Thu, 22 May 2014 15:52:41 -0400
8037816: Fix for 8036122 breaks build with Xcode5/clang
8043029: Change 8037816 breaks HS build with older GCC versions which don't support diagnostic pragmas
8043164: Format warning in traceStream.hpp
Summary: Backport of main fix + two corrections, enables clang compilation, turns on format attributes, corrects/mutes warnings
Reviewed-by: kvn, coleenp, iveresov, twisti
1 /*
2 * Copyright (c) 2001, 2014, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
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/padded.hpp"
54 #include "memory/referencePolicy.hpp"
55 #include "memory/resourceArea.hpp"
56 #include "memory/tenuredGeneration.hpp"
57 #include "oops/oop.inline.hpp"
58 #include "prims/jvmtiExport.hpp"
59 #include "runtime/globals_extension.hpp"
60 #include "runtime/handles.inline.hpp"
61 #include "runtime/java.hpp"
62 #include "runtime/vmThread.hpp"
63 #include "services/memoryService.hpp"
64 #include "services/runtimeService.hpp"
66 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
68 // statics
69 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
70 bool CMSCollector::_full_gc_requested = false;
71 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
73 //////////////////////////////////////////////////////////////////
74 // In support of CMS/VM thread synchronization
75 //////////////////////////////////////////////////////////////////
76 // We split use of the CGC_lock into 2 "levels".
77 // The low-level locking is of the usual CGC_lock monitor. We introduce
78 // a higher level "token" (hereafter "CMS token") built on top of the
79 // low level monitor (hereafter "CGC lock").
80 // The token-passing protocol gives priority to the VM thread. The
81 // CMS-lock doesn't provide any fairness guarantees, but clients
82 // should ensure that it is only held for very short, bounded
83 // durations.
84 //
85 // When either of the CMS thread or the VM thread is involved in
86 // collection operations during which it does not want the other
87 // thread to interfere, it obtains the CMS token.
88 //
89 // If either thread tries to get the token while the other has
90 // it, that thread waits. However, if the VM thread and CMS thread
91 // both want the token, then the VM thread gets priority while the
92 // CMS thread waits. This ensures, for instance, that the "concurrent"
93 // phases of the CMS thread's work do not block out the VM thread
94 // for long periods of time as the CMS thread continues to hog
95 // the token. (See bug 4616232).
96 //
97 // The baton-passing functions are, however, controlled by the
98 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
99 // and here the low-level CMS lock, not the high level token,
100 // ensures mutual exclusion.
101 //
102 // Two important conditions that we have to satisfy:
103 // 1. if a thread does a low-level wait on the CMS lock, then it
104 // relinquishes the CMS token if it were holding that token
105 // when it acquired the low-level CMS lock.
106 // 2. any low-level notifications on the low-level lock
107 // should only be sent when a thread has relinquished the token.
108 //
109 // In the absence of either property, we'd have potential deadlock.
110 //
111 // We protect each of the CMS (concurrent and sequential) phases
112 // with the CMS _token_, not the CMS _lock_.
113 //
114 // The only code protected by CMS lock is the token acquisition code
115 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
116 // baton-passing code.
117 //
118 // Unfortunately, i couldn't come up with a good abstraction to factor and
119 // hide the naked CGC_lock manipulation in the baton-passing code
120 // further below. That's something we should try to do. Also, the proof
121 // of correctness of this 2-level locking scheme is far from obvious,
122 // and potentially quite slippery. We have an uneasy supsicion, for instance,
123 // that there may be a theoretical possibility of delay/starvation in the
124 // low-level lock/wait/notify scheme used for the baton-passing because of
125 // potential intereference with the priority scheme embodied in the
126 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
127 // invocation further below and marked with "XXX 20011219YSR".
128 // Indeed, as we note elsewhere, this may become yet more slippery
129 // in the presence of multiple CMS and/or multiple VM threads. XXX
131 class CMSTokenSync: public StackObj {
132 private:
133 bool _is_cms_thread;
134 public:
135 CMSTokenSync(bool is_cms_thread):
136 _is_cms_thread(is_cms_thread) {
137 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
138 "Incorrect argument to constructor");
139 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
140 }
142 ~CMSTokenSync() {
143 assert(_is_cms_thread ?
144 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
145 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
146 "Incorrect state");
147 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
148 }
149 };
151 // Convenience class that does a CMSTokenSync, and then acquires
152 // upto three locks.
153 class CMSTokenSyncWithLocks: public CMSTokenSync {
154 private:
155 // Note: locks are acquired in textual declaration order
156 // and released in the opposite order
157 MutexLockerEx _locker1, _locker2, _locker3;
158 public:
159 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
160 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
161 CMSTokenSync(is_cms_thread),
162 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
163 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
164 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
165 { }
166 };
169 // Wrapper class to temporarily disable icms during a foreground cms collection.
170 class ICMSDisabler: public StackObj {
171 public:
172 // The ctor disables icms and wakes up the thread so it notices the change;
173 // the dtor re-enables icms. Note that the CMSCollector methods will check
174 // CMSIncrementalMode.
175 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); }
176 ~ICMSDisabler() { CMSCollector::enable_icms(); }
177 };
179 //////////////////////////////////////////////////////////////////
180 // Concurrent Mark-Sweep Generation /////////////////////////////
181 //////////////////////////////////////////////////////////////////
183 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
185 // This struct contains per-thread things necessary to support parallel
186 // young-gen collection.
187 class CMSParGCThreadState: public CHeapObj<mtGC> {
188 public:
189 CFLS_LAB lab;
190 PromotionInfo promo;
192 // Constructor.
193 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
194 promo.setSpace(cfls);
195 }
196 };
198 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
199 ReservedSpace rs, size_t initial_byte_size, int level,
200 CardTableRS* ct, bool use_adaptive_freelists,
201 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) :
202 CardGeneration(rs, initial_byte_size, level, ct),
203 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
204 _debug_collection_type(Concurrent_collection_type),
205 _did_compact(false)
206 {
207 HeapWord* bottom = (HeapWord*) _virtual_space.low();
208 HeapWord* end = (HeapWord*) _virtual_space.high();
210 _direct_allocated_words = 0;
211 NOT_PRODUCT(
212 _numObjectsPromoted = 0;
213 _numWordsPromoted = 0;
214 _numObjectsAllocated = 0;
215 _numWordsAllocated = 0;
216 )
218 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
219 use_adaptive_freelists,
220 dictionaryChoice);
221 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
222 if (_cmsSpace == NULL) {
223 vm_exit_during_initialization(
224 "CompactibleFreeListSpace allocation failure");
225 }
226 _cmsSpace->_gen = this;
228 _gc_stats = new CMSGCStats();
230 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
231 // offsets match. The ability to tell free chunks from objects
232 // depends on this property.
233 debug_only(
234 FreeChunk* junk = NULL;
235 assert(UseCompressedClassPointers ||
236 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
237 "Offset of FreeChunk::_prev within FreeChunk must match"
238 " that of OopDesc::_klass within OopDesc");
239 )
240 if (CollectedHeap::use_parallel_gc_threads()) {
241 typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
242 _par_gc_thread_states =
243 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC);
244 if (_par_gc_thread_states == NULL) {
245 vm_exit_during_initialization("Could not allocate par gc structs");
246 }
247 for (uint i = 0; i < ParallelGCThreads; i++) {
248 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
249 if (_par_gc_thread_states[i] == NULL) {
250 vm_exit_during_initialization("Could not allocate par gc structs");
251 }
252 }
253 } else {
254 _par_gc_thread_states = NULL;
255 }
256 _incremental_collection_failed = false;
257 // The "dilatation_factor" is the expansion that can occur on
258 // account of the fact that the minimum object size in the CMS
259 // generation may be larger than that in, say, a contiguous young
260 // generation.
261 // Ideally, in the calculation below, we'd compute the dilatation
262 // factor as: MinChunkSize/(promoting_gen's min object size)
263 // Since we do not have such a general query interface for the
264 // promoting generation, we'll instead just use the mimimum
265 // object size (which today is a header's worth of space);
266 // note that all arithmetic is in units of HeapWords.
267 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
268 assert(_dilatation_factor >= 1.0, "from previous assert");
269 }
272 // The field "_initiating_occupancy" represents the occupancy percentage
273 // at which we trigger a new collection cycle. Unless explicitly specified
274 // via CMSInitiatingOccupancyFraction (argument "io" below), it
275 // is calculated by:
276 //
277 // Let "f" be MinHeapFreeRatio in
278 //
279 // _intiating_occupancy = 100-f +
280 // f * (CMSTriggerRatio/100)
281 // where CMSTriggerRatio is the argument "tr" below.
282 //
283 // That is, if we assume the heap is at its desired maximum occupancy at the
284 // end of a collection, we let CMSTriggerRatio of the (purported) free
285 // space be allocated before initiating a new collection cycle.
286 //
287 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
288 assert(io <= 100 && tr <= 100, "Check the arguments");
289 if (io >= 0) {
290 _initiating_occupancy = (double)io / 100.0;
291 } else {
292 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
293 (double)(tr * MinHeapFreeRatio) / 100.0)
294 / 100.0;
295 }
296 }
298 void ConcurrentMarkSweepGeneration::ref_processor_init() {
299 assert(collector() != NULL, "no collector");
300 collector()->ref_processor_init();
301 }
303 void CMSCollector::ref_processor_init() {
304 if (_ref_processor == NULL) {
305 // Allocate and initialize a reference processor
306 _ref_processor =
307 new ReferenceProcessor(_span, // span
308 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
309 (int) ParallelGCThreads, // mt processing degree
310 _cmsGen->refs_discovery_is_mt(), // mt discovery
311 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
312 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
313 &_is_alive_closure, // closure for liveness info
314 false); // next field updates do not need write barrier
315 // Initialize the _ref_processor field of CMSGen
316 _cmsGen->set_ref_processor(_ref_processor);
318 }
319 }
321 CMSAdaptiveSizePolicy* CMSCollector::size_policy() {
322 GenCollectedHeap* gch = GenCollectedHeap::heap();
323 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
324 "Wrong type of heap");
325 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
326 gch->gen_policy()->size_policy();
327 assert(sp->is_gc_cms_adaptive_size_policy(),
328 "Wrong type of size policy");
329 return sp;
330 }
332 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() {
333 CMSGCAdaptivePolicyCounters* results =
334 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters();
335 assert(
336 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
337 "Wrong gc policy counter kind");
338 return results;
339 }
342 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
344 const char* gen_name = "old";
346 // Generation Counters - generation 1, 1 subspace
347 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space);
349 _space_counters = new GSpaceCounters(gen_name, 0,
350 _virtual_space.reserved_size(),
351 this, _gen_counters);
352 }
354 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
355 _cms_gen(cms_gen)
356 {
357 assert(alpha <= 100, "bad value");
358 _saved_alpha = alpha;
360 // Initialize the alphas to the bootstrap value of 100.
361 _gc0_alpha = _cms_alpha = 100;
363 _cms_begin_time.update();
364 _cms_end_time.update();
366 _gc0_duration = 0.0;
367 _gc0_period = 0.0;
368 _gc0_promoted = 0;
370 _cms_duration = 0.0;
371 _cms_period = 0.0;
372 _cms_allocated = 0;
374 _cms_used_at_gc0_begin = 0;
375 _cms_used_at_gc0_end = 0;
376 _allow_duty_cycle_reduction = false;
377 _valid_bits = 0;
378 _icms_duty_cycle = CMSIncrementalDutyCycle;
379 }
381 double CMSStats::cms_free_adjustment_factor(size_t free) const {
382 // TBD: CR 6909490
383 return 1.0;
384 }
386 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
387 }
389 // If promotion failure handling is on use
390 // the padded average size of the promotion for each
391 // young generation collection.
392 double CMSStats::time_until_cms_gen_full() const {
393 size_t cms_free = _cms_gen->cmsSpace()->free();
394 GenCollectedHeap* gch = GenCollectedHeap::heap();
395 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(),
396 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
397 if (cms_free > expected_promotion) {
398 // Start a cms collection if there isn't enough space to promote
399 // for the next minor collection. Use the padded average as
400 // a safety factor.
401 cms_free -= expected_promotion;
403 // Adjust by the safety factor.
404 double cms_free_dbl = (double)cms_free;
405 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
406 // Apply a further correction factor which tries to adjust
407 // for recent occurance of concurrent mode failures.
408 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
409 cms_free_dbl = cms_free_dbl * cms_adjustment;
411 if (PrintGCDetails && Verbose) {
412 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
413 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
414 cms_free, expected_promotion);
415 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
416 cms_free_dbl, cms_consumption_rate() + 1.0);
417 }
418 // Add 1 in case the consumption rate goes to zero.
419 return cms_free_dbl / (cms_consumption_rate() + 1.0);
420 }
421 return 0.0;
422 }
424 // Compare the duration of the cms collection to the
425 // time remaining before the cms generation is empty.
426 // Note that the time from the start of the cms collection
427 // to the start of the cms sweep (less than the total
428 // duration of the cms collection) can be used. This
429 // has been tried and some applications experienced
430 // promotion failures early in execution. This was
431 // possibly because the averages were not accurate
432 // enough at the beginning.
433 double CMSStats::time_until_cms_start() const {
434 // We add "gc0_period" to the "work" calculation
435 // below because this query is done (mostly) at the
436 // end of a scavenge, so we need to conservatively
437 // account for that much possible delay
438 // in the query so as to avoid concurrent mode failures
439 // due to starting the collection just a wee bit too
440 // late.
441 double work = cms_duration() + gc0_period();
442 double deadline = time_until_cms_gen_full();
443 // If a concurrent mode failure occurred recently, we want to be
444 // more conservative and halve our expected time_until_cms_gen_full()
445 if (work > deadline) {
446 if (Verbose && PrintGCDetails) {
447 gclog_or_tty->print(
448 " CMSCollector: collect because of anticipated promotion "
449 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
450 gc0_period(), time_until_cms_gen_full());
451 }
452 return 0.0;
453 }
454 return work - deadline;
455 }
457 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the
458 // amount of change to prevent wild oscillation.
459 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle,
460 unsigned int new_duty_cycle) {
461 assert(old_duty_cycle <= 100, "bad input value");
462 assert(new_duty_cycle <= 100, "bad input value");
464 // Note: use subtraction with caution since it may underflow (values are
465 // unsigned). Addition is safe since we're in the range 0-100.
466 unsigned int damped_duty_cycle = new_duty_cycle;
467 if (new_duty_cycle < old_duty_cycle) {
468 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U);
469 if (new_duty_cycle + largest_delta < old_duty_cycle) {
470 damped_duty_cycle = old_duty_cycle - largest_delta;
471 }
472 } else if (new_duty_cycle > old_duty_cycle) {
473 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U);
474 if (new_duty_cycle > old_duty_cycle + largest_delta) {
475 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U);
476 }
477 }
478 assert(damped_duty_cycle <= 100, "invalid duty cycle computed");
480 if (CMSTraceIncrementalPacing) {
481 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ",
482 old_duty_cycle, new_duty_cycle, damped_duty_cycle);
483 }
484 return damped_duty_cycle;
485 }
487 unsigned int CMSStats::icms_update_duty_cycle_impl() {
488 assert(CMSIncrementalPacing && valid(),
489 "should be handled in icms_update_duty_cycle()");
491 double cms_time_so_far = cms_timer().seconds();
492 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M;
493 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far);
495 // Avoid division by 0.
496 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01);
497 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full;
499 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U);
500 if (new_duty_cycle > _icms_duty_cycle) {
501 // Avoid very small duty cycles (1 or 2); 0 is allowed.
502 if (new_duty_cycle > 2) {
503 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle,
504 new_duty_cycle);
505 }
506 } else if (_allow_duty_cycle_reduction) {
507 // The duty cycle is reduced only once per cms cycle (see record_cms_end()).
508 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle);
509 // Respect the minimum duty cycle.
510 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin;
511 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle);
512 }
514 if (PrintGCDetails || CMSTraceIncrementalPacing) {
515 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle);
516 }
518 _allow_duty_cycle_reduction = false;
519 return _icms_duty_cycle;
520 }
522 #ifndef PRODUCT
523 void CMSStats::print_on(outputStream *st) const {
524 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
525 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
526 gc0_duration(), gc0_period(), gc0_promoted());
527 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
528 cms_duration(), cms_duration_per_mb(),
529 cms_period(), cms_allocated());
530 st->print(",cms_since_beg=%g,cms_since_end=%g",
531 cms_time_since_begin(), cms_time_since_end());
532 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
533 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
534 if (CMSIncrementalMode) {
535 st->print(",dc=%d", icms_duty_cycle());
536 }
538 if (valid()) {
539 st->print(",promo_rate=%g,cms_alloc_rate=%g",
540 promotion_rate(), cms_allocation_rate());
541 st->print(",cms_consumption_rate=%g,time_until_full=%g",
542 cms_consumption_rate(), time_until_cms_gen_full());
543 }
544 st->print(" ");
545 }
546 #endif // #ifndef PRODUCT
548 CMSCollector::CollectorState CMSCollector::_collectorState =
549 CMSCollector::Idling;
550 bool CMSCollector::_foregroundGCIsActive = false;
551 bool CMSCollector::_foregroundGCShouldWait = false;
553 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
554 CardTableRS* ct,
555 ConcurrentMarkSweepPolicy* cp):
556 _cmsGen(cmsGen),
557 _ct(ct),
558 _ref_processor(NULL), // will be set later
559 _conc_workers(NULL), // may be set later
560 _abort_preclean(false),
561 _start_sampling(false),
562 _between_prologue_and_epilogue(false),
563 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
564 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
565 -1 /* lock-free */, "No_lock" /* dummy */),
566 _modUnionClosure(&_modUnionTable),
567 _modUnionClosurePar(&_modUnionTable),
568 // Adjust my span to cover old (cms) gen
569 _span(cmsGen->reserved()),
570 // Construct the is_alive_closure with _span & markBitMap
571 _is_alive_closure(_span, &_markBitMap),
572 _restart_addr(NULL),
573 _overflow_list(NULL),
574 _stats(cmsGen),
575 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true)),
576 _eden_chunk_array(NULL), // may be set in ctor body
577 _eden_chunk_capacity(0), // -- ditto --
578 _eden_chunk_index(0), // -- ditto --
579 _survivor_plab_array(NULL), // -- ditto --
580 _survivor_chunk_array(NULL), // -- ditto --
581 _survivor_chunk_capacity(0), // -- ditto --
582 _survivor_chunk_index(0), // -- ditto --
583 _ser_pmc_preclean_ovflw(0),
584 _ser_kac_preclean_ovflw(0),
585 _ser_pmc_remark_ovflw(0),
586 _par_pmc_remark_ovflw(0),
587 _ser_kac_ovflw(0),
588 _par_kac_ovflw(0),
589 #ifndef PRODUCT
590 _num_par_pushes(0),
591 #endif
592 _collection_count_start(0),
593 _verifying(false),
594 _icms_start_limit(NULL),
595 _icms_stop_limit(NULL),
596 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
597 _completed_initialization(false),
598 _collector_policy(cp),
599 _should_unload_classes(CMSClassUnloadingEnabled),
600 _concurrent_cycles_since_last_unload(0),
601 _roots_scanning_options(SharedHeap::SO_None),
602 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
603 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
604 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
605 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
606 _cms_start_registered(false)
607 {
608 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
609 ExplicitGCInvokesConcurrent = true;
610 }
611 // Now expand the span and allocate the collection support structures
612 // (MUT, marking bit map etc.) to cover both generations subject to
613 // collection.
615 // For use by dirty card to oop closures.
616 _cmsGen->cmsSpace()->set_collector(this);
618 // Allocate MUT and marking bit map
619 {
620 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
621 if (!_markBitMap.allocate(_span)) {
622 warning("Failed to allocate CMS Bit Map");
623 return;
624 }
625 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
626 }
627 {
628 _modUnionTable.allocate(_span);
629 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
630 }
632 if (!_markStack.allocate(MarkStackSize)) {
633 warning("Failed to allocate CMS Marking Stack");
634 return;
635 }
637 // Support for multi-threaded concurrent phases
638 if (CMSConcurrentMTEnabled) {
639 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
640 // just for now
641 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
642 }
643 if (ConcGCThreads > 1) {
644 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
645 ConcGCThreads, true);
646 if (_conc_workers == NULL) {
647 warning("GC/CMS: _conc_workers allocation failure: "
648 "forcing -CMSConcurrentMTEnabled");
649 CMSConcurrentMTEnabled = false;
650 } else {
651 _conc_workers->initialize_workers();
652 }
653 } else {
654 CMSConcurrentMTEnabled = false;
655 }
656 }
657 if (!CMSConcurrentMTEnabled) {
658 ConcGCThreads = 0;
659 } else {
660 // Turn off CMSCleanOnEnter optimization temporarily for
661 // the MT case where it's not fixed yet; see 6178663.
662 CMSCleanOnEnter = false;
663 }
664 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
665 "Inconsistency");
667 // Parallel task queues; these are shared for the
668 // concurrent and stop-world phases of CMS, but
669 // are not shared with parallel scavenge (ParNew).
670 {
671 uint i;
672 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
674 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
675 || ParallelRefProcEnabled)
676 && num_queues > 0) {
677 _task_queues = new OopTaskQueueSet(num_queues);
678 if (_task_queues == NULL) {
679 warning("task_queues allocation failure.");
680 return;
681 }
682 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
683 if (_hash_seed == NULL) {
684 warning("_hash_seed array allocation failure");
685 return;
686 }
688 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
689 for (i = 0; i < num_queues; i++) {
690 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
691 if (q == NULL) {
692 warning("work_queue allocation failure.");
693 return;
694 }
695 _task_queues->register_queue(i, q);
696 }
697 for (i = 0; i < num_queues; i++) {
698 _task_queues->queue(i)->initialize();
699 _hash_seed[i] = 17; // copied from ParNew
700 }
701 }
702 }
704 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
706 // Clip CMSBootstrapOccupancy between 0 and 100.
707 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100;
709 _full_gcs_since_conc_gc = 0;
711 // Now tell CMS generations the identity of their collector
712 ConcurrentMarkSweepGeneration::set_collector(this);
714 // Create & start a CMS thread for this CMS collector
715 _cmsThread = ConcurrentMarkSweepThread::start(this);
716 assert(cmsThread() != NULL, "CMS Thread should have been created");
717 assert(cmsThread()->collector() == this,
718 "CMS Thread should refer to this gen");
719 assert(CGC_lock != NULL, "Where's the CGC_lock?");
721 // Support for parallelizing young gen rescan
722 GenCollectedHeap* gch = GenCollectedHeap::heap();
723 _young_gen = gch->prev_gen(_cmsGen);
724 if (gch->supports_inline_contig_alloc()) {
725 _top_addr = gch->top_addr();
726 _end_addr = gch->end_addr();
727 assert(_young_gen != NULL, "no _young_gen");
728 _eden_chunk_index = 0;
729 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
730 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
731 if (_eden_chunk_array == NULL) {
732 _eden_chunk_capacity = 0;
733 warning("GC/CMS: _eden_chunk_array allocation failure");
734 }
735 }
736 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
738 // Support for parallelizing survivor space rescan
739 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
740 const size_t max_plab_samples =
741 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize;
743 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
744 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC);
745 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
746 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
747 || _cursor == NULL) {
748 warning("Failed to allocate survivor plab/chunk array");
749 if (_survivor_plab_array != NULL) {
750 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC);
751 _survivor_plab_array = NULL;
752 }
753 if (_survivor_chunk_array != NULL) {
754 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC);
755 _survivor_chunk_array = NULL;
756 }
757 if (_cursor != NULL) {
758 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC);
759 _cursor = NULL;
760 }
761 } else {
762 _survivor_chunk_capacity = 2*max_plab_samples;
763 for (uint i = 0; i < ParallelGCThreads; i++) {
764 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
765 if (vec == NULL) {
766 warning("Failed to allocate survivor plab array");
767 for (int j = i; j > 0; j--) {
768 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC);
769 }
770 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC);
771 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC);
772 _survivor_plab_array = NULL;
773 _survivor_chunk_array = NULL;
774 _survivor_chunk_capacity = 0;
775 break;
776 } else {
777 ChunkArray* cur =
778 ::new (&_survivor_plab_array[i]) ChunkArray(vec,
779 max_plab_samples);
780 assert(cur->end() == 0, "Should be 0");
781 assert(cur->array() == vec, "Should be vec");
782 assert(cur->capacity() == max_plab_samples, "Error");
783 }
784 }
785 }
786 }
787 assert( ( _survivor_plab_array != NULL
788 && _survivor_chunk_array != NULL)
789 || ( _survivor_chunk_capacity == 0
790 && _survivor_chunk_index == 0),
791 "Error");
793 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
794 _gc_counters = new CollectorCounters("CMS", 1);
795 _completed_initialization = true;
796 _inter_sweep_timer.start(); // start of time
797 }
799 const char* ConcurrentMarkSweepGeneration::name() const {
800 return "concurrent mark-sweep generation";
801 }
802 void ConcurrentMarkSweepGeneration::update_counters() {
803 if (UsePerfData) {
804 _space_counters->update_all();
805 _gen_counters->update_all();
806 }
807 }
809 // this is an optimized version of update_counters(). it takes the
810 // used value as a parameter rather than computing it.
811 //
812 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
813 if (UsePerfData) {
814 _space_counters->update_used(used);
815 _space_counters->update_capacity();
816 _gen_counters->update_all();
817 }
818 }
820 void ConcurrentMarkSweepGeneration::print() const {
821 Generation::print();
822 cmsSpace()->print();
823 }
825 #ifndef PRODUCT
826 void ConcurrentMarkSweepGeneration::print_statistics() {
827 cmsSpace()->printFLCensus(0);
828 }
829 #endif
831 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
832 GenCollectedHeap* gch = GenCollectedHeap::heap();
833 if (PrintGCDetails) {
834 if (Verbose) {
835 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
836 level(), short_name(), s, used(), capacity());
837 } else {
838 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
839 level(), short_name(), s, used() / K, capacity() / K);
840 }
841 }
842 if (Verbose) {
843 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
844 gch->used(), gch->capacity());
845 } else {
846 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
847 gch->used() / K, gch->capacity() / K);
848 }
849 }
851 size_t
852 ConcurrentMarkSweepGeneration::contiguous_available() const {
853 // dld proposes an improvement in precision here. If the committed
854 // part of the space ends in a free block we should add that to
855 // uncommitted size in the calculation below. Will make this
856 // change later, staying with the approximation below for the
857 // time being. -- ysr.
858 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
859 }
861 size_t
862 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
863 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
864 }
866 size_t ConcurrentMarkSweepGeneration::max_available() const {
867 return free() + _virtual_space.uncommitted_size();
868 }
870 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
871 size_t available = max_available();
872 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
873 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
874 if (Verbose && PrintGCDetails) {
875 gclog_or_tty->print_cr(
876 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
877 "max_promo("SIZE_FORMAT")",
878 res? "":" not", available, res? ">=":"<",
879 av_promo, max_promotion_in_bytes);
880 }
881 return res;
882 }
884 // At a promotion failure dump information on block layout in heap
885 // (cms old generation).
886 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
887 if (CMSDumpAtPromotionFailure) {
888 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
889 }
890 }
892 CompactibleSpace*
893 ConcurrentMarkSweepGeneration::first_compaction_space() const {
894 return _cmsSpace;
895 }
897 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
898 // Clear the promotion information. These pointers can be adjusted
899 // along with all the other pointers into the heap but
900 // compaction is expected to be a rare event with
901 // a heap using cms so don't do it without seeing the need.
902 if (CollectedHeap::use_parallel_gc_threads()) {
903 for (uint i = 0; i < ParallelGCThreads; i++) {
904 _par_gc_thread_states[i]->promo.reset();
905 }
906 }
907 }
909 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
910 blk->do_space(_cmsSpace);
911 }
913 void ConcurrentMarkSweepGeneration::compute_new_size() {
914 assert_locked_or_safepoint(Heap_lock);
916 // If incremental collection failed, we just want to expand
917 // to the limit.
918 if (incremental_collection_failed()) {
919 clear_incremental_collection_failed();
920 grow_to_reserved();
921 return;
922 }
924 // The heap has been compacted but not reset yet.
925 // Any metric such as free() or used() will be incorrect.
927 CardGeneration::compute_new_size();
929 // Reset again after a possible resizing
930 if (did_compact()) {
931 cmsSpace()->reset_after_compaction();
932 }
933 }
935 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
936 assert_locked_or_safepoint(Heap_lock);
938 // If incremental collection failed, we just want to expand
939 // to the limit.
940 if (incremental_collection_failed()) {
941 clear_incremental_collection_failed();
942 grow_to_reserved();
943 return;
944 }
946 double free_percentage = ((double) free()) / capacity();
947 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
948 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
950 // compute expansion delta needed for reaching desired free percentage
951 if (free_percentage < desired_free_percentage) {
952 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
953 assert(desired_capacity >= capacity(), "invalid expansion size");
954 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
955 if (PrintGCDetails && Verbose) {
956 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
957 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
958 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
959 gclog_or_tty->print_cr(" Desired free fraction %f",
960 desired_free_percentage);
961 gclog_or_tty->print_cr(" Maximum free fraction %f",
962 maximum_free_percentage);
963 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000);
964 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
965 desired_capacity/1000);
966 int prev_level = level() - 1;
967 if (prev_level >= 0) {
968 size_t prev_size = 0;
969 GenCollectedHeap* gch = GenCollectedHeap::heap();
970 Generation* prev_gen = gch->_gens[prev_level];
971 prev_size = prev_gen->capacity();
972 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
973 prev_size/1000);
974 }
975 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
976 unsafe_max_alloc_nogc()/1000);
977 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
978 contiguous_available()/1000);
979 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
980 expand_bytes);
981 }
982 // safe if expansion fails
983 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
984 if (PrintGCDetails && Verbose) {
985 gclog_or_tty->print_cr(" Expanded free fraction %f",
986 ((double) free()) / capacity());
987 }
988 } else {
989 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
990 assert(desired_capacity <= capacity(), "invalid expansion size");
991 size_t shrink_bytes = capacity() - desired_capacity;
992 // Don't shrink unless the delta is greater than the minimum shrink we want
993 if (shrink_bytes >= MinHeapDeltaBytes) {
994 shrink_free_list_by(shrink_bytes);
995 }
996 }
997 }
999 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
1000 return cmsSpace()->freelistLock();
1001 }
1003 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
1004 bool tlab) {
1005 CMSSynchronousYieldRequest yr;
1006 MutexLockerEx x(freelistLock(),
1007 Mutex::_no_safepoint_check_flag);
1008 return have_lock_and_allocate(size, tlab);
1009 }
1011 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
1012 bool tlab /* ignored */) {
1013 assert_lock_strong(freelistLock());
1014 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
1015 HeapWord* res = cmsSpace()->allocate(adjustedSize);
1016 // Allocate the object live (grey) if the background collector has
1017 // started marking. This is necessary because the marker may
1018 // have passed this address and consequently this object will
1019 // not otherwise be greyed and would be incorrectly swept up.
1020 // Note that if this object contains references, the writing
1021 // of those references will dirty the card containing this object
1022 // allowing the object to be blackened (and its references scanned)
1023 // either during a preclean phase or at the final checkpoint.
1024 if (res != NULL) {
1025 // We may block here with an uninitialized object with
1026 // its mark-bit or P-bits not yet set. Such objects need
1027 // to be safely navigable by block_start().
1028 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
1029 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
1030 collector()->direct_allocated(res, adjustedSize);
1031 _direct_allocated_words += adjustedSize;
1032 // allocation counters
1033 NOT_PRODUCT(
1034 _numObjectsAllocated++;
1035 _numWordsAllocated += (int)adjustedSize;
1036 )
1037 }
1038 return res;
1039 }
1041 // In the case of direct allocation by mutators in a generation that
1042 // is being concurrently collected, the object must be allocated
1043 // live (grey) if the background collector has started marking.
1044 // This is necessary because the marker may
1045 // have passed this address and consequently this object will
1046 // not otherwise be greyed and would be incorrectly swept up.
1047 // Note that if this object contains references, the writing
1048 // of those references will dirty the card containing this object
1049 // allowing the object to be blackened (and its references scanned)
1050 // either during a preclean phase or at the final checkpoint.
1051 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
1052 assert(_markBitMap.covers(start, size), "Out of bounds");
1053 if (_collectorState >= Marking) {
1054 MutexLockerEx y(_markBitMap.lock(),
1055 Mutex::_no_safepoint_check_flag);
1056 // [see comments preceding SweepClosure::do_blk() below for details]
1057 //
1058 // Can the P-bits be deleted now? JJJ
1059 //
1060 // 1. need to mark the object as live so it isn't collected
1061 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
1062 // 3. need to mark the end of the object so marking, precleaning or sweeping
1063 // can skip over uninitialized or unparsable objects. An allocated
1064 // object is considered uninitialized for our purposes as long as
1065 // its klass word is NULL. All old gen objects are parsable
1066 // as soon as they are initialized.)
1067 _markBitMap.mark(start); // object is live
1068 _markBitMap.mark(start + 1); // object is potentially uninitialized?
1069 _markBitMap.mark(start + size - 1);
1070 // mark end of object
1071 }
1072 // check that oop looks uninitialized
1073 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
1074 }
1076 void CMSCollector::promoted(bool par, HeapWord* start,
1077 bool is_obj_array, size_t obj_size) {
1078 assert(_markBitMap.covers(start), "Out of bounds");
1079 // See comment in direct_allocated() about when objects should
1080 // be allocated live.
1081 if (_collectorState >= Marking) {
1082 // we already hold the marking bit map lock, taken in
1083 // the prologue
1084 if (par) {
1085 _markBitMap.par_mark(start);
1086 } else {
1087 _markBitMap.mark(start);
1088 }
1089 // We don't need to mark the object as uninitialized (as
1090 // in direct_allocated above) because this is being done with the
1091 // world stopped and the object will be initialized by the
1092 // time the marking, precleaning or sweeping get to look at it.
1093 // But see the code for copying objects into the CMS generation,
1094 // where we need to ensure that concurrent readers of the
1095 // block offset table are able to safely navigate a block that
1096 // is in flux from being free to being allocated (and in
1097 // transition while being copied into) and subsequently
1098 // becoming a bona-fide object when the copy/promotion is complete.
1099 assert(SafepointSynchronize::is_at_safepoint(),
1100 "expect promotion only at safepoints");
1102 if (_collectorState < Sweeping) {
1103 // Mark the appropriate cards in the modUnionTable, so that
1104 // this object gets scanned before the sweep. If this is
1105 // not done, CMS generation references in the object might
1106 // not get marked.
1107 // For the case of arrays, which are otherwise precisely
1108 // marked, we need to dirty the entire array, not just its head.
1109 if (is_obj_array) {
1110 // The [par_]mark_range() method expects mr.end() below to
1111 // be aligned to the granularity of a bit's representation
1112 // in the heap. In the case of the MUT below, that's a
1113 // card size.
1114 MemRegion mr(start,
1115 (HeapWord*)round_to((intptr_t)(start + obj_size),
1116 CardTableModRefBS::card_size /* bytes */));
1117 if (par) {
1118 _modUnionTable.par_mark_range(mr);
1119 } else {
1120 _modUnionTable.mark_range(mr);
1121 }
1122 } else { // not an obj array; we can just mark the head
1123 if (par) {
1124 _modUnionTable.par_mark(start);
1125 } else {
1126 _modUnionTable.mark(start);
1127 }
1128 }
1129 }
1130 }
1131 }
1133 static inline size_t percent_of_space(Space* space, HeapWord* addr)
1134 {
1135 size_t delta = pointer_delta(addr, space->bottom());
1136 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
1137 }
1139 void CMSCollector::icms_update_allocation_limits()
1140 {
1141 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
1142 EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
1144 const unsigned int duty_cycle = stats().icms_update_duty_cycle();
1145 if (CMSTraceIncrementalPacing) {
1146 stats().print();
1147 }
1149 assert(duty_cycle <= 100, "invalid duty cycle");
1150 if (duty_cycle != 0) {
1151 // The duty_cycle is a percentage between 0 and 100; convert to words and
1152 // then compute the offset from the endpoints of the space.
1153 size_t free_words = eden->free() / HeapWordSize;
1154 double free_words_dbl = (double)free_words;
1155 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
1156 size_t offset_words = (free_words - duty_cycle_words) / 2;
1158 _icms_start_limit = eden->top() + offset_words;
1159 _icms_stop_limit = eden->end() - offset_words;
1161 // The limits may be adjusted (shifted to the right) by
1162 // CMSIncrementalOffset, to allow the application more mutator time after a
1163 // young gen gc (when all mutators were stopped) and before CMS starts and
1164 // takes away one or more cpus.
1165 if (CMSIncrementalOffset != 0) {
1166 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
1167 size_t adjustment = (size_t)adjustment_dbl;
1168 HeapWord* tmp_stop = _icms_stop_limit + adjustment;
1169 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
1170 _icms_start_limit += adjustment;
1171 _icms_stop_limit = tmp_stop;
1172 }
1173 }
1174 }
1175 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
1176 _icms_start_limit = _icms_stop_limit = eden->end();
1177 }
1179 // Install the new start limit.
1180 eden->set_soft_end(_icms_start_limit);
1182 if (CMSTraceIncrementalMode) {
1183 gclog_or_tty->print(" icms alloc limits: "
1184 PTR_FORMAT "," PTR_FORMAT
1185 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
1186 p2i(_icms_start_limit), p2i(_icms_stop_limit),
1187 percent_of_space(eden, _icms_start_limit),
1188 percent_of_space(eden, _icms_stop_limit));
1189 if (Verbose) {
1190 gclog_or_tty->print("eden: ");
1191 eden->print_on(gclog_or_tty);
1192 }
1193 }
1194 }
1196 // Any changes here should try to maintain the invariant
1197 // that if this method is called with _icms_start_limit
1198 // and _icms_stop_limit both NULL, then it should return NULL
1199 // and not notify the icms thread.
1200 HeapWord*
1201 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
1202 size_t word_size)
1203 {
1204 // A start_limit equal to end() means the duty cycle is 0, so treat that as a
1205 // nop.
1206 if (CMSIncrementalMode && _icms_start_limit != space->end()) {
1207 if (top <= _icms_start_limit) {
1208 if (CMSTraceIncrementalMode) {
1209 space->print_on(gclog_or_tty);
1210 gclog_or_tty->stamp();
1211 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
1212 ", new limit=" PTR_FORMAT
1213 " (" SIZE_FORMAT "%%)",
1214 p2i(top), p2i(_icms_stop_limit),
1215 percent_of_space(space, _icms_stop_limit));
1216 }
1217 ConcurrentMarkSweepThread::start_icms();
1218 assert(top < _icms_stop_limit, "Tautology");
1219 if (word_size < pointer_delta(_icms_stop_limit, top)) {
1220 return _icms_stop_limit;
1221 }
1223 // The allocation will cross both the _start and _stop limits, so do the
1224 // stop notification also and return end().
1225 if (CMSTraceIncrementalMode) {
1226 space->print_on(gclog_or_tty);
1227 gclog_or_tty->stamp();
1228 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
1229 ", new limit=" PTR_FORMAT
1230 " (" SIZE_FORMAT "%%)",
1231 p2i(top), p2i(space->end()),
1232 percent_of_space(space, space->end()));
1233 }
1234 ConcurrentMarkSweepThread::stop_icms();
1235 return space->end();
1236 }
1238 if (top <= _icms_stop_limit) {
1239 if (CMSTraceIncrementalMode) {
1240 space->print_on(gclog_or_tty);
1241 gclog_or_tty->stamp();
1242 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
1243 ", new limit=" PTR_FORMAT
1244 " (" SIZE_FORMAT "%%)",
1245 top, space->end(),
1246 percent_of_space(space, space->end()));
1247 }
1248 ConcurrentMarkSweepThread::stop_icms();
1249 return space->end();
1250 }
1252 if (CMSTraceIncrementalMode) {
1253 space->print_on(gclog_or_tty);
1254 gclog_or_tty->stamp();
1255 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
1256 ", new limit=" PTR_FORMAT,
1257 top, NULL);
1258 }
1259 }
1261 return NULL;
1262 }
1264 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
1265 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1266 // allocate, copy and if necessary update promoinfo --
1267 // delegate to underlying space.
1268 assert_lock_strong(freelistLock());
1270 #ifndef PRODUCT
1271 if (Universe::heap()->promotion_should_fail()) {
1272 return NULL;
1273 }
1274 #endif // #ifndef PRODUCT
1276 oop res = _cmsSpace->promote(obj, obj_size);
1277 if (res == NULL) {
1278 // expand and retry
1279 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
1280 expand(s*HeapWordSize, MinHeapDeltaBytes,
1281 CMSExpansionCause::_satisfy_promotion);
1282 // Since there's currently no next generation, we don't try to promote
1283 // into a more senior generation.
1284 assert(next_gen() == NULL, "assumption, based upon which no attempt "
1285 "is made to pass on a possibly failing "
1286 "promotion to next generation");
1287 res = _cmsSpace->promote(obj, obj_size);
1288 }
1289 if (res != NULL) {
1290 // See comment in allocate() about when objects should
1291 // be allocated live.
1292 assert(obj->is_oop(), "Will dereference klass pointer below");
1293 collector()->promoted(false, // Not parallel
1294 (HeapWord*)res, obj->is_objArray(), obj_size);
1295 // promotion counters
1296 NOT_PRODUCT(
1297 _numObjectsPromoted++;
1298 _numWordsPromoted +=
1299 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1300 )
1301 }
1302 return res;
1303 }
1306 HeapWord*
1307 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
1308 HeapWord* top,
1309 size_t word_sz)
1310 {
1311 return collector()->allocation_limit_reached(space, top, word_sz);
1312 }
1314 // IMPORTANT: Notes on object size recognition in CMS.
1315 // ---------------------------------------------------
1316 // A block of storage in the CMS generation is always in
1317 // one of three states. A free block (FREE), an allocated
1318 // object (OBJECT) whose size() method reports the correct size,
1319 // and an intermediate state (TRANSIENT) in which its size cannot
1320 // be accurately determined.
1321 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1322 // -----------------------------------------------------
1323 // FREE: klass_word & 1 == 1; mark_word holds block size
1324 //
1325 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1326 // obj->size() computes correct size
1327 //
1328 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1329 //
1330 // STATE IDENTIFICATION: (64 bit+COOPS)
1331 // ------------------------------------
1332 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1333 //
1334 // OBJECT: klass_word installed; klass_word != 0;
1335 // obj->size() computes correct size
1336 //
1337 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1338 //
1339 //
1340 // STATE TRANSITION DIAGRAM
1341 //
1342 // mut / parnew mut / parnew
1343 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1344 // ^ |
1345 // |------------------------ DEAD <------------------------------------|
1346 // sweep mut
1347 //
1348 // While a block is in TRANSIENT state its size cannot be determined
1349 // so readers will either need to come back later or stall until
1350 // the size can be determined. Note that for the case of direct
1351 // allocation, P-bits, when available, may be used to determine the
1352 // size of an object that may not yet have been initialized.
1354 // Things to support parallel young-gen collection.
1355 oop
1356 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1357 oop old, markOop m,
1358 size_t word_sz) {
1359 #ifndef PRODUCT
1360 if (Universe::heap()->promotion_should_fail()) {
1361 return NULL;
1362 }
1363 #endif // #ifndef PRODUCT
1365 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1366 PromotionInfo* promoInfo = &ps->promo;
1367 // if we are tracking promotions, then first ensure space for
1368 // promotion (including spooling space for saving header if necessary).
1369 // then allocate and copy, then track promoted info if needed.
1370 // When tracking (see PromotionInfo::track()), the mark word may
1371 // be displaced and in this case restoration of the mark word
1372 // occurs in the (oop_since_save_marks_)iterate phase.
1373 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1374 // Out of space for allocating spooling buffers;
1375 // try expanding and allocating spooling buffers.
1376 if (!expand_and_ensure_spooling_space(promoInfo)) {
1377 return NULL;
1378 }
1379 }
1380 assert(promoInfo->has_spooling_space(), "Control point invariant");
1381 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1382 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1383 if (obj_ptr == NULL) {
1384 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1385 if (obj_ptr == NULL) {
1386 return NULL;
1387 }
1388 }
1389 oop obj = oop(obj_ptr);
1390 OrderAccess::storestore();
1391 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1392 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1393 // IMPORTANT: See note on object initialization for CMS above.
1394 // Otherwise, copy the object. Here we must be careful to insert the
1395 // klass pointer last, since this marks the block as an allocated object.
1396 // Except with compressed oops it's the mark word.
1397 HeapWord* old_ptr = (HeapWord*)old;
1398 // Restore the mark word copied above.
1399 obj->set_mark(m);
1400 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1401 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1402 OrderAccess::storestore();
1404 if (UseCompressedClassPointers) {
1405 // Copy gap missed by (aligned) header size calculation below
1406 obj->set_klass_gap(old->klass_gap());
1407 }
1408 if (word_sz > (size_t)oopDesc::header_size()) {
1409 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1410 obj_ptr + oopDesc::header_size(),
1411 word_sz - oopDesc::header_size());
1412 }
1414 // Now we can track the promoted object, if necessary. We take care
1415 // to delay the transition from uninitialized to full object
1416 // (i.e., insertion of klass pointer) until after, so that it
1417 // atomically becomes a promoted object.
1418 if (promoInfo->tracking()) {
1419 promoInfo->track((PromotedObject*)obj, old->klass());
1420 }
1421 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1422 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1423 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1425 // Finally, install the klass pointer (this should be volatile).
1426 OrderAccess::storestore();
1427 obj->set_klass(old->klass());
1428 // We should now be able to calculate the right size for this object
1429 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1431 collector()->promoted(true, // parallel
1432 obj_ptr, old->is_objArray(), word_sz);
1434 NOT_PRODUCT(
1435 Atomic::inc_ptr(&_numObjectsPromoted);
1436 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1437 )
1439 return obj;
1440 }
1442 void
1443 ConcurrentMarkSweepGeneration::
1444 par_promote_alloc_undo(int thread_num,
1445 HeapWord* obj, size_t word_sz) {
1446 // CMS does not support promotion undo.
1447 ShouldNotReachHere();
1448 }
1450 void
1451 ConcurrentMarkSweepGeneration::
1452 par_promote_alloc_done(int thread_num) {
1453 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1454 ps->lab.retire(thread_num);
1455 }
1457 void
1458 ConcurrentMarkSweepGeneration::
1459 par_oop_since_save_marks_iterate_done(int thread_num) {
1460 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1461 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1462 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1463 }
1465 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1466 size_t size,
1467 bool tlab)
1468 {
1469 // We allow a STW collection only if a full
1470 // collection was requested.
1471 return full || should_allocate(size, tlab); // FIX ME !!!
1472 // This and promotion failure handling are connected at the
1473 // hip and should be fixed by untying them.
1474 }
1476 bool CMSCollector::shouldConcurrentCollect() {
1477 if (_full_gc_requested) {
1478 if (Verbose && PrintGCDetails) {
1479 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1480 " gc request (or gc_locker)");
1481 }
1482 return true;
1483 }
1485 // For debugging purposes, change the type of collection.
1486 // If the rotation is not on the concurrent collection
1487 // type, don't start a concurrent collection.
1488 NOT_PRODUCT(
1489 if (RotateCMSCollectionTypes &&
1490 (_cmsGen->debug_collection_type() !=
1491 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
1492 assert(_cmsGen->debug_collection_type() !=
1493 ConcurrentMarkSweepGeneration::Unknown_collection_type,
1494 "Bad cms collection type");
1495 return false;
1496 }
1497 )
1499 FreelistLocker x(this);
1500 // ------------------------------------------------------------------
1501 // Print out lots of information which affects the initiation of
1502 // a collection.
1503 if (PrintCMSInitiationStatistics && stats().valid()) {
1504 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1505 gclog_or_tty->stamp();
1506 gclog_or_tty->cr();
1507 stats().print_on(gclog_or_tty);
1508 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1509 stats().time_until_cms_gen_full());
1510 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1511 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1512 _cmsGen->contiguous_available());
1513 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1514 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1515 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1516 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1517 gclog_or_tty->print_cr("metadata initialized %d",
1518 MetaspaceGC::should_concurrent_collect());
1519 }
1520 // ------------------------------------------------------------------
1522 // If the estimated time to complete a cms collection (cms_duration())
1523 // is less than the estimated time remaining until the cms generation
1524 // is full, start a collection.
1525 if (!UseCMSInitiatingOccupancyOnly) {
1526 if (stats().valid()) {
1527 if (stats().time_until_cms_start() == 0.0) {
1528 return true;
1529 }
1530 } else {
1531 // We want to conservatively collect somewhat early in order
1532 // to try and "bootstrap" our CMS/promotion statistics;
1533 // this branch will not fire after the first successful CMS
1534 // collection because the stats should then be valid.
1535 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1536 if (Verbose && PrintGCDetails) {
1537 gclog_or_tty->print_cr(
1538 " CMSCollector: collect for bootstrapping statistics:"
1539 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1540 _bootstrap_occupancy);
1541 }
1542 return true;
1543 }
1544 }
1545 }
1547 // Otherwise, we start a collection cycle if
1548 // old gen want a collection cycle started. Each may use
1549 // an appropriate criterion for making this decision.
1550 // XXX We need to make sure that the gen expansion
1551 // criterion dovetails well with this. XXX NEED TO FIX THIS
1552 if (_cmsGen->should_concurrent_collect()) {
1553 if (Verbose && PrintGCDetails) {
1554 gclog_or_tty->print_cr("CMS old gen initiated");
1555 }
1556 return true;
1557 }
1559 // We start a collection if we believe an incremental collection may fail;
1560 // this is not likely to be productive in practice because it's probably too
1561 // late anyway.
1562 GenCollectedHeap* gch = GenCollectedHeap::heap();
1563 assert(gch->collector_policy()->is_two_generation_policy(),
1564 "You may want to check the correctness of the following");
1565 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1566 if (Verbose && PrintGCDetails) {
1567 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1568 }
1569 return true;
1570 }
1572 if (MetaspaceGC::should_concurrent_collect()) {
1573 if (Verbose && PrintGCDetails) {
1574 gclog_or_tty->print("CMSCollector: collect for metadata allocation ");
1575 }
1576 return true;
1577 }
1579 return false;
1580 }
1582 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1584 // Clear _expansion_cause fields of constituent generations
1585 void CMSCollector::clear_expansion_cause() {
1586 _cmsGen->clear_expansion_cause();
1587 }
1589 // We should be conservative in starting a collection cycle. To
1590 // start too eagerly runs the risk of collecting too often in the
1591 // extreme. To collect too rarely falls back on full collections,
1592 // which works, even if not optimum in terms of concurrent work.
1593 // As a work around for too eagerly collecting, use the flag
1594 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1595 // giving the user an easily understandable way of controlling the
1596 // collections.
1597 // We want to start a new collection cycle if any of the following
1598 // conditions hold:
1599 // . our current occupancy exceeds the configured initiating occupancy
1600 // for this generation, or
1601 // . we recently needed to expand this space and have not, since that
1602 // expansion, done a collection of this generation, or
1603 // . the underlying space believes that it may be a good idea to initiate
1604 // a concurrent collection (this may be based on criteria such as the
1605 // following: the space uses linear allocation and linear allocation is
1606 // going to fail, or there is believed to be excessive fragmentation in
1607 // the generation, etc... or ...
1608 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1609 // the case of the old generation; see CR 6543076):
1610 // we may be approaching a point at which allocation requests may fail because
1611 // we will be out of sufficient free space given allocation rate estimates.]
1612 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1614 assert_lock_strong(freelistLock());
1615 if (occupancy() > initiating_occupancy()) {
1616 if (PrintGCDetails && Verbose) {
1617 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1618 short_name(), occupancy(), initiating_occupancy());
1619 }
1620 return true;
1621 }
1622 if (UseCMSInitiatingOccupancyOnly) {
1623 return false;
1624 }
1625 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1626 if (PrintGCDetails && Verbose) {
1627 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1628 short_name());
1629 }
1630 return true;
1631 }
1632 if (_cmsSpace->should_concurrent_collect()) {
1633 if (PrintGCDetails && Verbose) {
1634 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1635 short_name());
1636 }
1637 return true;
1638 }
1639 return false;
1640 }
1642 void ConcurrentMarkSweepGeneration::collect(bool full,
1643 bool clear_all_soft_refs,
1644 size_t size,
1645 bool tlab)
1646 {
1647 collector()->collect(full, clear_all_soft_refs, size, tlab);
1648 }
1650 void CMSCollector::collect(bool full,
1651 bool clear_all_soft_refs,
1652 size_t size,
1653 bool tlab)
1654 {
1655 if (!UseCMSCollectionPassing && _collectorState > Idling) {
1656 // For debugging purposes skip the collection if the state
1657 // is not currently idle
1658 if (TraceCMSState) {
1659 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
1660 Thread::current(), full, _collectorState);
1661 }
1662 return;
1663 }
1665 // The following "if" branch is present for defensive reasons.
1666 // In the current uses of this interface, it can be replaced with:
1667 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1668 // But I am not placing that assert here to allow future
1669 // generality in invoking this interface.
1670 if (GC_locker::is_active()) {
1671 // A consistency test for GC_locker
1672 assert(GC_locker::needs_gc(), "Should have been set already");
1673 // Skip this foreground collection, instead
1674 // expanding the heap if necessary.
1675 // Need the free list locks for the call to free() in compute_new_size()
1676 compute_new_size();
1677 return;
1678 }
1679 acquire_control_and_collect(full, clear_all_soft_refs);
1680 _full_gcs_since_conc_gc++;
1681 }
1683 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1684 GenCollectedHeap* gch = GenCollectedHeap::heap();
1685 unsigned int gc_count = gch->total_full_collections();
1686 if (gc_count == full_gc_count) {
1687 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1688 _full_gc_requested = true;
1689 _full_gc_cause = cause;
1690 CGC_lock->notify(); // nudge CMS thread
1691 } else {
1692 assert(gc_count > full_gc_count, "Error: causal loop");
1693 }
1694 }
1696 bool CMSCollector::is_external_interruption() {
1697 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1698 return GCCause::is_user_requested_gc(cause) ||
1699 GCCause::is_serviceability_requested_gc(cause);
1700 }
1702 void CMSCollector::report_concurrent_mode_interruption() {
1703 if (is_external_interruption()) {
1704 if (PrintGCDetails) {
1705 gclog_or_tty->print(" (concurrent mode interrupted)");
1706 }
1707 } else {
1708 if (PrintGCDetails) {
1709 gclog_or_tty->print(" (concurrent mode failure)");
1710 }
1711 _gc_tracer_cm->report_concurrent_mode_failure();
1712 }
1713 }
1716 // The foreground and background collectors need to coordinate in order
1717 // to make sure that they do not mutually interfere with CMS collections.
1718 // When a background collection is active,
1719 // the foreground collector may need to take over (preempt) and
1720 // synchronously complete an ongoing collection. Depending on the
1721 // frequency of the background collections and the heap usage
1722 // of the application, this preemption can be seldom or frequent.
1723 // There are only certain
1724 // points in the background collection that the "collection-baton"
1725 // can be passed to the foreground collector.
1726 //
1727 // The foreground collector will wait for the baton before
1728 // starting any part of the collection. The foreground collector
1729 // will only wait at one location.
1730 //
1731 // The background collector will yield the baton before starting a new
1732 // phase of the collection (e.g., before initial marking, marking from roots,
1733 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1734 // of the loop which switches the phases. The background collector does some
1735 // of the phases (initial mark, final re-mark) with the world stopped.
1736 // Because of locking involved in stopping the world,
1737 // the foreground collector should not block waiting for the background
1738 // collector when it is doing a stop-the-world phase. The background
1739 // collector will yield the baton at an additional point just before
1740 // it enters a stop-the-world phase. Once the world is stopped, the
1741 // background collector checks the phase of the collection. If the
1742 // phase has not changed, it proceeds with the collection. If the
1743 // phase has changed, it skips that phase of the collection. See
1744 // the comments on the use of the Heap_lock in collect_in_background().
1745 //
1746 // Variable used in baton passing.
1747 // _foregroundGCIsActive - Set to true by the foreground collector when
1748 // it wants the baton. The foreground clears it when it has finished
1749 // the collection.
1750 // _foregroundGCShouldWait - Set to true by the background collector
1751 // when it is running. The foreground collector waits while
1752 // _foregroundGCShouldWait is true.
1753 // CGC_lock - monitor used to protect access to the above variables
1754 // and to notify the foreground and background collectors.
1755 // _collectorState - current state of the CMS collection.
1756 //
1757 // The foreground collector
1758 // acquires the CGC_lock
1759 // sets _foregroundGCIsActive
1760 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1761 // various locks acquired in preparation for the collection
1762 // are released so as not to block the background collector
1763 // that is in the midst of a collection
1764 // proceeds with the collection
1765 // clears _foregroundGCIsActive
1766 // returns
1767 //
1768 // The background collector in a loop iterating on the phases of the
1769 // collection
1770 // acquires the CGC_lock
1771 // sets _foregroundGCShouldWait
1772 // if _foregroundGCIsActive is set
1773 // clears _foregroundGCShouldWait, notifies _CGC_lock
1774 // waits on _CGC_lock for _foregroundGCIsActive to become false
1775 // and exits the loop.
1776 // otherwise
1777 // proceed with that phase of the collection
1778 // if the phase is a stop-the-world phase,
1779 // yield the baton once more just before enqueueing
1780 // the stop-world CMS operation (executed by the VM thread).
1781 // returns after all phases of the collection are done
1782 //
1784 void CMSCollector::acquire_control_and_collect(bool full,
1785 bool clear_all_soft_refs) {
1786 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1787 assert(!Thread::current()->is_ConcurrentGC_thread(),
1788 "shouldn't try to acquire control from self!");
1790 // Start the protocol for acquiring control of the
1791 // collection from the background collector (aka CMS thread).
1792 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1793 "VM thread should have CMS token");
1794 // Remember the possibly interrupted state of an ongoing
1795 // concurrent collection
1796 CollectorState first_state = _collectorState;
1798 // Signal to a possibly ongoing concurrent collection that
1799 // we want to do a foreground collection.
1800 _foregroundGCIsActive = true;
1802 // Disable incremental mode during a foreground collection.
1803 ICMSDisabler icms_disabler;
1805 // release locks and wait for a notify from the background collector
1806 // releasing the locks in only necessary for phases which
1807 // do yields to improve the granularity of the collection.
1808 assert_lock_strong(bitMapLock());
1809 // We need to lock the Free list lock for the space that we are
1810 // currently collecting.
1811 assert(haveFreelistLocks(), "Must be holding free list locks");
1812 bitMapLock()->unlock();
1813 releaseFreelistLocks();
1814 {
1815 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1816 if (_foregroundGCShouldWait) {
1817 // We are going to be waiting for action for the CMS thread;
1818 // it had better not be gone (for instance at shutdown)!
1819 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1820 "CMS thread must be running");
1821 // Wait here until the background collector gives us the go-ahead
1822 ConcurrentMarkSweepThread::clear_CMS_flag(
1823 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1824 // Get a possibly blocked CMS thread going:
1825 // Note that we set _foregroundGCIsActive true above,
1826 // without protection of the CGC_lock.
1827 CGC_lock->notify();
1828 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1829 "Possible deadlock");
1830 while (_foregroundGCShouldWait) {
1831 // wait for notification
1832 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1833 // Possibility of delay/starvation here, since CMS token does
1834 // not know to give priority to VM thread? Actually, i think
1835 // there wouldn't be any delay/starvation, but the proof of
1836 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1837 }
1838 ConcurrentMarkSweepThread::set_CMS_flag(
1839 ConcurrentMarkSweepThread::CMS_vm_has_token);
1840 }
1841 }
1842 // The CMS_token is already held. Get back the other locks.
1843 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1844 "VM thread should have CMS token");
1845 getFreelistLocks();
1846 bitMapLock()->lock_without_safepoint_check();
1847 if (TraceCMSState) {
1848 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1849 INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1850 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1851 }
1853 // Check if we need to do a compaction, or if not, whether
1854 // we need to start the mark-sweep from scratch.
1855 bool should_compact = false;
1856 bool should_start_over = false;
1857 decide_foreground_collection_type(clear_all_soft_refs,
1858 &should_compact, &should_start_over);
1860 NOT_PRODUCT(
1861 if (RotateCMSCollectionTypes) {
1862 if (_cmsGen->debug_collection_type() ==
1863 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1864 should_compact = true;
1865 } else if (_cmsGen->debug_collection_type() ==
1866 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1867 should_compact = false;
1868 }
1869 }
1870 )
1872 if (first_state > Idling) {
1873 report_concurrent_mode_interruption();
1874 }
1876 set_did_compact(should_compact);
1877 if (should_compact) {
1878 // If the collection is being acquired from the background
1879 // collector, there may be references on the discovered
1880 // references lists that have NULL referents (being those
1881 // that were concurrently cleared by a mutator) or
1882 // that are no longer active (having been enqueued concurrently
1883 // by the mutator).
1884 // Scrub the list of those references because Mark-Sweep-Compact
1885 // code assumes referents are not NULL and that all discovered
1886 // Reference objects are active.
1887 ref_processor()->clean_up_discovered_references();
1889 if (first_state > Idling) {
1890 save_heap_summary();
1891 }
1893 do_compaction_work(clear_all_soft_refs);
1895 // Has the GC time limit been exceeded?
1896 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration();
1897 size_t max_eden_size = young_gen->max_capacity() -
1898 young_gen->to()->capacity() -
1899 young_gen->from()->capacity();
1900 GenCollectedHeap* gch = GenCollectedHeap::heap();
1901 GCCause::Cause gc_cause = gch->gc_cause();
1902 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1903 young_gen->eden()->used(),
1904 _cmsGen->max_capacity(),
1905 max_eden_size,
1906 full,
1907 gc_cause,
1908 gch->collector_policy());
1909 } else {
1910 do_mark_sweep_work(clear_all_soft_refs, first_state,
1911 should_start_over);
1912 }
1913 // Reset the expansion cause, now that we just completed
1914 // a collection cycle.
1915 clear_expansion_cause();
1916 _foregroundGCIsActive = false;
1917 return;
1918 }
1920 // Resize the tenured generation
1921 // after obtaining the free list locks for the
1922 // two generations.
1923 void CMSCollector::compute_new_size() {
1924 assert_locked_or_safepoint(Heap_lock);
1925 FreelistLocker z(this);
1926 MetaspaceGC::compute_new_size();
1927 _cmsGen->compute_new_size_free_list();
1928 }
1930 // A work method used by foreground collection to determine
1931 // what type of collection (compacting or not, continuing or fresh)
1932 // it should do.
1933 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1934 // and CMSCompactWhenClearAllSoftRefs the default in the future
1935 // and do away with the flags after a suitable period.
1936 void CMSCollector::decide_foreground_collection_type(
1937 bool clear_all_soft_refs, bool* should_compact,
1938 bool* should_start_over) {
1939 // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1940 // flag is set, and we have either requested a System.gc() or
1941 // the number of full gc's since the last concurrent cycle
1942 // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1943 // or if an incremental collection has failed
1944 GenCollectedHeap* gch = GenCollectedHeap::heap();
1945 assert(gch->collector_policy()->is_two_generation_policy(),
1946 "You may want to check the correctness of the following");
1947 // Inform cms gen if this was due to partial collection failing.
1948 // The CMS gen may use this fact to determine its expansion policy.
1949 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1950 assert(!_cmsGen->incremental_collection_failed(),
1951 "Should have been noticed, reacted to and cleared");
1952 _cmsGen->set_incremental_collection_failed();
1953 }
1954 *should_compact =
1955 UseCMSCompactAtFullCollection &&
1956 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1957 GCCause::is_user_requested_gc(gch->gc_cause()) ||
1958 gch->incremental_collection_will_fail(true /* consult_young */));
1959 *should_start_over = false;
1960 if (clear_all_soft_refs && !*should_compact) {
1961 // We are about to do a last ditch collection attempt
1962 // so it would normally make sense to do a compaction
1963 // to reclaim as much space as possible.
1964 if (CMSCompactWhenClearAllSoftRefs) {
1965 // Default: The rationale is that in this case either
1966 // we are past the final marking phase, in which case
1967 // we'd have to start over, or so little has been done
1968 // that there's little point in saving that work. Compaction
1969 // appears to be the sensible choice in either case.
1970 *should_compact = true;
1971 } else {
1972 // We have been asked to clear all soft refs, but not to
1973 // compact. Make sure that we aren't past the final checkpoint
1974 // phase, for that is where we process soft refs. If we are already
1975 // past that phase, we'll need to redo the refs discovery phase and
1976 // if necessary clear soft refs that weren't previously
1977 // cleared. We do so by remembering the phase in which
1978 // we came in, and if we are past the refs processing
1979 // phase, we'll choose to just redo the mark-sweep
1980 // collection from scratch.
1981 if (_collectorState > FinalMarking) {
1982 // We are past the refs processing phase;
1983 // start over and do a fresh synchronous CMS cycle
1984 _collectorState = Resetting; // skip to reset to start new cycle
1985 reset(false /* == !asynch */);
1986 *should_start_over = true;
1987 } // else we can continue a possibly ongoing current cycle
1988 }
1989 }
1990 }
1992 // A work method used by the foreground collector to do
1993 // a mark-sweep-compact.
1994 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1995 GenCollectedHeap* gch = GenCollectedHeap::heap();
1997 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1998 gc_timer->register_gc_start();
2000 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
2001 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
2003 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL);
2004 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
2005 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
2006 "collections passed to foreground collector", _full_gcs_since_conc_gc);
2007 }
2009 // Sample collection interval time and reset for collection pause.
2010 if (UseAdaptiveSizePolicy) {
2011 size_policy()->msc_collection_begin();
2012 }
2014 // Temporarily widen the span of the weak reference processing to
2015 // the entire heap.
2016 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
2017 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
2018 // Temporarily, clear the "is_alive_non_header" field of the
2019 // reference processor.
2020 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
2021 // Temporarily make reference _processing_ single threaded (non-MT).
2022 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
2023 // Temporarily make refs discovery atomic
2024 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
2025 // Temporarily make reference _discovery_ single threaded (non-MT)
2026 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
2028 ref_processor()->set_enqueuing_is_done(false);
2029 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/);
2030 ref_processor()->setup_policy(clear_all_soft_refs);
2031 // If an asynchronous collection finishes, the _modUnionTable is
2032 // all clear. If we are assuming the collection from an asynchronous
2033 // collection, clear the _modUnionTable.
2034 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
2035 "_modUnionTable should be clear if the baton was not passed");
2036 _modUnionTable.clear_all();
2037 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
2038 "mod union for klasses should be clear if the baton was passed");
2039 _ct->klass_rem_set()->clear_mod_union();
2041 // We must adjust the allocation statistics being maintained
2042 // in the free list space. We do so by reading and clearing
2043 // the sweep timer and updating the block flux rate estimates below.
2044 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
2045 if (_inter_sweep_timer.is_active()) {
2046 _inter_sweep_timer.stop();
2047 // Note that we do not use this sample to update the _inter_sweep_estimate.
2048 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
2049 _inter_sweep_estimate.padded_average(),
2050 _intra_sweep_estimate.padded_average());
2051 }
2053 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
2054 ref_processor(), clear_all_soft_refs);
2055 #ifdef ASSERT
2056 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
2057 size_t free_size = cms_space->free();
2058 assert(free_size ==
2059 pointer_delta(cms_space->end(), cms_space->compaction_top())
2060 * HeapWordSize,
2061 "All the free space should be compacted into one chunk at top");
2062 assert(cms_space->dictionary()->total_chunk_size(
2063 debug_only(cms_space->freelistLock())) == 0 ||
2064 cms_space->totalSizeInIndexedFreeLists() == 0,
2065 "All the free space should be in a single chunk");
2066 size_t num = cms_space->totalCount();
2067 assert((free_size == 0 && num == 0) ||
2068 (free_size > 0 && (num == 1 || num == 2)),
2069 "There should be at most 2 free chunks after compaction");
2070 #endif // ASSERT
2071 _collectorState = Resetting;
2072 assert(_restart_addr == NULL,
2073 "Should have been NULL'd before baton was passed");
2074 reset(false /* == !asynch */);
2075 _cmsGen->reset_after_compaction();
2076 _concurrent_cycles_since_last_unload = 0;
2078 // Clear any data recorded in the PLAB chunk arrays.
2079 if (_survivor_plab_array != NULL) {
2080 reset_survivor_plab_arrays();
2081 }
2083 // Adjust the per-size allocation stats for the next epoch.
2084 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
2085 // Restart the "inter sweep timer" for the next epoch.
2086 _inter_sweep_timer.reset();
2087 _inter_sweep_timer.start();
2089 // Sample collection pause time and reset for collection interval.
2090 if (UseAdaptiveSizePolicy) {
2091 size_policy()->msc_collection_end(gch->gc_cause());
2092 }
2094 gc_timer->register_gc_end();
2096 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
2098 // For a mark-sweep-compact, compute_new_size() will be called
2099 // in the heap's do_collection() method.
2100 }
2102 // A work method used by the foreground collector to do
2103 // a mark-sweep, after taking over from a possibly on-going
2104 // concurrent mark-sweep collection.
2105 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2106 CollectorState first_state, bool should_start_over) {
2107 if (PrintGC && Verbose) {
2108 gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2109 "collector with count %d",
2110 _full_gcs_since_conc_gc);
2111 }
2112 switch (_collectorState) {
2113 case Idling:
2114 if (first_state == Idling || should_start_over) {
2115 // The background GC was not active, or should
2116 // restarted from scratch; start the cycle.
2117 _collectorState = InitialMarking;
2118 }
2119 // If first_state was not Idling, then a background GC
2120 // was in progress and has now finished. No need to do it
2121 // again. Leave the state as Idling.
2122 break;
2123 case Precleaning:
2124 // In the foreground case don't do the precleaning since
2125 // it is not done concurrently and there is extra work
2126 // required.
2127 _collectorState = FinalMarking;
2128 }
2129 collect_in_foreground(clear_all_soft_refs, GenCollectedHeap::heap()->gc_cause());
2131 // For a mark-sweep, compute_new_size() will be called
2132 // in the heap's do_collection() method.
2133 }
2136 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
2137 DefNewGeneration* dng = _young_gen->as_DefNewGeneration();
2138 EdenSpace* eden_space = dng->eden();
2139 ContiguousSpace* from_space = dng->from();
2140 ContiguousSpace* to_space = dng->to();
2141 // Eden
2142 if (_eden_chunk_array != NULL) {
2143 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
2144 eden_space->bottom(), eden_space->top(),
2145 eden_space->end(), eden_space->capacity());
2146 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", "
2147 "_eden_chunk_capacity=" SIZE_FORMAT,
2148 _eden_chunk_index, _eden_chunk_capacity);
2149 for (size_t i = 0; i < _eden_chunk_index; i++) {
2150 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
2151 i, _eden_chunk_array[i]);
2152 }
2153 }
2154 // Survivor
2155 if (_survivor_chunk_array != NULL) {
2156 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
2157 from_space->bottom(), from_space->top(),
2158 from_space->end(), from_space->capacity());
2159 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", "
2160 "_survivor_chunk_capacity=" SIZE_FORMAT,
2161 _survivor_chunk_index, _survivor_chunk_capacity);
2162 for (size_t i = 0; i < _survivor_chunk_index; i++) {
2163 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
2164 i, _survivor_chunk_array[i]);
2165 }
2166 }
2167 }
2169 void CMSCollector::getFreelistLocks() const {
2170 // Get locks for all free lists in all generations that this
2171 // collector is responsible for
2172 _cmsGen->freelistLock()->lock_without_safepoint_check();
2173 }
2175 void CMSCollector::releaseFreelistLocks() const {
2176 // Release locks for all free lists in all generations that this
2177 // collector is responsible for
2178 _cmsGen->freelistLock()->unlock();
2179 }
2181 bool CMSCollector::haveFreelistLocks() const {
2182 // Check locks for all free lists in all generations that this
2183 // collector is responsible for
2184 assert_lock_strong(_cmsGen->freelistLock());
2185 PRODUCT_ONLY(ShouldNotReachHere());
2186 return true;
2187 }
2189 // A utility class that is used by the CMS collector to
2190 // temporarily "release" the foreground collector from its
2191 // usual obligation to wait for the background collector to
2192 // complete an ongoing phase before proceeding.
2193 class ReleaseForegroundGC: public StackObj {
2194 private:
2195 CMSCollector* _c;
2196 public:
2197 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2198 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2199 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2200 // allow a potentially blocked foreground collector to proceed
2201 _c->_foregroundGCShouldWait = false;
2202 if (_c->_foregroundGCIsActive) {
2203 CGC_lock->notify();
2204 }
2205 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2206 "Possible deadlock");
2207 }
2209 ~ReleaseForegroundGC() {
2210 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2211 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2212 _c->_foregroundGCShouldWait = true;
2213 }
2214 };
2216 // There are separate collect_in_background and collect_in_foreground because of
2217 // the different locking requirements of the background collector and the
2218 // foreground collector. There was originally an attempt to share
2219 // one "collect" method between the background collector and the foreground
2220 // collector but the if-then-else required made it cleaner to have
2221 // separate methods.
2222 void CMSCollector::collect_in_background(bool clear_all_soft_refs, GCCause::Cause cause) {
2223 assert(Thread::current()->is_ConcurrentGC_thread(),
2224 "A CMS asynchronous collection is only allowed on a CMS thread.");
2226 GenCollectedHeap* gch = GenCollectedHeap::heap();
2227 {
2228 bool safepoint_check = Mutex::_no_safepoint_check_flag;
2229 MutexLockerEx hl(Heap_lock, safepoint_check);
2230 FreelistLocker fll(this);
2231 MutexLockerEx x(CGC_lock, safepoint_check);
2232 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2233 // The foreground collector is active or we're
2234 // not using asynchronous collections. Skip this
2235 // background collection.
2236 assert(!_foregroundGCShouldWait, "Should be clear");
2237 return;
2238 } else {
2239 assert(_collectorState == Idling, "Should be idling before start.");
2240 _collectorState = InitialMarking;
2241 register_gc_start(cause);
2242 // Reset the expansion cause, now that we are about to begin
2243 // a new cycle.
2244 clear_expansion_cause();
2246 // Clear the MetaspaceGC flag since a concurrent collection
2247 // is starting but also clear it after the collection.
2248 MetaspaceGC::set_should_concurrent_collect(false);
2249 }
2250 // Decide if we want to enable class unloading as part of the
2251 // ensuing concurrent GC cycle.
2252 update_should_unload_classes();
2253 _full_gc_requested = false; // acks all outstanding full gc requests
2254 _full_gc_cause = GCCause::_no_gc;
2255 // Signal that we are about to start a collection
2256 gch->increment_total_full_collections(); // ... starting a collection cycle
2257 _collection_count_start = gch->total_full_collections();
2258 }
2260 // Used for PrintGC
2261 size_t prev_used;
2262 if (PrintGC && Verbose) {
2263 prev_used = _cmsGen->used(); // XXXPERM
2264 }
2266 // The change of the collection state is normally done at this level;
2267 // the exceptions are phases that are executed while the world is
2268 // stopped. For those phases the change of state is done while the
2269 // world is stopped. For baton passing purposes this allows the
2270 // background collector to finish the phase and change state atomically.
2271 // The foreground collector cannot wait on a phase that is done
2272 // while the world is stopped because the foreground collector already
2273 // has the world stopped and would deadlock.
2274 while (_collectorState != Idling) {
2275 if (TraceCMSState) {
2276 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2277 Thread::current(), _collectorState);
2278 }
2279 // The foreground collector
2280 // holds the Heap_lock throughout its collection.
2281 // holds the CMS token (but not the lock)
2282 // except while it is waiting for the background collector to yield.
2283 //
2284 // The foreground collector should be blocked (not for long)
2285 // if the background collector is about to start a phase
2286 // executed with world stopped. If the background
2287 // collector has already started such a phase, the
2288 // foreground collector is blocked waiting for the
2289 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
2290 // are executed in the VM thread.
2291 //
2292 // The locking order is
2293 // PendingListLock (PLL) -- if applicable (FinalMarking)
2294 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
2295 // CMS token (claimed in
2296 // stop_world_and_do() -->
2297 // safepoint_synchronize() -->
2298 // CMSThread::synchronize())
2300 {
2301 // Check if the FG collector wants us to yield.
2302 CMSTokenSync x(true); // is cms thread
2303 if (waitForForegroundGC()) {
2304 // We yielded to a foreground GC, nothing more to be
2305 // done this round.
2306 assert(_foregroundGCShouldWait == false, "We set it to false in "
2307 "waitForForegroundGC()");
2308 if (TraceCMSState) {
2309 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2310 " exiting collection CMS state %d",
2311 Thread::current(), _collectorState);
2312 }
2313 return;
2314 } else {
2315 // The background collector can run but check to see if the
2316 // foreground collector has done a collection while the
2317 // background collector was waiting to get the CGC_lock
2318 // above. If yes, break so that _foregroundGCShouldWait
2319 // is cleared before returning.
2320 if (_collectorState == Idling) {
2321 break;
2322 }
2323 }
2324 }
2326 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2327 "should be waiting");
2329 switch (_collectorState) {
2330 case InitialMarking:
2331 {
2332 ReleaseForegroundGC x(this);
2333 stats().record_cms_begin();
2334 VM_CMS_Initial_Mark initial_mark_op(this);
2335 VMThread::execute(&initial_mark_op);
2336 }
2337 // The collector state may be any legal state at this point
2338 // since the background collector may have yielded to the
2339 // foreground collector.
2340 break;
2341 case Marking:
2342 // initial marking in checkpointRootsInitialWork has been completed
2343 if (markFromRoots(true)) { // we were successful
2344 assert(_collectorState == Precleaning, "Collector state should "
2345 "have changed");
2346 } else {
2347 assert(_foregroundGCIsActive, "Internal state inconsistency");
2348 }
2349 break;
2350 case Precleaning:
2351 if (UseAdaptiveSizePolicy) {
2352 size_policy()->concurrent_precleaning_begin();
2353 }
2354 // marking from roots in markFromRoots has been completed
2355 preclean();
2356 if (UseAdaptiveSizePolicy) {
2357 size_policy()->concurrent_precleaning_end();
2358 }
2359 assert(_collectorState == AbortablePreclean ||
2360 _collectorState == FinalMarking,
2361 "Collector state should have changed");
2362 break;
2363 case AbortablePreclean:
2364 if (UseAdaptiveSizePolicy) {
2365 size_policy()->concurrent_phases_resume();
2366 }
2367 abortable_preclean();
2368 if (UseAdaptiveSizePolicy) {
2369 size_policy()->concurrent_precleaning_end();
2370 }
2371 assert(_collectorState == FinalMarking, "Collector state should "
2372 "have changed");
2373 break;
2374 case FinalMarking:
2375 {
2376 ReleaseForegroundGC x(this);
2378 VM_CMS_Final_Remark final_remark_op(this);
2379 VMThread::execute(&final_remark_op);
2380 }
2381 assert(_foregroundGCShouldWait, "block post-condition");
2382 break;
2383 case Sweeping:
2384 if (UseAdaptiveSizePolicy) {
2385 size_policy()->concurrent_sweeping_begin();
2386 }
2387 // final marking in checkpointRootsFinal has been completed
2388 sweep(true);
2389 assert(_collectorState == Resizing, "Collector state change "
2390 "to Resizing must be done under the free_list_lock");
2391 _full_gcs_since_conc_gc = 0;
2393 // Stop the timers for adaptive size policy for the concurrent phases
2394 if (UseAdaptiveSizePolicy) {
2395 size_policy()->concurrent_sweeping_end();
2396 size_policy()->concurrent_phases_end(gch->gc_cause(),
2397 gch->prev_gen(_cmsGen)->capacity(),
2398 _cmsGen->free());
2399 }
2401 case Resizing: {
2402 // Sweeping has been completed...
2403 // At this point the background collection has completed.
2404 // Don't move the call to compute_new_size() down
2405 // into code that might be executed if the background
2406 // collection was preempted.
2407 {
2408 ReleaseForegroundGC x(this); // unblock FG collection
2409 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
2410 CMSTokenSync z(true); // not strictly needed.
2411 if (_collectorState == Resizing) {
2412 compute_new_size();
2413 save_heap_summary();
2414 _collectorState = Resetting;
2415 } else {
2416 assert(_collectorState == Idling, "The state should only change"
2417 " because the foreground collector has finished the collection");
2418 }
2419 }
2420 break;
2421 }
2422 case Resetting:
2423 // CMS heap resizing has been completed
2424 reset(true);
2425 assert(_collectorState == Idling, "Collector state should "
2426 "have changed");
2428 MetaspaceGC::set_should_concurrent_collect(false);
2430 stats().record_cms_end();
2431 // Don't move the concurrent_phases_end() and compute_new_size()
2432 // calls to here because a preempted background collection
2433 // has it's state set to "Resetting".
2434 break;
2435 case Idling:
2436 default:
2437 ShouldNotReachHere();
2438 break;
2439 }
2440 if (TraceCMSState) {
2441 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2442 Thread::current(), _collectorState);
2443 }
2444 assert(_foregroundGCShouldWait, "block post-condition");
2445 }
2447 // Should this be in gc_epilogue?
2448 collector_policy()->counters()->update_counters();
2450 {
2451 // Clear _foregroundGCShouldWait and, in the event that the
2452 // foreground collector is waiting, notify it, before
2453 // returning.
2454 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2455 _foregroundGCShouldWait = false;
2456 if (_foregroundGCIsActive) {
2457 CGC_lock->notify();
2458 }
2459 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2460 "Possible deadlock");
2461 }
2462 if (TraceCMSState) {
2463 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2464 " exiting collection CMS state %d",
2465 Thread::current(), _collectorState);
2466 }
2467 if (PrintGC && Verbose) {
2468 _cmsGen->print_heap_change(prev_used);
2469 }
2470 }
2472 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) {
2473 if (!_cms_start_registered) {
2474 register_gc_start(cause);
2475 }
2476 }
2478 void CMSCollector::register_gc_start(GCCause::Cause cause) {
2479 _cms_start_registered = true;
2480 _gc_timer_cm->register_gc_start();
2481 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
2482 }
2484 void CMSCollector::register_gc_end() {
2485 if (_cms_start_registered) {
2486 report_heap_summary(GCWhen::AfterGC);
2488 _gc_timer_cm->register_gc_end();
2489 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
2490 _cms_start_registered = false;
2491 }
2492 }
2494 void CMSCollector::save_heap_summary() {
2495 GenCollectedHeap* gch = GenCollectedHeap::heap();
2496 _last_heap_summary = gch->create_heap_summary();
2497 _last_metaspace_summary = gch->create_metaspace_summary();
2498 }
2500 void CMSCollector::report_heap_summary(GCWhen::Type when) {
2501 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
2502 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
2503 }
2505 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) {
2506 assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2507 "Foreground collector should be waiting, not executing");
2508 assert(Thread::current()->is_VM_thread(), "A foreground collection"
2509 "may only be done by the VM Thread with the world stopped");
2510 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2511 "VM thread should have CMS token");
2513 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2514 true, NULL);)
2515 if (UseAdaptiveSizePolicy) {
2516 size_policy()->ms_collection_begin();
2517 }
2518 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2520 HandleMark hm; // Discard invalid handles created during verification
2522 if (VerifyBeforeGC &&
2523 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2524 Universe::verify();
2525 }
2527 // Snapshot the soft reference policy to be used in this collection cycle.
2528 ref_processor()->setup_policy(clear_all_soft_refs);
2530 // Decide if class unloading should be done
2531 update_should_unload_classes();
2533 bool init_mark_was_synchronous = false; // until proven otherwise
2534 while (_collectorState != Idling) {
2535 if (TraceCMSState) {
2536 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2537 Thread::current(), _collectorState);
2538 }
2539 switch (_collectorState) {
2540 case InitialMarking:
2541 register_foreground_gc_start(cause);
2542 init_mark_was_synchronous = true; // fact to be exploited in re-mark
2543 checkpointRootsInitial(false);
2544 assert(_collectorState == Marking, "Collector state should have changed"
2545 " within checkpointRootsInitial()");
2546 break;
2547 case Marking:
2548 // initial marking in checkpointRootsInitialWork has been completed
2549 if (VerifyDuringGC &&
2550 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2551 Universe::verify("Verify before initial mark: ");
2552 }
2553 {
2554 bool res = markFromRoots(false);
2555 assert(res && _collectorState == FinalMarking, "Collector state should "
2556 "have changed");
2557 break;
2558 }
2559 case FinalMarking:
2560 if (VerifyDuringGC &&
2561 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2562 Universe::verify("Verify before re-mark: ");
2563 }
2564 checkpointRootsFinal(false, clear_all_soft_refs,
2565 init_mark_was_synchronous);
2566 assert(_collectorState == Sweeping, "Collector state should not "
2567 "have changed within checkpointRootsFinal()");
2568 break;
2569 case Sweeping:
2570 // final marking in checkpointRootsFinal has been completed
2571 if (VerifyDuringGC &&
2572 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2573 Universe::verify("Verify before sweep: ");
2574 }
2575 sweep(false);
2576 assert(_collectorState == Resizing, "Incorrect state");
2577 break;
2578 case Resizing: {
2579 // Sweeping has been completed; the actual resize in this case
2580 // is done separately; nothing to be done in this state.
2581 _collectorState = Resetting;
2582 break;
2583 }
2584 case Resetting:
2585 // The heap has been resized.
2586 if (VerifyDuringGC &&
2587 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2588 Universe::verify("Verify before reset: ");
2589 }
2590 save_heap_summary();
2591 reset(false);
2592 assert(_collectorState == Idling, "Collector state should "
2593 "have changed");
2594 break;
2595 case Precleaning:
2596 case AbortablePreclean:
2597 // Elide the preclean phase
2598 _collectorState = FinalMarking;
2599 break;
2600 default:
2601 ShouldNotReachHere();
2602 }
2603 if (TraceCMSState) {
2604 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2605 Thread::current(), _collectorState);
2606 }
2607 }
2609 if (UseAdaptiveSizePolicy) {
2610 GenCollectedHeap* gch = GenCollectedHeap::heap();
2611 size_policy()->ms_collection_end(gch->gc_cause());
2612 }
2614 if (VerifyAfterGC &&
2615 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2616 Universe::verify();
2617 }
2618 if (TraceCMSState) {
2619 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2620 " exiting collection CMS state %d",
2621 Thread::current(), _collectorState);
2622 }
2623 }
2625 bool CMSCollector::waitForForegroundGC() {
2626 bool res = false;
2627 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2628 "CMS thread should have CMS token");
2629 // Block the foreground collector until the
2630 // background collectors decides whether to
2631 // yield.
2632 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2633 _foregroundGCShouldWait = true;
2634 if (_foregroundGCIsActive) {
2635 // The background collector yields to the
2636 // foreground collector and returns a value
2637 // indicating that it has yielded. The foreground
2638 // collector can proceed.
2639 res = true;
2640 _foregroundGCShouldWait = false;
2641 ConcurrentMarkSweepThread::clear_CMS_flag(
2642 ConcurrentMarkSweepThread::CMS_cms_has_token);
2643 ConcurrentMarkSweepThread::set_CMS_flag(
2644 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2645 // Get a possibly blocked foreground thread going
2646 CGC_lock->notify();
2647 if (TraceCMSState) {
2648 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2649 Thread::current(), _collectorState);
2650 }
2651 while (_foregroundGCIsActive) {
2652 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2653 }
2654 ConcurrentMarkSweepThread::set_CMS_flag(
2655 ConcurrentMarkSweepThread::CMS_cms_has_token);
2656 ConcurrentMarkSweepThread::clear_CMS_flag(
2657 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2658 }
2659 if (TraceCMSState) {
2660 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2661 Thread::current(), _collectorState);
2662 }
2663 return res;
2664 }
2666 // Because of the need to lock the free lists and other structures in
2667 // the collector, common to all the generations that the collector is
2668 // collecting, we need the gc_prologues of individual CMS generations
2669 // delegate to their collector. It may have been simpler had the
2670 // current infrastructure allowed one to call a prologue on a
2671 // collector. In the absence of that we have the generation's
2672 // prologue delegate to the collector, which delegates back
2673 // some "local" work to a worker method in the individual generations
2674 // that it's responsible for collecting, while itself doing any
2675 // work common to all generations it's responsible for. A similar
2676 // comment applies to the gc_epilogue()'s.
2677 // The role of the varaible _between_prologue_and_epilogue is to
2678 // enforce the invocation protocol.
2679 void CMSCollector::gc_prologue(bool full) {
2680 // Call gc_prologue_work() for the CMSGen
2681 // we are responsible for.
2683 // The following locking discipline assumes that we are only called
2684 // when the world is stopped.
2685 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2687 // The CMSCollector prologue must call the gc_prologues for the
2688 // "generations" that it's responsible
2689 // for.
2691 assert( Thread::current()->is_VM_thread()
2692 || ( CMSScavengeBeforeRemark
2693 && Thread::current()->is_ConcurrentGC_thread()),
2694 "Incorrect thread type for prologue execution");
2696 if (_between_prologue_and_epilogue) {
2697 // We have already been invoked; this is a gc_prologue delegation
2698 // from yet another CMS generation that we are responsible for, just
2699 // ignore it since all relevant work has already been done.
2700 return;
2701 }
2703 // set a bit saying prologue has been called; cleared in epilogue
2704 _between_prologue_and_epilogue = true;
2705 // Claim locks for common data structures, then call gc_prologue_work()
2706 // for each CMSGen.
2708 getFreelistLocks(); // gets free list locks on constituent spaces
2709 bitMapLock()->lock_without_safepoint_check();
2711 // Should call gc_prologue_work() for all cms gens we are responsible for
2712 bool duringMarking = _collectorState >= Marking
2713 && _collectorState < Sweeping;
2715 // The young collections clear the modified oops state, which tells if
2716 // there are any modified oops in the class. The remark phase also needs
2717 // that information. Tell the young collection to save the union of all
2718 // modified klasses.
2719 if (duringMarking) {
2720 _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2721 }
2723 bool registerClosure = duringMarking;
2725 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ?
2726 &_modUnionClosurePar
2727 : &_modUnionClosure;
2728 _cmsGen->gc_prologue_work(full, registerClosure, muc);
2730 if (!full) {
2731 stats().record_gc0_begin();
2732 }
2733 }
2735 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2737 _capacity_at_prologue = capacity();
2738 _used_at_prologue = used();
2740 // Delegate to CMScollector which knows how to coordinate between
2741 // this and any other CMS generations that it is responsible for
2742 // collecting.
2743 collector()->gc_prologue(full);
2744 }
2746 // This is a "private" interface for use by this generation's CMSCollector.
2747 // Not to be called directly by any other entity (for instance,
2748 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2749 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2750 bool registerClosure, ModUnionClosure* modUnionClosure) {
2751 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2752 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2753 "Should be NULL");
2754 if (registerClosure) {
2755 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2756 }
2757 cmsSpace()->gc_prologue();
2758 // Clear stat counters
2759 NOT_PRODUCT(
2760 assert(_numObjectsPromoted == 0, "check");
2761 assert(_numWordsPromoted == 0, "check");
2762 if (Verbose && PrintGC) {
2763 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2764 SIZE_FORMAT" bytes concurrently",
2765 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2766 }
2767 _numObjectsAllocated = 0;
2768 _numWordsAllocated = 0;
2769 )
2770 }
2772 void CMSCollector::gc_epilogue(bool full) {
2773 // The following locking discipline assumes that we are only called
2774 // when the world is stopped.
2775 assert(SafepointSynchronize::is_at_safepoint(),
2776 "world is stopped assumption");
2778 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2779 // if linear allocation blocks need to be appropriately marked to allow the
2780 // the blocks to be parsable. We also check here whether we need to nudge the
2781 // CMS collector thread to start a new cycle (if it's not already active).
2782 assert( Thread::current()->is_VM_thread()
2783 || ( CMSScavengeBeforeRemark
2784 && Thread::current()->is_ConcurrentGC_thread()),
2785 "Incorrect thread type for epilogue execution");
2787 if (!_between_prologue_and_epilogue) {
2788 // We have already been invoked; this is a gc_epilogue delegation
2789 // from yet another CMS generation that we are responsible for, just
2790 // ignore it since all relevant work has already been done.
2791 return;
2792 }
2793 assert(haveFreelistLocks(), "must have freelist locks");
2794 assert_lock_strong(bitMapLock());
2796 _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2798 _cmsGen->gc_epilogue_work(full);
2800 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2801 // in case sampling was not already enabled, enable it
2802 _start_sampling = true;
2803 }
2804 // reset _eden_chunk_array so sampling starts afresh
2805 _eden_chunk_index = 0;
2807 size_t cms_used = _cmsGen->cmsSpace()->used();
2809 // update performance counters - this uses a special version of
2810 // update_counters() that allows the utilization to be passed as a
2811 // parameter, avoiding multiple calls to used().
2812 //
2813 _cmsGen->update_counters(cms_used);
2815 if (CMSIncrementalMode) {
2816 icms_update_allocation_limits();
2817 }
2819 bitMapLock()->unlock();
2820 releaseFreelistLocks();
2822 if (!CleanChunkPoolAsync) {
2823 Chunk::clean_chunk_pool();
2824 }
2826 set_did_compact(false);
2827 _between_prologue_and_epilogue = false; // ready for next cycle
2828 }
2830 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2831 collector()->gc_epilogue(full);
2833 // Also reset promotion tracking in par gc thread states.
2834 if (CollectedHeap::use_parallel_gc_threads()) {
2835 for (uint i = 0; i < ParallelGCThreads; i++) {
2836 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2837 }
2838 }
2839 }
2841 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2842 assert(!incremental_collection_failed(), "Should have been cleared");
2843 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2844 cmsSpace()->gc_epilogue();
2845 // Print stat counters
2846 NOT_PRODUCT(
2847 assert(_numObjectsAllocated == 0, "check");
2848 assert(_numWordsAllocated == 0, "check");
2849 if (Verbose && PrintGC) {
2850 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2851 SIZE_FORMAT" bytes",
2852 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2853 }
2854 _numObjectsPromoted = 0;
2855 _numWordsPromoted = 0;
2856 )
2858 if (PrintGC && Verbose) {
2859 // Call down the chain in contiguous_available needs the freelistLock
2860 // so print this out before releasing the freeListLock.
2861 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2862 contiguous_available());
2863 }
2864 }
2866 #ifndef PRODUCT
2867 bool CMSCollector::have_cms_token() {
2868 Thread* thr = Thread::current();
2869 if (thr->is_VM_thread()) {
2870 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2871 } else if (thr->is_ConcurrentGC_thread()) {
2872 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2873 } else if (thr->is_GC_task_thread()) {
2874 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2875 ParGCRareEvent_lock->owned_by_self();
2876 }
2877 return false;
2878 }
2879 #endif
2881 // Check reachability of the given heap address in CMS generation,
2882 // treating all other generations as roots.
2883 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2884 // We could "guarantee" below, rather than assert, but i'll
2885 // leave these as "asserts" so that an adventurous debugger
2886 // could try this in the product build provided some subset of
2887 // the conditions were met, provided they were intersted in the
2888 // results and knew that the computation below wouldn't interfere
2889 // with other concurrent computations mutating the structures
2890 // being read or written.
2891 assert(SafepointSynchronize::is_at_safepoint(),
2892 "Else mutations in object graph will make answer suspect");
2893 assert(have_cms_token(), "Should hold cms token");
2894 assert(haveFreelistLocks(), "must hold free list locks");
2895 assert_lock_strong(bitMapLock());
2897 // Clear the marking bit map array before starting, but, just
2898 // for kicks, first report if the given address is already marked
2899 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
2900 _markBitMap.isMarked(addr) ? "" : " not");
2902 if (verify_after_remark()) {
2903 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2904 bool result = verification_mark_bm()->isMarked(addr);
2905 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
2906 result ? "IS" : "is NOT");
2907 return result;
2908 } else {
2909 gclog_or_tty->print_cr("Could not compute result");
2910 return false;
2911 }
2912 }
2915 void
2916 CMSCollector::print_on_error(outputStream* st) {
2917 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2918 if (collector != NULL) {
2919 CMSBitMap* bitmap = &collector->_markBitMap;
2920 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap);
2921 bitmap->print_on_error(st, " Bits: ");
2923 st->cr();
2925 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2926 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap);
2927 mut_bitmap->print_on_error(st, " Bits: ");
2928 }
2929 }
2931 ////////////////////////////////////////////////////////
2932 // CMS Verification Support
2933 ////////////////////////////////////////////////////////
2934 // Following the remark phase, the following invariant
2935 // should hold -- each object in the CMS heap which is
2936 // marked in markBitMap() should be marked in the verification_mark_bm().
2938 class VerifyMarkedClosure: public BitMapClosure {
2939 CMSBitMap* _marks;
2940 bool _failed;
2942 public:
2943 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2945 bool do_bit(size_t offset) {
2946 HeapWord* addr = _marks->offsetToHeapWord(offset);
2947 if (!_marks->isMarked(addr)) {
2948 oop(addr)->print_on(gclog_or_tty);
2949 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2950 _failed = true;
2951 }
2952 return true;
2953 }
2955 bool failed() { return _failed; }
2956 };
2958 bool CMSCollector::verify_after_remark(bool silent) {
2959 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... ");
2960 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2961 static bool init = false;
2963 assert(SafepointSynchronize::is_at_safepoint(),
2964 "Else mutations in object graph will make answer suspect");
2965 assert(have_cms_token(),
2966 "Else there may be mutual interference in use of "
2967 " verification data structures");
2968 assert(_collectorState > Marking && _collectorState <= Sweeping,
2969 "Else marking info checked here may be obsolete");
2970 assert(haveFreelistLocks(), "must hold free list locks");
2971 assert_lock_strong(bitMapLock());
2974 // Allocate marking bit map if not already allocated
2975 if (!init) { // first time
2976 if (!verification_mark_bm()->allocate(_span)) {
2977 return false;
2978 }
2979 init = true;
2980 }
2982 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2984 // Turn off refs discovery -- so we will be tracing through refs.
2985 // This is as intended, because by this time
2986 // GC must already have cleared any refs that need to be cleared,
2987 // and traced those that need to be marked; moreover,
2988 // the marking done here is not going to intefere in any
2989 // way with the marking information used by GC.
2990 NoRefDiscovery no_discovery(ref_processor());
2992 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2994 // Clear any marks from a previous round
2995 verification_mark_bm()->clear_all();
2996 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2997 verify_work_stacks_empty();
2999 GenCollectedHeap* gch = GenCollectedHeap::heap();
3000 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3001 // Update the saved marks which may affect the root scans.
3002 gch->save_marks();
3004 if (CMSRemarkVerifyVariant == 1) {
3005 // In this first variant of verification, we complete
3006 // all marking, then check if the new marks-verctor is
3007 // a subset of the CMS marks-vector.
3008 verify_after_remark_work_1();
3009 } else if (CMSRemarkVerifyVariant == 2) {
3010 // In this second variant of verification, we flag an error
3011 // (i.e. an object reachable in the new marks-vector not reachable
3012 // in the CMS marks-vector) immediately, also indicating the
3013 // identify of an object (A) that references the unmarked object (B) --
3014 // presumably, a mutation to A failed to be picked up by preclean/remark?
3015 verify_after_remark_work_2();
3016 } else {
3017 warning("Unrecognized value %d for CMSRemarkVerifyVariant",
3018 CMSRemarkVerifyVariant);
3019 }
3020 if (!silent) gclog_or_tty->print(" done] ");
3021 return true;
3022 }
3024 void CMSCollector::verify_after_remark_work_1() {
3025 ResourceMark rm;
3026 HandleMark hm;
3027 GenCollectedHeap* gch = GenCollectedHeap::heap();
3029 // Get a clear set of claim bits for the strong roots processing to work with.
3030 ClassLoaderDataGraph::clear_claimed_marks();
3032 // Mark from roots one level into CMS
3033 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
3034 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3036 gch->gen_process_strong_roots(_cmsGen->level(),
3037 true, // younger gens are roots
3038 true, // activate StrongRootsScope
3039 false, // not scavenging
3040 SharedHeap::ScanningOption(roots_scanning_options()),
3041 ¬Older,
3042 true, // walk code active on stacks
3043 NULL,
3044 NULL); // SSS: Provide correct closure
3046 // Now mark from the roots
3047 MarkFromRootsClosure markFromRootsClosure(this, _span,
3048 verification_mark_bm(), verification_mark_stack(),
3049 false /* don't yield */, true /* verifying */);
3050 assert(_restart_addr == NULL, "Expected pre-condition");
3051 verification_mark_bm()->iterate(&markFromRootsClosure);
3052 while (_restart_addr != NULL) {
3053 // Deal with stack overflow: by restarting at the indicated
3054 // address.
3055 HeapWord* ra = _restart_addr;
3056 markFromRootsClosure.reset(ra);
3057 _restart_addr = NULL;
3058 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
3059 }
3060 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
3061 verify_work_stacks_empty();
3063 // Marking completed -- now verify that each bit marked in
3064 // verification_mark_bm() is also marked in markBitMap(); flag all
3065 // errors by printing corresponding objects.
3066 VerifyMarkedClosure vcl(markBitMap());
3067 verification_mark_bm()->iterate(&vcl);
3068 if (vcl.failed()) {
3069 gclog_or_tty->print("Verification failed");
3070 Universe::heap()->print_on(gclog_or_tty);
3071 fatal("CMS: failed marking verification after remark");
3072 }
3073 }
3075 class VerifyKlassOopsKlassClosure : public KlassClosure {
3076 class VerifyKlassOopsClosure : public OopClosure {
3077 CMSBitMap* _bitmap;
3078 public:
3079 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
3080 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
3081 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
3082 } _oop_closure;
3083 public:
3084 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
3085 void do_klass(Klass* k) {
3086 k->oops_do(&_oop_closure);
3087 }
3088 };
3090 void CMSCollector::verify_after_remark_work_2() {
3091 ResourceMark rm;
3092 HandleMark hm;
3093 GenCollectedHeap* gch = GenCollectedHeap::heap();
3095 // Get a clear set of claim bits for the strong roots processing to work with.
3096 ClassLoaderDataGraph::clear_claimed_marks();
3098 // Mark from roots one level into CMS
3099 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
3100 markBitMap());
3101 CMKlassClosure klass_closure(¬Older);
3103 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3104 gch->gen_process_strong_roots(_cmsGen->level(),
3105 true, // younger gens are roots
3106 true, // activate StrongRootsScope
3107 false, // not scavenging
3108 SharedHeap::ScanningOption(roots_scanning_options()),
3109 ¬Older,
3110 true, // walk code active on stacks
3111 NULL,
3112 &klass_closure);
3114 // Now mark from the roots
3115 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
3116 verification_mark_bm(), markBitMap(), verification_mark_stack());
3117 assert(_restart_addr == NULL, "Expected pre-condition");
3118 verification_mark_bm()->iterate(&markFromRootsClosure);
3119 while (_restart_addr != NULL) {
3120 // Deal with stack overflow: by restarting at the indicated
3121 // address.
3122 HeapWord* ra = _restart_addr;
3123 markFromRootsClosure.reset(ra);
3124 _restart_addr = NULL;
3125 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
3126 }
3127 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
3128 verify_work_stacks_empty();
3130 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
3131 ClassLoaderDataGraph::classes_do(&verify_klass_oops);
3133 // Marking completed -- now verify that each bit marked in
3134 // verification_mark_bm() is also marked in markBitMap(); flag all
3135 // errors by printing corresponding objects.
3136 VerifyMarkedClosure vcl(markBitMap());
3137 verification_mark_bm()->iterate(&vcl);
3138 assert(!vcl.failed(), "Else verification above should not have succeeded");
3139 }
3141 void ConcurrentMarkSweepGeneration::save_marks() {
3142 // delegate to CMS space
3143 cmsSpace()->save_marks();
3144 for (uint i = 0; i < ParallelGCThreads; i++) {
3145 _par_gc_thread_states[i]->promo.startTrackingPromotions();
3146 }
3147 }
3149 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
3150 return cmsSpace()->no_allocs_since_save_marks();
3151 }
3153 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
3154 \
3155 void ConcurrentMarkSweepGeneration:: \
3156 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
3157 cl->set_generation(this); \
3158 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
3159 cl->reset_generation(); \
3160 save_marks(); \
3161 }
3163 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
3165 void
3166 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
3167 cl->set_generation(this);
3168 younger_refs_in_space_iterate(_cmsSpace, cl);
3169 cl->reset_generation();
3170 }
3172 void
3173 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) {
3174 if (freelistLock()->owned_by_self()) {
3175 Generation::oop_iterate(mr, cl);
3176 } else {
3177 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3178 Generation::oop_iterate(mr, cl);
3179 }
3180 }
3182 void
3183 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
3184 if (freelistLock()->owned_by_self()) {
3185 Generation::oop_iterate(cl);
3186 } else {
3187 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3188 Generation::oop_iterate(cl);
3189 }
3190 }
3192 void
3193 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
3194 if (freelistLock()->owned_by_self()) {
3195 Generation::object_iterate(cl);
3196 } else {
3197 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3198 Generation::object_iterate(cl);
3199 }
3200 }
3202 void
3203 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
3204 if (freelistLock()->owned_by_self()) {
3205 Generation::safe_object_iterate(cl);
3206 } else {
3207 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3208 Generation::safe_object_iterate(cl);
3209 }
3210 }
3212 void
3213 ConcurrentMarkSweepGeneration::post_compact() {
3214 }
3216 void
3217 ConcurrentMarkSweepGeneration::prepare_for_verify() {
3218 // Fix the linear allocation blocks to look like free blocks.
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()->prepare_for_verify();
3225 } else {
3226 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3227 cmsSpace()->prepare_for_verify();
3228 }
3229 }
3231 void
3232 ConcurrentMarkSweepGeneration::verify() {
3233 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3234 // are not called when the heap is verified during universe initialization and
3235 // at vm shutdown.
3236 if (freelistLock()->owned_by_self()) {
3237 cmsSpace()->verify();
3238 } else {
3239 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3240 cmsSpace()->verify();
3241 }
3242 }
3244 void CMSCollector::verify() {
3245 _cmsGen->verify();
3246 }
3248 #ifndef PRODUCT
3249 bool CMSCollector::overflow_list_is_empty() const {
3250 assert(_num_par_pushes >= 0, "Inconsistency");
3251 if (_overflow_list == NULL) {
3252 assert(_num_par_pushes == 0, "Inconsistency");
3253 }
3254 return _overflow_list == NULL;
3255 }
3257 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3258 // merely consolidate assertion checks that appear to occur together frequently.
3259 void CMSCollector::verify_work_stacks_empty() const {
3260 assert(_markStack.isEmpty(), "Marking stack should be empty");
3261 assert(overflow_list_is_empty(), "Overflow list should be empty");
3262 }
3264 void CMSCollector::verify_overflow_empty() const {
3265 assert(overflow_list_is_empty(), "Overflow list should be empty");
3266 assert(no_preserved_marks(), "No preserved marks");
3267 }
3268 #endif // PRODUCT
3270 // Decide if we want to enable class unloading as part of the
3271 // ensuing concurrent GC cycle. We will collect and
3272 // unload classes if it's the case that:
3273 // (1) an explicit gc request has been made and the flag
3274 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3275 // (2) (a) class unloading is enabled at the command line, and
3276 // (b) old gen is getting really full
3277 // NOTE: Provided there is no change in the state of the heap between
3278 // calls to this method, it should have idempotent results. Moreover,
3279 // its results should be monotonically increasing (i.e. going from 0 to 1,
3280 // but not 1 to 0) between successive calls between which the heap was
3281 // not collected. For the implementation below, it must thus rely on
3282 // the property that concurrent_cycles_since_last_unload()
3283 // will not decrease unless a collection cycle happened and that
3284 // _cmsGen->is_too_full() are
3285 // themselves also monotonic in that sense. See check_monotonicity()
3286 // below.
3287 void CMSCollector::update_should_unload_classes() {
3288 _should_unload_classes = false;
3289 // Condition 1 above
3290 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3291 _should_unload_classes = true;
3292 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3293 // Disjuncts 2.b.(i,ii,iii) above
3294 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3295 CMSClassUnloadingMaxInterval)
3296 || _cmsGen->is_too_full();
3297 }
3298 }
3300 bool ConcurrentMarkSweepGeneration::is_too_full() const {
3301 bool res = should_concurrent_collect();
3302 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
3303 return res;
3304 }
3306 void CMSCollector::setup_cms_unloading_and_verification_state() {
3307 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3308 || VerifyBeforeExit;
3309 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
3311 // We set the proper root for this CMS cycle here.
3312 if (should_unload_classes()) { // Should unload classes this cycle
3313 remove_root_scanning_option(SharedHeap::SO_AllClasses);
3314 add_root_scanning_option(SharedHeap::SO_SystemClasses);
3315 remove_root_scanning_option(rso); // Shrink the root set appropriately
3316 set_verifying(should_verify); // Set verification state for this cycle
3317 return; // Nothing else needs to be done at this time
3318 }
3320 // Not unloading classes this cycle
3321 assert(!should_unload_classes(), "Inconsitency!");
3322 remove_root_scanning_option(SharedHeap::SO_SystemClasses);
3323 add_root_scanning_option(SharedHeap::SO_AllClasses);
3325 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3326 // Include symbols, strings and code cache elements to prevent their resurrection.
3327 add_root_scanning_option(rso);
3328 set_verifying(true);
3329 } else if (verifying() && !should_verify) {
3330 // We were verifying, but some verification flags got disabled.
3331 set_verifying(false);
3332 // Exclude symbols, strings and code cache elements from root scanning to
3333 // reduce IM and RM pauses.
3334 remove_root_scanning_option(rso);
3335 }
3336 }
3339 #ifndef PRODUCT
3340 HeapWord* CMSCollector::block_start(const void* p) const {
3341 const HeapWord* addr = (HeapWord*)p;
3342 if (_span.contains(p)) {
3343 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3344 return _cmsGen->cmsSpace()->block_start(p);
3345 }
3346 }
3347 return NULL;
3348 }
3349 #endif
3351 HeapWord*
3352 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3353 bool tlab,
3354 bool parallel) {
3355 CMSSynchronousYieldRequest yr;
3356 assert(!tlab, "Can't deal with TLAB allocation");
3357 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3358 expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3359 CMSExpansionCause::_satisfy_allocation);
3360 if (GCExpandToAllocateDelayMillis > 0) {
3361 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3362 }
3363 return have_lock_and_allocate(word_size, tlab);
3364 }
3366 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3367 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3368 // to CardGeneration and share it...
3369 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) {
3370 return CardGeneration::expand(bytes, expand_bytes);
3371 }
3373 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3374 CMSExpansionCause::Cause cause)
3375 {
3377 bool success = expand(bytes, expand_bytes);
3379 // remember why we expanded; this information is used
3380 // by shouldConcurrentCollect() when making decisions on whether to start
3381 // a new CMS cycle.
3382 if (success) {
3383 set_expansion_cause(cause);
3384 if (PrintGCDetails && Verbose) {
3385 gclog_or_tty->print_cr("Expanded CMS gen for %s",
3386 CMSExpansionCause::to_string(cause));
3387 }
3388 }
3389 }
3391 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3392 HeapWord* res = NULL;
3393 MutexLocker x(ParGCRareEvent_lock);
3394 while (true) {
3395 // Expansion by some other thread might make alloc OK now:
3396 res = ps->lab.alloc(word_sz);
3397 if (res != NULL) return res;
3398 // If there's not enough expansion space available, give up.
3399 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3400 return NULL;
3401 }
3402 // Otherwise, we try expansion.
3403 expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3404 CMSExpansionCause::_allocate_par_lab);
3405 // Now go around the loop and try alloc again;
3406 // A competing par_promote might beat us to the expansion space,
3407 // so we may go around the loop again if promotion fails agaion.
3408 if (GCExpandToAllocateDelayMillis > 0) {
3409 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3410 }
3411 }
3412 }
3415 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3416 PromotionInfo* promo) {
3417 MutexLocker x(ParGCRareEvent_lock);
3418 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3419 while (true) {
3420 // Expansion by some other thread might make alloc OK now:
3421 if (promo->ensure_spooling_space()) {
3422 assert(promo->has_spooling_space(),
3423 "Post-condition of successful ensure_spooling_space()");
3424 return true;
3425 }
3426 // If there's not enough expansion space available, give up.
3427 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3428 return false;
3429 }
3430 // Otherwise, we try expansion.
3431 expand(refill_size_bytes, MinHeapDeltaBytes,
3432 CMSExpansionCause::_allocate_par_spooling_space);
3433 // Now go around the loop and try alloc again;
3434 // A competing allocation might beat us to the expansion space,
3435 // so we may go around the loop again if allocation fails again.
3436 if (GCExpandToAllocateDelayMillis > 0) {
3437 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3438 }
3439 }
3440 }
3443 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3444 assert_locked_or_safepoint(ExpandHeap_lock);
3445 // Shrink committed space
3446 _virtual_space.shrink_by(bytes);
3447 // Shrink space; this also shrinks the space's BOT
3448 _cmsSpace->set_end((HeapWord*) _virtual_space.high());
3449 size_t new_word_size = heap_word_size(_cmsSpace->capacity());
3450 // Shrink the shared block offset array
3451 _bts->resize(new_word_size);
3452 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3453 // Shrink the card table
3454 Universe::heap()->barrier_set()->resize_covered_region(mr);
3456 if (Verbose && PrintGC) {
3457 size_t new_mem_size = _virtual_space.committed_size();
3458 size_t old_mem_size = new_mem_size + bytes;
3459 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
3460 name(), old_mem_size/K, new_mem_size/K);
3461 }
3462 }
3464 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3465 assert_locked_or_safepoint(Heap_lock);
3466 size_t size = ReservedSpace::page_align_size_down(bytes);
3467 // Only shrink if a compaction was done so that all the free space
3468 // in the generation is in a contiguous block at the end.
3469 if (size > 0 && did_compact()) {
3470 shrink_by(size);
3471 }
3472 }
3474 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3475 assert_locked_or_safepoint(Heap_lock);
3476 bool result = _virtual_space.expand_by(bytes);
3477 if (result) {
3478 size_t new_word_size =
3479 heap_word_size(_virtual_space.committed_size());
3480 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3481 _bts->resize(new_word_size); // resize the block offset shared array
3482 Universe::heap()->barrier_set()->resize_covered_region(mr);
3483 // Hmmmm... why doesn't CFLS::set_end verify locking?
3484 // This is quite ugly; FIX ME XXX
3485 _cmsSpace->assert_locked(freelistLock());
3486 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3488 // update the space and generation capacity counters
3489 if (UsePerfData) {
3490 _space_counters->update_capacity();
3491 _gen_counters->update_all();
3492 }
3494 if (Verbose && PrintGC) {
3495 size_t new_mem_size = _virtual_space.committed_size();
3496 size_t old_mem_size = new_mem_size - bytes;
3497 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K",
3498 name(), old_mem_size/K, bytes/K, new_mem_size/K);
3499 }
3500 }
3501 return result;
3502 }
3504 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3505 assert_locked_or_safepoint(Heap_lock);
3506 bool success = true;
3507 const size_t remaining_bytes = _virtual_space.uncommitted_size();
3508 if (remaining_bytes > 0) {
3509 success = grow_by(remaining_bytes);
3510 DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3511 }
3512 return success;
3513 }
3515 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
3516 assert_locked_or_safepoint(Heap_lock);
3517 assert_lock_strong(freelistLock());
3518 if (PrintGCDetails && Verbose) {
3519 warning("Shrinking of CMS not yet implemented");
3520 }
3521 return;
3522 }
3525 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3526 // phases.
3527 class CMSPhaseAccounting: public StackObj {
3528 public:
3529 CMSPhaseAccounting(CMSCollector *collector,
3530 const char *phase,
3531 bool print_cr = true);
3532 ~CMSPhaseAccounting();
3534 private:
3535 CMSCollector *_collector;
3536 const char *_phase;
3537 elapsedTimer _wallclock;
3538 bool _print_cr;
3540 public:
3541 // Not MT-safe; so do not pass around these StackObj's
3542 // where they may be accessed by other threads.
3543 jlong wallclock_millis() {
3544 assert(_wallclock.is_active(), "Wall clock should not stop");
3545 _wallclock.stop(); // to record time
3546 jlong ret = _wallclock.milliseconds();
3547 _wallclock.start(); // restart
3548 return ret;
3549 }
3550 };
3552 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3553 const char *phase,
3554 bool print_cr) :
3555 _collector(collector), _phase(phase), _print_cr(print_cr) {
3557 if (PrintCMSStatistics != 0) {
3558 _collector->resetYields();
3559 }
3560 if (PrintGCDetails) {
3561 gclog_or_tty->date_stamp(PrintGCDateStamps);
3562 gclog_or_tty->stamp(PrintGCTimeStamps);
3563 gclog_or_tty->print_cr("[%s-concurrent-%s-start]",
3564 _collector->cmsGen()->short_name(), _phase);
3565 }
3566 _collector->resetTimer();
3567 _wallclock.start();
3568 _collector->startTimer();
3569 }
3571 CMSPhaseAccounting::~CMSPhaseAccounting() {
3572 assert(_wallclock.is_active(), "Wall clock should not have stopped");
3573 _collector->stopTimer();
3574 _wallclock.stop();
3575 if (PrintGCDetails) {
3576 gclog_or_tty->date_stamp(PrintGCDateStamps);
3577 gclog_or_tty->stamp(PrintGCTimeStamps);
3578 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3579 _collector->cmsGen()->short_name(),
3580 _phase, _collector->timerValue(), _wallclock.seconds());
3581 if (_print_cr) {
3582 gclog_or_tty->cr();
3583 }
3584 if (PrintCMSStatistics != 0) {
3585 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3586 _collector->yields());
3587 }
3588 }
3589 }
3591 // CMS work
3593 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
3594 class CMSParMarkTask : public AbstractGangTask {
3595 protected:
3596 CMSCollector* _collector;
3597 int _n_workers;
3598 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) :
3599 AbstractGangTask(name),
3600 _collector(collector),
3601 _n_workers(n_workers) {}
3602 // Work method in support of parallel rescan ... of young gen spaces
3603 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
3604 ContiguousSpace* space,
3605 HeapWord** chunk_array, size_t chunk_top);
3606 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
3607 };
3609 // Parallel initial mark task
3610 class CMSParInitialMarkTask: public CMSParMarkTask {
3611 public:
3612 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) :
3613 CMSParMarkTask("Scan roots and young gen for initial mark in parallel",
3614 collector, n_workers) {}
3615 void work(uint worker_id);
3616 };
3618 // Checkpoint the roots into this generation from outside
3619 // this generation. [Note this initial checkpoint need only
3620 // be approximate -- we'll do a catch up phase subsequently.]
3621 void CMSCollector::checkpointRootsInitial(bool asynch) {
3622 assert(_collectorState == InitialMarking, "Wrong collector state");
3623 check_correct_thread_executing();
3624 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
3626 save_heap_summary();
3627 report_heap_summary(GCWhen::BeforeGC);
3629 ReferenceProcessor* rp = ref_processor();
3630 SpecializationStats::clear();
3631 assert(_restart_addr == NULL, "Control point invariant");
3632 if (asynch) {
3633 // acquire locks for subsequent manipulations
3634 MutexLockerEx x(bitMapLock(),
3635 Mutex::_no_safepoint_check_flag);
3636 checkpointRootsInitialWork(asynch);
3637 // enable ("weak") refs discovery
3638 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/);
3639 _collectorState = Marking;
3640 } else {
3641 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3642 // which recognizes if we are a CMS generation, and doesn't try to turn on
3643 // discovery; verify that they aren't meddling.
3644 assert(!rp->discovery_is_atomic(),
3645 "incorrect setting of discovery predicate");
3646 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3647 "ref discovery for this generation kind");
3648 // already have locks
3649 checkpointRootsInitialWork(asynch);
3650 // now enable ("weak") refs discovery
3651 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/);
3652 _collectorState = Marking;
3653 }
3654 SpecializationStats::print();
3655 }
3657 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3658 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3659 assert(_collectorState == InitialMarking, "just checking");
3661 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3662 // precede our marking with a collection of all
3663 // younger generations to keep floating garbage to a minimum.
3664 // XXX: we won't do this for now -- it's an optimization to be done later.
3666 // already have locks
3667 assert_lock_strong(bitMapLock());
3668 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3670 // Setup the verification and class unloading state for this
3671 // CMS collection cycle.
3672 setup_cms_unloading_and_verification_state();
3674 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
3675 PrintGCDetails && Verbose, true, _gc_timer_cm);)
3676 if (UseAdaptiveSizePolicy) {
3677 size_policy()->checkpoint_roots_initial_begin();
3678 }
3680 // Reset all the PLAB chunk arrays if necessary.
3681 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3682 reset_survivor_plab_arrays();
3683 }
3685 ResourceMark rm;
3686 HandleMark hm;
3688 FalseClosure falseClosure;
3689 // In the case of a synchronous collection, we will elide the
3690 // remark step, so it's important to catch all the nmethod oops
3691 // in this step.
3692 // The final 'true' flag to gen_process_strong_roots will ensure this.
3693 // If 'async' is true, we can relax the nmethod tracing.
3694 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
3695 GenCollectedHeap* gch = GenCollectedHeap::heap();
3697 verify_work_stacks_empty();
3698 verify_overflow_empty();
3700 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3701 // Update the saved marks which may affect the root scans.
3702 gch->save_marks();
3704 // weak reference processing has not started yet.
3705 ref_processor()->set_enqueuing_is_done(false);
3707 // Need to remember all newly created CLDs,
3708 // so that we can guarantee that the remark finds them.
3709 ClassLoaderDataGraph::remember_new_clds(true);
3711 // Whenever a CLD is found, it will be claimed before proceeding to mark
3712 // the klasses. The claimed marks need to be cleared before marking starts.
3713 ClassLoaderDataGraph::clear_claimed_marks();
3715 if (CMSPrintEdenSurvivorChunks) {
3716 print_eden_and_survivor_chunk_arrays();
3717 }
3719 {
3720 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3721 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
3722 // The parallel version.
3723 FlexibleWorkGang* workers = gch->workers();
3724 assert(workers != NULL, "Need parallel worker threads.");
3725 int n_workers = workers->active_workers();
3726 CMSParInitialMarkTask tsk(this, n_workers);
3727 gch->set_par_threads(n_workers);
3728 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
3729 if (n_workers > 1) {
3730 GenCollectedHeap::StrongRootsScope srs(gch);
3731 workers->run_task(&tsk);
3732 } else {
3733 GenCollectedHeap::StrongRootsScope srs(gch);
3734 tsk.work(0);
3735 }
3736 gch->set_par_threads(0);
3737 } else {
3738 // The serial version.
3739 CMKlassClosure klass_closure(¬Older);
3740 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3741 gch->gen_process_strong_roots(_cmsGen->level(),
3742 true, // younger gens are roots
3743 true, // activate StrongRootsScope
3744 false, // not scavenging
3745 SharedHeap::ScanningOption(roots_scanning_options()),
3746 ¬Older,
3747 true, // walk all of code cache if (so & SO_CodeCache)
3748 NULL,
3749 &klass_closure);
3750 }
3751 }
3753 // Clear mod-union table; it will be dirtied in the prologue of
3754 // CMS generation per each younger generation collection.
3756 assert(_modUnionTable.isAllClear(),
3757 "Was cleared in most recent final checkpoint phase"
3758 " or no bits are set in the gc_prologue before the start of the next "
3759 "subsequent marking phase.");
3761 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
3763 // Save the end of the used_region of the constituent generations
3764 // to be used to limit the extent of sweep in each generation.
3765 save_sweep_limits();
3766 if (UseAdaptiveSizePolicy) {
3767 size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3768 }
3769 verify_overflow_empty();
3770 }
3772 bool CMSCollector::markFromRoots(bool asynch) {
3773 // we might be tempted to assert that:
3774 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3775 // "inconsistent argument?");
3776 // However that wouldn't be right, because it's possible that
3777 // a safepoint is indeed in progress as a younger generation
3778 // stop-the-world GC happens even as we mark in this generation.
3779 assert(_collectorState == Marking, "inconsistent state?");
3780 check_correct_thread_executing();
3781 verify_overflow_empty();
3783 bool res;
3784 if (asynch) {
3786 // Start the timers for adaptive size policy for the concurrent phases
3787 // Do it here so that the foreground MS can use the concurrent
3788 // timer since a foreground MS might has the sweep done concurrently
3789 // or STW.
3790 if (UseAdaptiveSizePolicy) {
3791 size_policy()->concurrent_marking_begin();
3792 }
3794 // Weak ref discovery note: We may be discovering weak
3795 // refs in this generation concurrent (but interleaved) with
3796 // weak ref discovery by a younger generation collector.
3798 CMSTokenSyncWithLocks ts(true, bitMapLock());
3799 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3800 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3801 res = markFromRootsWork(asynch);
3802 if (res) {
3803 _collectorState = Precleaning;
3804 } else { // We failed and a foreground collection wants to take over
3805 assert(_foregroundGCIsActive, "internal state inconsistency");
3806 assert(_restart_addr == NULL, "foreground will restart from scratch");
3807 if (PrintGCDetails) {
3808 gclog_or_tty->print_cr("bailing out to foreground collection");
3809 }
3810 }
3811 if (UseAdaptiveSizePolicy) {
3812 size_policy()->concurrent_marking_end();
3813 }
3814 } else {
3815 assert(SafepointSynchronize::is_at_safepoint(),
3816 "inconsistent with asynch == false");
3817 if (UseAdaptiveSizePolicy) {
3818 size_policy()->ms_collection_marking_begin();
3819 }
3820 // already have locks
3821 res = markFromRootsWork(asynch);
3822 _collectorState = FinalMarking;
3823 if (UseAdaptiveSizePolicy) {
3824 GenCollectedHeap* gch = GenCollectedHeap::heap();
3825 size_policy()->ms_collection_marking_end(gch->gc_cause());
3826 }
3827 }
3828 verify_overflow_empty();
3829 return res;
3830 }
3832 bool CMSCollector::markFromRootsWork(bool asynch) {
3833 // iterate over marked bits in bit map, doing a full scan and mark
3834 // from these roots using the following algorithm:
3835 // . if oop is to the right of the current scan pointer,
3836 // mark corresponding bit (we'll process it later)
3837 // . else (oop is to left of current scan pointer)
3838 // push oop on marking stack
3839 // . drain the marking stack
3841 // Note that when we do a marking step we need to hold the
3842 // bit map lock -- recall that direct allocation (by mutators)
3843 // and promotion (by younger generation collectors) is also
3844 // marking the bit map. [the so-called allocate live policy.]
3845 // Because the implementation of bit map marking is not
3846 // robust wrt simultaneous marking of bits in the same word,
3847 // we need to make sure that there is no such interference
3848 // between concurrent such updates.
3850 // already have locks
3851 assert_lock_strong(bitMapLock());
3853 verify_work_stacks_empty();
3854 verify_overflow_empty();
3855 bool result = false;
3856 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3857 result = do_marking_mt(asynch);
3858 } else {
3859 result = do_marking_st(asynch);
3860 }
3861 return result;
3862 }
3864 // Forward decl
3865 class CMSConcMarkingTask;
3867 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3868 CMSCollector* _collector;
3869 CMSConcMarkingTask* _task;
3870 public:
3871 virtual void yield();
3873 // "n_threads" is the number of threads to be terminated.
3874 // "queue_set" is a set of work queues of other threads.
3875 // "collector" is the CMS collector associated with this task terminator.
3876 // "yield" indicates whether we need the gang as a whole to yield.
3877 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3878 ParallelTaskTerminator(n_threads, queue_set),
3879 _collector(collector) { }
3881 void set_task(CMSConcMarkingTask* task) {
3882 _task = task;
3883 }
3884 };
3886 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3887 CMSConcMarkingTask* _task;
3888 public:
3889 bool should_exit_termination();
3890 void set_task(CMSConcMarkingTask* task) {
3891 _task = task;
3892 }
3893 };
3895 // MT Concurrent Marking Task
3896 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3897 CMSCollector* _collector;
3898 int _n_workers; // requested/desired # workers
3899 bool _asynch;
3900 bool _result;
3901 CompactibleFreeListSpace* _cms_space;
3902 char _pad_front[64]; // padding to ...
3903 HeapWord* _global_finger; // ... avoid sharing cache line
3904 char _pad_back[64];
3905 HeapWord* _restart_addr;
3907 // Exposed here for yielding support
3908 Mutex* const _bit_map_lock;
3910 // The per thread work queues, available here for stealing
3911 OopTaskQueueSet* _task_queues;
3913 // Termination (and yielding) support
3914 CMSConcMarkingTerminator _term;
3915 CMSConcMarkingTerminatorTerminator _term_term;
3917 public:
3918 CMSConcMarkingTask(CMSCollector* collector,
3919 CompactibleFreeListSpace* cms_space,
3920 bool asynch,
3921 YieldingFlexibleWorkGang* workers,
3922 OopTaskQueueSet* task_queues):
3923 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3924 _collector(collector),
3925 _cms_space(cms_space),
3926 _asynch(asynch), _n_workers(0), _result(true),
3927 _task_queues(task_queues),
3928 _term(_n_workers, task_queues, _collector),
3929 _bit_map_lock(collector->bitMapLock())
3930 {
3931 _requested_size = _n_workers;
3932 _term.set_task(this);
3933 _term_term.set_task(this);
3934 _restart_addr = _global_finger = _cms_space->bottom();
3935 }
3938 OopTaskQueueSet* task_queues() { return _task_queues; }
3940 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3942 HeapWord** global_finger_addr() { return &_global_finger; }
3944 CMSConcMarkingTerminator* terminator() { return &_term; }
3946 virtual void set_for_termination(int active_workers) {
3947 terminator()->reset_for_reuse(active_workers);
3948 }
3950 void work(uint worker_id);
3951 bool should_yield() {
3952 return ConcurrentMarkSweepThread::should_yield()
3953 && !_collector->foregroundGCIsActive()
3954 && _asynch;
3955 }
3957 virtual void coordinator_yield(); // stuff done by coordinator
3958 bool result() { return _result; }
3960 void reset(HeapWord* ra) {
3961 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3962 _restart_addr = _global_finger = ra;
3963 _term.reset_for_reuse();
3964 }
3966 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3967 OopTaskQueue* work_q);
3969 private:
3970 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3971 void do_work_steal(int i);
3972 void bump_global_finger(HeapWord* f);
3973 };
3975 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3976 assert(_task != NULL, "Error");
3977 return _task->yielding();
3978 // Note that we do not need the disjunct || _task->should_yield() above
3979 // because we want terminating threads to yield only if the task
3980 // is already in the midst of yielding, which happens only after at least one
3981 // thread has yielded.
3982 }
3984 void CMSConcMarkingTerminator::yield() {
3985 if (_task->should_yield()) {
3986 _task->yield();
3987 } else {
3988 ParallelTaskTerminator::yield();
3989 }
3990 }
3992 ////////////////////////////////////////////////////////////////
3993 // Concurrent Marking Algorithm Sketch
3994 ////////////////////////////////////////////////////////////////
3995 // Until all tasks exhausted (both spaces):
3996 // -- claim next available chunk
3997 // -- bump global finger via CAS
3998 // -- find first object that starts in this chunk
3999 // and start scanning bitmap from that position
4000 // -- scan marked objects for oops
4001 // -- CAS-mark target, and if successful:
4002 // . if target oop is above global finger (volatile read)
4003 // nothing to do
4004 // . if target oop is in chunk and above local finger
4005 // then nothing to do
4006 // . else push on work-queue
4007 // -- Deal with possible overflow issues:
4008 // . local work-queue overflow causes stuff to be pushed on
4009 // global (common) overflow queue
4010 // . always first empty local work queue
4011 // . then get a batch of oops from global work queue if any
4012 // . then do work stealing
4013 // -- When all tasks claimed (both spaces)
4014 // and local work queue empty,
4015 // then in a loop do:
4016 // . check global overflow stack; steal a batch of oops and trace
4017 // . try to steal from other threads oif GOS is empty
4018 // . if neither is available, offer termination
4019 // -- Terminate and return result
4020 //
4021 void CMSConcMarkingTask::work(uint worker_id) {
4022 elapsedTimer _timer;
4023 ResourceMark rm;
4024 HandleMark hm;
4026 DEBUG_ONLY(_collector->verify_overflow_empty();)
4028 // Before we begin work, our work queue should be empty
4029 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
4030 // Scan the bitmap covering _cms_space, tracing through grey objects.
4031 _timer.start();
4032 do_scan_and_mark(worker_id, _cms_space);
4033 _timer.stop();
4034 if (PrintCMSStatistics != 0) {
4035 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
4036 worker_id, _timer.seconds());
4037 // XXX: need xxx/xxx type of notation, two timers
4038 }
4040 // ... do work stealing
4041 _timer.reset();
4042 _timer.start();
4043 do_work_steal(worker_id);
4044 _timer.stop();
4045 if (PrintCMSStatistics != 0) {
4046 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
4047 worker_id, _timer.seconds());
4048 // XXX: need xxx/xxx type of notation, two timers
4049 }
4050 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
4051 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
4052 // Note that under the current task protocol, the
4053 // following assertion is true even of the spaces
4054 // expanded since the completion of the concurrent
4055 // marking. XXX This will likely change under a strict
4056 // ABORT semantics.
4057 // After perm removal the comparison was changed to
4058 // greater than or equal to from strictly greater than.
4059 // Before perm removal the highest address sweep would
4060 // have been at the end of perm gen but now is at the
4061 // end of the tenured gen.
4062 assert(_global_finger >= _cms_space->end(),
4063 "All tasks have been completed");
4064 DEBUG_ONLY(_collector->verify_overflow_empty();)
4065 }
4067 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
4068 HeapWord* read = _global_finger;
4069 HeapWord* cur = read;
4070 while (f > read) {
4071 cur = read;
4072 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
4073 if (cur == read) {
4074 // our cas succeeded
4075 assert(_global_finger >= f, "protocol consistency");
4076 break;
4077 }
4078 }
4079 }
4081 // This is really inefficient, and should be redone by
4082 // using (not yet available) block-read and -write interfaces to the
4083 // stack and the work_queue. XXX FIX ME !!!
4084 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
4085 OopTaskQueue* work_q) {
4086 // Fast lock-free check
4087 if (ovflw_stk->length() == 0) {
4088 return false;
4089 }
4090 assert(work_q->size() == 0, "Shouldn't steal");
4091 MutexLockerEx ml(ovflw_stk->par_lock(),
4092 Mutex::_no_safepoint_check_flag);
4093 // Grab up to 1/4 the size of the work queue
4094 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4095 (size_t)ParGCDesiredObjsFromOverflowList);
4096 num = MIN2(num, ovflw_stk->length());
4097 for (int i = (int) num; i > 0; i--) {
4098 oop cur = ovflw_stk->pop();
4099 assert(cur != NULL, "Counted wrong?");
4100 work_q->push(cur);
4101 }
4102 return num > 0;
4103 }
4105 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
4106 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4107 int n_tasks = pst->n_tasks();
4108 // We allow that there may be no tasks to do here because
4109 // we are restarting after a stack overflow.
4110 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
4111 uint nth_task = 0;
4113 HeapWord* aligned_start = sp->bottom();
4114 if (sp->used_region().contains(_restart_addr)) {
4115 // Align down to a card boundary for the start of 0th task
4116 // for this space.
4117 aligned_start =
4118 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
4119 CardTableModRefBS::card_size);
4120 }
4122 size_t chunk_size = sp->marking_task_size();
4123 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4124 // Having claimed the nth task in this space,
4125 // compute the chunk that it corresponds to:
4126 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
4127 aligned_start + (nth_task+1)*chunk_size);
4128 // Try and bump the global finger via a CAS;
4129 // note that we need to do the global finger bump
4130 // _before_ taking the intersection below, because
4131 // the task corresponding to that region will be
4132 // deemed done even if the used_region() expands
4133 // because of allocation -- as it almost certainly will
4134 // during start-up while the threads yield in the
4135 // closure below.
4136 HeapWord* finger = span.end();
4137 bump_global_finger(finger); // atomically
4138 // There are null tasks here corresponding to chunks
4139 // beyond the "top" address of the space.
4140 span = span.intersection(sp->used_region());
4141 if (!span.is_empty()) { // Non-null task
4142 HeapWord* prev_obj;
4143 assert(!span.contains(_restart_addr) || nth_task == 0,
4144 "Inconsistency");
4145 if (nth_task == 0) {
4146 // For the 0th task, we'll not need to compute a block_start.
4147 if (span.contains(_restart_addr)) {
4148 // In the case of a restart because of stack overflow,
4149 // we might additionally skip a chunk prefix.
4150 prev_obj = _restart_addr;
4151 } else {
4152 prev_obj = span.start();
4153 }
4154 } else {
4155 // We want to skip the first object because
4156 // the protocol is to scan any object in its entirety
4157 // that _starts_ in this span; a fortiori, any
4158 // object starting in an earlier span is scanned
4159 // as part of an earlier claimed task.
4160 // Below we use the "careful" version of block_start
4161 // so we do not try to navigate uninitialized objects.
4162 prev_obj = sp->block_start_careful(span.start());
4163 // Below we use a variant of block_size that uses the
4164 // Printezis bits to avoid waiting for allocated
4165 // objects to become initialized/parsable.
4166 while (prev_obj < span.start()) {
4167 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
4168 if (sz > 0) {
4169 prev_obj += sz;
4170 } else {
4171 // In this case we may end up doing a bit of redundant
4172 // scanning, but that appears unavoidable, short of
4173 // locking the free list locks; see bug 6324141.
4174 break;
4175 }
4176 }
4177 }
4178 if (prev_obj < span.end()) {
4179 MemRegion my_span = MemRegion(prev_obj, span.end());
4180 // Do the marking work within a non-empty span --
4181 // the last argument to the constructor indicates whether the
4182 // iteration should be incremental with periodic yields.
4183 Par_MarkFromRootsClosure cl(this, _collector, my_span,
4184 &_collector->_markBitMap,
4185 work_queue(i),
4186 &_collector->_markStack,
4187 _asynch);
4188 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
4189 } // else nothing to do for this task
4190 } // else nothing to do for this task
4191 }
4192 // We'd be tempted to assert here that since there are no
4193 // more tasks left to claim in this space, the global_finger
4194 // must exceed space->top() and a fortiori space->end(). However,
4195 // that would not quite be correct because the bumping of
4196 // global_finger occurs strictly after the claiming of a task,
4197 // so by the time we reach here the global finger may not yet
4198 // have been bumped up by the thread that claimed the last
4199 // task.
4200 pst->all_tasks_completed();
4201 }
4203 class Par_ConcMarkingClosure: public CMSOopClosure {
4204 private:
4205 CMSCollector* _collector;
4206 CMSConcMarkingTask* _task;
4207 MemRegion _span;
4208 CMSBitMap* _bit_map;
4209 CMSMarkStack* _overflow_stack;
4210 OopTaskQueue* _work_queue;
4211 protected:
4212 DO_OOP_WORK_DEFN
4213 public:
4214 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
4215 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
4216 CMSOopClosure(collector->ref_processor()),
4217 _collector(collector),
4218 _task(task),
4219 _span(collector->_span),
4220 _work_queue(work_queue),
4221 _bit_map(bit_map),
4222 _overflow_stack(overflow_stack)
4223 { }
4224 virtual void do_oop(oop* p);
4225 virtual void do_oop(narrowOop* p);
4227 void trim_queue(size_t max);
4228 void handle_stack_overflow(HeapWord* lost);
4229 void do_yield_check() {
4230 if (_task->should_yield()) {
4231 _task->yield();
4232 }
4233 }
4234 };
4236 // Grey object scanning during work stealing phase --
4237 // the salient assumption here is that any references
4238 // that are in these stolen objects being scanned must
4239 // already have been initialized (else they would not have
4240 // been published), so we do not need to check for
4241 // uninitialized objects before pushing here.
4242 void Par_ConcMarkingClosure::do_oop(oop obj) {
4243 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
4244 HeapWord* addr = (HeapWord*)obj;
4245 // Check if oop points into the CMS generation
4246 // and is not marked
4247 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
4248 // a white object ...
4249 // If we manage to "claim" the object, by being the
4250 // first thread to mark it, then we push it on our
4251 // marking stack
4252 if (_bit_map->par_mark(addr)) { // ... now grey
4253 // push on work queue (grey set)
4254 bool simulate_overflow = false;
4255 NOT_PRODUCT(
4256 if (CMSMarkStackOverflowALot &&
4257 _collector->simulate_overflow()) {
4258 // simulate a stack overflow
4259 simulate_overflow = true;
4260 }
4261 )
4262 if (simulate_overflow ||
4263 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
4264 // stack overflow
4265 if (PrintCMSStatistics != 0) {
4266 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
4267 SIZE_FORMAT, _overflow_stack->capacity());
4268 }
4269 // We cannot assert that the overflow stack is full because
4270 // it may have been emptied since.
4271 assert(simulate_overflow ||
4272 _work_queue->size() == _work_queue->max_elems(),
4273 "Else push should have succeeded");
4274 handle_stack_overflow(addr);
4275 }
4276 } // Else, some other thread got there first
4277 do_yield_check();
4278 }
4279 }
4281 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4282 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4284 void Par_ConcMarkingClosure::trim_queue(size_t max) {
4285 while (_work_queue->size() > max) {
4286 oop new_oop;
4287 if (_work_queue->pop_local(new_oop)) {
4288 assert(new_oop->is_oop(), "Should be an oop");
4289 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
4290 assert(_span.contains((HeapWord*)new_oop), "Not in span");
4291 new_oop->oop_iterate(this); // do_oop() above
4292 do_yield_check();
4293 }
4294 }
4295 }
4297 // Upon stack overflow, we discard (part of) the stack,
4298 // remembering the least address amongst those discarded
4299 // in CMSCollector's _restart_address.
4300 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
4301 // We need to do this under a mutex to prevent other
4302 // workers from interfering with the work done below.
4303 MutexLockerEx ml(_overflow_stack->par_lock(),
4304 Mutex::_no_safepoint_check_flag);
4305 // Remember the least grey address discarded
4306 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
4307 _collector->lower_restart_addr(ra);
4308 _overflow_stack->reset(); // discard stack contents
4309 _overflow_stack->expand(); // expand the stack if possible
4310 }
4313 void CMSConcMarkingTask::do_work_steal(int i) {
4314 OopTaskQueue* work_q = work_queue(i);
4315 oop obj_to_scan;
4316 CMSBitMap* bm = &(_collector->_markBitMap);
4317 CMSMarkStack* ovflw = &(_collector->_markStack);
4318 int* seed = _collector->hash_seed(i);
4319 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
4320 while (true) {
4321 cl.trim_queue(0);
4322 assert(work_q->size() == 0, "Should have been emptied above");
4323 if (get_work_from_overflow_stack(ovflw, work_q)) {
4324 // Can't assert below because the work obtained from the
4325 // overflow stack may already have been stolen from us.
4326 // assert(work_q->size() > 0, "Work from overflow stack");
4327 continue;
4328 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4329 assert(obj_to_scan->is_oop(), "Should be an oop");
4330 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
4331 obj_to_scan->oop_iterate(&cl);
4332 } else if (terminator()->offer_termination(&_term_term)) {
4333 assert(work_q->size() == 0, "Impossible!");
4334 break;
4335 } else if (yielding() || should_yield()) {
4336 yield();
4337 }
4338 }
4339 }
4341 // This is run by the CMS (coordinator) thread.
4342 void CMSConcMarkingTask::coordinator_yield() {
4343 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4344 "CMS thread should hold CMS token");
4345 // First give up the locks, then yield, then re-lock
4346 // We should probably use a constructor/destructor idiom to
4347 // do this unlock/lock or modify the MutexUnlocker class to
4348 // serve our purpose. XXX
4349 assert_lock_strong(_bit_map_lock);
4350 _bit_map_lock->unlock();
4351 ConcurrentMarkSweepThread::desynchronize(true);
4352 ConcurrentMarkSweepThread::acknowledge_yield_request();
4353 _collector->stopTimer();
4354 if (PrintCMSStatistics != 0) {
4355 _collector->incrementYields();
4356 }
4357 _collector->icms_wait();
4359 // It is possible for whichever thread initiated the yield request
4360 // not to get a chance to wake up and take the bitmap lock between
4361 // this thread releasing it and reacquiring it. So, while the
4362 // should_yield() flag is on, let's sleep for a bit to give the
4363 // other thread a chance to wake up. The limit imposed on the number
4364 // of iterations is defensive, to avoid any unforseen circumstances
4365 // putting us into an infinite loop. Since it's always been this
4366 // (coordinator_yield()) method that was observed to cause the
4367 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4368 // which is by default non-zero. For the other seven methods that
4369 // also perform the yield operation, as are using a different
4370 // parameter (CMSYieldSleepCount) which is by default zero. This way we
4371 // can enable the sleeping for those methods too, if necessary.
4372 // See 6442774.
4373 //
4374 // We really need to reconsider the synchronization between the GC
4375 // thread and the yield-requesting threads in the future and we
4376 // should really use wait/notify, which is the recommended
4377 // way of doing this type of interaction. Additionally, we should
4378 // consolidate the eight methods that do the yield operation and they
4379 // are almost identical into one for better maintenability and
4380 // readability. See 6445193.
4381 //
4382 // Tony 2006.06.29
4383 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4384 ConcurrentMarkSweepThread::should_yield() &&
4385 !CMSCollector::foregroundGCIsActive(); ++i) {
4386 os::sleep(Thread::current(), 1, false);
4387 ConcurrentMarkSweepThread::acknowledge_yield_request();
4388 }
4390 ConcurrentMarkSweepThread::synchronize(true);
4391 _bit_map_lock->lock_without_safepoint_check();
4392 _collector->startTimer();
4393 }
4395 bool CMSCollector::do_marking_mt(bool asynch) {
4396 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
4397 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers(
4398 conc_workers()->total_workers(),
4399 conc_workers()->active_workers(),
4400 Threads::number_of_non_daemon_threads());
4401 conc_workers()->set_active_workers(num_workers);
4403 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4405 CMSConcMarkingTask tsk(this,
4406 cms_space,
4407 asynch,
4408 conc_workers(),
4409 task_queues());
4411 // Since the actual number of workers we get may be different
4412 // from the number we requested above, do we need to do anything different
4413 // below? In particular, may be we need to subclass the SequantialSubTasksDone
4414 // class?? XXX
4415 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4417 // Refs discovery is already non-atomic.
4418 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4419 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
4420 conc_workers()->start_task(&tsk);
4421 while (tsk.yielded()) {
4422 tsk.coordinator_yield();
4423 conc_workers()->continue_task(&tsk);
4424 }
4425 // If the task was aborted, _restart_addr will be non-NULL
4426 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4427 while (_restart_addr != NULL) {
4428 // XXX For now we do not make use of ABORTED state and have not
4429 // yet implemented the right abort semantics (even in the original
4430 // single-threaded CMS case). That needs some more investigation
4431 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4432 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4433 // If _restart_addr is non-NULL, a marking stack overflow
4434 // occurred; we need to do a fresh marking iteration from the
4435 // indicated restart address.
4436 if (_foregroundGCIsActive && asynch) {
4437 // We may be running into repeated stack overflows, having
4438 // reached the limit of the stack size, while making very
4439 // slow forward progress. It may be best to bail out and
4440 // let the foreground collector do its job.
4441 // Clear _restart_addr, so that foreground GC
4442 // works from scratch. This avoids the headache of
4443 // a "rescan" which would otherwise be needed because
4444 // of the dirty mod union table & card table.
4445 _restart_addr = NULL;
4446 return false;
4447 }
4448 // Adjust the task to restart from _restart_addr
4449 tsk.reset(_restart_addr);
4450 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4451 _restart_addr);
4452 _restart_addr = NULL;
4453 // Get the workers going again
4454 conc_workers()->start_task(&tsk);
4455 while (tsk.yielded()) {
4456 tsk.coordinator_yield();
4457 conc_workers()->continue_task(&tsk);
4458 }
4459 }
4460 assert(tsk.completed(), "Inconsistency");
4461 assert(tsk.result() == true, "Inconsistency");
4462 return true;
4463 }
4465 bool CMSCollector::do_marking_st(bool asynch) {
4466 ResourceMark rm;
4467 HandleMark hm;
4469 // Temporarily make refs discovery single threaded (non-MT)
4470 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
4471 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4472 &_markStack, CMSYield && asynch);
4473 // the last argument to iterate indicates whether the iteration
4474 // should be incremental with periodic yields.
4475 _markBitMap.iterate(&markFromRootsClosure);
4476 // If _restart_addr is non-NULL, a marking stack overflow
4477 // occurred; we need to do a fresh iteration from the
4478 // indicated restart address.
4479 while (_restart_addr != NULL) {
4480 if (_foregroundGCIsActive && asynch) {
4481 // We may be running into repeated stack overflows, having
4482 // reached the limit of the stack size, while making very
4483 // slow forward progress. It may be best to bail out and
4484 // let the foreground collector do its job.
4485 // Clear _restart_addr, so that foreground GC
4486 // works from scratch. This avoids the headache of
4487 // a "rescan" which would otherwise be needed because
4488 // of the dirty mod union table & card table.
4489 _restart_addr = NULL;
4490 return false; // indicating failure to complete marking
4491 }
4492 // Deal with stack overflow:
4493 // we restart marking from _restart_addr
4494 HeapWord* ra = _restart_addr;
4495 markFromRootsClosure.reset(ra);
4496 _restart_addr = NULL;
4497 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4498 }
4499 return true;
4500 }
4502 void CMSCollector::preclean() {
4503 check_correct_thread_executing();
4504 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4505 verify_work_stacks_empty();
4506 verify_overflow_empty();
4507 _abort_preclean = false;
4508 if (CMSPrecleaningEnabled) {
4509 if (!CMSEdenChunksRecordAlways) {
4510 _eden_chunk_index = 0;
4511 }
4512 size_t used = get_eden_used();
4513 size_t capacity = get_eden_capacity();
4514 // Don't start sampling unless we will get sufficiently
4515 // many samples.
4516 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4517 * CMSScheduleRemarkEdenPenetration)) {
4518 _start_sampling = true;
4519 } else {
4520 _start_sampling = false;
4521 }
4522 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4523 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
4524 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4525 }
4526 CMSTokenSync x(true); // is cms thread
4527 if (CMSPrecleaningEnabled) {
4528 sample_eden();
4529 _collectorState = AbortablePreclean;
4530 } else {
4531 _collectorState = FinalMarking;
4532 }
4533 verify_work_stacks_empty();
4534 verify_overflow_empty();
4535 }
4537 // Try and schedule the remark such that young gen
4538 // occupancy is CMSScheduleRemarkEdenPenetration %.
4539 void CMSCollector::abortable_preclean() {
4540 check_correct_thread_executing();
4541 assert(CMSPrecleaningEnabled, "Inconsistent control state");
4542 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4544 // If Eden's current occupancy is below this threshold,
4545 // immediately schedule the remark; else preclean
4546 // past the next scavenge in an effort to
4547 // schedule the pause as described avove. By choosing
4548 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4549 // we will never do an actual abortable preclean cycle.
4550 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4551 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4552 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
4553 // We need more smarts in the abortable preclean
4554 // loop below to deal with cases where allocation
4555 // in young gen is very very slow, and our precleaning
4556 // is running a losing race against a horde of
4557 // mutators intent on flooding us with CMS updates
4558 // (dirty cards).
4559 // One, admittedly dumb, strategy is to give up
4560 // after a certain number of abortable precleaning loops
4561 // or after a certain maximum time. We want to make
4562 // this smarter in the next iteration.
4563 // XXX FIX ME!!! YSR
4564 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4565 while (!(should_abort_preclean() ||
4566 ConcurrentMarkSweepThread::should_terminate())) {
4567 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4568 cumworkdone += workdone;
4569 loops++;
4570 // Voluntarily terminate abortable preclean phase if we have
4571 // been at it for too long.
4572 if ((CMSMaxAbortablePrecleanLoops != 0) &&
4573 loops >= CMSMaxAbortablePrecleanLoops) {
4574 if (PrintGCDetails) {
4575 gclog_or_tty->print(" CMS: abort preclean due to loops ");
4576 }
4577 break;
4578 }
4579 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4580 if (PrintGCDetails) {
4581 gclog_or_tty->print(" CMS: abort preclean due to time ");
4582 }
4583 break;
4584 }
4585 // If we are doing little work each iteration, we should
4586 // take a short break.
4587 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4588 // Sleep for some time, waiting for work to accumulate
4589 stopTimer();
4590 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4591 startTimer();
4592 waited++;
4593 }
4594 }
4595 if (PrintCMSStatistics > 0) {
4596 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4597 loops, waited, cumworkdone);
4598 }
4599 }
4600 CMSTokenSync x(true); // is cms thread
4601 if (_collectorState != Idling) {
4602 assert(_collectorState == AbortablePreclean,
4603 "Spontaneous state transition?");
4604 _collectorState = FinalMarking;
4605 } // Else, a foreground collection completed this CMS cycle.
4606 return;
4607 }
4609 // Respond to an Eden sampling opportunity
4610 void CMSCollector::sample_eden() {
4611 // Make sure a young gc cannot sneak in between our
4612 // reading and recording of a sample.
4613 assert(Thread::current()->is_ConcurrentGC_thread(),
4614 "Only the cms thread may collect Eden samples");
4615 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4616 "Should collect samples while holding CMS token");
4617 if (!_start_sampling) {
4618 return;
4619 }
4620 // When CMSEdenChunksRecordAlways is true, the eden chunk array
4621 // is populated by the young generation.
4622 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
4623 if (_eden_chunk_index < _eden_chunk_capacity) {
4624 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
4625 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4626 "Unexpected state of Eden");
4627 // We'd like to check that what we just sampled is an oop-start address;
4628 // however, we cannot do that here since the object may not yet have been
4629 // initialized. So we'll instead do the check when we _use_ this sample
4630 // later.
4631 if (_eden_chunk_index == 0 ||
4632 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4633 _eden_chunk_array[_eden_chunk_index-1])
4634 >= CMSSamplingGrain)) {
4635 _eden_chunk_index++; // commit sample
4636 }
4637 }
4638 }
4639 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4640 size_t used = get_eden_used();
4641 size_t capacity = get_eden_capacity();
4642 assert(used <= capacity, "Unexpected state of Eden");
4643 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4644 _abort_preclean = true;
4645 }
4646 }
4647 }
4650 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4651 assert(_collectorState == Precleaning ||
4652 _collectorState == AbortablePreclean, "incorrect state");
4653 ResourceMark rm;
4654 HandleMark hm;
4656 // Precleaning is currently not MT but the reference processor
4657 // may be set for MT. Disable it temporarily here.
4658 ReferenceProcessor* rp = ref_processor();
4659 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
4661 // Do one pass of scrubbing the discovered reference lists
4662 // to remove any reference objects with strongly-reachable
4663 // referents.
4664 if (clean_refs) {
4665 CMSPrecleanRefsYieldClosure yield_cl(this);
4666 assert(rp->span().equals(_span), "Spans should be equal");
4667 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4668 &_markStack, true /* preclean */);
4669 CMSDrainMarkingStackClosure complete_trace(this,
4670 _span, &_markBitMap, &_markStack,
4671 &keep_alive, true /* preclean */);
4673 // We don't want this step to interfere with a young
4674 // collection because we don't want to take CPU
4675 // or memory bandwidth away from the young GC threads
4676 // (which may be as many as there are CPUs).
4677 // Note that we don't need to protect ourselves from
4678 // interference with mutators because they can't
4679 // manipulate the discovered reference lists nor affect
4680 // the computed reachability of the referents, the
4681 // only properties manipulated by the precleaning
4682 // of these reference lists.
4683 stopTimer();
4684 CMSTokenSyncWithLocks x(true /* is cms thread */,
4685 bitMapLock());
4686 startTimer();
4687 sample_eden();
4689 // The following will yield to allow foreground
4690 // collection to proceed promptly. XXX YSR:
4691 // The code in this method may need further
4692 // tweaking for better performance and some restructuring
4693 // for cleaner interfaces.
4694 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
4695 rp->preclean_discovered_references(
4696 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
4697 gc_timer);
4698 }
4700 if (clean_survivor) { // preclean the active survivor space(s)
4701 assert(_young_gen->kind() == Generation::DefNew ||
4702 _young_gen->kind() == Generation::ParNew ||
4703 _young_gen->kind() == Generation::ASParNew,
4704 "incorrect type for cast");
4705 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4706 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4707 &_markBitMap, &_modUnionTable,
4708 &_markStack, true /* precleaning phase */);
4709 stopTimer();
4710 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4711 bitMapLock());
4712 startTimer();
4713 unsigned int before_count =
4714 GenCollectedHeap::heap()->total_collections();
4715 SurvivorSpacePrecleanClosure
4716 sss_cl(this, _span, &_markBitMap, &_markStack,
4717 &pam_cl, before_count, CMSYield);
4718 dng->from()->object_iterate_careful(&sss_cl);
4719 dng->to()->object_iterate_careful(&sss_cl);
4720 }
4721 MarkRefsIntoAndScanClosure
4722 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4723 &_markStack, this, CMSYield,
4724 true /* precleaning phase */);
4725 // CAUTION: The following closure has persistent state that may need to
4726 // be reset upon a decrease in the sequence of addresses it
4727 // processes.
4728 ScanMarkedObjectsAgainCarefullyClosure
4729 smoac_cl(this, _span,
4730 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
4732 // Preclean dirty cards in ModUnionTable and CardTable using
4733 // appropriate convergence criterion;
4734 // repeat CMSPrecleanIter times unless we find that
4735 // we are losing.
4736 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4737 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4738 "Bad convergence multiplier");
4739 assert(CMSPrecleanThreshold >= 100,
4740 "Unreasonably low CMSPrecleanThreshold");
4742 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4743 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4744 numIter < CMSPrecleanIter;
4745 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4746 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4747 if (Verbose && PrintGCDetails) {
4748 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4749 }
4750 // Either there are very few dirty cards, so re-mark
4751 // pause will be small anyway, or our pre-cleaning isn't
4752 // that much faster than the rate at which cards are being
4753 // dirtied, so we might as well stop and re-mark since
4754 // precleaning won't improve our re-mark time by much.
4755 if (curNumCards <= CMSPrecleanThreshold ||
4756 (numIter > 0 &&
4757 (curNumCards * CMSPrecleanDenominator >
4758 lastNumCards * CMSPrecleanNumerator))) {
4759 numIter++;
4760 cumNumCards += curNumCards;
4761 break;
4762 }
4763 }
4765 preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
4767 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4768 cumNumCards += curNumCards;
4769 if (PrintGCDetails && PrintCMSStatistics != 0) {
4770 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4771 curNumCards, cumNumCards, numIter);
4772 }
4773 return cumNumCards; // as a measure of useful work done
4774 }
4776 // PRECLEANING NOTES:
4777 // Precleaning involves:
4778 // . reading the bits of the modUnionTable and clearing the set bits.
4779 // . For the cards corresponding to the set bits, we scan the
4780 // objects on those cards. This means we need the free_list_lock
4781 // so that we can safely iterate over the CMS space when scanning
4782 // for oops.
4783 // . When we scan the objects, we'll be both reading and setting
4784 // marks in the marking bit map, so we'll need the marking bit map.
4785 // . For protecting _collector_state transitions, we take the CGC_lock.
4786 // Note that any races in the reading of of card table entries by the
4787 // CMS thread on the one hand and the clearing of those entries by the
4788 // VM thread or the setting of those entries by the mutator threads on the
4789 // other are quite benign. However, for efficiency it makes sense to keep
4790 // the VM thread from racing with the CMS thread while the latter is
4791 // dirty card info to the modUnionTable. We therefore also use the
4792 // CGC_lock to protect the reading of the card table and the mod union
4793 // table by the CM thread.
4794 // . We run concurrently with mutator updates, so scanning
4795 // needs to be done carefully -- we should not try to scan
4796 // potentially uninitialized objects.
4797 //
4798 // Locking strategy: While holding the CGC_lock, we scan over and
4799 // reset a maximal dirty range of the mod union / card tables, then lock
4800 // the free_list_lock and bitmap lock to do a full marking, then
4801 // release these locks; and repeat the cycle. This allows for a
4802 // certain amount of fairness in the sharing of these locks between
4803 // the CMS collector on the one hand, and the VM thread and the
4804 // mutators on the other.
4806 // NOTE: preclean_mod_union_table() and preclean_card_table()
4807 // further below are largely identical; if you need to modify
4808 // one of these methods, please check the other method too.
4810 size_t CMSCollector::preclean_mod_union_table(
4811 ConcurrentMarkSweepGeneration* gen,
4812 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4813 verify_work_stacks_empty();
4814 verify_overflow_empty();
4816 // strategy: starting with the first card, accumulate contiguous
4817 // ranges of dirty cards; clear these cards, then scan the region
4818 // covered by these cards.
4820 // Since all of the MUT is committed ahead, we can just use
4821 // that, in case the generations expand while we are precleaning.
4822 // It might also be fine to just use the committed part of the
4823 // generation, but we might potentially miss cards when the
4824 // generation is rapidly expanding while we are in the midst
4825 // of precleaning.
4826 HeapWord* startAddr = gen->reserved().start();
4827 HeapWord* endAddr = gen->reserved().end();
4829 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4831 size_t numDirtyCards, cumNumDirtyCards;
4832 HeapWord *nextAddr, *lastAddr;
4833 for (cumNumDirtyCards = numDirtyCards = 0,
4834 nextAddr = lastAddr = startAddr;
4835 nextAddr < endAddr;
4836 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4838 ResourceMark rm;
4839 HandleMark hm;
4841 MemRegion dirtyRegion;
4842 {
4843 stopTimer();
4844 // Potential yield point
4845 CMSTokenSync ts(true);
4846 startTimer();
4847 sample_eden();
4848 // Get dirty region starting at nextOffset (inclusive),
4849 // simultaneously clearing it.
4850 dirtyRegion =
4851 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4852 assert(dirtyRegion.start() >= nextAddr,
4853 "returned region inconsistent?");
4854 }
4855 // Remember where the next search should begin.
4856 // The returned region (if non-empty) is a right open interval,
4857 // so lastOffset is obtained from the right end of that
4858 // interval.
4859 lastAddr = dirtyRegion.end();
4860 // Should do something more transparent and less hacky XXX
4861 numDirtyCards =
4862 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4864 // We'll scan the cards in the dirty region (with periodic
4865 // yields for foreground GC as needed).
4866 if (!dirtyRegion.is_empty()) {
4867 assert(numDirtyCards > 0, "consistency check");
4868 HeapWord* stop_point = NULL;
4869 stopTimer();
4870 // Potential yield point
4871 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4872 bitMapLock());
4873 startTimer();
4874 {
4875 verify_work_stacks_empty();
4876 verify_overflow_empty();
4877 sample_eden();
4878 stop_point =
4879 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4880 }
4881 if (stop_point != NULL) {
4882 // The careful iteration stopped early either because it found an
4883 // uninitialized object, or because we were in the midst of an
4884 // "abortable preclean", which should now be aborted. Redirty
4885 // the bits corresponding to the partially-scanned or unscanned
4886 // cards. We'll either restart at the next block boundary or
4887 // abort the preclean.
4888 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4889 "Should only be AbortablePreclean.");
4890 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4891 if (should_abort_preclean()) {
4892 break; // out of preclean loop
4893 } else {
4894 // Compute the next address at which preclean should pick up;
4895 // might need bitMapLock in order to read P-bits.
4896 lastAddr = next_card_start_after_block(stop_point);
4897 }
4898 }
4899 } else {
4900 assert(lastAddr == endAddr, "consistency check");
4901 assert(numDirtyCards == 0, "consistency check");
4902 break;
4903 }
4904 }
4905 verify_work_stacks_empty();
4906 verify_overflow_empty();
4907 return cumNumDirtyCards;
4908 }
4910 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4911 // below are largely identical; if you need to modify
4912 // one of these methods, please check the other method too.
4914 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4915 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4916 // strategy: it's similar to precleamModUnionTable above, in that
4917 // we accumulate contiguous ranges of dirty cards, mark these cards
4918 // precleaned, then scan the region covered by these cards.
4919 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4920 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4922 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4924 size_t numDirtyCards, cumNumDirtyCards;
4925 HeapWord *lastAddr, *nextAddr;
4927 for (cumNumDirtyCards = numDirtyCards = 0,
4928 nextAddr = lastAddr = startAddr;
4929 nextAddr < endAddr;
4930 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4932 ResourceMark rm;
4933 HandleMark hm;
4935 MemRegion dirtyRegion;
4936 {
4937 // See comments in "Precleaning notes" above on why we
4938 // do this locking. XXX Could the locking overheads be
4939 // too high when dirty cards are sparse? [I don't think so.]
4940 stopTimer();
4941 CMSTokenSync x(true); // is cms thread
4942 startTimer();
4943 sample_eden();
4944 // Get and clear dirty region from card table
4945 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4946 MemRegion(nextAddr, endAddr),
4947 true,
4948 CardTableModRefBS::precleaned_card_val());
4950 assert(dirtyRegion.start() >= nextAddr,
4951 "returned region inconsistent?");
4952 }
4953 lastAddr = dirtyRegion.end();
4954 numDirtyCards =
4955 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4957 if (!dirtyRegion.is_empty()) {
4958 stopTimer();
4959 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4960 startTimer();
4961 sample_eden();
4962 verify_work_stacks_empty();
4963 verify_overflow_empty();
4964 HeapWord* stop_point =
4965 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4966 if (stop_point != NULL) {
4967 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4968 "Should only be AbortablePreclean.");
4969 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4970 if (should_abort_preclean()) {
4971 break; // out of preclean loop
4972 } else {
4973 // Compute the next address at which preclean should pick up.
4974 lastAddr = next_card_start_after_block(stop_point);
4975 }
4976 }
4977 } else {
4978 break;
4979 }
4980 }
4981 verify_work_stacks_empty();
4982 verify_overflow_empty();
4983 return cumNumDirtyCards;
4984 }
4986 class PrecleanKlassClosure : public KlassClosure {
4987 CMKlassClosure _cm_klass_closure;
4988 public:
4989 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4990 void do_klass(Klass* k) {
4991 if (k->has_accumulated_modified_oops()) {
4992 k->clear_accumulated_modified_oops();
4994 _cm_klass_closure.do_klass(k);
4995 }
4996 }
4997 };
4999 // The freelist lock is needed to prevent asserts, is it really needed?
5000 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
5002 cl->set_freelistLock(freelistLock);
5004 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
5006 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
5007 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
5008 PrecleanKlassClosure preclean_klass_closure(cl);
5009 ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
5011 verify_work_stacks_empty();
5012 verify_overflow_empty();
5013 }
5015 void CMSCollector::checkpointRootsFinal(bool asynch,
5016 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
5017 assert(_collectorState == FinalMarking, "incorrect state transition?");
5018 check_correct_thread_executing();
5019 // world is stopped at this checkpoint
5020 assert(SafepointSynchronize::is_at_safepoint(),
5021 "world should be stopped");
5022 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5024 verify_work_stacks_empty();
5025 verify_overflow_empty();
5027 SpecializationStats::clear();
5028 if (PrintGCDetails) {
5029 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
5030 _young_gen->used() / K,
5031 _young_gen->capacity() / K);
5032 }
5033 if (asynch) {
5034 if (CMSScavengeBeforeRemark) {
5035 GenCollectedHeap* gch = GenCollectedHeap::heap();
5036 // Temporarily set flag to false, GCH->do_collection will
5037 // expect it to be false and set to true
5038 FlagSetting fl(gch->_is_gc_active, false);
5039 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
5040 PrintGCDetails && Verbose, true, _gc_timer_cm);)
5041 int level = _cmsGen->level() - 1;
5042 if (level >= 0) {
5043 gch->do_collection(true, // full (i.e. force, see below)
5044 false, // !clear_all_soft_refs
5045 0, // size
5046 false, // is_tlab
5047 level // max_level
5048 );
5049 }
5050 }
5051 FreelistLocker x(this);
5052 MutexLockerEx y(bitMapLock(),
5053 Mutex::_no_safepoint_check_flag);
5054 assert(!init_mark_was_synchronous, "but that's impossible!");
5055 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
5056 } else {
5057 // already have all the locks
5058 checkpointRootsFinalWork(asynch, clear_all_soft_refs,
5059 init_mark_was_synchronous);
5060 }
5061 verify_work_stacks_empty();
5062 verify_overflow_empty();
5063 SpecializationStats::print();
5064 }
5066 void CMSCollector::checkpointRootsFinalWork(bool asynch,
5067 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
5069 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);)
5071 assert(haveFreelistLocks(), "must have free list locks");
5072 assert_lock_strong(bitMapLock());
5074 if (UseAdaptiveSizePolicy) {
5075 size_policy()->checkpoint_roots_final_begin();
5076 }
5078 ResourceMark rm;
5079 HandleMark hm;
5081 GenCollectedHeap* gch = GenCollectedHeap::heap();
5083 if (should_unload_classes()) {
5084 CodeCache::gc_prologue();
5085 }
5086 assert(haveFreelistLocks(), "must have free list locks");
5087 assert_lock_strong(bitMapLock());
5089 if (!init_mark_was_synchronous) {
5090 // We might assume that we need not fill TLAB's when
5091 // CMSScavengeBeforeRemark is set, because we may have just done
5092 // a scavenge which would have filled all TLAB's -- and besides
5093 // Eden would be empty. This however may not always be the case --
5094 // for instance although we asked for a scavenge, it may not have
5095 // happened because of a JNI critical section. We probably need
5096 // a policy for deciding whether we can in that case wait until
5097 // the critical section releases and then do the remark following
5098 // the scavenge, and skip it here. In the absence of that policy,
5099 // or of an indication of whether the scavenge did indeed occur,
5100 // we cannot rely on TLAB's having been filled and must do
5101 // so here just in case a scavenge did not happen.
5102 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
5103 // Update the saved marks which may affect the root scans.
5104 gch->save_marks();
5106 if (CMSPrintEdenSurvivorChunks) {
5107 print_eden_and_survivor_chunk_arrays();
5108 }
5110 {
5111 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
5113 // Note on the role of the mod union table:
5114 // Since the marker in "markFromRoots" marks concurrently with
5115 // mutators, it is possible for some reachable objects not to have been
5116 // scanned. For instance, an only reference to an object A was
5117 // placed in object B after the marker scanned B. Unless B is rescanned,
5118 // A would be collected. Such updates to references in marked objects
5119 // are detected via the mod union table which is the set of all cards
5120 // dirtied since the first checkpoint in this GC cycle and prior to
5121 // the most recent young generation GC, minus those cleaned up by the
5122 // concurrent precleaning.
5123 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
5124 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm);
5125 do_remark_parallel();
5126 } else {
5127 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
5128 _gc_timer_cm);
5129 do_remark_non_parallel();
5130 }
5131 }
5132 } else {
5133 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
5134 // The initial mark was stop-world, so there's no rescanning to
5135 // do; go straight on to the next step below.
5136 }
5137 verify_work_stacks_empty();
5138 verify_overflow_empty();
5140 {
5141 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);)
5142 refProcessingWork(asynch, clear_all_soft_refs);
5143 }
5144 verify_work_stacks_empty();
5145 verify_overflow_empty();
5147 if (should_unload_classes()) {
5148 CodeCache::gc_epilogue();
5149 }
5150 JvmtiExport::gc_epilogue();
5152 // If we encountered any (marking stack / work queue) overflow
5153 // events during the current CMS cycle, take appropriate
5154 // remedial measures, where possible, so as to try and avoid
5155 // recurrence of that condition.
5156 assert(_markStack.isEmpty(), "No grey objects");
5157 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
5158 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
5159 if (ser_ovflw > 0) {
5160 if (PrintCMSStatistics != 0) {
5161 gclog_or_tty->print_cr("Marking stack overflow (benign) "
5162 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
5163 ", kac_preclean="SIZE_FORMAT")",
5164 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
5165 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
5166 }
5167 _markStack.expand();
5168 _ser_pmc_remark_ovflw = 0;
5169 _ser_pmc_preclean_ovflw = 0;
5170 _ser_kac_preclean_ovflw = 0;
5171 _ser_kac_ovflw = 0;
5172 }
5173 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
5174 if (PrintCMSStatistics != 0) {
5175 gclog_or_tty->print_cr("Work queue overflow (benign) "
5176 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
5177 _par_pmc_remark_ovflw, _par_kac_ovflw);
5178 }
5179 _par_pmc_remark_ovflw = 0;
5180 _par_kac_ovflw = 0;
5181 }
5182 if (PrintCMSStatistics != 0) {
5183 if (_markStack._hit_limit > 0) {
5184 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
5185 _markStack._hit_limit);
5186 }
5187 if (_markStack._failed_double > 0) {
5188 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
5189 " current capacity "SIZE_FORMAT,
5190 _markStack._failed_double,
5191 _markStack.capacity());
5192 }
5193 }
5194 _markStack._hit_limit = 0;
5195 _markStack._failed_double = 0;
5197 if ((VerifyAfterGC || VerifyDuringGC) &&
5198 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5199 verify_after_remark();
5200 }
5202 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
5204 // Change under the freelistLocks.
5205 _collectorState = Sweeping;
5206 // Call isAllClear() under bitMapLock
5207 assert(_modUnionTable.isAllClear(),
5208 "Should be clear by end of the final marking");
5209 assert(_ct->klass_rem_set()->mod_union_is_clear(),
5210 "Should be clear by end of the final marking");
5211 if (UseAdaptiveSizePolicy) {
5212 size_policy()->checkpoint_roots_final_end(gch->gc_cause());
5213 }
5214 }
5216 void CMSParInitialMarkTask::work(uint worker_id) {
5217 elapsedTimer _timer;
5218 ResourceMark rm;
5219 HandleMark hm;
5221 // ---------- scan from roots --------------
5222 _timer.start();
5223 GenCollectedHeap* gch = GenCollectedHeap::heap();
5224 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
5225 CMKlassClosure klass_closure(&par_mri_cl);
5227 // ---------- young gen roots --------------
5228 {
5229 work_on_young_gen_roots(worker_id, &par_mri_cl);
5230 _timer.stop();
5231 if (PrintCMSStatistics != 0) {
5232 gclog_or_tty->print_cr(
5233 "Finished young gen initial mark scan work in %dth thread: %3.3f sec",
5234 worker_id, _timer.seconds());
5235 }
5236 }
5238 // ---------- remaining roots --------------
5239 _timer.reset();
5240 _timer.start();
5241 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
5242 false, // yg was scanned above
5243 false, // this is parallel code
5244 false, // not scavenging
5245 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5246 &par_mri_cl,
5247 true, // walk all of code cache if (so & SO_CodeCache)
5248 NULL,
5249 &klass_closure);
5250 assert(_collector->should_unload_classes()
5251 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache),
5252 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5253 _timer.stop();
5254 if (PrintCMSStatistics != 0) {
5255 gclog_or_tty->print_cr(
5256 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec",
5257 worker_id, _timer.seconds());
5258 }
5259 }
5261 // Parallel remark task
5262 class CMSParRemarkTask: public CMSParMarkTask {
5263 CompactibleFreeListSpace* _cms_space;
5265 // The per-thread work queues, available here for stealing.
5266 OopTaskQueueSet* _task_queues;
5267 ParallelTaskTerminator _term;
5269 public:
5270 // A value of 0 passed to n_workers will cause the number of
5271 // workers to be taken from the active workers in the work gang.
5272 CMSParRemarkTask(CMSCollector* collector,
5273 CompactibleFreeListSpace* cms_space,
5274 int n_workers, FlexibleWorkGang* workers,
5275 OopTaskQueueSet* task_queues):
5276 CMSParMarkTask("Rescan roots and grey objects in parallel",
5277 collector, n_workers),
5278 _cms_space(cms_space),
5279 _task_queues(task_queues),
5280 _term(n_workers, task_queues) { }
5282 OopTaskQueueSet* task_queues() { return _task_queues; }
5284 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5286 ParallelTaskTerminator* terminator() { return &_term; }
5287 int n_workers() { return _n_workers; }
5289 void work(uint worker_id);
5291 private:
5292 // ... of dirty cards in old space
5293 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
5294 Par_MarkRefsIntoAndScanClosure* cl);
5296 // ... work stealing for the above
5297 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
5298 };
5300 class RemarkKlassClosure : public KlassClosure {
5301 CMKlassClosure _cm_klass_closure;
5302 public:
5303 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
5304 void do_klass(Klass* k) {
5305 // Check if we have modified any oops in the Klass during the concurrent marking.
5306 if (k->has_accumulated_modified_oops()) {
5307 k->clear_accumulated_modified_oops();
5309 // We could have transfered the current modified marks to the accumulated marks,
5310 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
5311 } else if (k->has_modified_oops()) {
5312 // Don't clear anything, this info is needed by the next young collection.
5313 } else {
5314 // No modified oops in the Klass.
5315 return;
5316 }
5318 // The klass has modified fields, need to scan the klass.
5319 _cm_klass_closure.do_klass(k);
5320 }
5321 };
5323 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
5324 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
5325 EdenSpace* eden_space = dng->eden();
5326 ContiguousSpace* from_space = dng->from();
5327 ContiguousSpace* to_space = dng->to();
5329 HeapWord** eca = _collector->_eden_chunk_array;
5330 size_t ect = _collector->_eden_chunk_index;
5331 HeapWord** sca = _collector->_survivor_chunk_array;
5332 size_t sct = _collector->_survivor_chunk_index;
5334 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
5335 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
5337 do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
5338 do_young_space_rescan(worker_id, cl, from_space, sca, sct);
5339 do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
5340 }
5342 // work_queue(i) is passed to the closure
5343 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
5344 // also is passed to do_dirty_card_rescan_tasks() and to
5345 // do_work_steal() to select the i-th task_queue.
5347 void CMSParRemarkTask::work(uint worker_id) {
5348 elapsedTimer _timer;
5349 ResourceMark rm;
5350 HandleMark hm;
5352 // ---------- rescan from roots --------------
5353 _timer.start();
5354 GenCollectedHeap* gch = GenCollectedHeap::heap();
5355 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
5356 _collector->_span, _collector->ref_processor(),
5357 &(_collector->_markBitMap),
5358 work_queue(worker_id));
5360 // Rescan young gen roots first since these are likely
5361 // coarsely partitioned and may, on that account, constitute
5362 // the critical path; thus, it's best to start off that
5363 // work first.
5364 // ---------- young gen roots --------------
5365 {
5366 work_on_young_gen_roots(worker_id, &par_mrias_cl);
5367 _timer.stop();
5368 if (PrintCMSStatistics != 0) {
5369 gclog_or_tty->print_cr(
5370 "Finished young gen rescan work in %dth thread: %3.3f sec",
5371 worker_id, _timer.seconds());
5372 }
5373 }
5375 // ---------- remaining roots --------------
5376 _timer.reset();
5377 _timer.start();
5378 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
5379 false, // yg was scanned above
5380 false, // this is parallel code
5381 false, // not scavenging
5382 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5383 &par_mrias_cl,
5384 true, // walk all of code cache if (so & SO_CodeCache)
5385 NULL,
5386 NULL); // The dirty klasses will be handled below
5387 assert(_collector->should_unload_classes()
5388 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache),
5389 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5390 _timer.stop();
5391 if (PrintCMSStatistics != 0) {
5392 gclog_or_tty->print_cr(
5393 "Finished remaining root rescan work in %dth thread: %3.3f sec",
5394 worker_id, _timer.seconds());
5395 }
5397 // ---------- unhandled CLD scanning ----------
5398 if (worker_id == 0) { // Single threaded at the moment.
5399 _timer.reset();
5400 _timer.start();
5402 // Scan all new class loader data objects and new dependencies that were
5403 // introduced during concurrent marking.
5404 ResourceMark rm;
5405 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5406 for (int i = 0; i < array->length(); i++) {
5407 par_mrias_cl.do_class_loader_data(array->at(i));
5408 }
5410 // We don't need to keep track of new CLDs anymore.
5411 ClassLoaderDataGraph::remember_new_clds(false);
5413 _timer.stop();
5414 if (PrintCMSStatistics != 0) {
5415 gclog_or_tty->print_cr(
5416 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec",
5417 worker_id, _timer.seconds());
5418 }
5419 }
5421 // ---------- dirty klass scanning ----------
5422 if (worker_id == 0) { // Single threaded at the moment.
5423 _timer.reset();
5424 _timer.start();
5426 // Scan all classes that was dirtied during the concurrent marking phase.
5427 RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
5428 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5430 _timer.stop();
5431 if (PrintCMSStatistics != 0) {
5432 gclog_or_tty->print_cr(
5433 "Finished dirty klass scanning work in %dth thread: %3.3f sec",
5434 worker_id, _timer.seconds());
5435 }
5436 }
5438 // We might have added oops to ClassLoaderData::_handles during the
5439 // concurrent marking phase. These oops point to newly allocated objects
5440 // that are guaranteed to be kept alive. Either by the direct allocation
5441 // code, or when the young collector processes the strong roots. Hence,
5442 // we don't have to revisit the _handles block during the remark phase.
5444 // ---------- rescan dirty cards ------------
5445 _timer.reset();
5446 _timer.start();
5448 // Do the rescan tasks for each of the two spaces
5449 // (cms_space) in turn.
5450 // "worker_id" is passed to select the task_queue for "worker_id"
5451 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
5452 _timer.stop();
5453 if (PrintCMSStatistics != 0) {
5454 gclog_or_tty->print_cr(
5455 "Finished dirty card rescan work in %dth thread: %3.3f sec",
5456 worker_id, _timer.seconds());
5457 }
5459 // ---------- steal work from other threads ...
5460 // ---------- ... and drain overflow list.
5461 _timer.reset();
5462 _timer.start();
5463 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
5464 _timer.stop();
5465 if (PrintCMSStatistics != 0) {
5466 gclog_or_tty->print_cr(
5467 "Finished work stealing in %dth thread: %3.3f sec",
5468 worker_id, _timer.seconds());
5469 }
5470 }
5472 // Note that parameter "i" is not used.
5473 void
5474 CMSParMarkTask::do_young_space_rescan(uint worker_id,
5475 OopsInGenClosure* cl, ContiguousSpace* space,
5476 HeapWord** chunk_array, size_t chunk_top) {
5477 // Until all tasks completed:
5478 // . claim an unclaimed task
5479 // . compute region boundaries corresponding to task claimed
5480 // using chunk_array
5481 // . par_oop_iterate(cl) over that region
5483 ResourceMark rm;
5484 HandleMark hm;
5486 SequentialSubTasksDone* pst = space->par_seq_tasks();
5488 uint nth_task = 0;
5489 uint n_tasks = pst->n_tasks();
5491 if (n_tasks > 0) {
5492 assert(pst->valid(), "Uninitialized use?");
5493 HeapWord *start, *end;
5494 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5495 // We claimed task # nth_task; compute its boundaries.
5496 if (chunk_top == 0) { // no samples were taken
5497 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
5498 start = space->bottom();
5499 end = space->top();
5500 } else if (nth_task == 0) {
5501 start = space->bottom();
5502 end = chunk_array[nth_task];
5503 } else if (nth_task < (uint)chunk_top) {
5504 assert(nth_task >= 1, "Control point invariant");
5505 start = chunk_array[nth_task - 1];
5506 end = chunk_array[nth_task];
5507 } else {
5508 assert(nth_task == (uint)chunk_top, "Control point invariant");
5509 start = chunk_array[chunk_top - 1];
5510 end = space->top();
5511 }
5512 MemRegion mr(start, end);
5513 // Verify that mr is in space
5514 assert(mr.is_empty() || space->used_region().contains(mr),
5515 "Should be in space");
5516 // Verify that "start" is an object boundary
5517 assert(mr.is_empty() || oop(mr.start())->is_oop(),
5518 "Should be an oop");
5519 space->par_oop_iterate(mr, cl);
5520 }
5521 pst->all_tasks_completed();
5522 }
5523 }
5525 void
5526 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5527 CompactibleFreeListSpace* sp, int i,
5528 Par_MarkRefsIntoAndScanClosure* cl) {
5529 // Until all tasks completed:
5530 // . claim an unclaimed task
5531 // . compute region boundaries corresponding to task claimed
5532 // . transfer dirty bits ct->mut for that region
5533 // . apply rescanclosure to dirty mut bits for that region
5535 ResourceMark rm;
5536 HandleMark hm;
5538 OopTaskQueue* work_q = work_queue(i);
5539 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5540 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5541 // CAUTION: This closure has state that persists across calls to
5542 // the work method dirty_range_iterate_clear() in that it has
5543 // imbedded in it a (subtype of) UpwardsObjectClosure. The
5544 // use of that state in the imbedded UpwardsObjectClosure instance
5545 // assumes that the cards are always iterated (even if in parallel
5546 // by several threads) in monotonically increasing order per each
5547 // thread. This is true of the implementation below which picks
5548 // card ranges (chunks) in monotonically increasing order globally
5549 // and, a-fortiori, in monotonically increasing order per thread
5550 // (the latter order being a subsequence of the former).
5551 // If the work code below is ever reorganized into a more chaotic
5552 // work-partitioning form than the current "sequential tasks"
5553 // paradigm, the use of that persistent state will have to be
5554 // revisited and modified appropriately. See also related
5555 // bug 4756801 work on which should examine this code to make
5556 // sure that the changes there do not run counter to the
5557 // assumptions made here and necessary for correctness and
5558 // efficiency. Note also that this code might yield inefficient
5559 // behaviour in the case of very large objects that span one or
5560 // more work chunks. Such objects would potentially be scanned
5561 // several times redundantly. Work on 4756801 should try and
5562 // address that performance anomaly if at all possible. XXX
5563 MemRegion full_span = _collector->_span;
5564 CMSBitMap* bm = &(_collector->_markBitMap); // shared
5565 MarkFromDirtyCardsClosure
5566 greyRescanClosure(_collector, full_span, // entire span of interest
5567 sp, bm, work_q, cl);
5569 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5570 assert(pst->valid(), "Uninitialized use?");
5571 uint nth_task = 0;
5572 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5573 MemRegion span = sp->used_region();
5574 HeapWord* start_addr = span.start();
5575 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5576 alignment);
5577 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5578 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5579 start_addr, "Check alignment");
5580 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5581 chunk_size, "Check alignment");
5583 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5584 // Having claimed the nth_task, compute corresponding mem-region,
5585 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
5586 // The alignment restriction ensures that we do not need any
5587 // synchronization with other gang-workers while setting or
5588 // clearing bits in thus chunk of the MUT.
5589 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5590 start_addr + (nth_task+1)*chunk_size);
5591 // The last chunk's end might be way beyond end of the
5592 // used region. In that case pull back appropriately.
5593 if (this_span.end() > end_addr) {
5594 this_span.set_end(end_addr);
5595 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5596 }
5597 // Iterate over the dirty cards covering this chunk, marking them
5598 // precleaned, and setting the corresponding bits in the mod union
5599 // table. Since we have been careful to partition at Card and MUT-word
5600 // boundaries no synchronization is needed between parallel threads.
5601 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5602 &modUnionClosure);
5604 // Having transferred these marks into the modUnionTable,
5605 // rescan the marked objects on the dirty cards in the modUnionTable.
5606 // Even if this is at a synchronous collection, the initial marking
5607 // may have been done during an asynchronous collection so there
5608 // may be dirty bits in the mod-union table.
5609 _collector->_modUnionTable.dirty_range_iterate_clear(
5610 this_span, &greyRescanClosure);
5611 _collector->_modUnionTable.verifyNoOneBitsInRange(
5612 this_span.start(),
5613 this_span.end());
5614 }
5615 pst->all_tasks_completed(); // declare that i am done
5616 }
5618 // . see if we can share work_queues with ParNew? XXX
5619 void
5620 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5621 int* seed) {
5622 OopTaskQueue* work_q = work_queue(i);
5623 NOT_PRODUCT(int num_steals = 0;)
5624 oop obj_to_scan;
5625 CMSBitMap* bm = &(_collector->_markBitMap);
5627 while (true) {
5628 // Completely finish any left over work from (an) earlier round(s)
5629 cl->trim_queue(0);
5630 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5631 (size_t)ParGCDesiredObjsFromOverflowList);
5632 // Now check if there's any work in the overflow list
5633 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5634 // only affects the number of attempts made to get work from the
5635 // overflow list and does not affect the number of workers. Just
5636 // pass ParallelGCThreads so this behavior is unchanged.
5637 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5638 work_q,
5639 ParallelGCThreads)) {
5640 // found something in global overflow list;
5641 // not yet ready to go stealing work from others.
5642 // We'd like to assert(work_q->size() != 0, ...)
5643 // because we just took work from the overflow list,
5644 // but of course we can't since all of that could have
5645 // been already stolen from us.
5646 // "He giveth and He taketh away."
5647 continue;
5648 }
5649 // Verify that we have no work before we resort to stealing
5650 assert(work_q->size() == 0, "Have work, shouldn't steal");
5651 // Try to steal from other queues that have work
5652 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5653 NOT_PRODUCT(num_steals++;)
5654 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5655 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5656 // Do scanning work
5657 obj_to_scan->oop_iterate(cl);
5658 // Loop around, finish this work, and try to steal some more
5659 } else if (terminator()->offer_termination()) {
5660 break; // nirvana from the infinite cycle
5661 }
5662 }
5663 NOT_PRODUCT(
5664 if (PrintCMSStatistics != 0) {
5665 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5666 }
5667 )
5668 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5669 "Else our work is not yet done");
5670 }
5672 // Record object boundaries in _eden_chunk_array by sampling the eden
5673 // top in the slow-path eden object allocation code path and record
5674 // the boundaries, if CMSEdenChunksRecordAlways is true. If
5675 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
5676 // sampling in sample_eden() that activates during the part of the
5677 // preclean phase.
5678 void CMSCollector::sample_eden_chunk() {
5679 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
5680 if (_eden_chunk_lock->try_lock()) {
5681 // Record a sample. This is the critical section. The contents
5682 // of the _eden_chunk_array have to be non-decreasing in the
5683 // address order.
5684 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
5685 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
5686 "Unexpected state of Eden");
5687 if (_eden_chunk_index == 0 ||
5688 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
5689 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
5690 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
5691 _eden_chunk_index++; // commit sample
5692 }
5693 _eden_chunk_lock->unlock();
5694 }
5695 }
5696 }
5698 // Return a thread-local PLAB recording array, as appropriate.
5699 void* CMSCollector::get_data_recorder(int thr_num) {
5700 if (_survivor_plab_array != NULL &&
5701 (CMSPLABRecordAlways ||
5702 (_collectorState > Marking && _collectorState < FinalMarking))) {
5703 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5704 ChunkArray* ca = &_survivor_plab_array[thr_num];
5705 ca->reset(); // clear it so that fresh data is recorded
5706 return (void*) ca;
5707 } else {
5708 return NULL;
5709 }
5710 }
5712 // Reset all the thread-local PLAB recording arrays
5713 void CMSCollector::reset_survivor_plab_arrays() {
5714 for (uint i = 0; i < ParallelGCThreads; i++) {
5715 _survivor_plab_array[i].reset();
5716 }
5717 }
5719 // Merge the per-thread plab arrays into the global survivor chunk
5720 // array which will provide the partitioning of the survivor space
5721 // for CMS initial scan and rescan.
5722 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
5723 int no_of_gc_threads) {
5724 assert(_survivor_plab_array != NULL, "Error");
5725 assert(_survivor_chunk_array != NULL, "Error");
5726 assert(_collectorState == FinalMarking ||
5727 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
5728 for (int j = 0; j < no_of_gc_threads; j++) {
5729 _cursor[j] = 0;
5730 }
5731 HeapWord* top = surv->top();
5732 size_t i;
5733 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
5734 HeapWord* min_val = top; // Higher than any PLAB address
5735 uint min_tid = 0; // position of min_val this round
5736 for (int j = 0; j < no_of_gc_threads; j++) {
5737 ChunkArray* cur_sca = &_survivor_plab_array[j];
5738 if (_cursor[j] == cur_sca->end()) {
5739 continue;
5740 }
5741 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5742 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5743 assert(surv->used_region().contains(cur_val), "Out of bounds value");
5744 if (cur_val < min_val) {
5745 min_tid = j;
5746 min_val = cur_val;
5747 } else {
5748 assert(cur_val < top, "All recorded addresses should be less");
5749 }
5750 }
5751 // At this point min_val and min_tid are respectively
5752 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5753 // and the thread (j) that witnesses that address.
5754 // We record this address in the _survivor_chunk_array[i]
5755 // and increment _cursor[min_tid] prior to the next round i.
5756 if (min_val == top) {
5757 break;
5758 }
5759 _survivor_chunk_array[i] = min_val;
5760 _cursor[min_tid]++;
5761 }
5762 // We are all done; record the size of the _survivor_chunk_array
5763 _survivor_chunk_index = i; // exclusive: [0, i)
5764 if (PrintCMSStatistics > 0) {
5765 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5766 }
5767 // Verify that we used up all the recorded entries
5768 #ifdef ASSERT
5769 size_t total = 0;
5770 for (int j = 0; j < no_of_gc_threads; j++) {
5771 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5772 total += _cursor[j];
5773 }
5774 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5775 // Check that the merged array is in sorted order
5776 if (total > 0) {
5777 for (size_t i = 0; i < total - 1; i++) {
5778 if (PrintCMSStatistics > 0) {
5779 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5780 i, _survivor_chunk_array[i]);
5781 }
5782 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5783 "Not sorted");
5784 }
5785 }
5786 #endif // ASSERT
5787 }
5789 // Set up the space's par_seq_tasks structure for work claiming
5790 // for parallel initial scan and rescan of young gen.
5791 // See ParRescanTask where this is currently used.
5792 void
5793 CMSCollector::
5794 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5795 assert(n_threads > 0, "Unexpected n_threads argument");
5796 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5798 // Eden space
5799 if (!dng->eden()->is_empty()) {
5800 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5801 assert(!pst->valid(), "Clobbering existing data?");
5802 // Each valid entry in [0, _eden_chunk_index) represents a task.
5803 size_t n_tasks = _eden_chunk_index + 1;
5804 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5805 // Sets the condition for completion of the subtask (how many threads
5806 // need to finish in order to be done).
5807 pst->set_n_threads(n_threads);
5808 pst->set_n_tasks((int)n_tasks);
5809 }
5811 // Merge the survivor plab arrays into _survivor_chunk_array
5812 if (_survivor_plab_array != NULL) {
5813 merge_survivor_plab_arrays(dng->from(), n_threads);
5814 } else {
5815 assert(_survivor_chunk_index == 0, "Error");
5816 }
5818 // To space
5819 {
5820 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5821 assert(!pst->valid(), "Clobbering existing data?");
5822 // Sets the condition for completion of the subtask (how many threads
5823 // need to finish in order to be done).
5824 pst->set_n_threads(n_threads);
5825 pst->set_n_tasks(1);
5826 assert(pst->valid(), "Error");
5827 }
5829 // From space
5830 {
5831 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5832 assert(!pst->valid(), "Clobbering existing data?");
5833 size_t n_tasks = _survivor_chunk_index + 1;
5834 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5835 // Sets the condition for completion of the subtask (how many threads
5836 // need to finish in order to be done).
5837 pst->set_n_threads(n_threads);
5838 pst->set_n_tasks((int)n_tasks);
5839 assert(pst->valid(), "Error");
5840 }
5841 }
5843 // Parallel version of remark
5844 void CMSCollector::do_remark_parallel() {
5845 GenCollectedHeap* gch = GenCollectedHeap::heap();
5846 FlexibleWorkGang* workers = gch->workers();
5847 assert(workers != NULL, "Need parallel worker threads.");
5848 // Choose to use the number of GC workers most recently set
5849 // into "active_workers". If active_workers is not set, set it
5850 // to ParallelGCThreads.
5851 int n_workers = workers->active_workers();
5852 if (n_workers == 0) {
5853 assert(n_workers > 0, "Should have been set during scavenge");
5854 n_workers = ParallelGCThreads;
5855 workers->set_active_workers(n_workers);
5856 }
5857 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5859 CMSParRemarkTask tsk(this,
5860 cms_space,
5861 n_workers, workers, task_queues());
5863 // Set up for parallel process_strong_roots work.
5864 gch->set_par_threads(n_workers);
5865 // We won't be iterating over the cards in the card table updating
5866 // the younger_gen cards, so we shouldn't call the following else
5867 // the verification code as well as subsequent younger_refs_iterate
5868 // code would get confused. XXX
5869 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5871 // The young gen rescan work will not be done as part of
5872 // process_strong_roots (which currently doesn't knw how to
5873 // parallelize such a scan), but rather will be broken up into
5874 // a set of parallel tasks (via the sampling that the [abortable]
5875 // preclean phase did of EdenSpace, plus the [two] tasks of
5876 // scanning the [two] survivor spaces. Further fine-grain
5877 // parallelization of the scanning of the survivor spaces
5878 // themselves, and of precleaning of the younger gen itself
5879 // is deferred to the future.
5880 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5882 // The dirty card rescan work is broken up into a "sequence"
5883 // of parallel tasks (per constituent space) that are dynamically
5884 // claimed by the parallel threads.
5885 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5887 // It turns out that even when we're using 1 thread, doing the work in a
5888 // separate thread causes wide variance in run times. We can't help this
5889 // in the multi-threaded case, but we special-case n=1 here to get
5890 // repeatable measurements of the 1-thread overhead of the parallel code.
5891 if (n_workers > 1) {
5892 // Make refs discovery MT-safe, if it isn't already: it may not
5893 // necessarily be so, since it's possible that we are doing
5894 // ST marking.
5895 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5896 GenCollectedHeap::StrongRootsScope srs(gch);
5897 workers->run_task(&tsk);
5898 } else {
5899 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5900 GenCollectedHeap::StrongRootsScope srs(gch);
5901 tsk.work(0);
5902 }
5904 gch->set_par_threads(0); // 0 ==> non-parallel.
5905 // restore, single-threaded for now, any preserved marks
5906 // as a result of work_q overflow
5907 restore_preserved_marks_if_any();
5908 }
5910 // Non-parallel version of remark
5911 void CMSCollector::do_remark_non_parallel() {
5912 ResourceMark rm;
5913 HandleMark hm;
5914 GenCollectedHeap* gch = GenCollectedHeap::heap();
5915 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5917 MarkRefsIntoAndScanClosure
5918 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
5919 &_markStack, this,
5920 false /* should_yield */, false /* not precleaning */);
5921 MarkFromDirtyCardsClosure
5922 markFromDirtyCardsClosure(this, _span,
5923 NULL, // space is set further below
5924 &_markBitMap, &_markStack, &mrias_cl);
5925 {
5926 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm);
5927 // Iterate over the dirty cards, setting the corresponding bits in the
5928 // mod union table.
5929 {
5930 ModUnionClosure modUnionClosure(&_modUnionTable);
5931 _ct->ct_bs()->dirty_card_iterate(
5932 _cmsGen->used_region(),
5933 &modUnionClosure);
5934 }
5935 // Having transferred these marks into the modUnionTable, we just need
5936 // to rescan the marked objects on the dirty cards in the modUnionTable.
5937 // The initial marking may have been done during an asynchronous
5938 // collection so there may be dirty bits in the mod-union table.
5939 const int alignment =
5940 CardTableModRefBS::card_size * BitsPerWord;
5941 {
5942 // ... First handle dirty cards in CMS gen
5943 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5944 MemRegion ur = _cmsGen->used_region();
5945 HeapWord* lb = ur.start();
5946 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5947 MemRegion cms_span(lb, ub);
5948 _modUnionTable.dirty_range_iterate_clear(cms_span,
5949 &markFromDirtyCardsClosure);
5950 verify_work_stacks_empty();
5951 if (PrintCMSStatistics != 0) {
5952 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5953 markFromDirtyCardsClosure.num_dirty_cards());
5954 }
5955 }
5956 }
5957 if (VerifyDuringGC &&
5958 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5959 HandleMark hm; // Discard invalid handles created during verification
5960 Universe::verify();
5961 }
5962 {
5963 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm);
5965 verify_work_stacks_empty();
5967 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5968 GenCollectedHeap::StrongRootsScope srs(gch);
5969 gch->gen_process_strong_roots(_cmsGen->level(),
5970 true, // younger gens as roots
5971 false, // use the local StrongRootsScope
5972 false, // not scavenging
5973 SharedHeap::ScanningOption(roots_scanning_options()),
5974 &mrias_cl,
5975 true, // walk code active on stacks
5976 NULL,
5977 NULL); // The dirty klasses will be handled below
5979 assert(should_unload_classes()
5980 || (roots_scanning_options() & SharedHeap::SO_CodeCache),
5981 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5982 }
5984 {
5985 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm);
5987 verify_work_stacks_empty();
5989 // Scan all class loader data objects that might have been introduced
5990 // during concurrent marking.
5991 ResourceMark rm;
5992 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5993 for (int i = 0; i < array->length(); i++) {
5994 mrias_cl.do_class_loader_data(array->at(i));
5995 }
5997 // We don't need to keep track of new CLDs anymore.
5998 ClassLoaderDataGraph::remember_new_clds(false);
6000 verify_work_stacks_empty();
6001 }
6003 {
6004 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm);
6006 verify_work_stacks_empty();
6008 RemarkKlassClosure remark_klass_closure(&mrias_cl);
6009 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
6011 verify_work_stacks_empty();
6012 }
6014 // We might have added oops to ClassLoaderData::_handles during the
6015 // concurrent marking phase. These oops point to newly allocated objects
6016 // that are guaranteed to be kept alive. Either by the direct allocation
6017 // code, or when the young collector processes the strong roots. Hence,
6018 // we don't have to revisit the _handles block during the remark phase.
6020 verify_work_stacks_empty();
6021 // Restore evacuated mark words, if any, used for overflow list links
6022 if (!CMSOverflowEarlyRestoration) {
6023 restore_preserved_marks_if_any();
6024 }
6025 verify_overflow_empty();
6026 }
6028 ////////////////////////////////////////////////////////
6029 // Parallel Reference Processing Task Proxy Class
6030 ////////////////////////////////////////////////////////
6031 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
6032 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
6033 CMSCollector* _collector;
6034 CMSBitMap* _mark_bit_map;
6035 const MemRegion _span;
6036 ProcessTask& _task;
6038 public:
6039 CMSRefProcTaskProxy(ProcessTask& task,
6040 CMSCollector* collector,
6041 const MemRegion& span,
6042 CMSBitMap* mark_bit_map,
6043 AbstractWorkGang* workers,
6044 OopTaskQueueSet* task_queues):
6045 // XXX Should superclass AGTWOQ also know about AWG since it knows
6046 // about the task_queues used by the AWG? Then it could initialize
6047 // the terminator() object. See 6984287. The set_for_termination()
6048 // below is a temporary band-aid for the regression in 6984287.
6049 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
6050 task_queues),
6051 _task(task),
6052 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
6053 {
6054 assert(_collector->_span.equals(_span) && !_span.is_empty(),
6055 "Inconsistency in _span");
6056 set_for_termination(workers->active_workers());
6057 }
6059 OopTaskQueueSet* task_queues() { return queues(); }
6061 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
6063 void do_work_steal(int i,
6064 CMSParDrainMarkingStackClosure* drain,
6065 CMSParKeepAliveClosure* keep_alive,
6066 int* seed);
6068 virtual void work(uint worker_id);
6069 };
6071 void CMSRefProcTaskProxy::work(uint worker_id) {
6072 assert(_collector->_span.equals(_span), "Inconsistency in _span");
6073 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
6074 _mark_bit_map,
6075 work_queue(worker_id));
6076 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
6077 _mark_bit_map,
6078 work_queue(worker_id));
6079 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
6080 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
6081 if (_task.marks_oops_alive()) {
6082 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
6083 _collector->hash_seed(worker_id));
6084 }
6085 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
6086 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
6087 }
6089 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
6090 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
6091 EnqueueTask& _task;
6093 public:
6094 CMSRefEnqueueTaskProxy(EnqueueTask& task)
6095 : AbstractGangTask("Enqueue reference objects in parallel"),
6096 _task(task)
6097 { }
6099 virtual void work(uint worker_id)
6100 {
6101 _task.work(worker_id);
6102 }
6103 };
6105 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
6106 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
6107 _span(span),
6108 _bit_map(bit_map),
6109 _work_queue(work_queue),
6110 _mark_and_push(collector, span, bit_map, work_queue),
6111 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6112 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
6113 { }
6115 // . see if we can share work_queues with ParNew? XXX
6116 void CMSRefProcTaskProxy::do_work_steal(int i,
6117 CMSParDrainMarkingStackClosure* drain,
6118 CMSParKeepAliveClosure* keep_alive,
6119 int* seed) {
6120 OopTaskQueue* work_q = work_queue(i);
6121 NOT_PRODUCT(int num_steals = 0;)
6122 oop obj_to_scan;
6124 while (true) {
6125 // Completely finish any left over work from (an) earlier round(s)
6126 drain->trim_queue(0);
6127 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
6128 (size_t)ParGCDesiredObjsFromOverflowList);
6129 // Now check if there's any work in the overflow list
6130 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
6131 // only affects the number of attempts made to get work from the
6132 // overflow list and does not affect the number of workers. Just
6133 // pass ParallelGCThreads so this behavior is unchanged.
6134 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
6135 work_q,
6136 ParallelGCThreads)) {
6137 // Found something in global overflow list;
6138 // not yet ready to go stealing work from others.
6139 // We'd like to assert(work_q->size() != 0, ...)
6140 // because we just took work from the overflow list,
6141 // but of course we can't, since all of that might have
6142 // been already stolen from us.
6143 continue;
6144 }
6145 // Verify that we have no work before we resort to stealing
6146 assert(work_q->size() == 0, "Have work, shouldn't steal");
6147 // Try to steal from other queues that have work
6148 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
6149 NOT_PRODUCT(num_steals++;)
6150 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
6151 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
6152 // Do scanning work
6153 obj_to_scan->oop_iterate(keep_alive);
6154 // Loop around, finish this work, and try to steal some more
6155 } else if (terminator()->offer_termination()) {
6156 break; // nirvana from the infinite cycle
6157 }
6158 }
6159 NOT_PRODUCT(
6160 if (PrintCMSStatistics != 0) {
6161 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
6162 }
6163 )
6164 }
6166 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
6167 {
6168 GenCollectedHeap* gch = GenCollectedHeap::heap();
6169 FlexibleWorkGang* workers = gch->workers();
6170 assert(workers != NULL, "Need parallel worker threads.");
6171 CMSRefProcTaskProxy rp_task(task, &_collector,
6172 _collector.ref_processor()->span(),
6173 _collector.markBitMap(),
6174 workers, _collector.task_queues());
6175 workers->run_task(&rp_task);
6176 }
6178 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
6179 {
6181 GenCollectedHeap* gch = GenCollectedHeap::heap();
6182 FlexibleWorkGang* workers = gch->workers();
6183 assert(workers != NULL, "Need parallel worker threads.");
6184 CMSRefEnqueueTaskProxy enq_task(task);
6185 workers->run_task(&enq_task);
6186 }
6188 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
6190 ResourceMark rm;
6191 HandleMark hm;
6193 ReferenceProcessor* rp = ref_processor();
6194 assert(rp->span().equals(_span), "Spans should be equal");
6195 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
6196 // Process weak references.
6197 rp->setup_policy(clear_all_soft_refs);
6198 verify_work_stacks_empty();
6200 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
6201 &_markStack, false /* !preclean */);
6202 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
6203 _span, &_markBitMap, &_markStack,
6204 &cmsKeepAliveClosure, false /* !preclean */);
6205 {
6206 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm);
6208 ReferenceProcessorStats stats;
6209 if (rp->processing_is_mt()) {
6210 // Set the degree of MT here. If the discovery is done MT, there
6211 // may have been a different number of threads doing the discovery
6212 // and a different number of discovered lists may have Ref objects.
6213 // That is OK as long as the Reference lists are balanced (see
6214 // balance_all_queues() and balance_queues()).
6215 GenCollectedHeap* gch = GenCollectedHeap::heap();
6216 int active_workers = ParallelGCThreads;
6217 FlexibleWorkGang* workers = gch->workers();
6218 if (workers != NULL) {
6219 active_workers = workers->active_workers();
6220 // The expectation is that active_workers will have already
6221 // been set to a reasonable value. If it has not been set,
6222 // investigate.
6223 assert(active_workers > 0, "Should have been set during scavenge");
6224 }
6225 rp->set_active_mt_degree(active_workers);
6226 CMSRefProcTaskExecutor task_executor(*this);
6227 stats = rp->process_discovered_references(&_is_alive_closure,
6228 &cmsKeepAliveClosure,
6229 &cmsDrainMarkingStackClosure,
6230 &task_executor,
6231 _gc_timer_cm);
6232 } else {
6233 stats = rp->process_discovered_references(&_is_alive_closure,
6234 &cmsKeepAliveClosure,
6235 &cmsDrainMarkingStackClosure,
6236 NULL,
6237 _gc_timer_cm);
6238 }
6239 _gc_tracer_cm->report_gc_reference_stats(stats);
6241 }
6243 // This is the point where the entire marking should have completed.
6244 verify_work_stacks_empty();
6246 if (should_unload_classes()) {
6247 {
6248 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm);
6250 // Unload classes and purge the SystemDictionary.
6251 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
6253 // Unload nmethods.
6254 CodeCache::do_unloading(&_is_alive_closure, purged_class);
6256 // Prune dead klasses from subklass/sibling/implementor lists.
6257 Klass::clean_weak_klass_links(&_is_alive_closure);
6258 }
6260 {
6261 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm);
6262 // Clean up unreferenced symbols in symbol table.
6263 SymbolTable::unlink();
6264 }
6265 }
6267 // CMS doesn't use the StringTable as hard roots when class unloading is turned off.
6268 // Need to check if we really scanned the StringTable.
6269 if ((roots_scanning_options() & SharedHeap::SO_Strings) == 0) {
6270 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm);
6271 // Delete entries for dead interned strings.
6272 StringTable::unlink(&_is_alive_closure);
6273 }
6275 // Restore any preserved marks as a result of mark stack or
6276 // work queue overflow
6277 restore_preserved_marks_if_any(); // done single-threaded for now
6279 rp->set_enqueuing_is_done(true);
6280 if (rp->processing_is_mt()) {
6281 rp->balance_all_queues();
6282 CMSRefProcTaskExecutor task_executor(*this);
6283 rp->enqueue_discovered_references(&task_executor);
6284 } else {
6285 rp->enqueue_discovered_references(NULL);
6286 }
6287 rp->verify_no_references_recorded();
6288 assert(!rp->discovery_enabled(), "should have been disabled");
6289 }
6291 #ifndef PRODUCT
6292 void CMSCollector::check_correct_thread_executing() {
6293 Thread* t = Thread::current();
6294 // Only the VM thread or the CMS thread should be here.
6295 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
6296 "Unexpected thread type");
6297 // If this is the vm thread, the foreground process
6298 // should not be waiting. Note that _foregroundGCIsActive is
6299 // true while the foreground collector is waiting.
6300 if (_foregroundGCShouldWait) {
6301 // We cannot be the VM thread
6302 assert(t->is_ConcurrentGC_thread(),
6303 "Should be CMS thread");
6304 } else {
6305 // We can be the CMS thread only if we are in a stop-world
6306 // phase of CMS collection.
6307 if (t->is_ConcurrentGC_thread()) {
6308 assert(_collectorState == InitialMarking ||
6309 _collectorState == FinalMarking,
6310 "Should be a stop-world phase");
6311 // The CMS thread should be holding the CMS_token.
6312 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6313 "Potential interference with concurrently "
6314 "executing VM thread");
6315 }
6316 }
6317 }
6318 #endif
6320 void CMSCollector::sweep(bool asynch) {
6321 assert(_collectorState == Sweeping, "just checking");
6322 check_correct_thread_executing();
6323 verify_work_stacks_empty();
6324 verify_overflow_empty();
6325 increment_sweep_count();
6326 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
6328 _inter_sweep_timer.stop();
6329 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
6330 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
6332 assert(!_intra_sweep_timer.is_active(), "Should not be active");
6333 _intra_sweep_timer.reset();
6334 _intra_sweep_timer.start();
6335 if (asynch) {
6336 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6337 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
6338 // First sweep the old gen
6339 {
6340 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6341 bitMapLock());
6342 sweepWork(_cmsGen, asynch);
6343 }
6345 // Update Universe::_heap_*_at_gc figures.
6346 // We need all the free list locks to make the abstract state
6347 // transition from Sweeping to Resetting. See detailed note
6348 // further below.
6349 {
6350 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
6351 // Update heap occupancy information which is used as
6352 // input to soft ref clearing policy at the next gc.
6353 Universe::update_heap_info_at_gc();
6354 _collectorState = Resizing;
6355 }
6356 } else {
6357 // already have needed locks
6358 sweepWork(_cmsGen, asynch);
6359 // Update heap occupancy information which is used as
6360 // input to soft ref clearing policy at the next gc.
6361 Universe::update_heap_info_at_gc();
6362 _collectorState = Resizing;
6363 }
6364 verify_work_stacks_empty();
6365 verify_overflow_empty();
6367 if (should_unload_classes()) {
6368 // Delay purge to the beginning of the next safepoint. Metaspace::contains
6369 // requires that the virtual spaces are stable and not deleted.
6370 ClassLoaderDataGraph::set_should_purge(true);
6371 }
6373 _intra_sweep_timer.stop();
6374 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
6376 _inter_sweep_timer.reset();
6377 _inter_sweep_timer.start();
6379 // We need to use a monotonically non-deccreasing time in ms
6380 // or we will see time-warp warnings and os::javaTimeMillis()
6381 // does not guarantee monotonicity.
6382 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
6383 update_time_of_last_gc(now);
6385 // NOTE on abstract state transitions:
6386 // Mutators allocate-live and/or mark the mod-union table dirty
6387 // based on the state of the collection. The former is done in
6388 // the interval [Marking, Sweeping] and the latter in the interval
6389 // [Marking, Sweeping). Thus the transitions into the Marking state
6390 // and out of the Sweeping state must be synchronously visible
6391 // globally to the mutators.
6392 // The transition into the Marking state happens with the world
6393 // stopped so the mutators will globally see it. Sweeping is
6394 // done asynchronously by the background collector so the transition
6395 // from the Sweeping state to the Resizing state must be done
6396 // under the freelistLock (as is the check for whether to
6397 // allocate-live and whether to dirty the mod-union table).
6398 assert(_collectorState == Resizing, "Change of collector state to"
6399 " Resizing must be done under the freelistLocks (plural)");
6401 // Now that sweeping has been completed, we clear
6402 // the incremental_collection_failed flag,
6403 // thus inviting a younger gen collection to promote into
6404 // this generation. If such a promotion may still fail,
6405 // the flag will be set again when a young collection is
6406 // attempted.
6407 GenCollectedHeap* gch = GenCollectedHeap::heap();
6408 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
6409 gch->update_full_collections_completed(_collection_count_start);
6410 }
6412 // FIX ME!!! Looks like this belongs in CFLSpace, with
6413 // CMSGen merely delegating to it.
6414 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
6415 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
6416 HeapWord* minAddr = _cmsSpace->bottom();
6417 HeapWord* largestAddr =
6418 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
6419 if (largestAddr == NULL) {
6420 // The dictionary appears to be empty. In this case
6421 // try to coalesce at the end of the heap.
6422 largestAddr = _cmsSpace->end();
6423 }
6424 size_t largestOffset = pointer_delta(largestAddr, minAddr);
6425 size_t nearLargestOffset =
6426 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
6427 if (PrintFLSStatistics != 0) {
6428 gclog_or_tty->print_cr(
6429 "CMS: Large Block: " PTR_FORMAT ";"
6430 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
6431 largestAddr,
6432 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset);
6433 }
6434 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
6435 }
6437 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
6438 return addr >= _cmsSpace->nearLargestChunk();
6439 }
6441 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
6442 return _cmsSpace->find_chunk_at_end();
6443 }
6445 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
6446 bool full) {
6447 // The next lower level has been collected. Gather any statistics
6448 // that are of interest at this point.
6449 if (!full && (current_level + 1) == level()) {
6450 // Gather statistics on the young generation collection.
6451 collector()->stats().record_gc0_end(used());
6452 }
6453 }
6455 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
6456 GenCollectedHeap* gch = GenCollectedHeap::heap();
6457 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
6458 "Wrong type of heap");
6459 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
6460 gch->gen_policy()->size_policy();
6461 assert(sp->is_gc_cms_adaptive_size_policy(),
6462 "Wrong type of size policy");
6463 return sp;
6464 }
6466 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
6467 if (PrintGCDetails && Verbose) {
6468 gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
6469 }
6470 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
6471 _debug_collection_type =
6472 (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
6473 if (PrintGCDetails && Verbose) {
6474 gclog_or_tty->print_cr("to %d ", _debug_collection_type);
6475 }
6476 }
6478 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
6479 bool asynch) {
6480 // We iterate over the space(s) underlying this generation,
6481 // checking the mark bit map to see if the bits corresponding
6482 // to specific blocks are marked or not. Blocks that are
6483 // marked are live and are not swept up. All remaining blocks
6484 // are swept up, with coalescing on-the-fly as we sweep up
6485 // contiguous free and/or garbage blocks:
6486 // We need to ensure that the sweeper synchronizes with allocators
6487 // and stop-the-world collectors. In particular, the following
6488 // locks are used:
6489 // . CMS token: if this is held, a stop the world collection cannot occur
6490 // . freelistLock: if this is held no allocation can occur from this
6491 // generation by another thread
6492 // . bitMapLock: if this is held, no other thread can access or update
6493 //
6495 // Note that we need to hold the freelistLock if we use
6496 // block iterate below; else the iterator might go awry if
6497 // a mutator (or promotion) causes block contents to change
6498 // (for instance if the allocator divvies up a block).
6499 // If we hold the free list lock, for all practical purposes
6500 // young generation GC's can't occur (they'll usually need to
6501 // promote), so we might as well prevent all young generation
6502 // GC's while we do a sweeping step. For the same reason, we might
6503 // as well take the bit map lock for the entire duration
6505 // check that we hold the requisite locks
6506 assert(have_cms_token(), "Should hold cms token");
6507 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
6508 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
6509 "Should possess CMS token to sweep");
6510 assert_lock_strong(gen->freelistLock());
6511 assert_lock_strong(bitMapLock());
6513 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
6514 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
6515 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
6516 _inter_sweep_estimate.padded_average(),
6517 _intra_sweep_estimate.padded_average());
6518 gen->setNearLargestChunk();
6520 {
6521 SweepClosure sweepClosure(this, gen, &_markBitMap,
6522 CMSYield && asynch);
6523 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
6524 // We need to free-up/coalesce garbage/blocks from a
6525 // co-terminal free run. This is done in the SweepClosure
6526 // destructor; so, do not remove this scope, else the
6527 // end-of-sweep-census below will be off by a little bit.
6528 }
6529 gen->cmsSpace()->sweep_completed();
6530 gen->cmsSpace()->endSweepFLCensus(sweep_count());
6531 if (should_unload_classes()) { // unloaded classes this cycle,
6532 _concurrent_cycles_since_last_unload = 0; // ... reset count
6533 } else { // did not unload classes,
6534 _concurrent_cycles_since_last_unload++; // ... increment count
6535 }
6536 }
6538 // Reset CMS data structures (for now just the marking bit map)
6539 // preparatory for the next cycle.
6540 void CMSCollector::reset(bool asynch) {
6541 GenCollectedHeap* gch = GenCollectedHeap::heap();
6542 CMSAdaptiveSizePolicy* sp = size_policy();
6543 AdaptiveSizePolicyOutput(sp, gch->total_collections());
6544 if (asynch) {
6545 CMSTokenSyncWithLocks ts(true, bitMapLock());
6547 // If the state is not "Resetting", the foreground thread
6548 // has done a collection and the resetting.
6549 if (_collectorState != Resetting) {
6550 assert(_collectorState == Idling, "The state should only change"
6551 " because the foreground collector has finished the collection");
6552 return;
6553 }
6555 // Clear the mark bitmap (no grey objects to start with)
6556 // for the next cycle.
6557 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6558 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
6560 HeapWord* curAddr = _markBitMap.startWord();
6561 while (curAddr < _markBitMap.endWord()) {
6562 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
6563 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
6564 _markBitMap.clear_large_range(chunk);
6565 if (ConcurrentMarkSweepThread::should_yield() &&
6566 !foregroundGCIsActive() &&
6567 CMSYield) {
6568 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6569 "CMS thread should hold CMS token");
6570 assert_lock_strong(bitMapLock());
6571 bitMapLock()->unlock();
6572 ConcurrentMarkSweepThread::desynchronize(true);
6573 ConcurrentMarkSweepThread::acknowledge_yield_request();
6574 stopTimer();
6575 if (PrintCMSStatistics != 0) {
6576 incrementYields();
6577 }
6578 icms_wait();
6580 // See the comment in coordinator_yield()
6581 for (unsigned i = 0; i < CMSYieldSleepCount &&
6582 ConcurrentMarkSweepThread::should_yield() &&
6583 !CMSCollector::foregroundGCIsActive(); ++i) {
6584 os::sleep(Thread::current(), 1, false);
6585 ConcurrentMarkSweepThread::acknowledge_yield_request();
6586 }
6588 ConcurrentMarkSweepThread::synchronize(true);
6589 bitMapLock()->lock_without_safepoint_check();
6590 startTimer();
6591 }
6592 curAddr = chunk.end();
6593 }
6594 // A successful mostly concurrent collection has been done.
6595 // Because only the full (i.e., concurrent mode failure) collections
6596 // are being measured for gc overhead limits, clean the "near" flag
6597 // and count.
6598 sp->reset_gc_overhead_limit_count();
6599 _collectorState = Idling;
6600 } else {
6601 // already have the lock
6602 assert(_collectorState == Resetting, "just checking");
6603 assert_lock_strong(bitMapLock());
6604 _markBitMap.clear_all();
6605 _collectorState = Idling;
6606 }
6608 // Stop incremental mode after a cycle completes, so that any future cycles
6609 // are triggered by allocation.
6610 stop_icms();
6612 NOT_PRODUCT(
6613 if (RotateCMSCollectionTypes) {
6614 _cmsGen->rotate_debug_collection_type();
6615 }
6616 )
6618 register_gc_end();
6619 }
6621 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
6622 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
6623 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6624 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL);
6625 TraceCollectorStats tcs(counters());
6627 switch (op) {
6628 case CMS_op_checkpointRootsInitial: {
6629 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6630 checkpointRootsInitial(true); // asynch
6631 if (PrintGC) {
6632 _cmsGen->printOccupancy("initial-mark");
6633 }
6634 break;
6635 }
6636 case CMS_op_checkpointRootsFinal: {
6637 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6638 checkpointRootsFinal(true, // asynch
6639 false, // !clear_all_soft_refs
6640 false); // !init_mark_was_synchronous
6641 if (PrintGC) {
6642 _cmsGen->printOccupancy("remark");
6643 }
6644 break;
6645 }
6646 default:
6647 fatal("No such CMS_op");
6648 }
6649 }
6651 #ifndef PRODUCT
6652 size_t const CMSCollector::skip_header_HeapWords() {
6653 return FreeChunk::header_size();
6654 }
6656 // Try and collect here conditions that should hold when
6657 // CMS thread is exiting. The idea is that the foreground GC
6658 // thread should not be blocked if it wants to terminate
6659 // the CMS thread and yet continue to run the VM for a while
6660 // after that.
6661 void CMSCollector::verify_ok_to_terminate() const {
6662 assert(Thread::current()->is_ConcurrentGC_thread(),
6663 "should be called by CMS thread");
6664 assert(!_foregroundGCShouldWait, "should be false");
6665 // We could check here that all the various low-level locks
6666 // are not held by the CMS thread, but that is overkill; see
6667 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6668 // is checked.
6669 }
6670 #endif
6672 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6673 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6674 "missing Printezis mark?");
6675 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6676 size_t size = pointer_delta(nextOneAddr + 1, addr);
6677 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6678 "alignment problem");
6679 assert(size >= 3, "Necessary for Printezis marks to work");
6680 return size;
6681 }
6683 // A variant of the above (block_size_using_printezis_bits()) except
6684 // that we return 0 if the P-bits are not yet set.
6685 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6686 if (_markBitMap.isMarked(addr + 1)) {
6687 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
6688 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6689 size_t size = pointer_delta(nextOneAddr + 1, addr);
6690 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6691 "alignment problem");
6692 assert(size >= 3, "Necessary for Printezis marks to work");
6693 return size;
6694 }
6695 return 0;
6696 }
6698 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6699 size_t sz = 0;
6700 oop p = (oop)addr;
6701 if (p->klass_or_null() != NULL) {
6702 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6703 } else {
6704 sz = block_size_using_printezis_bits(addr);
6705 }
6706 assert(sz > 0, "size must be nonzero");
6707 HeapWord* next_block = addr + sz;
6708 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
6709 CardTableModRefBS::card_size);
6710 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
6711 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6712 "must be different cards");
6713 return next_card;
6714 }
6717 // CMS Bit Map Wrapper /////////////////////////////////////////
6719 // Construct a CMS bit map infrastructure, but don't create the
6720 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6721 // further below.
6722 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6723 _bm(),
6724 _shifter(shifter),
6725 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6726 {
6727 _bmStartWord = 0;
6728 _bmWordSize = 0;
6729 }
6731 bool CMSBitMap::allocate(MemRegion mr) {
6732 _bmStartWord = mr.start();
6733 _bmWordSize = mr.word_size();
6734 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6735 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6736 if (!brs.is_reserved()) {
6737 warning("CMS bit map allocation failure");
6738 return false;
6739 }
6740 // For now we'll just commit all of the bit map up fromt.
6741 // Later on we'll try to be more parsimonious with swap.
6742 if (!_virtual_space.initialize(brs, brs.size())) {
6743 warning("CMS bit map backing store failure");
6744 return false;
6745 }
6746 assert(_virtual_space.committed_size() == brs.size(),
6747 "didn't reserve backing store for all of CMS bit map?");
6748 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
6749 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6750 _bmWordSize, "inconsistency in bit map sizing");
6751 _bm.set_size(_bmWordSize >> _shifter);
6753 // bm.clear(); // can we rely on getting zero'd memory? verify below
6754 assert(isAllClear(),
6755 "Expected zero'd memory from ReservedSpace constructor");
6756 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6757 "consistency check");
6758 return true;
6759 }
6761 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6762 HeapWord *next_addr, *end_addr, *last_addr;
6763 assert_locked();
6764 assert(covers(mr), "out-of-range error");
6765 // XXX assert that start and end are appropriately aligned
6766 for (next_addr = mr.start(), end_addr = mr.end();
6767 next_addr < end_addr; next_addr = last_addr) {
6768 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6769 last_addr = dirty_region.end();
6770 if (!dirty_region.is_empty()) {
6771 cl->do_MemRegion(dirty_region);
6772 } else {
6773 assert(last_addr == end_addr, "program logic");
6774 return;
6775 }
6776 }
6777 }
6779 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
6780 _bm.print_on_error(st, prefix);
6781 }
6783 #ifndef PRODUCT
6784 void CMSBitMap::assert_locked() const {
6785 CMSLockVerifier::assert_locked(lock());
6786 }
6788 bool CMSBitMap::covers(MemRegion mr) const {
6789 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6790 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6791 "size inconsistency");
6792 return (mr.start() >= _bmStartWord) &&
6793 (mr.end() <= endWord());
6794 }
6796 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6797 return (start >= _bmStartWord && (start + size) <= endWord());
6798 }
6800 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6801 // verify that there are no 1 bits in the interval [left, right)
6802 FalseBitMapClosure falseBitMapClosure;
6803 iterate(&falseBitMapClosure, left, right);
6804 }
6806 void CMSBitMap::region_invariant(MemRegion mr)
6807 {
6808 assert_locked();
6809 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6810 assert(!mr.is_empty(), "unexpected empty region");
6811 assert(covers(mr), "mr should be covered by bit map");
6812 // convert address range into offset range
6813 size_t start_ofs = heapWordToOffset(mr.start());
6814 // Make sure that end() is appropriately aligned
6815 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6816 (1 << (_shifter+LogHeapWordSize))),
6817 "Misaligned mr.end()");
6818 size_t end_ofs = heapWordToOffset(mr.end());
6819 assert(end_ofs > start_ofs, "Should mark at least one bit");
6820 }
6822 #endif
6824 bool CMSMarkStack::allocate(size_t size) {
6825 // allocate a stack of the requisite depth
6826 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6827 size * sizeof(oop)));
6828 if (!rs.is_reserved()) {
6829 warning("CMSMarkStack allocation failure");
6830 return false;
6831 }
6832 if (!_virtual_space.initialize(rs, rs.size())) {
6833 warning("CMSMarkStack backing store failure");
6834 return false;
6835 }
6836 assert(_virtual_space.committed_size() == rs.size(),
6837 "didn't reserve backing store for all of CMS stack?");
6838 _base = (oop*)(_virtual_space.low());
6839 _index = 0;
6840 _capacity = size;
6841 NOT_PRODUCT(_max_depth = 0);
6842 return true;
6843 }
6845 // XXX FIX ME !!! In the MT case we come in here holding a
6846 // leaf lock. For printing we need to take a further lock
6847 // which has lower rank. We need to recallibrate the two
6848 // lock-ranks involved in order to be able to rpint the
6849 // messages below. (Or defer the printing to the caller.
6850 // For now we take the expedient path of just disabling the
6851 // messages for the problematic case.)
6852 void CMSMarkStack::expand() {
6853 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6854 if (_capacity == MarkStackSizeMax) {
6855 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6856 // We print a warning message only once per CMS cycle.
6857 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6858 }
6859 return;
6860 }
6861 // Double capacity if possible
6862 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6863 // Do not give up existing stack until we have managed to
6864 // get the double capacity that we desired.
6865 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6866 new_capacity * sizeof(oop)));
6867 if (rs.is_reserved()) {
6868 // Release the backing store associated with old stack
6869 _virtual_space.release();
6870 // Reinitialize virtual space for new stack
6871 if (!_virtual_space.initialize(rs, rs.size())) {
6872 fatal("Not enough swap for expanded marking stack");
6873 }
6874 _base = (oop*)(_virtual_space.low());
6875 _index = 0;
6876 _capacity = new_capacity;
6877 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6878 // Failed to double capacity, continue;
6879 // we print a detail message only once per CMS cycle.
6880 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6881 SIZE_FORMAT"K",
6882 _capacity / K, new_capacity / K);
6883 }
6884 }
6887 // Closures
6888 // XXX: there seems to be a lot of code duplication here;
6889 // should refactor and consolidate common code.
6891 // This closure is used to mark refs into the CMS generation in
6892 // the CMS bit map. Called at the first checkpoint. This closure
6893 // assumes that we do not need to re-mark dirty cards; if the CMS
6894 // generation on which this is used is not an oldest
6895 // generation then this will lose younger_gen cards!
6897 MarkRefsIntoClosure::MarkRefsIntoClosure(
6898 MemRegion span, CMSBitMap* bitMap):
6899 _span(span),
6900 _bitMap(bitMap)
6901 {
6902 assert(_ref_processor == NULL, "deliberately left NULL");
6903 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6904 }
6906 void MarkRefsIntoClosure::do_oop(oop obj) {
6907 // if p points into _span, then mark corresponding bit in _markBitMap
6908 assert(obj->is_oop(), "expected an oop");
6909 HeapWord* addr = (HeapWord*)obj;
6910 if (_span.contains(addr)) {
6911 // this should be made more efficient
6912 _bitMap->mark(addr);
6913 }
6914 }
6916 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6917 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6919 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
6920 MemRegion span, CMSBitMap* bitMap):
6921 _span(span),
6922 _bitMap(bitMap)
6923 {
6924 assert(_ref_processor == NULL, "deliberately left NULL");
6925 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6926 }
6928 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
6929 // if p points into _span, then mark corresponding bit in _markBitMap
6930 assert(obj->is_oop(), "expected an oop");
6931 HeapWord* addr = (HeapWord*)obj;
6932 if (_span.contains(addr)) {
6933 // this should be made more efficient
6934 _bitMap->par_mark(addr);
6935 }
6936 }
6938 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6939 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6941 // A variant of the above, used for CMS marking verification.
6942 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6943 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6944 _span(span),
6945 _verification_bm(verification_bm),
6946 _cms_bm(cms_bm)
6947 {
6948 assert(_ref_processor == NULL, "deliberately left NULL");
6949 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6950 }
6952 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6953 // if p points into _span, then mark corresponding bit in _markBitMap
6954 assert(obj->is_oop(), "expected an oop");
6955 HeapWord* addr = (HeapWord*)obj;
6956 if (_span.contains(addr)) {
6957 _verification_bm->mark(addr);
6958 if (!_cms_bm->isMarked(addr)) {
6959 oop(addr)->print();
6960 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6961 fatal("... aborting");
6962 }
6963 }
6964 }
6966 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6967 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6969 //////////////////////////////////////////////////
6970 // MarkRefsIntoAndScanClosure
6971 //////////////////////////////////////////////////
6973 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6974 ReferenceProcessor* rp,
6975 CMSBitMap* bit_map,
6976 CMSBitMap* mod_union_table,
6977 CMSMarkStack* mark_stack,
6978 CMSCollector* collector,
6979 bool should_yield,
6980 bool concurrent_precleaning):
6981 _collector(collector),
6982 _span(span),
6983 _bit_map(bit_map),
6984 _mark_stack(mark_stack),
6985 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6986 mark_stack, concurrent_precleaning),
6987 _yield(should_yield),
6988 _concurrent_precleaning(concurrent_precleaning),
6989 _freelistLock(NULL)
6990 {
6991 _ref_processor = rp;
6992 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6993 }
6995 // This closure is used to mark refs into the CMS generation at the
6996 // second (final) checkpoint, and to scan and transitively follow
6997 // the unmarked oops. It is also used during the concurrent precleaning
6998 // phase while scanning objects on dirty cards in the CMS generation.
6999 // The marks are made in the marking bit map and the marking stack is
7000 // used for keeping the (newly) grey objects during the scan.
7001 // The parallel version (Par_...) appears further below.
7002 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
7003 if (obj != NULL) {
7004 assert(obj->is_oop(), "expected an oop");
7005 HeapWord* addr = (HeapWord*)obj;
7006 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
7007 assert(_collector->overflow_list_is_empty(),
7008 "overflow list should be empty");
7009 if (_span.contains(addr) &&
7010 !_bit_map->isMarked(addr)) {
7011 // mark bit map (object is now grey)
7012 _bit_map->mark(addr);
7013 // push on marking stack (stack should be empty), and drain the
7014 // stack by applying this closure to the oops in the oops popped
7015 // from the stack (i.e. blacken the grey objects)
7016 bool res = _mark_stack->push(obj);
7017 assert(res, "Should have space to push on empty stack");
7018 do {
7019 oop new_oop = _mark_stack->pop();
7020 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
7021 assert(_bit_map->isMarked((HeapWord*)new_oop),
7022 "only grey objects on this stack");
7023 // iterate over the oops in this oop, marking and pushing
7024 // the ones in CMS heap (i.e. in _span).
7025 new_oop->oop_iterate(&_pushAndMarkClosure);
7026 // check if it's time to yield
7027 do_yield_check();
7028 } while (!_mark_stack->isEmpty() ||
7029 (!_concurrent_precleaning && take_from_overflow_list()));
7030 // if marking stack is empty, and we are not doing this
7031 // during precleaning, then check the overflow list
7032 }
7033 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
7034 assert(_collector->overflow_list_is_empty(),
7035 "overflow list was drained above");
7036 // We could restore evacuated mark words, if any, used for
7037 // overflow list links here because the overflow list is
7038 // provably empty here. That would reduce the maximum
7039 // size requirements for preserved_{oop,mark}_stack.
7040 // But we'll just postpone it until we are all done
7041 // so we can just stream through.
7042 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
7043 _collector->restore_preserved_marks_if_any();
7044 assert(_collector->no_preserved_marks(), "No preserved marks");
7045 }
7046 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
7047 "All preserved marks should have been restored above");
7048 }
7049 }
7051 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
7052 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
7054 void MarkRefsIntoAndScanClosure::do_yield_work() {
7055 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7056 "CMS thread should hold CMS token");
7057 assert_lock_strong(_freelistLock);
7058 assert_lock_strong(_bit_map->lock());
7059 // relinquish the free_list_lock and bitMaplock()
7060 _bit_map->lock()->unlock();
7061 _freelistLock->unlock();
7062 ConcurrentMarkSweepThread::desynchronize(true);
7063 ConcurrentMarkSweepThread::acknowledge_yield_request();
7064 _collector->stopTimer();
7065 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7066 if (PrintCMSStatistics != 0) {
7067 _collector->incrementYields();
7068 }
7069 _collector->icms_wait();
7071 // See the comment in coordinator_yield()
7072 for (unsigned i = 0;
7073 i < CMSYieldSleepCount &&
7074 ConcurrentMarkSweepThread::should_yield() &&
7075 !CMSCollector::foregroundGCIsActive();
7076 ++i) {
7077 os::sleep(Thread::current(), 1, false);
7078 ConcurrentMarkSweepThread::acknowledge_yield_request();
7079 }
7081 ConcurrentMarkSweepThread::synchronize(true);
7082 _freelistLock->lock_without_safepoint_check();
7083 _bit_map->lock()->lock_without_safepoint_check();
7084 _collector->startTimer();
7085 }
7087 ///////////////////////////////////////////////////////////
7088 // Par_MarkRefsIntoAndScanClosure: a parallel version of
7089 // MarkRefsIntoAndScanClosure
7090 ///////////////////////////////////////////////////////////
7091 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
7092 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
7093 CMSBitMap* bit_map, OopTaskQueue* work_queue):
7094 _span(span),
7095 _bit_map(bit_map),
7096 _work_queue(work_queue),
7097 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
7098 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
7099 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
7100 {
7101 _ref_processor = rp;
7102 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7103 }
7105 // This closure is used to mark refs into the CMS generation at the
7106 // second (final) checkpoint, and to scan and transitively follow
7107 // the unmarked oops. The marks are made in the marking bit map and
7108 // the work_queue is used for keeping the (newly) grey objects during
7109 // the scan phase whence they are also available for stealing by parallel
7110 // threads. Since the marking bit map is shared, updates are
7111 // synchronized (via CAS).
7112 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
7113 if (obj != NULL) {
7114 // Ignore mark word because this could be an already marked oop
7115 // that may be chained at the end of the overflow list.
7116 assert(obj->is_oop(true), "expected an oop");
7117 HeapWord* addr = (HeapWord*)obj;
7118 if (_span.contains(addr) &&
7119 !_bit_map->isMarked(addr)) {
7120 // mark bit map (object will become grey):
7121 // It is possible for several threads to be
7122 // trying to "claim" this object concurrently;
7123 // the unique thread that succeeds in marking the
7124 // object first will do the subsequent push on
7125 // to the work queue (or overflow list).
7126 if (_bit_map->par_mark(addr)) {
7127 // push on work_queue (which may not be empty), and trim the
7128 // queue to an appropriate length by applying this closure to
7129 // the oops in the oops popped from the stack (i.e. blacken the
7130 // grey objects)
7131 bool res = _work_queue->push(obj);
7132 assert(res, "Low water mark should be less than capacity?");
7133 trim_queue(_low_water_mark);
7134 } // Else, another thread claimed the object
7135 }
7136 }
7137 }
7139 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
7140 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
7142 // This closure is used to rescan the marked objects on the dirty cards
7143 // in the mod union table and the card table proper.
7144 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
7145 oop p, MemRegion mr) {
7147 size_t size = 0;
7148 HeapWord* addr = (HeapWord*)p;
7149 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7150 assert(_span.contains(addr), "we are scanning the CMS generation");
7151 // check if it's time to yield
7152 if (do_yield_check()) {
7153 // We yielded for some foreground stop-world work,
7154 // and we have been asked to abort this ongoing preclean cycle.
7155 return 0;
7156 }
7157 if (_bitMap->isMarked(addr)) {
7158 // it's marked; is it potentially uninitialized?
7159 if (p->klass_or_null() != NULL) {
7160 // an initialized object; ignore mark word in verification below
7161 // since we are running concurrent with mutators
7162 assert(p->is_oop(true), "should be an oop");
7163 if (p->is_objArray()) {
7164 // objArrays are precisely marked; restrict scanning
7165 // to dirty cards only.
7166 size = CompactibleFreeListSpace::adjustObjectSize(
7167 p->oop_iterate(_scanningClosure, mr));
7168 } else {
7169 // A non-array may have been imprecisely marked; we need
7170 // to scan object in its entirety.
7171 size = CompactibleFreeListSpace::adjustObjectSize(
7172 p->oop_iterate(_scanningClosure));
7173 }
7174 #ifdef ASSERT
7175 size_t direct_size =
7176 CompactibleFreeListSpace::adjustObjectSize(p->size());
7177 assert(size == direct_size, "Inconsistency in size");
7178 assert(size >= 3, "Necessary for Printezis marks to work");
7179 if (!_bitMap->isMarked(addr+1)) {
7180 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
7181 } else {
7182 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
7183 assert(_bitMap->isMarked(addr+size-1),
7184 "inconsistent Printezis mark");
7185 }
7186 #endif // ASSERT
7187 } else {
7188 // an unitialized object
7189 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
7190 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7191 size = pointer_delta(nextOneAddr + 1, addr);
7192 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7193 "alignment problem");
7194 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
7195 // will dirty the card when the klass pointer is installed in the
7196 // object (signalling the completion of initialization).
7197 }
7198 } else {
7199 // Either a not yet marked object or an uninitialized object
7200 if (p->klass_or_null() == NULL) {
7201 // An uninitialized object, skip to the next card, since
7202 // we may not be able to read its P-bits yet.
7203 assert(size == 0, "Initial value");
7204 } else {
7205 // An object not (yet) reached by marking: we merely need to
7206 // compute its size so as to go look at the next block.
7207 assert(p->is_oop(true), "should be an oop");
7208 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
7209 }
7210 }
7211 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7212 return size;
7213 }
7215 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
7216 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7217 "CMS thread should hold CMS token");
7218 assert_lock_strong(_freelistLock);
7219 assert_lock_strong(_bitMap->lock());
7220 // relinquish the free_list_lock and bitMaplock()
7221 _bitMap->lock()->unlock();
7222 _freelistLock->unlock();
7223 ConcurrentMarkSweepThread::desynchronize(true);
7224 ConcurrentMarkSweepThread::acknowledge_yield_request();
7225 _collector->stopTimer();
7226 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7227 if (PrintCMSStatistics != 0) {
7228 _collector->incrementYields();
7229 }
7230 _collector->icms_wait();
7232 // See the comment in coordinator_yield()
7233 for (unsigned i = 0; i < CMSYieldSleepCount &&
7234 ConcurrentMarkSweepThread::should_yield() &&
7235 !CMSCollector::foregroundGCIsActive(); ++i) {
7236 os::sleep(Thread::current(), 1, false);
7237 ConcurrentMarkSweepThread::acknowledge_yield_request();
7238 }
7240 ConcurrentMarkSweepThread::synchronize(true);
7241 _freelistLock->lock_without_safepoint_check();
7242 _bitMap->lock()->lock_without_safepoint_check();
7243 _collector->startTimer();
7244 }
7247 //////////////////////////////////////////////////////////////////
7248 // SurvivorSpacePrecleanClosure
7249 //////////////////////////////////////////////////////////////////
7250 // This (single-threaded) closure is used to preclean the oops in
7251 // the survivor spaces.
7252 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
7254 HeapWord* addr = (HeapWord*)p;
7255 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7256 assert(!_span.contains(addr), "we are scanning the survivor spaces");
7257 assert(p->klass_or_null() != NULL, "object should be initializd");
7258 // an initialized object; ignore mark word in verification below
7259 // since we are running concurrent with mutators
7260 assert(p->is_oop(true), "should be an oop");
7261 // Note that we do not yield while we iterate over
7262 // the interior oops of p, pushing the relevant ones
7263 // on our marking stack.
7264 size_t size = p->oop_iterate(_scanning_closure);
7265 do_yield_check();
7266 // Observe that below, we do not abandon the preclean
7267 // phase as soon as we should; rather we empty the
7268 // marking stack before returning. This is to satisfy
7269 // some existing assertions. In general, it may be a
7270 // good idea to abort immediately and complete the marking
7271 // from the grey objects at a later time.
7272 while (!_mark_stack->isEmpty()) {
7273 oop new_oop = _mark_stack->pop();
7274 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
7275 assert(_bit_map->isMarked((HeapWord*)new_oop),
7276 "only grey objects on this stack");
7277 // iterate over the oops in this oop, marking and pushing
7278 // the ones in CMS heap (i.e. in _span).
7279 new_oop->oop_iterate(_scanning_closure);
7280 // check if it's time to yield
7281 do_yield_check();
7282 }
7283 unsigned int after_count =
7284 GenCollectedHeap::heap()->total_collections();
7285 bool abort = (_before_count != after_count) ||
7286 _collector->should_abort_preclean();
7287 return abort ? 0 : size;
7288 }
7290 void SurvivorSpacePrecleanClosure::do_yield_work() {
7291 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7292 "CMS thread should hold CMS token");
7293 assert_lock_strong(_bit_map->lock());
7294 // Relinquish the bit map lock
7295 _bit_map->lock()->unlock();
7296 ConcurrentMarkSweepThread::desynchronize(true);
7297 ConcurrentMarkSweepThread::acknowledge_yield_request();
7298 _collector->stopTimer();
7299 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7300 if (PrintCMSStatistics != 0) {
7301 _collector->incrementYields();
7302 }
7303 _collector->icms_wait();
7305 // See the comment in coordinator_yield()
7306 for (unsigned i = 0; i < CMSYieldSleepCount &&
7307 ConcurrentMarkSweepThread::should_yield() &&
7308 !CMSCollector::foregroundGCIsActive(); ++i) {
7309 os::sleep(Thread::current(), 1, false);
7310 ConcurrentMarkSweepThread::acknowledge_yield_request();
7311 }
7313 ConcurrentMarkSweepThread::synchronize(true);
7314 _bit_map->lock()->lock_without_safepoint_check();
7315 _collector->startTimer();
7316 }
7318 // This closure is used to rescan the marked objects on the dirty cards
7319 // in the mod union table and the card table proper. In the parallel
7320 // case, although the bitMap is shared, we do a single read so the
7321 // isMarked() query is "safe".
7322 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
7323 // Ignore mark word because we are running concurrent with mutators
7324 assert(p->is_oop_or_null(true), "expected an oop or null");
7325 HeapWord* addr = (HeapWord*)p;
7326 assert(_span.contains(addr), "we are scanning the CMS generation");
7327 bool is_obj_array = false;
7328 #ifdef ASSERT
7329 if (!_parallel) {
7330 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
7331 assert(_collector->overflow_list_is_empty(),
7332 "overflow list should be empty");
7334 }
7335 #endif // ASSERT
7336 if (_bit_map->isMarked(addr)) {
7337 // Obj arrays are precisely marked, non-arrays are not;
7338 // so we scan objArrays precisely and non-arrays in their
7339 // entirety.
7340 if (p->is_objArray()) {
7341 is_obj_array = true;
7342 if (_parallel) {
7343 p->oop_iterate(_par_scan_closure, mr);
7344 } else {
7345 p->oop_iterate(_scan_closure, mr);
7346 }
7347 } else {
7348 if (_parallel) {
7349 p->oop_iterate(_par_scan_closure);
7350 } else {
7351 p->oop_iterate(_scan_closure);
7352 }
7353 }
7354 }
7355 #ifdef ASSERT
7356 if (!_parallel) {
7357 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
7358 assert(_collector->overflow_list_is_empty(),
7359 "overflow list should be empty");
7361 }
7362 #endif // ASSERT
7363 return is_obj_array;
7364 }
7366 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
7367 MemRegion span,
7368 CMSBitMap* bitMap, CMSMarkStack* markStack,
7369 bool should_yield, bool verifying):
7370 _collector(collector),
7371 _span(span),
7372 _bitMap(bitMap),
7373 _mut(&collector->_modUnionTable),
7374 _markStack(markStack),
7375 _yield(should_yield),
7376 _skipBits(0)
7377 {
7378 assert(_markStack->isEmpty(), "stack should be empty");
7379 _finger = _bitMap->startWord();
7380 _threshold = _finger;
7381 assert(_collector->_restart_addr == NULL, "Sanity check");
7382 assert(_span.contains(_finger), "Out of bounds _finger?");
7383 DEBUG_ONLY(_verifying = verifying;)
7384 }
7386 void MarkFromRootsClosure::reset(HeapWord* addr) {
7387 assert(_markStack->isEmpty(), "would cause duplicates on stack");
7388 assert(_span.contains(addr), "Out of bounds _finger?");
7389 _finger = addr;
7390 _threshold = (HeapWord*)round_to(
7391 (intptr_t)_finger, CardTableModRefBS::card_size);
7392 }
7394 // Should revisit to see if this should be restructured for
7395 // greater efficiency.
7396 bool MarkFromRootsClosure::do_bit(size_t offset) {
7397 if (_skipBits > 0) {
7398 _skipBits--;
7399 return true;
7400 }
7401 // convert offset into a HeapWord*
7402 HeapWord* addr = _bitMap->startWord() + offset;
7403 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
7404 "address out of range");
7405 assert(_bitMap->isMarked(addr), "tautology");
7406 if (_bitMap->isMarked(addr+1)) {
7407 // this is an allocated but not yet initialized object
7408 assert(_skipBits == 0, "tautology");
7409 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
7410 oop p = oop(addr);
7411 if (p->klass_or_null() == NULL) {
7412 DEBUG_ONLY(if (!_verifying) {)
7413 // We re-dirty the cards on which this object lies and increase
7414 // the _threshold so that we'll come back to scan this object
7415 // during the preclean or remark phase. (CMSCleanOnEnter)
7416 if (CMSCleanOnEnter) {
7417 size_t sz = _collector->block_size_using_printezis_bits(addr);
7418 HeapWord* end_card_addr = (HeapWord*)round_to(
7419 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7420 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7421 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7422 // Bump _threshold to end_card_addr; note that
7423 // _threshold cannot possibly exceed end_card_addr, anyhow.
7424 // This prevents future clearing of the card as the scan proceeds
7425 // to the right.
7426 assert(_threshold <= end_card_addr,
7427 "Because we are just scanning into this object");
7428 if (_threshold < end_card_addr) {
7429 _threshold = end_card_addr;
7430 }
7431 if (p->klass_or_null() != NULL) {
7432 // Redirty the range of cards...
7433 _mut->mark_range(redirty_range);
7434 } // ...else the setting of klass will dirty the card anyway.
7435 }
7436 DEBUG_ONLY(})
7437 return true;
7438 }
7439 }
7440 scanOopsInOop(addr);
7441 return true;
7442 }
7444 // We take a break if we've been at this for a while,
7445 // so as to avoid monopolizing the locks involved.
7446 void MarkFromRootsClosure::do_yield_work() {
7447 // First give up the locks, then yield, then re-lock
7448 // We should probably use a constructor/destructor idiom to
7449 // do this unlock/lock or modify the MutexUnlocker class to
7450 // serve our purpose. XXX
7451 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7452 "CMS thread should hold CMS token");
7453 assert_lock_strong(_bitMap->lock());
7454 _bitMap->lock()->unlock();
7455 ConcurrentMarkSweepThread::desynchronize(true);
7456 ConcurrentMarkSweepThread::acknowledge_yield_request();
7457 _collector->stopTimer();
7458 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7459 if (PrintCMSStatistics != 0) {
7460 _collector->incrementYields();
7461 }
7462 _collector->icms_wait();
7464 // See the comment in coordinator_yield()
7465 for (unsigned i = 0; i < CMSYieldSleepCount &&
7466 ConcurrentMarkSweepThread::should_yield() &&
7467 !CMSCollector::foregroundGCIsActive(); ++i) {
7468 os::sleep(Thread::current(), 1, false);
7469 ConcurrentMarkSweepThread::acknowledge_yield_request();
7470 }
7472 ConcurrentMarkSweepThread::synchronize(true);
7473 _bitMap->lock()->lock_without_safepoint_check();
7474 _collector->startTimer();
7475 }
7477 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
7478 assert(_bitMap->isMarked(ptr), "expected bit to be set");
7479 assert(_markStack->isEmpty(),
7480 "should drain stack to limit stack usage");
7481 // convert ptr to an oop preparatory to scanning
7482 oop obj = oop(ptr);
7483 // Ignore mark word in verification below, since we
7484 // may be running concurrent with mutators.
7485 assert(obj->is_oop(true), "should be an oop");
7486 assert(_finger <= ptr, "_finger runneth ahead");
7487 // advance the finger to right end of this object
7488 _finger = ptr + obj->size();
7489 assert(_finger > ptr, "we just incremented it above");
7490 // On large heaps, it may take us some time to get through
7491 // the marking phase (especially if running iCMS). During
7492 // this time it's possible that a lot of mutations have
7493 // accumulated in the card table and the mod union table --
7494 // these mutation records are redundant until we have
7495 // actually traced into the corresponding card.
7496 // Here, we check whether advancing the finger would make
7497 // us cross into a new card, and if so clear corresponding
7498 // cards in the MUT (preclean them in the card-table in the
7499 // future).
7501 DEBUG_ONLY(if (!_verifying) {)
7502 // The clean-on-enter optimization is disabled by default,
7503 // until we fix 6178663.
7504 if (CMSCleanOnEnter && (_finger > _threshold)) {
7505 // [_threshold, _finger) represents the interval
7506 // of cards to be cleared in MUT (or precleaned in card table).
7507 // The set of cards to be cleared is all those that overlap
7508 // with the interval [_threshold, _finger); note that
7509 // _threshold is always kept card-aligned but _finger isn't
7510 // always card-aligned.
7511 HeapWord* old_threshold = _threshold;
7512 assert(old_threshold == (HeapWord*)round_to(
7513 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7514 "_threshold should always be card-aligned");
7515 _threshold = (HeapWord*)round_to(
7516 (intptr_t)_finger, CardTableModRefBS::card_size);
7517 MemRegion mr(old_threshold, _threshold);
7518 assert(!mr.is_empty(), "Control point invariant");
7519 assert(_span.contains(mr), "Should clear within span");
7520 _mut->clear_range(mr);
7521 }
7522 DEBUG_ONLY(})
7523 // Note: the finger doesn't advance while we drain
7524 // the stack below.
7525 PushOrMarkClosure pushOrMarkClosure(_collector,
7526 _span, _bitMap, _markStack,
7527 _finger, this);
7528 bool res = _markStack->push(obj);
7529 assert(res, "Empty non-zero size stack should have space for single push");
7530 while (!_markStack->isEmpty()) {
7531 oop new_oop = _markStack->pop();
7532 // Skip verifying header mark word below because we are
7533 // running concurrent with mutators.
7534 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7535 // now scan this oop's oops
7536 new_oop->oop_iterate(&pushOrMarkClosure);
7537 do_yield_check();
7538 }
7539 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
7540 }
7542 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
7543 CMSCollector* collector, MemRegion span,
7544 CMSBitMap* bit_map,
7545 OopTaskQueue* work_queue,
7546 CMSMarkStack* overflow_stack,
7547 bool should_yield):
7548 _collector(collector),
7549 _whole_span(collector->_span),
7550 _span(span),
7551 _bit_map(bit_map),
7552 _mut(&collector->_modUnionTable),
7553 _work_queue(work_queue),
7554 _overflow_stack(overflow_stack),
7555 _yield(should_yield),
7556 _skip_bits(0),
7557 _task(task)
7558 {
7559 assert(_work_queue->size() == 0, "work_queue should be empty");
7560 _finger = span.start();
7561 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
7562 assert(_span.contains(_finger), "Out of bounds _finger?");
7563 }
7565 // Should revisit to see if this should be restructured for
7566 // greater efficiency.
7567 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
7568 if (_skip_bits > 0) {
7569 _skip_bits--;
7570 return true;
7571 }
7572 // convert offset into a HeapWord*
7573 HeapWord* addr = _bit_map->startWord() + offset;
7574 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
7575 "address out of range");
7576 assert(_bit_map->isMarked(addr), "tautology");
7577 if (_bit_map->isMarked(addr+1)) {
7578 // this is an allocated object that might not yet be initialized
7579 assert(_skip_bits == 0, "tautology");
7580 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
7581 oop p = oop(addr);
7582 if (p->klass_or_null() == NULL) {
7583 // in the case of Clean-on-Enter optimization, redirty card
7584 // and avoid clearing card by increasing the threshold.
7585 return true;
7586 }
7587 }
7588 scan_oops_in_oop(addr);
7589 return true;
7590 }
7592 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
7593 assert(_bit_map->isMarked(ptr), "expected bit to be set");
7594 // Should we assert that our work queue is empty or
7595 // below some drain limit?
7596 assert(_work_queue->size() == 0,
7597 "should drain stack to limit stack usage");
7598 // convert ptr to an oop preparatory to scanning
7599 oop obj = oop(ptr);
7600 // Ignore mark word in verification below, since we
7601 // may be running concurrent with mutators.
7602 assert(obj->is_oop(true), "should be an oop");
7603 assert(_finger <= ptr, "_finger runneth ahead");
7604 // advance the finger to right end of this object
7605 _finger = ptr + obj->size();
7606 assert(_finger > ptr, "we just incremented it above");
7607 // On large heaps, it may take us some time to get through
7608 // the marking phase (especially if running iCMS). During
7609 // this time it's possible that a lot of mutations have
7610 // accumulated in the card table and the mod union table --
7611 // these mutation records are redundant until we have
7612 // actually traced into the corresponding card.
7613 // Here, we check whether advancing the finger would make
7614 // us cross into a new card, and if so clear corresponding
7615 // cards in the MUT (preclean them in the card-table in the
7616 // future).
7618 // The clean-on-enter optimization is disabled by default,
7619 // until we fix 6178663.
7620 if (CMSCleanOnEnter && (_finger > _threshold)) {
7621 // [_threshold, _finger) represents the interval
7622 // of cards to be cleared in MUT (or precleaned in card table).
7623 // The set of cards to be cleared is all those that overlap
7624 // with the interval [_threshold, _finger); note that
7625 // _threshold is always kept card-aligned but _finger isn't
7626 // always card-aligned.
7627 HeapWord* old_threshold = _threshold;
7628 assert(old_threshold == (HeapWord*)round_to(
7629 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7630 "_threshold should always be card-aligned");
7631 _threshold = (HeapWord*)round_to(
7632 (intptr_t)_finger, CardTableModRefBS::card_size);
7633 MemRegion mr(old_threshold, _threshold);
7634 assert(!mr.is_empty(), "Control point invariant");
7635 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7636 _mut->clear_range(mr);
7637 }
7639 // Note: the local finger doesn't advance while we drain
7640 // the stack below, but the global finger sure can and will.
7641 HeapWord** gfa = _task->global_finger_addr();
7642 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7643 _span, _bit_map,
7644 _work_queue,
7645 _overflow_stack,
7646 _finger,
7647 gfa, this);
7648 bool res = _work_queue->push(obj); // overflow could occur here
7649 assert(res, "Will hold once we use workqueues");
7650 while (true) {
7651 oop new_oop;
7652 if (!_work_queue->pop_local(new_oop)) {
7653 // We emptied our work_queue; check if there's stuff that can
7654 // be gotten from the overflow stack.
7655 if (CMSConcMarkingTask::get_work_from_overflow_stack(
7656 _overflow_stack, _work_queue)) {
7657 do_yield_check();
7658 continue;
7659 } else { // done
7660 break;
7661 }
7662 }
7663 // Skip verifying header mark word below because we are
7664 // running concurrent with mutators.
7665 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7666 // now scan this oop's oops
7667 new_oop->oop_iterate(&pushOrMarkClosure);
7668 do_yield_check();
7669 }
7670 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7671 }
7673 // Yield in response to a request from VM Thread or
7674 // from mutators.
7675 void Par_MarkFromRootsClosure::do_yield_work() {
7676 assert(_task != NULL, "sanity");
7677 _task->yield();
7678 }
7680 // A variant of the above used for verifying CMS marking work.
7681 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7682 MemRegion span,
7683 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7684 CMSMarkStack* mark_stack):
7685 _collector(collector),
7686 _span(span),
7687 _verification_bm(verification_bm),
7688 _cms_bm(cms_bm),
7689 _mark_stack(mark_stack),
7690 _pam_verify_closure(collector, span, verification_bm, cms_bm,
7691 mark_stack)
7692 {
7693 assert(_mark_stack->isEmpty(), "stack should be empty");
7694 _finger = _verification_bm->startWord();
7695 assert(_collector->_restart_addr == NULL, "Sanity check");
7696 assert(_span.contains(_finger), "Out of bounds _finger?");
7697 }
7699 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7700 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7701 assert(_span.contains(addr), "Out of bounds _finger?");
7702 _finger = addr;
7703 }
7705 // Should revisit to see if this should be restructured for
7706 // greater efficiency.
7707 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7708 // convert offset into a HeapWord*
7709 HeapWord* addr = _verification_bm->startWord() + offset;
7710 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7711 "address out of range");
7712 assert(_verification_bm->isMarked(addr), "tautology");
7713 assert(_cms_bm->isMarked(addr), "tautology");
7715 assert(_mark_stack->isEmpty(),
7716 "should drain stack to limit stack usage");
7717 // convert addr to an oop preparatory to scanning
7718 oop obj = oop(addr);
7719 assert(obj->is_oop(), "should be an oop");
7720 assert(_finger <= addr, "_finger runneth ahead");
7721 // advance the finger to right end of this object
7722 _finger = addr + obj->size();
7723 assert(_finger > addr, "we just incremented it above");
7724 // Note: the finger doesn't advance while we drain
7725 // the stack below.
7726 bool res = _mark_stack->push(obj);
7727 assert(res, "Empty non-zero size stack should have space for single push");
7728 while (!_mark_stack->isEmpty()) {
7729 oop new_oop = _mark_stack->pop();
7730 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7731 // now scan this oop's oops
7732 new_oop->oop_iterate(&_pam_verify_closure);
7733 }
7734 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7735 return true;
7736 }
7738 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7739 CMSCollector* collector, MemRegion span,
7740 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7741 CMSMarkStack* mark_stack):
7742 CMSOopClosure(collector->ref_processor()),
7743 _collector(collector),
7744 _span(span),
7745 _verification_bm(verification_bm),
7746 _cms_bm(cms_bm),
7747 _mark_stack(mark_stack)
7748 { }
7750 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7751 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7753 // Upon stack overflow, we discard (part of) the stack,
7754 // remembering the least address amongst those discarded
7755 // in CMSCollector's _restart_address.
7756 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7757 // Remember the least grey address discarded
7758 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7759 _collector->lower_restart_addr(ra);
7760 _mark_stack->reset(); // discard stack contents
7761 _mark_stack->expand(); // expand the stack if possible
7762 }
7764 void PushAndMarkVerifyClosure::do_oop(oop obj) {
7765 assert(obj->is_oop_or_null(), "expected an oop or NULL");
7766 HeapWord* addr = (HeapWord*)obj;
7767 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7768 // Oop lies in _span and isn't yet grey or black
7769 _verification_bm->mark(addr); // now grey
7770 if (!_cms_bm->isMarked(addr)) {
7771 oop(addr)->print();
7772 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
7773 addr);
7774 fatal("... aborting");
7775 }
7777 if (!_mark_stack->push(obj)) { // stack overflow
7778 if (PrintCMSStatistics != 0) {
7779 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7780 SIZE_FORMAT, _mark_stack->capacity());
7781 }
7782 assert(_mark_stack->isFull(), "Else push should have succeeded");
7783 handle_stack_overflow(addr);
7784 }
7785 // anything including and to the right of _finger
7786 // will be scanned as we iterate over the remainder of the
7787 // bit map
7788 }
7789 }
7791 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7792 MemRegion span,
7793 CMSBitMap* bitMap, CMSMarkStack* markStack,
7794 HeapWord* finger, MarkFromRootsClosure* parent) :
7795 CMSOopClosure(collector->ref_processor()),
7796 _collector(collector),
7797 _span(span),
7798 _bitMap(bitMap),
7799 _markStack(markStack),
7800 _finger(finger),
7801 _parent(parent)
7802 { }
7804 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7805 MemRegion span,
7806 CMSBitMap* bit_map,
7807 OopTaskQueue* work_queue,
7808 CMSMarkStack* overflow_stack,
7809 HeapWord* finger,
7810 HeapWord** global_finger_addr,
7811 Par_MarkFromRootsClosure* parent) :
7812 CMSOopClosure(collector->ref_processor()),
7813 _collector(collector),
7814 _whole_span(collector->_span),
7815 _span(span),
7816 _bit_map(bit_map),
7817 _work_queue(work_queue),
7818 _overflow_stack(overflow_stack),
7819 _finger(finger),
7820 _global_finger_addr(global_finger_addr),
7821 _parent(parent)
7822 { }
7824 // Assumes thread-safe access by callers, who are
7825 // responsible for mutual exclusion.
7826 void CMSCollector::lower_restart_addr(HeapWord* low) {
7827 assert(_span.contains(low), "Out of bounds addr");
7828 if (_restart_addr == NULL) {
7829 _restart_addr = low;
7830 } else {
7831 _restart_addr = MIN2(_restart_addr, low);
7832 }
7833 }
7835 // Upon stack overflow, we discard (part of) the stack,
7836 // remembering the least address amongst those discarded
7837 // in CMSCollector's _restart_address.
7838 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7839 // Remember the least grey address discarded
7840 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7841 _collector->lower_restart_addr(ra);
7842 _markStack->reset(); // discard stack contents
7843 _markStack->expand(); // expand the stack if possible
7844 }
7846 // Upon stack overflow, we discard (part of) the stack,
7847 // remembering the least address amongst those discarded
7848 // in CMSCollector's _restart_address.
7849 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7850 // We need to do this under a mutex to prevent other
7851 // workers from interfering with the work done below.
7852 MutexLockerEx ml(_overflow_stack->par_lock(),
7853 Mutex::_no_safepoint_check_flag);
7854 // Remember the least grey address discarded
7855 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7856 _collector->lower_restart_addr(ra);
7857 _overflow_stack->reset(); // discard stack contents
7858 _overflow_stack->expand(); // expand the stack if possible
7859 }
7861 void CMKlassClosure::do_klass(Klass* k) {
7862 assert(_oop_closure != NULL, "Not initialized?");
7863 k->oops_do(_oop_closure);
7864 }
7866 void PushOrMarkClosure::do_oop(oop obj) {
7867 // Ignore mark word because we are running concurrent with mutators.
7868 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7869 HeapWord* addr = (HeapWord*)obj;
7870 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7871 // Oop lies in _span and isn't yet grey or black
7872 _bitMap->mark(addr); // now grey
7873 if (addr < _finger) {
7874 // the bit map iteration has already either passed, or
7875 // sampled, this bit in the bit map; we'll need to
7876 // use the marking stack to scan this oop's oops.
7877 bool simulate_overflow = false;
7878 NOT_PRODUCT(
7879 if (CMSMarkStackOverflowALot &&
7880 _collector->simulate_overflow()) {
7881 // simulate a stack overflow
7882 simulate_overflow = true;
7883 }
7884 )
7885 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7886 if (PrintCMSStatistics != 0) {
7887 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7888 SIZE_FORMAT, _markStack->capacity());
7889 }
7890 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7891 handle_stack_overflow(addr);
7892 }
7893 }
7894 // anything including and to the right of _finger
7895 // will be scanned as we iterate over the remainder of the
7896 // bit map
7897 do_yield_check();
7898 }
7899 }
7901 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7902 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7904 void Par_PushOrMarkClosure::do_oop(oop obj) {
7905 // Ignore mark word because we are running concurrent with mutators.
7906 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7907 HeapWord* addr = (HeapWord*)obj;
7908 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7909 // Oop lies in _span and isn't yet grey or black
7910 // We read the global_finger (volatile read) strictly after marking oop
7911 bool res = _bit_map->par_mark(addr); // now grey
7912 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7913 // Should we push this marked oop on our stack?
7914 // -- if someone else marked it, nothing to do
7915 // -- if target oop is above global finger nothing to do
7916 // -- if target oop is in chunk and above local finger
7917 // then nothing to do
7918 // -- else push on work queue
7919 if ( !res // someone else marked it, they will deal with it
7920 || (addr >= *gfa) // will be scanned in a later task
7921 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7922 return;
7923 }
7924 // the bit map iteration has already either passed, or
7925 // sampled, this bit in the bit map; we'll need to
7926 // use the marking stack to scan this oop's oops.
7927 bool simulate_overflow = false;
7928 NOT_PRODUCT(
7929 if (CMSMarkStackOverflowALot &&
7930 _collector->simulate_overflow()) {
7931 // simulate a stack overflow
7932 simulate_overflow = true;
7933 }
7934 )
7935 if (simulate_overflow ||
7936 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7937 // stack overflow
7938 if (PrintCMSStatistics != 0) {
7939 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7940 SIZE_FORMAT, _overflow_stack->capacity());
7941 }
7942 // We cannot assert that the overflow stack is full because
7943 // it may have been emptied since.
7944 assert(simulate_overflow ||
7945 _work_queue->size() == _work_queue->max_elems(),
7946 "Else push should have succeeded");
7947 handle_stack_overflow(addr);
7948 }
7949 do_yield_check();
7950 }
7951 }
7953 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7954 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7956 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7957 MemRegion span,
7958 ReferenceProcessor* rp,
7959 CMSBitMap* bit_map,
7960 CMSBitMap* mod_union_table,
7961 CMSMarkStack* mark_stack,
7962 bool concurrent_precleaning):
7963 CMSOopClosure(rp),
7964 _collector(collector),
7965 _span(span),
7966 _bit_map(bit_map),
7967 _mod_union_table(mod_union_table),
7968 _mark_stack(mark_stack),
7969 _concurrent_precleaning(concurrent_precleaning)
7970 {
7971 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7972 }
7974 // Grey object rescan during pre-cleaning and second checkpoint phases --
7975 // the non-parallel version (the parallel version appears further below.)
7976 void PushAndMarkClosure::do_oop(oop obj) {
7977 // Ignore mark word verification. If during concurrent precleaning,
7978 // the object monitor may be locked. If during the checkpoint
7979 // phases, the object may already have been reached by a different
7980 // path and may be at the end of the global overflow list (so
7981 // the mark word may be NULL).
7982 assert(obj->is_oop_or_null(true /* ignore mark word */),
7983 "expected an oop or NULL");
7984 HeapWord* addr = (HeapWord*)obj;
7985 // Check if oop points into the CMS generation
7986 // and is not marked
7987 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7988 // a white object ...
7989 _bit_map->mark(addr); // ... now grey
7990 // push on the marking stack (grey set)
7991 bool simulate_overflow = false;
7992 NOT_PRODUCT(
7993 if (CMSMarkStackOverflowALot &&
7994 _collector->simulate_overflow()) {
7995 // simulate a stack overflow
7996 simulate_overflow = true;
7997 }
7998 )
7999 if (simulate_overflow || !_mark_stack->push(obj)) {
8000 if (_concurrent_precleaning) {
8001 // During precleaning we can just dirty the appropriate card(s)
8002 // in the mod union table, thus ensuring that the object remains
8003 // in the grey set and continue. In the case of object arrays
8004 // we need to dirty all of the cards that the object spans,
8005 // since the rescan of object arrays will be limited to the
8006 // dirty cards.
8007 // Note that no one can be intefering with us in this action
8008 // of dirtying the mod union table, so no locking or atomics
8009 // are required.
8010 if (obj->is_objArray()) {
8011 size_t sz = obj->size();
8012 HeapWord* end_card_addr = (HeapWord*)round_to(
8013 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
8014 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8015 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8016 _mod_union_table->mark_range(redirty_range);
8017 } else {
8018 _mod_union_table->mark(addr);
8019 }
8020 _collector->_ser_pmc_preclean_ovflw++;
8021 } else {
8022 // During the remark phase, we need to remember this oop
8023 // in the overflow list.
8024 _collector->push_on_overflow_list(obj);
8025 _collector->_ser_pmc_remark_ovflw++;
8026 }
8027 }
8028 }
8029 }
8031 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
8032 MemRegion span,
8033 ReferenceProcessor* rp,
8034 CMSBitMap* bit_map,
8035 OopTaskQueue* work_queue):
8036 CMSOopClosure(rp),
8037 _collector(collector),
8038 _span(span),
8039 _bit_map(bit_map),
8040 _work_queue(work_queue)
8041 {
8042 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
8043 }
8045 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
8046 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
8048 // Grey object rescan during second checkpoint phase --
8049 // the parallel version.
8050 void Par_PushAndMarkClosure::do_oop(oop obj) {
8051 // In the assert below, we ignore the mark word because
8052 // this oop may point to an already visited object that is
8053 // on the overflow stack (in which case the mark word has
8054 // been hijacked for chaining into the overflow stack --
8055 // if this is the last object in the overflow stack then
8056 // its mark word will be NULL). Because this object may
8057 // have been subsequently popped off the global overflow
8058 // stack, and the mark word possibly restored to the prototypical
8059 // value, by the time we get to examined this failing assert in
8060 // the debugger, is_oop_or_null(false) may subsequently start
8061 // to hold.
8062 assert(obj->is_oop_or_null(true),
8063 "expected an oop or NULL");
8064 HeapWord* addr = (HeapWord*)obj;
8065 // Check if oop points into the CMS generation
8066 // and is not marked
8067 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
8068 // a white object ...
8069 // If we manage to "claim" the object, by being the
8070 // first thread to mark it, then we push it on our
8071 // marking stack
8072 if (_bit_map->par_mark(addr)) { // ... now grey
8073 // push on work queue (grey set)
8074 bool simulate_overflow = false;
8075 NOT_PRODUCT(
8076 if (CMSMarkStackOverflowALot &&
8077 _collector->par_simulate_overflow()) {
8078 // simulate a stack overflow
8079 simulate_overflow = true;
8080 }
8081 )
8082 if (simulate_overflow || !_work_queue->push(obj)) {
8083 _collector->par_push_on_overflow_list(obj);
8084 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
8085 }
8086 } // Else, some other thread got there first
8087 }
8088 }
8090 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
8091 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
8093 void CMSPrecleanRefsYieldClosure::do_yield_work() {
8094 Mutex* bml = _collector->bitMapLock();
8095 assert_lock_strong(bml);
8096 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8097 "CMS thread should hold CMS token");
8099 bml->unlock();
8100 ConcurrentMarkSweepThread::desynchronize(true);
8102 ConcurrentMarkSweepThread::acknowledge_yield_request();
8104 _collector->stopTimer();
8105 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8106 if (PrintCMSStatistics != 0) {
8107 _collector->incrementYields();
8108 }
8109 _collector->icms_wait();
8111 // See the comment in coordinator_yield()
8112 for (unsigned i = 0; i < CMSYieldSleepCount &&
8113 ConcurrentMarkSweepThread::should_yield() &&
8114 !CMSCollector::foregroundGCIsActive(); ++i) {
8115 os::sleep(Thread::current(), 1, false);
8116 ConcurrentMarkSweepThread::acknowledge_yield_request();
8117 }
8119 ConcurrentMarkSweepThread::synchronize(true);
8120 bml->lock();
8122 _collector->startTimer();
8123 }
8125 bool CMSPrecleanRefsYieldClosure::should_return() {
8126 if (ConcurrentMarkSweepThread::should_yield()) {
8127 do_yield_work();
8128 }
8129 return _collector->foregroundGCIsActive();
8130 }
8132 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
8133 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
8134 "mr should be aligned to start at a card boundary");
8135 // We'd like to assert:
8136 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
8137 // "mr should be a range of cards");
8138 // However, that would be too strong in one case -- the last
8139 // partition ends at _unallocated_block which, in general, can be
8140 // an arbitrary boundary, not necessarily card aligned.
8141 if (PrintCMSStatistics != 0) {
8142 _num_dirty_cards +=
8143 mr.word_size()/CardTableModRefBS::card_size_in_words;
8144 }
8145 _space->object_iterate_mem(mr, &_scan_cl);
8146 }
8148 SweepClosure::SweepClosure(CMSCollector* collector,
8149 ConcurrentMarkSweepGeneration* g,
8150 CMSBitMap* bitMap, bool should_yield) :
8151 _collector(collector),
8152 _g(g),
8153 _sp(g->cmsSpace()),
8154 _limit(_sp->sweep_limit()),
8155 _freelistLock(_sp->freelistLock()),
8156 _bitMap(bitMap),
8157 _yield(should_yield),
8158 _inFreeRange(false), // No free range at beginning of sweep
8159 _freeRangeInFreeLists(false), // No free range at beginning of sweep
8160 _lastFreeRangeCoalesced(false),
8161 _freeFinger(g->used_region().start())
8162 {
8163 NOT_PRODUCT(
8164 _numObjectsFreed = 0;
8165 _numWordsFreed = 0;
8166 _numObjectsLive = 0;
8167 _numWordsLive = 0;
8168 _numObjectsAlreadyFree = 0;
8169 _numWordsAlreadyFree = 0;
8170 _last_fc = NULL;
8172 _sp->initializeIndexedFreeListArrayReturnedBytes();
8173 _sp->dictionary()->initialize_dict_returned_bytes();
8174 )
8175 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8176 "sweep _limit out of bounds");
8177 if (CMSTraceSweeper) {
8178 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
8179 _limit);
8180 }
8181 }
8183 void SweepClosure::print_on(outputStream* st) const {
8184 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
8185 _sp->bottom(), _sp->end());
8186 tty->print_cr("_limit = " PTR_FORMAT, _limit);
8187 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger);
8188 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);)
8189 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
8190 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
8191 }
8193 #ifndef PRODUCT
8194 // Assertion checking only: no useful work in product mode --
8195 // however, if any of the flags below become product flags,
8196 // you may need to review this code to see if it needs to be
8197 // enabled in product mode.
8198 SweepClosure::~SweepClosure() {
8199 assert_lock_strong(_freelistLock);
8200 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8201 "sweep _limit out of bounds");
8202 if (inFreeRange()) {
8203 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
8204 print();
8205 ShouldNotReachHere();
8206 }
8207 if (Verbose && PrintGC) {
8208 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes",
8209 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
8210 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
8211 SIZE_FORMAT" bytes "
8212 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
8213 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
8214 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
8215 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
8216 * sizeof(HeapWord);
8217 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
8219 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
8220 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
8221 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
8222 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
8223 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes);
8224 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
8225 indexListReturnedBytes);
8226 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
8227 dict_returned_bytes);
8228 }
8229 }
8230 if (CMSTraceSweeper) {
8231 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
8232 _limit);
8233 }
8234 }
8235 #endif // PRODUCT
8237 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
8238 bool freeRangeInFreeLists) {
8239 if (CMSTraceSweeper) {
8240 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n",
8241 freeFinger, freeRangeInFreeLists);
8242 }
8243 assert(!inFreeRange(), "Trampling existing free range");
8244 set_inFreeRange(true);
8245 set_lastFreeRangeCoalesced(false);
8247 set_freeFinger(freeFinger);
8248 set_freeRangeInFreeLists(freeRangeInFreeLists);
8249 if (CMSTestInFreeList) {
8250 if (freeRangeInFreeLists) {
8251 FreeChunk* fc = (FreeChunk*) freeFinger;
8252 assert(fc->is_free(), "A chunk on the free list should be free.");
8253 assert(fc->size() > 0, "Free range should have a size");
8254 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
8255 }
8256 }
8257 }
8259 // Note that the sweeper runs concurrently with mutators. Thus,
8260 // it is possible for direct allocation in this generation to happen
8261 // in the middle of the sweep. Note that the sweeper also coalesces
8262 // contiguous free blocks. Thus, unless the sweeper and the allocator
8263 // synchronize appropriately freshly allocated blocks may get swept up.
8264 // This is accomplished by the sweeper locking the free lists while
8265 // it is sweeping. Thus blocks that are determined to be free are
8266 // indeed free. There is however one additional complication:
8267 // blocks that have been allocated since the final checkpoint and
8268 // mark, will not have been marked and so would be treated as
8269 // unreachable and swept up. To prevent this, the allocator marks
8270 // the bit map when allocating during the sweep phase. This leads,
8271 // however, to a further complication -- objects may have been allocated
8272 // but not yet initialized -- in the sense that the header isn't yet
8273 // installed. The sweeper can not then determine the size of the block
8274 // in order to skip over it. To deal with this case, we use a technique
8275 // (due to Printezis) to encode such uninitialized block sizes in the
8276 // bit map. Since the bit map uses a bit per every HeapWord, but the
8277 // CMS generation has a minimum object size of 3 HeapWords, it follows
8278 // that "normal marks" won't be adjacent in the bit map (there will
8279 // always be at least two 0 bits between successive 1 bits). We make use
8280 // of these "unused" bits to represent uninitialized blocks -- the bit
8281 // corresponding to the start of the uninitialized object and the next
8282 // bit are both set. Finally, a 1 bit marks the end of the object that
8283 // started with the two consecutive 1 bits to indicate its potentially
8284 // uninitialized state.
8286 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
8287 FreeChunk* fc = (FreeChunk*)addr;
8288 size_t res;
8290 // Check if we are done sweeping. Below we check "addr >= _limit" rather
8291 // than "addr == _limit" because although _limit was a block boundary when
8292 // we started the sweep, it may no longer be one because heap expansion
8293 // may have caused us to coalesce the block ending at the address _limit
8294 // with a newly expanded chunk (this happens when _limit was set to the
8295 // previous _end of the space), so we may have stepped past _limit:
8296 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
8297 if (addr >= _limit) { // we have swept up to or past the limit: finish up
8298 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8299 "sweep _limit out of bounds");
8300 assert(addr < _sp->end(), "addr out of bounds");
8301 // Flush any free range we might be holding as a single
8302 // coalesced chunk to the appropriate free list.
8303 if (inFreeRange()) {
8304 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
8305 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger()));
8306 flush_cur_free_chunk(freeFinger(),
8307 pointer_delta(addr, freeFinger()));
8308 if (CMSTraceSweeper) {
8309 gclog_or_tty->print("Sweep: last chunk: ");
8310 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") "
8311 "[coalesced:"SIZE_FORMAT"]\n",
8312 freeFinger(), pointer_delta(addr, freeFinger()),
8313 lastFreeRangeCoalesced());
8314 }
8315 }
8317 // help the iterator loop finish
8318 return pointer_delta(_sp->end(), addr);
8319 }
8321 assert(addr < _limit, "sweep invariant");
8322 // check if we should yield
8323 do_yield_check(addr);
8324 if (fc->is_free()) {
8325 // Chunk that is already free
8326 res = fc->size();
8327 do_already_free_chunk(fc);
8328 debug_only(_sp->verifyFreeLists());
8329 // If we flush the chunk at hand in lookahead_and_flush()
8330 // and it's coalesced with a preceding chunk, then the
8331 // process of "mangling" the payload of the coalesced block
8332 // will cause erasure of the size information from the
8333 // (erstwhile) header of all the coalesced blocks but the
8334 // first, so the first disjunct in the assert will not hold
8335 // in that specific case (in which case the second disjunct
8336 // will hold).
8337 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
8338 "Otherwise the size info doesn't change at this step");
8339 NOT_PRODUCT(
8340 _numObjectsAlreadyFree++;
8341 _numWordsAlreadyFree += res;
8342 )
8343 NOT_PRODUCT(_last_fc = fc;)
8344 } else if (!_bitMap->isMarked(addr)) {
8345 // Chunk is fresh garbage
8346 res = do_garbage_chunk(fc);
8347 debug_only(_sp->verifyFreeLists());
8348 NOT_PRODUCT(
8349 _numObjectsFreed++;
8350 _numWordsFreed += res;
8351 )
8352 } else {
8353 // Chunk that is alive.
8354 res = do_live_chunk(fc);
8355 debug_only(_sp->verifyFreeLists());
8356 NOT_PRODUCT(
8357 _numObjectsLive++;
8358 _numWordsLive += res;
8359 )
8360 }
8361 return res;
8362 }
8364 // For the smart allocation, record following
8365 // split deaths - a free chunk is removed from its free list because
8366 // it is being split into two or more chunks.
8367 // split birth - a free chunk is being added to its free list because
8368 // a larger free chunk has been split and resulted in this free chunk.
8369 // coal death - a free chunk is being removed from its free list because
8370 // it is being coalesced into a large free chunk.
8371 // coal birth - a free chunk is being added to its free list because
8372 // it was created when two or more free chunks where coalesced into
8373 // this free chunk.
8374 //
8375 // These statistics are used to determine the desired number of free
8376 // chunks of a given size. The desired number is chosen to be relative
8377 // to the end of a CMS sweep. The desired number at the end of a sweep
8378 // is the
8379 // count-at-end-of-previous-sweep (an amount that was enough)
8380 // - count-at-beginning-of-current-sweep (the excess)
8381 // + split-births (gains in this size during interval)
8382 // - split-deaths (demands on this size during interval)
8383 // where the interval is from the end of one sweep to the end of the
8384 // next.
8385 //
8386 // When sweeping the sweeper maintains an accumulated chunk which is
8387 // the chunk that is made up of chunks that have been coalesced. That
8388 // will be termed the left-hand chunk. A new chunk of garbage that
8389 // is being considered for coalescing will be referred to as the
8390 // right-hand chunk.
8391 //
8392 // When making a decision on whether to coalesce a right-hand chunk with
8393 // the current left-hand chunk, the current count vs. the desired count
8394 // of the left-hand chunk is considered. Also if the right-hand chunk
8395 // is near the large chunk at the end of the heap (see
8396 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
8397 // left-hand chunk is coalesced.
8398 //
8399 // When making a decision about whether to split a chunk, the desired count
8400 // vs. the current count of the candidate to be split is also considered.
8401 // If the candidate is underpopulated (currently fewer chunks than desired)
8402 // a chunk of an overpopulated (currently more chunks than desired) size may
8403 // be chosen. The "hint" associated with a free list, if non-null, points
8404 // to a free list which may be overpopulated.
8405 //
8407 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
8408 const size_t size = fc->size();
8409 // Chunks that cannot be coalesced are not in the
8410 // free lists.
8411 if (CMSTestInFreeList && !fc->cantCoalesce()) {
8412 assert(_sp->verify_chunk_in_free_list(fc),
8413 "free chunk should be in free lists");
8414 }
8415 // a chunk that is already free, should not have been
8416 // marked in the bit map
8417 HeapWord* const addr = (HeapWord*) fc;
8418 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
8419 // Verify that the bit map has no bits marked between
8420 // addr and purported end of this block.
8421 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8423 // Some chunks cannot be coalesced under any circumstances.
8424 // See the definition of cantCoalesce().
8425 if (!fc->cantCoalesce()) {
8426 // This chunk can potentially be coalesced.
8427 if (_sp->adaptive_freelists()) {
8428 // All the work is done in
8429 do_post_free_or_garbage_chunk(fc, size);
8430 } else { // Not adaptive free lists
8431 // this is a free chunk that can potentially be coalesced by the sweeper;
8432 if (!inFreeRange()) {
8433 // if the next chunk is a free block that can't be coalesced
8434 // it doesn't make sense to remove this chunk from the free lists
8435 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
8436 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
8437 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
8438 nextChunk->is_free() && // ... which is free...
8439 nextChunk->cantCoalesce()) { // ... but can't be coalesced
8440 // nothing to do
8441 } else {
8442 // Potentially the start of a new free range:
8443 // Don't eagerly remove it from the free lists.
8444 // No need to remove it if it will just be put
8445 // back again. (Also from a pragmatic point of view
8446 // if it is a free block in a region that is beyond
8447 // any allocated blocks, an assertion will fail)
8448 // Remember the start of a free run.
8449 initialize_free_range(addr, true);
8450 // end - can coalesce with next chunk
8451 }
8452 } else {
8453 // the midst of a free range, we are coalescing
8454 print_free_block_coalesced(fc);
8455 if (CMSTraceSweeper) {
8456 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size);
8457 }
8458 // remove it from the free lists
8459 _sp->removeFreeChunkFromFreeLists(fc);
8460 set_lastFreeRangeCoalesced(true);
8461 // If the chunk is being coalesced and the current free range is
8462 // in the free lists, remove the current free range so that it
8463 // will be returned to the free lists in its entirety - all
8464 // the coalesced pieces included.
8465 if (freeRangeInFreeLists()) {
8466 FreeChunk* ffc = (FreeChunk*) freeFinger();
8467 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8468 "Size of free range is inconsistent with chunk size.");
8469 if (CMSTestInFreeList) {
8470 assert(_sp->verify_chunk_in_free_list(ffc),
8471 "free range is not in free lists");
8472 }
8473 _sp->removeFreeChunkFromFreeLists(ffc);
8474 set_freeRangeInFreeLists(false);
8475 }
8476 }
8477 }
8478 // Note that if the chunk is not coalescable (the else arm
8479 // below), we unconditionally flush, without needing to do
8480 // a "lookahead," as we do below.
8481 if (inFreeRange()) lookahead_and_flush(fc, size);
8482 } else {
8483 // Code path common to both original and adaptive free lists.
8485 // cant coalesce with previous block; this should be treated
8486 // as the end of a free run if any
8487 if (inFreeRange()) {
8488 // we kicked some butt; time to pick up the garbage
8489 assert(freeFinger() < addr, "freeFinger points too high");
8490 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8491 }
8492 // else, nothing to do, just continue
8493 }
8494 }
8496 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
8497 // This is a chunk of garbage. It is not in any free list.
8498 // Add it to a free list or let it possibly be coalesced into
8499 // a larger chunk.
8500 HeapWord* const addr = (HeapWord*) fc;
8501 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8503 if (_sp->adaptive_freelists()) {
8504 // Verify that the bit map has no bits marked between
8505 // addr and purported end of just dead object.
8506 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8508 do_post_free_or_garbage_chunk(fc, size);
8509 } else {
8510 if (!inFreeRange()) {
8511 // start of a new free range
8512 assert(size > 0, "A free range should have a size");
8513 initialize_free_range(addr, false);
8514 } else {
8515 // this will be swept up when we hit the end of the
8516 // free range
8517 if (CMSTraceSweeper) {
8518 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size);
8519 }
8520 // If the chunk is being coalesced and the current free range is
8521 // in the free lists, remove the current free range so that it
8522 // will be returned to the free lists in its entirety - all
8523 // the coalesced pieces included.
8524 if (freeRangeInFreeLists()) {
8525 FreeChunk* ffc = (FreeChunk*)freeFinger();
8526 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8527 "Size of free range is inconsistent with chunk size.");
8528 if (CMSTestInFreeList) {
8529 assert(_sp->verify_chunk_in_free_list(ffc),
8530 "free range is not in free lists");
8531 }
8532 _sp->removeFreeChunkFromFreeLists(ffc);
8533 set_freeRangeInFreeLists(false);
8534 }
8535 set_lastFreeRangeCoalesced(true);
8536 }
8537 // this will be swept up when we hit the end of the free range
8539 // Verify that the bit map has no bits marked between
8540 // addr and purported end of just dead object.
8541 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8542 }
8543 assert(_limit >= addr + size,
8544 "A freshly garbage chunk can't possibly straddle over _limit");
8545 if (inFreeRange()) lookahead_and_flush(fc, size);
8546 return size;
8547 }
8549 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
8550 HeapWord* addr = (HeapWord*) fc;
8551 // The sweeper has just found a live object. Return any accumulated
8552 // left hand chunk to the free lists.
8553 if (inFreeRange()) {
8554 assert(freeFinger() < addr, "freeFinger points too high");
8555 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8556 }
8558 // This object is live: we'd normally expect this to be
8559 // an oop, and like to assert the following:
8560 // assert(oop(addr)->is_oop(), "live block should be an oop");
8561 // However, as we commented above, this may be an object whose
8562 // header hasn't yet been initialized.
8563 size_t size;
8564 assert(_bitMap->isMarked(addr), "Tautology for this control point");
8565 if (_bitMap->isMarked(addr + 1)) {
8566 // Determine the size from the bit map, rather than trying to
8567 // compute it from the object header.
8568 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
8569 size = pointer_delta(nextOneAddr + 1, addr);
8570 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
8571 "alignment problem");
8573 #ifdef ASSERT
8574 if (oop(addr)->klass_or_null() != NULL) {
8575 // Ignore mark word because we are running concurrent with mutators
8576 assert(oop(addr)->is_oop(true), "live block should be an oop");
8577 assert(size ==
8578 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
8579 "P-mark and computed size do not agree");
8580 }
8581 #endif
8583 } else {
8584 // This should be an initialized object that's alive.
8585 assert(oop(addr)->klass_or_null() != NULL,
8586 "Should be an initialized object");
8587 // Ignore mark word because we are running concurrent with mutators
8588 assert(oop(addr)->is_oop(true), "live block should be an oop");
8589 // Verify that the bit map has no bits marked between
8590 // addr and purported end of this block.
8591 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8592 assert(size >= 3, "Necessary for Printezis marks to work");
8593 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
8594 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
8595 }
8596 return size;
8597 }
8599 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
8600 size_t chunkSize) {
8601 // do_post_free_or_garbage_chunk() should only be called in the case
8602 // of the adaptive free list allocator.
8603 const bool fcInFreeLists = fc->is_free();
8604 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
8605 assert((HeapWord*)fc <= _limit, "sweep invariant");
8606 if (CMSTestInFreeList && fcInFreeLists) {
8607 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
8608 }
8610 if (CMSTraceSweeper) {
8611 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
8612 }
8614 HeapWord* const fc_addr = (HeapWord*) fc;
8616 bool coalesce;
8617 const size_t left = pointer_delta(fc_addr, freeFinger());
8618 const size_t right = chunkSize;
8619 switch (FLSCoalescePolicy) {
8620 // numeric value forms a coalition aggressiveness metric
8621 case 0: { // never coalesce
8622 coalesce = false;
8623 break;
8624 }
8625 case 1: { // coalesce if left & right chunks on overpopulated lists
8626 coalesce = _sp->coalOverPopulated(left) &&
8627 _sp->coalOverPopulated(right);
8628 break;
8629 }
8630 case 2: { // coalesce if left chunk on overpopulated list (default)
8631 coalesce = _sp->coalOverPopulated(left);
8632 break;
8633 }
8634 case 3: { // coalesce if left OR right chunk on overpopulated list
8635 coalesce = _sp->coalOverPopulated(left) ||
8636 _sp->coalOverPopulated(right);
8637 break;
8638 }
8639 case 4: { // always coalesce
8640 coalesce = true;
8641 break;
8642 }
8643 default:
8644 ShouldNotReachHere();
8645 }
8647 // Should the current free range be coalesced?
8648 // If the chunk is in a free range and either we decided to coalesce above
8649 // or the chunk is near the large block at the end of the heap
8650 // (isNearLargestChunk() returns true), then coalesce this chunk.
8651 const bool doCoalesce = inFreeRange()
8652 && (coalesce || _g->isNearLargestChunk(fc_addr));
8653 if (doCoalesce) {
8654 // Coalesce the current free range on the left with the new
8655 // chunk on the right. If either is on a free list,
8656 // it must be removed from the list and stashed in the closure.
8657 if (freeRangeInFreeLists()) {
8658 FreeChunk* const ffc = (FreeChunk*)freeFinger();
8659 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
8660 "Size of free range is inconsistent with chunk size.");
8661 if (CMSTestInFreeList) {
8662 assert(_sp->verify_chunk_in_free_list(ffc),
8663 "Chunk is not in free lists");
8664 }
8665 _sp->coalDeath(ffc->size());
8666 _sp->removeFreeChunkFromFreeLists(ffc);
8667 set_freeRangeInFreeLists(false);
8668 }
8669 if (fcInFreeLists) {
8670 _sp->coalDeath(chunkSize);
8671 assert(fc->size() == chunkSize,
8672 "The chunk has the wrong size or is not in the free lists");
8673 _sp->removeFreeChunkFromFreeLists(fc);
8674 }
8675 set_lastFreeRangeCoalesced(true);
8676 print_free_block_coalesced(fc);
8677 } else { // not in a free range and/or should not coalesce
8678 // Return the current free range and start a new one.
8679 if (inFreeRange()) {
8680 // In a free range but cannot coalesce with the right hand chunk.
8681 // Put the current free range into the free lists.
8682 flush_cur_free_chunk(freeFinger(),
8683 pointer_delta(fc_addr, freeFinger()));
8684 }
8685 // Set up for new free range. Pass along whether the right hand
8686 // chunk is in the free lists.
8687 initialize_free_range((HeapWord*)fc, fcInFreeLists);
8688 }
8689 }
8691 // Lookahead flush:
8692 // If we are tracking a free range, and this is the last chunk that
8693 // we'll look at because its end crosses past _limit, we'll preemptively
8694 // flush it along with any free range we may be holding on to. Note that
8695 // this can be the case only for an already free or freshly garbage
8696 // chunk. If this block is an object, it can never straddle
8697 // over _limit. The "straddling" occurs when _limit is set at
8698 // the previous end of the space when this cycle started, and
8699 // a subsequent heap expansion caused the previously co-terminal
8700 // free block to be coalesced with the newly expanded portion,
8701 // thus rendering _limit a non-block-boundary making it dangerous
8702 // for the sweeper to step over and examine.
8703 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
8704 assert(inFreeRange(), "Should only be called if currently in a free range.");
8705 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
8706 assert(_sp->used_region().contains(eob - 1),
8707 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
8708 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
8709 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
8710 eob, eob-1, _limit, _sp->bottom(), _sp->end(), fc, chunk_size));
8711 if (eob >= _limit) {
8712 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
8713 if (CMSTraceSweeper) {
8714 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
8715 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
8716 "[" PTR_FORMAT "," PTR_FORMAT ")",
8717 _limit, fc, eob, _sp->bottom(), _sp->end());
8718 }
8719 // Return the storage we are tracking back into the free lists.
8720 if (CMSTraceSweeper) {
8721 gclog_or_tty->print_cr("Flushing ... ");
8722 }
8723 assert(freeFinger() < eob, "Error");
8724 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
8725 }
8726 }
8728 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
8729 assert(inFreeRange(), "Should only be called if currently in a free range.");
8730 assert(size > 0,
8731 "A zero sized chunk cannot be added to the free lists.");
8732 if (!freeRangeInFreeLists()) {
8733 if (CMSTestInFreeList) {
8734 FreeChunk* fc = (FreeChunk*) chunk;
8735 fc->set_size(size);
8736 assert(!_sp->verify_chunk_in_free_list(fc),
8737 "chunk should not be in free lists yet");
8738 }
8739 if (CMSTraceSweeper) {
8740 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8741 chunk, size);
8742 }
8743 // A new free range is going to be starting. The current
8744 // free range has not been added to the free lists yet or
8745 // was removed so add it back.
8746 // If the current free range was coalesced, then the death
8747 // of the free range was recorded. Record a birth now.
8748 if (lastFreeRangeCoalesced()) {
8749 _sp->coalBirth(size);
8750 }
8751 _sp->addChunkAndRepairOffsetTable(chunk, size,
8752 lastFreeRangeCoalesced());
8753 } else if (CMSTraceSweeper) {
8754 gclog_or_tty->print_cr("Already in free list: nothing to flush");
8755 }
8756 set_inFreeRange(false);
8757 set_freeRangeInFreeLists(false);
8758 }
8760 // We take a break if we've been at this for a while,
8761 // so as to avoid monopolizing the locks involved.
8762 void SweepClosure::do_yield_work(HeapWord* addr) {
8763 // Return current free chunk being used for coalescing (if any)
8764 // to the appropriate freelist. After yielding, the next
8765 // free block encountered will start a coalescing range of
8766 // free blocks. If the next free block is adjacent to the
8767 // chunk just flushed, they will need to wait for the next
8768 // sweep to be coalesced.
8769 if (inFreeRange()) {
8770 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8771 }
8773 // First give up the locks, then yield, then re-lock.
8774 // We should probably use a constructor/destructor idiom to
8775 // do this unlock/lock or modify the MutexUnlocker class to
8776 // serve our purpose. XXX
8777 assert_lock_strong(_bitMap->lock());
8778 assert_lock_strong(_freelistLock);
8779 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8780 "CMS thread should hold CMS token");
8781 _bitMap->lock()->unlock();
8782 _freelistLock->unlock();
8783 ConcurrentMarkSweepThread::desynchronize(true);
8784 ConcurrentMarkSweepThread::acknowledge_yield_request();
8785 _collector->stopTimer();
8786 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8787 if (PrintCMSStatistics != 0) {
8788 _collector->incrementYields();
8789 }
8790 _collector->icms_wait();
8792 // See the comment in coordinator_yield()
8793 for (unsigned i = 0; i < CMSYieldSleepCount &&
8794 ConcurrentMarkSweepThread::should_yield() &&
8795 !CMSCollector::foregroundGCIsActive(); ++i) {
8796 os::sleep(Thread::current(), 1, false);
8797 ConcurrentMarkSweepThread::acknowledge_yield_request();
8798 }
8800 ConcurrentMarkSweepThread::synchronize(true);
8801 _freelistLock->lock();
8802 _bitMap->lock()->lock_without_safepoint_check();
8803 _collector->startTimer();
8804 }
8806 #ifndef PRODUCT
8807 // This is actually very useful in a product build if it can
8808 // be called from the debugger. Compile it into the product
8809 // as needed.
8810 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
8811 return debug_cms_space->verify_chunk_in_free_list(fc);
8812 }
8813 #endif
8815 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
8816 if (CMSTraceSweeper) {
8817 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
8818 fc, fc->size());
8819 }
8820 }
8822 // CMSIsAliveClosure
8823 bool CMSIsAliveClosure::do_object_b(oop obj) {
8824 HeapWord* addr = (HeapWord*)obj;
8825 return addr != NULL &&
8826 (!_span.contains(addr) || _bit_map->isMarked(addr));
8827 }
8830 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
8831 MemRegion span,
8832 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
8833 bool cpc):
8834 _collector(collector),
8835 _span(span),
8836 _bit_map(bit_map),
8837 _mark_stack(mark_stack),
8838 _concurrent_precleaning(cpc) {
8839 assert(!_span.is_empty(), "Empty span could spell trouble");
8840 }
8843 // CMSKeepAliveClosure: the serial version
8844 void CMSKeepAliveClosure::do_oop(oop obj) {
8845 HeapWord* addr = (HeapWord*)obj;
8846 if (_span.contains(addr) &&
8847 !_bit_map->isMarked(addr)) {
8848 _bit_map->mark(addr);
8849 bool simulate_overflow = false;
8850 NOT_PRODUCT(
8851 if (CMSMarkStackOverflowALot &&
8852 _collector->simulate_overflow()) {
8853 // simulate a stack overflow
8854 simulate_overflow = true;
8855 }
8856 )
8857 if (simulate_overflow || !_mark_stack->push(obj)) {
8858 if (_concurrent_precleaning) {
8859 // We dirty the overflown object and let the remark
8860 // phase deal with it.
8861 assert(_collector->overflow_list_is_empty(), "Error");
8862 // In the case of object arrays, we need to dirty all of
8863 // the cards that the object spans. No locking or atomics
8864 // are needed since no one else can be mutating the mod union
8865 // table.
8866 if (obj->is_objArray()) {
8867 size_t sz = obj->size();
8868 HeapWord* end_card_addr =
8869 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
8870 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8871 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8872 _collector->_modUnionTable.mark_range(redirty_range);
8873 } else {
8874 _collector->_modUnionTable.mark(addr);
8875 }
8876 _collector->_ser_kac_preclean_ovflw++;
8877 } else {
8878 _collector->push_on_overflow_list(obj);
8879 _collector->_ser_kac_ovflw++;
8880 }
8881 }
8882 }
8883 }
8885 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8886 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8888 // CMSParKeepAliveClosure: a parallel version of the above.
8889 // The work queues are private to each closure (thread),
8890 // but (may be) available for stealing by other threads.
8891 void CMSParKeepAliveClosure::do_oop(oop obj) {
8892 HeapWord* addr = (HeapWord*)obj;
8893 if (_span.contains(addr) &&
8894 !_bit_map->isMarked(addr)) {
8895 // In general, during recursive tracing, several threads
8896 // may be concurrently getting here; the first one to
8897 // "tag" it, claims it.
8898 if (_bit_map->par_mark(addr)) {
8899 bool res = _work_queue->push(obj);
8900 assert(res, "Low water mark should be much less than capacity");
8901 // Do a recursive trim in the hope that this will keep
8902 // stack usage lower, but leave some oops for potential stealers
8903 trim_queue(_low_water_mark);
8904 } // Else, another thread got there first
8905 }
8906 }
8908 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8909 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8911 void CMSParKeepAliveClosure::trim_queue(uint max) {
8912 while (_work_queue->size() > max) {
8913 oop new_oop;
8914 if (_work_queue->pop_local(new_oop)) {
8915 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8916 assert(_bit_map->isMarked((HeapWord*)new_oop),
8917 "no white objects on this stack!");
8918 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8919 // iterate over the oops in this oop, marking and pushing
8920 // the ones in CMS heap (i.e. in _span).
8921 new_oop->oop_iterate(&_mark_and_push);
8922 }
8923 }
8924 }
8926 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8927 CMSCollector* collector,
8928 MemRegion span, CMSBitMap* bit_map,
8929 OopTaskQueue* work_queue):
8930 _collector(collector),
8931 _span(span),
8932 _bit_map(bit_map),
8933 _work_queue(work_queue) { }
8935 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8936 HeapWord* addr = (HeapWord*)obj;
8937 if (_span.contains(addr) &&
8938 !_bit_map->isMarked(addr)) {
8939 if (_bit_map->par_mark(addr)) {
8940 bool simulate_overflow = false;
8941 NOT_PRODUCT(
8942 if (CMSMarkStackOverflowALot &&
8943 _collector->par_simulate_overflow()) {
8944 // simulate a stack overflow
8945 simulate_overflow = true;
8946 }
8947 )
8948 if (simulate_overflow || !_work_queue->push(obj)) {
8949 _collector->par_push_on_overflow_list(obj);
8950 _collector->_par_kac_ovflw++;
8951 }
8952 } // Else another thread got there already
8953 }
8954 }
8956 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8957 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8959 //////////////////////////////////////////////////////////////////
8960 // CMSExpansionCause /////////////////////////////
8961 //////////////////////////////////////////////////////////////////
8962 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8963 switch (cause) {
8964 case _no_expansion:
8965 return "No expansion";
8966 case _satisfy_free_ratio:
8967 return "Free ratio";
8968 case _satisfy_promotion:
8969 return "Satisfy promotion";
8970 case _satisfy_allocation:
8971 return "allocation";
8972 case _allocate_par_lab:
8973 return "Par LAB";
8974 case _allocate_par_spooling_space:
8975 return "Par Spooling Space";
8976 case _adaptive_size_policy:
8977 return "Ergonomics";
8978 default:
8979 return "unknown";
8980 }
8981 }
8983 void CMSDrainMarkingStackClosure::do_void() {
8984 // the max number to take from overflow list at a time
8985 const size_t num = _mark_stack->capacity()/4;
8986 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8987 "Overflow list should be NULL during concurrent phases");
8988 while (!_mark_stack->isEmpty() ||
8989 // if stack is empty, check the overflow list
8990 _collector->take_from_overflow_list(num, _mark_stack)) {
8991 oop obj = _mark_stack->pop();
8992 HeapWord* addr = (HeapWord*)obj;
8993 assert(_span.contains(addr), "Should be within span");
8994 assert(_bit_map->isMarked(addr), "Should be marked");
8995 assert(obj->is_oop(), "Should be an oop");
8996 obj->oop_iterate(_keep_alive);
8997 }
8998 }
9000 void CMSParDrainMarkingStackClosure::do_void() {
9001 // drain queue
9002 trim_queue(0);
9003 }
9005 // Trim our work_queue so its length is below max at return
9006 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
9007 while (_work_queue->size() > max) {
9008 oop new_oop;
9009 if (_work_queue->pop_local(new_oop)) {
9010 assert(new_oop->is_oop(), "Expected an oop");
9011 assert(_bit_map->isMarked((HeapWord*)new_oop),
9012 "no white objects on this stack!");
9013 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
9014 // iterate over the oops in this oop, marking and pushing
9015 // the ones in CMS heap (i.e. in _span).
9016 new_oop->oop_iterate(&_mark_and_push);
9017 }
9018 }
9019 }
9021 ////////////////////////////////////////////////////////////////////
9022 // Support for Marking Stack Overflow list handling and related code
9023 ////////////////////////////////////////////////////////////////////
9024 // Much of the following code is similar in shape and spirit to the
9025 // code used in ParNewGC. We should try and share that code
9026 // as much as possible in the future.
9028 #ifndef PRODUCT
9029 // Debugging support for CMSStackOverflowALot
9031 // It's OK to call this multi-threaded; the worst thing
9032 // that can happen is that we'll get a bunch of closely
9033 // spaced simulated oveflows, but that's OK, in fact
9034 // probably good as it would exercise the overflow code
9035 // under contention.
9036 bool CMSCollector::simulate_overflow() {
9037 if (_overflow_counter-- <= 0) { // just being defensive
9038 _overflow_counter = CMSMarkStackOverflowInterval;
9039 return true;
9040 } else {
9041 return false;
9042 }
9043 }
9045 bool CMSCollector::par_simulate_overflow() {
9046 return simulate_overflow();
9047 }
9048 #endif
9050 // Single-threaded
9051 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
9052 assert(stack->isEmpty(), "Expected precondition");
9053 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
9054 size_t i = num;
9055 oop cur = _overflow_list;
9056 const markOop proto = markOopDesc::prototype();
9057 NOT_PRODUCT(ssize_t n = 0;)
9058 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
9059 next = oop(cur->mark());
9060 cur->set_mark(proto); // until proven otherwise
9061 assert(cur->is_oop(), "Should be an oop");
9062 bool res = stack->push(cur);
9063 assert(res, "Bit off more than can chew?");
9064 NOT_PRODUCT(n++;)
9065 }
9066 _overflow_list = cur;
9067 #ifndef PRODUCT
9068 assert(_num_par_pushes >= n, "Too many pops?");
9069 _num_par_pushes -=n;
9070 #endif
9071 return !stack->isEmpty();
9072 }
9074 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
9075 // (MT-safe) Get a prefix of at most "num" from the list.
9076 // The overflow list is chained through the mark word of
9077 // each object in the list. We fetch the entire list,
9078 // break off a prefix of the right size and return the
9079 // remainder. If other threads try to take objects from
9080 // the overflow list at that time, they will wait for
9081 // some time to see if data becomes available. If (and
9082 // only if) another thread places one or more object(s)
9083 // on the global list before we have returned the suffix
9084 // to the global list, we will walk down our local list
9085 // to find its end and append the global list to
9086 // our suffix before returning it. This suffix walk can
9087 // prove to be expensive (quadratic in the amount of traffic)
9088 // when there are many objects in the overflow list and
9089 // there is much producer-consumer contention on the list.
9090 // *NOTE*: The overflow list manipulation code here and
9091 // in ParNewGeneration:: are very similar in shape,
9092 // except that in the ParNew case we use the old (from/eden)
9093 // copy of the object to thread the list via its klass word.
9094 // Because of the common code, if you make any changes in
9095 // the code below, please check the ParNew version to see if
9096 // similar changes might be needed.
9097 // CR 6797058 has been filed to consolidate the common code.
9098 bool CMSCollector::par_take_from_overflow_list(size_t num,
9099 OopTaskQueue* work_q,
9100 int no_of_gc_threads) {
9101 assert(work_q->size() == 0, "First empty local work queue");
9102 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
9103 if (_overflow_list == NULL) {
9104 return false;
9105 }
9106 // Grab the entire list; we'll put back a suffix
9107 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
9108 Thread* tid = Thread::current();
9109 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
9110 // set to ParallelGCThreads.
9111 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
9112 size_t sleep_time_millis = MAX2((size_t)1, num/100);
9113 // If the list is busy, we spin for a short while,
9114 // sleeping between attempts to get the list.
9115 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
9116 os::sleep(tid, sleep_time_millis, false);
9117 if (_overflow_list == NULL) {
9118 // Nothing left to take
9119 return false;
9120 } else if (_overflow_list != BUSY) {
9121 // Try and grab the prefix
9122 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
9123 }
9124 }
9125 // If the list was found to be empty, or we spun long
9126 // enough, we give up and return empty-handed. If we leave
9127 // the list in the BUSY state below, it must be the case that
9128 // some other thread holds the overflow list and will set it
9129 // to a non-BUSY state in the future.
9130 if (prefix == NULL || prefix == BUSY) {
9131 // Nothing to take or waited long enough
9132 if (prefix == NULL) {
9133 // Write back the NULL in case we overwrote it with BUSY above
9134 // and it is still the same value.
9135 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
9136 }
9137 return false;
9138 }
9139 assert(prefix != NULL && prefix != BUSY, "Error");
9140 size_t i = num;
9141 oop cur = prefix;
9142 // Walk down the first "num" objects, unless we reach the end.
9143 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
9144 if (cur->mark() == NULL) {
9145 // We have "num" or fewer elements in the list, so there
9146 // is nothing to return to the global list.
9147 // Write back the NULL in lieu of the BUSY we wrote
9148 // above, if it is still the same value.
9149 if (_overflow_list == BUSY) {
9150 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
9151 }
9152 } else {
9153 // Chop off the suffix and rerturn it to the global list.
9154 assert(cur->mark() != BUSY, "Error");
9155 oop suffix_head = cur->mark(); // suffix will be put back on global list
9156 cur->set_mark(NULL); // break off suffix
9157 // It's possible that the list is still in the empty(busy) state
9158 // we left it in a short while ago; in that case we may be
9159 // able to place back the suffix without incurring the cost
9160 // of a walk down the list.
9161 oop observed_overflow_list = _overflow_list;
9162 oop cur_overflow_list = observed_overflow_list;
9163 bool attached = false;
9164 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
9165 observed_overflow_list =
9166 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
9167 if (cur_overflow_list == observed_overflow_list) {
9168 attached = true;
9169 break;
9170 } else cur_overflow_list = observed_overflow_list;
9171 }
9172 if (!attached) {
9173 // Too bad, someone else sneaked in (at least) an element; we'll need
9174 // to do a splice. Find tail of suffix so we can prepend suffix to global
9175 // list.
9176 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
9177 oop suffix_tail = cur;
9178 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
9179 "Tautology");
9180 observed_overflow_list = _overflow_list;
9181 do {
9182 cur_overflow_list = observed_overflow_list;
9183 if (cur_overflow_list != BUSY) {
9184 // Do the splice ...
9185 suffix_tail->set_mark(markOop(cur_overflow_list));
9186 } else { // cur_overflow_list == BUSY
9187 suffix_tail->set_mark(NULL);
9188 }
9189 // ... and try to place spliced list back on overflow_list ...
9190 observed_overflow_list =
9191 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
9192 } while (cur_overflow_list != observed_overflow_list);
9193 // ... until we have succeeded in doing so.
9194 }
9195 }
9197 // Push the prefix elements on work_q
9198 assert(prefix != NULL, "control point invariant");
9199 const markOop proto = markOopDesc::prototype();
9200 oop next;
9201 NOT_PRODUCT(ssize_t n = 0;)
9202 for (cur = prefix; cur != NULL; cur = next) {
9203 next = oop(cur->mark());
9204 cur->set_mark(proto); // until proven otherwise
9205 assert(cur->is_oop(), "Should be an oop");
9206 bool res = work_q->push(cur);
9207 assert(res, "Bit off more than we can chew?");
9208 NOT_PRODUCT(n++;)
9209 }
9210 #ifndef PRODUCT
9211 assert(_num_par_pushes >= n, "Too many pops?");
9212 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
9213 #endif
9214 return true;
9215 }
9217 // Single-threaded
9218 void CMSCollector::push_on_overflow_list(oop p) {
9219 NOT_PRODUCT(_num_par_pushes++;)
9220 assert(p->is_oop(), "Not an oop");
9221 preserve_mark_if_necessary(p);
9222 p->set_mark((markOop)_overflow_list);
9223 _overflow_list = p;
9224 }
9226 // Multi-threaded; use CAS to prepend to overflow list
9227 void CMSCollector::par_push_on_overflow_list(oop p) {
9228 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
9229 assert(p->is_oop(), "Not an oop");
9230 par_preserve_mark_if_necessary(p);
9231 oop observed_overflow_list = _overflow_list;
9232 oop cur_overflow_list;
9233 do {
9234 cur_overflow_list = observed_overflow_list;
9235 if (cur_overflow_list != BUSY) {
9236 p->set_mark(markOop(cur_overflow_list));
9237 } else {
9238 p->set_mark(NULL);
9239 }
9240 observed_overflow_list =
9241 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
9242 } while (cur_overflow_list != observed_overflow_list);
9243 }
9244 #undef BUSY
9246 // Single threaded
9247 // General Note on GrowableArray: pushes may silently fail
9248 // because we are (temporarily) out of C-heap for expanding
9249 // the stack. The problem is quite ubiquitous and affects
9250 // a lot of code in the JVM. The prudent thing for GrowableArray
9251 // to do (for now) is to exit with an error. However, that may
9252 // be too draconian in some cases because the caller may be
9253 // able to recover without much harm. For such cases, we
9254 // should probably introduce a "soft_push" method which returns
9255 // an indication of success or failure with the assumption that
9256 // the caller may be able to recover from a failure; code in
9257 // the VM can then be changed, incrementally, to deal with such
9258 // failures where possible, thus, incrementally hardening the VM
9259 // in such low resource situations.
9260 void CMSCollector::preserve_mark_work(oop p, markOop m) {
9261 _preserved_oop_stack.push(p);
9262 _preserved_mark_stack.push(m);
9263 assert(m == p->mark(), "Mark word changed");
9264 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9265 "bijection");
9266 }
9268 // Single threaded
9269 void CMSCollector::preserve_mark_if_necessary(oop p) {
9270 markOop m = p->mark();
9271 if (m->must_be_preserved(p)) {
9272 preserve_mark_work(p, m);
9273 }
9274 }
9276 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
9277 markOop m = p->mark();
9278 if (m->must_be_preserved(p)) {
9279 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
9280 // Even though we read the mark word without holding
9281 // the lock, we are assured that it will not change
9282 // because we "own" this oop, so no other thread can
9283 // be trying to push it on the overflow list; see
9284 // the assertion in preserve_mark_work() that checks
9285 // that m == p->mark().
9286 preserve_mark_work(p, m);
9287 }
9288 }
9290 // We should be able to do this multi-threaded,
9291 // a chunk of stack being a task (this is
9292 // correct because each oop only ever appears
9293 // once in the overflow list. However, it's
9294 // not very easy to completely overlap this with
9295 // other operations, so will generally not be done
9296 // until all work's been completed. Because we
9297 // expect the preserved oop stack (set) to be small,
9298 // it's probably fine to do this single-threaded.
9299 // We can explore cleverer concurrent/overlapped/parallel
9300 // processing of preserved marks if we feel the
9301 // need for this in the future. Stack overflow should
9302 // be so rare in practice and, when it happens, its
9303 // effect on performance so great that this will
9304 // likely just be in the noise anyway.
9305 void CMSCollector::restore_preserved_marks_if_any() {
9306 assert(SafepointSynchronize::is_at_safepoint(),
9307 "world should be stopped");
9308 assert(Thread::current()->is_ConcurrentGC_thread() ||
9309 Thread::current()->is_VM_thread(),
9310 "should be single-threaded");
9311 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9312 "bijection");
9314 while (!_preserved_oop_stack.is_empty()) {
9315 oop p = _preserved_oop_stack.pop();
9316 assert(p->is_oop(), "Should be an oop");
9317 assert(_span.contains(p), "oop should be in _span");
9318 assert(p->mark() == markOopDesc::prototype(),
9319 "Set when taken from overflow list");
9320 markOop m = _preserved_mark_stack.pop();
9321 p->set_mark(m);
9322 }
9323 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
9324 "stacks were cleared above");
9325 }
9327 #ifndef PRODUCT
9328 bool CMSCollector::no_preserved_marks() const {
9329 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
9330 }
9331 #endif
9333 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
9334 {
9335 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
9336 CMSAdaptiveSizePolicy* size_policy =
9337 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
9338 assert(size_policy->is_gc_cms_adaptive_size_policy(),
9339 "Wrong type for size policy");
9340 return size_policy;
9341 }
9343 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
9344 size_t desired_promo_size) {
9345 if (cur_promo_size < desired_promo_size) {
9346 size_t expand_bytes = desired_promo_size - cur_promo_size;
9347 if (PrintAdaptiveSizePolicy && Verbose) {
9348 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9349 "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
9350 expand_bytes);
9351 }
9352 expand(expand_bytes,
9353 MinHeapDeltaBytes,
9354 CMSExpansionCause::_adaptive_size_policy);
9355 } else if (desired_promo_size < cur_promo_size) {
9356 size_t shrink_bytes = cur_promo_size - desired_promo_size;
9357 if (PrintAdaptiveSizePolicy && Verbose) {
9358 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9359 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
9360 shrink_bytes);
9361 }
9362 shrink(shrink_bytes);
9363 }
9364 }
9366 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
9367 GenCollectedHeap* gch = GenCollectedHeap::heap();
9368 CMSGCAdaptivePolicyCounters* counters =
9369 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
9370 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
9371 "Wrong kind of counters");
9372 return counters;
9373 }
9376 void ASConcurrentMarkSweepGeneration::update_counters() {
9377 if (UsePerfData) {
9378 _space_counters->update_all();
9379 _gen_counters->update_all();
9380 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9381 GenCollectedHeap* gch = GenCollectedHeap::heap();
9382 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9383 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9384 "Wrong gc statistics type");
9385 counters->update_counters(gc_stats_l);
9386 }
9387 }
9389 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
9390 if (UsePerfData) {
9391 _space_counters->update_used(used);
9392 _space_counters->update_capacity();
9393 _gen_counters->update_all();
9395 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9396 GenCollectedHeap* gch = GenCollectedHeap::heap();
9397 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9398 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9399 "Wrong gc statistics type");
9400 counters->update_counters(gc_stats_l);
9401 }
9402 }
9404 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
9405 assert_locked_or_safepoint(Heap_lock);
9406 assert_lock_strong(freelistLock());
9407 HeapWord* old_end = _cmsSpace->end();
9408 HeapWord* unallocated_start = _cmsSpace->unallocated_block();
9409 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
9410 FreeChunk* chunk_at_end = find_chunk_at_end();
9411 if (chunk_at_end == NULL) {
9412 // No room to shrink
9413 if (PrintGCDetails && Verbose) {
9414 gclog_or_tty->print_cr("No room to shrink: old_end "
9415 PTR_FORMAT " unallocated_start " PTR_FORMAT
9416 " chunk_at_end " PTR_FORMAT,
9417 old_end, unallocated_start, chunk_at_end);
9418 }
9419 return;
9420 } else {
9422 // Find the chunk at the end of the space and determine
9423 // how much it can be shrunk.
9424 size_t shrinkable_size_in_bytes = chunk_at_end->size();
9425 size_t aligned_shrinkable_size_in_bytes =
9426 align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
9427 assert(unallocated_start <= (HeapWord*) chunk_at_end->end(),
9428 "Inconsistent chunk at end of space");
9429 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
9430 size_t word_size_before = heap_word_size(_virtual_space.committed_size());
9432 // Shrink the underlying space
9433 _virtual_space.shrink_by(bytes);
9434 if (PrintGCDetails && Verbose) {
9435 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
9436 " desired_bytes " SIZE_FORMAT
9437 " shrinkable_size_in_bytes " SIZE_FORMAT
9438 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
9439 " bytes " SIZE_FORMAT,
9440 desired_bytes, shrinkable_size_in_bytes,
9441 aligned_shrinkable_size_in_bytes, bytes);
9442 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT
9443 " unallocated_start " SIZE_FORMAT,
9444 old_end, unallocated_start);
9445 }
9447 // If the space did shrink (shrinking is not guaranteed),
9448 // shrink the chunk at the end by the appropriate amount.
9449 if (((HeapWord*)_virtual_space.high()) < old_end) {
9450 size_t new_word_size =
9451 heap_word_size(_virtual_space.committed_size());
9453 // Have to remove the chunk from the dictionary because it is changing
9454 // size and might be someplace elsewhere in the dictionary.
9456 // Get the chunk at end, shrink it, and put it
9457 // back.
9458 _cmsSpace->removeChunkFromDictionary(chunk_at_end);
9459 size_t word_size_change = word_size_before - new_word_size;
9460 size_t chunk_at_end_old_size = chunk_at_end->size();
9461 assert(chunk_at_end_old_size >= word_size_change,
9462 "Shrink is too large");
9463 chunk_at_end->set_size(chunk_at_end_old_size -
9464 word_size_change);
9465 _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
9466 word_size_change);
9468 _cmsSpace->returnChunkToDictionary(chunk_at_end);
9470 MemRegion mr(_cmsSpace->bottom(), new_word_size);
9471 _bts->resize(new_word_size); // resize the block offset shared array
9472 Universe::heap()->barrier_set()->resize_covered_region(mr);
9473 _cmsSpace->assert_locked();
9474 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
9476 NOT_PRODUCT(_cmsSpace->dictionary()->verify());
9478 // update the space and generation capacity counters
9479 if (UsePerfData) {
9480 _space_counters->update_capacity();
9481 _gen_counters->update_all();
9482 }
9484 if (Verbose && PrintGCDetails) {
9485 size_t new_mem_size = _virtual_space.committed_size();
9486 size_t old_mem_size = new_mem_size + bytes;
9487 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K",
9488 name(), old_mem_size/K, bytes/K, new_mem_size/K);
9489 }
9490 }
9492 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
9493 "Inconsistency at end of space");
9494 assert(chunk_at_end->end() == (uintptr_t*) _cmsSpace->end(),
9495 "Shrinking is inconsistent");
9496 return;
9497 }
9498 }
9499 // Transfer some number of overflown objects to usual marking
9500 // stack. Return true if some objects were transferred.
9501 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
9502 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
9503 (size_t)ParGCDesiredObjsFromOverflowList);
9505 bool res = _collector->take_from_overflow_list(num, _mark_stack);
9506 assert(_collector->overflow_list_is_empty() || res,
9507 "If list is not empty, we should have taken something");
9508 assert(!res || !_mark_stack->isEmpty(),
9509 "If we took something, it should now be on our stack");
9510 return res;
9511 }
9513 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
9514 size_t res = _sp->block_size_no_stall(addr, _collector);
9515 if (_sp->block_is_obj(addr)) {
9516 if (_live_bit_map->isMarked(addr)) {
9517 // It can't have been dead in a previous cycle
9518 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
9519 } else {
9520 _dead_bit_map->mark(addr); // mark the dead object
9521 }
9522 }
9523 // Could be 0, if the block size could not be computed without stalling.
9524 return res;
9525 }
9527 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
9529 switch (phase) {
9530 case CMSCollector::InitialMarking:
9531 initialize(true /* fullGC */ ,
9532 cause /* cause of the GC */,
9533 true /* recordGCBeginTime */,
9534 true /* recordPreGCUsage */,
9535 false /* recordPeakUsage */,
9536 false /* recordPostGCusage */,
9537 true /* recordAccumulatedGCTime */,
9538 false /* recordGCEndTime */,
9539 false /* countCollection */ );
9540 break;
9542 case CMSCollector::FinalMarking:
9543 initialize(true /* fullGC */ ,
9544 cause /* cause of the GC */,
9545 false /* recordGCBeginTime */,
9546 false /* recordPreGCUsage */,
9547 false /* recordPeakUsage */,
9548 false /* recordPostGCusage */,
9549 true /* recordAccumulatedGCTime */,
9550 false /* recordGCEndTime */,
9551 false /* countCollection */ );
9552 break;
9554 case CMSCollector::Sweeping:
9555 initialize(true /* fullGC */ ,
9556 cause /* cause of the GC */,
9557 false /* recordGCBeginTime */,
9558 false /* recordPreGCUsage */,
9559 true /* recordPeakUsage */,
9560 true /* recordPostGCusage */,
9561 false /* recordAccumulatedGCTime */,
9562 true /* recordGCEndTime */,
9563 true /* countCollection */ );
9564 break;
9566 default:
9567 ShouldNotReachHere();
9568 }
9569 }