in CMS heap (i.e. in _span).
7211 new_oop->oop_iterate(_scanning_closure);
7212 // check if it's time to yield
7213 do_yield_check();
7214 }
7215 unsigned int after_count =
7216 GenCollectedHeap::heap()->total_collections();
7217 bool abort = (_before_count != after_count) ||
7218 _collector->should_abort_preclean();
7219 return abort ? 0 : size;
7220 }
7221
7222 void SurvivorSpacePrecleanClosure::do_yield_work() {
7223 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7224 "CMS thread should hold CMS token");
7225 assert_lock_strong(_bit_map->lock());
7226 // Relinquish the bit map lock
7227 _bit_map->lock()->unlock();
7228 ConcurrentMarkSweepThread::desynchronize(true);
7229 ConcurrentMarkSweepThread::acknowledge_yield_request();
7230 _collector->stopTimer();
7231 if (PrintCMSStatistics != 0) {
7232 _collector->incrementYields();
7233 }
7234 _collector->icms_wait();
7235
7236 // See the comment in coordinator_yield()
7237 for (unsigned i = 0; i < CMSYieldSleepCount &&
7238 ConcurrentMarkSweepThread::should_yield() &&
7239 !CMSCollector::foregroundGCIsActive(); ++i) {
7240 os::sleep(Thread::current(), 1, false);
7241 ConcurrentMarkSweepThread::acknowledge_yield_request();
7242 }
7243
7244 ConcurrentMarkSweepThread::synchronize(true);
7245 _bit_map->lock()->lock_without_safepoint_check();
7246 _collector->startTimer();
7247 }
7248
7249 // This closure is used to rescan the marked objects on the dirty cards
7250 // in the mod union table and the card table proper. In the parallel
7251 // case, although the bitMap is shared, we do a single read so the
7252 // isMarked() query is "safe".
7253 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
7254 // Ignore mark word because we are running concurrent with mutators
7255 assert(p->is_oop_or_null(true), "expected an oop or null");
7256 HeapWord* addr = (HeapWord*)p;
7257 assert(_span.contains(addr), "we are scanning the CMS generation");
7258 bool is_obj_array = false;
7259 #ifdef ASSERT
7260 if (!_parallel) {
7261 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
7262 assert(_collector->overflow_list_is_empty(),
7263 "overflow list should be empty");
7264
7265 }
7266 #endif // ASSERT
7267 if (_bit_map->isMarked(addr)) {
7268 // Obj arrays are precisely marked, non-arrays are not;
7269 // so we scan objArrays precisely and non-arrays in their
7270 // entirety.
7271 if (p->is_objArray()) {
7272 is_obj_array = true;
7273 if (_parallel) {
7274 p->oop_iterate(_par_scan_closure, mr);
7275 } else {
7276 p->oop_iterate(_scan_closure, mr);
7277 }
7278 } else {
7279 if (_parallel) {
7280 p->oop_iterate(_par_scan_closure);
7281 } else {
7282 p->oop_iterate(_scan_closure);
7283 }
7284 }
7285 }
7286 #ifdef ASSERT
7287 if (!_parallel) {
7288 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
7289 assert(_collector->overflow_list_is_empty(),
7290 "overflow list should be empty");
7291
7292 }
7293 #endif // ASSERT
7294 return is_obj_array;
7295 }
7296
7297 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
7298 MemRegion span,
7299 CMSBitMap* bitMap, CMSMarkStack* markStack,
7300 bool should_yield, bool verifying):
7301 _collector(collector),
7302 _span(span),
7303 _bitMap(bitMap),
7304 _mut(&collector->_modUnionTable),
7305 _markStack(markStack),
7306 _yield(should_yield),
7307 _skipBits(0)
7308 {
7309 assert(_markStack->isEmpty(), "stack should be empty");
7310 _finger = _bitMap->startWord();
7311 _threshold = _finger;
7312 assert(_collector->_restart_addr == NULL, "Sanity check");
7313 assert(_span.contains(_finger), "Out of bounds _finger?");
7314 DEBUG_ONLY(_verifying = verifying;)
7315 }
7316
7317 void MarkFromRootsClosure::reset(HeapWord* addr) {
7318 assert(_markStack->isEmpty(), "would cause duplicates on stack");
7319 assert(_span.contains(addr), "Out of bounds _finger?");
7320 _finger = addr;
7321 _threshold = (HeapWord*)round_to(
7322 (intptr_t)_finger, CardTableModRefBS::card_size);
7323 }
7324
7325 // Should revisit to see if this should be restructured for
7326 // greater efficiency.
7327 bool MarkFromRootsClosure::do_bit(size_t offset) {
7328 if (_skipBits > 0) {
7329 _skipBits--;
7330 return true;
7331 }
7332 // convert offset into a HeapWord*
7333 HeapWord* addr = _bitMap->startWord() + offset;
7334 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
7335 "address out of range");
7336 assert(_bitMap->isMarked(addr), "tautology");
7337 if (_bitMap->isMarked(addr+1)) {
7338 // this is an allocated but not yet initialized object
7339 assert(_skipBits == 0, "tautology");
7340 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
7341 oop p = oop(addr);
7342 if (p->klass_or_null() == NULL) {
7343 DEBUG_ONLY(if (!_verifying) {)
7344 // We re-dirty the cards on which this object lies and increase
7345 // the _threshold so that we'll come back to scan this object
7346 // during the preclean or remark phase. (CMSCleanOnEnter)
7347 if (CMSCleanOnEnter) {
7348 size_t sz = _collector->block_size_using_printezis_bits(addr);
7349 HeapWord* end_card_addr = (HeapWord*)round_to(
7350 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7351 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7352 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7353 // Bump _threshold to end_card_addr; note that
7354 // _threshold cannot possibly exceed end_card_addr, anyhow.
7355 // This prevents future clearing of the card as the scan proceeds
7356 // to the right.
7357 assert(_threshold <= end_card_addr,
7358 "Because we are just scanning into this object");
7359 if (_threshold < end_card_addr) {
7360 _threshold = end_card_addr;
7361 }
7362 if (p->klass_or_null() != NULL) {
7363 // Redirty the range of cards...
7364 _mut->mark_range(redirty_range);
7365 } // ...else the setting of klass will dirty the card anyway.
7366 }
7367 DEBUG_ONLY(})
7368 return true;
7369 }
7370 }
7371 scanOopsInOop(addr);
7372 return true;
7373 }
7374
7375 // We take a break if we've been at this for a while,
7376 // so as to avoid monopolizing the locks involved.
7377 void MarkFromRootsClosure::do_yield_work() {
7378 // First give up the locks, then yield, then re-lock
7379 // We should probably use a constructor/destructor idiom to
7380 // do this unlock/lock or modify the MutexUnlocker class to
7381 // serve our purpose. XXX
7382 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7383 "CMS thread should hold CMS token");
7384 assert_lock_strong(_bitMap->lock());
7385 _bitMap->lock()->unlock();
7386 ConcurrentMarkSweepThread::desynchronize(true);
7387 ConcurrentMarkSweepThread::acknowledge_yield_request();
7388 _collector->stopTimer();
7389 if (PrintCMSStatistics != 0) {
7390 _collector->incrementYields();
7391 }
7392 _collector->icms_wait();
7393
7394 // See the comment in coordinator_yield()
7395 for (unsigned i = 0; i < CMSYieldSleepCount &&
7396 ConcurrentMarkSweepThread::should_yield() &&
7397 !CMSCollector::foregroundGCIsActive(); ++i) {
7398 os::sleep(Thread::current(), 1, false);
7399 ConcurrentMarkSweepThread::acknowledge_yield_request();
7400 }
7401
7402 ConcurrentMarkSweepThread::synchronize(true);
7403 _bitMap->lock()->lock_without_safepoint_check();
7404 _collector->startTimer();
7405 }
7406
7407 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
7408 assert(_bitMap->isMarked(ptr), "expected bit to be set");
7409 assert(_markStack->isEmpty(),
7410 "should drain stack to limit stack usage");
7411 // convert ptr to an oop preparatory to scanning
7412 oop obj = oop(ptr);
7413 // Ignore mark word in verification below, since we
7414 // may be running concurrent with mutators.
7415 assert(obj->is_oop(true), "should be an oop");
7416 assert(_finger <= ptr, "_finger runneth ahead");
7417 // advance the finger to right end of this object
7418 _finger = ptr + obj->size();
7419 assert(_finger > ptr, "we just incremented it above");
7420 // On large heaps, it may take us some time to get through
7421 // the marking phase (especially if running iCMS). During
7422 // this time it's possible that a lot of mutations have
7423 // accumulated in the card table and the mod union table --
7424 // these mutation records are redundant until we have
7425 // actually traced into the corresponding card.
7426 // Here, we check whether advancing the finger would make
7427 // us cross into a new card, and if so clear corresponding
7428 // cards in the MUT (preclean them in the card-table in the
7429 // future).
7430
7431 DEBUG_ONLY(if (!_verifying) {)
7432 // The clean-on-enter optimization is disabled by default,
7433 // until we fix 6178663.
7434 if (CMSCleanOnEnter && (_finger > _threshold)) {
7435 // [_threshold, _finger) represents the interval
7436 // of cards to be cleared in MUT (or precleaned in card table).
7437 // The set of cards to be cleared is all those that overlap
7438 // with the interval [_threshold, _finger); note that
7439 // _threshold is always kept card-aligned but _finger isn't
7440 // always card-aligned.
7441 HeapWord* old_threshold = _threshold;
7442 assert(old_threshold == (HeapWord*)round_to(
7443 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7444 "_threshold should always be card-aligned");
7445 _threshold = (HeapWord*)round_to(
7446 (intptr_t)_finger, CardTableModRefBS::card_size);
7447 MemRegion mr(old_threshold, _threshold);
7448 assert(!mr.is_empty(), "Control point invariant");
7449 assert(_span.contains(mr), "Should clear within span");
7450 _mut->clear_range(mr);
7451 }
7452 DEBUG_ONLY(})
7453 // Note: the finger doesn't advance while we drain
7454 // the stack below.
7455 PushOrMarkClosure pushOrMarkClosure(_collector,
7456 _span, _bitMap, _markStack,
7457 _finger, this);
7458 bool res = _markStack->push(obj);
7459 assert(res, "Empty non-zero size stack should have space for single push");
7460 while (!_markStack->isEmpty()) {
7461 oop new_oop = _markStack->pop();
7462 // Skip verifying header mark word below because we are
7463 // running concurrent with mutators.
7464 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7465 // now scan this oop's oops
7466 new_oop->oop_iterate(&pushOrMarkClosure);
7467 do_yield_check();
7468 }
7469 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
7470 }
7471
7472 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
7473 CMSCollector* collector, MemRegion span,
7474 CMSBitMap* bit_map,
7475 OopTaskQueue* work_queue,
7476 CMSMarkStack* overflow_stack,
7477 bool should_yield):
7478 _collector(collector),
7479 _whole_span(collector->_span),
7480 _span(span),
7481 _bit_map(bit_map),
7482 _mut(&collector->_modUnionTable),
7483 _work_queue(work_queue),
7484 _overflow_stack(overflow_stack),
7485 _yield(should_yield),
7486 _skip_bits(0),
7487 _task(task)
7488 {
7489 assert(_work_queue->size() == 0, "work_queue should be empty");
7490 _finger = span.start();
7491 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
7492 assert(_span.contains(_finger), "Out of bounds _finger?");
7493 }
7494
7495 // Should revisit to see if this should be restructured for
7496 // greater efficiency.
7497 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
7498 if (_skip_bits > 0) {
7499 _skip_bits--;
7500 return true;
7501 }
7502 // convert offset into a HeapWord*
7503 HeapWord* addr = _bit_map->startWord() + offset;
7504 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
7505 "address out of range");
7506 assert(_bit_map->isMarked(addr), "tautology");
7507 if (_bit_map->isMarked(addr+1)) {
7508 // this is an allocated object that might not yet be initialized
7509 assert(_skip_bits == 0, "tautology");
7510 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
7511 oop p = oop(addr);
7512 if (p->klass_or_null() == NULL) {
7513 // in the case of Clean-on-Enter optimization, redirty card
7514 // and avoid clearing card by increasing the threshold.
7515 return true;
7516 }
7517 }
7518 scan_oops_in_oop(addr);
7519 return true;
7520 }
7521
7522 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
7523 assert(_bit_map->isMarked(ptr), "expected bit to be set");
7524 // Should we assert that our work queue is empty or
7525 // below some drain limit?
7526 assert(_work_queue->size() == 0,
7527 "should drain stack to limit stack usage");
7528 // convert ptr to an oop preparatory to scanning
7529 oop obj = oop(ptr);
7530 // Ignore mark word in verification below, since we
7531 // may be running concurrent with mutators.
7532 assert(obj->is_oop(true), "should be an oop");
7533 assert(_finger <= ptr, "_finger runneth ahead");
7534 // advance the finger to right end of this object
7535 _finger = ptr + obj->size();
7536 assert(_finger > ptr, "we just incremented it above");
7537 // On large heaps, it may take us some time to get through
7538 // the marking phase (especially if running iCMS). During
7539 // this time it's possible that a lot of mutations have
7540 // accumulated in the card table and the mod union table --
7541 // these mutation records are redundant until we have
7542 // actually traced into the corresponding card.
7543 // Here, we check whether advancing the finger would make
7544 // us cross into a new card, and if so clear corresponding
7545 // cards in the MUT (preclean them in the card-table in the
7546 // future).
7547
7548 // The clean-on-enter optimization is disabled by default,
7549 // until we fix 6178663.
7550 if (CMSCleanOnEnter && (_finger > _threshold)) {
7551 // [_threshold, _finger) represents the interval
7552 // of cards to be cleared in MUT (or precleaned in card table).
7553 // The set of cards to be cleared is all those that overlap
7554 // with the interval [_threshold, _finger); note that
7555 // _threshold is always kept card-aligned but _finger isn't
7556 // always card-aligned.
7557 HeapWord* old_threshold = _threshold;
7558 assert(old_threshold == (HeapWord*)round_to(
7559 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7560 "_threshold should always be card-aligned");
7561 _threshold = (HeapWord*)round_to(
7562 (intptr_t)_finger, CardTableModRefBS::card_size);
7563 MemRegion mr(old_threshold, _threshold);
7564 assert(!mr.is_empty(), "Control point invariant");
7565 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7566 _mut->clear_range(mr);
7567 }
7568
7569 // Note: the local finger doesn't advance while we drain
7570 // the stack below, but the global finger sure can and will.
7571 HeapWord** gfa = _task->global_finger_addr();
7572 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7573 _span, _bit_map,
7574 _work_queue,
7575 _overflow_stack,
7576 _finger,
7577 gfa, this);
7578 bool res = _work_queue->push(obj); // overflow could occur here
7579 assert(res, "Will hold once we use workqueues");
7580 while (true) {
7581 oop new_oop;
7582 if (!_work_queue->pop_local(new_oop)) {
7583 // We emptied our work_queue; check if there's stuff that can
7584 // be gotten from the overflow stack.
7585 if (CMSConcMarkingTask::get_work_from_overflow_stack(
7586 _overflow_stack, _work_queue)) {
7587 do_yield_check();
7588 continue;
7589 } else { // done
7590 break;
7591 }
7592 }
7593 // Skip verifying header mark word below because we are
7594 // running concurrent with mutators.
7595 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7596 // now scan this oop's oops
7597 new_oop->oop_iterate(&pushOrMarkClosure);
7598 do_yield_check();
7599 }
7600 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7601 }
7602
7603 // Yield in response to a request from VM Thread or
7604 // from mutators.
7605 void Par_MarkFromRootsClosure::do_yield_work() {
7606 assert(_task != NULL, "sanity");
7607 _task->yield();
7608 }
7609
7610 // A variant of the above used for verifying CMS marking work.
7611 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7612 MemRegion span,
7613 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7614 CMSMarkStack* mark_stack):
7615 _collector(collector),
7616 _span(span),
7617 _verification_bm(verification_bm),
7618 _cms_bm(cms_bm),
7619 _mark_stack(mark_stack),
7620 _pam_verify_closure(collector, span, verification_bm, cms_bm,
7621 mark_stack)
7622 {
7623 assert(_mark_stack->isEmpty(), "stack should be empty");
7624 _finger = _verification_bm->startWord();
7625 assert(_collector->_restart_addr == NULL, "Sanity check");
7626 assert(_span.contains(_finger), "Out of bounds _finger?");
7627 }
7628
7629 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7630 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7631 assert(_span.contains(addr), "Out of bounds _finger?");
7632 _finger = addr;
7633 }
7634
7635 // Should revisit to see if this should be restructured for
7636 // greater efficiency.
7637 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7638 // convert offset into a HeapWord*
7639 HeapWord* addr = _verification_bm->startWord() + offset;
7640 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7641 "address out of range");
7642 assert(_verification_bm->isMarked(addr), "tautology");
7643 assert(_cms_bm->isMarked(addr), "tautology");
7644
7645 assert(_mark_stack->isEmpty(),
7646 "should drain stack to limit stack usage");
7647 // convert addr to an oop preparatory to scanning
7648 oop obj = oop(addr);
7649 assert(obj->is_oop(), "should be an oop");
7650 assert(_finger <= addr, "_finger runneth ahead");
7651 // advance the finger to right end of this object
7652 _finger = addr + obj->size();
7653 assert(_finger > addr, "we just incremented it above");
7654 // Note: the finger doesn't advance while we drain
7655 // the stack below.
7656 bool res = _mark_stack->push(obj);
7657 assert(res, "Empty non-zero size stack should have space for single push");
7658 while (!_mark_stack->isEmpty()) {
7659 oop new_oop = _mark_stack->pop();
7660 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7661 // now scan this oop's oops
7662 new_oop->oop_iterate(&_pam_verify_closure);
7663 }
7664 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7665 return true;
7666 }
7667
7668 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7669 CMSCollector* collector, MemRegion span,
7670 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7671 CMSMarkStack* mark_stack):
7672 MetadataAwareOopClosure(collector->ref_processor()),
7673 _collector(collector),
7674 _span(span),
7675 _verification_bm(verification_bm),
7676 _cms_bm(cms_bm),
7677 _mark_stack(mark_stack)
7678 { }
7679
7680 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7681 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7682
7683 // Upon stack overflow, we discard (part of) the stack,
7684 // remembering the least address amongst those discarded
7685 // in CMSCollector's _restart_address.
7686 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7687 // Remember the least grey address discarded
7688 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7689 _collector->lower_restart_addr(ra);
7690 _mark_stack->reset(); // discard stack contents
7691 _mark_stack->expand(); // expand the stack if possible
7692 }
7693
7694 void PushAndMarkVerifyClosure::do_oop(oop obj) {
7695 assert(obj->is_oop_or_null(), "expected an oop or NULL");
7696 HeapWord* addr = (HeapWord*)obj;
7697 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7698 // Oop lies in _span and isn't yet grey or black
7699 _verification_bm->mark(addr); // now grey
7700 if (!_cms_bm->isMarked(addr)) {
7701 oop(addr)->print();
7702 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
7703 addr);
7704 fatal("... aborting");
7705 }
7706
7707 if (!_mark_stack->push(obj)) { // stack overflow
7708 if (PrintCMSStatistics != 0) {
7709 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7710 SIZE_FORMAT, _mark_stack->capacity());
7711 }
7712 assert(_mark_stack->isFull(), "Else push should have succeeded");
7713 handle_stack_overflow(addr);
7714 }
7715 // anything including and to the right of _finger
7716 // will be scanned as we iterate over the remainder of the
7717 // bit map
7718 }
7719 }
7720
7721 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7722 MemRegion span,
7723 CMSBitMap* bitMap, CMSMarkStack* markStack,
7724 HeapWord* finger, MarkFromRootsClosure* parent) :
7725 MetadataAwareOopClosure(collector->ref_processor()),
7726 _collector(collector),
7727 _span(span),
7728 _bitMap(bitMap),
7729 _markStack(markStack),
7730 _finger(finger),
7731 _parent(parent)
7732 { }
7733
7734 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7735 MemRegion span,
7736 CMSBitMap* bit_map,
7737 OopTaskQueue* work_queue,
7738 CMSMarkStack* overflow_stack,
7739 HeapWord* finger,
7740 HeapWord** global_finger_addr,
7741 Par_MarkFromRootsClosure* parent) :
7742 MetadataAwareOopClosure(collector->ref_processor()),
7743 _collector(collector),
7744 _whole_span(collector->_span),
7745 _span(span),
7746 _bit_map(bit_map),
7747 _work_queue(work_queue),
7748 _overflow_stack(overflow_stack),
7749 _finger(finger),
7750 _global_finger_addr(global_finger_addr),
7751 _parent(parent)
7752 { }
7753
7754 // Assumes thread-safe access by callers, who are
7755 // responsible for mutual exclusion.
7756 void CMSCollector::lower_restart_addr(HeapWord* low) {
7757 assert(_span.contains(low), "Out of bounds addr");
7758 if (_restart_addr == NULL) {
7759 _restart_addr = low;
7760 } else {
7761 _restart_addr = MIN2(_restart_addr, low);
7762 }
7763 }
7764
7765 // Upon stack overflow, we discard (part of) the stack,
7766 // remembering the least address amongst those discarded
7767 // in CMSCollector's _restart_address.
7768 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7769 // Remember the least grey address discarded
7770 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7771 _collector->lower_restart_addr(ra);
7772 _markStack->reset(); // discard stack contents
7773 _markStack->expand(); // expand the stack if possible
7774 }
7775
7776 // Upon stack overflow, we discard (part of) the stack,
7777 // remembering the least address amongst those discarded
7778 // in CMSCollector's _restart_address.
7779 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7780 // We need to do this under a mutex to prevent other
7781 // workers from interfering with the work done below.
7782 MutexLockerEx ml(_overflow_stack->par_lock(),
7783 Mutex::_no_safepoint_check_flag);
7784 // Remember the least grey address discarded
7785 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7786 _collector->lower_restart_addr(ra);
7787 _overflow_stack->reset(); // discard stack contents
7788 _overflow_stack->expand(); // expand the stack if possible
7789 }
7790
7791 void PushOrMarkClosure::do_oop(oop obj) {
7792 // Ignore mark word because we are running concurrent with mutators.
7793 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7794 HeapWord* addr = (HeapWord*)obj;
7795 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7796 // Oop lies in _span and isn't yet grey or black
7797 _bitMap->mark(addr); // now grey
7798 if (addr < _finger) {
7799 // the bit map iteration has already either passed, or
7800 // sampled, this bit in the bit map; we'll need to
7801 // use the marking stack to scan this oop's oops.
7802 bool simulate_overflow = false;
7803 NOT_PRODUCT(
7804 if (CMSMarkStackOverflowALot &&
7805 _collector->simulate_overflow()) {
7806 // simulate a stack overflow
7807 simulate_overflow = true;
7808 }
7809 )
7810 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7811 if (PrintCMSStatistics != 0) {
7812 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7813 SIZE_FORMAT, _markStack->capacity());
7814 }
7815 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7816 handle_stack_overflow(addr);
7817 }
7818 }
7819 // anything including and to the right of _finger
7820 // will be scanned as we iterate over the remainder of the
7821 // bit map
7822 do_yield_check();
7823 }
7824 }
7825
7826 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7827 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7828
7829 void Par_PushOrMarkClosure::do_oop(oop obj) {
7830 // Ignore mark word because we are running concurrent with mutators.
7831 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7832 HeapWord* addr = (HeapWord*)obj;
7833 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7834 // Oop lies in _span and isn't yet grey or black
7835 // We read the global_finger (volatile read) strictly after marking oop
7836 bool res = _bit_map->par_mark(addr); // now grey
7837 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7838 // Should we push this marked oop on our stack?
7839 // -- if someone else marked it, nothing to do
7840 // -- if target oop is above global finger nothing to do
7841 // -- if target oop is in chunk and above local finger
7842 // then nothing to do
7843 // -- else push on work queue
7844 if ( !res // someone else marked it, they will deal with it
7845 || (addr >= *gfa) // will be scanned in a later task
7846 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7847 return;
7848 }
7849 // the bit map iteration has already either passed, or
7850 // sampled, this bit in the bit map; we'll need to
7851 // use the marking stack to scan this oop's oops.
7852 bool simulate_overflow = false;
7853 NOT_PRODUCT(
7854 if (CMSMarkStackOverflowALot &&
7855 _collector->simulate_overflow()) {
7856 // simulate a stack overflow
7857 simulate_overflow = true;
7858 }
7859 )
7860 if (simulate_overflow ||
7861 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7862 // stack overflow
7863 if (PrintCMSStatistics != 0) {
7864 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7865 SIZE_FORMAT, _overflow_stack->capacity());
7866 }
7867 // We cannot assert that the overflow stack is full because
7868 // it may have been emptied since.
7869 assert(simulate_overflow ||
7870 _work_queue->size() == _work_queue->max_elems(),
7871 "Else push should have succeeded");
7872 handle_stack_overflow(addr);
7873 }
7874 do_yield_check();
7875 }
7876 }
7877
7878 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7879 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7880
7881 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7882 MemRegion span,
7883 ReferenceProcessor* rp,
7884 CMSBitMap* bit_map,
7885 CMSBitMap* mod_union_table,
7886 CMSMarkStack* mark_stack,
7887 bool concurrent_precleaning):
7888 MetadataAwareOopClosure(rp),
7889 _collector(collector),
7890 _span(span),
7891 _bit_map(bit_map),
7892 _mod_union_table(mod_union_table),
7893 _mark_stack(mark_stack),
7894 _concurrent_precleaning(concurrent_precleaning)
7895 {
7896 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7897 }
7898
7899 // Grey object rescan during pre-cleaning and second checkpoint phases --
7900 // the non-parallel version (the parallel version appears further below.)
7901 void PushAndMarkClosure::do_oop(oop obj) {
7902 // Ignore mark word verification. If during concurrent precleaning,
7903 // the object monitor may be locked. If during the checkpoint
7904 // phases, the object may already have been reached by a different
7905 // path and may be at the end of the global overflow list (so
7906 // the mark word may be NULL).
7907 assert(obj->is_oop_or_null(true /* ignore mark word */),
7908 "expected an oop or NULL");
7909 HeapWord* addr = (HeapWord*)obj;
7910 // Check if oop points into the CMS generation
7911 // and is not marked
7912 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7913 // a white object ...
7914 _bit_map->mark(addr); // ... now grey
7915 // push on the marking stack (grey set)
7916 bool simulate_overflow = false;
7917 NOT_PRODUCT(
7918 if (CMSMarkStackOverflowALot &&
7919 _collector->simulate_overflow()) {
7920 // simulate a stack overflow
7921 simulate_overflow = true;
7922 }
7923 )
7924 if (simulate_overflow || !_mark_stack->push(obj)) {
7925 if (_concurrent_precleaning) {
7926 // During precleaning we can just dirty the appropriate card(s)
7927 // in the mod union table, thus ensuring that the object remains
7928 // in the grey set and continue. In the case of object arrays
7929 // we need to dirty all of the cards that the object spans,
7930 // since the rescan of object arrays will be limited to the
7931 // dirty cards.
7932 // Note that no one can be interfering with us in this action
7933 // of dirtying the mod union table, so no locking or atomics
7934 // are required.
7935 if (obj->is_objArray()) {
7936 size_t sz = obj->size();
7937 HeapWord* end_card_addr = (HeapWord*)round_to(
7938 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7939 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7940 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7941 _mod_union_table->mark_range(redirty_range);
7942 } else {
7943 _mod_union_table->mark(addr);
7944 }
7945 _collector->_ser_pmc_preclean_ovflw++;
7946 } else {
7947 // During the remark phase, we need to remember this oop
7948 // in the overflow list.
7949 _collector->push_on_overflow_list(obj);
7950 _collector->_ser_pmc_remark_ovflw++;
7951 }
7952 }
7953 }
7954 }
7955
7956 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7957 MemRegion span,
7958 ReferenceProcessor* rp,
7959 CMSBitMap* bit_map,
7960 OopTaskQueue* work_queue):
7961 MetadataAwareOopClosure(rp),
7962 _collector(collector),
7963 _span(span),
7964 _bit_map(bit_map),
7965 _work_queue(work_queue)
7966 {
7967 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7968 }
7969
7970 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7971 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7972
7973 // Grey object rescan during second checkpoint phase --
7974 // the parallel version.
7975 void Par_PushAndMarkClosure::do_oop(oop obj) {
7976 // In the assert below, we ignore the mark word because
7977 // this oop may point to an already visited object that is
7978 // on the overflow stack (in which case the mark word has
7979 // been hijacked for chaining into the overflow stack --
7980 // if this is the last object in the overflow stack then
7981 // its mark word will be NULL). Because this object may
7982 // have been subsequently popped off the global overflow
7983 // stack, and the mark word possibly restored to the prototypical
7984 // value, by the time we get to examined this failing assert in
7985 // the debugger, is_oop_or_null(false) may subsequently start
7986 // to hold.
7987 assert(obj->is_oop_or_null(true),
7988 "expected an oop or NULL");
7989 HeapWord* addr = (HeapWord*)obj;
7990 // Check if oop points into the CMS generation
7991 // and is not marked
7992 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7993 // a white object ...
7994 // If we manage to "claim" the object, by being the
7995 // first thread to mark it, then we push it on our
7996 // marking stack
7997 if (_bit_map->par_mark(addr)) { // ... now grey
7998 // push on work queue (grey set)
7999 bool simulate_overflow = false;
8000 NOT_PRODUCT(
8001 if (CMSMarkStackOverflowALot &&
8002 _collector->par_simulate_overflow()) {
8003 // simulate a stack overflow
8004 simulate_overflow = true;
8005 }
8006 )
8007 if (simulate_overflow || !_work_queue->push(obj)) {
8008 _collector->par_push_on_overflow_list(obj);
8009 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
8010 }
8011 } // Else, some other thread got there first
8012 }
8013 }
8014
8015 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
8016 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
8017
8018 void CMSPrecleanRefsYieldClosure::do_yield_work() {
8019 Mutex* bml = _collector->bitMapLock();
8020 assert_lock_strong(bml);
8021 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8022 "CMS thread should hold CMS token");
8023
8024 bml->unlock();
8025 ConcurrentMarkSweepThread::desynchronize(true);
8026
8027 ConcurrentMarkSweepThread::acknowledge_yield_request();
8028
8029 _collector->stopTimer();
8030 if (PrintCMSStatistics != 0) {
8031 _collector->incrementYields();
8032 }
8033 _collector->icms_wait();
8034
8035 // See the comment in coordinator_yield()
8036 for (unsigned i = 0; i < CMSYieldSleepCount &&
8037 ConcurrentMarkSweepThread::should_yield() &&
8038 !CMSCollector::foregroundGCIsActive(); ++i) {
8039 os::sleep(Thread::current(), 1, false);
8040 ConcurrentMarkSweepThread::acknowledge_yield_request();
8041 }
8042
8043 ConcurrentMarkSweepThread::synchronize(true);
8044 bml->lock();
8045
8046 _collector->startTimer();
8047 }
8048
8049 bool CMSPrecleanRefsYieldClosure::should_return() {
8050 if (ConcurrentMarkSweepThread::should_yield()) {
8051 do_yield_work();
8052 }
8053 return _collector->foregroundGCIsActive();
8054 }
8055
8056 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
8057 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
8058 "mr should be aligned to start at a card boundary");
8059 // We'd like to assert:
8060 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
8061 // "mr should be a range of cards");
8062 // However, that would be too strong in one case -- the last
8063 // partition ends at _unallocated_block which, in general, can be
8064 // an arbitrary boundary, not necessarily card aligned.
8065 if (PrintCMSStatistics != 0) {
8066 _num_dirty_cards +=
8067 mr.word_size()/CardTableModRefBS::card_size_in_words;
8068 }
8069 _space->object_iterate_mem(mr, &_scan_cl);
8070 }
8071
8072 SweepClosure::SweepClosure(CMSCollector* collector,
8073 ConcurrentMarkSweepGeneration* g,
8074 CMSBitMap* bitMap, bool should_yield) :
8075 _collector(collector),
8076 _g(g),
8077 _sp(g->cmsSpace()),
8078 _limit(_sp->sweep_limit()),
8079 _freelistLock(_sp->freelistLock()),
8080 _bitMap(bitMap),
8081 _yield(should_yield),
8082 _inFreeRange(false), // No free range at beginning of sweep
8083 _freeRangeInFreeLists(false), // No free range at beginning of sweep
8084 _lastFreeRangeCoalesced(false),
8085 _freeFinger(g->used_region().start())
8086 {
8087 NOT_PRODUCT(
8088 _numObjectsFreed = 0;
8089 _numWordsFreed = 0;
8090 _numObjectsLive = 0;
8091 _numWordsLive = 0;
8092 _numObjectsAlreadyFree = 0;
8093 _numWordsAlreadyFree = 0;
8094 _last_fc = NULL;
8095
8096 _sp->initializeIndexedFreeListArrayReturnedBytes();
8097 _sp->dictionary()->initialize_dict_returned_bytes();
8098 )
8099 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8100 "sweep _limit out of bounds");
8101 if (CMSTraceSweeper) {
8102 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
8103 _limit);
8104 }
8105 }
8106
8107 void SweepClosure::print_on(outputStream* st) const {
8108 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
8109 _sp->bottom(), _sp->end());
8110 tty->print_cr("_limit = " PTR_FORMAT, _limit);
8111 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger);
8112 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);)
8113 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
8114 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
8115 }
8116
8117 #ifndef PRODUCT
8118 // Assertion checking only: no useful work in product mode --
8119 // however, if any of the flags below become product flags,
8120 // you may need to review this code to see if it needs to be
8121 // enabled in product mode.
8122 SweepClosure::~SweepClosure() {
8123 assert_lock_strong(_freelistLock);
8124 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8125 "sweep _limit out of bounds");
8126 if (inFreeRange()) {
8127 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
8128 print();
8129 ShouldNotReachHere();
8130 }
8131 if (Verbose && PrintGC) {
8132 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes",
8133 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
8134 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
8135 SIZE_FORMAT" bytes "
8136 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
8137 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
8138 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
8139 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
8140 * sizeof(HeapWord);
8141 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
8142
8143 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
8144 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
8145 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
8146 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
8147 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes);
8148 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
8149 indexListReturnedBytes);
8150 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
8151 dict_returned_bytes);
8152 }
8153 }
8154 if (CMSTraceSweeper) {
8155 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
8156 _limit);
8157 }
8158 }
8159 #endif // PRODUCT
8160
8161 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
8162 bool freeRangeInFreeLists) {
8163 if (CMSTraceSweeper) {
8164 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n",
8165 freeFinger, freeRangeInFreeLists);
8166 }
8167 assert(!inFreeRange(), "Trampling existing free range");
8168 set_inFreeRange(true);
8169 set_lastFreeRangeCoalesced(false);
8170
8171 set_freeFinger(freeFinger);
8172 set_freeRangeInFreeLists(freeRangeInFreeLists);
8173 if (CMSTestInFreeList) {
8174 if (freeRangeInFreeLists) {
8175 FreeChunk* fc = (FreeChunk*) freeFinger;
8176 assert(fc->is_free(), "A chunk on the free list should be free.");
8177 assert(fc->size() > 0, "Free range should have a size");
8178 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
8179 }
8180 }
8181 }
8182
8183 // Note that the sweeper runs concurrently with mutators. Thus,
8184 // it is possible for direct allocation in this generation to happen
8185 // in the middle of the sweep. Note that the sweeper also coalesces
8186 // contiguous free blocks. Thus, unless the sweeper and the allocator
8187 // synchronize appropriately freshly allocated blocks may get swept up.
8188 // This is accomplished by the sweeper locking the free lists while
8189 // it is sweeping. Thus blocks that are determined to be free are
8190 // indeed free. There is however one additional complication:
8191 // blocks that have been allocated since the final checkpoint and
8192 // mark, will not have been marked and so would be treated as
8193 // unreachable and swept up. To prevent this, the allocator marks
8194 // the bit map when allocating during the sweep phase. This leads,
8195 // however, to a further complication -- objects may have been allocated
8196 // but not yet initialized -- in the sense that the header isn't yet
8197 // installed. The sweeper can not then determine the size of the block
8198 // in order to skip over it. To deal with this case, we use a technique
8199 // (due to Printezis) to encode such uninitialized block sizes in the
8200 // bit map. Since the bit map uses a bit per every HeapWord, but the
8201 // CMS generation has a minimum object size of 3 HeapWords, it follows
8202 // that "normal marks" won't be adjacent in the bit map (there will
8203 // always be at least two 0 bits between successive 1 bits). We make use
8204 // of these "unused" bits to represent uninitialized blocks -- the bit
8205 // corresponding to the start of the uninitialized object and the next
8206 // bit are both set. Finally, a 1 bit marks the end of the object that
8207 // started with the two consecutive 1 bits to indicate its potentially
8208 // uninitialized state.
8209
8210 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
8211 FreeChunk* fc = (FreeChunk*)addr;
8212 size_t res;
8213
8214 // Check if we are done sweeping. Below we check "addr >= _limit" rather
8215 // than "addr == _limit" because although _limit was a block boundary when
8216 // we started the sweep, it may no longer be one because heap expansion
8217 // may have caused us to coalesce the block ending at the address _limit
8218 // with a newly expanded chunk (this happens when _limit was set to the
8219 // previous _end of the space), so we may have stepped past _limit:
8220 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
8221 if (addr >= _limit) { // we have swept up to or past the limit: finish up
8222 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8223 "sweep _limit out of bounds");
8224 assert(addr < _sp->end(), "addr out of bounds");
8225 // Flush any free range we might be holding as a single
8226 // coalesced chunk to the appropriate free list.
8227 if (inFreeRange()) {
8228 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
8229 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger()));
8230 flush_cur_free_chunk(freeFinger(),
8231 pointer_delta(addr, freeFinger()));
8232 if (CMSTraceSweeper) {
8233 gclog_or_tty->print("Sweep: last chunk: ");
8234 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") "
8235 "[coalesced:%d]\n",
8236 freeFinger(), pointer_delta(addr, freeFinger()),
8237 lastFreeRangeCoalesced() ? 1 : 0);
8238 }
8239 }
8240
8241 // help the iterator loop finish
8242 return pointer_delta(_sp->end(), addr);
8243 }
8244
8245 assert(addr < _limit, "sweep invariant");
8246 // check if we should yield
8247 do_yield_check(addr);
8248 if (fc->is_free()) {
8249 // Chunk that is already free
8250 res = fc->size();
8251 do_already_free_chunk(fc);
8252 debug_only(_sp->verifyFreeLists());
8253 // If we flush the chunk at hand in lookahead_and_flush()
8254 // and it's coalesced with a preceding chunk, then the
8255 // process of "mangling" the payload of the coalesced block
8256 // will cause erasure of the size information from the
8257 // (erstwhile) header of all the coalesced blocks but the
8258 // first, so the first disjunct in the assert will not hold
8259 // in that specific case (in which case the second disjunct
8260 // will hold).
8261 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
8262 "Otherwise the size info doesn't change at this step");
8263 NOT_PRODUCT(
8264 _numObjectsAlreadyFree++;
8265 _numWordsAlreadyFree += res;
8266 )
8267 NOT_PRODUCT(_last_fc = fc;)
8268 } else if (!_bitMap->isMarked(addr)) {
8269 // Chunk is fresh garbage
8270 res = do_garbage_chunk(fc);
8271 debug_only(_sp->verifyFreeLists());
8272 NOT_PRODUCT(
8273 _numObjectsFreed++;
8274 _numWordsFreed += res;
8275 )
8276 } else {
8277 // Chunk that is alive.
8278 res = do_live_chunk(fc);
8279 debug_only(_sp->verifyFreeLists());
8280 NOT_PRODUCT(
8281 _numObjectsLive++;
8282 _numWordsLive += res;
8283 )
8284 }
8285 return res;
8286 }
8287
8288 // For the smart allocation, record following
8289 // split deaths - a free chunk is removed from its free list because
8290 // it is being split into two or more chunks.
8291 // split birth - a free chunk is being added to its free list because
8292 // a larger free chunk has been split and resulted in this free chunk.
8293 // coal death - a free chunk is being removed from its free list because
8294 // it is being coalesced into a large free chunk.
8295 // coal birth - a free chunk is being added to its free list because
8296 // it was created when two or more free chunks where coalesced into
8297 // this free chunk.
8298 //
8299 // These statistics are used to determine the desired number of free
8300 // chunks of a given size. The desired number is chosen to be relative
8301 // to the end of a CMS sweep. The desired number at the end of a sweep
8302 // is the
8303 // count-at-end-of-previous-sweep (an amount that was enough)
8304 // - count-at-beginning-of-current-sweep (the excess)
8305 // + split-births (gains in this size during interval)
8306 // - split-deaths (demands on this size during interval)
8307 // where the interval is from the end of one sweep to the end of the
8308 // next.
8309 //
8310 // When sweeping the sweeper maintains an accumulated chunk which is
8311 // the chunk that is made up of chunks that have been coalesced. That
8312 // will be termed the left-hand chunk. A new chunk of garbage that
8313 // is being considered for coalescing will be referred to as the
8314 // right-hand chunk.
8315 //
8316 // When making a decision on whether to coalesce a right-hand chunk with
8317 // the current left-hand chunk, the current count vs. the desired count
8318 // of the left-hand chunk is considered. Also if the right-hand chunk
8319 // is near the large chunk at the end of the heap (see
8320 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
8321 // left-hand chunk is coalesced.
8322 //
8323 // When making a decision about whether to split a chunk, the desired count
8324 // vs. the current count of the candidate to be split is also considered.
8325 // If the candidate is underpopulated (currently fewer chunks than desired)
8326 // a chunk of an overpopulated (currently more chunks than desired) size may
8327 // be chosen. The "hint" associated with a free list, if non-null, points
8328 // to a free list which may be overpopulated.
8329 //
8330
8331 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
8332 const size_t size = fc->size();
8333 // Chunks that cannot be coalesced are not in the
8334 // free lists.
8335 if (CMSTestInFreeList && !fc->cantCoalesce()) {
8336 assert(_sp->verify_chunk_in_free_list(fc),
8337 "free chunk should be in free lists");
8338 }
8339 // a chunk that is already free, should not have been
8340 // marked in the bit map
8341 HeapWord* const addr = (HeapWord*) fc;
8342 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
8343 // Verify that the bit map has no bits marked between
8344 // addr and purported end of this block.
8345 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8346
8347 // Some chunks cannot be coalesced under any circumstances.
8348 // See the definition of cantCoalesce().
8349 if (!fc->cantCoalesce()) {
8350 // This chunk can potentially be coalesced.
8351 if (_sp->adaptive_freelists()) {
8352 // All the work is done in
8353 do_post_free_or_garbage_chunk(fc, size);
8354 } else { // Not adaptive free lists
8355 // this is a free chunk that can potentially be coalesced by the sweeper;
8356 if (!inFreeRange()) {
8357 // if the next chunk is a free block that can't be coalesced
8358 // it doesn't make sense to remove this chunk from the free lists
8359 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
8360 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
8361 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
8362 nextChunk->is_free() && // ... which is free...
8363 nextChunk->cantCoalesce()) { // ... but can't be coalesced
8364 // nothing to do
8365 } else {
8366 // Potentially the start of a new free range:
8367 // Don't eagerly remove it from the free lists.
8368 // No need to remove it if it will just be put
8369 // back again. (Also from a pragmatic point of view
8370 // if it is a free block in a region that is beyond
8371 // any allocated blocks, an assertion will fail)
8372 // Remember the start of a free run.
8373 initialize_free_range(addr, true);
8374 // end - can coalesce with next chunk
8375 }
8376 } else {
8377 // the midst of a free range, we are coalescing
8378 print_free_block_coalesced(fc);
8379 if (CMSTraceSweeper) {
8380 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", fc, size);
8381 }
8382 // remove it from the free lists
8383 _sp->removeFreeChunkFromFreeLists(fc);
8384 set_lastFreeRangeCoalesced(true);
8385 // If the chunk is being coalesced and the current free range is
8386 // in the free lists, remove the current free range so that it
8387 // will be returned to the free lists in its entirety - all
8388 // the coalesced pieces included.
8389 if (freeRangeInFreeLists()) {
8390 FreeChunk* ffc = (FreeChunk*) freeFinger();
8391 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8392 "Size of free range is inconsistent with chunk size.");
8393 if (CMSTestInFreeList) {
8394 assert(_sp->verify_chunk_in_free_list(ffc),
8395 "free range is not in free lists");
8396 }
8397 _sp->removeFreeChunkFromFreeLists(ffc);
8398 set_freeRangeInFreeLists(false);
8399 }
8400 }
8401 }
8402 // Note that if the chunk is not coalescable (the else arm
8403 // below), we unconditionally flush, without needing to do
8404 // a "lookahead," as we do below.
8405 if (inFreeRange()) lookahead_and_flush(fc, size);
8406 } else {
8407 // Code path common to both original and adaptive free lists.
8408
8409 // cant coalesce with previous block; this should be treated
8410 // as the end of a free run if any
8411 if (inFreeRange()) {
8412 // we kicked some butt; time to pick up the garbage
8413 assert(freeFinger() < addr, "freeFinger points too high");
8414 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8415 }
8416 // else, nothing to do, just continue
8417 }
8418 }
8419
8420 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
8421 // This is a chunk of garbage. It is not in any free list.
8422 // Add it to a free list or let it possibly be coalesced into
8423 // a larger chunk.
8424 HeapWord* const addr = (HeapWord*) fc;
8425 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8426
8427 if (_sp->adaptive_freelists()) {
8428 // Verify that the bit map has no bits marked between
8429 // addr and purported end of just dead object.
8430 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8431
8432 do_post_free_or_garbage_chunk(fc, size);
8433 } else {
8434 if (!inFreeRange()) {
8435 // start of a new free range
8436 assert(size > 0, "A free range should have a size");
8437 initialize_free_range(addr, false);
8438 } else {
8439 // this will be swept up when we hit the end of the
8440 // free range
8441 if (CMSTraceSweeper) {
8442 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", fc, size);
8443 }
8444 // If the chunk is being coalesced and the current free range is
8445 // in the free lists, remove the current free range so that it
8446 // will be returned to the free lists in its entirety - all
8447 // the coalesced pieces included.
8448 if (freeRangeInFreeLists()) {
8449 FreeChunk* ffc = (FreeChunk*)freeFinger();
8450 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8451 "Size of free range is inconsistent with chunk size.");
8452 if (CMSTestInFreeList) {
8453 assert(_sp->verify_chunk_in_free_list(ffc),
8454 "free range is not in free lists");
8455 }
8456 _sp->removeFreeChunkFromFreeLists(ffc);
8457 set_freeRangeInFreeLists(false);
8458 }
8459 set_lastFreeRangeCoalesced(true);
8460 }
8461 // this will be swept up when we hit the end of the free range
8462
8463 // Verify that the bit map has no bits marked between
8464 // addr and purported end of just dead object.
8465 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8466 }
8467 assert(_limit >= addr + size,
8468 "A freshly garbage chunk can't possibly straddle over _limit");
8469 if (inFreeRange()) lookahead_and_flush(fc, size);
8470 return size;
8471 }
8472
8473 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
8474 HeapWord* addr = (HeapWord*) fc;
8475 // The sweeper has just found a live object. Return any accumulated
8476 // left hand chunk to the free lists.
8477 if (inFreeRange()) {
8478 assert(freeFinger() < addr, "freeFinger points too high");
8479 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8480 }
8481
8482 // This object is live: we'd normally expect this to be
8483 // an oop, and like to assert the following:
8484 // assert(oop(addr)->is_oop(), "live block should be an oop");
8485 // However, as we commented above, this may be an object whose
8486 // header hasn't yet been initialized.
8487 size_t size;
8488 assert(_bitMap->isMarked(addr), "Tautology for this control point");
8489 if (_bitMap->isMarked(addr + 1)) {
8490 // Determine the size from the bit map, rather than trying to
8491 // compute it from the object header.
8492 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
8493 size = pointer_delta(nextOneAddr + 1, addr);
8494 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
8495 "alignment problem");
8496
8497 #ifdef ASSERT
8498 if (oop(addr)->klass_or_null() != NULL) {
8499 // Ignore mark word because we are running concurrent with mutators
8500 assert(oop(addr)->is_oop(true), "live block should be an oop");
8501 assert(size ==
8502 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
8503 "P-mark and computed size do not agree");
8504 }
8505 #endif
8506
8507 } else {
8508 // This should be an initialized object that's alive.
8509 assert(oop(addr)->klass_or_null() != NULL,
8510 "Should be an initialized object");
8511 // Ignore mark word because we are running concurrent with mutators
8512 assert(oop(addr)->is_oop(true), "live block should be an oop");
8513 // Verify that the bit map has no bits marked between
8514 // addr and purported end of this block.
8515 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8516 assert(size >= 3, "Necessary for Printezis marks to work");
8517 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
8518 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
8519 }
8520 return size;
8521 }
8522
8523 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
8524 size_t chunkSize) {
8525 // do_post_free_or_garbage_chunk() should only be called in the case
8526 // of the adaptive free list allocator.
8527 const bool fcInFreeLists = fc->is_free();
8528 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
8529 assert((HeapWord*)fc <= _limit, "sweep invariant");
8530 if (CMSTestInFreeList && fcInFreeLists) {
8531 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
8532 }
8533
8534 if (CMSTraceSweeper) {
8535 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", fc, chunkSize);
8536 }
8537
8538 HeapWord* const fc_addr = (HeapWord*) fc;
8539
8540 bool coalesce;
8541 const size_t left = pointer_delta(fc_addr, freeFinger());
8542 const size_t right = chunkSize;
8543 switch (FLSCoalescePolicy) {
8544 // numeric value forms a coalition aggressiveness metric
8545 case 0: { // never coalesce
8546 coalesce = false;
8547 break;
8548 }
8549 case 1: { // coalesce if left & right chunks on overpopulated lists
8550 coalesce = _sp->coalOverPopulated(left) &&
8551 _sp->coalOverPopulated(right);
8552 break;
8553 }
8554 case 2: { // coalesce if left chunk on overpopulated list (default)
8555 coalesce = _sp->coalOverPopulated(left);
8556 break;
8557 }
8558 case 3: { // coalesce if left OR right chunk on overpopulated list
8559 coalesce = _sp->coalOverPopulated(left) ||
8560 _sp->coalOverPopulated(right);
8561 break;
8562 }
8563 case 4: { // always coalesce
8564 coalesce = true;
8565 break;
8566 }
8567 default:
8568 ShouldNotReachHere();
8569 }
8570
8571 // Should the current free range be coalesced?
8572 // If the chunk is in a free range and either we decided to coalesce above
8573 // or the chunk is near the large block at the end of the heap
8574 // (isNearLargestChunk() returns true), then coalesce this chunk.
8575 const bool doCoalesce = inFreeRange()
8576 && (coalesce || _g->isNearLargestChunk(fc_addr));
8577 if (doCoalesce) {
8578 // Coalesce the current free range on the left with the new
8579 // chunk on the right. If either is on a free list,
8580 // it must be removed from the list and stashed in the closure.
8581 if (freeRangeInFreeLists()) {
8582 FreeChunk* const ffc = (FreeChunk*)freeFinger();
8583 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
8584 "Size of free range is inconsistent with chunk size.");
8585 if (CMSTestInFreeList) {
8586 assert(_sp->verify_chunk_in_free_list(ffc),
8587 "Chunk is not in free lists");
8588 }
8589 _sp->coalDeath(ffc->size());
8590 _sp->removeFreeChunkFromFreeLists(ffc);
8591 set_freeRangeInFreeLists(false);
8592 }
8593 if (fcInFreeLists) {
8594 _sp->coalDeath(chunkSize);
8595 assert(fc->size() == chunkSize,
8596 "The chunk has the wrong size or is not in the free lists");
8597 _sp->removeFreeChunkFromFreeLists(fc);
8598 }
8599 set_lastFreeRangeCoalesced(true);
8600 print_free_block_coalesced(fc);
8601 } else { // not in a free range and/or should not coalesce
8602 // Return the current free range and start a new one.
8603 if (inFreeRange()) {
8604 // In a free range but cannot coalesce with the right hand chunk.
8605 // Put the current free range into the free lists.
8606 flush_cur_free_chunk(freeFinger(),
8607 pointer_delta(fc_addr, freeFinger()));
8608 }
8609 // Set up for new free range. Pass along whether the right hand
8610 // chunk is in the free lists.
8611 initialize_free_range((HeapWord*)fc, fcInFreeLists);
8612 }
8613 }
8614
8615 // Lookahead flush:
8616 // If we are tracking a free range, and this is the last chunk that
8617 // we'll look at because its end crosses past _limit, we'll preemptively
8618 // flush it along with any free range we may be holding on to. Note that
8619 // this can be the case only for an already free or freshly garbage
8620 // chunk. If this block is an object, it can never straddle
8621 // over _limit. The "straddling" occurs when _limit is set at
8622 // the previous end of the space when this cycle started, and
8623 // a subsequent heap expansion caused the previously co-terminal
8624 // free block to be coalesced with the newly expanded portion,
8625 // thus rendering _limit a non-block-boundary making it dangerous
8626 // for the sweeper to step over and examine.
8627 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
8628 assert(inFreeRange(), "Should only be called if currently in a free range.");
8629 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
8630 assert(_sp->used_region().contains(eob - 1),
8631 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
8632 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
8633 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
8634 eob, eob-1, _limit, _sp->bottom(), _sp->end(), fc, chunk_size));
8635 if (eob >= _limit) {
8636 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
8637 if (CMSTraceSweeper) {
8638 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
8639 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
8640 "[" PTR_FORMAT "," PTR_FORMAT ")",
8641 _limit, fc, eob, _sp->bottom(), _sp->end());
8642 }
8643 // Return the storage we are tracking back into the free lists.
8644 if (CMSTraceSweeper) {
8645 gclog_or_tty->print_cr("Flushing ... ");
8646 }
8647 assert(freeFinger() < eob, "Error");
8648 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
8649 }
8650 }
8651
8652 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
8653 assert(inFreeRange(), "Should only be called if currently in a free range.");
8654 assert(size > 0,
8655 "A zero sized chunk cannot be added to the free lists.");
8656 if (!freeRangeInFreeLists()) {
8657 if (CMSTestInFreeList) {
8658 FreeChunk* fc = (FreeChunk*) chunk;
8659 fc->set_size(size);
8660 assert(!_sp->verify_chunk_in_free_list(fc),
8661 "chunk should not be in free lists yet");
8662 }
8663 if (CMSTraceSweeper) {
8664 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists",
8665 chunk, size);
8666 }
8667 // A new free range is going to be starting. The current
8668 // free range has not been added to the free lists yet or
8669 // was removed so add it back.
8670 // If the current free range was coalesced, then the death
8671 // of the free range was recorded. Record a birth now.
8672 if (lastFreeRangeCoalesced()) {
8673 _sp->coalBirth(size);
8674 }
8675 _sp->addChunkAndRepairOffsetTable(chunk, size,
8676 lastFreeRangeCoalesced());
8677 } else if (CMSTraceSweeper) {
8678 gclog_or_tty->print_cr("Already in free list: nothing to flush");
8679 }
8680 set_inFreeRange(false);
8681 set_freeRangeInFreeLists(false);
8682 }
8683
8684 // We take a break if we've been at this for a while,
8685 // so as to avoid monopolizing the locks involved.
8686 void SweepClosure::do_yield_work(HeapWord* addr) {
8687 // Return current free chunk being used for coalescing (if any)
8688 // to the appropriate freelist. After yielding, the next
8689 // free block encountered will start a coalescing range of
8690 // free blocks. If the next free block is adjacent to the
8691 // chunk just flushed, they will need to wait for the next
8692 // sweep to be coalesced.
8693 if (inFreeRange()) {
8694 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8695 }
8696
8697 // First give up the locks, then yield, then re-lock.
8698 // We should probably use a constructor/destructor idiom to
8699 // do this unlock/lock or modify the MutexUnlocker class to
8700 // serve our purpose. XXX
8701 assert_lock_strong(_bitMap->lock());
8702 assert_lock_strong(_freelistLock);
8703 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8704 "CMS thread should hold CMS token");
8705 _bitMap->lock()->unlock();
8706 _freelistLock->unlock();
8707 ConcurrentMarkSweepThread::desynchronize(true);
8708 ConcurrentMarkSweepThread::acknowledge_yield_request();
8709 _collector->stopTimer();
8710 if (PrintCMSStatistics != 0) {
8711 _collector->incrementYields();
8712 }
8713 _collector->icms_wait();
8714
8715 // See the comment in coordinator_yield()
8716 for (unsigned i = 0; i < CMSYieldSleepCount &&
8717 ConcurrentMarkSweepThread::should_yield() &&
8718 !CMSCollector::foregroundGCIsActive(); ++i) {
8719 os::sleep(Thread::current(), 1, false);
8720 ConcurrentMarkSweepThread::acknowledge_yield_request();
8721 }
8722
8723 ConcurrentMarkSweepThread::synchronize(true);
8724 _freelistLock->lock();
8725 _bitMap->lock()->lock_without_safepoint_check();
8726 _collector->startTimer();
8727 }
8728
8729 #ifndef PRODUCT
8730 // This is actually very useful in a product build if it can
8731 // be called from the debugger. Compile it into the product
8732 // as needed.
8733 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
8734 return debug_cms_space->verify_chunk_in_free_list(fc);
8735 }
8736 #endif
8737
8738 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
8739 if (CMSTraceSweeper) {
8740 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
8741 fc, fc->size());
8742 }
8743 }
8744
8745 // CMSIsAliveClosure
8746 bool CMSIsAliveClosure::do_object_b(oop obj) {
8747 HeapWord* addr = (HeapWord*)obj;
8748 return addr != NULL &&
8749 (!_span.contains(addr) || _bit_map->isMarked(addr));
8750 }
8751
8752
8753 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
8754 MemRegion span,
8755 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
8756 bool cpc):
8757 _collector(collector),
8758 _span(span),
8759 _bit_map(bit_map),
8760 _mark_stack(mark_stack),
8761 _concurrent_precleaning(cpc) {
8762 assert(!_span.is_empty(), "Empty span could spell trouble");
8763 }
8764
8765
8766 // CMSKeepAliveClosure: the serial version
8767 void CMSKeepAliveClosure::do_oop(oop obj) {
8768 HeapWord* addr = (HeapWord*)obj;
8769 if (_span.contains(addr) &&
8770 !_bit_map->isMarked(addr)) {
8771 _bit_map->mark(addr);
8772 bool simulate_overflow = false;
8773 NOT_PRODUCT(
8774 if (CMSMarkStackOverflowALot &&
8775 _collector->simulate_overflow()) {
8776 // simulate a stack overflow
8777 simulate_overflow = true;
8778 }
8779 )
8780 if (simulate_overflow || !_mark_stack->push(obj)) {
8781 if (_concurrent_precleaning) {
8782 // We dirty the overflown object and let the remark
8783 // phase deal with it.
8784 assert(_collector->overflow_list_is_empty(), "Error");
8785 // In the case of object arrays, we need to dirty all of
8786 // the cards that the object spans. No locking or atomics
8787 // are needed since no one else can be mutating the mod union
8788 // table.
8789 if (obj->is_objArray()) {
8790 size_t sz = obj->size();
8791 HeapWord* end_card_addr =
8792 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
8793 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8794 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8795 _collector->_modUnionTable.mark_range(redirty_range);
8796 } else {
8797 _collector->_modUnionTable.mark(addr);
8798 }
8799 _collector->_ser_kac_preclean_ovflw++;
8800 } else {
8801 _collector->push_on_overflow_list(obj);
8802 _collector->_ser_kac_ovflw++;
8803 }
8804 }
8805 }
8806 }
8807
8808 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8809 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8810
8811 // CMSParKeepAliveClosure: a parallel version of the above.
8812 // The work queues are private to each closure (thread),
8813 // but (may be) available for stealing by other threads.
8814 void CMSParKeepAliveClosure::do_oop(oop obj) {
8815 HeapWord* addr = (HeapWord*)obj;
8816 if (_span.contains(addr) &&
8817 !_bit_map->isMarked(addr)) {
8818 // In general, during recursive tracing, several threads
8819 // may be concurrently getting here; the first one to
8820 // "tag" it, claims it.
8821 if (_bit_map->par_mark(addr)) {
8822 bool res = _work_queue->push(obj);
8823 assert(res, "Low water mark should be much less than capacity");
8824 // Do a recursive trim in the hope that this will keep
8825 // stack usage lower, but leave some oops for potential stealers
8826 trim_queue(_low_water_mark);
8827 } // Else, another thread got there first
8828 }
8829 }
8830
8831 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8832 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8833
8834 void CMSParKeepAliveClosure::trim_queue(uint max) {
8835 while (_work_queue->size() > max) {
8836 oop new_oop;
8837 if (_work_queue->pop_local(new_oop)) {
8838 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8839 assert(_bit_map->isMarked((HeapWord*)new_oop),
8840 "no white objects on this stack!");
8841 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8842 // iterate over the oops in this oop, marking and pushing
8843 // the ones in CMS heap (i.e. in _span).
8844 new_oop->oop_iterate(&_mark_and_push);
8845 }
8846 }
8847 }
8848
8849 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8850 CMSCollector* collector,
8851 MemRegion span, CMSBitMap* bit_map,
8852 OopTaskQueue* work_queue):
8853 _collector(collector),
8854 _span(span),
8855 _bit_map(bit_map),
8856 _work_queue(work_queue) { }
8857
8858 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8859 HeapWord* addr = (HeapWord*)obj;
8860 if (_span.contains(addr) &&
8861 !_bit_map->isMarked(addr)) {
8862 if (_bit_map->par_mark(addr)) {
8863 bool simulate_overflow = false;
8864 NOT_PRODUCT(
8865 if (CMSMarkStackOverflowALot &&
8866 _collector->par_simulate_overflow()) {
8867 // simulate a stack overflow
8868 simulate_overflow = true;
8869 }
8870 )
8871 if (simulate_overflow || !_work_queue->push(obj)) {
8872 _collector->par_push_on_overflow_list(obj);
8873 _collector->_par_kac_ovflw++;
8874 }
8875 } // Else another thread got there already
8876 }
8877 }
8878
8879 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8880 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8881
8882 //////////////////////////////////////////////////////////////////
8883 // CMSExpansionCause /////////////////////////////
8884 //////////////////////////////////////////////////////////////////
8885 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8886 switch (cause) {
8887 case _no_expansion:
8888 return "No expansion";
8889 case _satisfy_free_ratio:
8890 return "Free ratio";
8891 case _satisfy_promotion:
8892 return "Satisfy promotion";
8893 case _satisfy_allocation:
8894 return "allocation";
8895 case _allocate_par_lab:
8896 return "Par LAB";
8897 case _allocate_par_spooling_space:
8898 return "Par Spooling Space";
8899 case _adaptive_size_policy:
8900 return "Ergonomics";
8901 default:
8902 return "unknown";
8903 }
8904 }
8905
8906 void CMSDrainMarkingStackClosure::do_void() {
8907 // the max number to take from overflow list at a time
8908 const size_t num = _mark_stack->capacity()/4;
8909 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8910 "Overflow list should be NULL during concurrent phases");
8911 while (!_mark_stack->isEmpty() ||
8912 // if stack is empty, check the overflow list
8913 _collector->take_from_overflow_list(num, _mark_stack)) {
8914 oop obj = _mark_stack->pop();
8915 HeapWord* addr = (HeapWord*)obj;
8916 assert(_span.contains(addr), "Should be within span");
8917 assert(_bit_map->isMarked(addr), "Should be marked");
8918 assert(obj->is_oop(), "Should be an oop");
8919 obj->oop_iterate(_keep_alive);
8920 }
8921 }
8922
8923 void CMSParDrainMarkingStackClosure::do_void() {
8924 // drain queue
8925 trim_queue(0);
8926 }
8927
8928 // Trim our work_queue so its length is below max at return
8929 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8930 while (_work_queue->size() > max) {
8931 oop new_oop;
8932 if (_work_queue->pop_local(new_oop)) {
8933 assert(new_oop->is_oop(), "Expected an oop");
8934 assert(_bit_map->isMarked((HeapWord*)new_oop),
8935 "no white objects on this stack!");
8936 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8937 // iterate over the oops in this oop, marking and pushing
8938 // the ones in CMS heap (i.e. in _span).
8939 new_oop->oop_iterate(&_mark_and_push);
8940 }
8941 }
8942 }
8943
8944 ////////////////////////////////////////////////////////////////////
8945 // Support for Marking Stack Overflow list handling and related code
8946 ////////////////////////////////////////////////////////////////////
8947 // Much of the following code is similar in shape and spirit to the
8948 // code used in ParNewGC. We should try and share that code
8949 // as much as possible in the future.
8950
8951 #ifndef PRODUCT
8952 // Debugging support for CMSStackOverflowALot
8953
8954 // It's OK to call this multi-threaded; the worst thing
8955 // that can happen is that we'll get a bunch of closely
8956 // spaced simulated overflows, but that's OK, in fact
8957 // probably good as it would exercise the overflow code
8958 // under contention.
8959 bool CMSCollector::simulate_overflow() {
8960 if (_overflow_counter-- <= 0) { // just being defensive
8961 _overflow_counter = CMSMarkStackOverflowInterval;
8962 return true;
8963 } else {
8964 return false;
8965 }
8966 }
8967
8968 bool CMSCollector::par_simulate_overflow() {
8969 return simulate_overflow();
8970 }
8971 #endif
8972
8973 // Single-threaded
8974 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8975 assert(stack->isEmpty(), "Expected precondition");
8976 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8977 size_t i = num;
8978 oop cur = _overflow_list;
8979 const markOop proto = markOopDesc::prototype();
8980 NOT_PRODUCT(ssize_t n = 0;)
8981 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8982 next = oop(cur->mark());
8983 cur->set_mark(proto); // until proven otherwise
8984 assert(cur->is_oop(), "Should be an oop");
8985 bool res = stack->push(cur);
8986 assert(res, "Bit off more than can chew?");
8987 NOT_PRODUCT(n++;)
8988 }
8989 _overflow_list = cur;
8990 #ifndef PRODUCT
8991 assert(_num_par_pushes >= n, "Too many pops?");
8992 _num_par_pushes -=n;
8993 #endif
8994 return !stack->isEmpty();
8995 }
8996
8997 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
8998 // (MT-safe) Get a prefix of at most "num" from the list.
8999 // The overflow list is chained through the mark word of
9000 // each object in the list. We fetch the entire list,
9001 // break off a prefix of the right size and return the
9002 // remainder. If other threads try to take objects from
9003 // the overflow list at that time, they will wait for
9004 // some time to see if data becomes available. If (and
9005 // only if) another thread places one or more object(s)
9006 // on the global list before we have returned the suffix
9007 // to the global list, we will walk down our local list
9008 // to find its end and append the global list to
9009 // our suffix before returning it. This suffix walk can
9010 // prove to be expensive (quadratic in the amount of traffic)
9011 // when there are many objects in the overflow list and
9012 // there is much producer-consumer contention on the list.
9013 // *NOTE*: The overflow list manipulation code here and
9014 // in ParNewGeneration:: are very similar in shape,
9015 // except that in the ParNew case we use the old (from/eden)
9016 // copy of the object to thread the list via its klass word.
9017 // Because of the common code, if you make any changes in
9018 // the code below, please check the ParNew version to see if
9019 // similar changes might be needed.
9020 // CR 6797058 has been filed to consolidate the common code.
9021 bool CMSCollector::par_take_from_overflow_list(size_t num,
9022 OopTaskQueue* work_q,
9023 int no_of_gc_threads) {
9024 assert(work_q->size() == 0, "First empty local work queue");
9025 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
9026 if (_overflow_list == NULL) {
9027 return false;
9028 }
9029 // Grab the entire list; we'll put back a suffix
9030 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
9031 Thread* tid = Thread::current();
9032 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
9033 // set to ParallelGCThreads.
9034 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
9035 size_t sleep_time_millis = MAX2((size_t)1, num/100);
9036 // If the list is busy, we spin for a short while,
9037 // sleeping between attempts to get the list.
9038 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
9039 os::sleep(tid, sleep_time_millis, false);
9040 if (_overflow_list == NULL) {
9041 // Nothing left to take
9042 return false;
9043 } else if (_overflow_list != BUSY) {
9044 // Try and grab the prefix
9045 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
9046 }
9047 }
9048 // If the list was found to be empty, or we spun long
9049 // enough, we give up and return empty-handed. If we leave
9050 // the list in the BUSY state below, it must be the case that
9051 // some other thread holds the overflow list and will set it
9052 // to a non-BUSY state in the future.
9053 if (prefix == NULL || prefix == BUSY) {
9054 // Nothing to take or waited long enough
9055 if (prefix == NULL) {
9056 // Write back the NULL in case we overwrote it with BUSY above
9057 // and it is still the same value.
9058 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
9059 }
9060 return false;
9061 }
9062 assert(prefix != NULL && prefix != BUSY, "Error");
9063 size_t i = num;
9064 oop cur = prefix;
9065 // Walk down the first "num" objects, unless we reach the end.
9066 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
9067 if (cur->mark() == NULL) {
9068 // We have "num" or fewer elements in the list, so there
9069 // is nothing to return to the global list.
9070 // Write back the NULL in lieu of the BUSY we wrote
9071 // above, if it is still the same value.
9072 if (_overflow_list == BUSY) {
9073 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
9074 }
9075 } else {
9076 // Chop off the suffix and return it to the global list.
9077 assert(cur->mark() != BUSY, "Error");
9078 oop suffix_head = cur->mark(); // suffix will be put back on global list
9079 cur->set_mark(NULL); // break off suffix
9080 // It's possible that the list is still in the empty(busy) state
9081 // we left it in a short while ago; in that case we may be
9082 // able to place back the suffix without incurring the cost
9083 // of a walk down the list.
9084 oop observed_overflow_list = _overflow_list;
9085 oop cur_overflow_list = observed_overflow_list;
9086 bool attached = false;
9087 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
9088 observed_overflow_list =
9089 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
9090 if (cur_overflow_list == observed_overflow_list) {
9091 attached = true;
9092 break;
9093 } else cur_overflow_list = observed_overflow_list;
9094 }
9095 if (!attached) {
9096 // Too bad, someone else sneaked in (at least) an element; we'll need
9097 // to do a splice. Find tail of suffix so we can prepend suffix to global
9098 // list.
9099 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
9100 oop suffix_tail = cur;
9101 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
9102 "Tautology");
9103 observed_overflow_list = _overflow_list;
9104 do {
9105 cur_overflow_list = observed_overflow_list;
9106 if (cur_overflow_list != BUSY) {
9107 // Do the splice ...
9108 suffix_tail->set_mark(markOop(cur_overflow_list));
9109 } else { // cur_overflow_list == BUSY
9110 suffix_tail->set_mark(NULL);
9111 }
9112 // ... and try to place spliced list back on overflow_list ...
9113 observed_overflow_list =
9114 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
9115 } while (cur_overflow_list != observed_overflow_list);
9116 // ... until we have succeeded in doing so.
9117 }
9118 }
9119
9120 // Push the prefix elements on work_q
9121 assert(prefix != NULL, "control point invariant");
9122 const markOop proto = markOopDesc::prototype();
9123 oop next;
9124 NOT_PRODUCT(ssize_t n = 0;)
9125 for (cur = prefix; cur != NULL; cur = next) {
9126 next = oop(cur->mark());
9127 cur->set_mark(proto); // until proven otherwise
9128 assert(cur->is_oop(), "Should be an oop");
9129 bool res = work_q->push(cur);
9130 assert(res, "Bit off more than we can chew?");
9131 NOT_PRODUCT(n++;)
9132 }
9133 #ifndef PRODUCT
9134 assert(_num_par_pushes >= n, "Too many pops?");
9135 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
9136 #endif
9137 return true;
9138 }
9139
9140 // Single-threaded
9141 void CMSCollector::push_on_overflow_list(oop p) {
9142 NOT_PRODUCT(_num_par_pushes++;)
9143 assert(p->is_oop(), "Not an oop");
9144 preserve_mark_if_necessary(p);
9145 p->set_mark((markOop)_overflow_list);
9146 _overflow_list = p;
9147 }
9148
9149 // Multi-threaded; use CAS to prepend to overflow list
9150 void CMSCollector::par_push_on_overflow_list(oop p) {
9151 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
9152 assert(p->is_oop(), "Not an oop");
9153 par_preserve_mark_if_necessary(p);
9154 oop observed_overflow_list = _overflow_list;
9155 oop cur_overflow_list;
9156 do {
9157 cur_overflow_list = observed_overflow_list;
9158 if (cur_overflow_list != BUSY) {
9159 p->set_mark(markOop(cur_overflow_list));
9160 } else {
9161 p->set_mark(NULL);
9162 }
9163 observed_overflow_list =
9164 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
9165 } while (cur_overflow_list != observed_overflow_list);
9166 }
9167 #undef BUSY
9168
9169 // Single threaded
9170 // General Note on GrowableArray: pushes may silently fail
9171 // because we are (temporarily) out of C-heap for expanding
9172 // the stack. The problem is quite ubiquitous and affects
9173 // a lot of code in the JVM. The prudent thing for GrowableArray
9174 // to do (for now) is to exit with an error. However, that may
9175 // be too draconian in some cases because the caller may be
9176 // able to recover without much harm. For such cases, we
9177 // should probably introduce a "soft_push" method which returns
9178 // an indication of success or failure with the assumption that
9179 // the caller may be able to recover from a failure; code in
9180 // the VM can then be changed, incrementally, to deal with such
9181 // failures where possible, thus, incrementally hardening the VM
9182 // in such low resource situations.
9183 void CMSCollector::preserve_mark_work(oop p, markOop m) {
9184 _preserved_oop_stack.push(p);
9185 _preserved_mark_stack.push(m);
9186 assert(m == p->mark(), "Mark word changed");
9187 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9188 "bijection");
9189 }
9190
9191 // Single threaded
9192 void CMSCollector::preserve_mark_if_necessary(oop p) {
9193 markOop m = p->mark();
9194 if (m->must_be_preserved(p)) {
9195 preserve_mark_work(p, m);
9196 }
9197 }
9198
9199 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
9200 markOop m = p->mark();
9201 if (m->must_be_preserved(p)) {
9202 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
9203 // Even though we read the mark word without holding
9204 // the lock, we are assured that it will not change
9205 // because we "own" this oop, so no other thread can
9206 // be trying to push it on the overflow list; see
9207 // the assertion in preserve_mark_work() that checks
9208 // that m == p->mark().
9209 preserve_mark_work(p, m);
9210 }
9211 }
9212
9213 // We should be able to do this multi-threaded,
9214 // a chunk of stack being a task (this is
9215 // correct because each oop only ever appears
9216 // once in the overflow list. However, it's
9217 // not very easy to completely overlap this with
9218 // other operations, so will generally not be done
9219 // until all work's been completed. Because we
9220 // expect the preserved oop stack (set) to be small,
9221 // it's probably fine to do this single-threaded.
9222 // We can explore cleverer concurrent/overlapped/parallel
9223 // processing of preserved marks if we feel the
9224 // need for this in the future. Stack overflow should
9225 // be so rare in practice and, when it happens, its
9226 // effect on performance so great that this will
9227 // likely just be in the noise anyway.
9228 void CMSCollector::restore_preserved_marks_if_any() {
9229 assert(SafepointSynchronize::is_at_safepoint(),
9230 "world should be stopped");
9231 assert(Thread::current()->is_ConcurrentGC_thread() ||
9232 Thread::current()->is_VM_thread(),
9233 "should be single-threaded");
9234 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9235 "bijection");
9236
9237 while (!_preserved_oop_stack.is_empty()) {
9238 oop p = _preserved_oop_stack.pop();
9239 assert(p->is_oop(), "Should be an oop");
9240 assert(_span.contains(p), "oop should be in _span");
9241 assert(p->mark() == markOopDesc::prototype(),
9242 "Set when taken from overflow list");
9243 markOop m = _preserved_mark_stack.pop();
9244 p->set_mark(m);
9245 }
9246 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
9247 "stacks were cleared above");
9248 }
9249
9250 #ifndef PRODUCT
9251 bool CMSCollector::no_preserved_marks() const {
9252 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
9253 }
9254 #endif
9255
9256 // Transfer some number of overflown objects to usual marking
9257 // stack. Return true if some objects were transferred.
9258 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
9259 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
9260 (size_t)ParGCDesiredObjsFromOverflowList);
9261
9262 bool res = _collector->take_from_overflow_list(num, _mark_stack);
9263 assert(_collector->overflow_list_is_empty() || res,
9264 "If list is not empty, we should have taken something");
9265 assert(!res || !_mark_stack->isEmpty(),
9266 "If we took something, it should now be on our stack");
9267 return res;
9268 }
9269
9270 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
9271 size_t res = _sp->block_size_no_stall(addr, _collector);
9272 if (_sp->block_is_obj(addr)) {
9273 if (_live_bit_map->isMarked(addr)) {
9274 // It can't have been dead in a previous cycle
9275 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
9276 } else {
9277 _dead_bit_map->mark(addr); // mark the dead object
9278 }
9279 }
9280 // Could be 0, if the block size could not be computed without stalling.
9281 return res;
9282 }
9283
9284 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
9285
9286 switch (phase) {
9287 case CMSCollector::InitialMarking:
9288 initialize(true /* fullGC */ ,
9289 cause /* cause of the GC */,
9290 true /* recordGCBeginTime */,
9291 true /* recordPreGCUsage */,
9292 false /* recordPeakUsage */,
9293 false /* recordPostGCusage */,
9294 true /* recordAccumulatedGCTime */,
9295 false /* recordGCEndTime */,
9296 false /* countCollection */ );
9297 break;
9298
9299 case CMSCollector::FinalMarking:
9300 initialize(true /* fullGC */ ,
9301 cause /* cause of the GC */,
9302 false /* recordGCBeginTime */,
9303 false /* recordPreGCUsage */,
9304 false /* recordPeakUsage */,
9305 false /* recordPostGCusage */,
9306 true /* recordAccumulatedGCTime */,
9307 false /* recordGCEndTime */,
9308 false /* countCollection */ );
9309 break;
9310
9311 case CMSCollector::Sweeping:
9312 initialize(true /* fullGC */ ,
9313 cause /* cause of the GC */,
9314 false /* recordGCBeginTime */,
9315 false /* recordPreGCUsage */,
9316 true /* recordPeakUsage */,
9317 true /* recordPostGCusage */,
9318 false /* recordAccumulatedGCTime */,
9319 true /* recordGCEndTime */,
9320 true /* countCollection */ );
9321 break;
9322
9323 default:
9324 ShouldNotReachHere();
9325 }
9326 }