1 /*
2 * Copyright (c) 2001, 2013, 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/symbolTable.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "code/codeCache.hpp"
30 #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp"
31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp"
32 #include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.hpp"
33 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp"
34 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp"
35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp"
36 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp"
37 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp"
38 #include "gc_implementation/parNew/parNewGeneration.hpp"
39 #include "gc_implementation/shared/collectorCounters.hpp"
40 #include "gc_implementation/shared/gcTimer.hpp"
41 #include "gc_implementation/shared/gcTrace.hpp"
42 #include "gc_implementation/shared/gcTraceTime.hpp"
43 #include "gc_implementation/shared/isGCActiveMark.hpp"
44 #include "gc_interface/collectedHeap.inline.hpp"
45 #include "memory/allocation.hpp"
46 #include "memory/cardTableRS.hpp"
47 #include "memory/collectorPolicy.hpp"
48 #include "memory/gcLocker.inline.hpp"
49 #include "memory/genCollectedHeap.hpp"
50 #include "memory/genMarkSweep.hpp"
51 #include "memory/genOopClosures.inline.hpp"
52 #include "memory/iterator.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/globals_extension.hpp"
59 #include "runtime/handles.inline.hpp"
60 #include "runtime/java.hpp"
61 #include "runtime/vmThread.hpp"
62 #include "services/memoryService.hpp"
63 #include "services/runtimeService.hpp"
64
65 // statics
66 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
67 bool CMSCollector::_full_gc_requested = false;
68 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
69
70 //////////////////////////////////////////////////////////////////
71 // In support of CMS/VM thread synchronization
72 //////////////////////////////////////////////////////////////////
73 // We split use of the CGC_lock into 2 "levels".
74 // The low-level locking is of the usual CGC_lock monitor. We introduce
75 // a higher level "token" (hereafter "CMS token") built on top of the
76 // low level monitor (hereafter "CGC lock").
77 // The token-passing protocol gives priority to the VM thread. The
78 // CMS-lock doesn't provide any fairness guarantees, but clients
79 // should ensure that it is only held for very short, bounded
80 // durations.
81 //
82 // When either of the CMS thread or the VM thread is involved in
83 // collection operations during which it does not want the other
84 // thread to interfere, it obtains the CMS token.
85 //
86 // If either thread tries to get the token while the other has
87 // it, that thread waits. However, if the VM thread and CMS thread
88 // both want the token, then the VM thread gets priority while the
89 // CMS thread waits. This ensures, for instance, that the "concurrent"
90 // phases of the CMS thread's work do not block out the VM thread
91 // for long periods of time as the CMS thread continues to hog
92 // the token. (See bug 4616232).
93 //
94 // The baton-passing functions are, however, controlled by the
95 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
96 // and here the low-level CMS lock, not the high level token,
97 // ensures mutual exclusion.
98 //
99 // Two important conditions that we have to satisfy:
100 // 1. if a thread does a low-level wait on the CMS lock, then it
101 // relinquishes the CMS token if it were holding that token
102 // when it acquired the low-level CMS lock.
103 // 2. any low-level notifications on the low-level lock
104 // should only be sent when a thread has relinquished the token.
105 //
106 // In the absence of either property, we'd have potential deadlock.
107 //
108 // We protect each of the CMS (concurrent and sequential) phases
109 // with the CMS _token_, not the CMS _lock_.
110 //
111 // The only code protected by CMS lock is the token acquisition code
112 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
113 // baton-passing code.
114 //
115 // Unfortunately, i couldn't come up with a good abstraction to factor and
116 // hide the naked CGC_lock manipulation in the baton-passing code
117 // further below. That's something we should try to do. Also, the proof
118 // of correctness of this 2-level locking scheme is far from obvious,
119 // and potentially quite slippery. We have an uneasy supsicion, for instance,
120 // that there may be a theoretical possibility of delay/starvation in the
121 // low-level lock/wait/notify scheme used for the baton-passing because of
122 // potential intereference with the priority scheme embodied in the
123 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
124 // invocation further below and marked with "XXX 20011219YSR".
125 // Indeed, as we note elsewhere, this may become yet more slippery
126 // in the presence of multiple CMS and/or multiple VM threads. XXX
127
128 class CMSTokenSync: public StackObj {
129 private:
130 bool _is_cms_thread;
131 public:
132 CMSTokenSync(bool is_cms_thread):
133 _is_cms_thread(is_cms_thread) {
134 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
135 "Incorrect argument to constructor");
136 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
137 }
138
139 ~CMSTokenSync() {
140 assert(_is_cms_thread ?
141 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
142 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
143 "Incorrect state");
144 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
145 }
146 };
147
148 // Convenience class that does a CMSTokenSync, and then acquires
149 // upto three locks.
150 class CMSTokenSyncWithLocks: public CMSTokenSync {
151 private:
152 // Note: locks are acquired in textual declaration order
153 // and released in the opposite order
154 MutexLockerEx _locker1, _locker2, _locker3;
155 public:
156 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
157 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
158 CMSTokenSync(is_cms_thread),
159 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
160 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
161 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
162 { }
163 };
164
165
166 // Wrapper class to temporarily disable icms during a foreground cms collection.
167 class ICMSDisabler: public StackObj {
168 public:
169 // The ctor disables icms and wakes up the thread so it notices the change;
170 // the dtor re-enables icms. Note that the CMSCollector methods will check
171 // CMSIncrementalMode.
172 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); }
173 ~ICMSDisabler() { CMSCollector::enable_icms(); }
174 };
175
176 //////////////////////////////////////////////////////////////////
177 // Concurrent Mark-Sweep Generation /////////////////////////////
178 //////////////////////////////////////////////////////////////////
179
180 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
181
182 // This struct contains per-thread things necessary to support parallel
183 // young-gen collection.
184 class CMSParGCThreadState: public CHeapObj<mtGC> {
185 public:
186 CFLS_LAB lab;
187 PromotionInfo promo;
188
189 // Constructor.
190 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
191 promo.setSpace(cfls);
192 }
193 };
194
195 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
196 ReservedSpace rs, size_t initial_byte_size, int level,
197 CardTableRS* ct, bool use_adaptive_freelists,
198 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) :
199 CardGeneration(rs, initial_byte_size, level, ct),
200 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
201 _debug_collection_type(Concurrent_collection_type),
202 _did_compact(false)
203 {
204 HeapWord* bottom = (HeapWord*) _virtual_space.low();
205 HeapWord* end = (HeapWord*) _virtual_space.high();
206
207 _direct_allocated_words = 0;
208 NOT_PRODUCT(
209 _numObjectsPromoted = 0;
210 _numWordsPromoted = 0;
211 _numObjectsAllocated = 0;
212 _numWordsAllocated = 0;
213 )
214
215 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
216 use_adaptive_freelists,
217 dictionaryChoice);
218 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
219 if (_cmsSpace == NULL) {
220 vm_exit_during_initialization(
221 "CompactibleFreeListSpace allocation failure");
222 }
223 _cmsSpace->_gen = this;
224
225 _gc_stats = new CMSGCStats();
226
227 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
228 // offsets match. The ability to tell free chunks from objects
229 // depends on this property.
230 debug_only(
231 FreeChunk* junk = NULL;
232 assert(UseCompressedKlassPointers ||
233 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
234 "Offset of FreeChunk::_prev within FreeChunk must match"
235 " that of OopDesc::_klass within OopDesc");
236 )
237 if (CollectedHeap::use_parallel_gc_threads()) {
238 typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
239 _par_gc_thread_states =
240 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC);
241 if (_par_gc_thread_states == NULL) {
242 vm_exit_during_initialization("Could not allocate par gc structs");
243 }
244 for (uint i = 0; i < ParallelGCThreads; i++) {
245 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
246 if (_par_gc_thread_states[i] == NULL) {
247 vm_exit_during_initialization("Could not allocate par gc structs");
248 }
249 }
250 } else {
251 _par_gc_thread_states = NULL;
252 }
253 _incremental_collection_failed = false;
254 // The "dilatation_factor" is the expansion that can occur on
255 // account of the fact that the minimum object size in the CMS
256 // generation may be larger than that in, say, a contiguous young
257 // generation.
258 // Ideally, in the calculation below, we'd compute the dilatation
259 // factor as: MinChunkSize/(promoting_gen's min object size)
260 // Since we do not have such a general query interface for the
261 // promoting generation, we'll instead just use the mimimum
262 // object size (which today is a header's worth of space);
263 // note that all arithmetic is in units of HeapWords.
264 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
265 assert(_dilatation_factor >= 1.0, "from previous assert");
266 }
267
268
269 // The field "_initiating_occupancy" represents the occupancy percentage
270 // at which we trigger a new collection cycle. Unless explicitly specified
271 // via CMSInitiatingOccupancyFraction (argument "io" below), it
272 // is calculated by:
273 //
274 // Let "f" be MinHeapFreeRatio in
275 //
276 // _intiating_occupancy = 100-f +
277 // f * (CMSTriggerRatio/100)
278 // where CMSTriggerRatio is the argument "tr" below.
279 //
280 // That is, if we assume the heap is at its desired maximum occupancy at the
281 // end of a collection, we let CMSTriggerRatio of the (purported) free
282 // space be allocated before initiating a new collection cycle.
283 //
284 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
285 assert(io <= 100 && tr <= 100, "Check the arguments");
286 if (io >= 0) {
287 _initiating_occupancy = (double)io / 100.0;
288 } else {
289 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
290 (double)(tr * MinHeapFreeRatio) / 100.0)
291 / 100.0;
292 }
293 }
294
295 void ConcurrentMarkSweepGeneration::ref_processor_init() {
296 assert(collector() != NULL, "no collector");
297 collector()->ref_processor_init();
298 }
299
300 void CMSCollector::ref_processor_init() {
301 if (_ref_processor == NULL) {
302 // Allocate and initialize a reference processor
303 _ref_processor =
304 new ReferenceProcessor(_span, // span
305 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
306 (int) ParallelGCThreads, // mt processing degree
307 _cmsGen->refs_discovery_is_mt(), // mt discovery
308 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
309 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
310 &_is_alive_closure, // closure for liveness info
311 false); // next field updates do not need write barrier
312 // Initialize the _ref_processor field of CMSGen
313 _cmsGen->set_ref_processor(_ref_processor);
314
315 }
316 }
317
318 CMSAdaptiveSizePolicy* CMSCollector::size_policy() {
319 GenCollectedHeap* gch = GenCollectedHeap::heap();
320 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
321 "Wrong type of heap");
322 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
323 gch->gen_policy()->size_policy();
324 assert(sp->is_gc_cms_adaptive_size_policy(),
325 "Wrong type of size policy");
326 return sp;
327 }
328
329 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() {
330 CMSGCAdaptivePolicyCounters* results =
331 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters();
332 assert(
333 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
334 "Wrong gc policy counter kind");
335 return results;
336 }
337
338
339 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
340
341 const char* gen_name = "old";
342
343 // Generation Counters - generation 1, 1 subspace
344 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space);
345
346 _space_counters = new GSpaceCounters(gen_name, 0,
347 _virtual_space.reserved_size(),
348 this, _gen_counters);
349 }
350
351 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
352 _cms_gen(cms_gen)
353 {
354 assert(alpha <= 100, "bad value");
355 _saved_alpha = alpha;
356
357 // Initialize the alphas to the bootstrap value of 100.
358 _gc0_alpha = _cms_alpha = 100;
359
360 _cms_begin_time.update();
361 _cms_end_time.update();
362
363 _gc0_duration = 0.0;
364 _gc0_period = 0.0;
365 _gc0_promoted = 0;
366
367 _cms_duration = 0.0;
368 _cms_period = 0.0;
369 _cms_allocated = 0;
370
371 _cms_used_at_gc0_begin = 0;
372 _cms_used_at_gc0_end = 0;
373 _allow_duty_cycle_reduction = false;
374 _valid_bits = 0;
375 _icms_duty_cycle = CMSIncrementalDutyCycle;
376 }
377
378 double CMSStats::cms_free_adjustment_factor(size_t free) const {
379 // TBD: CR 6909490
380 return 1.0;
381 }
382
383 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
384 }
385
386 // If promotion failure handling is on use
387 // the padded average size of the promotion for each
388 // young generation collection.
389 double CMSStats::time_until_cms_gen_full() const {
390 size_t cms_free = _cms_gen->cmsSpace()->free();
391 GenCollectedHeap* gch = GenCollectedHeap::heap();
392 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(),
393 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
394 if (cms_free > expected_promotion) {
395 // Start a cms collection if there isn't enough space to promote
396 // for the next minor collection. Use the padded average as
397 // a safety factor.
398 cms_free -= expected_promotion;
399
400 // Adjust by the safety factor.
401 double cms_free_dbl = (double)cms_free;
402 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
403 // Apply a further correction factor which tries to adjust
404 // for recent occurance of concurrent mode failures.
405 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
406 cms_free_dbl = cms_free_dbl * cms_adjustment;
407
408 if (PrintGCDetails && Verbose) {
409 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
410 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
411 cms_free, expected_promotion);
412 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
413 cms_free_dbl, cms_consumption_rate() + 1.0);
414 }
415 // Add 1 in case the consumption rate goes to zero.
416 return cms_free_dbl / (cms_consumption_rate() + 1.0);
417 }
418 return 0.0;
419 }
420
421 // Compare the duration of the cms collection to the
422 // time remaining before the cms generation is empty.
423 // Note that the time from the start of the cms collection
424 // to the start of the cms sweep (less than the total
425 // duration of the cms collection) can be used. This
426 // has been tried and some applications experienced
427 // promotion failures early in execution. This was
428 // possibly because the averages were not accurate
429 // enough at the beginning.
430 double CMSStats::time_until_cms_start() const {
431 // We add "gc0_period" to the "work" calculation
432 // below because this query is done (mostly) at the
433 // end of a scavenge, so we need to conservatively
434 // account for that much possible delay
435 // in the query so as to avoid concurrent mode failures
436 // due to starting the collection just a wee bit too
437 // late.
438 double work = cms_duration() + gc0_period();
439 double deadline = time_until_cms_gen_full();
440 // If a concurrent mode failure occurred recently, we want to be
441 // more conservative and halve our expected time_until_cms_gen_full()
442 if (work > deadline) {
443 if (Verbose && PrintGCDetails) {
444 gclog_or_tty->print(
445 " CMSCollector: collect because of anticipated promotion "
446 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
447 gc0_period(), time_until_cms_gen_full());
448 }
449 return 0.0;
450 }
451 return work - deadline;
452 }
453
454 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the
455 // amount of change to prevent wild oscillation.
456 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle,
457 unsigned int new_duty_cycle) {
458 assert(old_duty_cycle <= 100, "bad input value");
459 assert(new_duty_cycle <= 100, "bad input value");
460
461 // Note: use subtraction with caution since it may underflow (values are
462 // unsigned). Addition is safe since we're in the range 0-100.
463 unsigned int damped_duty_cycle = new_duty_cycle;
464 if (new_duty_cycle < old_duty_cycle) {
465 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U);
466 if (new_duty_cycle + largest_delta < old_duty_cycle) {
467 damped_duty_cycle = old_duty_cycle - largest_delta;
468 }
469 } else if (new_duty_cycle > old_duty_cycle) {
470 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U);
471 if (new_duty_cycle > old_duty_cycle + largest_delta) {
472 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U);
473 }
474 }
475 assert(damped_duty_cycle <= 100, "invalid duty cycle computed");
476
477 if (CMSTraceIncrementalPacing) {
478 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ",
479 old_duty_cycle, new_duty_cycle, damped_duty_cycle);
480 }
481 return damped_duty_cycle;
482 }
483
484 unsigned int CMSStats::icms_update_duty_cycle_impl() {
485 assert(CMSIncrementalPacing && valid(),
486 "should be handled in icms_update_duty_cycle()");
487
488 double cms_time_so_far = cms_timer().seconds();
489 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M;
490 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far);
491
492 // Avoid division by 0.
493 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01);
494 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full;
495
496 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U);
497 if (new_duty_cycle > _icms_duty_cycle) {
498 // Avoid very small duty cycles (1 or 2); 0 is allowed.
499 if (new_duty_cycle > 2) {
500 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle,
501 new_duty_cycle);
502 }
503 } else if (_allow_duty_cycle_reduction) {
504 // The duty cycle is reduced only once per cms cycle (see record_cms_end()).
505 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle);
506 // Respect the minimum duty cycle.
507 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin;
508 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle);
509 }
510
511 if (PrintGCDetails || CMSTraceIncrementalPacing) {
512 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle);
513 }
514
515 _allow_duty_cycle_reduction = false;
516 return _icms_duty_cycle;
517 }
518
519 #ifndef PRODUCT
520 void CMSStats::print_on(outputStream *st) const {
521 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
522 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
523 gc0_duration(), gc0_period(), gc0_promoted());
524 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
525 cms_duration(), cms_duration_per_mb(),
526 cms_period(), cms_allocated());
527 st->print(",cms_since_beg=%g,cms_since_end=%g",
528 cms_time_since_begin(), cms_time_since_end());
529 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
530 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
531 if (CMSIncrementalMode) {
532 st->print(",dc=%d", icms_duty_cycle());
533 }
534
535 if (valid()) {
536 st->print(",promo_rate=%g,cms_alloc_rate=%g",
537 promotion_rate(), cms_allocation_rate());
538 st->print(",cms_consumption_rate=%g,time_until_full=%g",
539 cms_consumption_rate(), time_until_cms_gen_full());
540 }
541 st->print(" ");
542 }
543 #endif // #ifndef PRODUCT
544
545 CMSCollector::CollectorState CMSCollector::_collectorState =
546 CMSCollector::Idling;
547 bool CMSCollector::_foregroundGCIsActive = false;
548 bool CMSCollector::_foregroundGCShouldWait = false;
549
550 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
551 CardTableRS* ct,
552 ConcurrentMarkSweepPolicy* cp):
553 _cmsGen(cmsGen),
554 _ct(ct),
555 _ref_processor(NULL), // will be set later
556 _conc_workers(NULL), // may be set later
557 _abort_preclean(false),
558 _start_sampling(false),
559 _between_prologue_and_epilogue(false),
560 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
561 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
562 -1 /* lock-free */, "No_lock" /* dummy */),
563 _modUnionClosure(&_modUnionTable),
564 _modUnionClosurePar(&_modUnionTable),
565 // Adjust my span to cover old (cms) gen
566 _span(cmsGen->reserved()),
567 // Construct the is_alive_closure with _span & markBitMap
568 _is_alive_closure(_span, &_markBitMap),
569 _restart_addr(NULL),
570 _overflow_list(NULL),
571 _stats(cmsGen),
572 _eden_chunk_array(NULL), // may be set in ctor body
573 _eden_chunk_capacity(0), // -- ditto --
574 _eden_chunk_index(0), // -- ditto --
575 _survivor_plab_array(NULL), // -- ditto --
576 _survivor_chunk_array(NULL), // -- ditto --
577 _survivor_chunk_capacity(0), // -- ditto --
578 _survivor_chunk_index(0), // -- ditto --
579 _ser_pmc_preclean_ovflw(0),
580 _ser_kac_preclean_ovflw(0),
581 _ser_pmc_remark_ovflw(0),
582 _par_pmc_remark_ovflw(0),
583 _ser_kac_ovflw(0),
584 _par_kac_ovflw(0),
585 #ifndef PRODUCT
586 _num_par_pushes(0),
587 #endif
588 _collection_count_start(0),
589 _verifying(false),
590 _icms_start_limit(NULL),
591 _icms_stop_limit(NULL),
592 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
593 _completed_initialization(false),
594 _collector_policy(cp),
595 _should_unload_classes(false),
596 _concurrent_cycles_since_last_unload(0),
597 _roots_scanning_options(0),
598 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
599 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
600 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
601 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
602 _cms_start_registered(false)
603 {
604 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
605 ExplicitGCInvokesConcurrent = true;
606 }
607 // Now expand the span and allocate the collection support structures
608 // (MUT, marking bit map etc.) to cover both generations subject to
609 // collection.
610
611 // For use by dirty card to oop closures.
612 _cmsGen->cmsSpace()->set_collector(this);
613
614 // Allocate MUT and marking bit map
615 {
616 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
617 if (!_markBitMap.allocate(_span)) {
618 warning("Failed to allocate CMS Bit Map");
619 return;
620 }
621 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
622 }
623 {
624 _modUnionTable.allocate(_span);
625 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
626 }
627
628 if (!_markStack.allocate(MarkStackSize)) {
629 warning("Failed to allocate CMS Marking Stack");
630 return;
631 }
632
633 // Support for multi-threaded concurrent phases
634 if (CMSConcurrentMTEnabled) {
635 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
636 // just for now
637 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
638 }
639 if (ConcGCThreads > 1) {
640 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
641 ConcGCThreads, true);
642 if (_conc_workers == NULL) {
643 warning("GC/CMS: _conc_workers allocation failure: "
644 "forcing -CMSConcurrentMTEnabled");
645 CMSConcurrentMTEnabled = false;
646 } else {
647 _conc_workers->initialize_workers();
648 }
649 } else {
650 CMSConcurrentMTEnabled = false;
651 }
652 }
653 if (!CMSConcurrentMTEnabled) {
654 ConcGCThreads = 0;
655 } else {
656 // Turn off CMSCleanOnEnter optimization temporarily for
657 // the MT case where it's not fixed yet; see 6178663.
658 CMSCleanOnEnter = false;
659 }
660 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
661 "Inconsistency");
662
663 // Parallel task queues; these are shared for the
664 // concurrent and stop-world phases of CMS, but
665 // are not shared with parallel scavenge (ParNew).
666 {
667 uint i;
668 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
669
670 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
671 || ParallelRefProcEnabled)
672 && num_queues > 0) {
673 _task_queues = new OopTaskQueueSet(num_queues);
674 if (_task_queues == NULL) {
675 warning("task_queues allocation failure.");
676 return;
677 }
678 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
679 if (_hash_seed == NULL) {
680 warning("_hash_seed array allocation failure");
681 return;
682 }
683
684 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
685 for (i = 0; i < num_queues; i++) {
686 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
687 if (q == NULL) {
688 warning("work_queue allocation failure.");
689 return;
690 }
691 _task_queues->register_queue(i, q);
692 }
693 for (i = 0; i < num_queues; i++) {
694 _task_queues->queue(i)->initialize();
695 _hash_seed[i] = 17; // copied from ParNew
696 }
697 }
698 }
699
700 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
701
702 // Clip CMSBootstrapOccupancy between 0 and 100.
703 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100;
704
705 _full_gcs_since_conc_gc = 0;
706
707 // Now tell CMS generations the identity of their collector
708 ConcurrentMarkSweepGeneration::set_collector(this);
709
710 // Create & start a CMS thread for this CMS collector
711 _cmsThread = ConcurrentMarkSweepThread::start(this);
712 assert(cmsThread() != NULL, "CMS Thread should have been created");
713 assert(cmsThread()->collector() == this,
714 "CMS Thread should refer to this gen");
715 assert(CGC_lock != NULL, "Where's the CGC_lock?");
716
717 // Support for parallelizing young gen rescan
718 GenCollectedHeap* gch = GenCollectedHeap::heap();
719 _young_gen = gch->prev_gen(_cmsGen);
720 if (gch->supports_inline_contig_alloc()) {
721 _top_addr = gch->top_addr();
722 _end_addr = gch->end_addr();
723 assert(_young_gen != NULL, "no _young_gen");
724 _eden_chunk_index = 0;
725 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
726 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
727 if (_eden_chunk_array == NULL) {
728 _eden_chunk_capacity = 0;
729 warning("GC/CMS: _eden_chunk_array allocation failure");
730 }
731 }
732 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
733
734 // Support for parallelizing survivor space rescan
735 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) {
736 const size_t max_plab_samples =
737 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize;
738
739 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
740 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC);
741 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
742 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
743 || _cursor == NULL) {
744 warning("Failed to allocate survivor plab/chunk array");
745 if (_survivor_plab_array != NULL) {
746 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC);
747 _survivor_plab_array = NULL;
748 }
749 if (_survivor_chunk_array != NULL) {
750 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC);
751 _survivor_chunk_array = NULL;
752 }
753 if (_cursor != NULL) {
754 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC);
755 _cursor = NULL;
756 }
757 } else {
758 _survivor_chunk_capacity = 2*max_plab_samples;
759 for (uint i = 0; i < ParallelGCThreads; i++) {
760 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
761 if (vec == NULL) {
762 warning("Failed to allocate survivor plab array");
763 for (int j = i; j > 0; j--) {
764 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC);
765 }
766 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC);
767 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC);
768 _survivor_plab_array = NULL;
769 _survivor_chunk_array = NULL;
770 _survivor_chunk_capacity = 0;
771 break;
772 } else {
773 ChunkArray* cur =
774 ::new (&_survivor_plab_array[i]) ChunkArray(vec,
775 max_plab_samples);
776 assert(cur->end() == 0, "Should be 0");
777 assert(cur->array() == vec, "Should be vec");
778 assert(cur->capacity() == max_plab_samples, "Error");
779 }
780 }
781 }
782 }
783 assert( ( _survivor_plab_array != NULL
784 && _survivor_chunk_array != NULL)
785 || ( _survivor_chunk_capacity == 0
786 && _survivor_chunk_index == 0),
787 "Error");
788
789 // Choose what strong roots should be scanned depending on verification options
790 if (!CMSClassUnloadingEnabled) {
791 // If class unloading is disabled we want to include all classes into the root set.
792 add_root_scanning_option(SharedHeap::SO_AllClasses);
793 } else {
794 add_root_scanning_option(SharedHeap::SO_SystemClasses);
795 }
796
797 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
798 _gc_counters = new CollectorCounters("CMS", 1);
799 _completed_initialization = true;
800 _inter_sweep_timer.start(); // start of time
801 }
802
803 const char* ConcurrentMarkSweepGeneration::name() const {
804 return "concurrent mark-sweep generation";
805 }
806 void ConcurrentMarkSweepGeneration::update_counters() {
807 if (UsePerfData) {
808 _space_counters->update_all();
809 _gen_counters->update_all();
810 }
811 }
812
813 // this is an optimized version of update_counters(). it takes the
814 // used value as a parameter rather than computing it.
815 //
816 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
817 if (UsePerfData) {
818 _space_counters->update_used(used);
819 _space_counters->update_capacity();
820 _gen_counters->update_all();
821 }
822 }
823
824 void ConcurrentMarkSweepGeneration::print() const {
825 Generation::print();
826 cmsSpace()->print();
827 }
828
829 #ifndef PRODUCT
830 void ConcurrentMarkSweepGeneration::print_statistics() {
831 cmsSpace()->printFLCensus(0);
832 }
833 #endif
834
835 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
836 GenCollectedHeap* gch = GenCollectedHeap::heap();
837 if (PrintGCDetails) {
838 if (Verbose) {
839 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
840 level(), short_name(), s, used(), capacity());
841 } else {
842 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
843 level(), short_name(), s, used() / K, capacity() / K);
844 }
845 }
846 if (Verbose) {
847 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
848 gch->used(), gch->capacity());
849 } else {
850 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
851 gch->used() / K, gch->capacity() / K);
852 }
853 }
854
855 size_t
856 ConcurrentMarkSweepGeneration::contiguous_available() const {
857 // dld proposes an improvement in precision here. If the committed
858 // part of the space ends in a free block we should add that to
859 // uncommitted size in the calculation below. Will make this
860 // change later, staying with the approximation below for the
861 // time being. -- ysr.
862 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
863 }
864
865 size_t
866 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
867 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
868 }
869
870 size_t ConcurrentMarkSweepGeneration::max_available() const {
871 return free() + _virtual_space.uncommitted_size();
872 }
873
874 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
875 size_t available = max_available();
876 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
877 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
878 if (Verbose && PrintGCDetails) {
879 gclog_or_tty->print_cr(
880 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
881 "max_promo("SIZE_FORMAT")",
882 res? "":" not", available, res? ">=":"<",
883 av_promo, max_promotion_in_bytes);
884 }
885 return res;
886 }
887
888 // At a promotion failure dump information on block layout in heap
889 // (cms old generation).
890 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
891 if (CMSDumpAtPromotionFailure) {
892 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
893 }
894 }
895
896 CompactibleSpace*
897 ConcurrentMarkSweepGeneration::first_compaction_space() const {
898 return _cmsSpace;
899 }
900
901 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
902 // Clear the promotion information. These pointers can be adjusted
903 // along with all the other pointers into the heap but
904 // compaction is expected to be a rare event with
905 // a heap using cms so don't do it without seeing the need.
906 if (CollectedHeap::use_parallel_gc_threads()) {
907 for (uint i = 0; i < ParallelGCThreads; i++) {
908 _par_gc_thread_states[i]->promo.reset();
909 }
910 }
911 }
912
913 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
914 blk->do_space(_cmsSpace);
915 }
916
917 void ConcurrentMarkSweepGeneration::compute_new_size() {
918 assert_locked_or_safepoint(Heap_lock);
919
920 // If incremental collection failed, we just want to expand
921 // to the limit.
922 if (incremental_collection_failed()) {
923 clear_incremental_collection_failed();
924 grow_to_reserved();
925 return;
926 }
927
928 // The heap has been compacted but not reset yet.
929 // Any metric such as free() or used() will be incorrect.
930
931 CardGeneration::compute_new_size();
932
933 // Reset again after a possible resizing
934 if (did_compact()) {
935 cmsSpace()->reset_after_compaction();
936 }
937 }
938
939 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
940 assert_locked_or_safepoint(Heap_lock);
941
942 // If incremental collection failed, we just want to expand
943 // to the limit.
944 if (incremental_collection_failed()) {
945 clear_incremental_collection_failed();
946 grow_to_reserved();
947 return;
948 }
949
950 double free_percentage = ((double) free()) / capacity();
951 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
952 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
953
954 // compute expansion delta needed for reaching desired free percentage
955 if (free_percentage < desired_free_percentage) {
956 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
957 assert(desired_capacity >= capacity(), "invalid expansion size");
958 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
959 if (PrintGCDetails && Verbose) {
960 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
961 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
962 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
963 gclog_or_tty->print_cr(" Desired free fraction %f",
964 desired_free_percentage);
965 gclog_or_tty->print_cr(" Maximum free fraction %f",
966 maximum_free_percentage);
967 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000);
968 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
969 desired_capacity/1000);
970 int prev_level = level() - 1;
971 if (prev_level >= 0) {
972 size_t prev_size = 0;
973 GenCollectedHeap* gch = GenCollectedHeap::heap();
974 Generation* prev_gen = gch->_gens[prev_level];
975 prev_size = prev_gen->capacity();
976 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
977 prev_size/1000);
978 }
979 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
980 unsafe_max_alloc_nogc()/1000);
981 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
982 contiguous_available()/1000);
983 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
984 expand_bytes);
985 }
986 // safe if expansion fails
987 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
988 if (PrintGCDetails && Verbose) {
989 gclog_or_tty->print_cr(" Expanded free fraction %f",
990 ((double) free()) / capacity());
991 }
992 } else {
993 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
994 assert(desired_capacity <= capacity(), "invalid expansion size");
995 size_t shrink_bytes = capacity() - desired_capacity;
996 // Don't shrink unless the delta is greater than the minimum shrink we want
997 if (shrink_bytes >= MinHeapDeltaBytes) {
998 shrink_free_list_by(shrink_bytes);
999 }
1000 }
1001 }
1002
1003 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
1004 return cmsSpace()->freelistLock();
1005 }
1006
1007 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
1008 bool tlab) {
1009 CMSSynchronousYieldRequest yr;
1010 MutexLockerEx x(freelistLock(),
1011 Mutex::_no_safepoint_check_flag);
1012 return have_lock_and_allocate(size, tlab);
1013 }
1014
1015 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
1016 bool tlab /* ignored */) {
1017 assert_lock_strong(freelistLock());
1018 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
1019 HeapWord* res = cmsSpace()->allocate(adjustedSize);
1020 // Allocate the object live (grey) if the background collector has
1021 // started marking. This is necessary because the marker may
1022 // have passed this address and consequently this object will
1023 // not otherwise be greyed and would be incorrectly swept up.
1024 // Note that if this object contains references, the writing
1025 // of those references will dirty the card containing this object
1026 // allowing the object to be blackened (and its references scanned)
1027 // either during a preclean phase or at the final checkpoint.
1028 if (res != NULL) {
1029 // We may block here with an uninitialized object with
1030 // its mark-bit or P-bits not yet set. Such objects need
1031 // to be safely navigable by block_start().
1032 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
1033 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
1034 collector()->direct_allocated(res, adjustedSize);
1035 _direct_allocated_words += adjustedSize;
1036 // allocation counters
1037 NOT_PRODUCT(
1038 _numObjectsAllocated++;
1039 _numWordsAllocated += (int)adjustedSize;
1040 )
1041 }
1042 return res;
1043 }
1044
1045 // In the case of direct allocation by mutators in a generation that
1046 // is being concurrently collected, the object must be allocated
1047 // live (grey) if the background collector has started marking.
1048 // This is necessary because the marker may
1049 // have passed this address and consequently this object will
1050 // not otherwise be greyed and would be incorrectly swept up.
1051 // Note that if this object contains references, the writing
1052 // of those references will dirty the card containing this object
1053 // allowing the object to be blackened (and its references scanned)
1054 // either during a preclean phase or at the final checkpoint.
1055 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
1056 assert(_markBitMap.covers(start, size), "Out of bounds");
1057 if (_collectorState >= Marking) {
1058 MutexLockerEx y(_markBitMap.lock(),
1059 Mutex::_no_safepoint_check_flag);
1060 // [see comments preceding SweepClosure::do_blk() below for details]
1061 //
1062 // Can the P-bits be deleted now? JJJ
1063 //
1064 // 1. need to mark the object as live so it isn't collected
1065 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
1066 // 3. need to mark the end of the object so marking, precleaning or sweeping
1067 // can skip over uninitialized or unparsable objects. An allocated
1068 // object is considered uninitialized for our purposes as long as
1069 // its klass word is NULL. All old gen objects are parsable
1070 // as soon as they are initialized.)
1071 _markBitMap.mark(start); // object is live
1072 _markBitMap.mark(start + 1); // object is potentially uninitialized?
1073 _markBitMap.mark(start + size - 1);
1074 // mark end of object
1075 }
1076 // check that oop looks uninitialized
1077 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
1078 }
1079
1080 void CMSCollector::promoted(bool par, HeapWord* start,
1081 bool is_obj_array, size_t obj_size) {
1082 assert(_markBitMap.covers(start), "Out of bounds");
1083 // See comment in direct_allocated() about when objects should
1084 // be allocated live.
1085 if (_collectorState >= Marking) {
1086 // we already hold the marking bit map lock, taken in
1087 // the prologue
1088 if (par) {
1089 _markBitMap.par_mark(start);
1090 } else {
1091 _markBitMap.mark(start);
1092 }
1093 // We don't need to mark the object as uninitialized (as
1094 // in direct_allocated above) because this is being done with the
1095 // world stopped and the object will be initialized by the
1096 // time the marking, precleaning or sweeping get to look at it.
1097 // But see the code for copying objects into the CMS generation,
1098 // where we need to ensure that concurrent readers of the
1099 // block offset table are able to safely navigate a block that
1100 // is in flux from being free to being allocated (and in
1101 // transition while being copied into) and subsequently
1102 // becoming a bona-fide object when the copy/promotion is complete.
1103 assert(SafepointSynchronize::is_at_safepoint(),
1104 "expect promotion only at safepoints");
1105
1106 if (_collectorState < Sweeping) {
1107 // Mark the appropriate cards in the modUnionTable, so that
1108 // this object gets scanned before the sweep. If this is
1109 // not done, CMS generation references in the object might
1110 // not get marked.
1111 // For the case of arrays, which are otherwise precisely
1112 // marked, we need to dirty the entire array, not just its head.
1113 if (is_obj_array) {
1114 // The [par_]mark_range() method expects mr.end() below to
1115 // be aligned to the granularity of a bit's representation
1116 // in the heap. In the case of the MUT below, that's a
1117 // card size.
1118 MemRegion mr(start,
1119 (HeapWord*)round_to((intptr_t)(start + obj_size),
1120 CardTableModRefBS::card_size /* bytes */));
1121 if (par) {
1122 _modUnionTable.par_mark_range(mr);
1123 } else {
1124 _modUnionTable.mark_range(mr);
1125 }
1126 } else { // not an obj array; we can just mark the head
1127 if (par) {
1128 _modUnionTable.par_mark(start);
1129 } else {
1130 _modUnionTable.mark(start);
1131 }
1132 }
1133 }
1134 }
1135 }
1136
1137 static inline size_t percent_of_space(Space* space, HeapWord* addr)
1138 {
1139 size_t delta = pointer_delta(addr, space->bottom());
1140 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
1141 }
1142
1143 void CMSCollector::icms_update_allocation_limits()
1144 {
1145 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
1146 EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
1147
1148 const unsigned int duty_cycle = stats().icms_update_duty_cycle();
1149 if (CMSTraceIncrementalPacing) {
1150 stats().print();
1151 }
1152
1153 assert(duty_cycle <= 100, "invalid duty cycle");
1154 if (duty_cycle != 0) {
1155 // The duty_cycle is a percentage between 0 and 100; convert to words and
1156 // then compute the offset from the endpoints of the space.
1157 size_t free_words = eden->free() / HeapWordSize;
1158 double free_words_dbl = (double)free_words;
1159 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
1160 size_t offset_words = (free_words - duty_cycle_words) / 2;
1161
1162 _icms_start_limit = eden->top() + offset_words;
1163 _icms_stop_limit = eden->end() - offset_words;
1164
1165 // The limits may be adjusted (shifted to the right) by
1166 // CMSIncrementalOffset, to allow the application more mutator time after a
1167 // young gen gc (when all mutators were stopped) and before CMS starts and
1168 // takes away one or more cpus.
1169 if (CMSIncrementalOffset != 0) {
1170 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
1171 size_t adjustment = (size_t)adjustment_dbl;
1172 HeapWord* tmp_stop = _icms_stop_limit + adjustment;
1173 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
1174 _icms_start_limit += adjustment;
1175 _icms_stop_limit = tmp_stop;
1176 }
1177 }
1178 }
1179 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
1180 _icms_start_limit = _icms_stop_limit = eden->end();
1181 }
1182
1183 // Install the new start limit.
1184 eden->set_soft_end(_icms_start_limit);
1185
1186 if (CMSTraceIncrementalMode) {
1187 gclog_or_tty->print(" icms alloc limits: "
1188 PTR_FORMAT "," PTR_FORMAT
1189 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
1190 _icms_start_limit, _icms_stop_limit,
1191 percent_of_space(eden, _icms_start_limit),
1192 percent_of_space(eden, _icms_stop_limit));
1193 if (Verbose) {
1194 gclog_or_tty->print("eden: ");
1195 eden->print_on(gclog_or_tty);
1196 }
1197 }
1198 }
1199
1200 // Any changes here should try to maintain the invariant
1201 // that if this method is called with _icms_start_limit
1202 // and _icms_stop_limit both NULL, then it should return NULL
1203 // and not notify the icms thread.
1204 HeapWord*
1205 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
1206 size_t word_size)
1207 {
1208 // A start_limit equal to end() means the duty cycle is 0, so treat that as a
1209 // nop.
1210 if (CMSIncrementalMode && _icms_start_limit != space->end()) {
1211 if (top <= _icms_start_limit) {
1212 if (CMSTraceIncrementalMode) {
1213 space->print_on(gclog_or_tty);
1214 gclog_or_tty->stamp();
1215 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
1216 ", new limit=" PTR_FORMAT
1217 " (" SIZE_FORMAT "%%)",
1218 top, _icms_stop_limit,
1219 percent_of_space(space, _icms_stop_limit));
1220 }
1221 ConcurrentMarkSweepThread::start_icms();
1222 assert(top < _icms_stop_limit, "Tautology");
1223 if (word_size < pointer_delta(_icms_stop_limit, top)) {
1224 return _icms_stop_limit;
1225 }
1226
1227 // The allocation will cross both the _start and _stop limits, so do the
1228 // stop notification also and return end().
1229 if (CMSTraceIncrementalMode) {
1230 space->print_on(gclog_or_tty);
1231 gclog_or_tty->stamp();
1232 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
1233 ", new limit=" PTR_FORMAT
1234 " (" SIZE_FORMAT "%%)",
1235 top, space->end(),
1236 percent_of_space(space, space->end()));
1237 }
1238 ConcurrentMarkSweepThread::stop_icms();
1239 return space->end();
1240 }
1241
1242 if (top <= _icms_stop_limit) {
1243 if (CMSTraceIncrementalMode) {
1244 space->print_on(gclog_or_tty);
1245 gclog_or_tty->stamp();
1246 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
1247 ", new limit=" PTR_FORMAT
1248 " (" SIZE_FORMAT "%%)",
1249 top, space->end(),
1250 percent_of_space(space, space->end()));
1251 }
1252 ConcurrentMarkSweepThread::stop_icms();
1253 return space->end();
1254 }
1255
1256 if (CMSTraceIncrementalMode) {
1257 space->print_on(gclog_or_tty);
1258 gclog_or_tty->stamp();
1259 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
1260 ", new limit=" PTR_FORMAT,
1261 top, NULL);
1262 }
1263 }
1264
1265 return NULL;
1266 }
1267
1268 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
1269 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1270 // allocate, copy and if necessary update promoinfo --
1271 // delegate to underlying space.
1272 assert_lock_strong(freelistLock());
1273
1274 #ifndef PRODUCT
1275 if (Universe::heap()->promotion_should_fail()) {
1276 return NULL;
1277 }
1278 #endif // #ifndef PRODUCT
1279
1280 oop res = _cmsSpace->promote(obj, obj_size);
1281 if (res == NULL) {
1282 // expand and retry
1283 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
1284 expand(s*HeapWordSize, MinHeapDeltaBytes,
1285 CMSExpansionCause::_satisfy_promotion);
1286 // Since there's currently no next generation, we don't try to promote
1287 // into a more senior generation.
1288 assert(next_gen() == NULL, "assumption, based upon which no attempt "
1289 "is made to pass on a possibly failing "
1290 "promotion to next generation");
1291 res = _cmsSpace->promote(obj, obj_size);
1292 }
1293 if (res != NULL) {
1294 // See comment in allocate() about when objects should
1295 // be allocated live.
1296 assert(obj->is_oop(), "Will dereference klass pointer below");
1297 collector()->promoted(false, // Not parallel
1298 (HeapWord*)res, obj->is_objArray(), obj_size);
1299 // promotion counters
1300 NOT_PRODUCT(
1301 _numObjectsPromoted++;
1302 _numWordsPromoted +=
1303 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1304 )
1305 }
1306 return res;
1307 }
1308
1309
1310 HeapWord*
1311 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
1312 HeapWord* top,
1313 size_t word_sz)
1314 {
1315 return collector()->allocation_limit_reached(space, top, word_sz);
1316 }
1317
1318 // IMPORTANT: Notes on object size recognition in CMS.
1319 // ---------------------------------------------------
1320 // A block of storage in the CMS generation is always in
1321 // one of three states. A free block (FREE), an allocated
1322 // object (OBJECT) whose size() method reports the correct size,
1323 // and an intermediate state (TRANSIENT) in which its size cannot
1324 // be accurately determined.
1325 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1326 // -----------------------------------------------------
1327 // FREE: klass_word & 1 == 1; mark_word holds block size
1328 //
1329 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1330 // obj->size() computes correct size
1331 //
1332 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1333 //
1334 // STATE IDENTIFICATION: (64 bit+COOPS)
1335 // ------------------------------------
1336 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1337 //
1338 // OBJECT: klass_word installed; klass_word != 0;
1339 // obj->size() computes correct size
1340 //
1341 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1342 //
1343 //
1344 // STATE TRANSITION DIAGRAM
1345 //
1346 // mut / parnew mut / parnew
1347 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1348 // ^ |
1349 // |------------------------ DEAD <------------------------------------|
1350 // sweep mut
1351 //
1352 // While a block is in TRANSIENT state its size cannot be determined
1353 // so readers will either need to come back later or stall until
1354 // the size can be determined. Note that for the case of direct
1355 // allocation, P-bits, when available, may be used to determine the
1356 // size of an object that may not yet have been initialized.
1357
1358 // Things to support parallel young-gen collection.
1359 oop
1360 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1361 oop old, markOop m,
1362 size_t word_sz) {
1363 #ifndef PRODUCT
1364 if (Universe::heap()->promotion_should_fail()) {
1365 return NULL;
1366 }
1367 #endif // #ifndef PRODUCT
1368
1369 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1370 PromotionInfo* promoInfo = &ps->promo;
1371 // if we are tracking promotions, then first ensure space for
1372 // promotion (including spooling space for saving header if necessary).
1373 // then allocate and copy, then track promoted info if needed.
1374 // When tracking (see PromotionInfo::track()), the mark word may
1375 // be displaced and in this case restoration of the mark word
1376 // occurs in the (oop_since_save_marks_)iterate phase.
1377 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1378 // Out of space for allocating spooling buffers;
1379 // try expanding and allocating spooling buffers.
1380 if (!expand_and_ensure_spooling_space(promoInfo)) {
1381 return NULL;
1382 }
1383 }
1384 assert(promoInfo->has_spooling_space(), "Control point invariant");
1385 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1386 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1387 if (obj_ptr == NULL) {
1388 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1389 if (obj_ptr == NULL) {
1390 return NULL;
1391 }
1392 }
1393 oop obj = oop(obj_ptr);
1394 OrderAccess::storestore();
1395 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1396 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1397 // IMPORTANT: See note on object initialization for CMS above.
1398 // Otherwise, copy the object. Here we must be careful to insert the
1399 // klass pointer last, since this marks the block as an allocated object.
1400 // Except with compressed oops it's the mark word.
1401 HeapWord* old_ptr = (HeapWord*)old;
1402 // Restore the mark word copied above.
1403 obj->set_mark(m);
1404 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1405 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1406 OrderAccess::storestore();
1407
1408 if (UseCompressedKlassPointers) {
1409 // Copy gap missed by (aligned) header size calculation below
1410 obj->set_klass_gap(old->klass_gap());
1411 }
1412 if (word_sz > (size_t)oopDesc::header_size()) {
1413 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1414 obj_ptr + oopDesc::header_size(),
1415 word_sz - oopDesc::header_size());
1416 }
1417
1418 // Now we can track the promoted object, if necessary. We take care
1419 // to delay the transition from uninitialized to full object
1420 // (i.e., insertion of klass pointer) until after, so that it
1421 // atomically becomes a promoted object.
1422 if (promoInfo->tracking()) {
1423 promoInfo->track((PromotedObject*)obj, old->klass());
1424 }
1425 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1426 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1427 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1428
1429 // Finally, install the klass pointer (this should be volatile).
1430 OrderAccess::storestore();
1431 obj->set_klass(old->klass());
1432 // We should now be able to calculate the right size for this object
1433 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1434
1435 collector()->promoted(true, // parallel
1436 obj_ptr, old->is_objArray(), word_sz);
1437
1438 NOT_PRODUCT(
1439 Atomic::inc_ptr(&_numObjectsPromoted);
1440 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1441 )
1442
1443 return obj;
1444 }
1445
1446 void
1447 ConcurrentMarkSweepGeneration::
1448 par_promote_alloc_undo(int thread_num,
1449 HeapWord* obj, size_t word_sz) {
1450 // CMS does not support promotion undo.
1451 ShouldNotReachHere();
1452 }
1453
1454 void
1455 ConcurrentMarkSweepGeneration::
1456 par_promote_alloc_done(int thread_num) {
1457 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1458 ps->lab.retire(thread_num);
1459 }
1460
1461 void
1462 ConcurrentMarkSweepGeneration::
1463 par_oop_since_save_marks_iterate_done(int thread_num) {
1464 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1465 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1466 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1467 }
1468
1469 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1470 size_t size,
1471 bool tlab)
1472 {
1473 // We allow a STW collection only if a full
1474 // collection was requested.
1475 return full || should_allocate(size, tlab); // FIX ME !!!
1476 // This and promotion failure handling are connected at the
1477 // hip and should be fixed by untying them.
1478 }
1479
1480 bool CMSCollector::shouldConcurrentCollect() {
1481 if (_full_gc_requested) {
1482 if (Verbose && PrintGCDetails) {
1483 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1484 " gc request (or gc_locker)");
1485 }
1486 return true;
1487 }
1488
1489 // For debugging purposes, change the type of collection.
1490 // If the rotation is not on the concurrent collection
1491 // type, don't start a concurrent collection.
1492 NOT_PRODUCT(
1493 if (RotateCMSCollectionTypes &&
1494 (_cmsGen->debug_collection_type() !=
1495 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
1496 assert(_cmsGen->debug_collection_type() !=
1497 ConcurrentMarkSweepGeneration::Unknown_collection_type,
1498 "Bad cms collection type");
1499 return false;
1500 }
1501 )
1502
1503 FreelistLocker x(this);
1504 // ------------------------------------------------------------------
1505 // Print out lots of information which affects the initiation of
1506 // a collection.
1507 if (PrintCMSInitiationStatistics && stats().valid()) {
1508 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1509 gclog_or_tty->stamp();
1510 gclog_or_tty->print_cr("");
1511 stats().print_on(gclog_or_tty);
1512 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1513 stats().time_until_cms_gen_full());
1514 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1515 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1516 _cmsGen->contiguous_available());
1517 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1518 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1519 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1520 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1521 gclog_or_tty->print_cr("metadata initialized %d",
1522 MetaspaceGC::should_concurrent_collect());
1523 }
1524 // ------------------------------------------------------------------
1525
1526 // If the estimated time to complete a cms collection (cms_duration())
1527 // is less than the estimated time remaining until the cms generation
1528 // is full, start a collection.
1529 if (!UseCMSInitiatingOccupancyOnly) {
1530 if (stats().valid()) {
1531 if (stats().time_until_cms_start() == 0.0) {
1532 return true;
1533 }
1534 } else {
1535 // We want to conservatively collect somewhat early in order
1536 // to try and "bootstrap" our CMS/promotion statistics;
1537 // this branch will not fire after the first successful CMS
1538 // collection because the stats should then be valid.
1539 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1540 if (Verbose && PrintGCDetails) {
1541 gclog_or_tty->print_cr(
1542 " CMSCollector: collect for bootstrapping statistics:"
1543 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1544 _bootstrap_occupancy);
1545 }
1546 return true;
1547 }
1548 }
1549 }
1550
1551 // Otherwise, we start a collection cycle if
1552 // old gen want a collection cycle started. Each may use
1553 // an appropriate criterion for making this decision.
1554 // XXX We need to make sure that the gen expansion
1555 // criterion dovetails well with this. XXX NEED TO FIX THIS
1556 if (_cmsGen->should_concurrent_collect()) {
1557 if (Verbose && PrintGCDetails) {
1558 gclog_or_tty->print_cr("CMS old gen initiated");
1559 }
1560 return true;
1561 }
1562
1563 // We start a collection if we believe an incremental collection may fail;
1564 // this is not likely to be productive in practice because it's probably too
1565 // late anyway.
1566 GenCollectedHeap* gch = GenCollectedHeap::heap();
1567 assert(gch->collector_policy()->is_two_generation_policy(),
1568 "You may want to check the correctness of the following");
1569 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1570 if (Verbose && PrintGCDetails) {
1571 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1572 }
1573 return true;
1574 }
1575
1576 if (MetaspaceGC::should_concurrent_collect()) {
1577 if (Verbose && PrintGCDetails) {
1578 gclog_or_tty->print("CMSCollector: collect for metadata allocation ");
1579 }
1580 return true;
1581 }
1582
1583 return false;
1584 }
1585
1586 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1587
1588 // Clear _expansion_cause fields of constituent generations
1589 void CMSCollector::clear_expansion_cause() {
1590 _cmsGen->clear_expansion_cause();
1591 }
1592
1593 // We should be conservative in starting a collection cycle. To
1594 // start too eagerly runs the risk of collecting too often in the
1595 // extreme. To collect too rarely falls back on full collections,
1596 // which works, even if not optimum in terms of concurrent work.
1597 // As a work around for too eagerly collecting, use the flag
1598 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1599 // giving the user an easily understandable way of controlling the
1600 // collections.
1601 // We want to start a new collection cycle if any of the following
1602 // conditions hold:
1603 // . our current occupancy exceeds the configured initiating occupancy
1604 // for this generation, or
1605 // . we recently needed to expand this space and have not, since that
1606 // expansion, done a collection of this generation, or
1607 // . the underlying space believes that it may be a good idea to initiate
1608 // a concurrent collection (this may be based on criteria such as the
1609 // following: the space uses linear allocation and linear allocation is
1610 // going to fail, or there is believed to be excessive fragmentation in
1611 // the generation, etc... or ...
1612 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1613 // the case of the old generation; see CR 6543076):
1614 // we may be approaching a point at which allocation requests may fail because
1615 // we will be out of sufficient free space given allocation rate estimates.]
1616 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1617
1618 assert_lock_strong(freelistLock());
1619 if (occupancy() > initiating_occupancy()) {
1620 if (PrintGCDetails && Verbose) {
1621 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1622 short_name(), occupancy(), initiating_occupancy());
1623 }
1624 return true;
1625 }
1626 if (UseCMSInitiatingOccupancyOnly) {
1627 return false;
1628 }
1629 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1630 if (PrintGCDetails && Verbose) {
1631 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1632 short_name());
1633 }
1634 return true;
1635 }
1636 if (_cmsSpace->should_concurrent_collect()) {
1637 if (PrintGCDetails && Verbose) {
1638 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1639 short_name());
1640 }
1641 return true;
1642 }
1643 return false;
1644 }
1645
1646 void ConcurrentMarkSweepGeneration::collect(bool full,
1647 bool clear_all_soft_refs,
1648 size_t size,
1649 bool tlab)
1650 {
1651 collector()->collect(full, clear_all_soft_refs, size, tlab);
1652 }
1653
1654 void CMSCollector::collect(bool full,
1655 bool clear_all_soft_refs,
1656 size_t size,
1657 bool tlab)
1658 {
1659 if (!UseCMSCollectionPassing && _collectorState > Idling) {
1660 // For debugging purposes skip the collection if the state
1661 // is not currently idle
1662 if (TraceCMSState) {
1663 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
1664 Thread::current(), full, _collectorState);
1665 }
1666 return;
1667 }
1668
1669 // The following "if" branch is present for defensive reasons.
1670 // In the current uses of this interface, it can be replaced with:
1671 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1672 // But I am not placing that assert here to allow future
1673 // generality in invoking this interface.
1674 if (GC_locker::is_active()) {
1675 // A consistency test for GC_locker
1676 assert(GC_locker::needs_gc(), "Should have been set already");
1677 // Skip this foreground collection, instead
1678 // expanding the heap if necessary.
1679 // Need the free list locks for the call to free() in compute_new_size()
1680 compute_new_size();
1681 return;
1682 }
1683 acquire_control_and_collect(full, clear_all_soft_refs);
1684 _full_gcs_since_conc_gc++;
1685 }
1686
1687 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1688 GenCollectedHeap* gch = GenCollectedHeap::heap();
1689 unsigned int gc_count = gch->total_full_collections();
1690 if (gc_count == full_gc_count) {
1691 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1692 _full_gc_requested = true;
1693 _full_gc_cause = cause;
1694 CGC_lock->notify(); // nudge CMS thread
1695 } else {
1696 assert(gc_count > full_gc_count, "Error: causal loop");
1697 }
1698 }
1699
1700 bool CMSCollector::is_external_interruption() {
1701 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1702 return GCCause::is_user_requested_gc(cause) ||
1703 GCCause::is_serviceability_requested_gc(cause);
1704 }
1705
1706 void CMSCollector::report_concurrent_mode_interruption() {
1707 if (is_external_interruption()) {
1708 if (PrintGCDetails) {
1709 gclog_or_tty->print(" (concurrent mode interrupted)");
1710 }
1711 } else {
1712 if (PrintGCDetails) {
1713 gclog_or_tty->print(" (concurrent mode failure)");
1714 }
1715 _gc_tracer_cm->report_concurrent_mode_failure();
1716 }
1717 }
1718
1719
1720 // The foreground and background collectors need to coordinate in order
1721 // to make sure that they do not mutually interfere with CMS collections.
1722 // When a background collection is active,
1723 // the foreground collector may need to take over (preempt) and
1724 // synchronously complete an ongoing collection. Depending on the
1725 // frequency of the background collections and the heap usage
1726 // of the application, this preemption can be seldom or frequent.
1727 // There are only certain
1728 // points in the background collection that the "collection-baton"
1729 // can be passed to the foreground collector.
1730 //
1731 // The foreground collector will wait for the baton before
1732 // starting any part of the collection. The foreground collector
1733 // will only wait at one location.
1734 //
1735 // The background collector will yield the baton before starting a new
1736 // phase of the collection (e.g., before initial marking, marking from roots,
1737 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1738 // of the loop which switches the phases. The background collector does some
1739 // of the phases (initial mark, final re-mark) with the world stopped.
1740 // Because of locking involved in stopping the world,
1741 // the foreground collector should not block waiting for the background
1742 // collector when it is doing a stop-the-world phase. The background
1743 // collector will yield the baton at an additional point just before
1744 // it enters a stop-the-world phase. Once the world is stopped, the
1745 // background collector checks the phase of the collection. If the
1746 // phase has not changed, it proceeds with the collection. If the
1747 // phase has changed, it skips that phase of the collection. See
1748 // the comments on the use of the Heap_lock in collect_in_background().
1749 //
1750 // Variable used in baton passing.
1751 // _foregroundGCIsActive - Set to true by the foreground collector when
1752 // it wants the baton. The foreground clears it when it has finished
1753 // the collection.
1754 // _foregroundGCShouldWait - Set to true by the background collector
1755 // when it is running. The foreground collector waits while
1756 // _foregroundGCShouldWait is true.
1757 // CGC_lock - monitor used to protect access to the above variables
1758 // and to notify the foreground and background collectors.
1759 // _collectorState - current state of the CMS collection.
1760 //
1761 // The foreground collector
1762 // acquires the CGC_lock
1763 // sets _foregroundGCIsActive
1764 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1765 // various locks acquired in preparation for the collection
1766 // are released so as not to block the background collector
1767 // that is in the midst of a collection
1768 // proceeds with the collection
1769 // clears _foregroundGCIsActive
1770 // returns
1771 //
1772 // The background collector in a loop iterating on the phases of the
1773 // collection
1774 // acquires the CGC_lock
1775 // sets _foregroundGCShouldWait
1776 // if _foregroundGCIsActive is set
1777 // clears _foregroundGCShouldWait, notifies _CGC_lock
1778 // waits on _CGC_lock for _foregroundGCIsActive to become false
1779 // and exits the loop.
1780 // otherwise
1781 // proceed with that phase of the collection
1782 // if the phase is a stop-the-world phase,
1783 // yield the baton once more just before enqueueing
1784 // the stop-world CMS operation (executed by the VM thread).
1785 // returns after all phases of the collection are done
1786 //
1787
1788 void CMSCollector::acquire_control_and_collect(bool full,
1789 bool clear_all_soft_refs) {
1790 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1791 assert(!Thread::current()->is_ConcurrentGC_thread(),
1792 "shouldn't try to acquire control from self!");
1793
1794 // Start the protocol for acquiring control of the
1795 // collection from the background collector (aka CMS thread).
1796 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1797 "VM thread should have CMS token");
1798 // Remember the possibly interrupted state of an ongoing
1799 // concurrent collection
1800 CollectorState first_state = _collectorState;
1801
1802 // Signal to a possibly ongoing concurrent collection that
1803 // we want to do a foreground collection.
1804 _foregroundGCIsActive = true;
1805
1806 // Disable incremental mode during a foreground collection.
1807 ICMSDisabler icms_disabler;
1808
1809 // release locks and wait for a notify from the background collector
1810 // releasing the locks in only necessary for phases which
1811 // do yields to improve the granularity of the collection.
1812 assert_lock_strong(bitMapLock());
1813 // We need to lock the Free list lock for the space that we are
1814 // currently collecting.
1815 assert(haveFreelistLocks(), "Must be holding free list locks");
1816 bitMapLock()->unlock();
1817 releaseFreelistLocks();
1818 {
1819 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1820 if (_foregroundGCShouldWait) {
1821 // We are going to be waiting for action for the CMS thread;
1822 // it had better not be gone (for instance at shutdown)!
1823 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1824 "CMS thread must be running");
1825 // Wait here until the background collector gives us the go-ahead
1826 ConcurrentMarkSweepThread::clear_CMS_flag(
1827 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1828 // Get a possibly blocked CMS thread going:
1829 // Note that we set _foregroundGCIsActive true above,
1830 // without protection of the CGC_lock.
1831 CGC_lock->notify();
1832 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1833 "Possible deadlock");
1834 while (_foregroundGCShouldWait) {
1835 // wait for notification
1836 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1837 // Possibility of delay/starvation here, since CMS token does
1838 // not know to give priority to VM thread? Actually, i think
1839 // there wouldn't be any delay/starvation, but the proof of
1840 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1841 }
1842 ConcurrentMarkSweepThread::set_CMS_flag(
1843 ConcurrentMarkSweepThread::CMS_vm_has_token);
1844 }
1845 }
1846 // The CMS_token is already held. Get back the other locks.
1847 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1848 "VM thread should have CMS token");
1849 getFreelistLocks();
1850 bitMapLock()->lock_without_safepoint_check();
1851 if (TraceCMSState) {
1852 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1853 INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1854 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1855 }
1856
1857 // Check if we need to do a compaction, or if not, whether
1858 // we need to start the mark-sweep from scratch.
1859 bool should_compact = false;
1860 bool should_start_over = false;
1861 decide_foreground_collection_type(clear_all_soft_refs,
1862 &should_compact, &should_start_over);
1863
1864 NOT_PRODUCT(
1865 if (RotateCMSCollectionTypes) {
1866 if (_cmsGen->debug_collection_type() ==
1867 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1868 should_compact = true;
1869 } else if (_cmsGen->debug_collection_type() ==
1870 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1871 should_compact = false;
1872 }
1873 }
1874 )
1875
1876 if (first_state > Idling) {
1877 report_concurrent_mode_interruption();
1878 }
1879
1880 set_did_compact(should_compact);
1881 if (should_compact) {
1882 // If the collection is being acquired from the background
1883 // collector, there may be references on the discovered
1884 // references lists that have NULL referents (being those
1885 // that were concurrently cleared by a mutator) or
1886 // that are no longer active (having been enqueued concurrently
1887 // by the mutator).
1888 // Scrub the list of those references because Mark-Sweep-Compact
1889 // code assumes referents are not NULL and that all discovered
1890 // Reference objects are active.
1891 ref_processor()->clean_up_discovered_references();
1892
1893 if (first_state > Idling) {
1894 save_heap_summary();
1895 }
1896
1897 do_compaction_work(clear_all_soft_refs);
1898
1899 // Has the GC time limit been exceeded?
1900 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration();
1901 size_t max_eden_size = young_gen->max_capacity() -
1902 young_gen->to()->capacity() -
1903 young_gen->from()->capacity();
1904 GenCollectedHeap* gch = GenCollectedHeap::heap();
1905 GCCause::Cause gc_cause = gch->gc_cause();
1906 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1907 young_gen->eden()->used(),
1908 _cmsGen->max_capacity(),
1909 max_eden_size,
1910 full,
1911 gc_cause,
1912 gch->collector_policy());
1913 } else {
1914 do_mark_sweep_work(clear_all_soft_refs, first_state,
1915 should_start_over);
1916 }
1917 // Reset the expansion cause, now that we just completed
1918 // a collection cycle.
1919 clear_expansion_cause();
1920 _foregroundGCIsActive = false;
1921 return;
1922 }
1923
1924 // Resize the tenured generation
1925 // after obtaining the free list locks for the
1926 // two generations.
1927 void CMSCollector::compute_new_size() {
1928 assert_locked_or_safepoint(Heap_lock);
1929 FreelistLocker z(this);
1930 MetaspaceGC::compute_new_size();
1931 _cmsGen->compute_new_size_free_list();
1932 }
1933
1934 // A work method used by foreground collection to determine
1935 // what type of collection (compacting or not, continuing or fresh)
1936 // it should do.
1937 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1938 // and CMSCompactWhenClearAllSoftRefs the default in the future
1939 // and do away with the flags after a suitable period.
1940 void CMSCollector::decide_foreground_collection_type(
1941 bool clear_all_soft_refs, bool* should_compact,
1942 bool* should_start_over) {
1943 // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1944 // flag is set, and we have either requested a System.gc() or
1945 // the number of full gc's since the last concurrent cycle
1946 // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1947 // or if an incremental collection has failed
1948 GenCollectedHeap* gch = GenCollectedHeap::heap();
1949 assert(gch->collector_policy()->is_two_generation_policy(),
1950 "You may want to check the correctness of the following");
1951 // Inform cms gen if this was due to partial collection failing.
1952 // The CMS gen may use this fact to determine its expansion policy.
1953 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1954 assert(!_cmsGen->incremental_collection_failed(),
1955 "Should have been noticed, reacted to and cleared");
1956 _cmsGen->set_incremental_collection_failed();
1957 }
1958 *should_compact =
1959 UseCMSCompactAtFullCollection &&
1960 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1961 GCCause::is_user_requested_gc(gch->gc_cause()) ||
1962 gch->incremental_collection_will_fail(true /* consult_young */));
1963 *should_start_over = false;
1964 if (clear_all_soft_refs && !*should_compact) {
1965 // We are about to do a last ditch collection attempt
1966 // so it would normally make sense to do a compaction
1967 // to reclaim as much space as possible.
1968 if (CMSCompactWhenClearAllSoftRefs) {
1969 // Default: The rationale is that in this case either
1970 // we are past the final marking phase, in which case
1971 // we'd have to start over, or so little has been done
1972 // that there's little point in saving that work. Compaction
1973 // appears to be the sensible choice in either case.
1974 *should_compact = true;
1975 } else {
1976 // We have been asked to clear all soft refs, but not to
1977 // compact. Make sure that we aren't past the final checkpoint
1978 // phase, for that is where we process soft refs. If we are already
1979 // past that phase, we'll need to redo the refs discovery phase and
1980 // if necessary clear soft refs that weren't previously
1981 // cleared. We do so by remembering the phase in which
1982 // we came in, and if we are past the refs processing
1983 // phase, we'll choose to just redo the mark-sweep
1984 // collection from scratch.
1985 if (_collectorState > FinalMarking) {
1986 // We are past the refs processing phase;
1987 // start over and do a fresh synchronous CMS cycle
1988 _collectorState = Resetting; // skip to reset to start new cycle
1989 reset(false /* == !asynch */);
1990 *should_start_over = true;
1991 } // else we can continue a possibly ongoing current cycle
1992 }
1993 }
1994 }
1995
1996 // A work method used by the foreground collector to do
1997 // a mark-sweep-compact.
1998 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1999 GenCollectedHeap* gch = GenCollectedHeap::heap();
2000
2001 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
2002 gc_timer->register_gc_start(os::elapsed_counter());
2003
2004 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
2005 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
2006
2007 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL);
2008 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
2009 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
2010 "collections passed to foreground collector", _full_gcs_since_conc_gc);
2011 }
2012
2013 // Sample collection interval time and reset for collection pause.
2014 if (UseAdaptiveSizePolicy) {
2015 size_policy()->msc_collection_begin();
2016 }
2017
2018 // Temporarily widen the span of the weak reference processing to
2019 // the entire heap.
2020 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
2021 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
2022 // Temporarily, clear the "is_alive_non_header" field of the
2023 // reference processor.
2024 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
2025 // Temporarily make reference _processing_ single threaded (non-MT).
2026 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
2027 // Temporarily make refs discovery atomic
2028 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
2029 // Temporarily make reference _discovery_ single threaded (non-MT)
2030 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
2031
2032 ref_processor()->set_enqueuing_is_done(false);
2033 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/);
2034 ref_processor()->setup_policy(clear_all_soft_refs);
2035 // If an asynchronous collection finishes, the _modUnionTable is
2036 // all clear. If we are assuming the collection from an asynchronous
2037 // collection, clear the _modUnionTable.
2038 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
2039 "_modUnionTable should be clear if the baton was not passed");
2040 _modUnionTable.clear_all();
2041 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
2042 "mod union for klasses should be clear if the baton was passed");
2043 _ct->klass_rem_set()->clear_mod_union();
2044
2045 // We must adjust the allocation statistics being maintained
2046 // in the free list space. We do so by reading and clearing
2047 // the sweep timer and updating the block flux rate estimates below.
2048 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
2049 if (_inter_sweep_timer.is_active()) {
2050 _inter_sweep_timer.stop();
2051 // Note that we do not use this sample to update the _inter_sweep_estimate.
2052 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
2053 _inter_sweep_estimate.padded_average(),
2054 _intra_sweep_estimate.padded_average());
2055 }
2056
2057 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
2058 ref_processor(), clear_all_soft_refs);
2059 #ifdef ASSERT
2060 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
2061 size_t free_size = cms_space->free();
2062 assert(free_size ==
2063 pointer_delta(cms_space->end(), cms_space->compaction_top())
2064 * HeapWordSize,
2065 "All the free space should be compacted into one chunk at top");
2066 assert(cms_space->dictionary()->total_chunk_size(
2067 debug_only(cms_space->freelistLock())) == 0 ||
2068 cms_space->totalSizeInIndexedFreeLists() == 0,
2069 "All the free space should be in a single chunk");
2070 size_t num = cms_space->totalCount();
2071 assert((free_size == 0 && num == 0) ||
2072 (free_size > 0 && (num == 1 || num == 2)),
2073 "There should be at most 2 free chunks after compaction");
2074 #endif // ASSERT
2075 _collectorState = Resetting;
2076 assert(_restart_addr == NULL,
2077 "Should have been NULL'd before baton was passed");
2078 reset(false /* == !asynch */);
2079 _cmsGen->reset_after_compaction();
2080 _concurrent_cycles_since_last_unload = 0;
2081
2082 // Clear any data recorded in the PLAB chunk arrays.
2083 if (_survivor_plab_array != NULL) {
2084 reset_survivor_plab_arrays();
2085 }
2086
2087 // Adjust the per-size allocation stats for the next epoch.
2088 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
2089 // Restart the "inter sweep timer" for the next epoch.
2090 _inter_sweep_timer.reset();
2091 _inter_sweep_timer.start();
2092
2093 // Sample collection pause time and reset for collection interval.
2094 if (UseAdaptiveSizePolicy) {
2095 size_policy()->msc_collection_end(gch->gc_cause());
2096 }
2097
2098 gc_timer->register_gc_end(os::elapsed_counter());
2099
2100 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
2101
2102 // For a mark-sweep-compact, compute_new_size() will be called
2103 // in the heap's do_collection() method.
2104 }
2105
2106 // A work method used by the foreground collector to do
2107 // a mark-sweep, after taking over from a possibly on-going
2108 // concurrent mark-sweep collection.
2109 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2110 CollectorState first_state, bool should_start_over) {
2111 if (PrintGC && Verbose) {
2112 gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2113 "collector with count %d",
2114 _full_gcs_since_conc_gc);
2115 }
2116 switch (_collectorState) {
2117 case Idling:
2118 if (first_state == Idling || should_start_over) {
2119 // The background GC was not active, or should
2120 // restarted from scratch; start the cycle.
2121 _collectorState = InitialMarking;
2122 }
2123 // If first_state was not Idling, then a background GC
2124 // was in progress and has now finished. No need to do it
2125 // again. Leave the state as Idling.
2126 break;
2127 case Precleaning:
2128 // In the foreground case don't do the precleaning since
2129 // it is not done concurrently and there is extra work
2130 // required.
2131 _collectorState = FinalMarking;
2132 }
2133 collect_in_foreground(clear_all_soft_refs, GenCollectedHeap::heap()->gc_cause());
2134
2135 // For a mark-sweep, compute_new_size() will be called
2136 // in the heap's do_collection() method.
2137 }
2138
2139
2140 void CMSCollector::getFreelistLocks() const {
2141 // Get locks for all free lists in all generations that this
2142 // collector is responsible for
2143 _cmsGen->freelistLock()->lock_without_safepoint_check();
2144 }
2145
2146 void CMSCollector::releaseFreelistLocks() const {
2147 // Release locks for all free lists in all generations that this
2148 // collector is responsible for
2149 _cmsGen->freelistLock()->unlock();
2150 }
2151
2152 bool CMSCollector::haveFreelistLocks() const {
2153 // Check locks for all free lists in all generations that this
2154 // collector is responsible for
2155 assert_lock_strong(_cmsGen->freelistLock());
2156 PRODUCT_ONLY(ShouldNotReachHere());
2157 return true;
2158 }
2159
2160 // A utility class that is used by the CMS collector to
2161 // temporarily "release" the foreground collector from its
2162 // usual obligation to wait for the background collector to
2163 // complete an ongoing phase before proceeding.
2164 class ReleaseForegroundGC: public StackObj {
2165 private:
2166 CMSCollector* _c;
2167 public:
2168 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2169 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2170 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2171 // allow a potentially blocked foreground collector to proceed
2172 _c->_foregroundGCShouldWait = false;
2173 if (_c->_foregroundGCIsActive) {
2174 CGC_lock->notify();
2175 }
2176 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2177 "Possible deadlock");
2178 }
2179
2180 ~ReleaseForegroundGC() {
2181 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2182 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2183 _c->_foregroundGCShouldWait = true;
2184 }
2185 };
2186
2187 // There are separate collect_in_background and collect_in_foreground because of
2188 // the different locking requirements of the background collector and the
2189 // foreground collector. There was originally an attempt to share
2190 // one "collect" method between the background collector and the foreground
2191 // collector but the if-then-else required made it cleaner to have
2192 // separate methods.
2193 void CMSCollector::collect_in_background(bool clear_all_soft_refs, GCCause::Cause cause) {
2194 assert(Thread::current()->is_ConcurrentGC_thread(),
2195 "A CMS asynchronous collection is only allowed on a CMS thread.");
2196
2197 GenCollectedHeap* gch = GenCollectedHeap::heap();
2198 {
2199 bool safepoint_check = Mutex::_no_safepoint_check_flag;
2200 MutexLockerEx hl(Heap_lock, safepoint_check);
2201 FreelistLocker fll(this);
2202 MutexLockerEx x(CGC_lock, safepoint_check);
2203 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2204 // The foreground collector is active or we're
2205 // not using asynchronous collections. Skip this
2206 // background collection.
2207 assert(!_foregroundGCShouldWait, "Should be clear");
2208 return;
2209 } else {
2210 assert(_collectorState == Idling, "Should be idling before start.");
2211 _collectorState = InitialMarking;
2212 register_gc_start(cause);
2213 // Reset the expansion cause, now that we are about to begin
2214 // a new cycle.
2215 clear_expansion_cause();
2216
2217 // Clear the MetaspaceGC flag since a concurrent collection
2218 // is starting but also clear it after the collection.
2219 MetaspaceGC::set_should_concurrent_collect(false);
2220 }
2221 // Decide if we want to enable class unloading as part of the
2222 // ensuing concurrent GC cycle.
2223 update_should_unload_classes();
2224 _full_gc_requested = false; // acks all outstanding full gc requests
2225 _full_gc_cause = GCCause::_no_gc;
2226 // Signal that we are about to start a collection
2227 gch->increment_total_full_collections(); // ... starting a collection cycle
2228 _collection_count_start = gch->total_full_collections();
2229 }
2230
2231 // Used for PrintGC
2232 size_t prev_used;
2233 if (PrintGC && Verbose) {
2234 prev_used = _cmsGen->used(); // XXXPERM
2235 }
2236
2237 // The change of the collection state is normally done at this level;
2238 // the exceptions are phases that are executed while the world is
2239 // stopped. For those phases the change of state is done while the
2240 // world is stopped. For baton passing purposes this allows the
2241 // background collector to finish the phase and change state atomically.
2242 // The foreground collector cannot wait on a phase that is done
2243 // while the world is stopped because the foreground collector already
2244 // has the world stopped and would deadlock.
2245 while (_collectorState != Idling) {
2246 if (TraceCMSState) {
2247 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2248 Thread::current(), _collectorState);
2249 }
2250 // The foreground collector
2251 // holds the Heap_lock throughout its collection.
2252 // holds the CMS token (but not the lock)
2253 // except while it is waiting for the background collector to yield.
2254 //
2255 // The foreground collector should be blocked (not for long)
2256 // if the background collector is about to start a phase
2257 // executed with world stopped. If the background
2258 // collector has already started such a phase, the
2259 // foreground collector is blocked waiting for the
2260 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
2261 // are executed in the VM thread.
2262 //
2263 // The locking order is
2264 // PendingListLock (PLL) -- if applicable (FinalMarking)
2265 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
2266 // CMS token (claimed in
2267 // stop_world_and_do() -->
2268 // safepoint_synchronize() -->
2269 // CMSThread::synchronize())
2270
2271 {
2272 // Check if the FG collector wants us to yield.
2273 CMSTokenSync x(true); // is cms thread
2274 if (waitForForegroundGC()) {
2275 // We yielded to a foreground GC, nothing more to be
2276 // done this round.
2277 assert(_foregroundGCShouldWait == false, "We set it to false in "
2278 "waitForForegroundGC()");
2279 if (TraceCMSState) {
2280 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2281 " exiting collection CMS state %d",
2282 Thread::current(), _collectorState);
2283 }
2284 return;
2285 } else {
2286 // The background collector can run but check to see if the
2287 // foreground collector has done a collection while the
2288 // background collector was waiting to get the CGC_lock
2289 // above. If yes, break so that _foregroundGCShouldWait
2290 // is cleared before returning.
2291 if (_collectorState == Idling) {
2292 break;
2293 }
2294 }
2295 }
2296
2297 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2298 "should be waiting");
2299
2300 switch (_collectorState) {
2301 case InitialMarking:
2302 {
2303 ReleaseForegroundGC x(this);
2304 stats().record_cms_begin();
2305 VM_CMS_Initial_Mark initial_mark_op(this);
2306 VMThread::execute(&initial_mark_op);
2307 }
2308 // The collector state may be any legal state at this point
2309 // since the background collector may have yielded to the
2310 // foreground collector.
2311 break;
2312 case Marking:
2313 // initial marking in checkpointRootsInitialWork has been completed
2314 if (markFromRoots(true)) { // we were successful
2315 assert(_collectorState == Precleaning, "Collector state should "
2316 "have changed");
2317 } else {
2318 assert(_foregroundGCIsActive, "Internal state inconsistency");
2319 }
2320 break;
2321 case Precleaning:
2322 if (UseAdaptiveSizePolicy) {
2323 size_policy()->concurrent_precleaning_begin();
2324 }
2325 // marking from roots in markFromRoots has been completed
2326 preclean();
2327 if (UseAdaptiveSizePolicy) {
2328 size_policy()->concurrent_precleaning_end();
2329 }
2330 assert(_collectorState == AbortablePreclean ||
2331 _collectorState == FinalMarking,
2332 "Collector state should have changed");
2333 break;
2334 case AbortablePreclean:
2335 if (UseAdaptiveSizePolicy) {
2336 size_policy()->concurrent_phases_resume();
2337 }
2338 abortable_preclean();
2339 if (UseAdaptiveSizePolicy) {
2340 size_policy()->concurrent_precleaning_end();
2341 }
2342 assert(_collectorState == FinalMarking, "Collector state should "
2343 "have changed");
2344 break;
2345 case FinalMarking:
2346 {
2347 ReleaseForegroundGC x(this);
2348
2349 VM_CMS_Final_Remark final_remark_op(this);
2350 VMThread::execute(&final_remark_op);
2351 }
2352 assert(_foregroundGCShouldWait, "block post-condition");
2353 break;
2354 case Sweeping:
2355 if (UseAdaptiveSizePolicy) {
2356 size_policy()->concurrent_sweeping_begin();
2357 }
2358 // final marking in checkpointRootsFinal has been completed
2359 sweep(true);
2360 assert(_collectorState == Resizing, "Collector state change "
2361 "to Resizing must be done under the free_list_lock");
2362 _full_gcs_since_conc_gc = 0;
2363
2364 // Stop the timers for adaptive size policy for the concurrent phases
2365 if (UseAdaptiveSizePolicy) {
2366 size_policy()->concurrent_sweeping_end();
2367 size_policy()->concurrent_phases_end(gch->gc_cause(),
2368 gch->prev_gen(_cmsGen)->capacity(),
2369 _cmsGen->free());
2370 }
2371
2372 case Resizing: {
2373 // Sweeping has been completed...
2374 // At this point the background collection has completed.
2375 // Don't move the call to compute_new_size() down
2376 // into code that might be executed if the background
2377 // collection was preempted.
2378 {
2379 ReleaseForegroundGC x(this); // unblock FG collection
2380 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
2381 CMSTokenSync z(true); // not strictly needed.
2382 if (_collectorState == Resizing) {
2383 compute_new_size();
2384 save_heap_summary();
2385 _collectorState = Resetting;
2386 } else {
2387 assert(_collectorState == Idling, "The state should only change"
2388 " because the foreground collector has finished the collection");
2389 }
2390 }
2391 break;
2392 }
2393 case Resetting:
2394 // CMS heap resizing has been completed
2395 reset(true);
2396 assert(_collectorState == Idling, "Collector state should "
2397 "have changed");
2398
2399 MetaspaceGC::set_should_concurrent_collect(false);
2400
2401 stats().record_cms_end();
2402 // Don't move the concurrent_phases_end() and compute_new_size()
2403 // calls to here because a preempted background collection
2404 // has it's state set to "Resetting".
2405 break;
2406 case Idling:
2407 default:
2408 ShouldNotReachHere();
2409 break;
2410 }
2411 if (TraceCMSState) {
2412 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2413 Thread::current(), _collectorState);
2414 }
2415 assert(_foregroundGCShouldWait, "block post-condition");
2416 }
2417
2418 // Should this be in gc_epilogue?
2419 collector_policy()->counters()->update_counters();
2420
2421 {
2422 // Clear _foregroundGCShouldWait and, in the event that the
2423 // foreground collector is waiting, notify it, before
2424 // returning.
2425 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2426 _foregroundGCShouldWait = false;
2427 if (_foregroundGCIsActive) {
2428 CGC_lock->notify();
2429 }
2430 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2431 "Possible deadlock");
2432 }
2433 if (TraceCMSState) {
2434 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2435 " exiting collection CMS state %d",
2436 Thread::current(), _collectorState);
2437 }
2438 if (PrintGC && Verbose) {
2439 _cmsGen->print_heap_change(prev_used);
2440 }
2441 }
2442
2443 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) {
2444 if (!_cms_start_registered) {
2445 register_gc_start(cause);
2446 }
2447 }
2448
2449 void CMSCollector::register_gc_start(GCCause::Cause cause) {
2450 _cms_start_registered = true;
2451 _gc_timer_cm->register_gc_start(os::elapsed_counter());
2452 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
2453 }
2454
2455 void CMSCollector::register_gc_end() {
2456 if (_cms_start_registered) {
2457 report_heap_summary(GCWhen::AfterGC);
2458
2459 _gc_timer_cm->register_gc_end(os::elapsed_counter());
2460 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
2461 _cms_start_registered = false;
2462 }
2463 }
2464
2465 void CMSCollector::save_heap_summary() {
2466 GenCollectedHeap* gch = GenCollectedHeap::heap();
2467 _last_heap_summary = gch->create_heap_summary();
2468 _last_metaspace_summary = gch->create_metaspace_summary();
2469 }
2470
2471 void CMSCollector::report_heap_summary(GCWhen::Type when) {
2472 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary, _last_metaspace_summary);
2473 }
2474
2475 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) {
2476 assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2477 "Foreground collector should be waiting, not executing");
2478 assert(Thread::current()->is_VM_thread(), "A foreground collection"
2479 "may only be done by the VM Thread with the world stopped");
2480 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2481 "VM thread should have CMS token");
2482
2483 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2484 true, NULL);)
2485 if (UseAdaptiveSizePolicy) {
2486 size_policy()->ms_collection_begin();
2487 }
2488 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2489
2490 HandleMark hm; // Discard invalid handles created during verification
2491
2492 if (VerifyBeforeGC &&
2493 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2494 Universe::verify();
2495 }
2496
2497 // Snapshot the soft reference policy to be used in this collection cycle.
2498 ref_processor()->setup_policy(clear_all_soft_refs);
2499
2500 bool init_mark_was_synchronous = false; // until proven otherwise
2501 while (_collectorState != Idling) {
2502 if (TraceCMSState) {
2503 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2504 Thread::current(), _collectorState);
2505 }
2506 switch (_collectorState) {
2507 case InitialMarking:
2508 register_foreground_gc_start(cause);
2509 init_mark_was_synchronous = true; // fact to be exploited in re-mark
2510 checkpointRootsInitial(false);
2511 assert(_collectorState == Marking, "Collector state should have changed"
2512 " within checkpointRootsInitial()");
2513 break;
2514 case Marking:
2515 // initial marking in checkpointRootsInitialWork has been completed
2516 if (VerifyDuringGC &&
2517 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2518 Universe::verify("Verify before initial mark: ");
2519 }
2520 {
2521 bool res = markFromRoots(false);
2522 assert(res && _collectorState == FinalMarking, "Collector state should "
2523 "have changed");
2524 break;
2525 }
2526 case FinalMarking:
2527 if (VerifyDuringGC &&
2528 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2529 Universe::verify("Verify before re-mark: ");
2530 }
2531 checkpointRootsFinal(false, clear_all_soft_refs,
2532 init_mark_was_synchronous);
2533 assert(_collectorState == Sweeping, "Collector state should not "
2534 "have changed within checkpointRootsFinal()");
2535 break;
2536 case Sweeping:
2537 // final marking in checkpointRootsFinal has been completed
2538 if (VerifyDuringGC &&
2539 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2540 Universe::verify("Verify before sweep: ");
2541 }
2542 sweep(false);
2543 assert(_collectorState == Resizing, "Incorrect state");
2544 break;
2545 case Resizing: {
2546 // Sweeping has been completed; the actual resize in this case
2547 // is done separately; nothing to be done in this state.
2548 _collectorState = Resetting;
2549 break;
2550 }
2551 case Resetting:
2552 // The heap has been resized.
2553 if (VerifyDuringGC &&
2554 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2555 Universe::verify("Verify before reset: ");
2556 }
2557 save_heap_summary();
2558 reset(false);
2559 assert(_collectorState == Idling, "Collector state should "
2560 "have changed");
2561 break;
2562 case Precleaning:
2563 case AbortablePreclean:
2564 // Elide the preclean phase
2565 _collectorState = FinalMarking;
2566 break;
2567 default:
2568 ShouldNotReachHere();
2569 }
2570 if (TraceCMSState) {
2571 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2572 Thread::current(), _collectorState);
2573 }
2574 }
2575
2576 if (UseAdaptiveSizePolicy) {
2577 GenCollectedHeap* gch = GenCollectedHeap::heap();
2578 size_policy()->ms_collection_end(gch->gc_cause());
2579 }
2580
2581 if (VerifyAfterGC &&
2582 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2583 Universe::verify();
2584 }
2585 if (TraceCMSState) {
2586 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2587 " exiting collection CMS state %d",
2588 Thread::current(), _collectorState);
2589 }
2590 }
2591
2592 bool CMSCollector::waitForForegroundGC() {
2593 bool res = false;
2594 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2595 "CMS thread should have CMS token");
2596 // Block the foreground collector until the
2597 // background collectors decides whether to
2598 // yield.
2599 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2600 _foregroundGCShouldWait = true;
2601 if (_foregroundGCIsActive) {
2602 // The background collector yields to the
2603 // foreground collector and returns a value
2604 // indicating that it has yielded. The foreground
2605 // collector can proceed.
2606 res = true;
2607 _foregroundGCShouldWait = false;
2608 ConcurrentMarkSweepThread::clear_CMS_flag(
2609 ConcurrentMarkSweepThread::CMS_cms_has_token);
2610 ConcurrentMarkSweepThread::set_CMS_flag(
2611 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2612 // Get a possibly blocked foreground thread going
2613 CGC_lock->notify();
2614 if (TraceCMSState) {
2615 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2616 Thread::current(), _collectorState);
2617 }
2618 while (_foregroundGCIsActive) {
2619 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2620 }
2621 ConcurrentMarkSweepThread::set_CMS_flag(
2622 ConcurrentMarkSweepThread::CMS_cms_has_token);
2623 ConcurrentMarkSweepThread::clear_CMS_flag(
2624 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2625 }
2626 if (TraceCMSState) {
2627 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2628 Thread::current(), _collectorState);
2629 }
2630 return res;
2631 }
2632
2633 // Because of the need to lock the free lists and other structures in
2634 // the collector, common to all the generations that the collector is
2635 // collecting, we need the gc_prologues of individual CMS generations
2636 // delegate to their collector. It may have been simpler had the
2637 // current infrastructure allowed one to call a prologue on a
2638 // collector. In the absence of that we have the generation's
2639 // prologue delegate to the collector, which delegates back
2640 // some "local" work to a worker method in the individual generations
2641 // that it's responsible for collecting, while itself doing any
2642 // work common to all generations it's responsible for. A similar
2643 // comment applies to the gc_epilogue()'s.
2644 // The role of the varaible _between_prologue_and_epilogue is to
2645 // enforce the invocation protocol.
2646 void CMSCollector::gc_prologue(bool full) {
2647 // Call gc_prologue_work() for the CMSGen
2648 // we are responsible for.
2649
2650 // The following locking discipline assumes that we are only called
2651 // when the world is stopped.
2652 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2653
2654 // The CMSCollector prologue must call the gc_prologues for the
2655 // "generations" that it's responsible
2656 // for.
2657
2658 assert( Thread::current()->is_VM_thread()
2659 || ( CMSScavengeBeforeRemark
2660 && Thread::current()->is_ConcurrentGC_thread()),
2661 "Incorrect thread type for prologue execution");
2662
2663 if (_between_prologue_and_epilogue) {
2664 // We have already been invoked; this is a gc_prologue delegation
2665 // from yet another CMS generation that we are responsible for, just
2666 // ignore it since all relevant work has already been done.
2667 return;
2668 }
2669
2670 // set a bit saying prologue has been called; cleared in epilogue
2671 _between_prologue_and_epilogue = true;
2672 // Claim locks for common data structures, then call gc_prologue_work()
2673 // for each CMSGen.
2674
2675 getFreelistLocks(); // gets free list locks on constituent spaces
2676 bitMapLock()->lock_without_safepoint_check();
2677
2678 // Should call gc_prologue_work() for all cms gens we are responsible for
2679 bool duringMarking = _collectorState >= Marking
2680 && _collectorState < Sweeping;
2681
2682 // The young collections clear the modified oops state, which tells if
2683 // there are any modified oops in the class. The remark phase also needs
2684 // that information. Tell the young collection to save the union of all
2685 // modified klasses.
2686 if (duringMarking) {
2687 _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2688 }
2689
2690 bool registerClosure = duringMarking;
2691
2692 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ?
2693 &_modUnionClosurePar
2694 : &_modUnionClosure;
2695 _cmsGen->gc_prologue_work(full, registerClosure, muc);
2696
2697 if (!full) {
2698 stats().record_gc0_begin();
2699 }
2700 }
2701
2702 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2703
2704 _capacity_at_prologue = capacity();
2705 _used_at_prologue = used();
2706
2707 // Delegate to CMScollector which knows how to coordinate between
2708 // this and any other CMS generations that it is responsible for
2709 // collecting.
2710 collector()->gc_prologue(full);
2711 }
2712
2713 // This is a "private" interface for use by this generation's CMSCollector.
2714 // Not to be called directly by any other entity (for instance,
2715 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2716 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2717 bool registerClosure, ModUnionClosure* modUnionClosure) {
2718 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2719 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2720 "Should be NULL");
2721 if (registerClosure) {
2722 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2723 }
2724 cmsSpace()->gc_prologue();
2725 // Clear stat counters
2726 NOT_PRODUCT(
2727 assert(_numObjectsPromoted == 0, "check");
2728 assert(_numWordsPromoted == 0, "check");
2729 if (Verbose && PrintGC) {
2730 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2731 SIZE_FORMAT" bytes concurrently",
2732 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2733 }
2734 _numObjectsAllocated = 0;
2735 _numWordsAllocated = 0;
2736 )
2737 }
2738
2739 void CMSCollector::gc_epilogue(bool full) {
2740 // The following locking discipline assumes that we are only called
2741 // when the world is stopped.
2742 assert(SafepointSynchronize::is_at_safepoint(),
2743 "world is stopped assumption");
2744
2745 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2746 // if linear allocation blocks need to be appropriately marked to allow the
2747 // the blocks to be parsable. We also check here whether we need to nudge the
2748 // CMS collector thread to start a new cycle (if it's not already active).
2749 assert( Thread::current()->is_VM_thread()
2750 || ( CMSScavengeBeforeRemark
2751 && Thread::current()->is_ConcurrentGC_thread()),
2752 "Incorrect thread type for epilogue execution");
2753
2754 if (!_between_prologue_and_epilogue) {
2755 // We have already been invoked; this is a gc_epilogue delegation
2756 // from yet another CMS generation that we are responsible for, just
2757 // ignore it since all relevant work has already been done.
2758 return;
2759 }
2760 assert(haveFreelistLocks(), "must have freelist locks");
2761 assert_lock_strong(bitMapLock());
2762
2763 _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2764
2765 _cmsGen->gc_epilogue_work(full);
2766
2767 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2768 // in case sampling was not already enabled, enable it
2769 _start_sampling = true;
2770 }
2771 // reset _eden_chunk_array so sampling starts afresh
2772 _eden_chunk_index = 0;
2773
2774 size_t cms_used = _cmsGen->cmsSpace()->used();
2775
2776 // update performance counters - this uses a special version of
2777 // update_counters() that allows the utilization to be passed as a
2778 // parameter, avoiding multiple calls to used().
2779 //
2780 _cmsGen->update_counters(cms_used);
2781
2782 if (CMSIncrementalMode) {
2783 icms_update_allocation_limits();
2784 }
2785
2786 bitMapLock()->unlock();
2787 releaseFreelistLocks();
2788
2789 if (!CleanChunkPoolAsync) {
2790 Chunk::clean_chunk_pool();
2791 }
2792
2793 set_did_compact(false);
2794 _between_prologue_and_epilogue = false; // ready for next cycle
2795 }
2796
2797 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2798 collector()->gc_epilogue(full);
2799
2800 // Also reset promotion tracking in par gc thread states.
2801 if (CollectedHeap::use_parallel_gc_threads()) {
2802 for (uint i = 0; i < ParallelGCThreads; i++) {
2803 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2804 }
2805 }
2806 }
2807
2808 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2809 assert(!incremental_collection_failed(), "Should have been cleared");
2810 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2811 cmsSpace()->gc_epilogue();
2812 // Print stat counters
2813 NOT_PRODUCT(
2814 assert(_numObjectsAllocated == 0, "check");
2815 assert(_numWordsAllocated == 0, "check");
2816 if (Verbose && PrintGC) {
2817 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2818 SIZE_FORMAT" bytes",
2819 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2820 }
2821 _numObjectsPromoted = 0;
2822 _numWordsPromoted = 0;
2823 )
2824
2825 if (PrintGC && Verbose) {
2826 // Call down the chain in contiguous_available needs the freelistLock
2827 // so print this out before releasing the freeListLock.
2828 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2829 contiguous_available());
2830 }
2831 }
2832
2833 #ifndef PRODUCT
2834 bool CMSCollector::have_cms_token() {
2835 Thread* thr = Thread::current();
2836 if (thr->is_VM_thread()) {
2837 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2838 } else if (thr->is_ConcurrentGC_thread()) {
2839 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2840 } else if (thr->is_GC_task_thread()) {
2841 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2842 ParGCRareEvent_lock->owned_by_self();
2843 }
2844 return false;
2845 }
2846 #endif
2847
2848 // Check reachability of the given heap address in CMS generation,
2849 // treating all other generations as roots.
2850 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2851 // We could "guarantee" below, rather than assert, but i'll
2852 // leave these as "asserts" so that an adventurous debugger
2853 // could try this in the product build provided some subset of
2854 // the conditions were met, provided they were intersted in the
2855 // results and knew that the computation below wouldn't interfere
2856 // with other concurrent computations mutating the structures
2857 // being read or written.
2858 assert(SafepointSynchronize::is_at_safepoint(),
2859 "Else mutations in object graph will make answer suspect");
2860 assert(have_cms_token(), "Should hold cms token");
2861 assert(haveFreelistLocks(), "must hold free list locks");
2862 assert_lock_strong(bitMapLock());
2863
2864 // Clear the marking bit map array before starting, but, just
2865 // for kicks, first report if the given address is already marked
2866 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
2867 _markBitMap.isMarked(addr) ? "" : " not");
2868
2869 if (verify_after_remark()) {
2870 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2871 bool result = verification_mark_bm()->isMarked(addr);
2872 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
2873 result ? "IS" : "is NOT");
2874 return result;
2875 } else {
2876 gclog_or_tty->print_cr("Could not compute result");
2877 return false;
2878 }
2879 }
2880
2881
2882 void
2883 CMSCollector::print_on_error(outputStream* st) {
2884 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2885 if (collector != NULL) {
2886 CMSBitMap* bitmap = &collector->_markBitMap;
2887 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap);
2888 bitmap->print_on_error(st, " Bits: ");
2889
2890 st->cr();
2891
2892 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2893 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap);
2894 mut_bitmap->print_on_error(st, " Bits: ");
2895 }
2896 }
2897
2898 ////////////////////////////////////////////////////////
2899 // CMS Verification Support
2900 ////////////////////////////////////////////////////////
2901 // Following the remark phase, the following invariant
2902 // should hold -- each object in the CMS heap which is
2903 // marked in markBitMap() should be marked in the verification_mark_bm().
2904
2905 class VerifyMarkedClosure: public BitMapClosure {
2906 CMSBitMap* _marks;
2907 bool _failed;
2908
2909 public:
2910 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2911
2912 bool do_bit(size_t offset) {
2913 HeapWord* addr = _marks->offsetToHeapWord(offset);
2914 if (!_marks->isMarked(addr)) {
2915 oop(addr)->print_on(gclog_or_tty);
2916 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2917 _failed = true;
2918 }
2919 return true;
2920 }
2921
2922 bool failed() { return _failed; }
2923 };
2924
2925 bool CMSCollector::verify_after_remark(bool silent) {
2926 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... ");
2927 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2928 static bool init = false;
2929
2930 assert(SafepointSynchronize::is_at_safepoint(),
2931 "Else mutations in object graph will make answer suspect");
2932 assert(have_cms_token(),
2933 "Else there may be mutual interference in use of "
2934 " verification data structures");
2935 assert(_collectorState > Marking && _collectorState <= Sweeping,
2936 "Else marking info checked here may be obsolete");
2937 assert(haveFreelistLocks(), "must hold free list locks");
2938 assert_lock_strong(bitMapLock());
2939
2940
2941 // Allocate marking bit map if not already allocated
2942 if (!init) { // first time
2943 if (!verification_mark_bm()->allocate(_span)) {
2944 return false;
2945 }
2946 init = true;
2947 }
2948
2949 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2950
2951 // Turn off refs discovery -- so we will be tracing through refs.
2952 // This is as intended, because by this time
2953 // GC must already have cleared any refs that need to be cleared,
2954 // and traced those that need to be marked; moreover,
2955 // the marking done here is not going to intefere in any
2956 // way with the marking information used by GC.
2957 NoRefDiscovery no_discovery(ref_processor());
2958
2959 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2960
2961 // Clear any marks from a previous round
2962 verification_mark_bm()->clear_all();
2963 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2964 verify_work_stacks_empty();
2965
2966 GenCollectedHeap* gch = GenCollectedHeap::heap();
2967 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2968 // Update the saved marks which may affect the root scans.
2969 gch->save_marks();
2970
2971 if (CMSRemarkVerifyVariant == 1) {
2972 // In this first variant of verification, we complete
2973 // all marking, then check if the new marks-verctor is
2974 // a subset of the CMS marks-vector.
2975 verify_after_remark_work_1();
2976 } else if (CMSRemarkVerifyVariant == 2) {
2977 // In this second variant of verification, we flag an error
2978 // (i.e. an object reachable in the new marks-vector not reachable
2979 // in the CMS marks-vector) immediately, also indicating the
2980 // identify of an object (A) that references the unmarked object (B) --
2981 // presumably, a mutation to A failed to be picked up by preclean/remark?
2982 verify_after_remark_work_2();
2983 } else {
2984 warning("Unrecognized value %d for CMSRemarkVerifyVariant",
2985 CMSRemarkVerifyVariant);
2986 }
2987 if (!silent) gclog_or_tty->print(" done] ");
2988 return true;
2989 }
2990
2991 void CMSCollector::verify_after_remark_work_1() {
2992 ResourceMark rm;
2993 HandleMark hm;
2994 GenCollectedHeap* gch = GenCollectedHeap::heap();
2995
2996 // Get a clear set of claim bits for the strong roots processing to work with.
2997 ClassLoaderDataGraph::clear_claimed_marks();
2998
2999 // Mark from roots one level into CMS
3000 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
3001 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3002
3003 gch->gen_process_strong_roots(_cmsGen->level(),
3004 true, // younger gens are roots
3005 true, // activate StrongRootsScope
3006 false, // not scavenging
3007 SharedHeap::ScanningOption(roots_scanning_options()),
3008 ¬Older,
3009 true, // walk code active on stacks
3010 NULL,
3011 NULL); // SSS: Provide correct closure
3012
3013 // Now mark from the roots
3014 MarkFromRootsClosure markFromRootsClosure(this, _span,
3015 verification_mark_bm(), verification_mark_stack(),
3016 false /* don't yield */, true /* verifying */);
3017 assert(_restart_addr == NULL, "Expected pre-condition");
3018 verification_mark_bm()->iterate(&markFromRootsClosure);
3019 while (_restart_addr != NULL) {
3020 // Deal with stack overflow: by restarting at the indicated
3021 // address.
3022 HeapWord* ra = _restart_addr;
3023 markFromRootsClosure.reset(ra);
3024 _restart_addr = NULL;
3025 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
3026 }
3027 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
3028 verify_work_stacks_empty();
3029
3030 // Marking completed -- now verify that each bit marked in
3031 // verification_mark_bm() is also marked in markBitMap(); flag all
3032 // errors by printing corresponding objects.
3033 VerifyMarkedClosure vcl(markBitMap());
3034 verification_mark_bm()->iterate(&vcl);
3035 if (vcl.failed()) {
3036 gclog_or_tty->print("Verification failed");
3037 Universe::heap()->print_on(gclog_or_tty);
3038 fatal("CMS: failed marking verification after remark");
3039 }
3040 }
3041
3042 class VerifyKlassOopsKlassClosure : public KlassClosure {
3043 class VerifyKlassOopsClosure : public OopClosure {
3044 CMSBitMap* _bitmap;
3045 public:
3046 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
3047 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
3048 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
3049 } _oop_closure;
3050 public:
3051 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
3052 void do_klass(Klass* k) {
3053 k->oops_do(&_oop_closure);
3054 }
3055 };
3056
3057 void CMSCollector::verify_after_remark_work_2() {
3058 ResourceMark rm;
3059 HandleMark hm;
3060 GenCollectedHeap* gch = GenCollectedHeap::heap();
3061
3062 // Get a clear set of claim bits for the strong roots processing to work with.
3063 ClassLoaderDataGraph::clear_claimed_marks();
3064
3065 // Mark from roots one level into CMS
3066 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
3067 markBitMap());
3068 CMKlassClosure klass_closure(¬Older);
3069
3070 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3071 gch->gen_process_strong_roots(_cmsGen->level(),
3072 true, // younger gens are roots
3073 true, // activate StrongRootsScope
3074 false, // not scavenging
3075 SharedHeap::ScanningOption(roots_scanning_options()),
3076 ¬Older,
3077 true, // walk code active on stacks
3078 NULL,
3079 &klass_closure);
3080
3081 // Now mark from the roots
3082 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
3083 verification_mark_bm(), markBitMap(), verification_mark_stack());
3084 assert(_restart_addr == NULL, "Expected pre-condition");
3085 verification_mark_bm()->iterate(&markFromRootsClosure);
3086 while (_restart_addr != NULL) {
3087 // Deal with stack overflow: by restarting at the indicated
3088 // address.
3089 HeapWord* ra = _restart_addr;
3090 markFromRootsClosure.reset(ra);
3091 _restart_addr = NULL;
3092 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
3093 }
3094 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
3095 verify_work_stacks_empty();
3096
3097 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
3098 ClassLoaderDataGraph::classes_do(&verify_klass_oops);
3099
3100 // Marking completed -- now verify that each bit marked in
3101 // verification_mark_bm() is also marked in markBitMap(); flag all
3102 // errors by printing corresponding objects.
3103 VerifyMarkedClosure vcl(markBitMap());
3104 verification_mark_bm()->iterate(&vcl);
3105 assert(!vcl.failed(), "Else verification above should not have succeeded");
3106 }
3107
3108 void ConcurrentMarkSweepGeneration::save_marks() {
3109 // delegate to CMS space
3110 cmsSpace()->save_marks();
3111 for (uint i = 0; i < ParallelGCThreads; i++) {
3112 _par_gc_thread_states[i]->promo.startTrackingPromotions();
3113 }
3114 }
3115
3116 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
3117 return cmsSpace()->no_allocs_since_save_marks();
3118 }
3119
3120 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
3121 \
3122 void ConcurrentMarkSweepGeneration:: \
3123 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
3124 cl->set_generation(this); \
3125 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
3126 cl->reset_generation(); \
3127 save_marks(); \
3128 }
3129
3130 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
3131
3132 void
3133 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk)
3134 {
3135 // Not currently implemented; need to do the following. -- ysr.
3136 // dld -- I think that is used for some sort of allocation profiler. So it
3137 // really means the objects allocated by the mutator since the last
3138 // GC. We could potentially implement this cheaply by recording only
3139 // the direct allocations in a side data structure.
3140 //
3141 // I think we probably ought not to be required to support these
3142 // iterations at any arbitrary point; I think there ought to be some
3143 // call to enable/disable allocation profiling in a generation/space,
3144 // and the iterator ought to return the objects allocated in the
3145 // gen/space since the enable call, or the last iterator call (which
3146 // will probably be at a GC.) That way, for gens like CM&S that would
3147 // require some extra data structure to support this, we only pay the
3148 // cost when it's in use...
3149 cmsSpace()->object_iterate_since_last_GC(blk);
3150 }
3151
3152 void
3153 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
3154 cl->set_generation(this);
3155 younger_refs_in_space_iterate(_cmsSpace, cl);
3156 cl->reset_generation();
3157 }
3158
3159 void
3160 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) {
3161 if (freelistLock()->owned_by_self()) {
3162 Generation::oop_iterate(mr, cl);
3163 } else {
3164 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3165 Generation::oop_iterate(mr, cl);
3166 }
3167 }
3168
3169 void
3170 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
3171 if (freelistLock()->owned_by_self()) {
3172 Generation::oop_iterate(cl);
3173 } else {
3174 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3175 Generation::oop_iterate(cl);
3176 }
3177 }
3178
3179 void
3180 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
3181 if (freelistLock()->owned_by_self()) {
3182 Generation::object_iterate(cl);
3183 } else {
3184 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3185 Generation::object_iterate(cl);
3186 }
3187 }
3188
3189 void
3190 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
3191 if (freelistLock()->owned_by_self()) {
3192 Generation::safe_object_iterate(cl);
3193 } else {
3194 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3195 Generation::safe_object_iterate(cl);
3196 }
3197 }
3198
3199 void
3200 ConcurrentMarkSweepGeneration::post_compact() {
3201 }
3202
3203 void
3204 ConcurrentMarkSweepGeneration::prepare_for_verify() {
3205 // Fix the linear allocation blocks to look like free blocks.
3206
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()->prepare_for_verify();
3212 } else {
3213 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3214 cmsSpace()->prepare_for_verify();
3215 }
3216 }
3217
3218 void
3219 ConcurrentMarkSweepGeneration::verify() {
3220 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3221 // are not called when the heap is verified during universe initialization and
3222 // at vm shutdown.
3223 if (freelistLock()->owned_by_self()) {
3224 cmsSpace()->verify();
3225 } else {
3226 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3227 cmsSpace()->verify();
3228 }
3229 }
3230
3231 void CMSCollector::verify() {
3232 _cmsGen->verify();
3233 }
3234
3235 #ifndef PRODUCT
3236 bool CMSCollector::overflow_list_is_empty() const {
3237 assert(_num_par_pushes >= 0, "Inconsistency");
3238 if (_overflow_list == NULL) {
3239 assert(_num_par_pushes == 0, "Inconsistency");
3240 }
3241 return _overflow_list == NULL;
3242 }
3243
3244 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3245 // merely consolidate assertion checks that appear to occur together frequently.
3246 void CMSCollector::verify_work_stacks_empty() const {
3247 assert(_markStack.isEmpty(), "Marking stack should be empty");
3248 assert(overflow_list_is_empty(), "Overflow list should be empty");
3249 }
3250
3251 void CMSCollector::verify_overflow_empty() const {
3252 assert(overflow_list_is_empty(), "Overflow list should be empty");
3253 assert(no_preserved_marks(), "No preserved marks");
3254 }
3255 #endif // PRODUCT
3256
3257 // Decide if we want to enable class unloading as part of the
3258 // ensuing concurrent GC cycle. We will collect and
3259 // unload classes if it's the case that:
3260 // (1) an explicit gc request has been made and the flag
3261 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3262 // (2) (a) class unloading is enabled at the command line, and
3263 // (b) old gen is getting really full
3264 // NOTE: Provided there is no change in the state of the heap between
3265 // calls to this method, it should have idempotent results. Moreover,
3266 // its results should be monotonically increasing (i.e. going from 0 to 1,
3267 // but not 1 to 0) between successive calls between which the heap was
3268 // not collected. For the implementation below, it must thus rely on
3269 // the property that concurrent_cycles_since_last_unload()
3270 // will not decrease unless a collection cycle happened and that
3271 // _cmsGen->is_too_full() are
3272 // themselves also monotonic in that sense. See check_monotonicity()
3273 // below.
3274 void CMSCollector::update_should_unload_classes() {
3275 _should_unload_classes = false;
3276 // Condition 1 above
3277 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3278 _should_unload_classes = true;
3279 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3280 // Disjuncts 2.b.(i,ii,iii) above
3281 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3282 CMSClassUnloadingMaxInterval)
3283 || _cmsGen->is_too_full();
3284 }
3285 }
3286
3287 bool ConcurrentMarkSweepGeneration::is_too_full() const {
3288 bool res = should_concurrent_collect();
3289 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
3290 return res;
3291 }
3292
3293 void CMSCollector::setup_cms_unloading_and_verification_state() {
3294 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3295 || VerifyBeforeExit;
3296 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
3297
3298 if (should_unload_classes()) { // Should unload classes this cycle
3299 remove_root_scanning_option(rso); // Shrink the root set appropriately
3300 set_verifying(should_verify); // Set verification state for this cycle
3301 return; // Nothing else needs to be done at this time
3302 }
3303
3304 // Not unloading classes this cycle
3305 assert(!should_unload_classes(), "Inconsitency!");
3306 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3307 // Include symbols, strings and code cache elements to prevent their resurrection.
3308 add_root_scanning_option(rso);
3309 set_verifying(true);
3310 } else if (verifying() && !should_verify) {
3311 // We were verifying, but some verification flags got disabled.
3312 set_verifying(false);
3313 // Exclude symbols, strings and code cache elements from root scanning to
3314 // reduce IM and RM pauses.
3315 remove_root_scanning_option(rso);
3316 }
3317 }
3318
3319
3320 #ifndef PRODUCT
3321 HeapWord* CMSCollector::block_start(const void* p) const {
3322 const HeapWord* addr = (HeapWord*)p;
3323 if (_span.contains(p)) {
3324 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3325 return _cmsGen->cmsSpace()->block_start(p);
3326 }
3327 }
3328 return NULL;
3329 }
3330 #endif
3331
3332 HeapWord*
3333 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3334 bool tlab,
3335 bool parallel) {
3336 CMSSynchronousYieldRequest yr;
3337 assert(!tlab, "Can't deal with TLAB allocation");
3338 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3339 expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3340 CMSExpansionCause::_satisfy_allocation);
3341 if (GCExpandToAllocateDelayMillis > 0) {
3342 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3343 }
3344 return have_lock_and_allocate(word_size, tlab);
3345 }
3346
3347 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3348 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3349 // to CardGeneration and share it...
3350 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) {
3351 return CardGeneration::expand(bytes, expand_bytes);
3352 }
3353
3354 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3355 CMSExpansionCause::Cause cause)
3356 {
3357
3358 bool success = expand(bytes, expand_bytes);
3359
3360 // remember why we expanded; this information is used
3361 // by shouldConcurrentCollect() when making decisions on whether to start
3362 // a new CMS cycle.
3363 if (success) {
3364 set_expansion_cause(cause);
3365 if (PrintGCDetails && Verbose) {
3366 gclog_or_tty->print_cr("Expanded CMS gen for %s",
3367 CMSExpansionCause::to_string(cause));
3368 }
3369 }
3370 }
3371
3372 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3373 HeapWord* res = NULL;
3374 MutexLocker x(ParGCRareEvent_lock);
3375 while (true) {
3376 // Expansion by some other thread might make alloc OK now:
3377 res = ps->lab.alloc(word_sz);
3378 if (res != NULL) return res;
3379 // If there's not enough expansion space available, give up.
3380 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3381 return NULL;
3382 }
3383 // Otherwise, we try expansion.
3384 expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3385 CMSExpansionCause::_allocate_par_lab);
3386 // Now go around the loop and try alloc again;
3387 // A competing par_promote might beat us to the expansion space,
3388 // so we may go around the loop again if promotion fails agaion.
3389 if (GCExpandToAllocateDelayMillis > 0) {
3390 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3391 }
3392 }
3393 }
3394
3395
3396 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3397 PromotionInfo* promo) {
3398 MutexLocker x(ParGCRareEvent_lock);
3399 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3400 while (true) {
3401 // Expansion by some other thread might make alloc OK now:
3402 if (promo->ensure_spooling_space()) {
3403 assert(promo->has_spooling_space(),
3404 "Post-condition of successful ensure_spooling_space()");
3405 return true;
3406 }
3407 // If there's not enough expansion space available, give up.
3408 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3409 return false;
3410 }
3411 // Otherwise, we try expansion.
3412 expand(refill_size_bytes, MinHeapDeltaBytes,
3413 CMSExpansionCause::_allocate_par_spooling_space);
3414 // Now go around the loop and try alloc again;
3415 // A competing allocation might beat us to the expansion space,
3416 // so we may go around the loop again if allocation fails again.
3417 if (GCExpandToAllocateDelayMillis > 0) {
3418 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3419 }
3420 }
3421 }
3422
3423
3424 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3425 assert_locked_or_safepoint(ExpandHeap_lock);
3426 // Shrink committed space
3427 _virtual_space.shrink_by(bytes);
3428 // Shrink space; this also shrinks the space's BOT
3429 _cmsSpace->set_end((HeapWord*) _virtual_space.high());
3430 size_t new_word_size = heap_word_size(_cmsSpace->capacity());
3431 // Shrink the shared block offset array
3432 _bts->resize(new_word_size);
3433 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3434 // Shrink the card table
3435 Universe::heap()->barrier_set()->resize_covered_region(mr);
3436
3437 if (Verbose && PrintGC) {
3438 size_t new_mem_size = _virtual_space.committed_size();
3439 size_t old_mem_size = new_mem_size + bytes;
3440 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
3441 name(), old_mem_size/K, new_mem_size/K);
3442 }
3443 }
3444
3445 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3446 assert_locked_or_safepoint(Heap_lock);
3447 size_t size = ReservedSpace::page_align_size_down(bytes);
3448 if (size > 0) {
3449 shrink_by(size);
3450 }
3451 }
3452
3453 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3454 assert_locked_or_safepoint(Heap_lock);
3455 bool result = _virtual_space.expand_by(bytes);
3456 if (result) {
3457 size_t new_word_size =
3458 heap_word_size(_virtual_space.committed_size());
3459 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3460 _bts->resize(new_word_size); // resize the block offset shared array
3461 Universe::heap()->barrier_set()->resize_covered_region(mr);
3462 // Hmmmm... why doesn't CFLS::set_end verify locking?
3463 // This is quite ugly; FIX ME XXX
3464 _cmsSpace->assert_locked(freelistLock());
3465 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3466
3467 // update the space and generation capacity counters
3468 if (UsePerfData) {
3469 _space_counters->update_capacity();
3470 _gen_counters->update_all();
3471 }
3472
3473 if (Verbose && PrintGC) {
3474 size_t new_mem_size = _virtual_space.committed_size();
3475 size_t old_mem_size = new_mem_size - bytes;
3476 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K",
3477 name(), old_mem_size/K, bytes/K, new_mem_size/K);
3478 }
3479 }
3480 return result;
3481 }
3482
3483 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3484 assert_locked_or_safepoint(Heap_lock);
3485 bool success = true;
3486 const size_t remaining_bytes = _virtual_space.uncommitted_size();
3487 if (remaining_bytes > 0) {
3488 success = grow_by(remaining_bytes);
3489 DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3490 }
3491 return success;
3492 }
3493
3494 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
3495 assert_locked_or_safepoint(Heap_lock);
3496 assert_lock_strong(freelistLock());
3497 if (PrintGCDetails && Verbose) {
3498 warning("Shrinking of CMS not yet implemented");
3499 }
3500 return;
3501 }
3502
3503
3504 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3505 // phases.
3506 class CMSPhaseAccounting: public StackObj {
3507 public:
3508 CMSPhaseAccounting(CMSCollector *collector,
3509 const char *phase,
3510 bool print_cr = true);
3511 ~CMSPhaseAccounting();
3512
3513 private:
3514 CMSCollector *_collector;
3515 const char *_phase;
3516 elapsedTimer _wallclock;
3517 bool _print_cr;
3518
3519 public:
3520 // Not MT-safe; so do not pass around these StackObj's
3521 // where they may be accessed by other threads.
3522 jlong wallclock_millis() {
3523 assert(_wallclock.is_active(), "Wall clock should not stop");
3524 _wallclock.stop(); // to record time
3525 jlong ret = _wallclock.milliseconds();
3526 _wallclock.start(); // restart
3527 return ret;
3528 }
3529 };
3530
3531 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3532 const char *phase,
3533 bool print_cr) :
3534 _collector(collector), _phase(phase), _print_cr(print_cr) {
3535
3536 if (PrintCMSStatistics != 0) {
3537 _collector->resetYields();
3538 }
3539 if (PrintGCDetails) {
3540 gclog_or_tty->date_stamp(PrintGCDateStamps);
3541 gclog_or_tty->stamp(PrintGCTimeStamps);
3542 gclog_or_tty->print_cr("[%s-concurrent-%s-start]",
3543 _collector->cmsGen()->short_name(), _phase);
3544 }
3545 _collector->resetTimer();
3546 _wallclock.start();
3547 _collector->startTimer();
3548 }
3549
3550 CMSPhaseAccounting::~CMSPhaseAccounting() {
3551 assert(_wallclock.is_active(), "Wall clock should not have stopped");
3552 _collector->stopTimer();
3553 _wallclock.stop();
3554 if (PrintGCDetails) {
3555 gclog_or_tty->date_stamp(PrintGCDateStamps);
3556 gclog_or_tty->stamp(PrintGCTimeStamps);
3557 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3558 _collector->cmsGen()->short_name(),
3559 _phase, _collector->timerValue(), _wallclock.seconds());
3560 if (_print_cr) {
3561 gclog_or_tty->print_cr("");
3562 }
3563 if (PrintCMSStatistics != 0) {
3564 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3565 _collector->yields());
3566 }
3567 }
3568 }
3569
3570 // CMS work
3571
3572 // Checkpoint the roots into this generation from outside
3573 // this generation. [Note this initial checkpoint need only
3574 // be approximate -- we'll do a catch up phase subsequently.]
3575 void CMSCollector::checkpointRootsInitial(bool asynch) {
3576 assert(_collectorState == InitialMarking, "Wrong collector state");
3577 check_correct_thread_executing();
3578 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
3579
3580 save_heap_summary();
3581 report_heap_summary(GCWhen::BeforeGC);
3582
3583 ReferenceProcessor* rp = ref_processor();
3584 SpecializationStats::clear();
3585 assert(_restart_addr == NULL, "Control point invariant");
3586 if (asynch) {
3587 // acquire locks for subsequent manipulations
3588 MutexLockerEx x(bitMapLock(),
3589 Mutex::_no_safepoint_check_flag);
3590 checkpointRootsInitialWork(asynch);
3591 // enable ("weak") refs discovery
3592 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/);
3593 _collectorState = Marking;
3594 } else {
3595 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3596 // which recognizes if we are a CMS generation, and doesn't try to turn on
3597 // discovery; verify that they aren't meddling.
3598 assert(!rp->discovery_is_atomic(),
3599 "incorrect setting of discovery predicate");
3600 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3601 "ref discovery for this generation kind");
3602 // already have locks
3603 checkpointRootsInitialWork(asynch);
3604 // now enable ("weak") refs discovery
3605 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/);
3606 _collectorState = Marking;
3607 }
3608 SpecializationStats::print();
3609 }
3610
3611 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3612 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3613 assert(_collectorState == InitialMarking, "just checking");
3614
3615 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3616 // precede our marking with a collection of all
3617 // younger generations to keep floating garbage to a minimum.
3618 // XXX: we won't do this for now -- it's an optimization to be done later.
3619
3620 // already have locks
3621 assert_lock_strong(bitMapLock());
3622 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3623
3624 // Setup the verification and class unloading state for this
3625 // CMS collection cycle.
3626 setup_cms_unloading_and_verification_state();
3627
3628 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
3629 PrintGCDetails && Verbose, true, _gc_timer_cm);)
3630 if (UseAdaptiveSizePolicy) {
3631 size_policy()->checkpoint_roots_initial_begin();
3632 }
3633
3634 // Reset all the PLAB chunk arrays if necessary.
3635 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3636 reset_survivor_plab_arrays();
3637 }
3638
3639 ResourceMark rm;
3640 HandleMark hm;
3641
3642 FalseClosure falseClosure;
3643 // In the case of a synchronous collection, we will elide the
3644 // remark step, so it's important to catch all the nmethod oops
3645 // in this step.
3646 // The final 'true' flag to gen_process_strong_roots will ensure this.
3647 // If 'async' is true, we can relax the nmethod tracing.
3648 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
3649 GenCollectedHeap* gch = GenCollectedHeap::heap();
3650
3651 verify_work_stacks_empty();
3652 verify_overflow_empty();
3653
3654 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3655 // Update the saved marks which may affect the root scans.
3656 gch->save_marks();
3657
3658 // weak reference processing has not started yet.
3659 ref_processor()->set_enqueuing_is_done(false);
3660
3661 // Need to remember all newly created CLDs,
3662 // so that we can guarantee that the remark finds them.
3663 ClassLoaderDataGraph::remember_new_clds(true);
3664
3665 // Whenever a CLD is found, it will be claimed before proceeding to mark
3666 // the klasses. The claimed marks need to be cleared before marking starts.
3667 ClassLoaderDataGraph::clear_claimed_marks();
3668
3669 CMKlassClosure klass_closure(¬Older);
3670 {
3671 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3672 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3673 gch->gen_process_strong_roots(_cmsGen->level(),
3674 true, // younger gens are roots
3675 true, // activate StrongRootsScope
3676 false, // not scavenging
3677 SharedHeap::ScanningOption(roots_scanning_options()),
3678 ¬Older,
3679 true, // walk all of code cache if (so & SO_CodeCache)
3680 NULL,
3681 &klass_closure);
3682 }
3683
3684 // Clear mod-union table; it will be dirtied in the prologue of
3685 // CMS generation per each younger generation collection.
3686
3687 assert(_modUnionTable.isAllClear(),
3688 "Was cleared in most recent final checkpoint phase"
3689 " or no bits are set in the gc_prologue before the start of the next "
3690 "subsequent marking phase.");
3691
3692 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
3693
3694 // Save the end of the used_region of the constituent generations
3695 // to be used to limit the extent of sweep in each generation.
3696 save_sweep_limits();
3697 if (UseAdaptiveSizePolicy) {
3698 size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3699 }
3700 verify_overflow_empty();
3701 }
3702
3703 bool CMSCollector::markFromRoots(bool asynch) {
3704 // we might be tempted to assert that:
3705 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3706 // "inconsistent argument?");
3707 // However that wouldn't be right, because it's possible that
3708 // a safepoint is indeed in progress as a younger generation
3709 // stop-the-world GC happens even as we mark in this generation.
3710 assert(_collectorState == Marking, "inconsistent state?");
3711 check_correct_thread_executing();
3712 verify_overflow_empty();
3713
3714 bool res;
3715 if (asynch) {
3716
3717 // Start the timers for adaptive size policy for the concurrent phases
3718 // Do it here so that the foreground MS can use the concurrent
3719 // timer since a foreground MS might has the sweep done concurrently
3720 // or STW.
3721 if (UseAdaptiveSizePolicy) {
3722 size_policy()->concurrent_marking_begin();
3723 }
3724
3725 // Weak ref discovery note: We may be discovering weak
3726 // refs in this generation concurrent (but interleaved) with
3727 // weak ref discovery by a younger generation collector.
3728
3729 CMSTokenSyncWithLocks ts(true, bitMapLock());
3730 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3731 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3732 res = markFromRootsWork(asynch);
3733 if (res) {
3734 _collectorState = Precleaning;
3735 } else { // We failed and a foreground collection wants to take over
3736 assert(_foregroundGCIsActive, "internal state inconsistency");
3737 assert(_restart_addr == NULL, "foreground will restart from scratch");
3738 if (PrintGCDetails) {
3739 gclog_or_tty->print_cr("bailing out to foreground collection");
3740 }
3741 }
3742 if (UseAdaptiveSizePolicy) {
3743 size_policy()->concurrent_marking_end();
3744 }
3745 } else {
3746 assert(SafepointSynchronize::is_at_safepoint(),
3747 "inconsistent with asynch == false");
3748 if (UseAdaptiveSizePolicy) {
3749 size_policy()->ms_collection_marking_begin();
3750 }
3751 // already have locks
3752 res = markFromRootsWork(asynch);
3753 _collectorState = FinalMarking;
3754 if (UseAdaptiveSizePolicy) {
3755 GenCollectedHeap* gch = GenCollectedHeap::heap();
3756 size_policy()->ms_collection_marking_end(gch->gc_cause());
3757 }
3758 }
3759 verify_overflow_empty();
3760 return res;
3761 }
3762
3763 bool CMSCollector::markFromRootsWork(bool asynch) {
3764 // iterate over marked bits in bit map, doing a full scan and mark
3765 // from these roots using the following algorithm:
3766 // . if oop is to the right of the current scan pointer,
3767 // mark corresponding bit (we'll process it later)
3768 // . else (oop is to left of current scan pointer)
3769 // push oop on marking stack
3770 // . drain the marking stack
3771
3772 // Note that when we do a marking step we need to hold the
3773 // bit map lock -- recall that direct allocation (by mutators)
3774 // and promotion (by younger generation collectors) is also
3775 // marking the bit map. [the so-called allocate live policy.]
3776 // Because the implementation of bit map marking is not
3777 // robust wrt simultaneous marking of bits in the same word,
3778 // we need to make sure that there is no such interference
3779 // between concurrent such updates.
3780
3781 // already have locks
3782 assert_lock_strong(bitMapLock());
3783
3784 verify_work_stacks_empty();
3785 verify_overflow_empty();
3786 bool result = false;
3787 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3788 result = do_marking_mt(asynch);
3789 } else {
3790 result = do_marking_st(asynch);
3791 }
3792 return result;
3793 }
3794
3795 // Forward decl
3796 class CMSConcMarkingTask;
3797
3798 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3799 CMSCollector* _collector;
3800 CMSConcMarkingTask* _task;
3801 public:
3802 virtual void yield();
3803
3804 // "n_threads" is the number of threads to be terminated.
3805 // "queue_set" is a set of work queues of other threads.
3806 // "collector" is the CMS collector associated with this task terminator.
3807 // "yield" indicates whether we need the gang as a whole to yield.
3808 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3809 ParallelTaskTerminator(n_threads, queue_set),
3810 _collector(collector) { }
3811
3812 void set_task(CMSConcMarkingTask* task) {
3813 _task = task;
3814 }
3815 };
3816
3817 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3818 CMSConcMarkingTask* _task;
3819 public:
3820 bool should_exit_termination();
3821 void set_task(CMSConcMarkingTask* task) {
3822 _task = task;
3823 }
3824 };
3825
3826 // MT Concurrent Marking Task
3827 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3828 CMSCollector* _collector;
3829 int _n_workers; // requested/desired # workers
3830 bool _asynch;
3831 bool _result;
3832 CompactibleFreeListSpace* _cms_space;
3833 char _pad_front[64]; // padding to ...
3834 HeapWord* _global_finger; // ... avoid sharing cache line
3835 char _pad_back[64];
3836 HeapWord* _restart_addr;
3837
3838 // Exposed here for yielding support
3839 Mutex* const _bit_map_lock;
3840
3841 // The per thread work queues, available here for stealing
3842 OopTaskQueueSet* _task_queues;
3843
3844 // Termination (and yielding) support
3845 CMSConcMarkingTerminator _term;
3846 CMSConcMarkingTerminatorTerminator _term_term;
3847
3848 public:
3849 CMSConcMarkingTask(CMSCollector* collector,
3850 CompactibleFreeListSpace* cms_space,
3851 bool asynch,
3852 YieldingFlexibleWorkGang* workers,
3853 OopTaskQueueSet* task_queues):
3854 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3855 _collector(collector),
3856 _cms_space(cms_space),
3857 _asynch(asynch), _n_workers(0), _result(true),
3858 _task_queues(task_queues),
3859 _term(_n_workers, task_queues, _collector),
3860 _bit_map_lock(collector->bitMapLock())
3861 {
3862 _requested_size = _n_workers;
3863 _term.set_task(this);
3864 _term_term.set_task(this);
3865 _restart_addr = _global_finger = _cms_space->bottom();
3866 }
3867
3868
3869 OopTaskQueueSet* task_queues() { return _task_queues; }
3870
3871 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3872
3873 HeapWord** global_finger_addr() { return &_global_finger; }
3874
3875 CMSConcMarkingTerminator* terminator() { return &_term; }
3876
3877 virtual void set_for_termination(int active_workers) {
3878 terminator()->reset_for_reuse(active_workers);
3879 }
3880
3881 void work(uint worker_id);
3882 bool should_yield() {
3883 return ConcurrentMarkSweepThread::should_yield()
3884 && !_collector->foregroundGCIsActive()
3885 && _asynch;
3886 }
3887
3888 virtual void coordinator_yield(); // stuff done by coordinator
3889 bool result() { return _result; }
3890
3891 void reset(HeapWord* ra) {
3892 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3893 _restart_addr = _global_finger = ra;
3894 _term.reset_for_reuse();
3895 }
3896
3897 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3898 OopTaskQueue* work_q);
3899
3900 private:
3901 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3902 void do_work_steal(int i);
3903 void bump_global_finger(HeapWord* f);
3904 };
3905
3906 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3907 assert(_task != NULL, "Error");
3908 return _task->yielding();
3909 // Note that we do not need the disjunct || _task->should_yield() above
3910 // because we want terminating threads to yield only if the task
3911 // is already in the midst of yielding, which happens only after at least one
3912 // thread has yielded.
3913 }
3914
3915 void CMSConcMarkingTerminator::yield() {
3916 if (_task->should_yield()) {
3917 _task->yield();
3918 } else {
3919 ParallelTaskTerminator::yield();
3920 }
3921 }
3922
3923 ////////////////////////////////////////////////////////////////
3924 // Concurrent Marking Algorithm Sketch
3925 ////////////////////////////////////////////////////////////////
3926 // Until all tasks exhausted (both spaces):
3927 // -- claim next available chunk
3928 // -- bump global finger via CAS
3929 // -- find first object that starts in this chunk
3930 // and start scanning bitmap from that position
3931 // -- scan marked objects for oops
3932 // -- CAS-mark target, and if successful:
3933 // . if target oop is above global finger (volatile read)
3934 // nothing to do
3935 // . if target oop is in chunk and above local finger
3936 // then nothing to do
3937 // . else push on work-queue
3938 // -- Deal with possible overflow issues:
3939 // . local work-queue overflow causes stuff to be pushed on
3940 // global (common) overflow queue
3941 // . always first empty local work queue
3942 // . then get a batch of oops from global work queue if any
3943 // . then do work stealing
3944 // -- When all tasks claimed (both spaces)
3945 // and local work queue empty,
3946 // then in a loop do:
3947 // . check global overflow stack; steal a batch of oops and trace
3948 // . try to steal from other threads oif GOS is empty
3949 // . if neither is available, offer termination
3950 // -- Terminate and return result
3951 //
3952 void CMSConcMarkingTask::work(uint worker_id) {
3953 elapsedTimer _timer;
3954 ResourceMark rm;
3955 HandleMark hm;
3956
3957 DEBUG_ONLY(_collector->verify_overflow_empty();)
3958
3959 // Before we begin work, our work queue should be empty
3960 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3961 // Scan the bitmap covering _cms_space, tracing through grey objects.
3962 _timer.start();
3963 do_scan_and_mark(worker_id, _cms_space);
3964 _timer.stop();
3965 if (PrintCMSStatistics != 0) {
3966 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3967 worker_id, _timer.seconds());
3968 // XXX: need xxx/xxx type of notation, two timers
3969 }
3970
3971 // ... do work stealing
3972 _timer.reset();
3973 _timer.start();
3974 do_work_steal(worker_id);
3975 _timer.stop();
3976 if (PrintCMSStatistics != 0) {
3977 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3978 worker_id, _timer.seconds());
3979 // XXX: need xxx/xxx type of notation, two timers
3980 }
3981 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3982 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3983 // Note that under the current task protocol, the
3984 // following assertion is true even of the spaces
3985 // expanded since the completion of the concurrent
3986 // marking. XXX This will likely change under a strict
3987 // ABORT semantics.
3988 // After perm removal the comparison was changed to
3989 // greater than or equal to from strictly greater than.
3990 // Before perm removal the highest address sweep would
3991 // have been at the end of perm gen but now is at the
3992 // end of the tenured gen.
3993 assert(_global_finger >= _cms_space->end(),
3994 "All tasks have been completed");
3995 DEBUG_ONLY(_collector->verify_overflow_empty();)
3996 }
3997
3998 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3999 HeapWord* read = _global_finger;
4000 HeapWord* cur = read;
4001 while (f > read) {
4002 cur = read;
4003 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
4004 if (cur == read) {
4005 // our cas succeeded
4006 assert(_global_finger >= f, "protocol consistency");
4007 break;
4008 }
4009 }
4010 }
4011
4012 // This is really inefficient, and should be redone by
4013 // using (not yet available) block-read and -write interfaces to the
4014 // stack and the work_queue. XXX FIX ME !!!
4015 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
4016 OopTaskQueue* work_q) {
4017 // Fast lock-free check
4018 if (ovflw_stk->length() == 0) {
4019 return false;
4020 }
4021 assert(work_q->size() == 0, "Shouldn't steal");
4022 MutexLockerEx ml(ovflw_stk->par_lock(),
4023 Mutex::_no_safepoint_check_flag);
4024 // Grab up to 1/4 the size of the work queue
4025 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4026 (size_t)ParGCDesiredObjsFromOverflowList);
4027 num = MIN2(num, ovflw_stk->length());
4028 for (int i = (int) num; i > 0; i--) {
4029 oop cur = ovflw_stk->pop();
4030 assert(cur != NULL, "Counted wrong?");
4031 work_q->push(cur);
4032 }
4033 return num > 0;
4034 }
4035
4036 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
4037 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4038 int n_tasks = pst->n_tasks();
4039 // We allow that there may be no tasks to do here because
4040 // we are restarting after a stack overflow.
4041 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
4042 uint nth_task = 0;
4043
4044 HeapWord* aligned_start = sp->bottom();
4045 if (sp->used_region().contains(_restart_addr)) {
4046 // Align down to a card boundary for the start of 0th task
4047 // for this space.
4048 aligned_start =
4049 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
4050 CardTableModRefBS::card_size);
4051 }
4052
4053 size_t chunk_size = sp->marking_task_size();
4054 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4055 // Having claimed the nth task in this space,
4056 // compute the chunk that it corresponds to:
4057 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
4058 aligned_start + (nth_task+1)*chunk_size);
4059 // Try and bump the global finger via a CAS;
4060 // note that we need to do the global finger bump
4061 // _before_ taking the intersection below, because
4062 // the task corresponding to that region will be
4063 // deemed done even if the used_region() expands
4064 // because of allocation -- as it almost certainly will
4065 // during start-up while the threads yield in the
4066 // closure below.
4067 HeapWord* finger = span.end();
4068 bump_global_finger(finger); // atomically
4069 // There are null tasks here corresponding to chunks
4070 // beyond the "top" address of the space.
4071 span = span.intersection(sp->used_region());
4072 if (!span.is_empty()) { // Non-null task
4073 HeapWord* prev_obj;
4074 assert(!span.contains(_restart_addr) || nth_task == 0,
4075 "Inconsistency");
4076 if (nth_task == 0) {
4077 // For the 0th task, we'll not need to compute a block_start.
4078 if (span.contains(_restart_addr)) {
4079 // In the case of a restart because of stack overflow,
4080 // we might additionally skip a chunk prefix.
4081 prev_obj = _restart_addr;
4082 } else {
4083 prev_obj = span.start();
4084 }
4085 } else {
4086 // We want to skip the first object because
4087 // the protocol is to scan any object in its entirety
4088 // that _starts_ in this span; a fortiori, any
4089 // object starting in an earlier span is scanned
4090 // as part of an earlier claimed task.
4091 // Below we use the "careful" version of block_start
4092 // so we do not try to navigate uninitialized objects.
4093 prev_obj = sp->block_start_careful(span.start());
4094 // Below we use a variant of block_size that uses the
4095 // Printezis bits to avoid waiting for allocated
4096 // objects to become initialized/parsable.
4097 while (prev_obj < span.start()) {
4098 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
4099 if (sz > 0) {
4100 prev_obj += sz;
4101 } else {
4102 // In this case we may end up doing a bit of redundant
4103 // scanning, but that appears unavoidable, short of
4104 // locking the free list locks; see bug 6324141.
4105 break;
4106 }
4107 }
4108 }
4109 if (prev_obj < span.end()) {
4110 MemRegion my_span = MemRegion(prev_obj, span.end());
4111 // Do the marking work within a non-empty span --
4112 // the last argument to the constructor indicates whether the
4113 // iteration should be incremental with periodic yields.
4114 Par_MarkFromRootsClosure cl(this, _collector, my_span,
4115 &_collector->_markBitMap,
4116 work_queue(i),
4117 &_collector->_markStack,
4118 _asynch);
4119 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
4120 } // else nothing to do for this task
4121 } // else nothing to do for this task
4122 }
4123 // We'd be tempted to assert here that since there are no
4124 // more tasks left to claim in this space, the global_finger
4125 // must exceed space->top() and a fortiori space->end(). However,
4126 // that would not quite be correct because the bumping of
4127 // global_finger occurs strictly after the claiming of a task,
4128 // so by the time we reach here the global finger may not yet
4129 // have been bumped up by the thread that claimed the last
4130 // task.
4131 pst->all_tasks_completed();
4132 }
4133
4134 class Par_ConcMarkingClosure: public CMSOopClosure {
4135 private:
4136 CMSCollector* _collector;
4137 CMSConcMarkingTask* _task;
4138 MemRegion _span;
4139 CMSBitMap* _bit_map;
4140 CMSMarkStack* _overflow_stack;
4141 OopTaskQueue* _work_queue;
4142 protected:
4143 DO_OOP_WORK_DEFN
4144 public:
4145 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
4146 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
4147 CMSOopClosure(collector->ref_processor()),
4148 _collector(collector),
4149 _task(task),
4150 _span(collector->_span),
4151 _work_queue(work_queue),
4152 _bit_map(bit_map),
4153 _overflow_stack(overflow_stack)
4154 { }
4155 virtual void do_oop(oop* p);
4156 virtual void do_oop(narrowOop* p);
4157
4158 void trim_queue(size_t max);
4159 void handle_stack_overflow(HeapWord* lost);
4160 void do_yield_check() {
4161 if (_task->should_yield()) {
4162 _task->yield();
4163 }
4164 }
4165 };
4166
4167 // Grey object scanning during work stealing phase --
4168 // the salient assumption here is that any references
4169 // that are in these stolen objects being scanned must
4170 // already have been initialized (else they would not have
4171 // been published), so we do not need to check for
4172 // uninitialized objects before pushing here.
4173 void Par_ConcMarkingClosure::do_oop(oop obj) {
4174 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
4175 HeapWord* addr = (HeapWord*)obj;
4176 // Check if oop points into the CMS generation
4177 // and is not marked
4178 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
4179 // a white object ...
4180 // If we manage to "claim" the object, by being the
4181 // first thread to mark it, then we push it on our
4182 // marking stack
4183 if (_bit_map->par_mark(addr)) { // ... now grey
4184 // push on work queue (grey set)
4185 bool simulate_overflow = false;
4186 NOT_PRODUCT(
4187 if (CMSMarkStackOverflowALot &&
4188 _collector->simulate_overflow()) {
4189 // simulate a stack overflow
4190 simulate_overflow = true;
4191 }
4192 )
4193 if (simulate_overflow ||
4194 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
4195 // stack overflow
4196 if (PrintCMSStatistics != 0) {
4197 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
4198 SIZE_FORMAT, _overflow_stack->capacity());
4199 }
4200 // We cannot assert that the overflow stack is full because
4201 // it may have been emptied since.
4202 assert(simulate_overflow ||
4203 _work_queue->size() == _work_queue->max_elems(),
4204 "Else push should have succeeded");
4205 handle_stack_overflow(addr);
4206 }
4207 } // Else, some other thread got there first
4208 do_yield_check();
4209 }
4210 }
4211
4212 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4213 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4214
4215 void Par_ConcMarkingClosure::trim_queue(size_t max) {
4216 while (_work_queue->size() > max) {
4217 oop new_oop;
4218 if (_work_queue->pop_local(new_oop)) {
4219 assert(new_oop->is_oop(), "Should be an oop");
4220 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
4221 assert(_span.contains((HeapWord*)new_oop), "Not in span");
4222 new_oop->oop_iterate(this); // do_oop() above
4223 do_yield_check();
4224 }
4225 }
4226 }
4227
4228 // Upon stack overflow, we discard (part of) the stack,
4229 // remembering the least address amongst those discarded
4230 // in CMSCollector's _restart_address.
4231 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
4232 // We need to do this under a mutex to prevent other
4233 // workers from interfering with the work done below.
4234 MutexLockerEx ml(_overflow_stack->par_lock(),
4235 Mutex::_no_safepoint_check_flag);
4236 // Remember the least grey address discarded
4237 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
4238 _collector->lower_restart_addr(ra);
4239 _overflow_stack->reset(); // discard stack contents
4240 _overflow_stack->expand(); // expand the stack if possible
4241 }
4242
4243
4244 void CMSConcMarkingTask::do_work_steal(int i) {
4245 OopTaskQueue* work_q = work_queue(i);
4246 oop obj_to_scan;
4247 CMSBitMap* bm = &(_collector->_markBitMap);
4248 CMSMarkStack* ovflw = &(_collector->_markStack);
4249 int* seed = _collector->hash_seed(i);
4250 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
4251 while (true) {
4252 cl.trim_queue(0);
4253 assert(work_q->size() == 0, "Should have been emptied above");
4254 if (get_work_from_overflow_stack(ovflw, work_q)) {
4255 // Can't assert below because the work obtained from the
4256 // overflow stack may already have been stolen from us.
4257 // assert(work_q->size() > 0, "Work from overflow stack");
4258 continue;
4259 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4260 assert(obj_to_scan->is_oop(), "Should be an oop");
4261 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
4262 obj_to_scan->oop_iterate(&cl);
4263 } else if (terminator()->offer_termination(&_term_term)) {
4264 assert(work_q->size() == 0, "Impossible!");
4265 break;
4266 } else if (yielding() || should_yield()) {
4267 yield();
4268 }
4269 }
4270 }
4271
4272 // This is run by the CMS (coordinator) thread.
4273 void CMSConcMarkingTask::coordinator_yield() {
4274 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4275 "CMS thread should hold CMS token");
4276 // First give up the locks, then yield, then re-lock
4277 // We should probably use a constructor/destructor idiom to
4278 // do this unlock/lock or modify the MutexUnlocker class to
4279 // serve our purpose. XXX
4280 assert_lock_strong(_bit_map_lock);
4281 _bit_map_lock->unlock();
4282 ConcurrentMarkSweepThread::desynchronize(true);
4283 ConcurrentMarkSweepThread::acknowledge_yield_request();
4284 _collector->stopTimer();
4285 if (PrintCMSStatistics != 0) {
4286 _collector->incrementYields();
4287 }
4288 _collector->icms_wait();
4289
4290 // It is possible for whichever thread initiated the yield request
4291 // not to get a chance to wake up and take the bitmap lock between
4292 // this thread releasing it and reacquiring it. So, while the
4293 // should_yield() flag is on, let's sleep for a bit to give the
4294 // other thread a chance to wake up. The limit imposed on the number
4295 // of iterations is defensive, to avoid any unforseen circumstances
4296 // putting us into an infinite loop. Since it's always been this
4297 // (coordinator_yield()) method that was observed to cause the
4298 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4299 // which is by default non-zero. For the other seven methods that
4300 // also perform the yield operation, as are using a different
4301 // parameter (CMSYieldSleepCount) which is by default zero. This way we
4302 // can enable the sleeping for those methods too, if necessary.
4303 // See 6442774.
4304 //
4305 // We really need to reconsider the synchronization between the GC
4306 // thread and the yield-requesting threads in the future and we
4307 // should really use wait/notify, which is the recommended
4308 // way of doing this type of interaction. Additionally, we should
4309 // consolidate the eight methods that do the yield operation and they
4310 // are almost identical into one for better maintenability and
4311 // readability. See 6445193.
4312 //
4313 // Tony 2006.06.29
4314 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4315 ConcurrentMarkSweepThread::should_yield() &&
4316 !CMSCollector::foregroundGCIsActive(); ++i) {
4317 os::sleep(Thread::current(), 1, false);
4318 ConcurrentMarkSweepThread::acknowledge_yield_request();
4319 }
4320
4321 ConcurrentMarkSweepThread::synchronize(true);
4322 _bit_map_lock->lock_without_safepoint_check();
4323 _collector->startTimer();
4324 }
4325
4326 bool CMSCollector::do_marking_mt(bool asynch) {
4327 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
4328 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers(
4329 conc_workers()->total_workers(),
4330 conc_workers()->active_workers(),
4331 Threads::number_of_non_daemon_threads());
4332 conc_workers()->set_active_workers(num_workers);
4333
4334 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4335
4336 CMSConcMarkingTask tsk(this,
4337 cms_space,
4338 asynch,
4339 conc_workers(),
4340 task_queues());
4341
4342 // Since the actual number of workers we get may be different
4343 // from the number we requested above, do we need to do anything different
4344 // below? In particular, may be we need to subclass the SequantialSubTasksDone
4345 // class?? XXX
4346 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4347
4348 // Refs discovery is already non-atomic.
4349 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4350 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
4351 conc_workers()->start_task(&tsk);
4352 while (tsk.yielded()) {
4353 tsk.coordinator_yield();
4354 conc_workers()->continue_task(&tsk);
4355 }
4356 // If the task was aborted, _restart_addr will be non-NULL
4357 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4358 while (_restart_addr != NULL) {
4359 // XXX For now we do not make use of ABORTED state and have not
4360 // yet implemented the right abort semantics (even in the original
4361 // single-threaded CMS case). That needs some more investigation
4362 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4363 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4364 // If _restart_addr is non-NULL, a marking stack overflow
4365 // occurred; we need to do a fresh marking iteration from the
4366 // indicated restart address.
4367 if (_foregroundGCIsActive && asynch) {
4368 // We may be running into repeated stack overflows, having
4369 // reached the limit of the stack size, while making very
4370 // slow forward progress. It may be best to bail out and
4371 // let the foreground collector do its job.
4372 // Clear _restart_addr, so that foreground GC
4373 // works from scratch. This avoids the headache of
4374 // a "rescan" which would otherwise be needed because
4375 // of the dirty mod union table & card table.
4376 _restart_addr = NULL;
4377 return false;
4378 }
4379 // Adjust the task to restart from _restart_addr
4380 tsk.reset(_restart_addr);
4381 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4382 _restart_addr);
4383 _restart_addr = NULL;
4384 // Get the workers going again
4385 conc_workers()->start_task(&tsk);
4386 while (tsk.yielded()) {
4387 tsk.coordinator_yield();
4388 conc_workers()->continue_task(&tsk);
4389 }
4390 }
4391 assert(tsk.completed(), "Inconsistency");
4392 assert(tsk.result() == true, "Inconsistency");
4393 return true;
4394 }
4395
4396 bool CMSCollector::do_marking_st(bool asynch) {
4397 ResourceMark rm;
4398 HandleMark hm;
4399
4400 // Temporarily make refs discovery single threaded (non-MT)
4401 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
4402 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4403 &_markStack, CMSYield && asynch);
4404 // the last argument to iterate indicates whether the iteration
4405 // should be incremental with periodic yields.
4406 _markBitMap.iterate(&markFromRootsClosure);
4407 // If _restart_addr is non-NULL, a marking stack overflow
4408 // occurred; we need to do a fresh iteration from the
4409 // indicated restart address.
4410 while (_restart_addr != NULL) {
4411 if (_foregroundGCIsActive && asynch) {
4412 // We may be running into repeated stack overflows, having
4413 // reached the limit of the stack size, while making very
4414 // slow forward progress. It may be best to bail out and
4415 // let the foreground collector do its job.
4416 // Clear _restart_addr, so that foreground GC
4417 // works from scratch. This avoids the headache of
4418 // a "rescan" which would otherwise be needed because
4419 // of the dirty mod union table & card table.
4420 _restart_addr = NULL;
4421 return false; // indicating failure to complete marking
4422 }
4423 // Deal with stack overflow:
4424 // we restart marking from _restart_addr
4425 HeapWord* ra = _restart_addr;
4426 markFromRootsClosure.reset(ra);
4427 _restart_addr = NULL;
4428 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4429 }
4430 return true;
4431 }
4432
4433 void CMSCollector::preclean() {
4434 check_correct_thread_executing();
4435 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4436 verify_work_stacks_empty();
4437 verify_overflow_empty();
4438 _abort_preclean = false;
4439 if (CMSPrecleaningEnabled) {
4440 _eden_chunk_index = 0;
4441 size_t used = get_eden_used();
4442 size_t capacity = get_eden_capacity();
4443 // Don't start sampling unless we will get sufficiently
4444 // many samples.
4445 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4446 * CMSScheduleRemarkEdenPenetration)) {
4447 _start_sampling = true;
4448 } else {
4449 _start_sampling = false;
4450 }
4451 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4452 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
4453 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4454 }
4455 CMSTokenSync x(true); // is cms thread
4456 if (CMSPrecleaningEnabled) {
4457 sample_eden();
4458 _collectorState = AbortablePreclean;
4459 } else {
4460 _collectorState = FinalMarking;
4461 }
4462 verify_work_stacks_empty();
4463 verify_overflow_empty();
4464 }
4465
4466 // Try and schedule the remark such that young gen
4467 // occupancy is CMSScheduleRemarkEdenPenetration %.
4468 void CMSCollector::abortable_preclean() {
4469 check_correct_thread_executing();
4470 assert(CMSPrecleaningEnabled, "Inconsistent control state");
4471 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4472
4473 // If Eden's current occupancy is below this threshold,
4474 // immediately schedule the remark; else preclean
4475 // past the next scavenge in an effort to
4476 // schedule the pause as described avove. By choosing
4477 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4478 // we will never do an actual abortable preclean cycle.
4479 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4480 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4481 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
4482 // We need more smarts in the abortable preclean
4483 // loop below to deal with cases where allocation
4484 // in young gen is very very slow, and our precleaning
4485 // is running a losing race against a horde of
4486 // mutators intent on flooding us with CMS updates
4487 // (dirty cards).
4488 // One, admittedly dumb, strategy is to give up
4489 // after a certain number of abortable precleaning loops
4490 // or after a certain maximum time. We want to make
4491 // this smarter in the next iteration.
4492 // XXX FIX ME!!! YSR
4493 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4494 while (!(should_abort_preclean() ||
4495 ConcurrentMarkSweepThread::should_terminate())) {
4496 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4497 cumworkdone += workdone;
4498 loops++;
4499 // Voluntarily terminate abortable preclean phase if we have
4500 // been at it for too long.
4501 if ((CMSMaxAbortablePrecleanLoops != 0) &&
4502 loops >= CMSMaxAbortablePrecleanLoops) {
4503 if (PrintGCDetails) {
4504 gclog_or_tty->print(" CMS: abort preclean due to loops ");
4505 }
4506 break;
4507 }
4508 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4509 if (PrintGCDetails) {
4510 gclog_or_tty->print(" CMS: abort preclean due to time ");
4511 }
4512 break;
4513 }
4514 // If we are doing little work each iteration, we should
4515 // take a short break.
4516 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4517 // Sleep for some time, waiting for work to accumulate
4518 stopTimer();
4519 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4520 startTimer();
4521 waited++;
4522 }
4523 }
4524 if (PrintCMSStatistics > 0) {
4525 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4526 loops, waited, cumworkdone);
4527 }
4528 }
4529 CMSTokenSync x(true); // is cms thread
4530 if (_collectorState != Idling) {
4531 assert(_collectorState == AbortablePreclean,
4532 "Spontaneous state transition?");
4533 _collectorState = FinalMarking;
4534 } // Else, a foreground collection completed this CMS cycle.
4535 return;
4536 }
4537
4538 // Respond to an Eden sampling opportunity
4539 void CMSCollector::sample_eden() {
4540 // Make sure a young gc cannot sneak in between our
4541 // reading and recording of a sample.
4542 assert(Thread::current()->is_ConcurrentGC_thread(),
4543 "Only the cms thread may collect Eden samples");
4544 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4545 "Should collect samples while holding CMS token");
4546 if (!_start_sampling) {
4547 return;
4548 }
4549 if (_eden_chunk_array) {
4550 if (_eden_chunk_index < _eden_chunk_capacity) {
4551 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
4552 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4553 "Unexpected state of Eden");
4554 // We'd like to check that what we just sampled is an oop-start address;
4555 // however, we cannot do that here since the object may not yet have been
4556 // initialized. So we'll instead do the check when we _use_ this sample
4557 // later.
4558 if (_eden_chunk_index == 0 ||
4559 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4560 _eden_chunk_array[_eden_chunk_index-1])
4561 >= CMSSamplingGrain)) {
4562 _eden_chunk_index++; // commit sample
4563 }
4564 }
4565 }
4566 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4567 size_t used = get_eden_used();
4568 size_t capacity = get_eden_capacity();
4569 assert(used <= capacity, "Unexpected state of Eden");
4570 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4571 _abort_preclean = true;
4572 }
4573 }
4574 }
4575
4576
4577 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4578 assert(_collectorState == Precleaning ||
4579 _collectorState == AbortablePreclean, "incorrect state");
4580 ResourceMark rm;
4581 HandleMark hm;
4582
4583 // Precleaning is currently not MT but the reference processor
4584 // may be set for MT. Disable it temporarily here.
4585 ReferenceProcessor* rp = ref_processor();
4586 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
4587
4588 // Do one pass of scrubbing the discovered reference lists
4589 // to remove any reference objects with strongly-reachable
4590 // referents.
4591 if (clean_refs) {
4592 CMSPrecleanRefsYieldClosure yield_cl(this);
4593 assert(rp->span().equals(_span), "Spans should be equal");
4594 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4595 &_markStack, true /* preclean */);
4596 CMSDrainMarkingStackClosure complete_trace(this,
4597 _span, &_markBitMap, &_markStack,
4598 &keep_alive, true /* preclean */);
4599
4600 // We don't want this step to interfere with a young
4601 // collection because we don't want to take CPU
4602 // or memory bandwidth away from the young GC threads
4603 // (which may be as many as there are CPUs).
4604 // Note that we don't need to protect ourselves from
4605 // interference with mutators because they can't
4606 // manipulate the discovered reference lists nor affect
4607 // the computed reachability of the referents, the
4608 // only properties manipulated by the precleaning
4609 // of these reference lists.
4610 stopTimer();
4611 CMSTokenSyncWithLocks x(true /* is cms thread */,
4612 bitMapLock());
4613 startTimer();
4614 sample_eden();
4615
4616 // The following will yield to allow foreground
4617 // collection to proceed promptly. XXX YSR:
4618 // The code in this method may need further
4619 // tweaking for better performance and some restructuring
4620 // for cleaner interfaces.
4621 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
4622 rp->preclean_discovered_references(
4623 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
4624 gc_timer);
4625 }
4626
4627 if (clean_survivor) { // preclean the active survivor space(s)
4628 assert(_young_gen->kind() == Generation::DefNew ||
4629 _young_gen->kind() == Generation::ParNew ||
4630 _young_gen->kind() == Generation::ASParNew,
4631 "incorrect type for cast");
4632 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4633 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4634 &_markBitMap, &_modUnionTable,
4635 &_markStack, true /* precleaning phase */);
4636 stopTimer();
4637 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4638 bitMapLock());
4639 startTimer();
4640 unsigned int before_count =
4641 GenCollectedHeap::heap()->total_collections();
4642 SurvivorSpacePrecleanClosure
4643 sss_cl(this, _span, &_markBitMap, &_markStack,
4644 &pam_cl, before_count, CMSYield);
4645 dng->from()->object_iterate_careful(&sss_cl);
4646 dng->to()->object_iterate_careful(&sss_cl);
4647 }
4648 MarkRefsIntoAndScanClosure
4649 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4650 &_markStack, this, CMSYield,
4651 true /* precleaning phase */);
4652 // CAUTION: The following closure has persistent state that may need to
4653 // be reset upon a decrease in the sequence of addresses it
4654 // processes.
4655 ScanMarkedObjectsAgainCarefullyClosure
4656 smoac_cl(this, _span,
4657 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
4658
4659 // Preclean dirty cards in ModUnionTable and CardTable using
4660 // appropriate convergence criterion;
4661 // repeat CMSPrecleanIter times unless we find that
4662 // we are losing.
4663 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4664 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4665 "Bad convergence multiplier");
4666 assert(CMSPrecleanThreshold >= 100,
4667 "Unreasonably low CMSPrecleanThreshold");
4668
4669 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4670 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4671 numIter < CMSPrecleanIter;
4672 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4673 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4674 if (Verbose && PrintGCDetails) {
4675 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4676 }
4677 // Either there are very few dirty cards, so re-mark
4678 // pause will be small anyway, or our pre-cleaning isn't
4679 // that much faster than the rate at which cards are being
4680 // dirtied, so we might as well stop and re-mark since
4681 // precleaning won't improve our re-mark time by much.
4682 if (curNumCards <= CMSPrecleanThreshold ||
4683 (numIter > 0 &&
4684 (curNumCards * CMSPrecleanDenominator >
4685 lastNumCards * CMSPrecleanNumerator))) {
4686 numIter++;
4687 cumNumCards += curNumCards;
4688 break;
4689 }
4690 }
4691
4692 preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
4693
4694 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4695 cumNumCards += curNumCards;
4696 if (PrintGCDetails && PrintCMSStatistics != 0) {
4697 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4698 curNumCards, cumNumCards, numIter);
4699 }
4700 return cumNumCards; // as a measure of useful work done
4701 }
4702
4703 // PRECLEANING NOTES:
4704 // Precleaning involves:
4705 // . reading the bits of the modUnionTable and clearing the set bits.
4706 // . For the cards corresponding to the set bits, we scan the
4707 // objects on those cards. This means we need the free_list_lock
4708 // so that we can safely iterate over the CMS space when scanning
4709 // for oops.
4710 // . When we scan the objects, we'll be both reading and setting
4711 // marks in the marking bit map, so we'll need the marking bit map.
4712 // . For protecting _collector_state transitions, we take the CGC_lock.
4713 // Note that any races in the reading of of card table entries by the
4714 // CMS thread on the one hand and the clearing of those entries by the
4715 // VM thread or the setting of those entries by the mutator threads on the
4716 // other are quite benign. However, for efficiency it makes sense to keep
4717 // the VM thread from racing with the CMS thread while the latter is
4718 // dirty card info to the modUnionTable. We therefore also use the
4719 // CGC_lock to protect the reading of the card table and the mod union
4720 // table by the CM thread.
4721 // . We run concurrently with mutator updates, so scanning
4722 // needs to be done carefully -- we should not try to scan
4723 // potentially uninitialized objects.
4724 //
4725 // Locking strategy: While holding the CGC_lock, we scan over and
4726 // reset a maximal dirty range of the mod union / card tables, then lock
4727 // the free_list_lock and bitmap lock to do a full marking, then
4728 // release these locks; and repeat the cycle. This allows for a
4729 // certain amount of fairness in the sharing of these locks between
4730 // the CMS collector on the one hand, and the VM thread and the
4731 // mutators on the other.
4732
4733 // NOTE: preclean_mod_union_table() and preclean_card_table()
4734 // further below are largely identical; if you need to modify
4735 // one of these methods, please check the other method too.
4736
4737 size_t CMSCollector::preclean_mod_union_table(
4738 ConcurrentMarkSweepGeneration* gen,
4739 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4740 verify_work_stacks_empty();
4741 verify_overflow_empty();
4742
4743 // strategy: starting with the first card, accumulate contiguous
4744 // ranges of dirty cards; clear these cards, then scan the region
4745 // covered by these cards.
4746
4747 // Since all of the MUT is committed ahead, we can just use
4748 // that, in case the generations expand while we are precleaning.
4749 // It might also be fine to just use the committed part of the
4750 // generation, but we might potentially miss cards when the
4751 // generation is rapidly expanding while we are in the midst
4752 // of precleaning.
4753 HeapWord* startAddr = gen->reserved().start();
4754 HeapWord* endAddr = gen->reserved().end();
4755
4756 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4757
4758 size_t numDirtyCards, cumNumDirtyCards;
4759 HeapWord *nextAddr, *lastAddr;
4760 for (cumNumDirtyCards = numDirtyCards = 0,
4761 nextAddr = lastAddr = startAddr;
4762 nextAddr < endAddr;
4763 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4764
4765 ResourceMark rm;
4766 HandleMark hm;
4767
4768 MemRegion dirtyRegion;
4769 {
4770 stopTimer();
4771 // Potential yield point
4772 CMSTokenSync ts(true);
4773 startTimer();
4774 sample_eden();
4775 // Get dirty region starting at nextOffset (inclusive),
4776 // simultaneously clearing it.
4777 dirtyRegion =
4778 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4779 assert(dirtyRegion.start() >= nextAddr,
4780 "returned region inconsistent?");
4781 }
4782 // Remember where the next search should begin.
4783 // The returned region (if non-empty) is a right open interval,
4784 // so lastOffset is obtained from the right end of that
4785 // interval.
4786 lastAddr = dirtyRegion.end();
4787 // Should do something more transparent and less hacky XXX
4788 numDirtyCards =
4789 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4790
4791 // We'll scan the cards in the dirty region (with periodic
4792 // yields for foreground GC as needed).
4793 if (!dirtyRegion.is_empty()) {
4794 assert(numDirtyCards > 0, "consistency check");
4795 HeapWord* stop_point = NULL;
4796 stopTimer();
4797 // Potential yield point
4798 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4799 bitMapLock());
4800 startTimer();
4801 {
4802 verify_work_stacks_empty();
4803 verify_overflow_empty();
4804 sample_eden();
4805 stop_point =
4806 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4807 }
4808 if (stop_point != NULL) {
4809 // The careful iteration stopped early either because it found an
4810 // uninitialized object, or because we were in the midst of an
4811 // "abortable preclean", which should now be aborted. Redirty
4812 // the bits corresponding to the partially-scanned or unscanned
4813 // cards. We'll either restart at the next block boundary or
4814 // abort the preclean.
4815 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4816 "Should only be AbortablePreclean.");
4817 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4818 if (should_abort_preclean()) {
4819 break; // out of preclean loop
4820 } else {
4821 // Compute the next address at which preclean should pick up;
4822 // might need bitMapLock in order to read P-bits.
4823 lastAddr = next_card_start_after_block(stop_point);
4824 }
4825 }
4826 } else {
4827 assert(lastAddr == endAddr, "consistency check");
4828 assert(numDirtyCards == 0, "consistency check");
4829 break;
4830 }
4831 }
4832 verify_work_stacks_empty();
4833 verify_overflow_empty();
4834 return cumNumDirtyCards;
4835 }
4836
4837 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4838 // below are largely identical; if you need to modify
4839 // one of these methods, please check the other method too.
4840
4841 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4842 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4843 // strategy: it's similar to precleamModUnionTable above, in that
4844 // we accumulate contiguous ranges of dirty cards, mark these cards
4845 // precleaned, then scan the region covered by these cards.
4846 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4847 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4848
4849 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4850
4851 size_t numDirtyCards, cumNumDirtyCards;
4852 HeapWord *lastAddr, *nextAddr;
4853
4854 for (cumNumDirtyCards = numDirtyCards = 0,
4855 nextAddr = lastAddr = startAddr;
4856 nextAddr < endAddr;
4857 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4858
4859 ResourceMark rm;
4860 HandleMark hm;
4861
4862 MemRegion dirtyRegion;
4863 {
4864 // See comments in "Precleaning notes" above on why we
4865 // do this locking. XXX Could the locking overheads be
4866 // too high when dirty cards are sparse? [I don't think so.]
4867 stopTimer();
4868 CMSTokenSync x(true); // is cms thread
4869 startTimer();
4870 sample_eden();
4871 // Get and clear dirty region from card table
4872 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4873 MemRegion(nextAddr, endAddr),
4874 true,
4875 CardTableModRefBS::precleaned_card_val());
4876
4877 assert(dirtyRegion.start() >= nextAddr,
4878 "returned region inconsistent?");
4879 }
4880 lastAddr = dirtyRegion.end();
4881 numDirtyCards =
4882 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4883
4884 if (!dirtyRegion.is_empty()) {
4885 stopTimer();
4886 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4887 startTimer();
4888 sample_eden();
4889 verify_work_stacks_empty();
4890 verify_overflow_empty();
4891 HeapWord* stop_point =
4892 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4893 if (stop_point != NULL) {
4894 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4895 "Should only be AbortablePreclean.");
4896 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4897 if (should_abort_preclean()) {
4898 break; // out of preclean loop
4899 } else {
4900 // Compute the next address at which preclean should pick up.
4901 lastAddr = next_card_start_after_block(stop_point);
4902 }
4903 }
4904 } else {
4905 break;
4906 }
4907 }
4908 verify_work_stacks_empty();
4909 verify_overflow_empty();
4910 return cumNumDirtyCards;
4911 }
4912
4913 class PrecleanKlassClosure : public KlassClosure {
4914 CMKlassClosure _cm_klass_closure;
4915 public:
4916 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4917 void do_klass(Klass* k) {
4918 if (k->has_accumulated_modified_oops()) {
4919 k->clear_accumulated_modified_oops();
4920
4921 _cm_klass_closure.do_klass(k);
4922 }
4923 }
4924 };
4925
4926 // The freelist lock is needed to prevent asserts, is it really needed?
4927 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4928
4929 cl->set_freelistLock(freelistLock);
4930
4931 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4932
4933 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4934 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4935 PrecleanKlassClosure preclean_klass_closure(cl);
4936 ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
4937
4938 verify_work_stacks_empty();
4939 verify_overflow_empty();
4940 }
4941
4942 void CMSCollector::checkpointRootsFinal(bool asynch,
4943 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4944 assert(_collectorState == FinalMarking, "incorrect state transition?");
4945 check_correct_thread_executing();
4946 // world is stopped at this checkpoint
4947 assert(SafepointSynchronize::is_at_safepoint(),
4948 "world should be stopped");
4949 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4950
4951 verify_work_stacks_empty();
4952 verify_overflow_empty();
4953
4954 SpecializationStats::clear();
4955 if (PrintGCDetails) {
4956 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4957 _young_gen->used() / K,
4958 _young_gen->capacity() / K);
4959 }
4960 if (asynch) {
4961 if (CMSScavengeBeforeRemark) {
4962 GenCollectedHeap* gch = GenCollectedHeap::heap();
4963 // Temporarily set flag to false, GCH->do_collection will
4964 // expect it to be false and set to true
4965 FlagSetting fl(gch->_is_gc_active, false);
4966 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
4967 PrintGCDetails && Verbose, true, _gc_timer_cm);)
4968 int level = _cmsGen->level() - 1;
4969 if (level >= 0) {
4970 gch->do_collection(true, // full (i.e. force, see below)
4971 false, // !clear_all_soft_refs
4972 0, // size
4973 false, // is_tlab
4974 level // max_level
4975 );
4976 }
4977 }
4978 FreelistLocker x(this);
4979 MutexLockerEx y(bitMapLock(),
4980 Mutex::_no_safepoint_check_flag);
4981 assert(!init_mark_was_synchronous, "but that's impossible!");
4982 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
4983 } else {
4984 // already have all the locks
4985 checkpointRootsFinalWork(asynch, clear_all_soft_refs,
4986 init_mark_was_synchronous);
4987 }
4988 verify_work_stacks_empty();
4989 verify_overflow_empty();
4990 SpecializationStats::print();
4991 }
4992
4993 void CMSCollector::checkpointRootsFinalWork(bool asynch,
4994 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4995
4996 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);)
4997
4998 assert(haveFreelistLocks(), "must have free list locks");
4999 assert_lock_strong(bitMapLock());
5000
5001 if (UseAdaptiveSizePolicy) {
5002 size_policy()->checkpoint_roots_final_begin();
5003 }
5004
5005 ResourceMark rm;
5006 HandleMark hm;
5007
5008 GenCollectedHeap* gch = GenCollectedHeap::heap();
5009
5010 if (should_unload_classes()) {
5011 CodeCache::gc_prologue();
5012 }
5013 assert(haveFreelistLocks(), "must have free list locks");
5014 assert_lock_strong(bitMapLock());
5015
5016 if (!init_mark_was_synchronous) {
5017 // We might assume that we need not fill TLAB's when
5018 // CMSScavengeBeforeRemark is set, because we may have just done
5019 // a scavenge which would have filled all TLAB's -- and besides
5020 // Eden would be empty. This however may not always be the case --
5021 // for instance although we asked for a scavenge, it may not have
5022 // happened because of a JNI critical section. We probably need
5023 // a policy for deciding whether we can in that case wait until
5024 // the critical section releases and then do the remark following
5025 // the scavenge, and skip it here. In the absence of that policy,
5026 // or of an indication of whether the scavenge did indeed occur,
5027 // we cannot rely on TLAB's having been filled and must do
5028 // so here just in case a scavenge did not happen.
5029 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
5030 // Update the saved marks which may affect the root scans.
5031 gch->save_marks();
5032
5033 {
5034 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
5035
5036 // Note on the role of the mod union table:
5037 // Since the marker in "markFromRoots" marks concurrently with
5038 // mutators, it is possible for some reachable objects not to have been
5039 // scanned. For instance, an only reference to an object A was
5040 // placed in object B after the marker scanned B. Unless B is rescanned,
5041 // A would be collected. Such updates to references in marked objects
5042 // are detected via the mod union table which is the set of all cards
5043 // dirtied since the first checkpoint in this GC cycle and prior to
5044 // the most recent young generation GC, minus those cleaned up by the
5045 // concurrent precleaning.
5046 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
5047 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm);
5048 do_remark_parallel();
5049 } else {
5050 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
5051 _gc_timer_cm);
5052 do_remark_non_parallel();
5053 }
5054 }
5055 } else {
5056 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
5057 // The initial mark was stop-world, so there's no rescanning to
5058 // do; go straight on to the next step below.
5059 }
5060 verify_work_stacks_empty();
5061 verify_overflow_empty();
5062
5063 {
5064 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);)
5065 refProcessingWork(asynch, clear_all_soft_refs);
5066 }
5067 verify_work_stacks_empty();
5068 verify_overflow_empty();
5069
5070 if (should_unload_classes()) {
5071 CodeCache::gc_epilogue();
5072 }
5073 JvmtiExport::gc_epilogue();
5074
5075 // If we encountered any (marking stack / work queue) overflow
5076 // events during the current CMS cycle, take appropriate
5077 // remedial measures, where possible, so as to try and avoid
5078 // recurrence of that condition.
5079 assert(_markStack.isEmpty(), "No grey objects");
5080 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
5081 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
5082 if (ser_ovflw > 0) {
5083 if (PrintCMSStatistics != 0) {
5084 gclog_or_tty->print_cr("Marking stack overflow (benign) "
5085 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
5086 ", kac_preclean="SIZE_FORMAT")",
5087 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
5088 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
5089 }
5090 _markStack.expand();
5091 _ser_pmc_remark_ovflw = 0;
5092 _ser_pmc_preclean_ovflw = 0;
5093 _ser_kac_preclean_ovflw = 0;
5094 _ser_kac_ovflw = 0;
5095 }
5096 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
5097 if (PrintCMSStatistics != 0) {
5098 gclog_or_tty->print_cr("Work queue overflow (benign) "
5099 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
5100 _par_pmc_remark_ovflw, _par_kac_ovflw);
5101 }
5102 _par_pmc_remark_ovflw = 0;
5103 _par_kac_ovflw = 0;
5104 }
5105 if (PrintCMSStatistics != 0) {
5106 if (_markStack._hit_limit > 0) {
5107 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
5108 _markStack._hit_limit);
5109 }
5110 if (_markStack._failed_double > 0) {
5111 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
5112 " current capacity "SIZE_FORMAT,
5113 _markStack._failed_double,
5114 _markStack.capacity());
5115 }
5116 }
5117 _markStack._hit_limit = 0;
5118 _markStack._failed_double = 0;
5119
5120 if ((VerifyAfterGC || VerifyDuringGC) &&
5121 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5122 verify_after_remark();
5123 }
5124
5125 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
5126
5127 // Change under the freelistLocks.
5128 _collectorState = Sweeping;
5129 // Call isAllClear() under bitMapLock
5130 assert(_modUnionTable.isAllClear(),
5131 "Should be clear by end of the final marking");
5132 assert(_ct->klass_rem_set()->mod_union_is_clear(),
5133 "Should be clear by end of the final marking");
5134 if (UseAdaptiveSizePolicy) {
5135 size_policy()->checkpoint_roots_final_end(gch->gc_cause());
5136 }
5137 }
5138
5139 // Parallel remark task
5140 class CMSParRemarkTask: public AbstractGangTask {
5141 CMSCollector* _collector;
5142 int _n_workers;
5143 CompactibleFreeListSpace* _cms_space;
5144
5145 // The per-thread work queues, available here for stealing.
5146 OopTaskQueueSet* _task_queues;
5147 ParallelTaskTerminator _term;
5148
5149 public:
5150 // A value of 0 passed to n_workers will cause the number of
5151 // workers to be taken from the active workers in the work gang.
5152 CMSParRemarkTask(CMSCollector* collector,
5153 CompactibleFreeListSpace* cms_space,
5154 int n_workers, FlexibleWorkGang* workers,
5155 OopTaskQueueSet* task_queues):
5156 AbstractGangTask("Rescan roots and grey objects in parallel"),
5157 _collector(collector),
5158 _cms_space(cms_space),
5159 _n_workers(n_workers),
5160 _task_queues(task_queues),
5161 _term(n_workers, task_queues) { }
5162
5163 OopTaskQueueSet* task_queues() { return _task_queues; }
5164
5165 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5166
5167 ParallelTaskTerminator* terminator() { return &_term; }
5168 int n_workers() { return _n_workers; }
5169
5170 void work(uint worker_id);
5171
5172 private:
5173 // Work method in support of parallel rescan ... of young gen spaces
5174 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl,
5175 ContiguousSpace* space,
5176 HeapWord** chunk_array, size_t chunk_top);
5177
5178 // ... of dirty cards in old space
5179 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
5180 Par_MarkRefsIntoAndScanClosure* cl);
5181
5182 // ... work stealing for the above
5183 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
5184 };
5185
5186 class RemarkKlassClosure : public KlassClosure {
5187 CMKlassClosure _cm_klass_closure;
5188 public:
5189 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
5190 void do_klass(Klass* k) {
5191 // Check if we have modified any oops in the Klass during the concurrent marking.
5192 if (k->has_accumulated_modified_oops()) {
5193 k->clear_accumulated_modified_oops();
5194
5195 // We could have transfered the current modified marks to the accumulated marks,
5196 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
5197 } else if (k->has_modified_oops()) {
5198 // Don't clear anything, this info is needed by the next young collection.
5199 } else {
5200 // No modified oops in the Klass.
5201 return;
5202 }
5203
5204 // The klass has modified fields, need to scan the klass.
5205 _cm_klass_closure.do_klass(k);
5206 }
5207 };
5208
5209 // work_queue(i) is passed to the closure
5210 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
5211 // also is passed to do_dirty_card_rescan_tasks() and to
5212 // do_work_steal() to select the i-th task_queue.
5213
5214 void CMSParRemarkTask::work(uint worker_id) {
5215 elapsedTimer _timer;
5216 ResourceMark rm;
5217 HandleMark hm;
5218
5219 // ---------- rescan from roots --------------
5220 _timer.start();
5221 GenCollectedHeap* gch = GenCollectedHeap::heap();
5222 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
5223 _collector->_span, _collector->ref_processor(),
5224 &(_collector->_markBitMap),
5225 work_queue(worker_id));
5226
5227 // Rescan young gen roots first since these are likely
5228 // coarsely partitioned and may, on that account, constitute
5229 // the critical path; thus, it's best to start off that
5230 // work first.
5231 // ---------- young gen roots --------------
5232 {
5233 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
5234 EdenSpace* eden_space = dng->eden();
5235 ContiguousSpace* from_space = dng->from();
5236 ContiguousSpace* to_space = dng->to();
5237
5238 HeapWord** eca = _collector->_eden_chunk_array;
5239 size_t ect = _collector->_eden_chunk_index;
5240 HeapWord** sca = _collector->_survivor_chunk_array;
5241 size_t sct = _collector->_survivor_chunk_index;
5242
5243 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
5244 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
5245
5246 do_young_space_rescan(worker_id, &par_mrias_cl, to_space, NULL, 0);
5247 do_young_space_rescan(worker_id, &par_mrias_cl, from_space, sca, sct);
5248 do_young_space_rescan(worker_id, &par_mrias_cl, eden_space, eca, ect);
5249
5250 _timer.stop();
5251 if (PrintCMSStatistics != 0) {
5252 gclog_or_tty->print_cr(
5253 "Finished young gen rescan work in %dth thread: %3.3f sec",
5254 worker_id, _timer.seconds());
5255 }
5256 }
5257
5258 // ---------- remaining roots --------------
5259 _timer.reset();
5260 _timer.start();
5261 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
5262 false, // yg was scanned above
5263 false, // this is parallel code
5264 false, // not scavenging
5265 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5266 &par_mrias_cl,
5267 true, // walk all of code cache if (so & SO_CodeCache)
5268 NULL,
5269 NULL); // The dirty klasses will be handled below
5270 assert(_collector->should_unload_classes()
5271 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache),
5272 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5273 _timer.stop();
5274 if (PrintCMSStatistics != 0) {
5275 gclog_or_tty->print_cr(
5276 "Finished remaining root rescan work in %dth thread: %3.3f sec",
5277 worker_id, _timer.seconds());
5278 }
5279
5280 // ---------- unhandled CLD scanning ----------
5281 if (worker_id == 0) { // Single threaded at the moment.
5282 _timer.reset();
5283 _timer.start();
5284
5285 // Scan all new class loader data objects and new dependencies that were
5286 // introduced during concurrent marking.
5287 ResourceMark rm;
5288 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5289 for (int i = 0; i < array->length(); i++) {
5290 par_mrias_cl.do_class_loader_data(array->at(i));
5291 }
5292
5293 // We don't need to keep track of new CLDs anymore.
5294 ClassLoaderDataGraph::remember_new_clds(false);
5295
5296 _timer.stop();
5297 if (PrintCMSStatistics != 0) {
5298 gclog_or_tty->print_cr(
5299 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec",
5300 worker_id, _timer.seconds());
5301 }
5302 }
5303
5304 // ---------- dirty klass scanning ----------
5305 if (worker_id == 0) { // Single threaded at the moment.
5306 _timer.reset();
5307 _timer.start();
5308
5309 // Scan all classes that was dirtied during the concurrent marking phase.
5310 RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
5311 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5312
5313 _timer.stop();
5314 if (PrintCMSStatistics != 0) {
5315 gclog_or_tty->print_cr(
5316 "Finished dirty klass scanning work in %dth thread: %3.3f sec",
5317 worker_id, _timer.seconds());
5318 }
5319 }
5320
5321 // We might have added oops to ClassLoaderData::_handles during the
5322 // concurrent marking phase. These oops point to newly allocated objects
5323 // that are guaranteed to be kept alive. Either by the direct allocation
5324 // code, or when the young collector processes the strong roots. Hence,
5325 // we don't have to revisit the _handles block during the remark phase.
5326
5327 // ---------- rescan dirty cards ------------
5328 _timer.reset();
5329 _timer.start();
5330
5331 // Do the rescan tasks for each of the two spaces
5332 // (cms_space) in turn.
5333 // "worker_id" is passed to select the task_queue for "worker_id"
5334 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
5335 _timer.stop();
5336 if (PrintCMSStatistics != 0) {
5337 gclog_or_tty->print_cr(
5338 "Finished dirty card rescan work in %dth thread: %3.3f sec",
5339 worker_id, _timer.seconds());
5340 }
5341
5342 // ---------- steal work from other threads ...
5343 // ---------- ... and drain overflow list.
5344 _timer.reset();
5345 _timer.start();
5346 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
5347 _timer.stop();
5348 if (PrintCMSStatistics != 0) {
5349 gclog_or_tty->print_cr(
5350 "Finished work stealing in %dth thread: %3.3f sec",
5351 worker_id, _timer.seconds());
5352 }
5353 }
5354
5355 // Note that parameter "i" is not used.
5356 void
5357 CMSParRemarkTask::do_young_space_rescan(int i,
5358 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space,
5359 HeapWord** chunk_array, size_t chunk_top) {
5360 // Until all tasks completed:
5361 // . claim an unclaimed task
5362 // . compute region boundaries corresponding to task claimed
5363 // using chunk_array
5364 // . par_oop_iterate(cl) over that region
5365
5366 ResourceMark rm;
5367 HandleMark hm;
5368
5369 SequentialSubTasksDone* pst = space->par_seq_tasks();
5370 assert(pst->valid(), "Uninitialized use?");
5371
5372 uint nth_task = 0;
5373 uint n_tasks = pst->n_tasks();
5374
5375 HeapWord *start, *end;
5376 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5377 // We claimed task # nth_task; compute its boundaries.
5378 if (chunk_top == 0) { // no samples were taken
5379 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
5380 start = space->bottom();
5381 end = space->top();
5382 } else if (nth_task == 0) {
5383 start = space->bottom();
5384 end = chunk_array[nth_task];
5385 } else if (nth_task < (uint)chunk_top) {
5386 assert(nth_task >= 1, "Control point invariant");
5387 start = chunk_array[nth_task - 1];
5388 end = chunk_array[nth_task];
5389 } else {
5390 assert(nth_task == (uint)chunk_top, "Control point invariant");
5391 start = chunk_array[chunk_top - 1];
5392 end = space->top();
5393 }
5394 MemRegion mr(start, end);
5395 // Verify that mr is in space
5396 assert(mr.is_empty() || space->used_region().contains(mr),
5397 "Should be in space");
5398 // Verify that "start" is an object boundary
5399 assert(mr.is_empty() || oop(mr.start())->is_oop(),
5400 "Should be an oop");
5401 space->par_oop_iterate(mr, cl);
5402 }
5403 pst->all_tasks_completed();
5404 }
5405
5406 void
5407 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5408 CompactibleFreeListSpace* sp, int i,
5409 Par_MarkRefsIntoAndScanClosure* cl) {
5410 // Until all tasks completed:
5411 // . claim an unclaimed task
5412 // . compute region boundaries corresponding to task claimed
5413 // . transfer dirty bits ct->mut for that region
5414 // . apply rescanclosure to dirty mut bits for that region
5415
5416 ResourceMark rm;
5417 HandleMark hm;
5418
5419 OopTaskQueue* work_q = work_queue(i);
5420 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5421 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5422 // CAUTION: This closure has state that persists across calls to
5423 // the work method dirty_range_iterate_clear() in that it has
5424 // imbedded in it a (subtype of) UpwardsObjectClosure. The
5425 // use of that state in the imbedded UpwardsObjectClosure instance
5426 // assumes that the cards are always iterated (even if in parallel
5427 // by several threads) in monotonically increasing order per each
5428 // thread. This is true of the implementation below which picks
5429 // card ranges (chunks) in monotonically increasing order globally
5430 // and, a-fortiori, in monotonically increasing order per thread
5431 // (the latter order being a subsequence of the former).
5432 // If the work code below is ever reorganized into a more chaotic
5433 // work-partitioning form than the current "sequential tasks"
5434 // paradigm, the use of that persistent state will have to be
5435 // revisited and modified appropriately. See also related
5436 // bug 4756801 work on which should examine this code to make
5437 // sure that the changes there do not run counter to the
5438 // assumptions made here and necessary for correctness and
5439 // efficiency. Note also that this code might yield inefficient
5440 // behaviour in the case of very large objects that span one or
5441 // more work chunks. Such objects would potentially be scanned
5442 // several times redundantly. Work on 4756801 should try and
5443 // address that performance anomaly if at all possible. XXX
5444 MemRegion full_span = _collector->_span;
5445 CMSBitMap* bm = &(_collector->_markBitMap); // shared
5446 MarkFromDirtyCardsClosure
5447 greyRescanClosure(_collector, full_span, // entire span of interest
5448 sp, bm, work_q, cl);
5449
5450 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5451 assert(pst->valid(), "Uninitialized use?");
5452 uint nth_task = 0;
5453 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5454 MemRegion span = sp->used_region();
5455 HeapWord* start_addr = span.start();
5456 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5457 alignment);
5458 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5459 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5460 start_addr, "Check alignment");
5461 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5462 chunk_size, "Check alignment");
5463
5464 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5465 // Having claimed the nth_task, compute corresponding mem-region,
5466 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
5467 // The alignment restriction ensures that we do not need any
5468 // synchronization with other gang-workers while setting or
5469 // clearing bits in thus chunk of the MUT.
5470 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5471 start_addr + (nth_task+1)*chunk_size);
5472 // The last chunk's end might be way beyond end of the
5473 // used region. In that case pull back appropriately.
5474 if (this_span.end() > end_addr) {
5475 this_span.set_end(end_addr);
5476 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5477 }
5478 // Iterate over the dirty cards covering this chunk, marking them
5479 // precleaned, and setting the corresponding bits in the mod union
5480 // table. Since we have been careful to partition at Card and MUT-word
5481 // boundaries no synchronization is needed between parallel threads.
5482 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5483 &modUnionClosure);
5484
5485 // Having transferred these marks into the modUnionTable,
5486 // rescan the marked objects on the dirty cards in the modUnionTable.
5487 // Even if this is at a synchronous collection, the initial marking
5488 // may have been done during an asynchronous collection so there
5489 // may be dirty bits in the mod-union table.
5490 _collector->_modUnionTable.dirty_range_iterate_clear(
5491 this_span, &greyRescanClosure);
5492 _collector->_modUnionTable.verifyNoOneBitsInRange(
5493 this_span.start(),
5494 this_span.end());
5495 }
5496 pst->all_tasks_completed(); // declare that i am done
5497 }
5498
5499 // . see if we can share work_queues with ParNew? XXX
5500 void
5501 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5502 int* seed) {
5503 OopTaskQueue* work_q = work_queue(i);
5504 NOT_PRODUCT(int num_steals = 0;)
5505 oop obj_to_scan;
5506 CMSBitMap* bm = &(_collector->_markBitMap);
5507
5508 while (true) {
5509 // Completely finish any left over work from (an) earlier round(s)
5510 cl->trim_queue(0);
5511 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5512 (size_t)ParGCDesiredObjsFromOverflowList);
5513 // Now check if there's any work in the overflow list
5514 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5515 // only affects the number of attempts made to get work from the
5516 // overflow list and does not affect the number of workers. Just
5517 // pass ParallelGCThreads so this behavior is unchanged.
5518 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5519 work_q,
5520 ParallelGCThreads)) {
5521 // found something in global overflow list;
5522 // not yet ready to go stealing work from others.
5523 // We'd like to assert(work_q->size() != 0, ...)
5524 // because we just took work from the overflow list,
5525 // but of course we can't since all of that could have
5526 // been already stolen from us.
5527 // "He giveth and He taketh away."
5528 continue;
5529 }
5530 // Verify that we have no work before we resort to stealing
5531 assert(work_q->size() == 0, "Have work, shouldn't steal");
5532 // Try to steal from other queues that have work
5533 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5534 NOT_PRODUCT(num_steals++;)
5535 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5536 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5537 // Do scanning work
5538 obj_to_scan->oop_iterate(cl);
5539 // Loop around, finish this work, and try to steal some more
5540 } else if (terminator()->offer_termination()) {
5541 break; // nirvana from the infinite cycle
5542 }
5543 }
5544 NOT_PRODUCT(
5545 if (PrintCMSStatistics != 0) {
5546 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5547 }
5548 )
5549 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5550 "Else our work is not yet done");
5551 }
5552
5553 // Return a thread-local PLAB recording array, as appropriate.
5554 void* CMSCollector::get_data_recorder(int thr_num) {
5555 if (_survivor_plab_array != NULL &&
5556 (CMSPLABRecordAlways ||
5557 (_collectorState > Marking && _collectorState < FinalMarking))) {
5558 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5559 ChunkArray* ca = &_survivor_plab_array[thr_num];
5560 ca->reset(); // clear it so that fresh data is recorded
5561 return (void*) ca;
5562 } else {
5563 return NULL;
5564 }
5565 }
5566
5567 // Reset all the thread-local PLAB recording arrays
5568 void CMSCollector::reset_survivor_plab_arrays() {
5569 for (uint i = 0; i < ParallelGCThreads; i++) {
5570 _survivor_plab_array[i].reset();
5571 }
5572 }
5573
5574 // Merge the per-thread plab arrays into the global survivor chunk
5575 // array which will provide the partitioning of the survivor space
5576 // for CMS rescan.
5577 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
5578 int no_of_gc_threads) {
5579 assert(_survivor_plab_array != NULL, "Error");
5580 assert(_survivor_chunk_array != NULL, "Error");
5581 assert(_collectorState == FinalMarking, "Error");
5582 for (int j = 0; j < no_of_gc_threads; j++) {
5583 _cursor[j] = 0;
5584 }
5585 HeapWord* top = surv->top();
5586 size_t i;
5587 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
5588 HeapWord* min_val = top; // Higher than any PLAB address
5589 uint min_tid = 0; // position of min_val this round
5590 for (int j = 0; j < no_of_gc_threads; j++) {
5591 ChunkArray* cur_sca = &_survivor_plab_array[j];
5592 if (_cursor[j] == cur_sca->end()) {
5593 continue;
5594 }
5595 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5596 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5597 assert(surv->used_region().contains(cur_val), "Out of bounds value");
5598 if (cur_val < min_val) {
5599 min_tid = j;
5600 min_val = cur_val;
5601 } else {
5602 assert(cur_val < top, "All recorded addresses should be less");
5603 }
5604 }
5605 // At this point min_val and min_tid are respectively
5606 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5607 // and the thread (j) that witnesses that address.
5608 // We record this address in the _survivor_chunk_array[i]
5609 // and increment _cursor[min_tid] prior to the next round i.
5610 if (min_val == top) {
5611 break;
5612 }
5613 _survivor_chunk_array[i] = min_val;
5614 _cursor[min_tid]++;
5615 }
5616 // We are all done; record the size of the _survivor_chunk_array
5617 _survivor_chunk_index = i; // exclusive: [0, i)
5618 if (PrintCMSStatistics > 0) {
5619 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5620 }
5621 // Verify that we used up all the recorded entries
5622 #ifdef ASSERT
5623 size_t total = 0;
5624 for (int j = 0; j < no_of_gc_threads; j++) {
5625 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5626 total += _cursor[j];
5627 }
5628 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5629 // Check that the merged array is in sorted order
5630 if (total > 0) {
5631 for (size_t i = 0; i < total - 1; i++) {
5632 if (PrintCMSStatistics > 0) {
5633 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5634 i, _survivor_chunk_array[i]);
5635 }
5636 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5637 "Not sorted");
5638 }
5639 }
5640 #endif // ASSERT
5641 }
5642
5643 // Set up the space's par_seq_tasks structure for work claiming
5644 // for parallel rescan of young gen.
5645 // See ParRescanTask where this is currently used.
5646 void
5647 CMSCollector::
5648 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5649 assert(n_threads > 0, "Unexpected n_threads argument");
5650 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5651
5652 // Eden space
5653 {
5654 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5655 assert(!pst->valid(), "Clobbering existing data?");
5656 // Each valid entry in [0, _eden_chunk_index) represents a task.
5657 size_t n_tasks = _eden_chunk_index + 1;
5658 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5659 // Sets the condition for completion of the subtask (how many threads
5660 // need to finish in order to be done).
5661 pst->set_n_threads(n_threads);
5662 pst->set_n_tasks((int)n_tasks);
5663 }
5664
5665 // Merge the survivor plab arrays into _survivor_chunk_array
5666 if (_survivor_plab_array != NULL) {
5667 merge_survivor_plab_arrays(dng->from(), n_threads);
5668 } else {
5669 assert(_survivor_chunk_index == 0, "Error");
5670 }
5671
5672 // To space
5673 {
5674 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5675 assert(!pst->valid(), "Clobbering existing data?");
5676 // Sets the condition for completion of the subtask (how many threads
5677 // need to finish in order to be done).
5678 pst->set_n_threads(n_threads);
5679 pst->set_n_tasks(1);
5680 assert(pst->valid(), "Error");
5681 }
5682
5683 // From space
5684 {
5685 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5686 assert(!pst->valid(), "Clobbering existing data?");
5687 size_t n_tasks = _survivor_chunk_index + 1;
5688 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5689 // Sets the condition for completion of the subtask (how many threads
5690 // need to finish in order to be done).
5691 pst->set_n_threads(n_threads);
5692 pst->set_n_tasks((int)n_tasks);
5693 assert(pst->valid(), "Error");
5694 }
5695 }
5696
5697 // Parallel version of remark
5698 void CMSCollector::do_remark_parallel() {
5699 GenCollectedHeap* gch = GenCollectedHeap::heap();
5700 FlexibleWorkGang* workers = gch->workers();
5701 assert(workers != NULL, "Need parallel worker threads.");
5702 // Choose to use the number of GC workers most recently set
5703 // into "active_workers". If active_workers is not set, set it
5704 // to ParallelGCThreads.
5705 int n_workers = workers->active_workers();
5706 if (n_workers == 0) {
5707 assert(n_workers > 0, "Should have been set during scavenge");
5708 n_workers = ParallelGCThreads;
5709 workers->set_active_workers(n_workers);
5710 }
5711 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5712
5713 CMSParRemarkTask tsk(this,
5714 cms_space,
5715 n_workers, workers, task_queues());
5716
5717 // Set up for parallel process_strong_roots work.
5718 gch->set_par_threads(n_workers);
5719 // We won't be iterating over the cards in the card table updating
5720 // the younger_gen cards, so we shouldn't call the following else
5721 // the verification code as well as subsequent younger_refs_iterate
5722 // code would get confused. XXX
5723 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5724
5725 // The young gen rescan work will not be done as part of
5726 // process_strong_roots (which currently doesn't knw how to
5727 // parallelize such a scan), but rather will be broken up into
5728 // a set of parallel tasks (via the sampling that the [abortable]
5729 // preclean phase did of EdenSpace, plus the [two] tasks of
5730 // scanning the [two] survivor spaces. Further fine-grain
5731 // parallelization of the scanning of the survivor spaces
5732 // themselves, and of precleaning of the younger gen itself
5733 // is deferred to the future.
5734 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5735
5736 // The dirty card rescan work is broken up into a "sequence"
5737 // of parallel tasks (per constituent space) that are dynamically
5738 // claimed by the parallel threads.
5739 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5740
5741 // It turns out that even when we're using 1 thread, doing the work in a
5742 // separate thread causes wide variance in run times. We can't help this
5743 // in the multi-threaded case, but we special-case n=1 here to get
5744 // repeatable measurements of the 1-thread overhead of the parallel code.
5745 if (n_workers > 1) {
5746 // Make refs discovery MT-safe, if it isn't already: it may not
5747 // necessarily be so, since it's possible that we are doing
5748 // ST marking.
5749 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5750 GenCollectedHeap::StrongRootsScope srs(gch);
5751 workers->run_task(&tsk);
5752 } else {
5753 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5754 GenCollectedHeap::StrongRootsScope srs(gch);
5755 tsk.work(0);
5756 }
5757
5758 gch->set_par_threads(0); // 0 ==> non-parallel.
5759 // restore, single-threaded for now, any preserved marks
5760 // as a result of work_q overflow
5761 restore_preserved_marks_if_any();
5762 }
5763
5764 // Non-parallel version of remark
5765 void CMSCollector::do_remark_non_parallel() {
5766 ResourceMark rm;
5767 HandleMark hm;
5768 GenCollectedHeap* gch = GenCollectedHeap::heap();
5769 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5770
5771 MarkRefsIntoAndScanClosure
5772 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
5773 &_markStack, this,
5774 false /* should_yield */, false /* not precleaning */);
5775 MarkFromDirtyCardsClosure
5776 markFromDirtyCardsClosure(this, _span,
5777 NULL, // space is set further below
5778 &_markBitMap, &_markStack, &mrias_cl);
5779 {
5780 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm);
5781 // Iterate over the dirty cards, setting the corresponding bits in the
5782 // mod union table.
5783 {
5784 ModUnionClosure modUnionClosure(&_modUnionTable);
5785 _ct->ct_bs()->dirty_card_iterate(
5786 _cmsGen->used_region(),
5787 &modUnionClosure);
5788 }
5789 // Having transferred these marks into the modUnionTable, we just need
5790 // to rescan the marked objects on the dirty cards in the modUnionTable.
5791 // The initial marking may have been done during an asynchronous
5792 // collection so there may be dirty bits in the mod-union table.
5793 const int alignment =
5794 CardTableModRefBS::card_size * BitsPerWord;
5795 {
5796 // ... First handle dirty cards in CMS gen
5797 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5798 MemRegion ur = _cmsGen->used_region();
5799 HeapWord* lb = ur.start();
5800 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5801 MemRegion cms_span(lb, ub);
5802 _modUnionTable.dirty_range_iterate_clear(cms_span,
5803 &markFromDirtyCardsClosure);
5804 verify_work_stacks_empty();
5805 if (PrintCMSStatistics != 0) {
5806 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5807 markFromDirtyCardsClosure.num_dirty_cards());
5808 }
5809 }
5810 }
5811 if (VerifyDuringGC &&
5812 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5813 HandleMark hm; // Discard invalid handles created during verification
5814 Universe::verify();
5815 }
5816 {
5817 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm);
5818
5819 verify_work_stacks_empty();
5820
5821 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5822 GenCollectedHeap::StrongRootsScope srs(gch);
5823 gch->gen_process_strong_roots(_cmsGen->level(),
5824 true, // younger gens as roots
5825 false, // use the local StrongRootsScope
5826 false, // not scavenging
5827 SharedHeap::ScanningOption(roots_scanning_options()),
5828 &mrias_cl,
5829 true, // walk code active on stacks
5830 NULL,
5831 NULL); // The dirty klasses will be handled below
5832
5833 assert(should_unload_classes()
5834 || (roots_scanning_options() & SharedHeap::SO_CodeCache),
5835 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5836 }
5837
5838 {
5839 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm);
5840
5841 verify_work_stacks_empty();
5842
5843 // Scan all class loader data objects that might have been introduced
5844 // during concurrent marking.
5845 ResourceMark rm;
5846 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5847 for (int i = 0; i < array->length(); i++) {
5848 mrias_cl.do_class_loader_data(array->at(i));
5849 }
5850
5851 // We don't need to keep track of new CLDs anymore.
5852 ClassLoaderDataGraph::remember_new_clds(false);
5853
5854 verify_work_stacks_empty();
5855 }
5856
5857 {
5858 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm);
5859
5860 verify_work_stacks_empty();
5861
5862 RemarkKlassClosure remark_klass_closure(&mrias_cl);
5863 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5864
5865 verify_work_stacks_empty();
5866 }
5867
5868 // We might have added oops to ClassLoaderData::_handles during the
5869 // concurrent marking phase. These oops point to newly allocated objects
5870 // that are guaranteed to be kept alive. Either by the direct allocation
5871 // code, or when the young collector processes the strong roots. Hence,
5872 // we don't have to revisit the _handles block during the remark phase.
5873
5874 verify_work_stacks_empty();
5875 // Restore evacuated mark words, if any, used for overflow list links
5876 if (!CMSOverflowEarlyRestoration) {
5877 restore_preserved_marks_if_any();
5878 }
5879 verify_overflow_empty();
5880 }
5881
5882 ////////////////////////////////////////////////////////
5883 // Parallel Reference Processing Task Proxy Class
5884 ////////////////////////////////////////////////////////
5885 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5886 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5887 CMSCollector* _collector;
5888 CMSBitMap* _mark_bit_map;
5889 const MemRegion _span;
5890 ProcessTask& _task;
5891
5892 public:
5893 CMSRefProcTaskProxy(ProcessTask& task,
5894 CMSCollector* collector,
5895 const MemRegion& span,
5896 CMSBitMap* mark_bit_map,
5897 AbstractWorkGang* workers,
5898 OopTaskQueueSet* task_queues):
5899 // XXX Should superclass AGTWOQ also know about AWG since it knows
5900 // about the task_queues used by the AWG? Then it could initialize
5901 // the terminator() object. See 6984287. The set_for_termination()
5902 // below is a temporary band-aid for the regression in 6984287.
5903 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5904 task_queues),
5905 _task(task),
5906 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5907 {
5908 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5909 "Inconsistency in _span");
5910 set_for_termination(workers->active_workers());
5911 }
5912
5913 OopTaskQueueSet* task_queues() { return queues(); }
5914
5915 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5916
5917 void do_work_steal(int i,
5918 CMSParDrainMarkingStackClosure* drain,
5919 CMSParKeepAliveClosure* keep_alive,
5920 int* seed);
5921
5922 virtual void work(uint worker_id);
5923 };
5924
5925 void CMSRefProcTaskProxy::work(uint worker_id) {
5926 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5927 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5928 _mark_bit_map,
5929 work_queue(worker_id));
5930 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5931 _mark_bit_map,
5932 work_queue(worker_id));
5933 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5934 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5935 if (_task.marks_oops_alive()) {
5936 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5937 _collector->hash_seed(worker_id));
5938 }
5939 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5940 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5941 }
5942
5943 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5944 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5945 EnqueueTask& _task;
5946
5947 public:
5948 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5949 : AbstractGangTask("Enqueue reference objects in parallel"),
5950 _task(task)
5951 { }
5952
5953 virtual void work(uint worker_id)
5954 {
5955 _task.work(worker_id);
5956 }
5957 };
5958
5959 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5960 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5961 _span(span),
5962 _bit_map(bit_map),
5963 _work_queue(work_queue),
5964 _mark_and_push(collector, span, bit_map, work_queue),
5965 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5966 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5967 { }
5968
5969 // . see if we can share work_queues with ParNew? XXX
5970 void CMSRefProcTaskProxy::do_work_steal(int i,
5971 CMSParDrainMarkingStackClosure* drain,
5972 CMSParKeepAliveClosure* keep_alive,
5973 int* seed) {
5974 OopTaskQueue* work_q = work_queue(i);
5975 NOT_PRODUCT(int num_steals = 0;)
5976 oop obj_to_scan;
5977
5978 while (true) {
5979 // Completely finish any left over work from (an) earlier round(s)
5980 drain->trim_queue(0);
5981 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5982 (size_t)ParGCDesiredObjsFromOverflowList);
5983 // Now check if there's any work in the overflow list
5984 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5985 // only affects the number of attempts made to get work from the
5986 // overflow list and does not affect the number of workers. Just
5987 // pass ParallelGCThreads so this behavior is unchanged.
5988 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5989 work_q,
5990 ParallelGCThreads)) {
5991 // Found something in global overflow list;
5992 // not yet ready to go stealing work from others.
5993 // We'd like to assert(work_q->size() != 0, ...)
5994 // because we just took work from the overflow list,
5995 // but of course we can't, since all of that might have
5996 // been already stolen from us.
5997 continue;
5998 }
5999 // Verify that we have no work before we resort to stealing
6000 assert(work_q->size() == 0, "Have work, shouldn't steal");
6001 // Try to steal from other queues that have work
6002 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
6003 NOT_PRODUCT(num_steals++;)
6004 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
6005 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
6006 // Do scanning work
6007 obj_to_scan->oop_iterate(keep_alive);
6008 // Loop around, finish this work, and try to steal some more
6009 } else if (terminator()->offer_termination()) {
6010 break; // nirvana from the infinite cycle
6011 }
6012 }
6013 NOT_PRODUCT(
6014 if (PrintCMSStatistics != 0) {
6015 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
6016 }
6017 )
6018 }
6019
6020 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
6021 {
6022 GenCollectedHeap* gch = GenCollectedHeap::heap();
6023 FlexibleWorkGang* workers = gch->workers();
6024 assert(workers != NULL, "Need parallel worker threads.");
6025 CMSRefProcTaskProxy rp_task(task, &_collector,
6026 _collector.ref_processor()->span(),
6027 _collector.markBitMap(),
6028 workers, _collector.task_queues());
6029 workers->run_task(&rp_task);
6030 }
6031
6032 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
6033 {
6034
6035 GenCollectedHeap* gch = GenCollectedHeap::heap();
6036 FlexibleWorkGang* workers = gch->workers();
6037 assert(workers != NULL, "Need parallel worker threads.");
6038 CMSRefEnqueueTaskProxy enq_task(task);
6039 workers->run_task(&enq_task);
6040 }
6041
6042 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
6043
6044 ResourceMark rm;
6045 HandleMark hm;
6046
6047 ReferenceProcessor* rp = ref_processor();
6048 assert(rp->span().equals(_span), "Spans should be equal");
6049 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
6050 // Process weak references.
6051 rp->setup_policy(clear_all_soft_refs);
6052 verify_work_stacks_empty();
6053
6054 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
6055 &_markStack, false /* !preclean */);
6056 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
6057 _span, &_markBitMap, &_markStack,
6058 &cmsKeepAliveClosure, false /* !preclean */);
6059 {
6060 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm);
6061
6062 ReferenceProcessorStats stats;
6063 if (rp->processing_is_mt()) {
6064 // Set the degree of MT here. If the discovery is done MT, there
6065 // may have been a different number of threads doing the discovery
6066 // and a different number of discovered lists may have Ref objects.
6067 // That is OK as long as the Reference lists are balanced (see
6068 // balance_all_queues() and balance_queues()).
6069 GenCollectedHeap* gch = GenCollectedHeap::heap();
6070 int active_workers = ParallelGCThreads;
6071 FlexibleWorkGang* workers = gch->workers();
6072 if (workers != NULL) {
6073 active_workers = workers->active_workers();
6074 // The expectation is that active_workers will have already
6075 // been set to a reasonable value. If it has not been set,
6076 // investigate.
6077 assert(active_workers > 0, "Should have been set during scavenge");
6078 }
6079 rp->set_active_mt_degree(active_workers);
6080 CMSRefProcTaskExecutor task_executor(*this);
6081 stats = rp->process_discovered_references(&_is_alive_closure,
6082 &cmsKeepAliveClosure,
6083 &cmsDrainMarkingStackClosure,
6084 &task_executor,
6085 _gc_timer_cm);
6086 } else {
6087 stats = rp->process_discovered_references(&_is_alive_closure,
6088 &cmsKeepAliveClosure,
6089 &cmsDrainMarkingStackClosure,
6090 NULL,
6091 _gc_timer_cm);
6092 }
6093 _gc_tracer_cm->report_gc_reference_stats(stats);
6094
6095 }
6096
6097 // This is the point where the entire marking should have completed.
6098 verify_work_stacks_empty();
6099
6100 if (should_unload_classes()) {
6101 {
6102 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm);
6103
6104 // Unload classes and purge the SystemDictionary.
6105 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
6106
6107 // Unload nmethods.
6108 CodeCache::do_unloading(&_is_alive_closure, purged_class);
6109
6110 // Prune dead klasses from subklass/sibling/implementor lists.
6111 Klass::clean_weak_klass_links(&_is_alive_closure);
6112 }
6113
6114 {
6115 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm);
6116 // Clean up unreferenced symbols in symbol table.
6117 SymbolTable::unlink();
6118 }
6119 }
6120
6121 // CMS doesn't use the StringTable as hard roots when class unloading is turned off.
6122 // Need to check if we really scanned the StringTable.
6123 if ((roots_scanning_options() & SharedHeap::SO_Strings) == 0) {
6124 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm);
6125 // Delete entries for dead interned strings.
6126 StringTable::unlink(&_is_alive_closure);
6127 }
6128
6129 // Restore any preserved marks as a result of mark stack or
6130 // work queue overflow
6131 restore_preserved_marks_if_any(); // done single-threaded for now
6132
6133 rp->set_enqueuing_is_done(true);
6134 if (rp->processing_is_mt()) {
6135 rp->balance_all_queues();
6136 CMSRefProcTaskExecutor task_executor(*this);
6137 rp->enqueue_discovered_references(&task_executor);
6138 } else {
6139 rp->enqueue_discovered_references(NULL);
6140 }
6141 rp->verify_no_references_recorded();
6142 assert(!rp->discovery_enabled(), "should have been disabled");
6143 }
6144
6145 #ifndef PRODUCT
6146 void CMSCollector::check_correct_thread_executing() {
6147 Thread* t = Thread::current();
6148 // Only the VM thread or the CMS thread should be here.
6149 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
6150 "Unexpected thread type");
6151 // If this is the vm thread, the foreground process
6152 // should not be waiting. Note that _foregroundGCIsActive is
6153 // true while the foreground collector is waiting.
6154 if (_foregroundGCShouldWait) {
6155 // We cannot be the VM thread
6156 assert(t->is_ConcurrentGC_thread(),
6157 "Should be CMS thread");
6158 } else {
6159 // We can be the CMS thread only if we are in a stop-world
6160 // phase of CMS collection.
6161 if (t->is_ConcurrentGC_thread()) {
6162 assert(_collectorState == InitialMarking ||
6163 _collectorState == FinalMarking,
6164 "Should be a stop-world phase");
6165 // The CMS thread should be holding the CMS_token.
6166 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6167 "Potential interference with concurrently "
6168 "executing VM thread");
6169 }
6170 }
6171 }
6172 #endif
6173
6174 void CMSCollector::sweep(bool asynch) {
6175 assert(_collectorState == Sweeping, "just checking");
6176 check_correct_thread_executing();
6177 verify_work_stacks_empty();
6178 verify_overflow_empty();
6179 increment_sweep_count();
6180 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
6181
6182 _inter_sweep_timer.stop();
6183 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
6184 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
6185
6186 assert(!_intra_sweep_timer.is_active(), "Should not be active");
6187 _intra_sweep_timer.reset();
6188 _intra_sweep_timer.start();
6189 if (asynch) {
6190 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6191 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
6192 // First sweep the old gen
6193 {
6194 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6195 bitMapLock());
6196 sweepWork(_cmsGen, asynch);
6197 }
6198
6199 // Update Universe::_heap_*_at_gc figures.
6200 // We need all the free list locks to make the abstract state
6201 // transition from Sweeping to Resetting. See detailed note
6202 // further below.
6203 {
6204 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
6205 // Update heap occupancy information which is used as
6206 // input to soft ref clearing policy at the next gc.
6207 Universe::update_heap_info_at_gc();
6208 _collectorState = Resizing;
6209 }
6210 } else {
6211 // already have needed locks
6212 sweepWork(_cmsGen, asynch);
6213 // Update heap occupancy information which is used as
6214 // input to soft ref clearing policy at the next gc.
6215 Universe::update_heap_info_at_gc();
6216 _collectorState = Resizing;
6217 }
6218 verify_work_stacks_empty();
6219 verify_overflow_empty();
6220
6221 if (should_unload_classes()) {
6222 ClassLoaderDataGraph::purge();
6223 }
6224
6225 _intra_sweep_timer.stop();
6226 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
6227
6228 _inter_sweep_timer.reset();
6229 _inter_sweep_timer.start();
6230
6231 // We need to use a monotonically non-deccreasing time in ms
6232 // or we will see time-warp warnings and os::javaTimeMillis()
6233 // does not guarantee monotonicity.
6234 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
6235 update_time_of_last_gc(now);
6236
6237 // NOTE on abstract state transitions:
6238 // Mutators allocate-live and/or mark the mod-union table dirty
6239 // based on the state of the collection. The former is done in
6240 // the interval [Marking, Sweeping] and the latter in the interval
6241 // [Marking, Sweeping). Thus the transitions into the Marking state
6242 // and out of the Sweeping state must be synchronously visible
6243 // globally to the mutators.
6244 // The transition into the Marking state happens with the world
6245 // stopped so the mutators will globally see it. Sweeping is
6246 // done asynchronously by the background collector so the transition
6247 // from the Sweeping state to the Resizing state must be done
6248 // under the freelistLock (as is the check for whether to
6249 // allocate-live and whether to dirty the mod-union table).
6250 assert(_collectorState == Resizing, "Change of collector state to"
6251 " Resizing must be done under the freelistLocks (plural)");
6252
6253 // Now that sweeping has been completed, we clear
6254 // the incremental_collection_failed flag,
6255 // thus inviting a younger gen collection to promote into
6256 // this generation. If such a promotion may still fail,
6257 // the flag will be set again when a young collection is
6258 // attempted.
6259 GenCollectedHeap* gch = GenCollectedHeap::heap();
6260 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
6261 gch->update_full_collections_completed(_collection_count_start);
6262 }
6263
6264 // FIX ME!!! Looks like this belongs in CFLSpace, with
6265 // CMSGen merely delegating to it.
6266 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
6267 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
6268 HeapWord* minAddr = _cmsSpace->bottom();
6269 HeapWord* largestAddr =
6270 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
6271 if (largestAddr == NULL) {
6272 // The dictionary appears to be empty. In this case
6273 // try to coalesce at the end of the heap.
6274 largestAddr = _cmsSpace->end();
6275 }
6276 size_t largestOffset = pointer_delta(largestAddr, minAddr);
6277 size_t nearLargestOffset =
6278 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
6279 if (PrintFLSStatistics != 0) {
6280 gclog_or_tty->print_cr(
6281 "CMS: Large Block: " PTR_FORMAT ";"
6282 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
6283 largestAddr,
6284 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset);
6285 }
6286 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
6287 }
6288
6289 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
6290 return addr >= _cmsSpace->nearLargestChunk();
6291 }
6292
6293 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
6294 return _cmsSpace->find_chunk_at_end();
6295 }
6296
6297 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
6298 bool full) {
6299 // The next lower level has been collected. Gather any statistics
6300 // that are of interest at this point.
6301 if (!full && (current_level + 1) == level()) {
6302 // Gather statistics on the young generation collection.
6303 collector()->stats().record_gc0_end(used());
6304 }
6305 }
6306
6307 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
6308 GenCollectedHeap* gch = GenCollectedHeap::heap();
6309 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
6310 "Wrong type of heap");
6311 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
6312 gch->gen_policy()->size_policy();
6313 assert(sp->is_gc_cms_adaptive_size_policy(),
6314 "Wrong type of size policy");
6315 return sp;
6316 }
6317
6318 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
6319 if (PrintGCDetails && Verbose) {
6320 gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
6321 }
6322 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
6323 _debug_collection_type =
6324 (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
6325 if (PrintGCDetails && Verbose) {
6326 gclog_or_tty->print_cr("to %d ", _debug_collection_type);
6327 }
6328 }
6329
6330 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
6331 bool asynch) {
6332 // We iterate over the space(s) underlying this generation,
6333 // checking the mark bit map to see if the bits corresponding
6334 // to specific blocks are marked or not. Blocks that are
6335 // marked are live and are not swept up. All remaining blocks
6336 // are swept up, with coalescing on-the-fly as we sweep up
6337 // contiguous free and/or garbage blocks:
6338 // We need to ensure that the sweeper synchronizes with allocators
6339 // and stop-the-world collectors. In particular, the following
6340 // locks are used:
6341 // . CMS token: if this is held, a stop the world collection cannot occur
6342 // . freelistLock: if this is held no allocation can occur from this
6343 // generation by another thread
6344 // . bitMapLock: if this is held, no other thread can access or update
6345 //
6346
6347 // Note that we need to hold the freelistLock if we use
6348 // block iterate below; else the iterator might go awry if
6349 // a mutator (or promotion) causes block contents to change
6350 // (for instance if the allocator divvies up a block).
6351 // If we hold the free list lock, for all practical purposes
6352 // young generation GC's can't occur (they'll usually need to
6353 // promote), so we might as well prevent all young generation
6354 // GC's while we do a sweeping step. For the same reason, we might
6355 // as well take the bit map lock for the entire duration
6356
6357 // check that we hold the requisite locks
6358 assert(have_cms_token(), "Should hold cms token");
6359 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
6360 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
6361 "Should possess CMS token to sweep");
6362 assert_lock_strong(gen->freelistLock());
6363 assert_lock_strong(bitMapLock());
6364
6365 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
6366 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
6367 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
6368 _inter_sweep_estimate.padded_average(),
6369 _intra_sweep_estimate.padded_average());
6370 gen->setNearLargestChunk();
6371
6372 {
6373 SweepClosure sweepClosure(this, gen, &_markBitMap,
6374 CMSYield && asynch);
6375 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
6376 // We need to free-up/coalesce garbage/blocks from a
6377 // co-terminal free run. This is done in the SweepClosure
6378 // destructor; so, do not remove this scope, else the
6379 // end-of-sweep-census below will be off by a little bit.
6380 }
6381 gen->cmsSpace()->sweep_completed();
6382 gen->cmsSpace()->endSweepFLCensus(sweep_count());
6383 if (should_unload_classes()) { // unloaded classes this cycle,
6384 _concurrent_cycles_since_last_unload = 0; // ... reset count
6385 } else { // did not unload classes,
6386 _concurrent_cycles_since_last_unload++; // ... increment count
6387 }
6388 }
6389
6390 // Reset CMS data structures (for now just the marking bit map)
6391 // preparatory for the next cycle.
6392 void CMSCollector::reset(bool asynch) {
6393 GenCollectedHeap* gch = GenCollectedHeap::heap();
6394 CMSAdaptiveSizePolicy* sp = size_policy();
6395 AdaptiveSizePolicyOutput(sp, gch->total_collections());
6396 if (asynch) {
6397 CMSTokenSyncWithLocks ts(true, bitMapLock());
6398
6399 // If the state is not "Resetting", the foreground thread
6400 // has done a collection and the resetting.
6401 if (_collectorState != Resetting) {
6402 assert(_collectorState == Idling, "The state should only change"
6403 " because the foreground collector has finished the collection");
6404 return;
6405 }
6406
6407 // Clear the mark bitmap (no grey objects to start with)
6408 // for the next cycle.
6409 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6410 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
6411
6412 HeapWord* curAddr = _markBitMap.startWord();
6413 while (curAddr < _markBitMap.endWord()) {
6414 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
6415 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
6416 _markBitMap.clear_large_range(chunk);
6417 if (ConcurrentMarkSweepThread::should_yield() &&
6418 !foregroundGCIsActive() &&
6419 CMSYield) {
6420 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6421 "CMS thread should hold CMS token");
6422 assert_lock_strong(bitMapLock());
6423 bitMapLock()->unlock();
6424 ConcurrentMarkSweepThread::desynchronize(true);
6425 ConcurrentMarkSweepThread::acknowledge_yield_request();
6426 stopTimer();
6427 if (PrintCMSStatistics != 0) {
6428 incrementYields();
6429 }
6430 icms_wait();
6431
6432 // See the comment in coordinator_yield()
6433 for (unsigned i = 0; i < CMSYieldSleepCount &&
6434 ConcurrentMarkSweepThread::should_yield() &&
6435 !CMSCollector::foregroundGCIsActive(); ++i) {
6436 os::sleep(Thread::current(), 1, false);
6437 ConcurrentMarkSweepThread::acknowledge_yield_request();
6438 }
6439
6440 ConcurrentMarkSweepThread::synchronize(true);
6441 bitMapLock()->lock_without_safepoint_check();
6442 startTimer();
6443 }
6444 curAddr = chunk.end();
6445 }
6446 // A successful mostly concurrent collection has been done.
6447 // Because only the full (i.e., concurrent mode failure) collections
6448 // are being measured for gc overhead limits, clean the "near" flag
6449 // and count.
6450 sp->reset_gc_overhead_limit_count();
6451 _collectorState = Idling;
6452 } else {
6453 // already have the lock
6454 assert(_collectorState == Resetting, "just checking");
6455 assert_lock_strong(bitMapLock());
6456 _markBitMap.clear_all();
6457 _collectorState = Idling;
6458 }
6459
6460 // Stop incremental mode after a cycle completes, so that any future cycles
6461 // are triggered by allocation.
6462 stop_icms();
6463
6464 NOT_PRODUCT(
6465 if (RotateCMSCollectionTypes) {
6466 _cmsGen->rotate_debug_collection_type();
6467 }
6468 )
6469
6470 register_gc_end();
6471 }
6472
6473 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
6474 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
6475 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6476 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL);
6477 TraceCollectorStats tcs(counters());
6478
6479 switch (op) {
6480 case CMS_op_checkpointRootsInitial: {
6481 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6482 checkpointRootsInitial(true); // asynch
6483 if (PrintGC) {
6484 _cmsGen->printOccupancy("initial-mark");
6485 }
6486 break;
6487 }
6488 case CMS_op_checkpointRootsFinal: {
6489 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6490 checkpointRootsFinal(true, // asynch
6491 false, // !clear_all_soft_refs
6492 false); // !init_mark_was_synchronous
6493 if (PrintGC) {
6494 _cmsGen->printOccupancy("remark");
6495 }
6496 break;
6497 }
6498 default:
6499 fatal("No such CMS_op");
6500 }
6501 }
6502
6503 #ifndef PRODUCT
6504 size_t const CMSCollector::skip_header_HeapWords() {
6505 return FreeChunk::header_size();
6506 }
6507
6508 // Try and collect here conditions that should hold when
6509 // CMS thread is exiting. The idea is that the foreground GC
6510 // thread should not be blocked if it wants to terminate
6511 // the CMS thread and yet continue to run the VM for a while
6512 // after that.
6513 void CMSCollector::verify_ok_to_terminate() const {
6514 assert(Thread::current()->is_ConcurrentGC_thread(),
6515 "should be called by CMS thread");
6516 assert(!_foregroundGCShouldWait, "should be false");
6517 // We could check here that all the various low-level locks
6518 // are not held by the CMS thread, but that is overkill; see
6519 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6520 // is checked.
6521 }
6522 #endif
6523
6524 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6525 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6526 "missing Printezis mark?");
6527 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6528 size_t size = pointer_delta(nextOneAddr + 1, addr);
6529 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6530 "alignment problem");
6531 assert(size >= 3, "Necessary for Printezis marks to work");
6532 return size;
6533 }
6534
6535 // A variant of the above (block_size_using_printezis_bits()) except
6536 // that we return 0 if the P-bits are not yet set.
6537 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6538 if (_markBitMap.isMarked(addr + 1)) {
6539 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
6540 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6541 size_t size = pointer_delta(nextOneAddr + 1, addr);
6542 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6543 "alignment problem");
6544 assert(size >= 3, "Necessary for Printezis marks to work");
6545 return size;
6546 }
6547 return 0;
6548 }
6549
6550 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6551 size_t sz = 0;
6552 oop p = (oop)addr;
6553 if (p->klass_or_null() != NULL) {
6554 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6555 } else {
6556 sz = block_size_using_printezis_bits(addr);
6557 }
6558 assert(sz > 0, "size must be nonzero");
6559 HeapWord* next_block = addr + sz;
6560 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
6561 CardTableModRefBS::card_size);
6562 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
6563 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6564 "must be different cards");
6565 return next_card;
6566 }
6567
6568
6569 // CMS Bit Map Wrapper /////////////////////////////////////////
6570
6571 // Construct a CMS bit map infrastructure, but don't create the
6572 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6573 // further below.
6574 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6575 _bm(),
6576 _shifter(shifter),
6577 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6578 {
6579 _bmStartWord = 0;
6580 _bmWordSize = 0;
6581 }
6582
6583 bool CMSBitMap::allocate(MemRegion mr) {
6584 _bmStartWord = mr.start();
6585 _bmWordSize = mr.word_size();
6586 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6587 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6588 if (!brs.is_reserved()) {
6589 warning("CMS bit map allocation failure");
6590 return false;
6591 }
6592 // For now we'll just commit all of the bit map up fromt.
6593 // Later on we'll try to be more parsimonious with swap.
6594 if (!_virtual_space.initialize(brs, brs.size())) {
6595 warning("CMS bit map backing store failure");
6596 return false;
6597 }
6598 assert(_virtual_space.committed_size() == brs.size(),
6599 "didn't reserve backing store for all of CMS bit map?");
6600 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
6601 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6602 _bmWordSize, "inconsistency in bit map sizing");
6603 _bm.set_size(_bmWordSize >> _shifter);
6604
6605 // bm.clear(); // can we rely on getting zero'd memory? verify below
6606 assert(isAllClear(),
6607 "Expected zero'd memory from ReservedSpace constructor");
6608 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6609 "consistency check");
6610 return true;
6611 }
6612
6613 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6614 HeapWord *next_addr, *end_addr, *last_addr;
6615 assert_locked();
6616 assert(covers(mr), "out-of-range error");
6617 // XXX assert that start and end are appropriately aligned
6618 for (next_addr = mr.start(), end_addr = mr.end();
6619 next_addr < end_addr; next_addr = last_addr) {
6620 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6621 last_addr = dirty_region.end();
6622 if (!dirty_region.is_empty()) {
6623 cl->do_MemRegion(dirty_region);
6624 } else {
6625 assert(last_addr == end_addr, "program logic");
6626 return;
6627 }
6628 }
6629 }
6630
6631 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
6632 _bm.print_on_error(st, prefix);
6633 }
6634
6635 #ifndef PRODUCT
6636 void CMSBitMap::assert_locked() const {
6637 CMSLockVerifier::assert_locked(lock());
6638 }
6639
6640 bool CMSBitMap::covers(MemRegion mr) const {
6641 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6642 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6643 "size inconsistency");
6644 return (mr.start() >= _bmStartWord) &&
6645 (mr.end() <= endWord());
6646 }
6647
6648 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6649 return (start >= _bmStartWord && (start + size) <= endWord());
6650 }
6651
6652 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6653 // verify that there are no 1 bits in the interval [left, right)
6654 FalseBitMapClosure falseBitMapClosure;
6655 iterate(&falseBitMapClosure, left, right);
6656 }
6657
6658 void CMSBitMap::region_invariant(MemRegion mr)
6659 {
6660 assert_locked();
6661 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6662 assert(!mr.is_empty(), "unexpected empty region");
6663 assert(covers(mr), "mr should be covered by bit map");
6664 // convert address range into offset range
6665 size_t start_ofs = heapWordToOffset(mr.start());
6666 // Make sure that end() is appropriately aligned
6667 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6668 (1 << (_shifter+LogHeapWordSize))),
6669 "Misaligned mr.end()");
6670 size_t end_ofs = heapWordToOffset(mr.end());
6671 assert(end_ofs > start_ofs, "Should mark at least one bit");
6672 }
6673
6674 #endif
6675
6676 bool CMSMarkStack::allocate(size_t size) {
6677 // allocate a stack of the requisite depth
6678 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6679 size * sizeof(oop)));
6680 if (!rs.is_reserved()) {
6681 warning("CMSMarkStack allocation failure");
6682 return false;
6683 }
6684 if (!_virtual_space.initialize(rs, rs.size())) {
6685 warning("CMSMarkStack backing store failure");
6686 return false;
6687 }
6688 assert(_virtual_space.committed_size() == rs.size(),
6689 "didn't reserve backing store for all of CMS stack?");
6690 _base = (oop*)(_virtual_space.low());
6691 _index = 0;
6692 _capacity = size;
6693 NOT_PRODUCT(_max_depth = 0);
6694 return true;
6695 }
6696
6697 // XXX FIX ME !!! In the MT case we come in here holding a
6698 // leaf lock. For printing we need to take a further lock
6699 // which has lower rank. We need to recallibrate the two
6700 // lock-ranks involved in order to be able to rpint the
6701 // messages below. (Or defer the printing to the caller.
6702 // For now we take the expedient path of just disabling the
6703 // messages for the problematic case.)
6704 void CMSMarkStack::expand() {
6705 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6706 if (_capacity == MarkStackSizeMax) {
6707 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6708 // We print a warning message only once per CMS cycle.
6709 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6710 }
6711 return;
6712 }
6713 // Double capacity if possible
6714 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6715 // Do not give up existing stack until we have managed to
6716 // get the double capacity that we desired.
6717 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6718 new_capacity * sizeof(oop)));
6719 if (rs.is_reserved()) {
6720 // Release the backing store associated with old stack
6721 _virtual_space.release();
6722 // Reinitialize virtual space for new stack
6723 if (!_virtual_space.initialize(rs, rs.size())) {
6724 fatal("Not enough swap for expanded marking stack");
6725 }
6726 _base = (oop*)(_virtual_space.low());
6727 _index = 0;
6728 _capacity = new_capacity;
6729 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6730 // Failed to double capacity, continue;
6731 // we print a detail message only once per CMS cycle.
6732 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6733 SIZE_FORMAT"K",
6734 _capacity / K, new_capacity / K);
6735 }
6736 }
6737
6738
6739 // Closures
6740 // XXX: there seems to be a lot of code duplication here;
6741 // should refactor and consolidate common code.
6742
6743 // This closure is used to mark refs into the CMS generation in
6744 // the CMS bit map. Called at the first checkpoint. This closure
6745 // assumes that we do not need to re-mark dirty cards; if the CMS
6746 // generation on which this is used is not an oldest
6747 // generation then this will lose younger_gen cards!
6748
6749 MarkRefsIntoClosure::MarkRefsIntoClosure(
6750 MemRegion span, CMSBitMap* bitMap):
6751 _span(span),
6752 _bitMap(bitMap)
6753 {
6754 assert(_ref_processor == NULL, "deliberately left NULL");
6755 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6756 }
6757
6758 void MarkRefsIntoClosure::do_oop(oop obj) {
6759 // if p points into _span, then mark corresponding bit in _markBitMap
6760 assert(obj->is_oop(), "expected an oop");
6761 HeapWord* addr = (HeapWord*)obj;
6762 if (_span.contains(addr)) {
6763 // this should be made more efficient
6764 _bitMap->mark(addr);
6765 }
6766 }
6767
6768 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6769 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6770
6771 // A variant of the above, used for CMS marking verification.
6772 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6773 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6774 _span(span),
6775 _verification_bm(verification_bm),
6776 _cms_bm(cms_bm)
6777 {
6778 assert(_ref_processor == NULL, "deliberately left NULL");
6779 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6780 }
6781
6782 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6783 // if p points into _span, then mark corresponding bit in _markBitMap
6784 assert(obj->is_oop(), "expected an oop");
6785 HeapWord* addr = (HeapWord*)obj;
6786 if (_span.contains(addr)) {
6787 _verification_bm->mark(addr);
6788 if (!_cms_bm->isMarked(addr)) {
6789 oop(addr)->print();
6790 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6791 fatal("... aborting");
6792 }
6793 }
6794 }
6795
6796 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6797 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6798
6799 //////////////////////////////////////////////////
6800 // MarkRefsIntoAndScanClosure
6801 //////////////////////////////////////////////////
6802
6803 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6804 ReferenceProcessor* rp,
6805 CMSBitMap* bit_map,
6806 CMSBitMap* mod_union_table,
6807 CMSMarkStack* mark_stack,
6808 CMSCollector* collector,
6809 bool should_yield,
6810 bool concurrent_precleaning):
6811 _collector(collector),
6812 _span(span),
6813 _bit_map(bit_map),
6814 _mark_stack(mark_stack),
6815 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6816 mark_stack, concurrent_precleaning),
6817 _yield(should_yield),
6818 _concurrent_precleaning(concurrent_precleaning),
6819 _freelistLock(NULL)
6820 {
6821 _ref_processor = rp;
6822 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6823 }
6824
6825 // This closure is used to mark refs into the CMS generation at the
6826 // second (final) checkpoint, and to scan and transitively follow
6827 // the unmarked oops. It is also used during the concurrent precleaning
6828 // phase while scanning objects on dirty cards in the CMS generation.
6829 // The marks are made in the marking bit map and the marking stack is
6830 // used for keeping the (newly) grey objects during the scan.
6831 // The parallel version (Par_...) appears further below.
6832 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6833 if (obj != NULL) {
6834 assert(obj->is_oop(), "expected an oop");
6835 HeapWord* addr = (HeapWord*)obj;
6836 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6837 assert(_collector->overflow_list_is_empty(),
6838 "overflow list should be empty");
6839 if (_span.contains(addr) &&
6840 !_bit_map->isMarked(addr)) {
6841 // mark bit map (object is now grey)
6842 _bit_map->mark(addr);
6843 // push on marking stack (stack should be empty), and drain the
6844 // stack by applying this closure to the oops in the oops popped
6845 // from the stack (i.e. blacken the grey objects)
6846 bool res = _mark_stack->push(obj);
6847 assert(res, "Should have space to push on empty stack");
6848 do {
6849 oop new_oop = _mark_stack->pop();
6850 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6851 assert(_bit_map->isMarked((HeapWord*)new_oop),
6852 "only grey objects on this stack");
6853 // iterate over the oops in this oop, marking and pushing
6854 // the ones in CMS heap (i.e. in _span).
6855 new_oop->oop_iterate(&_pushAndMarkClosure);
6856 // check if it's time to yield
6857 do_yield_check();
6858 } while (!_mark_stack->isEmpty() ||
6859 (!_concurrent_precleaning && take_from_overflow_list()));
6860 // if marking stack is empty, and we are not doing this
6861 // during precleaning, then check the overflow list
6862 }
6863 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6864 assert(_collector->overflow_list_is_empty(),
6865 "overflow list was drained above");
6866 // We could restore evacuated mark words, if any, used for
6867 // overflow list links here because the overflow list is
6868 // provably empty here. That would reduce the maximum
6869 // size requirements for preserved_{oop,mark}_stack.
6870 // But we'll just postpone it until we are all done
6871 // so we can just stream through.
6872 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6873 _collector->restore_preserved_marks_if_any();
6874 assert(_collector->no_preserved_marks(), "No preserved marks");
6875 }
6876 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6877 "All preserved marks should have been restored above");
6878 }
6879 }
6880
6881 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6882 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6883
6884 void MarkRefsIntoAndScanClosure::do_yield_work() {
6885 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6886 "CMS thread should hold CMS token");
6887 assert_lock_strong(_freelistLock);
6888 assert_lock_strong(_bit_map->lock());
6889 // relinquish the free_list_lock and bitMaplock()
6890 _bit_map->lock()->unlock();
6891 _freelistLock->unlock();
6892 ConcurrentMarkSweepThread::desynchronize(true);
6893 ConcurrentMarkSweepThread::acknowledge_yield_request();
6894 _collector->stopTimer();
6895 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6896 if (PrintCMSStatistics != 0) {
6897 _collector->incrementYields();
6898 }
6899 _collector->icms_wait();
6900
6901 // See the comment in coordinator_yield()
6902 for (unsigned i = 0;
6903 i < CMSYieldSleepCount &&
6904 ConcurrentMarkSweepThread::should_yield() &&
6905 !CMSCollector::foregroundGCIsActive();
6906 ++i) {
6907 os::sleep(Thread::current(), 1, false);
6908 ConcurrentMarkSweepThread::acknowledge_yield_request();
6909 }
6910
6911 ConcurrentMarkSweepThread::synchronize(true);
6912 _freelistLock->lock_without_safepoint_check();
6913 _bit_map->lock()->lock_without_safepoint_check();
6914 _collector->startTimer();
6915 }
6916
6917 ///////////////////////////////////////////////////////////
6918 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6919 // MarkRefsIntoAndScanClosure
6920 ///////////////////////////////////////////////////////////
6921 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6922 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6923 CMSBitMap* bit_map, OopTaskQueue* work_queue):
6924 _span(span),
6925 _bit_map(bit_map),
6926 _work_queue(work_queue),
6927 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6928 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6929 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
6930 {
6931 _ref_processor = rp;
6932 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6933 }
6934
6935 // This closure is used to mark refs into the CMS generation at the
6936 // second (final) checkpoint, and to scan and transitively follow
6937 // the unmarked oops. The marks are made in the marking bit map and
6938 // the work_queue is used for keeping the (newly) grey objects during
6939 // the scan phase whence they are also available for stealing by parallel
6940 // threads. Since the marking bit map is shared, updates are
6941 // synchronized (via CAS).
6942 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6943 if (obj != NULL) {
6944 // Ignore mark word because this could be an already marked oop
6945 // that may be chained at the end of the overflow list.
6946 assert(obj->is_oop(true), "expected an oop");
6947 HeapWord* addr = (HeapWord*)obj;
6948 if (_span.contains(addr) &&
6949 !_bit_map->isMarked(addr)) {
6950 // mark bit map (object will become grey):
6951 // It is possible for several threads to be
6952 // trying to "claim" this object concurrently;
6953 // the unique thread that succeeds in marking the
6954 // object first will do the subsequent push on
6955 // to the work queue (or overflow list).
6956 if (_bit_map->par_mark(addr)) {
6957 // push on work_queue (which may not be empty), and trim the
6958 // queue to an appropriate length by applying this closure to
6959 // the oops in the oops popped from the stack (i.e. blacken the
6960 // grey objects)
6961 bool res = _work_queue->push(obj);
6962 assert(res, "Low water mark should be less than capacity?");
6963 trim_queue(_low_water_mark);
6964 } // Else, another thread claimed the object
6965 }
6966 }
6967 }
6968
6969 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6970 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6971
6972 // This closure is used to rescan the marked objects on the dirty cards
6973 // in the mod union table and the card table proper.
6974 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6975 oop p, MemRegion mr) {
6976
6977 size_t size = 0;
6978 HeapWord* addr = (HeapWord*)p;
6979 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6980 assert(_span.contains(addr), "we are scanning the CMS generation");
6981 // check if it's time to yield
6982 if (do_yield_check()) {
6983 // We yielded for some foreground stop-world work,
6984 // and we have been asked to abort this ongoing preclean cycle.
6985 return 0;
6986 }
6987 if (_bitMap->isMarked(addr)) {
6988 // it's marked; is it potentially uninitialized?
6989 if (p->klass_or_null() != NULL) {
6990 // an initialized object; ignore mark word in verification below
6991 // since we are running concurrent with mutators
6992 assert(p->is_oop(true), "should be an oop");
6993 if (p->is_objArray()) {
6994 // objArrays are precisely marked; restrict scanning
6995 // to dirty cards only.
6996 size = CompactibleFreeListSpace::adjustObjectSize(
6997 p->oop_iterate(_scanningClosure, mr));
6998 } else {
6999 // A non-array may have been imprecisely marked; we need
7000 // to scan object in its entirety.
7001 size = CompactibleFreeListSpace::adjustObjectSize(
7002 p->oop_iterate(_scanningClosure));
7003 }
7004 #ifdef ASSERT
7005 size_t direct_size =
7006 CompactibleFreeListSpace::adjustObjectSize(p->size());
7007 assert(size == direct_size, "Inconsistency in size");
7008 assert(size >= 3, "Necessary for Printezis marks to work");
7009 if (!_bitMap->isMarked(addr+1)) {
7010 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
7011 } else {
7012 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
7013 assert(_bitMap->isMarked(addr+size-1),
7014 "inconsistent Printezis mark");
7015 }
7016 #endif // ASSERT
7017 } else {
7018 // an unitialized object
7019 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
7020 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7021 size = pointer_delta(nextOneAddr + 1, addr);
7022 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7023 "alignment problem");
7024 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
7025 // will dirty the card when the klass pointer is installed in the
7026 // object (signalling the completion of initialization).
7027 }
7028 } else {
7029 // Either a not yet marked object or an uninitialized object
7030 if (p->klass_or_null() == NULL) {
7031 // An uninitialized object, skip to the next card, since
7032 // we may not be able to read its P-bits yet.
7033 assert(size == 0, "Initial value");
7034 } else {
7035 // An object not (yet) reached by marking: we merely need to
7036 // compute its size so as to go look at the next block.
7037 assert(p->is_oop(true), "should be an oop");
7038 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
7039 }
7040 }
7041 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7042 return size;
7043 }
7044
7045 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
7046 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7047 "CMS thread should hold CMS token");
7048 assert_lock_strong(_freelistLock);
7049 assert_lock_strong(_bitMap->lock());
7050 // relinquish the free_list_lock and bitMaplock()
7051 _bitMap->lock()->unlock();
7052 _freelistLock->unlock();
7053 ConcurrentMarkSweepThread::desynchronize(true);
7054 ConcurrentMarkSweepThread::acknowledge_yield_request();
7055 _collector->stopTimer();
7056 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7057 if (PrintCMSStatistics != 0) {
7058 _collector->incrementYields();
7059 }
7060 _collector->icms_wait();
7061
7062 // See the comment in coordinator_yield()
7063 for (unsigned i = 0; i < CMSYieldSleepCount &&
7064 ConcurrentMarkSweepThread::should_yield() &&
7065 !CMSCollector::foregroundGCIsActive(); ++i) {
7066 os::sleep(Thread::current(), 1, false);
7067 ConcurrentMarkSweepThread::acknowledge_yield_request();
7068 }
7069
7070 ConcurrentMarkSweepThread::synchronize(true);
7071 _freelistLock->lock_without_safepoint_check();
7072 _bitMap->lock()->lock_without_safepoint_check();
7073 _collector->startTimer();
7074 }
7075
7076
7077 //////////////////////////////////////////////////////////////////
7078 // SurvivorSpacePrecleanClosure
7079 //////////////////////////////////////////////////////////////////
7080 // This (single-threaded) closure is used to preclean the oops in
7081 // the survivor spaces.
7082 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
7083
7084 HeapWord* addr = (HeapWord*)p;
7085 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7086 assert(!_span.contains(addr), "we are scanning the survivor spaces");
7087 assert(p->klass_or_null() != NULL, "object should be initializd");
7088 // an initialized object; ignore mark word in verification below
7089 // since we are running concurrent with mutators
7090 assert(p->is_oop(true), "should be an oop");
7091 // Note that we do not yield while we iterate over
7092 // the interior oops of p, pushing the relevant ones
7093 // on our marking stack.
7094 size_t size = p->oop_iterate(_scanning_closure);
7095 do_yield_check();
7096 // Observe that below, we do not abandon the preclean
7097 // phase as soon as we should; rather we empty the
7098 // marking stack before returning. This is to satisfy
7099 // some existing assertions. In general, it may be a
7100 // good idea to abort immediately and complete the marking
7101 // from the grey objects at a later time.
7102 while (!_mark_stack->isEmpty()) {
7103 oop new_oop = _mark_stack->pop();
7104 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
7105 assert(_bit_map->isMarked((HeapWord*)new_oop),
7106 "only grey objects on this stack");
7107 // iterate over the oops in this oop, marking and pushing
7108 // the ones in CMS heap (i.e. in _span).
7109 new_oop->oop_iterate(_scanning_closure);
7110 // check if it's time to yield
7111 do_yield_check();
7112 }
7113 unsigned int after_count =
7114 GenCollectedHeap::heap()->total_collections();
7115 bool abort = (_before_count != after_count) ||
7116 _collector->should_abort_preclean();
7117 return abort ? 0 : size;
7118 }
7119
7120 void SurvivorSpacePrecleanClosure::do_yield_work() {
7121 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7122 "CMS thread should hold CMS token");
7123 assert_lock_strong(_bit_map->lock());
7124 // Relinquish the bit map lock
7125 _bit_map->lock()->unlock();
7126 ConcurrentMarkSweepThread::desynchronize(true);
7127 ConcurrentMarkSweepThread::acknowledge_yield_request();
7128 _collector->stopTimer();
7129 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7130 if (PrintCMSStatistics != 0) {
7131 _collector->incrementYields();
7132 }
7133 _collector->icms_wait();
7134
7135 // See the comment in coordinator_yield()
7136 for (unsigned i = 0; i < CMSYieldSleepCount &&
7137 ConcurrentMarkSweepThread::should_yield() &&
7138 !CMSCollector::foregroundGCIsActive(); ++i) {
7139 os::sleep(Thread::current(), 1, false);
7140 ConcurrentMarkSweepThread::acknowledge_yield_request();
7141 }
7142
7143 ConcurrentMarkSweepThread::synchronize(true);
7144 _bit_map->lock()->lock_without_safepoint_check();
7145 _collector->startTimer();
7146 }
7147
7148 // This closure is used to rescan the marked objects on the dirty cards
7149 // in the mod union table and the card table proper. In the parallel
7150 // case, although the bitMap is shared, we do a single read so the
7151 // isMarked() query is "safe".
7152 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
7153 // Ignore mark word because we are running concurrent with mutators
7154 assert(p->is_oop_or_null(true), "expected an oop or null");
7155 HeapWord* addr = (HeapWord*)p;
7156 assert(_span.contains(addr), "we are scanning the CMS generation");
7157 bool is_obj_array = false;
7158 #ifdef ASSERT
7159 if (!_parallel) {
7160 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
7161 assert(_collector->overflow_list_is_empty(),
7162 "overflow list should be empty");
7163
7164 }
7165 #endif // ASSERT
7166 if (_bit_map->isMarked(addr)) {
7167 // Obj arrays are precisely marked, non-arrays are not;
7168 // so we scan objArrays precisely and non-arrays in their
7169 // entirety.
7170 if (p->is_objArray()) {
7171 is_obj_array = true;
7172 if (_parallel) {
7173 p->oop_iterate(_par_scan_closure, mr);
7174 } else {
7175 p->oop_iterate(_scan_closure, mr);
7176 }
7177 } else {
7178 if (_parallel) {
7179 p->oop_iterate(_par_scan_closure);
7180 } else {
7181 p->oop_iterate(_scan_closure);
7182 }
7183 }
7184 }
7185 #ifdef ASSERT
7186 if (!_parallel) {
7187 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
7188 assert(_collector->overflow_list_is_empty(),
7189 "overflow list should be empty");
7190
7191 }
7192 #endif // ASSERT
7193 return is_obj_array;
7194 }
7195
7196 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
7197 MemRegion span,
7198 CMSBitMap* bitMap, CMSMarkStack* markStack,
7199 bool should_yield, bool verifying):
7200 _collector(collector),
7201 _span(span),
7202 _bitMap(bitMap),
7203 _mut(&collector->_modUnionTable),
7204 _markStack(markStack),
7205 _yield(should_yield),
7206 _skipBits(0)
7207 {
7208 assert(_markStack->isEmpty(), "stack should be empty");
7209 _finger = _bitMap->startWord();
7210 _threshold = _finger;
7211 assert(_collector->_restart_addr == NULL, "Sanity check");
7212 assert(_span.contains(_finger), "Out of bounds _finger?");
7213 DEBUG_ONLY(_verifying = verifying;)
7214 }
7215
7216 void MarkFromRootsClosure::reset(HeapWord* addr) {
7217 assert(_markStack->isEmpty(), "would cause duplicates on stack");
7218 assert(_span.contains(addr), "Out of bounds _finger?");
7219 _finger = addr;
7220 _threshold = (HeapWord*)round_to(
7221 (intptr_t)_finger, CardTableModRefBS::card_size);
7222 }
7223
7224 // Should revisit to see if this should be restructured for
7225 // greater efficiency.
7226 bool MarkFromRootsClosure::do_bit(size_t offset) {
7227 if (_skipBits > 0) {
7228 _skipBits--;
7229 return true;
7230 }
7231 // convert offset into a HeapWord*
7232 HeapWord* addr = _bitMap->startWord() + offset;
7233 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
7234 "address out of range");
7235 assert(_bitMap->isMarked(addr), "tautology");
7236 if (_bitMap->isMarked(addr+1)) {
7237 // this is an allocated but not yet initialized object
7238 assert(_skipBits == 0, "tautology");
7239 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
7240 oop p = oop(addr);
7241 if (p->klass_or_null() == NULL) {
7242 DEBUG_ONLY(if (!_verifying) {)
7243 // We re-dirty the cards on which this object lies and increase
7244 // the _threshold so that we'll come back to scan this object
7245 // during the preclean or remark phase. (CMSCleanOnEnter)
7246 if (CMSCleanOnEnter) {
7247 size_t sz = _collector->block_size_using_printezis_bits(addr);
7248 HeapWord* end_card_addr = (HeapWord*)round_to(
7249 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7250 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7251 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7252 // Bump _threshold to end_card_addr; note that
7253 // _threshold cannot possibly exceed end_card_addr, anyhow.
7254 // This prevents future clearing of the card as the scan proceeds
7255 // to the right.
7256 assert(_threshold <= end_card_addr,
7257 "Because we are just scanning into this object");
7258 if (_threshold < end_card_addr) {
7259 _threshold = end_card_addr;
7260 }
7261 if (p->klass_or_null() != NULL) {
7262 // Redirty the range of cards...
7263 _mut->mark_range(redirty_range);
7264 } // ...else the setting of klass will dirty the card anyway.
7265 }
7266 DEBUG_ONLY(})
7267 return true;
7268 }
7269 }
7270 scanOopsInOop(addr);
7271 return true;
7272 }
7273
7274 // We take a break if we've been at this for a while,
7275 // so as to avoid monopolizing the locks involved.
7276 void MarkFromRootsClosure::do_yield_work() {
7277 // First give up the locks, then yield, then re-lock
7278 // We should probably use a constructor/destructor idiom to
7279 // do this unlock/lock or modify the MutexUnlocker class to
7280 // serve our purpose. XXX
7281 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7282 "CMS thread should hold CMS token");
7283 assert_lock_strong(_bitMap->lock());
7284 _bitMap->lock()->unlock();
7285 ConcurrentMarkSweepThread::desynchronize(true);
7286 ConcurrentMarkSweepThread::acknowledge_yield_request();
7287 _collector->stopTimer();
7288 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7289 if (PrintCMSStatistics != 0) {
7290 _collector->incrementYields();
7291 }
7292 _collector->icms_wait();
7293
7294 // See the comment in coordinator_yield()
7295 for (unsigned i = 0; i < CMSYieldSleepCount &&
7296 ConcurrentMarkSweepThread::should_yield() &&
7297 !CMSCollector::foregroundGCIsActive(); ++i) {
7298 os::sleep(Thread::current(), 1, false);
7299 ConcurrentMarkSweepThread::acknowledge_yield_request();
7300 }
7301
7302 ConcurrentMarkSweepThread::synchronize(true);
7303 _bitMap->lock()->lock_without_safepoint_check();
7304 _collector->startTimer();
7305 }
7306
7307 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
7308 assert(_bitMap->isMarked(ptr), "expected bit to be set");
7309 assert(_markStack->isEmpty(),
7310 "should drain stack to limit stack usage");
7311 // convert ptr to an oop preparatory to scanning
7312 oop obj = oop(ptr);
7313 // Ignore mark word in verification below, since we
7314 // may be running concurrent with mutators.
7315 assert(obj->is_oop(true), "should be an oop");
7316 assert(_finger <= ptr, "_finger runneth ahead");
7317 // advance the finger to right end of this object
7318 _finger = ptr + obj->size();
7319 assert(_finger > ptr, "we just incremented it above");
7320 // On large heaps, it may take us some time to get through
7321 // the marking phase (especially if running iCMS). During
7322 // this time it's possible that a lot of mutations have
7323 // accumulated in the card table and the mod union table --
7324 // these mutation records are redundant until we have
7325 // actually traced into the corresponding card.
7326 // Here, we check whether advancing the finger would make
7327 // us cross into a new card, and if so clear corresponding
7328 // cards in the MUT (preclean them in the card-table in the
7329 // future).
7330
7331 DEBUG_ONLY(if (!_verifying) {)
7332 // The clean-on-enter optimization is disabled by default,
7333 // until we fix 6178663.
7334 if (CMSCleanOnEnter && (_finger > _threshold)) {
7335 // [_threshold, _finger) represents the interval
7336 // of cards to be cleared in MUT (or precleaned in card table).
7337 // The set of cards to be cleared is all those that overlap
7338 // with the interval [_threshold, _finger); note that
7339 // _threshold is always kept card-aligned but _finger isn't
7340 // always card-aligned.
7341 HeapWord* old_threshold = _threshold;
7342 assert(old_threshold == (HeapWord*)round_to(
7343 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7344 "_threshold should always be card-aligned");
7345 _threshold = (HeapWord*)round_to(
7346 (intptr_t)_finger, CardTableModRefBS::card_size);
7347 MemRegion mr(old_threshold, _threshold);
7348 assert(!mr.is_empty(), "Control point invariant");
7349 assert(_span.contains(mr), "Should clear within span");
7350 _mut->clear_range(mr);
7351 }
7352 DEBUG_ONLY(})
7353 // Note: the finger doesn't advance while we drain
7354 // the stack below.
7355 PushOrMarkClosure pushOrMarkClosure(_collector,
7356 _span, _bitMap, _markStack,
7357 _finger, this);
7358 bool res = _markStack->push(obj);
7359 assert(res, "Empty non-zero size stack should have space for single push");
7360 while (!_markStack->isEmpty()) {
7361 oop new_oop = _markStack->pop();
7362 // Skip verifying header mark word below because we are
7363 // running concurrent with mutators.
7364 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7365 // now scan this oop's oops
7366 new_oop->oop_iterate(&pushOrMarkClosure);
7367 do_yield_check();
7368 }
7369 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
7370 }
7371
7372 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
7373 CMSCollector* collector, MemRegion span,
7374 CMSBitMap* bit_map,
7375 OopTaskQueue* work_queue,
7376 CMSMarkStack* overflow_stack,
7377 bool should_yield):
7378 _collector(collector),
7379 _whole_span(collector->_span),
7380 _span(span),
7381 _bit_map(bit_map),
7382 _mut(&collector->_modUnionTable),
7383 _work_queue(work_queue),
7384 _overflow_stack(overflow_stack),
7385 _yield(should_yield),
7386 _skip_bits(0),
7387 _task(task)
7388 {
7389 assert(_work_queue->size() == 0, "work_queue should be empty");
7390 _finger = span.start();
7391 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
7392 assert(_span.contains(_finger), "Out of bounds _finger?");
7393 }
7394
7395 // Should revisit to see if this should be restructured for
7396 // greater efficiency.
7397 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
7398 if (_skip_bits > 0) {
7399 _skip_bits--;
7400 return true;
7401 }
7402 // convert offset into a HeapWord*
7403 HeapWord* addr = _bit_map->startWord() + offset;
7404 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
7405 "address out of range");
7406 assert(_bit_map->isMarked(addr), "tautology");
7407 if (_bit_map->isMarked(addr+1)) {
7408 // this is an allocated object that might not yet be initialized
7409 assert(_skip_bits == 0, "tautology");
7410 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
7411 oop p = oop(addr);
7412 if (p->klass_or_null() == NULL) {
7413 // in the case of Clean-on-Enter optimization, redirty card
7414 // and avoid clearing card by increasing the threshold.
7415 return true;
7416 }
7417 }
7418 scan_oops_in_oop(addr);
7419 return true;
7420 }
7421
7422 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
7423 assert(_bit_map->isMarked(ptr), "expected bit to be set");
7424 // Should we assert that our work queue is empty or
7425 // below some drain limit?
7426 assert(_work_queue->size() == 0,
7427 "should drain stack to limit stack usage");
7428 // convert ptr to an oop preparatory to scanning
7429 oop obj = oop(ptr);
7430 // Ignore mark word in verification below, since we
7431 // may be running concurrent with mutators.
7432 assert(obj->is_oop(true), "should be an oop");
7433 assert(_finger <= ptr, "_finger runneth ahead");
7434 // advance the finger to right end of this object
7435 _finger = ptr + obj->size();
7436 assert(_finger > ptr, "we just incremented it above");
7437 // On large heaps, it may take us some time to get through
7438 // the marking phase (especially if running iCMS). During
7439 // this time it's possible that a lot of mutations have
7440 // accumulated in the card table and the mod union table --
7441 // these mutation records are redundant until we have
7442 // actually traced into the corresponding card.
7443 // Here, we check whether advancing the finger would make
7444 // us cross into a new card, and if so clear corresponding
7445 // cards in the MUT (preclean them in the card-table in the
7446 // future).
7447
7448 // The clean-on-enter optimization is disabled by default,
7449 // until we fix 6178663.
7450 if (CMSCleanOnEnter && (_finger > _threshold)) {
7451 // [_threshold, _finger) represents the interval
7452 // of cards to be cleared in MUT (or precleaned in card table).
7453 // The set of cards to be cleared is all those that overlap
7454 // with the interval [_threshold, _finger); note that
7455 // _threshold is always kept card-aligned but _finger isn't
7456 // always card-aligned.
7457 HeapWord* old_threshold = _threshold;
7458 assert(old_threshold == (HeapWord*)round_to(
7459 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7460 "_threshold should always be card-aligned");
7461 _threshold = (HeapWord*)round_to(
7462 (intptr_t)_finger, CardTableModRefBS::card_size);
7463 MemRegion mr(old_threshold, _threshold);
7464 assert(!mr.is_empty(), "Control point invariant");
7465 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7466 _mut->clear_range(mr);
7467 }
7468
7469 // Note: the local finger doesn't advance while we drain
7470 // the stack below, but the global finger sure can and will.
7471 HeapWord** gfa = _task->global_finger_addr();
7472 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7473 _span, _bit_map,
7474 _work_queue,
7475 _overflow_stack,
7476 _finger,
7477 gfa, this);
7478 bool res = _work_queue->push(obj); // overflow could occur here
7479 assert(res, "Will hold once we use workqueues");
7480 while (true) {
7481 oop new_oop;
7482 if (!_work_queue->pop_local(new_oop)) {
7483 // We emptied our work_queue; check if there's stuff that can
7484 // be gotten from the overflow stack.
7485 if (CMSConcMarkingTask::get_work_from_overflow_stack(
7486 _overflow_stack, _work_queue)) {
7487 do_yield_check();
7488 continue;
7489 } else { // done
7490 break;
7491 }
7492 }
7493 // Skip verifying header mark word below because we are
7494 // running concurrent with mutators.
7495 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7496 // now scan this oop's oops
7497 new_oop->oop_iterate(&pushOrMarkClosure);
7498 do_yield_check();
7499 }
7500 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7501 }
7502
7503 // Yield in response to a request from VM Thread or
7504 // from mutators.
7505 void Par_MarkFromRootsClosure::do_yield_work() {
7506 assert(_task != NULL, "sanity");
7507 _task->yield();
7508 }
7509
7510 // A variant of the above used for verifying CMS marking work.
7511 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7512 MemRegion span,
7513 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7514 CMSMarkStack* mark_stack):
7515 _collector(collector),
7516 _span(span),
7517 _verification_bm(verification_bm),
7518 _cms_bm(cms_bm),
7519 _mark_stack(mark_stack),
7520 _pam_verify_closure(collector, span, verification_bm, cms_bm,
7521 mark_stack)
7522 {
7523 assert(_mark_stack->isEmpty(), "stack should be empty");
7524 _finger = _verification_bm->startWord();
7525 assert(_collector->_restart_addr == NULL, "Sanity check");
7526 assert(_span.contains(_finger), "Out of bounds _finger?");
7527 }
7528
7529 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7530 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7531 assert(_span.contains(addr), "Out of bounds _finger?");
7532 _finger = addr;
7533 }
7534
7535 // Should revisit to see if this should be restructured for
7536 // greater efficiency.
7537 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7538 // convert offset into a HeapWord*
7539 HeapWord* addr = _verification_bm->startWord() + offset;
7540 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7541 "address out of range");
7542 assert(_verification_bm->isMarked(addr), "tautology");
7543 assert(_cms_bm->isMarked(addr), "tautology");
7544
7545 assert(_mark_stack->isEmpty(),
7546 "should drain stack to limit stack usage");
7547 // convert addr to an oop preparatory to scanning
7548 oop obj = oop(addr);
7549 assert(obj->is_oop(), "should be an oop");
7550 assert(_finger <= addr, "_finger runneth ahead");
7551 // advance the finger to right end of this object
7552 _finger = addr + obj->size();
7553 assert(_finger > addr, "we just incremented it above");
7554 // Note: the finger doesn't advance while we drain
7555 // the stack below.
7556 bool res = _mark_stack->push(obj);
7557 assert(res, "Empty non-zero size stack should have space for single push");
7558 while (!_mark_stack->isEmpty()) {
7559 oop new_oop = _mark_stack->pop();
7560 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7561 // now scan this oop's oops
7562 new_oop->oop_iterate(&_pam_verify_closure);
7563 }
7564 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7565 return true;
7566 }
7567
7568 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7569 CMSCollector* collector, MemRegion span,
7570 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7571 CMSMarkStack* mark_stack):
7572 CMSOopClosure(collector->ref_processor()),
7573 _collector(collector),
7574 _span(span),
7575 _verification_bm(verification_bm),
7576 _cms_bm(cms_bm),
7577 _mark_stack(mark_stack)
7578 { }
7579
7580 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7581 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7582
7583 // Upon stack overflow, we discard (part of) the stack,
7584 // remembering the least address amongst those discarded
7585 // in CMSCollector's _restart_address.
7586 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7587 // Remember the least grey address discarded
7588 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7589 _collector->lower_restart_addr(ra);
7590 _mark_stack->reset(); // discard stack contents
7591 _mark_stack->expand(); // expand the stack if possible
7592 }
7593
7594 void PushAndMarkVerifyClosure::do_oop(oop obj) {
7595 assert(obj->is_oop_or_null(), "expected an oop or NULL");
7596 HeapWord* addr = (HeapWord*)obj;
7597 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7598 // Oop lies in _span and isn't yet grey or black
7599 _verification_bm->mark(addr); // now grey
7600 if (!_cms_bm->isMarked(addr)) {
7601 oop(addr)->print();
7602 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
7603 addr);
7604 fatal("... aborting");
7605 }
7606
7607 if (!_mark_stack->push(obj)) { // stack overflow
7608 if (PrintCMSStatistics != 0) {
7609 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7610 SIZE_FORMAT, _mark_stack->capacity());
7611 }
7612 assert(_mark_stack->isFull(), "Else push should have succeeded");
7613 handle_stack_overflow(addr);
7614 }
7615 // anything including and to the right of _finger
7616 // will be scanned as we iterate over the remainder of the
7617 // bit map
7618 }
7619 }
7620
7621 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7622 MemRegion span,
7623 CMSBitMap* bitMap, CMSMarkStack* markStack,
7624 HeapWord* finger, MarkFromRootsClosure* parent) :
7625 CMSOopClosure(collector->ref_processor()),
7626 _collector(collector),
7627 _span(span),
7628 _bitMap(bitMap),
7629 _markStack(markStack),
7630 _finger(finger),
7631 _parent(parent)
7632 { }
7633
7634 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7635 MemRegion span,
7636 CMSBitMap* bit_map,
7637 OopTaskQueue* work_queue,
7638 CMSMarkStack* overflow_stack,
7639 HeapWord* finger,
7640 HeapWord** global_finger_addr,
7641 Par_MarkFromRootsClosure* parent) :
7642 CMSOopClosure(collector->ref_processor()),
7643 _collector(collector),
7644 _whole_span(collector->_span),
7645 _span(span),
7646 _bit_map(bit_map),
7647 _work_queue(work_queue),
7648 _overflow_stack(overflow_stack),
7649 _finger(finger),
7650 _global_finger_addr(global_finger_addr),
7651 _parent(parent)
7652 { }
7653
7654 // Assumes thread-safe access by callers, who are
7655 // responsible for mutual exclusion.
7656 void CMSCollector::lower_restart_addr(HeapWord* low) {
7657 assert(_span.contains(low), "Out of bounds addr");
7658 if (_restart_addr == NULL) {
7659 _restart_addr = low;
7660 } else {
7661 _restart_addr = MIN2(_restart_addr, low);
7662 }
7663 }
7664
7665 // Upon stack overflow, we discard (part of) the stack,
7666 // remembering the least address amongst those discarded
7667 // in CMSCollector's _restart_address.
7668 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7669 // Remember the least grey address discarded
7670 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7671 _collector->lower_restart_addr(ra);
7672 _markStack->reset(); // discard stack contents
7673 _markStack->expand(); // expand the stack if possible
7674 }
7675
7676 // Upon stack overflow, we discard (part of) the stack,
7677 // remembering the least address amongst those discarded
7678 // in CMSCollector's _restart_address.
7679 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7680 // We need to do this under a mutex to prevent other
7681 // workers from interfering with the work done below.
7682 MutexLockerEx ml(_overflow_stack->par_lock(),
7683 Mutex::_no_safepoint_check_flag);
7684 // Remember the least grey address discarded
7685 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7686 _collector->lower_restart_addr(ra);
7687 _overflow_stack->reset(); // discard stack contents
7688 _overflow_stack->expand(); // expand the stack if possible
7689 }
7690
7691 void CMKlassClosure::do_klass(Klass* k) {
7692 assert(_oop_closure != NULL, "Not initialized?");
7693 k->oops_do(_oop_closure);
7694 }
7695
7696 void PushOrMarkClosure::do_oop(oop obj) {
7697 // Ignore mark word because we are running concurrent with mutators.
7698 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7699 HeapWord* addr = (HeapWord*)obj;
7700 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7701 // Oop lies in _span and isn't yet grey or black
7702 _bitMap->mark(addr); // now grey
7703 if (addr < _finger) {
7704 // the bit map iteration has already either passed, or
7705 // sampled, this bit in the bit map; we'll need to
7706 // use the marking stack to scan this oop's oops.
7707 bool simulate_overflow = false;
7708 NOT_PRODUCT(
7709 if (CMSMarkStackOverflowALot &&
7710 _collector->simulate_overflow()) {
7711 // simulate a stack overflow
7712 simulate_overflow = true;
7713 }
7714 )
7715 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7716 if (PrintCMSStatistics != 0) {
7717 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7718 SIZE_FORMAT, _markStack->capacity());
7719 }
7720 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7721 handle_stack_overflow(addr);
7722 }
7723 }
7724 // anything including and to the right of _finger
7725 // will be scanned as we iterate over the remainder of the
7726 // bit map
7727 do_yield_check();
7728 }
7729 }
7730
7731 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7732 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7733
7734 void Par_PushOrMarkClosure::do_oop(oop obj) {
7735 // Ignore mark word because we are running concurrent with mutators.
7736 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7737 HeapWord* addr = (HeapWord*)obj;
7738 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7739 // Oop lies in _span and isn't yet grey or black
7740 // We read the global_finger (volatile read) strictly after marking oop
7741 bool res = _bit_map->par_mark(addr); // now grey
7742 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7743 // Should we push this marked oop on our stack?
7744 // -- if someone else marked it, nothing to do
7745 // -- if target oop is above global finger nothing to do
7746 // -- if target oop is in chunk and above local finger
7747 // then nothing to do
7748 // -- else push on work queue
7749 if ( !res // someone else marked it, they will deal with it
7750 || (addr >= *gfa) // will be scanned in a later task
7751 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7752 return;
7753 }
7754 // the bit map iteration has already either passed, or
7755 // sampled, this bit in the bit map; we'll need to
7756 // use the marking stack to scan this oop's oops.
7757 bool simulate_overflow = false;
7758 NOT_PRODUCT(
7759 if (CMSMarkStackOverflowALot &&
7760 _collector->simulate_overflow()) {
7761 // simulate a stack overflow
7762 simulate_overflow = true;
7763 }
7764 )
7765 if (simulate_overflow ||
7766 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7767 // stack overflow
7768 if (PrintCMSStatistics != 0) {
7769 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7770 SIZE_FORMAT, _overflow_stack->capacity());
7771 }
7772 // We cannot assert that the overflow stack is full because
7773 // it may have been emptied since.
7774 assert(simulate_overflow ||
7775 _work_queue->size() == _work_queue->max_elems(),
7776 "Else push should have succeeded");
7777 handle_stack_overflow(addr);
7778 }
7779 do_yield_check();
7780 }
7781 }
7782
7783 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7784 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7785
7786 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7787 MemRegion span,
7788 ReferenceProcessor* rp,
7789 CMSBitMap* bit_map,
7790 CMSBitMap* mod_union_table,
7791 CMSMarkStack* mark_stack,
7792 bool concurrent_precleaning):
7793 CMSOopClosure(rp),
7794 _collector(collector),
7795 _span(span),
7796 _bit_map(bit_map),
7797 _mod_union_table(mod_union_table),
7798 _mark_stack(mark_stack),
7799 _concurrent_precleaning(concurrent_precleaning)
7800 {
7801 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7802 }
7803
7804 // Grey object rescan during pre-cleaning and second checkpoint phases --
7805 // the non-parallel version (the parallel version appears further below.)
7806 void PushAndMarkClosure::do_oop(oop obj) {
7807 // Ignore mark word verification. If during concurrent precleaning,
7808 // the object monitor may be locked. If during the checkpoint
7809 // phases, the object may already have been reached by a different
7810 // path and may be at the end of the global overflow list (so
7811 // the mark word may be NULL).
7812 assert(obj->is_oop_or_null(true /* ignore mark word */),
7813 "expected an oop or NULL");
7814 HeapWord* addr = (HeapWord*)obj;
7815 // Check if oop points into the CMS generation
7816 // and is not marked
7817 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7818 // a white object ...
7819 _bit_map->mark(addr); // ... now grey
7820 // push on the marking stack (grey set)
7821 bool simulate_overflow = false;
7822 NOT_PRODUCT(
7823 if (CMSMarkStackOverflowALot &&
7824 _collector->simulate_overflow()) {
7825 // simulate a stack overflow
7826 simulate_overflow = true;
7827 }
7828 )
7829 if (simulate_overflow || !_mark_stack->push(obj)) {
7830 if (_concurrent_precleaning) {
7831 // During precleaning we can just dirty the appropriate card(s)
7832 // in the mod union table, thus ensuring that the object remains
7833 // in the grey set and continue. In the case of object arrays
7834 // we need to dirty all of the cards that the object spans,
7835 // since the rescan of object arrays will be limited to the
7836 // dirty cards.
7837 // Note that no one can be intefering with us in this action
7838 // of dirtying the mod union table, so no locking or atomics
7839 // are required.
7840 if (obj->is_objArray()) {
7841 size_t sz = obj->size();
7842 HeapWord* end_card_addr = (HeapWord*)round_to(
7843 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7844 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7845 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7846 _mod_union_table->mark_range(redirty_range);
7847 } else {
7848 _mod_union_table->mark(addr);
7849 }
7850 _collector->_ser_pmc_preclean_ovflw++;
7851 } else {
7852 // During the remark phase, we need to remember this oop
7853 // in the overflow list.
7854 _collector->push_on_overflow_list(obj);
7855 _collector->_ser_pmc_remark_ovflw++;
7856 }
7857 }
7858 }
7859 }
7860
7861 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7862 MemRegion span,
7863 ReferenceProcessor* rp,
7864 CMSBitMap* bit_map,
7865 OopTaskQueue* work_queue):
7866 CMSOopClosure(rp),
7867 _collector(collector),
7868 _span(span),
7869 _bit_map(bit_map),
7870 _work_queue(work_queue)
7871 {
7872 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7873 }
7874
7875 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7876 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7877
7878 // Grey object rescan during second checkpoint phase --
7879 // the parallel version.
7880 void Par_PushAndMarkClosure::do_oop(oop obj) {
7881 // In the assert below, we ignore the mark word because
7882 // this oop may point to an already visited object that is
7883 // on the overflow stack (in which case the mark word has
7884 // been hijacked for chaining into the overflow stack --
7885 // if this is the last object in the overflow stack then
7886 // its mark word will be NULL). Because this object may
7887 // have been subsequently popped off the global overflow
7888 // stack, and the mark word possibly restored to the prototypical
7889 // value, by the time we get to examined this failing assert in
7890 // the debugger, is_oop_or_null(false) may subsequently start
7891 // to hold.
7892 assert(obj->is_oop_or_null(true),
7893 "expected an oop or NULL");
7894 HeapWord* addr = (HeapWord*)obj;
7895 // Check if oop points into the CMS generation
7896 // and is not marked
7897 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7898 // a white object ...
7899 // If we manage to "claim" the object, by being the
7900 // first thread to mark it, then we push it on our
7901 // marking stack
7902 if (_bit_map->par_mark(addr)) { // ... now grey
7903 // push on work queue (grey set)
7904 bool simulate_overflow = false;
7905 NOT_PRODUCT(
7906 if (CMSMarkStackOverflowALot &&
7907 _collector->par_simulate_overflow()) {
7908 // simulate a stack overflow
7909 simulate_overflow = true;
7910 }
7911 )
7912 if (simulate_overflow || !_work_queue->push(obj)) {
7913 _collector->par_push_on_overflow_list(obj);
7914 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7915 }
7916 } // Else, some other thread got there first
7917 }
7918 }
7919
7920 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7921 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7922
7923 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7924 Mutex* bml = _collector->bitMapLock();
7925 assert_lock_strong(bml);
7926 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7927 "CMS thread should hold CMS token");
7928
7929 bml->unlock();
7930 ConcurrentMarkSweepThread::desynchronize(true);
7931
7932 ConcurrentMarkSweepThread::acknowledge_yield_request();
7933
7934 _collector->stopTimer();
7935 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7936 if (PrintCMSStatistics != 0) {
7937 _collector->incrementYields();
7938 }
7939 _collector->icms_wait();
7940
7941 // See the comment in coordinator_yield()
7942 for (unsigned i = 0; i < CMSYieldSleepCount &&
7943 ConcurrentMarkSweepThread::should_yield() &&
7944 !CMSCollector::foregroundGCIsActive(); ++i) {
7945 os::sleep(Thread::current(), 1, false);
7946 ConcurrentMarkSweepThread::acknowledge_yield_request();
7947 }
7948
7949 ConcurrentMarkSweepThread::synchronize(true);
7950 bml->lock();
7951
7952 _collector->startTimer();
7953 }
7954
7955 bool CMSPrecleanRefsYieldClosure::should_return() {
7956 if (ConcurrentMarkSweepThread::should_yield()) {
7957 do_yield_work();
7958 }
7959 return _collector->foregroundGCIsActive();
7960 }
7961
7962 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7963 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7964 "mr should be aligned to start at a card boundary");
7965 // We'd like to assert:
7966 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7967 // "mr should be a range of cards");
7968 // However, that would be too strong in one case -- the last
7969 // partition ends at _unallocated_block which, in general, can be
7970 // an arbitrary boundary, not necessarily card aligned.
7971 if (PrintCMSStatistics != 0) {
7972 _num_dirty_cards +=
7973 mr.word_size()/CardTableModRefBS::card_size_in_words;
7974 }
7975 _space->object_iterate_mem(mr, &_scan_cl);
7976 }
7977
7978 SweepClosure::SweepClosure(CMSCollector* collector,
7979 ConcurrentMarkSweepGeneration* g,
7980 CMSBitMap* bitMap, bool should_yield) :
7981 _collector(collector),
7982 _g(g),
7983 _sp(g->cmsSpace()),
7984 _limit(_sp->sweep_limit()),
7985 _freelistLock(_sp->freelistLock()),
7986 _bitMap(bitMap),
7987 _yield(should_yield),
7988 _inFreeRange(false), // No free range at beginning of sweep
7989 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7990 _lastFreeRangeCoalesced(false),
7991 _freeFinger(g->used_region().start())
7992 {
7993 NOT_PRODUCT(
7994 _numObjectsFreed = 0;
7995 _numWordsFreed = 0;
7996 _numObjectsLive = 0;
7997 _numWordsLive = 0;
7998 _numObjectsAlreadyFree = 0;
7999 _numWordsAlreadyFree = 0;
8000 _last_fc = NULL;
8001
8002 _sp->initializeIndexedFreeListArrayReturnedBytes();
8003 _sp->dictionary()->initialize_dict_returned_bytes();
8004 )
8005 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8006 "sweep _limit out of bounds");
8007 if (CMSTraceSweeper) {
8008 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
8009 _limit);
8010 }
8011 }
8012
8013 void SweepClosure::print_on(outputStream* st) const {
8014 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
8015 _sp->bottom(), _sp->end());
8016 tty->print_cr("_limit = " PTR_FORMAT, _limit);
8017 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger);
8018 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);)
8019 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
8020 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
8021 }
8022
8023 #ifndef PRODUCT
8024 // Assertion checking only: no useful work in product mode --
8025 // however, if any of the flags below become product flags,
8026 // you may need to review this code to see if it needs to be
8027 // enabled in product mode.
8028 SweepClosure::~SweepClosure() {
8029 assert_lock_strong(_freelistLock);
8030 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8031 "sweep _limit out of bounds");
8032 if (inFreeRange()) {
8033 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
8034 print();
8035 ShouldNotReachHere();
8036 }
8037 if (Verbose && PrintGC) {
8038 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes",
8039 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
8040 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
8041 SIZE_FORMAT" bytes "
8042 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
8043 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
8044 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
8045 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
8046 * sizeof(HeapWord);
8047 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
8048
8049 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
8050 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
8051 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
8052 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
8053 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes);
8054 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
8055 indexListReturnedBytes);
8056 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
8057 dict_returned_bytes);
8058 }
8059 }
8060 if (CMSTraceSweeper) {
8061 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
8062 _limit);
8063 }
8064 }
8065 #endif // PRODUCT
8066
8067 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
8068 bool freeRangeInFreeLists) {
8069 if (CMSTraceSweeper) {
8070 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n",
8071 freeFinger, freeRangeInFreeLists);
8072 }
8073 assert(!inFreeRange(), "Trampling existing free range");
8074 set_inFreeRange(true);
8075 set_lastFreeRangeCoalesced(false);
8076
8077 set_freeFinger(freeFinger);
8078 set_freeRangeInFreeLists(freeRangeInFreeLists);
8079 if (CMSTestInFreeList) {
8080 if (freeRangeInFreeLists) {
8081 FreeChunk* fc = (FreeChunk*) freeFinger;
8082 assert(fc->is_free(), "A chunk on the free list should be free.");
8083 assert(fc->size() > 0, "Free range should have a size");
8084 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
8085 }
8086 }
8087 }
8088
8089 // Note that the sweeper runs concurrently with mutators. Thus,
8090 // it is possible for direct allocation in this generation to happen
8091 // in the middle of the sweep. Note that the sweeper also coalesces
8092 // contiguous free blocks. Thus, unless the sweeper and the allocator
8093 // synchronize appropriately freshly allocated blocks may get swept up.
8094 // This is accomplished by the sweeper locking the free lists while
8095 // it is sweeping. Thus blocks that are determined to be free are
8096 // indeed free. There is however one additional complication:
8097 // blocks that have been allocated since the final checkpoint and
8098 // mark, will not have been marked and so would be treated as
8099 // unreachable and swept up. To prevent this, the allocator marks
8100 // the bit map when allocating during the sweep phase. This leads,
8101 // however, to a further complication -- objects may have been allocated
8102 // but not yet initialized -- in the sense that the header isn't yet
8103 // installed. The sweeper can not then determine the size of the block
8104 // in order to skip over it. To deal with this case, we use a technique
8105 // (due to Printezis) to encode such uninitialized block sizes in the
8106 // bit map. Since the bit map uses a bit per every HeapWord, but the
8107 // CMS generation has a minimum object size of 3 HeapWords, it follows
8108 // that "normal marks" won't be adjacent in the bit map (there will
8109 // always be at least two 0 bits between successive 1 bits). We make use
8110 // of these "unused" bits to represent uninitialized blocks -- the bit
8111 // corresponding to the start of the uninitialized object and the next
8112 // bit are both set. Finally, a 1 bit marks the end of the object that
8113 // started with the two consecutive 1 bits to indicate its potentially
8114 // uninitialized state.
8115
8116 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
8117 FreeChunk* fc = (FreeChunk*)addr;
8118 size_t res;
8119
8120 // Check if we are done sweeping. Below we check "addr >= _limit" rather
8121 // than "addr == _limit" because although _limit was a block boundary when
8122 // we started the sweep, it may no longer be one because heap expansion
8123 // may have caused us to coalesce the block ending at the address _limit
8124 // with a newly expanded chunk (this happens when _limit was set to the
8125 // previous _end of the space), so we may have stepped past _limit:
8126 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
8127 if (addr >= _limit) { // we have swept up to or past the limit: finish up
8128 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8129 "sweep _limit out of bounds");
8130 assert(addr < _sp->end(), "addr out of bounds");
8131 // Flush any free range we might be holding as a single
8132 // coalesced chunk to the appropriate free list.
8133 if (inFreeRange()) {
8134 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
8135 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger()));
8136 flush_cur_free_chunk(freeFinger(),
8137 pointer_delta(addr, freeFinger()));
8138 if (CMSTraceSweeper) {
8139 gclog_or_tty->print("Sweep: last chunk: ");
8140 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") "
8141 "[coalesced:"SIZE_FORMAT"]\n",
8142 freeFinger(), pointer_delta(addr, freeFinger()),
8143 lastFreeRangeCoalesced());
8144 }
8145 }
8146
8147 // help the iterator loop finish
8148 return pointer_delta(_sp->end(), addr);
8149 }
8150
8151 assert(addr < _limit, "sweep invariant");
8152 // check if we should yield
8153 do_yield_check(addr);
8154 if (fc->is_free()) {
8155 // Chunk that is already free
8156 res = fc->size();
8157 do_already_free_chunk(fc);
8158 debug_only(_sp->verifyFreeLists());
8159 // If we flush the chunk at hand in lookahead_and_flush()
8160 // and it's coalesced with a preceding chunk, then the
8161 // process of "mangling" the payload of the coalesced block
8162 // will cause erasure of the size information from the
8163 // (erstwhile) header of all the coalesced blocks but the
8164 // first, so the first disjunct in the assert will not hold
8165 // in that specific case (in which case the second disjunct
8166 // will hold).
8167 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
8168 "Otherwise the size info doesn't change at this step");
8169 NOT_PRODUCT(
8170 _numObjectsAlreadyFree++;
8171 _numWordsAlreadyFree += res;
8172 )
8173 NOT_PRODUCT(_last_fc = fc;)
8174 } else if (!_bitMap->isMarked(addr)) {
8175 // Chunk is fresh garbage
8176 res = do_garbage_chunk(fc);
8177 debug_only(_sp->verifyFreeLists());
8178 NOT_PRODUCT(
8179 _numObjectsFreed++;
8180 _numWordsFreed += res;
8181 )
8182 } else {
8183 // Chunk that is alive.
8184 res = do_live_chunk(fc);
8185 debug_only(_sp->verifyFreeLists());
8186 NOT_PRODUCT(
8187 _numObjectsLive++;
8188 _numWordsLive += res;
8189 )
8190 }
8191 return res;
8192 }
8193
8194 // For the smart allocation, record following
8195 // split deaths - a free chunk is removed from its free list because
8196 // it is being split into two or more chunks.
8197 // split birth - a free chunk is being added to its free list because
8198 // a larger free chunk has been split and resulted in this free chunk.
8199 // coal death - a free chunk is being removed from its free list because
8200 // it is being coalesced into a large free chunk.
8201 // coal birth - a free chunk is being added to its free list because
8202 // it was created when two or more free chunks where coalesced into
8203 // this free chunk.
8204 //
8205 // These statistics are used to determine the desired number of free
8206 // chunks of a given size. The desired number is chosen to be relative
8207 // to the end of a CMS sweep. The desired number at the end of a sweep
8208 // is the
8209 // count-at-end-of-previous-sweep (an amount that was enough)
8210 // - count-at-beginning-of-current-sweep (the excess)
8211 // + split-births (gains in this size during interval)
8212 // - split-deaths (demands on this size during interval)
8213 // where the interval is from the end of one sweep to the end of the
8214 // next.
8215 //
8216 // When sweeping the sweeper maintains an accumulated chunk which is
8217 // the chunk that is made up of chunks that have been coalesced. That
8218 // will be termed the left-hand chunk. A new chunk of garbage that
8219 // is being considered for coalescing will be referred to as the
8220 // right-hand chunk.
8221 //
8222 // When making a decision on whether to coalesce a right-hand chunk with
8223 // the current left-hand chunk, the current count vs. the desired count
8224 // of the left-hand chunk is considered. Also if the right-hand chunk
8225 // is near the large chunk at the end of the heap (see
8226 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
8227 // left-hand chunk is coalesced.
8228 //
8229 // When making a decision about whether to split a chunk, the desired count
8230 // vs. the current count of the candidate to be split is also considered.
8231 // If the candidate is underpopulated (currently fewer chunks than desired)
8232 // a chunk of an overpopulated (currently more chunks than desired) size may
8233 // be chosen. The "hint" associated with a free list, if non-null, points
8234 // to a free list which may be overpopulated.
8235 //
8236
8237 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
8238 const size_t size = fc->size();
8239 // Chunks that cannot be coalesced are not in the
8240 // free lists.
8241 if (CMSTestInFreeList && !fc->cantCoalesce()) {
8242 assert(_sp->verify_chunk_in_free_list(fc),
8243 "free chunk should be in free lists");
8244 }
8245 // a chunk that is already free, should not have been
8246 // marked in the bit map
8247 HeapWord* const addr = (HeapWord*) fc;
8248 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
8249 // Verify that the bit map has no bits marked between
8250 // addr and purported end of this block.
8251 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8252
8253 // Some chunks cannot be coalesced under any circumstances.
8254 // See the definition of cantCoalesce().
8255 if (!fc->cantCoalesce()) {
8256 // This chunk can potentially be coalesced.
8257 if (_sp->adaptive_freelists()) {
8258 // All the work is done in
8259 do_post_free_or_garbage_chunk(fc, size);
8260 } else { // Not adaptive free lists
8261 // this is a free chunk that can potentially be coalesced by the sweeper;
8262 if (!inFreeRange()) {
8263 // if the next chunk is a free block that can't be coalesced
8264 // it doesn't make sense to remove this chunk from the free lists
8265 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
8266 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
8267 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
8268 nextChunk->is_free() && // ... which is free...
8269 nextChunk->cantCoalesce()) { // ... but can't be coalesced
8270 // nothing to do
8271 } else {
8272 // Potentially the start of a new free range:
8273 // Don't eagerly remove it from the free lists.
8274 // No need to remove it if it will just be put
8275 // back again. (Also from a pragmatic point of view
8276 // if it is a free block in a region that is beyond
8277 // any allocated blocks, an assertion will fail)
8278 // Remember the start of a free run.
8279 initialize_free_range(addr, true);
8280 // end - can coalesce with next chunk
8281 }
8282 } else {
8283 // the midst of a free range, we are coalescing
8284 print_free_block_coalesced(fc);
8285 if (CMSTraceSweeper) {
8286 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size);
8287 }
8288 // remove it from the free lists
8289 _sp->removeFreeChunkFromFreeLists(fc);
8290 set_lastFreeRangeCoalesced(true);
8291 // If the chunk is being coalesced and the current free range is
8292 // in the free lists, remove the current free range so that it
8293 // will be returned to the free lists in its entirety - all
8294 // the coalesced pieces included.
8295 if (freeRangeInFreeLists()) {
8296 FreeChunk* ffc = (FreeChunk*) freeFinger();
8297 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8298 "Size of free range is inconsistent with chunk size.");
8299 if (CMSTestInFreeList) {
8300 assert(_sp->verify_chunk_in_free_list(ffc),
8301 "free range is not in free lists");
8302 }
8303 _sp->removeFreeChunkFromFreeLists(ffc);
8304 set_freeRangeInFreeLists(false);
8305 }
8306 }
8307 }
8308 // Note that if the chunk is not coalescable (the else arm
8309 // below), we unconditionally flush, without needing to do
8310 // a "lookahead," as we do below.
8311 if (inFreeRange()) lookahead_and_flush(fc, size);
8312 } else {
8313 // Code path common to both original and adaptive free lists.
8314
8315 // cant coalesce with previous block; this should be treated
8316 // as the end of a free run if any
8317 if (inFreeRange()) {
8318 // we kicked some butt; time to pick up the garbage
8319 assert(freeFinger() < addr, "freeFinger points too high");
8320 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8321 }
8322 // else, nothing to do, just continue
8323 }
8324 }
8325
8326 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
8327 // This is a chunk of garbage. It is not in any free list.
8328 // Add it to a free list or let it possibly be coalesced into
8329 // a larger chunk.
8330 HeapWord* const addr = (HeapWord*) fc;
8331 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8332
8333 if (_sp->adaptive_freelists()) {
8334 // Verify that the bit map has no bits marked between
8335 // addr and purported end of just dead object.
8336 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8337
8338 do_post_free_or_garbage_chunk(fc, size);
8339 } else {
8340 if (!inFreeRange()) {
8341 // start of a new free range
8342 assert(size > 0, "A free range should have a size");
8343 initialize_free_range(addr, false);
8344 } else {
8345 // this will be swept up when we hit the end of the
8346 // free range
8347 if (CMSTraceSweeper) {
8348 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size);
8349 }
8350 // If the chunk is being coalesced and the current free range is
8351 // in the free lists, remove the current free range so that it
8352 // will be returned to the free lists in its entirety - all
8353 // the coalesced pieces included.
8354 if (freeRangeInFreeLists()) {
8355 FreeChunk* ffc = (FreeChunk*)freeFinger();
8356 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8357 "Size of free range is inconsistent with chunk size.");
8358 if (CMSTestInFreeList) {
8359 assert(_sp->verify_chunk_in_free_list(ffc),
8360 "free range is not in free lists");
8361 }
8362 _sp->removeFreeChunkFromFreeLists(ffc);
8363 set_freeRangeInFreeLists(false);
8364 }
8365 set_lastFreeRangeCoalesced(true);
8366 }
8367 // this will be swept up when we hit the end of the free range
8368
8369 // Verify that the bit map has no bits marked between
8370 // addr and purported end of just dead object.
8371 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8372 }
8373 assert(_limit >= addr + size,
8374 "A freshly garbage chunk can't possibly straddle over _limit");
8375 if (inFreeRange()) lookahead_and_flush(fc, size);
8376 return size;
8377 }
8378
8379 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
8380 HeapWord* addr = (HeapWord*) fc;
8381 // The sweeper has just found a live object. Return any accumulated
8382 // left hand chunk to the free lists.
8383 if (inFreeRange()) {
8384 assert(freeFinger() < addr, "freeFinger points too high");
8385 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8386 }
8387
8388 // This object is live: we'd normally expect this to be
8389 // an oop, and like to assert the following:
8390 // assert(oop(addr)->is_oop(), "live block should be an oop");
8391 // However, as we commented above, this may be an object whose
8392 // header hasn't yet been initialized.
8393 size_t size;
8394 assert(_bitMap->isMarked(addr), "Tautology for this control point");
8395 if (_bitMap->isMarked(addr + 1)) {
8396 // Determine the size from the bit map, rather than trying to
8397 // compute it from the object header.
8398 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
8399 size = pointer_delta(nextOneAddr + 1, addr);
8400 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
8401 "alignment problem");
8402
8403 #ifdef ASSERT
8404 if (oop(addr)->klass_or_null() != NULL) {
8405 // Ignore mark word because we are running concurrent with mutators
8406 assert(oop(addr)->is_oop(true), "live block should be an oop");
8407 assert(size ==
8408 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
8409 "P-mark and computed size do not agree");
8410 }
8411 #endif
8412
8413 } else {
8414 // This should be an initialized object that's alive.
8415 assert(oop(addr)->klass_or_null() != NULL,
8416 "Should be an initialized object");
8417 // Ignore mark word because we are running concurrent with mutators
8418 assert(oop(addr)->is_oop(true), "live block should be an oop");
8419 // Verify that the bit map has no bits marked between
8420 // addr and purported end of this block.
8421 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8422 assert(size >= 3, "Necessary for Printezis marks to work");
8423 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
8424 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
8425 }
8426 return size;
8427 }
8428
8429 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
8430 size_t chunkSize) {
8431 // do_post_free_or_garbage_chunk() should only be called in the case
8432 // of the adaptive free list allocator.
8433 const bool fcInFreeLists = fc->is_free();
8434 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
8435 assert((HeapWord*)fc <= _limit, "sweep invariant");
8436 if (CMSTestInFreeList && fcInFreeLists) {
8437 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
8438 }
8439
8440 if (CMSTraceSweeper) {
8441 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
8442 }
8443
8444 HeapWord* const fc_addr = (HeapWord*) fc;
8445
8446 bool coalesce;
8447 const size_t left = pointer_delta(fc_addr, freeFinger());
8448 const size_t right = chunkSize;
8449 switch (FLSCoalescePolicy) {
8450 // numeric value forms a coalition aggressiveness metric
8451 case 0: { // never coalesce
8452 coalesce = false;
8453 break;
8454 }
8455 case 1: { // coalesce if left & right chunks on overpopulated lists
8456 coalesce = _sp->coalOverPopulated(left) &&
8457 _sp->coalOverPopulated(right);
8458 break;
8459 }
8460 case 2: { // coalesce if left chunk on overpopulated list (default)
8461 coalesce = _sp->coalOverPopulated(left);
8462 break;
8463 }
8464 case 3: { // coalesce if left OR right chunk on overpopulated list
8465 coalesce = _sp->coalOverPopulated(left) ||
8466 _sp->coalOverPopulated(right);
8467 break;
8468 }
8469 case 4: { // always coalesce
8470 coalesce = true;
8471 break;
8472 }
8473 default:
8474 ShouldNotReachHere();
8475 }
8476
8477 // Should the current free range be coalesced?
8478 // If the chunk is in a free range and either we decided to coalesce above
8479 // or the chunk is near the large block at the end of the heap
8480 // (isNearLargestChunk() returns true), then coalesce this chunk.
8481 const bool doCoalesce = inFreeRange()
8482 && (coalesce || _g->isNearLargestChunk(fc_addr));
8483 if (doCoalesce) {
8484 // Coalesce the current free range on the left with the new
8485 // chunk on the right. If either is on a free list,
8486 // it must be removed from the list and stashed in the closure.
8487 if (freeRangeInFreeLists()) {
8488 FreeChunk* const ffc = (FreeChunk*)freeFinger();
8489 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
8490 "Size of free range is inconsistent with chunk size.");
8491 if (CMSTestInFreeList) {
8492 assert(_sp->verify_chunk_in_free_list(ffc),
8493 "Chunk is not in free lists");
8494 }
8495 _sp->coalDeath(ffc->size());
8496 _sp->removeFreeChunkFromFreeLists(ffc);
8497 set_freeRangeInFreeLists(false);
8498 }
8499 if (fcInFreeLists) {
8500 _sp->coalDeath(chunkSize);
8501 assert(fc->size() == chunkSize,
8502 "The chunk has the wrong size or is not in the free lists");
8503 _sp->removeFreeChunkFromFreeLists(fc);
8504 }
8505 set_lastFreeRangeCoalesced(true);
8506 print_free_block_coalesced(fc);
8507 } else { // not in a free range and/or should not coalesce
8508 // Return the current free range and start a new one.
8509 if (inFreeRange()) {
8510 // In a free range but cannot coalesce with the right hand chunk.
8511 // Put the current free range into the free lists.
8512 flush_cur_free_chunk(freeFinger(),
8513 pointer_delta(fc_addr, freeFinger()));
8514 }
8515 // Set up for new free range. Pass along whether the right hand
8516 // chunk is in the free lists.
8517 initialize_free_range((HeapWord*)fc, fcInFreeLists);
8518 }
8519 }
8520
8521 // Lookahead flush:
8522 // If we are tracking a free range, and this is the last chunk that
8523 // we'll look at because its end crosses past _limit, we'll preemptively
8524 // flush it along with any free range we may be holding on to. Note that
8525 // this can be the case only for an already free or freshly garbage
8526 // chunk. If this block is an object, it can never straddle
8527 // over _limit. The "straddling" occurs when _limit is set at
8528 // the previous end of the space when this cycle started, and
8529 // a subsequent heap expansion caused the previously co-terminal
8530 // free block to be coalesced with the newly expanded portion,
8531 // thus rendering _limit a non-block-boundary making it dangerous
8532 // for the sweeper to step over and examine.
8533 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
8534 assert(inFreeRange(), "Should only be called if currently in a free range.");
8535 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
8536 assert(_sp->used_region().contains(eob - 1),
8537 err_msg("eob = " PTR_FORMAT " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
8538 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
8539 _limit, _sp->bottom(), _sp->end(), fc, chunk_size));
8540 if (eob >= _limit) {
8541 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
8542 if (CMSTraceSweeper) {
8543 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
8544 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
8545 "[" PTR_FORMAT "," PTR_FORMAT ")",
8546 _limit, fc, eob, _sp->bottom(), _sp->end());
8547 }
8548 // Return the storage we are tracking back into the free lists.
8549 if (CMSTraceSweeper) {
8550 gclog_or_tty->print_cr("Flushing ... ");
8551 }
8552 assert(freeFinger() < eob, "Error");
8553 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
8554 }
8555 }
8556
8557 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
8558 assert(inFreeRange(), "Should only be called if currently in a free range.");
8559 assert(size > 0,
8560 "A zero sized chunk cannot be added to the free lists.");
8561 if (!freeRangeInFreeLists()) {
8562 if (CMSTestInFreeList) {
8563 FreeChunk* fc = (FreeChunk*) chunk;
8564 fc->set_size(size);
8565 assert(!_sp->verify_chunk_in_free_list(fc),
8566 "chunk should not be in free lists yet");
8567 }
8568 if (CMSTraceSweeper) {
8569 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8570 chunk, size);
8571 }
8572 // A new free range is going to be starting. The current
8573 // free range has not been added to the free lists yet or
8574 // was removed so add it back.
8575 // If the current free range was coalesced, then the death
8576 // of the free range was recorded. Record a birth now.
8577 if (lastFreeRangeCoalesced()) {
8578 _sp->coalBirth(size);
8579 }
8580 _sp->addChunkAndRepairOffsetTable(chunk, size,
8581 lastFreeRangeCoalesced());
8582 } else if (CMSTraceSweeper) {
8583 gclog_or_tty->print_cr("Already in free list: nothing to flush");
8584 }
8585 set_inFreeRange(false);
8586 set_freeRangeInFreeLists(false);
8587 }
8588
8589 // We take a break if we've been at this for a while,
8590 // so as to avoid monopolizing the locks involved.
8591 void SweepClosure::do_yield_work(HeapWord* addr) {
8592 // Return current free chunk being used for coalescing (if any)
8593 // to the appropriate freelist. After yielding, the next
8594 // free block encountered will start a coalescing range of
8595 // free blocks. If the next free block is adjacent to the
8596 // chunk just flushed, they will need to wait for the next
8597 // sweep to be coalesced.
8598 if (inFreeRange()) {
8599 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8600 }
8601
8602 // First give up the locks, then yield, then re-lock.
8603 // We should probably use a constructor/destructor idiom to
8604 // do this unlock/lock or modify the MutexUnlocker class to
8605 // serve our purpose. XXX
8606 assert_lock_strong(_bitMap->lock());
8607 assert_lock_strong(_freelistLock);
8608 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8609 "CMS thread should hold CMS token");
8610 _bitMap->lock()->unlock();
8611 _freelistLock->unlock();
8612 ConcurrentMarkSweepThread::desynchronize(true);
8613 ConcurrentMarkSweepThread::acknowledge_yield_request();
8614 _collector->stopTimer();
8615 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8616 if (PrintCMSStatistics != 0) {
8617 _collector->incrementYields();
8618 }
8619 _collector->icms_wait();
8620
8621 // See the comment in coordinator_yield()
8622 for (unsigned i = 0; i < CMSYieldSleepCount &&
8623 ConcurrentMarkSweepThread::should_yield() &&
8624 !CMSCollector::foregroundGCIsActive(); ++i) {
8625 os::sleep(Thread::current(), 1, false);
8626 ConcurrentMarkSweepThread::acknowledge_yield_request();
8627 }
8628
8629 ConcurrentMarkSweepThread::synchronize(true);
8630 _freelistLock->lock();
8631 _bitMap->lock()->lock_without_safepoint_check();
8632 _collector->startTimer();
8633 }
8634
8635 #ifndef PRODUCT
8636 // This is actually very useful in a product build if it can
8637 // be called from the debugger. Compile it into the product
8638 // as needed.
8639 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
8640 return debug_cms_space->verify_chunk_in_free_list(fc);
8641 }
8642 #endif
8643
8644 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
8645 if (CMSTraceSweeper) {
8646 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
8647 fc, fc->size());
8648 }
8649 }
8650
8651 // CMSIsAliveClosure
8652 bool CMSIsAliveClosure::do_object_b(oop obj) {
8653 HeapWord* addr = (HeapWord*)obj;
8654 return addr != NULL &&
8655 (!_span.contains(addr) || _bit_map->isMarked(addr));
8656 }
8657
8658
8659 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
8660 MemRegion span,
8661 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
8662 bool cpc):
8663 _collector(collector),
8664 _span(span),
8665 _bit_map(bit_map),
8666 _mark_stack(mark_stack),
8667 _concurrent_precleaning(cpc) {
8668 assert(!_span.is_empty(), "Empty span could spell trouble");
8669 }
8670
8671
8672 // CMSKeepAliveClosure: the serial version
8673 void CMSKeepAliveClosure::do_oop(oop obj) {
8674 HeapWord* addr = (HeapWord*)obj;
8675 if (_span.contains(addr) &&
8676 !_bit_map->isMarked(addr)) {
8677 _bit_map->mark(addr);
8678 bool simulate_overflow = false;
8679 NOT_PRODUCT(
8680 if (CMSMarkStackOverflowALot &&
8681 _collector->simulate_overflow()) {
8682 // simulate a stack overflow
8683 simulate_overflow = true;
8684 }
8685 )
8686 if (simulate_overflow || !_mark_stack->push(obj)) {
8687 if (_concurrent_precleaning) {
8688 // We dirty the overflown object and let the remark
8689 // phase deal with it.
8690 assert(_collector->overflow_list_is_empty(), "Error");
8691 // In the case of object arrays, we need to dirty all of
8692 // the cards that the object spans. No locking or atomics
8693 // are needed since no one else can be mutating the mod union
8694 // table.
8695 if (obj->is_objArray()) {
8696 size_t sz = obj->size();
8697 HeapWord* end_card_addr =
8698 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
8699 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8700 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8701 _collector->_modUnionTable.mark_range(redirty_range);
8702 } else {
8703 _collector->_modUnionTable.mark(addr);
8704 }
8705 _collector->_ser_kac_preclean_ovflw++;
8706 } else {
8707 _collector->push_on_overflow_list(obj);
8708 _collector->_ser_kac_ovflw++;
8709 }
8710 }
8711 }
8712 }
8713
8714 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8715 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8716
8717 // CMSParKeepAliveClosure: a parallel version of the above.
8718 // The work queues are private to each closure (thread),
8719 // but (may be) available for stealing by other threads.
8720 void CMSParKeepAliveClosure::do_oop(oop obj) {
8721 HeapWord* addr = (HeapWord*)obj;
8722 if (_span.contains(addr) &&
8723 !_bit_map->isMarked(addr)) {
8724 // In general, during recursive tracing, several threads
8725 // may be concurrently getting here; the first one to
8726 // "tag" it, claims it.
8727 if (_bit_map->par_mark(addr)) {
8728 bool res = _work_queue->push(obj);
8729 assert(res, "Low water mark should be much less than capacity");
8730 // Do a recursive trim in the hope that this will keep
8731 // stack usage lower, but leave some oops for potential stealers
8732 trim_queue(_low_water_mark);
8733 } // Else, another thread got there first
8734 }
8735 }
8736
8737 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8738 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8739
8740 void CMSParKeepAliveClosure::trim_queue(uint max) {
8741 while (_work_queue->size() > max) {
8742 oop new_oop;
8743 if (_work_queue->pop_local(new_oop)) {
8744 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8745 assert(_bit_map->isMarked((HeapWord*)new_oop),
8746 "no white objects on this stack!");
8747 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8748 // iterate over the oops in this oop, marking and pushing
8749 // the ones in CMS heap (i.e. in _span).
8750 new_oop->oop_iterate(&_mark_and_push);
8751 }
8752 }
8753 }
8754
8755 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8756 CMSCollector* collector,
8757 MemRegion span, CMSBitMap* bit_map,
8758 OopTaskQueue* work_queue):
8759 _collector(collector),
8760 _span(span),
8761 _bit_map(bit_map),
8762 _work_queue(work_queue) { }
8763
8764 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8765 HeapWord* addr = (HeapWord*)obj;
8766 if (_span.contains(addr) &&
8767 !_bit_map->isMarked(addr)) {
8768 if (_bit_map->par_mark(addr)) {
8769 bool simulate_overflow = false;
8770 NOT_PRODUCT(
8771 if (CMSMarkStackOverflowALot &&
8772 _collector->par_simulate_overflow()) {
8773 // simulate a stack overflow
8774 simulate_overflow = true;
8775 }
8776 )
8777 if (simulate_overflow || !_work_queue->push(obj)) {
8778 _collector->par_push_on_overflow_list(obj);
8779 _collector->_par_kac_ovflw++;
8780 }
8781 } // Else another thread got there already
8782 }
8783 }
8784
8785 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8786 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8787
8788 //////////////////////////////////////////////////////////////////
8789 // CMSExpansionCause /////////////////////////////
8790 //////////////////////////////////////////////////////////////////
8791 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8792 switch (cause) {
8793 case _no_expansion:
8794 return "No expansion";
8795 case _satisfy_free_ratio:
8796 return "Free ratio";
8797 case _satisfy_promotion:
8798 return "Satisfy promotion";
8799 case _satisfy_allocation:
8800 return "allocation";
8801 case _allocate_par_lab:
8802 return "Par LAB";
8803 case _allocate_par_spooling_space:
8804 return "Par Spooling Space";
8805 case _adaptive_size_policy:
8806 return "Ergonomics";
8807 default:
8808 return "unknown";
8809 }
8810 }
8811
8812 void CMSDrainMarkingStackClosure::do_void() {
8813 // the max number to take from overflow list at a time
8814 const size_t num = _mark_stack->capacity()/4;
8815 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8816 "Overflow list should be NULL during concurrent phases");
8817 while (!_mark_stack->isEmpty() ||
8818 // if stack is empty, check the overflow list
8819 _collector->take_from_overflow_list(num, _mark_stack)) {
8820 oop obj = _mark_stack->pop();
8821 HeapWord* addr = (HeapWord*)obj;
8822 assert(_span.contains(addr), "Should be within span");
8823 assert(_bit_map->isMarked(addr), "Should be marked");
8824 assert(obj->is_oop(), "Should be an oop");
8825 obj->oop_iterate(_keep_alive);
8826 }
8827 }
8828
8829 void CMSParDrainMarkingStackClosure::do_void() {
8830 // drain queue
8831 trim_queue(0);
8832 }
8833
8834 // Trim our work_queue so its length is below max at return
8835 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8836 while (_work_queue->size() > max) {
8837 oop new_oop;
8838 if (_work_queue->pop_local(new_oop)) {
8839 assert(new_oop->is_oop(), "Expected an oop");
8840 assert(_bit_map->isMarked((HeapWord*)new_oop),
8841 "no white objects on this stack!");
8842 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8843 // iterate over the oops in this oop, marking and pushing
8844 // the ones in CMS heap (i.e. in _span).
8845 new_oop->oop_iterate(&_mark_and_push);
8846 }
8847 }
8848 }
8849
8850 ////////////////////////////////////////////////////////////////////
8851 // Support for Marking Stack Overflow list handling and related code
8852 ////////////////////////////////////////////////////////////////////
8853 // Much of the following code is similar in shape and spirit to the
8854 // code used in ParNewGC. We should try and share that code
8855 // as much as possible in the future.
8856
8857 #ifndef PRODUCT
8858 // Debugging support for CMSStackOverflowALot
8859
8860 // It's OK to call this multi-threaded; the worst thing
8861 // that can happen is that we'll get a bunch of closely
8862 // spaced simulated oveflows, but that's OK, in fact
8863 // probably good as it would exercise the overflow code
8864 // under contention.
8865 bool CMSCollector::simulate_overflow() {
8866 if (_overflow_counter-- <= 0) { // just being defensive
8867 _overflow_counter = CMSMarkStackOverflowInterval;
8868 return true;
8869 } else {
8870 return false;
8871 }
8872 }
8873
8874 bool CMSCollector::par_simulate_overflow() {
8875 return simulate_overflow();
8876 }
8877 #endif
8878
8879 // Single-threaded
8880 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8881 assert(stack->isEmpty(), "Expected precondition");
8882 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8883 size_t i = num;
8884 oop cur = _overflow_list;
8885 const markOop proto = markOopDesc::prototype();
8886 NOT_PRODUCT(ssize_t n = 0;)
8887 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8888 next = oop(cur->mark());
8889 cur->set_mark(proto); // until proven otherwise
8890 assert(cur->is_oop(), "Should be an oop");
8891 bool res = stack->push(cur);
8892 assert(res, "Bit off more than can chew?");
8893 NOT_PRODUCT(n++;)
8894 }
8895 _overflow_list = cur;
8896 #ifndef PRODUCT
8897 assert(_num_par_pushes >= n, "Too many pops?");
8898 _num_par_pushes -=n;
8899 #endif
8900 return !stack->isEmpty();
8901 }
8902
8903 #define BUSY (oop(0x1aff1aff))
8904 // (MT-safe) Get a prefix of at most "num" from the list.
8905 // The overflow list is chained through the mark word of
8906 // each object in the list. We fetch the entire list,
8907 // break off a prefix of the right size and return the
8908 // remainder. If other threads try to take objects from
8909 // the overflow list at that time, they will wait for
8910 // some time to see if data becomes available. If (and
8911 // only if) another thread places one or more object(s)
8912 // on the global list before we have returned the suffix
8913 // to the global list, we will walk down our local list
8914 // to find its end and append the global list to
8915 // our suffix before returning it. This suffix walk can
8916 // prove to be expensive (quadratic in the amount of traffic)
8917 // when there are many objects in the overflow list and
8918 // there is much producer-consumer contention on the list.
8919 // *NOTE*: The overflow list manipulation code here and
8920 // in ParNewGeneration:: are very similar in shape,
8921 // except that in the ParNew case we use the old (from/eden)
8922 // copy of the object to thread the list via its klass word.
8923 // Because of the common code, if you make any changes in
8924 // the code below, please check the ParNew version to see if
8925 // similar changes might be needed.
8926 // CR 6797058 has been filed to consolidate the common code.
8927 bool CMSCollector::par_take_from_overflow_list(size_t num,
8928 OopTaskQueue* work_q,
8929 int no_of_gc_threads) {
8930 assert(work_q->size() == 0, "First empty local work queue");
8931 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8932 if (_overflow_list == NULL) {
8933 return false;
8934 }
8935 // Grab the entire list; we'll put back a suffix
8936 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
8937 Thread* tid = Thread::current();
8938 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
8939 // set to ParallelGCThreads.
8940 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
8941 size_t sleep_time_millis = MAX2((size_t)1, num/100);
8942 // If the list is busy, we spin for a short while,
8943 // sleeping between attempts to get the list.
8944 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
8945 os::sleep(tid, sleep_time_millis, false);
8946 if (_overflow_list == NULL) {
8947 // Nothing left to take
8948 return false;
8949 } else if (_overflow_list != BUSY) {
8950 // Try and grab the prefix
8951 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
8952 }
8953 }
8954 // If the list was found to be empty, or we spun long
8955 // enough, we give up and return empty-handed. If we leave
8956 // the list in the BUSY state below, it must be the case that
8957 // some other thread holds the overflow list and will set it
8958 // to a non-BUSY state in the future.
8959 if (prefix == NULL || prefix == BUSY) {
8960 // Nothing to take or waited long enough
8961 if (prefix == NULL) {
8962 // Write back the NULL in case we overwrote it with BUSY above
8963 // and it is still the same value.
8964 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8965 }
8966 return false;
8967 }
8968 assert(prefix != NULL && prefix != BUSY, "Error");
8969 size_t i = num;
8970 oop cur = prefix;
8971 // Walk down the first "num" objects, unless we reach the end.
8972 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8973 if (cur->mark() == NULL) {
8974 // We have "num" or fewer elements in the list, so there
8975 // is nothing to return to the global list.
8976 // Write back the NULL in lieu of the BUSY we wrote
8977 // above, if it is still the same value.
8978 if (_overflow_list == BUSY) {
8979 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8980 }
8981 } else {
8982 // Chop off the suffix and rerturn it to the global list.
8983 assert(cur->mark() != BUSY, "Error");
8984 oop suffix_head = cur->mark(); // suffix will be put back on global list
8985 cur->set_mark(NULL); // break off suffix
8986 // It's possible that the list is still in the empty(busy) state
8987 // we left it in a short while ago; in that case we may be
8988 // able to place back the suffix without incurring the cost
8989 // of a walk down the list.
8990 oop observed_overflow_list = _overflow_list;
8991 oop cur_overflow_list = observed_overflow_list;
8992 bool attached = false;
8993 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
8994 observed_overflow_list =
8995 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8996 if (cur_overflow_list == observed_overflow_list) {
8997 attached = true;
8998 break;
8999 } else cur_overflow_list = observed_overflow_list;
9000 }
9001 if (!attached) {
9002 // Too bad, someone else sneaked in (at least) an element; we'll need
9003 // to do a splice. Find tail of suffix so we can prepend suffix to global
9004 // list.
9005 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
9006 oop suffix_tail = cur;
9007 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
9008 "Tautology");
9009 observed_overflow_list = _overflow_list;
9010 do {
9011 cur_overflow_list = observed_overflow_list;
9012 if (cur_overflow_list != BUSY) {
9013 // Do the splice ...
9014 suffix_tail->set_mark(markOop(cur_overflow_list));
9015 } else { // cur_overflow_list == BUSY
9016 suffix_tail->set_mark(NULL);
9017 }
9018 // ... and try to place spliced list back on overflow_list ...
9019 observed_overflow_list =
9020 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
9021 } while (cur_overflow_list != observed_overflow_list);
9022 // ... until we have succeeded in doing so.
9023 }
9024 }
9025
9026 // Push the prefix elements on work_q
9027 assert(prefix != NULL, "control point invariant");
9028 const markOop proto = markOopDesc::prototype();
9029 oop next;
9030 NOT_PRODUCT(ssize_t n = 0;)
9031 for (cur = prefix; cur != NULL; cur = next) {
9032 next = oop(cur->mark());
9033 cur->set_mark(proto); // until proven otherwise
9034 assert(cur->is_oop(), "Should be an oop");
9035 bool res = work_q->push(cur);
9036 assert(res, "Bit off more than we can chew?");
9037 NOT_PRODUCT(n++;)
9038 }
9039 #ifndef PRODUCT
9040 assert(_num_par_pushes >= n, "Too many pops?");
9041 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
9042 #endif
9043 return true;
9044 }
9045
9046 // Single-threaded
9047 void CMSCollector::push_on_overflow_list(oop p) {
9048 NOT_PRODUCT(_num_par_pushes++;)
9049 assert(p->is_oop(), "Not an oop");
9050 preserve_mark_if_necessary(p);
9051 p->set_mark((markOop)_overflow_list);
9052 _overflow_list = p;
9053 }
9054
9055 // Multi-threaded; use CAS to prepend to overflow list
9056 void CMSCollector::par_push_on_overflow_list(oop p) {
9057 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
9058 assert(p->is_oop(), "Not an oop");
9059 par_preserve_mark_if_necessary(p);
9060 oop observed_overflow_list = _overflow_list;
9061 oop cur_overflow_list;
9062 do {
9063 cur_overflow_list = observed_overflow_list;
9064 if (cur_overflow_list != BUSY) {
9065 p->set_mark(markOop(cur_overflow_list));
9066 } else {
9067 p->set_mark(NULL);
9068 }
9069 observed_overflow_list =
9070 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
9071 } while (cur_overflow_list != observed_overflow_list);
9072 }
9073 #undef BUSY
9074
9075 // Single threaded
9076 // General Note on GrowableArray: pushes may silently fail
9077 // because we are (temporarily) out of C-heap for expanding
9078 // the stack. The problem is quite ubiquitous and affects
9079 // a lot of code in the JVM. The prudent thing for GrowableArray
9080 // to do (for now) is to exit with an error. However, that may
9081 // be too draconian in some cases because the caller may be
9082 // able to recover without much harm. For such cases, we
9083 // should probably introduce a "soft_push" method which returns
9084 // an indication of success or failure with the assumption that
9085 // the caller may be able to recover from a failure; code in
9086 // the VM can then be changed, incrementally, to deal with such
9087 // failures where possible, thus, incrementally hardening the VM
9088 // in such low resource situations.
9089 void CMSCollector::preserve_mark_work(oop p, markOop m) {
9090 _preserved_oop_stack.push(p);
9091 _preserved_mark_stack.push(m);
9092 assert(m == p->mark(), "Mark word changed");
9093 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9094 "bijection");
9095 }
9096
9097 // Single threaded
9098 void CMSCollector::preserve_mark_if_necessary(oop p) {
9099 markOop m = p->mark();
9100 if (m->must_be_preserved(p)) {
9101 preserve_mark_work(p, m);
9102 }
9103 }
9104
9105 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
9106 markOop m = p->mark();
9107 if (m->must_be_preserved(p)) {
9108 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
9109 // Even though we read the mark word without holding
9110 // the lock, we are assured that it will not change
9111 // because we "own" this oop, so no other thread can
9112 // be trying to push it on the overflow list; see
9113 // the assertion in preserve_mark_work() that checks
9114 // that m == p->mark().
9115 preserve_mark_work(p, m);
9116 }
9117 }
9118
9119 // We should be able to do this multi-threaded,
9120 // a chunk of stack being a task (this is
9121 // correct because each oop only ever appears
9122 // once in the overflow list. However, it's
9123 // not very easy to completely overlap this with
9124 // other operations, so will generally not be done
9125 // until all work's been completed. Because we
9126 // expect the preserved oop stack (set) to be small,
9127 // it's probably fine to do this single-threaded.
9128 // We can explore cleverer concurrent/overlapped/parallel
9129 // processing of preserved marks if we feel the
9130 // need for this in the future. Stack overflow should
9131 // be so rare in practice and, when it happens, its
9132 // effect on performance so great that this will
9133 // likely just be in the noise anyway.
9134 void CMSCollector::restore_preserved_marks_if_any() {
9135 assert(SafepointSynchronize::is_at_safepoint(),
9136 "world should be stopped");
9137 assert(Thread::current()->is_ConcurrentGC_thread() ||
9138 Thread::current()->is_VM_thread(),
9139 "should be single-threaded");
9140 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9141 "bijection");
9142
9143 while (!_preserved_oop_stack.is_empty()) {
9144 oop p = _preserved_oop_stack.pop();
9145 assert(p->is_oop(), "Should be an oop");
9146 assert(_span.contains(p), "oop should be in _span");
9147 assert(p->mark() == markOopDesc::prototype(),
9148 "Set when taken from overflow list");
9149 markOop m = _preserved_mark_stack.pop();
9150 p->set_mark(m);
9151 }
9152 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
9153 "stacks were cleared above");
9154 }
9155
9156 #ifndef PRODUCT
9157 bool CMSCollector::no_preserved_marks() const {
9158 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
9159 }
9160 #endif
9161
9162 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
9163 {
9164 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
9165 CMSAdaptiveSizePolicy* size_policy =
9166 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
9167 assert(size_policy->is_gc_cms_adaptive_size_policy(),
9168 "Wrong type for size policy");
9169 return size_policy;
9170 }
9171
9172 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
9173 size_t desired_promo_size) {
9174 if (cur_promo_size < desired_promo_size) {
9175 size_t expand_bytes = desired_promo_size - cur_promo_size;
9176 if (PrintAdaptiveSizePolicy && Verbose) {
9177 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9178 "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
9179 expand_bytes);
9180 }
9181 expand(expand_bytes,
9182 MinHeapDeltaBytes,
9183 CMSExpansionCause::_adaptive_size_policy);
9184 } else if (desired_promo_size < cur_promo_size) {
9185 size_t shrink_bytes = cur_promo_size - desired_promo_size;
9186 if (PrintAdaptiveSizePolicy && Verbose) {
9187 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9188 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
9189 shrink_bytes);
9190 }
9191 shrink(shrink_bytes);
9192 }
9193 }
9194
9195 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
9196 GenCollectedHeap* gch = GenCollectedHeap::heap();
9197 CMSGCAdaptivePolicyCounters* counters =
9198 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
9199 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
9200 "Wrong kind of counters");
9201 return counters;
9202 }
9203
9204
9205 void ASConcurrentMarkSweepGeneration::update_counters() {
9206 if (UsePerfData) {
9207 _space_counters->update_all();
9208 _gen_counters->update_all();
9209 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9210 GenCollectedHeap* gch = GenCollectedHeap::heap();
9211 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9212 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9213 "Wrong gc statistics type");
9214 counters->update_counters(gc_stats_l);
9215 }
9216 }
9217
9218 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
9219 if (UsePerfData) {
9220 _space_counters->update_used(used);
9221 _space_counters->update_capacity();
9222 _gen_counters->update_all();
9223
9224 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9225 GenCollectedHeap* gch = GenCollectedHeap::heap();
9226 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9227 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9228 "Wrong gc statistics type");
9229 counters->update_counters(gc_stats_l);
9230 }
9231 }
9232
9233 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
9234 assert_locked_or_safepoint(Heap_lock);
9235 assert_lock_strong(freelistLock());
9236 HeapWord* old_end = _cmsSpace->end();
9237 HeapWord* unallocated_start = _cmsSpace->unallocated_block();
9238 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
9239 FreeChunk* chunk_at_end = find_chunk_at_end();
9240 if (chunk_at_end == NULL) {
9241 // No room to shrink
9242 if (PrintGCDetails && Verbose) {
9243 gclog_or_tty->print_cr("No room to shrink: old_end "
9244 PTR_FORMAT " unallocated_start " PTR_FORMAT
9245 " chunk_at_end " PTR_FORMAT,
9246 old_end, unallocated_start, chunk_at_end);
9247 }
9248 return;
9249 } else {
9250
9251 // Find the chunk at the end of the space and determine
9252 // how much it can be shrunk.
9253 size_t shrinkable_size_in_bytes = chunk_at_end->size();
9254 size_t aligned_shrinkable_size_in_bytes =
9255 align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
9256 assert(unallocated_start <= (HeapWord*) chunk_at_end->end(),
9257 "Inconsistent chunk at end of space");
9258 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
9259 size_t word_size_before = heap_word_size(_virtual_space.committed_size());
9260
9261 // Shrink the underlying space
9262 _virtual_space.shrink_by(bytes);
9263 if (PrintGCDetails && Verbose) {
9264 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
9265 " desired_bytes " SIZE_FORMAT
9266 " shrinkable_size_in_bytes " SIZE_FORMAT
9267 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
9268 " bytes " SIZE_FORMAT,
9269 desired_bytes, shrinkable_size_in_bytes,
9270 aligned_shrinkable_size_in_bytes, bytes);
9271 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT
9272 " unallocated_start " SIZE_FORMAT,
9273 old_end, unallocated_start);
9274 }
9275
9276 // If the space did shrink (shrinking is not guaranteed),
9277 // shrink the chunk at the end by the appropriate amount.
9278 if (((HeapWord*)_virtual_space.high()) < old_end) {
9279 size_t new_word_size =
9280 heap_word_size(_virtual_space.committed_size());
9281
9282 // Have to remove the chunk from the dictionary because it is changing
9283 // size and might be someplace elsewhere in the dictionary.
9284
9285 // Get the chunk at end, shrink it, and put it
9286 // back.
9287 _cmsSpace->removeChunkFromDictionary(chunk_at_end);
9288 size_t word_size_change = word_size_before - new_word_size;
9289 size_t chunk_at_end_old_size = chunk_at_end->size();
9290 assert(chunk_at_end_old_size >= word_size_change,
9291 "Shrink is too large");
9292 chunk_at_end->set_size(chunk_at_end_old_size -
9293 word_size_change);
9294 _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
9295 word_size_change);
9296
9297 _cmsSpace->returnChunkToDictionary(chunk_at_end);
9298
9299 MemRegion mr(_cmsSpace->bottom(), new_word_size);
9300 _bts->resize(new_word_size); // resize the block offset shared array
9301 Universe::heap()->barrier_set()->resize_covered_region(mr);
9302 _cmsSpace->assert_locked();
9303 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
9304
9305 NOT_PRODUCT(_cmsSpace->dictionary()->verify());
9306
9307 // update the space and generation capacity counters
9308 if (UsePerfData) {
9309 _space_counters->update_capacity();
9310 _gen_counters->update_all();
9311 }
9312
9313 if (Verbose && PrintGCDetails) {
9314 size_t new_mem_size = _virtual_space.committed_size();
9315 size_t old_mem_size = new_mem_size + bytes;
9316 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K",
9317 name(), old_mem_size/K, bytes/K, new_mem_size/K);
9318 }
9319 }
9320
9321 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
9322 "Inconsistency at end of space");
9323 assert(chunk_at_end->end() == (uintptr_t*) _cmsSpace->end(),
9324 "Shrinking is inconsistent");
9325 return;
9326 }
9327 }
9328
9329 // Transfer some number of overflown objects to usual marking
9330 // stack. Return true if some objects were transferred.
9331 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
9332 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
9333 (size_t)ParGCDesiredObjsFromOverflowList);
9334
9335 bool res = _collector->take_from_overflow_list(num, _mark_stack);
9336 assert(_collector->overflow_list_is_empty() || res,
9337 "If list is not empty, we should have taken something");
9338 assert(!res || !_mark_stack->isEmpty(),
9339 "If we took something, it should now be on our stack");
9340 return res;
9341 }
9342
9343 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
9344 size_t res = _sp->block_size_no_stall(addr, _collector);
9345 if (_sp->block_is_obj(addr)) {
9346 if (_live_bit_map->isMarked(addr)) {
9347 // It can't have been dead in a previous cycle
9348 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
9349 } else {
9350 _dead_bit_map->mark(addr); // mark the dead object
9351 }
9352 }
9353 // Could be 0, if the block size could not be computed without stalling.
9354 return res;
9355 }
9356
9357 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
9358
9359 switch (phase) {
9360 case CMSCollector::InitialMarking:
9361 initialize(true /* fullGC */ ,
9362 cause /* cause of the GC */,
9363 true /* recordGCBeginTime */,
9364 true /* recordPreGCUsage */,
9365 false /* recordPeakUsage */,
9366 false /* recordPostGCusage */,
9367 true /* recordAccumulatedGCTime */,
9368 false /* recordGCEndTime */,
9369 false /* countCollection */ );
9370 break;
9371
9372 case CMSCollector::FinalMarking:
9373 initialize(true /* fullGC */ ,
9374 cause /* cause of the GC */,
9375 false /* recordGCBeginTime */,
9376 false /* recordPreGCUsage */,
9377 false /* recordPeakUsage */,
9378 false /* recordPostGCusage */,
9379 true /* recordAccumulatedGCTime */,
9380 false /* recordGCEndTime */,
9381 false /* countCollection */ );
9382 break;
9383
9384 case CMSCollector::Sweeping:
9385 initialize(true /* fullGC */ ,
9386 cause /* cause of the GC */,
9387 false /* recordGCBeginTime */,
9388 false /* recordPreGCUsage */,
9389 true /* recordPeakUsage */,
9390 true /* recordPostGCusage */,
9391 false /* recordAccumulatedGCTime */,
9392 true /* recordGCEndTime */,
9393 true /* countCollection */ );
9394 break;
9395
9396 default:
9397 ShouldNotReachHere();
9398 }
9399 }
9400