Print this page
| Split |
Close |
| Expand all |
| Collapse all |
--- old/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.cpp
+++ new/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.cpp
1 1 /*
2 2 * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 20 * or visit www.oracle.com if you need additional information or have any
21 21 * questions.
22 22 *
23 23 */
24 24
25 25 #include "precompiled.hpp"
26 26 #include "classfile/symbolTable.hpp"
27 27 #include "classfile/systemDictionary.hpp"
28 28 #include "code/codeCache.hpp"
29 29 #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp"
30 30 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp"
31 31 #include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.hpp"
32 32 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp"
33 33 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp"
34 34 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp"
35 35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp"
36 36 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp"
37 37 #include "gc_implementation/parNew/parNewGeneration.hpp"
38 38 #include "gc_implementation/shared/collectorCounters.hpp"
39 39 #include "gc_implementation/shared/isGCActiveMark.hpp"
40 40 #include "gc_interface/collectedHeap.inline.hpp"
41 41 #include "memory/cardTableRS.hpp"
42 42 #include "memory/collectorPolicy.hpp"
43 43 #include "memory/gcLocker.inline.hpp"
44 44 #include "memory/genCollectedHeap.hpp"
45 45 #include "memory/genMarkSweep.hpp"
46 46 #include "memory/genOopClosures.inline.hpp"
47 47 #include "memory/iterator.hpp"
48 48 #include "memory/referencePolicy.hpp"
49 49 #include "memory/resourceArea.hpp"
50 50 #include "oops/oop.inline.hpp"
51 51 #include "prims/jvmtiExport.hpp"
52 52 #include "runtime/globals_extension.hpp"
53 53 #include "runtime/handles.inline.hpp"
54 54 #include "runtime/java.hpp"
55 55 #include "runtime/vmThread.hpp"
56 56 #include "services/memoryService.hpp"
57 57 #include "services/runtimeService.hpp"
58 58
59 59 // statics
60 60 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
61 61 bool CMSCollector::_full_gc_requested = false;
62 62
63 63 //////////////////////////////////////////////////////////////////
64 64 // In support of CMS/VM thread synchronization
65 65 //////////////////////////////////////////////////////////////////
66 66 // We split use of the CGC_lock into 2 "levels".
67 67 // The low-level locking is of the usual CGC_lock monitor. We introduce
68 68 // a higher level "token" (hereafter "CMS token") built on top of the
69 69 // low level monitor (hereafter "CGC lock").
70 70 // The token-passing protocol gives priority to the VM thread. The
71 71 // CMS-lock doesn't provide any fairness guarantees, but clients
72 72 // should ensure that it is only held for very short, bounded
73 73 // durations.
74 74 //
75 75 // When either of the CMS thread or the VM thread is involved in
76 76 // collection operations during which it does not want the other
77 77 // thread to interfere, it obtains the CMS token.
78 78 //
79 79 // If either thread tries to get the token while the other has
80 80 // it, that thread waits. However, if the VM thread and CMS thread
81 81 // both want the token, then the VM thread gets priority while the
82 82 // CMS thread waits. This ensures, for instance, that the "concurrent"
83 83 // phases of the CMS thread's work do not block out the VM thread
84 84 // for long periods of time as the CMS thread continues to hog
85 85 // the token. (See bug 4616232).
86 86 //
87 87 // The baton-passing functions are, however, controlled by the
88 88 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
89 89 // and here the low-level CMS lock, not the high level token,
90 90 // ensures mutual exclusion.
91 91 //
92 92 // Two important conditions that we have to satisfy:
93 93 // 1. if a thread does a low-level wait on the CMS lock, then it
94 94 // relinquishes the CMS token if it were holding that token
95 95 // when it acquired the low-level CMS lock.
96 96 // 2. any low-level notifications on the low-level lock
97 97 // should only be sent when a thread has relinquished the token.
98 98 //
99 99 // In the absence of either property, we'd have potential deadlock.
100 100 //
101 101 // We protect each of the CMS (concurrent and sequential) phases
102 102 // with the CMS _token_, not the CMS _lock_.
103 103 //
104 104 // The only code protected by CMS lock is the token acquisition code
105 105 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
106 106 // baton-passing code.
107 107 //
108 108 // Unfortunately, i couldn't come up with a good abstraction to factor and
109 109 // hide the naked CGC_lock manipulation in the baton-passing code
110 110 // further below. That's something we should try to do. Also, the proof
111 111 // of correctness of this 2-level locking scheme is far from obvious,
112 112 // and potentially quite slippery. We have an uneasy supsicion, for instance,
113 113 // that there may be a theoretical possibility of delay/starvation in the
114 114 // low-level lock/wait/notify scheme used for the baton-passing because of
115 115 // potential intereference with the priority scheme embodied in the
116 116 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
117 117 // invocation further below and marked with "XXX 20011219YSR".
118 118 // Indeed, as we note elsewhere, this may become yet more slippery
119 119 // in the presence of multiple CMS and/or multiple VM threads. XXX
120 120
121 121 class CMSTokenSync: public StackObj {
122 122 private:
123 123 bool _is_cms_thread;
124 124 public:
125 125 CMSTokenSync(bool is_cms_thread):
126 126 _is_cms_thread(is_cms_thread) {
127 127 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
128 128 "Incorrect argument to constructor");
129 129 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
130 130 }
131 131
132 132 ~CMSTokenSync() {
133 133 assert(_is_cms_thread ?
134 134 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
135 135 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
136 136 "Incorrect state");
137 137 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
138 138 }
139 139 };
140 140
141 141 // Convenience class that does a CMSTokenSync, and then acquires
142 142 // upto three locks.
143 143 class CMSTokenSyncWithLocks: public CMSTokenSync {
144 144 private:
145 145 // Note: locks are acquired in textual declaration order
146 146 // and released in the opposite order
147 147 MutexLockerEx _locker1, _locker2, _locker3;
148 148 public:
149 149 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
150 150 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
151 151 CMSTokenSync(is_cms_thread),
152 152 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
153 153 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
154 154 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
155 155 { }
156 156 };
157 157
158 158
159 159 // Wrapper class to temporarily disable icms during a foreground cms collection.
160 160 class ICMSDisabler: public StackObj {
161 161 public:
162 162 // The ctor disables icms and wakes up the thread so it notices the change;
163 163 // the dtor re-enables icms. Note that the CMSCollector methods will check
164 164 // CMSIncrementalMode.
165 165 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); }
166 166 ~ICMSDisabler() { CMSCollector::enable_icms(); }
167 167 };
168 168
169 169 //////////////////////////////////////////////////////////////////
170 170 // Concurrent Mark-Sweep Generation /////////////////////////////
171 171 //////////////////////////////////////////////////////////////////
172 172
173 173 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
174 174
175 175 // This struct contains per-thread things necessary to support parallel
176 176 // young-gen collection.
177 177 class CMSParGCThreadState: public CHeapObj {
178 178 public:
179 179 CFLS_LAB lab;
180 180 PromotionInfo promo;
181 181
182 182 // Constructor.
183 183 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
184 184 promo.setSpace(cfls);
185 185 }
186 186 };
187 187
188 188 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
189 189 ReservedSpace rs, size_t initial_byte_size, int level,
190 190 CardTableRS* ct, bool use_adaptive_freelists,
191 191 FreeBlockDictionary::DictionaryChoice dictionaryChoice) :
192 192 CardGeneration(rs, initial_byte_size, level, ct),
193 193 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
194 194 _debug_collection_type(Concurrent_collection_type)
195 195 {
196 196 HeapWord* bottom = (HeapWord*) _virtual_space.low();
197 197 HeapWord* end = (HeapWord*) _virtual_space.high();
198 198
199 199 _direct_allocated_words = 0;
200 200 NOT_PRODUCT(
201 201 _numObjectsPromoted = 0;
202 202 _numWordsPromoted = 0;
203 203 _numObjectsAllocated = 0;
204 204 _numWordsAllocated = 0;
205 205 )
206 206
207 207 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
208 208 use_adaptive_freelists,
209 209 dictionaryChoice);
210 210 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
211 211 if (_cmsSpace == NULL) {
212 212 vm_exit_during_initialization(
213 213 "CompactibleFreeListSpace allocation failure");
214 214 }
215 215 _cmsSpace->_gen = this;
216 216
217 217 _gc_stats = new CMSGCStats();
218 218
219 219 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
220 220 // offsets match. The ability to tell free chunks from objects
221 221 // depends on this property.
222 222 debug_only(
223 223 FreeChunk* junk = NULL;
224 224 assert(UseCompressedOops ||
225 225 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
226 226 "Offset of FreeChunk::_prev within FreeChunk must match"
227 227 " that of OopDesc::_klass within OopDesc");
228 228 )
229 229 if (CollectedHeap::use_parallel_gc_threads()) {
230 230 typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
231 231 _par_gc_thread_states =
232 232 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads);
233 233 if (_par_gc_thread_states == NULL) {
234 234 vm_exit_during_initialization("Could not allocate par gc structs");
235 235 }
236 236 for (uint i = 0; i < ParallelGCThreads; i++) {
237 237 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
238 238 if (_par_gc_thread_states[i] == NULL) {
239 239 vm_exit_during_initialization("Could not allocate par gc structs");
240 240 }
241 241 }
242 242 } else {
243 243 _par_gc_thread_states = NULL;
244 244 }
245 245 _incremental_collection_failed = false;
246 246 // The "dilatation_factor" is the expansion that can occur on
247 247 // account of the fact that the minimum object size in the CMS
248 248 // generation may be larger than that in, say, a contiguous young
249 249 // generation.
250 250 // Ideally, in the calculation below, we'd compute the dilatation
251 251 // factor as: MinChunkSize/(promoting_gen's min object size)
252 252 // Since we do not have such a general query interface for the
253 253 // promoting generation, we'll instead just use the mimimum
254 254 // object size (which today is a header's worth of space);
255 255 // note that all arithmetic is in units of HeapWords.
256 256 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
257 257 assert(_dilatation_factor >= 1.0, "from previous assert");
258 258 }
259 259
260 260
261 261 // The field "_initiating_occupancy" represents the occupancy percentage
262 262 // at which we trigger a new collection cycle. Unless explicitly specified
263 263 // via CMSInitiating[Perm]OccupancyFraction (argument "io" below), it
264 264 // is calculated by:
265 265 //
266 266 // Let "f" be MinHeapFreeRatio in
267 267 //
268 268 // _intiating_occupancy = 100-f +
269 269 // f * (CMSTrigger[Perm]Ratio/100)
270 270 // where CMSTrigger[Perm]Ratio is the argument "tr" below.
271 271 //
272 272 // That is, if we assume the heap is at its desired maximum occupancy at the
273 273 // end of a collection, we let CMSTrigger[Perm]Ratio of the (purported) free
274 274 // space be allocated before initiating a new collection cycle.
275 275 //
276 276 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) {
277 277 assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments");
278 278 if (io >= 0) {
279 279 _initiating_occupancy = (double)io / 100.0;
280 280 } else {
281 281 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
282 282 (double)(tr * MinHeapFreeRatio) / 100.0)
283 283 / 100.0;
284 284 }
|
↓ open down ↓ |
284 lines elided |
↑ open up ↑ |
285 285 }
286 286
287 287 void ConcurrentMarkSweepGeneration::ref_processor_init() {
288 288 assert(collector() != NULL, "no collector");
289 289 collector()->ref_processor_init();
290 290 }
291 291
292 292 void CMSCollector::ref_processor_init() {
293 293 if (_ref_processor == NULL) {
294 294 // Allocate and initialize a reference processor
295 - _ref_processor = ReferenceProcessor::create_ref_processor(
296 - _span, // span
297 - _cmsGen->refs_discovery_is_atomic(), // atomic_discovery
298 - _cmsGen->refs_discovery_is_mt(), // mt_discovery
299 - &_is_alive_closure,
300 - ParallelGCThreads,
301 - ParallelRefProcEnabled);
295 + _ref_processor =
296 + new ReferenceProcessor(_span, // span
297 + (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
298 + ParallelGCThreads, // mt processing degree
299 + _cmsGen->refs_discovery_is_mt(), // mt discovery
300 + ConcGCThreads, // mt discovery degree
301 + _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
302 + &_is_alive_closure, // closure for liveness info
303 + false); // next field updates do not need write barrier
302 304 // Initialize the _ref_processor field of CMSGen
303 305 _cmsGen->set_ref_processor(_ref_processor);
304 306
305 307 // Allocate a dummy ref processor for perm gen.
306 308 ReferenceProcessor* rp2 = new ReferenceProcessor();
307 309 if (rp2 == NULL) {
308 310 vm_exit_during_initialization("Could not allocate ReferenceProcessor object");
309 311 }
310 312 _permGen->set_ref_processor(rp2);
311 313 }
312 314 }
313 315
314 316 CMSAdaptiveSizePolicy* CMSCollector::size_policy() {
315 317 GenCollectedHeap* gch = GenCollectedHeap::heap();
316 318 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
317 319 "Wrong type of heap");
318 320 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
319 321 gch->gen_policy()->size_policy();
320 322 assert(sp->is_gc_cms_adaptive_size_policy(),
321 323 "Wrong type of size policy");
322 324 return sp;
323 325 }
324 326
325 327 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() {
326 328 CMSGCAdaptivePolicyCounters* results =
327 329 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters();
328 330 assert(
329 331 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
330 332 "Wrong gc policy counter kind");
331 333 return results;
332 334 }
333 335
334 336
335 337 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
336 338
337 339 const char* gen_name = "old";
338 340
339 341 // Generation Counters - generation 1, 1 subspace
340 342 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space);
341 343
342 344 _space_counters = new GSpaceCounters(gen_name, 0,
343 345 _virtual_space.reserved_size(),
344 346 this, _gen_counters);
345 347 }
346 348
347 349 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
348 350 _cms_gen(cms_gen)
349 351 {
350 352 assert(alpha <= 100, "bad value");
351 353 _saved_alpha = alpha;
352 354
353 355 // Initialize the alphas to the bootstrap value of 100.
354 356 _gc0_alpha = _cms_alpha = 100;
355 357
356 358 _cms_begin_time.update();
357 359 _cms_end_time.update();
358 360
359 361 _gc0_duration = 0.0;
360 362 _gc0_period = 0.0;
361 363 _gc0_promoted = 0;
362 364
363 365 _cms_duration = 0.0;
364 366 _cms_period = 0.0;
365 367 _cms_allocated = 0;
366 368
367 369 _cms_used_at_gc0_begin = 0;
368 370 _cms_used_at_gc0_end = 0;
369 371 _allow_duty_cycle_reduction = false;
370 372 _valid_bits = 0;
371 373 _icms_duty_cycle = CMSIncrementalDutyCycle;
372 374 }
373 375
374 376 double CMSStats::cms_free_adjustment_factor(size_t free) const {
375 377 // TBD: CR 6909490
376 378 return 1.0;
377 379 }
378 380
379 381 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
380 382 }
381 383
382 384 // If promotion failure handling is on use
383 385 // the padded average size of the promotion for each
384 386 // young generation collection.
385 387 double CMSStats::time_until_cms_gen_full() const {
386 388 size_t cms_free = _cms_gen->cmsSpace()->free();
387 389 GenCollectedHeap* gch = GenCollectedHeap::heap();
388 390 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(),
389 391 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
390 392 if (cms_free > expected_promotion) {
391 393 // Start a cms collection if there isn't enough space to promote
392 394 // for the next minor collection. Use the padded average as
393 395 // a safety factor.
394 396 cms_free -= expected_promotion;
395 397
396 398 // Adjust by the safety factor.
397 399 double cms_free_dbl = (double)cms_free;
398 400 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
399 401 // Apply a further correction factor which tries to adjust
400 402 // for recent occurance of concurrent mode failures.
401 403 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
402 404 cms_free_dbl = cms_free_dbl * cms_adjustment;
403 405
404 406 if (PrintGCDetails && Verbose) {
405 407 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
406 408 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
407 409 cms_free, expected_promotion);
408 410 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
409 411 cms_free_dbl, cms_consumption_rate() + 1.0);
410 412 }
411 413 // Add 1 in case the consumption rate goes to zero.
412 414 return cms_free_dbl / (cms_consumption_rate() + 1.0);
413 415 }
414 416 return 0.0;
415 417 }
416 418
417 419 // Compare the duration of the cms collection to the
418 420 // time remaining before the cms generation is empty.
419 421 // Note that the time from the start of the cms collection
420 422 // to the start of the cms sweep (less than the total
421 423 // duration of the cms collection) can be used. This
422 424 // has been tried and some applications experienced
423 425 // promotion failures early in execution. This was
424 426 // possibly because the averages were not accurate
425 427 // enough at the beginning.
426 428 double CMSStats::time_until_cms_start() const {
427 429 // We add "gc0_period" to the "work" calculation
428 430 // below because this query is done (mostly) at the
429 431 // end of a scavenge, so we need to conservatively
430 432 // account for that much possible delay
431 433 // in the query so as to avoid concurrent mode failures
432 434 // due to starting the collection just a wee bit too
433 435 // late.
434 436 double work = cms_duration() + gc0_period();
435 437 double deadline = time_until_cms_gen_full();
436 438 // If a concurrent mode failure occurred recently, we want to be
437 439 // more conservative and halve our expected time_until_cms_gen_full()
438 440 if (work > deadline) {
439 441 if (Verbose && PrintGCDetails) {
440 442 gclog_or_tty->print(
441 443 " CMSCollector: collect because of anticipated promotion "
442 444 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
443 445 gc0_period(), time_until_cms_gen_full());
444 446 }
445 447 return 0.0;
446 448 }
447 449 return work - deadline;
448 450 }
449 451
450 452 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the
451 453 // amount of change to prevent wild oscillation.
452 454 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle,
453 455 unsigned int new_duty_cycle) {
454 456 assert(old_duty_cycle <= 100, "bad input value");
455 457 assert(new_duty_cycle <= 100, "bad input value");
456 458
457 459 // Note: use subtraction with caution since it may underflow (values are
458 460 // unsigned). Addition is safe since we're in the range 0-100.
459 461 unsigned int damped_duty_cycle = new_duty_cycle;
460 462 if (new_duty_cycle < old_duty_cycle) {
461 463 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U);
462 464 if (new_duty_cycle + largest_delta < old_duty_cycle) {
463 465 damped_duty_cycle = old_duty_cycle - largest_delta;
464 466 }
465 467 } else if (new_duty_cycle > old_duty_cycle) {
466 468 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U);
467 469 if (new_duty_cycle > old_duty_cycle + largest_delta) {
468 470 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U);
469 471 }
470 472 }
471 473 assert(damped_duty_cycle <= 100, "invalid duty cycle computed");
472 474
473 475 if (CMSTraceIncrementalPacing) {
474 476 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ",
475 477 old_duty_cycle, new_duty_cycle, damped_duty_cycle);
476 478 }
477 479 return damped_duty_cycle;
478 480 }
479 481
480 482 unsigned int CMSStats::icms_update_duty_cycle_impl() {
481 483 assert(CMSIncrementalPacing && valid(),
482 484 "should be handled in icms_update_duty_cycle()");
483 485
484 486 double cms_time_so_far = cms_timer().seconds();
485 487 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M;
486 488 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far);
487 489
488 490 // Avoid division by 0.
489 491 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01);
490 492 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full;
491 493
492 494 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U);
493 495 if (new_duty_cycle > _icms_duty_cycle) {
494 496 // Avoid very small duty cycles (1 or 2); 0 is allowed.
495 497 if (new_duty_cycle > 2) {
496 498 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle,
497 499 new_duty_cycle);
498 500 }
499 501 } else if (_allow_duty_cycle_reduction) {
500 502 // The duty cycle is reduced only once per cms cycle (see record_cms_end()).
501 503 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle);
502 504 // Respect the minimum duty cycle.
503 505 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin;
504 506 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle);
505 507 }
506 508
507 509 if (PrintGCDetails || CMSTraceIncrementalPacing) {
508 510 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle);
509 511 }
510 512
511 513 _allow_duty_cycle_reduction = false;
512 514 return _icms_duty_cycle;
513 515 }
514 516
515 517 #ifndef PRODUCT
516 518 void CMSStats::print_on(outputStream *st) const {
517 519 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
518 520 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
519 521 gc0_duration(), gc0_period(), gc0_promoted());
520 522 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
521 523 cms_duration(), cms_duration_per_mb(),
522 524 cms_period(), cms_allocated());
523 525 st->print(",cms_since_beg=%g,cms_since_end=%g",
524 526 cms_time_since_begin(), cms_time_since_end());
525 527 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
526 528 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
527 529 if (CMSIncrementalMode) {
528 530 st->print(",dc=%d", icms_duty_cycle());
529 531 }
530 532
531 533 if (valid()) {
532 534 st->print(",promo_rate=%g,cms_alloc_rate=%g",
533 535 promotion_rate(), cms_allocation_rate());
534 536 st->print(",cms_consumption_rate=%g,time_until_full=%g",
535 537 cms_consumption_rate(), time_until_cms_gen_full());
536 538 }
537 539 st->print(" ");
538 540 }
539 541 #endif // #ifndef PRODUCT
540 542
541 543 CMSCollector::CollectorState CMSCollector::_collectorState =
542 544 CMSCollector::Idling;
543 545 bool CMSCollector::_foregroundGCIsActive = false;
544 546 bool CMSCollector::_foregroundGCShouldWait = false;
545 547
546 548 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
547 549 ConcurrentMarkSweepGeneration* permGen,
548 550 CardTableRS* ct,
549 551 ConcurrentMarkSweepPolicy* cp):
550 552 _cmsGen(cmsGen),
551 553 _permGen(permGen),
552 554 _ct(ct),
553 555 _ref_processor(NULL), // will be set later
554 556 _conc_workers(NULL), // may be set later
555 557 _abort_preclean(false),
556 558 _start_sampling(false),
557 559 _between_prologue_and_epilogue(false),
558 560 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
559 561 _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"),
560 562 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
561 563 -1 /* lock-free */, "No_lock" /* dummy */),
562 564 _modUnionClosure(&_modUnionTable),
563 565 _modUnionClosurePar(&_modUnionTable),
564 566 // Adjust my span to cover old (cms) gen and perm gen
565 567 _span(cmsGen->reserved()._union(permGen->reserved())),
566 568 // Construct the is_alive_closure with _span & markBitMap
567 569 _is_alive_closure(_span, &_markBitMap),
568 570 _restart_addr(NULL),
569 571 _overflow_list(NULL),
570 572 _stats(cmsGen),
571 573 _eden_chunk_array(NULL), // may be set in ctor body
572 574 _eden_chunk_capacity(0), // -- ditto --
573 575 _eden_chunk_index(0), // -- ditto --
574 576 _survivor_plab_array(NULL), // -- ditto --
575 577 _survivor_chunk_array(NULL), // -- ditto --
576 578 _survivor_chunk_capacity(0), // -- ditto --
577 579 _survivor_chunk_index(0), // -- ditto --
578 580 _ser_pmc_preclean_ovflw(0),
579 581 _ser_kac_preclean_ovflw(0),
580 582 _ser_pmc_remark_ovflw(0),
581 583 _par_pmc_remark_ovflw(0),
582 584 _ser_kac_ovflw(0),
583 585 _par_kac_ovflw(0),
584 586 #ifndef PRODUCT
585 587 _num_par_pushes(0),
586 588 #endif
587 589 _collection_count_start(0),
588 590 _verifying(false),
589 591 _icms_start_limit(NULL),
590 592 _icms_stop_limit(NULL),
591 593 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
592 594 _completed_initialization(false),
593 595 _collector_policy(cp),
594 596 _should_unload_classes(false),
595 597 _concurrent_cycles_since_last_unload(0),
596 598 _roots_scanning_options(0),
597 599 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
598 600 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding)
599 601 {
600 602 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
601 603 ExplicitGCInvokesConcurrent = true;
602 604 }
603 605 // Now expand the span and allocate the collection support structures
604 606 // (MUT, marking bit map etc.) to cover both generations subject to
605 607 // collection.
606 608
607 609 // First check that _permGen is adjacent to _cmsGen and above it.
608 610 assert( _cmsGen->reserved().word_size() > 0
609 611 && _permGen->reserved().word_size() > 0,
610 612 "generations should not be of zero size");
611 613 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(),
612 614 "_cmsGen and _permGen should not overlap");
613 615 assert(_cmsGen->reserved().end() == _permGen->reserved().start(),
614 616 "_cmsGen->end() different from _permGen->start()");
615 617
616 618 // For use by dirty card to oop closures.
617 619 _cmsGen->cmsSpace()->set_collector(this);
618 620 _permGen->cmsSpace()->set_collector(this);
619 621
620 622 // Allocate MUT and marking bit map
621 623 {
622 624 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
623 625 if (!_markBitMap.allocate(_span)) {
624 626 warning("Failed to allocate CMS Bit Map");
625 627 return;
626 628 }
627 629 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
628 630 }
629 631 {
630 632 _modUnionTable.allocate(_span);
631 633 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
632 634 }
633 635
634 636 if (!_markStack.allocate(MarkStackSize)) {
635 637 warning("Failed to allocate CMS Marking Stack");
636 638 return;
637 639 }
638 640 if (!_revisitStack.allocate(CMSRevisitStackSize)) {
639 641 warning("Failed to allocate CMS Revisit Stack");
640 642 return;
641 643 }
642 644
643 645 // Support for multi-threaded concurrent phases
644 646 if (CollectedHeap::use_parallel_gc_threads() && CMSConcurrentMTEnabled) {
645 647 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
646 648 // just for now
647 649 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
648 650 }
649 651 if (ConcGCThreads > 1) {
650 652 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
651 653 ConcGCThreads, true);
652 654 if (_conc_workers == NULL) {
653 655 warning("GC/CMS: _conc_workers allocation failure: "
654 656 "forcing -CMSConcurrentMTEnabled");
655 657 CMSConcurrentMTEnabled = false;
656 658 } else {
657 659 _conc_workers->initialize_workers();
658 660 }
659 661 } else {
660 662 CMSConcurrentMTEnabled = false;
661 663 }
662 664 }
663 665 if (!CMSConcurrentMTEnabled) {
664 666 ConcGCThreads = 0;
665 667 } else {
666 668 // Turn off CMSCleanOnEnter optimization temporarily for
667 669 // the MT case where it's not fixed yet; see 6178663.
668 670 CMSCleanOnEnter = false;
669 671 }
670 672 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
671 673 "Inconsistency");
672 674
673 675 // Parallel task queues; these are shared for the
674 676 // concurrent and stop-world phases of CMS, but
675 677 // are not shared with parallel scavenge (ParNew).
676 678 {
677 679 uint i;
678 680 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
679 681
680 682 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
681 683 || ParallelRefProcEnabled)
682 684 && num_queues > 0) {
683 685 _task_queues = new OopTaskQueueSet(num_queues);
684 686 if (_task_queues == NULL) {
685 687 warning("task_queues allocation failure.");
686 688 return;
687 689 }
688 690 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues);
689 691 if (_hash_seed == NULL) {
690 692 warning("_hash_seed array allocation failure");
691 693 return;
692 694 }
693 695
694 696 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
695 697 for (i = 0; i < num_queues; i++) {
696 698 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
697 699 if (q == NULL) {
698 700 warning("work_queue allocation failure.");
699 701 return;
700 702 }
701 703 _task_queues->register_queue(i, q);
702 704 }
703 705 for (i = 0; i < num_queues; i++) {
704 706 _task_queues->queue(i)->initialize();
705 707 _hash_seed[i] = 17; // copied from ParNew
706 708 }
707 709 }
708 710 }
709 711
710 712 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
711 713 _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio);
712 714
713 715 // Clip CMSBootstrapOccupancy between 0 and 100.
714 716 _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy)))
715 717 /(double)100;
716 718
717 719 _full_gcs_since_conc_gc = 0;
718 720
719 721 // Now tell CMS generations the identity of their collector
720 722 ConcurrentMarkSweepGeneration::set_collector(this);
721 723
722 724 // Create & start a CMS thread for this CMS collector
723 725 _cmsThread = ConcurrentMarkSweepThread::start(this);
724 726 assert(cmsThread() != NULL, "CMS Thread should have been created");
725 727 assert(cmsThread()->collector() == this,
726 728 "CMS Thread should refer to this gen");
727 729 assert(CGC_lock != NULL, "Where's the CGC_lock?");
728 730
729 731 // Support for parallelizing young gen rescan
730 732 GenCollectedHeap* gch = GenCollectedHeap::heap();
731 733 _young_gen = gch->prev_gen(_cmsGen);
732 734 if (gch->supports_inline_contig_alloc()) {
733 735 _top_addr = gch->top_addr();
734 736 _end_addr = gch->end_addr();
735 737 assert(_young_gen != NULL, "no _young_gen");
736 738 _eden_chunk_index = 0;
737 739 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
738 740 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity);
739 741 if (_eden_chunk_array == NULL) {
740 742 _eden_chunk_capacity = 0;
741 743 warning("GC/CMS: _eden_chunk_array allocation failure");
742 744 }
743 745 }
744 746 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
745 747
746 748 // Support for parallelizing survivor space rescan
747 749 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) {
748 750 const size_t max_plab_samples =
749 751 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize;
750 752
751 753 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads);
752 754 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples);
753 755 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads);
754 756 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
755 757 || _cursor == NULL) {
756 758 warning("Failed to allocate survivor plab/chunk array");
757 759 if (_survivor_plab_array != NULL) {
758 760 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
759 761 _survivor_plab_array = NULL;
760 762 }
761 763 if (_survivor_chunk_array != NULL) {
762 764 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
763 765 _survivor_chunk_array = NULL;
764 766 }
765 767 if (_cursor != NULL) {
766 768 FREE_C_HEAP_ARRAY(size_t, _cursor);
767 769 _cursor = NULL;
768 770 }
769 771 } else {
770 772 _survivor_chunk_capacity = 2*max_plab_samples;
771 773 for (uint i = 0; i < ParallelGCThreads; i++) {
772 774 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples);
773 775 if (vec == NULL) {
774 776 warning("Failed to allocate survivor plab array");
775 777 for (int j = i; j > 0; j--) {
776 778 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array());
777 779 }
778 780 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
779 781 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
780 782 _survivor_plab_array = NULL;
781 783 _survivor_chunk_array = NULL;
782 784 _survivor_chunk_capacity = 0;
783 785 break;
784 786 } else {
785 787 ChunkArray* cur =
786 788 ::new (&_survivor_plab_array[i]) ChunkArray(vec,
787 789 max_plab_samples);
788 790 assert(cur->end() == 0, "Should be 0");
789 791 assert(cur->array() == vec, "Should be vec");
790 792 assert(cur->capacity() == max_plab_samples, "Error");
791 793 }
792 794 }
793 795 }
794 796 }
795 797 assert( ( _survivor_plab_array != NULL
796 798 && _survivor_chunk_array != NULL)
797 799 || ( _survivor_chunk_capacity == 0
798 800 && _survivor_chunk_index == 0),
799 801 "Error");
800 802
801 803 // Choose what strong roots should be scanned depending on verification options
802 804 // and perm gen collection mode.
803 805 if (!CMSClassUnloadingEnabled) {
804 806 // If class unloading is disabled we want to include all classes into the root set.
805 807 add_root_scanning_option(SharedHeap::SO_AllClasses);
806 808 } else {
807 809 add_root_scanning_option(SharedHeap::SO_SystemClasses);
808 810 }
809 811
810 812 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
811 813 _gc_counters = new CollectorCounters("CMS", 1);
812 814 _completed_initialization = true;
813 815 _inter_sweep_timer.start(); // start of time
814 816 #ifdef SPARC
815 817 // Issue a stern warning, but allow use for experimentation and debugging.
816 818 if (VM_Version::is_sun4v() && UseMemSetInBOT) {
817 819 assert(!FLAG_IS_DEFAULT(UseMemSetInBOT), "Error");
818 820 warning("Experimental flag -XX:+UseMemSetInBOT is known to cause instability"
819 821 " on sun4v; please understand that you are using at your own risk!");
820 822 }
821 823 #endif
822 824 }
823 825
824 826 const char* ConcurrentMarkSweepGeneration::name() const {
825 827 return "concurrent mark-sweep generation";
826 828 }
827 829 void ConcurrentMarkSweepGeneration::update_counters() {
828 830 if (UsePerfData) {
829 831 _space_counters->update_all();
830 832 _gen_counters->update_all();
831 833 }
832 834 }
833 835
834 836 // this is an optimized version of update_counters(). it takes the
835 837 // used value as a parameter rather than computing it.
836 838 //
837 839 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
838 840 if (UsePerfData) {
839 841 _space_counters->update_used(used);
840 842 _space_counters->update_capacity();
841 843 _gen_counters->update_all();
842 844 }
843 845 }
844 846
845 847 void ConcurrentMarkSweepGeneration::print() const {
846 848 Generation::print();
847 849 cmsSpace()->print();
848 850 }
849 851
850 852 #ifndef PRODUCT
851 853 void ConcurrentMarkSweepGeneration::print_statistics() {
852 854 cmsSpace()->printFLCensus(0);
853 855 }
854 856 #endif
855 857
856 858 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
857 859 GenCollectedHeap* gch = GenCollectedHeap::heap();
858 860 if (PrintGCDetails) {
859 861 if (Verbose) {
860 862 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
861 863 level(), short_name(), s, used(), capacity());
862 864 } else {
863 865 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
864 866 level(), short_name(), s, used() / K, capacity() / K);
865 867 }
866 868 }
867 869 if (Verbose) {
868 870 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
869 871 gch->used(), gch->capacity());
870 872 } else {
871 873 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
872 874 gch->used() / K, gch->capacity() / K);
873 875 }
874 876 }
875 877
876 878 size_t
877 879 ConcurrentMarkSweepGeneration::contiguous_available() const {
878 880 // dld proposes an improvement in precision here. If the committed
879 881 // part of the space ends in a free block we should add that to
880 882 // uncommitted size in the calculation below. Will make this
881 883 // change later, staying with the approximation below for the
882 884 // time being. -- ysr.
883 885 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
884 886 }
885 887
886 888 size_t
887 889 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
888 890 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
889 891 }
890 892
891 893 size_t ConcurrentMarkSweepGeneration::max_available() const {
892 894 return free() + _virtual_space.uncommitted_size();
893 895 }
894 896
895 897 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
896 898 size_t available = max_available();
897 899 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
898 900 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
899 901 if (Verbose && PrintGCDetails) {
900 902 gclog_or_tty->print_cr(
901 903 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
902 904 "max_promo("SIZE_FORMAT")",
903 905 res? "":" not", available, res? ">=":"<",
904 906 av_promo, max_promotion_in_bytes);
905 907 }
906 908 return res;
907 909 }
908 910
909 911 // At a promotion failure dump information on block layout in heap
910 912 // (cms old generation).
911 913 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
912 914 if (CMSDumpAtPromotionFailure) {
913 915 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
914 916 }
915 917 }
916 918
917 919 CompactibleSpace*
918 920 ConcurrentMarkSweepGeneration::first_compaction_space() const {
919 921 return _cmsSpace;
920 922 }
921 923
922 924 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
923 925 // Clear the promotion information. These pointers can be adjusted
924 926 // along with all the other pointers into the heap but
925 927 // compaction is expected to be a rare event with
926 928 // a heap using cms so don't do it without seeing the need.
927 929 if (CollectedHeap::use_parallel_gc_threads()) {
928 930 for (uint i = 0; i < ParallelGCThreads; i++) {
929 931 _par_gc_thread_states[i]->promo.reset();
930 932 }
931 933 }
932 934 }
933 935
934 936 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
935 937 blk->do_space(_cmsSpace);
936 938 }
937 939
938 940 void ConcurrentMarkSweepGeneration::compute_new_size() {
939 941 assert_locked_or_safepoint(Heap_lock);
940 942
941 943 // If incremental collection failed, we just want to expand
942 944 // to the limit.
943 945 if (incremental_collection_failed()) {
944 946 clear_incremental_collection_failed();
945 947 grow_to_reserved();
946 948 return;
947 949 }
948 950
949 951 size_t expand_bytes = 0;
950 952 double free_percentage = ((double) free()) / capacity();
951 953 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
952 954 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
953 955
954 956 // compute expansion delta needed for reaching desired free percentage
955 957 if (free_percentage < desired_free_percentage) {
956 958 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
957 959 assert(desired_capacity >= capacity(), "invalid expansion size");
958 960 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
959 961 }
960 962 if (expand_bytes > 0) {
961 963 if (PrintGCDetails && Verbose) {
962 964 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
963 965 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
964 966 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
965 967 gclog_or_tty->print_cr(" Desired free fraction %f",
966 968 desired_free_percentage);
967 969 gclog_or_tty->print_cr(" Maximum free fraction %f",
968 970 maximum_free_percentage);
969 971 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000);
970 972 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
971 973 desired_capacity/1000);
972 974 int prev_level = level() - 1;
973 975 if (prev_level >= 0) {
974 976 size_t prev_size = 0;
975 977 GenCollectedHeap* gch = GenCollectedHeap::heap();
976 978 Generation* prev_gen = gch->_gens[prev_level];
977 979 prev_size = prev_gen->capacity();
978 980 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
979 981 prev_size/1000);
980 982 }
981 983 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
982 984 unsafe_max_alloc_nogc()/1000);
983 985 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
984 986 contiguous_available()/1000);
985 987 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
986 988 expand_bytes);
987 989 }
988 990 // safe if expansion fails
989 991 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
990 992 if (PrintGCDetails && Verbose) {
991 993 gclog_or_tty->print_cr(" Expanded free fraction %f",
992 994 ((double) free()) / capacity());
993 995 }
994 996 }
995 997 }
996 998
997 999 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
998 1000 return cmsSpace()->freelistLock();
999 1001 }
1000 1002
1001 1003 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
1002 1004 bool tlab) {
1003 1005 CMSSynchronousYieldRequest yr;
1004 1006 MutexLockerEx x(freelistLock(),
1005 1007 Mutex::_no_safepoint_check_flag);
1006 1008 return have_lock_and_allocate(size, tlab);
1007 1009 }
1008 1010
1009 1011 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
1010 1012 bool tlab /* ignored */) {
1011 1013 assert_lock_strong(freelistLock());
1012 1014 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
1013 1015 HeapWord* res = cmsSpace()->allocate(adjustedSize);
1014 1016 // Allocate the object live (grey) if the background collector has
1015 1017 // started marking. This is necessary because the marker may
1016 1018 // have passed this address and consequently this object will
1017 1019 // not otherwise be greyed and would be incorrectly swept up.
1018 1020 // Note that if this object contains references, the writing
1019 1021 // of those references will dirty the card containing this object
1020 1022 // allowing the object to be blackened (and its references scanned)
1021 1023 // either during a preclean phase or at the final checkpoint.
1022 1024 if (res != NULL) {
1023 1025 // We may block here with an uninitialized object with
1024 1026 // its mark-bit or P-bits not yet set. Such objects need
1025 1027 // to be safely navigable by block_start().
1026 1028 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
1027 1029 assert(!((FreeChunk*)res)->isFree(), "Error, block will look free but show wrong size");
1028 1030 collector()->direct_allocated(res, adjustedSize);
1029 1031 _direct_allocated_words += adjustedSize;
1030 1032 // allocation counters
1031 1033 NOT_PRODUCT(
1032 1034 _numObjectsAllocated++;
1033 1035 _numWordsAllocated += (int)adjustedSize;
1034 1036 )
1035 1037 }
1036 1038 return res;
1037 1039 }
1038 1040
1039 1041 // In the case of direct allocation by mutators in a generation that
1040 1042 // is being concurrently collected, the object must be allocated
1041 1043 // live (grey) if the background collector has started marking.
1042 1044 // This is necessary because the marker may
1043 1045 // have passed this address and consequently this object will
1044 1046 // not otherwise be greyed and would be incorrectly swept up.
1045 1047 // Note that if this object contains references, the writing
1046 1048 // of those references will dirty the card containing this object
1047 1049 // allowing the object to be blackened (and its references scanned)
1048 1050 // either during a preclean phase or at the final checkpoint.
1049 1051 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
1050 1052 assert(_markBitMap.covers(start, size), "Out of bounds");
1051 1053 if (_collectorState >= Marking) {
1052 1054 MutexLockerEx y(_markBitMap.lock(),
1053 1055 Mutex::_no_safepoint_check_flag);
1054 1056 // [see comments preceding SweepClosure::do_blk() below for details]
1055 1057 // 1. need to mark the object as live so it isn't collected
1056 1058 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
1057 1059 // 3. need to mark the end of the object so marking, precleaning or sweeping
1058 1060 // can skip over uninitialized or unparsable objects. An allocated
1059 1061 // object is considered uninitialized for our purposes as long as
1060 1062 // its klass word is NULL. (Unparsable objects are those which are
1061 1063 // initialized in the sense just described, but whose sizes can still
1062 1064 // not be correctly determined. Note that the class of unparsable objects
1063 1065 // can only occur in the perm gen. All old gen objects are parsable
1064 1066 // as soon as they are initialized.)
1065 1067 _markBitMap.mark(start); // object is live
1066 1068 _markBitMap.mark(start + 1); // object is potentially uninitialized?
1067 1069 _markBitMap.mark(start + size - 1);
1068 1070 // mark end of object
1069 1071 }
1070 1072 // check that oop looks uninitialized
1071 1073 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
1072 1074 }
1073 1075
1074 1076 void CMSCollector::promoted(bool par, HeapWord* start,
1075 1077 bool is_obj_array, size_t obj_size) {
1076 1078 assert(_markBitMap.covers(start), "Out of bounds");
1077 1079 // See comment in direct_allocated() about when objects should
1078 1080 // be allocated live.
1079 1081 if (_collectorState >= Marking) {
1080 1082 // we already hold the marking bit map lock, taken in
1081 1083 // the prologue
1082 1084 if (par) {
1083 1085 _markBitMap.par_mark(start);
1084 1086 } else {
1085 1087 _markBitMap.mark(start);
1086 1088 }
1087 1089 // We don't need to mark the object as uninitialized (as
1088 1090 // in direct_allocated above) because this is being done with the
1089 1091 // world stopped and the object will be initialized by the
1090 1092 // time the marking, precleaning or sweeping get to look at it.
1091 1093 // But see the code for copying objects into the CMS generation,
1092 1094 // where we need to ensure that concurrent readers of the
1093 1095 // block offset table are able to safely navigate a block that
1094 1096 // is in flux from being free to being allocated (and in
1095 1097 // transition while being copied into) and subsequently
1096 1098 // becoming a bona-fide object when the copy/promotion is complete.
1097 1099 assert(SafepointSynchronize::is_at_safepoint(),
1098 1100 "expect promotion only at safepoints");
1099 1101
1100 1102 if (_collectorState < Sweeping) {
1101 1103 // Mark the appropriate cards in the modUnionTable, so that
1102 1104 // this object gets scanned before the sweep. If this is
1103 1105 // not done, CMS generation references in the object might
1104 1106 // not get marked.
1105 1107 // For the case of arrays, which are otherwise precisely
1106 1108 // marked, we need to dirty the entire array, not just its head.
1107 1109 if (is_obj_array) {
1108 1110 // The [par_]mark_range() method expects mr.end() below to
1109 1111 // be aligned to the granularity of a bit's representation
1110 1112 // in the heap. In the case of the MUT below, that's a
1111 1113 // card size.
1112 1114 MemRegion mr(start,
1113 1115 (HeapWord*)round_to((intptr_t)(start + obj_size),
1114 1116 CardTableModRefBS::card_size /* bytes */));
1115 1117 if (par) {
1116 1118 _modUnionTable.par_mark_range(mr);
1117 1119 } else {
1118 1120 _modUnionTable.mark_range(mr);
1119 1121 }
1120 1122 } else { // not an obj array; we can just mark the head
1121 1123 if (par) {
1122 1124 _modUnionTable.par_mark(start);
1123 1125 } else {
1124 1126 _modUnionTable.mark(start);
1125 1127 }
1126 1128 }
1127 1129 }
1128 1130 }
1129 1131 }
1130 1132
1131 1133 static inline size_t percent_of_space(Space* space, HeapWord* addr)
1132 1134 {
1133 1135 size_t delta = pointer_delta(addr, space->bottom());
1134 1136 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
1135 1137 }
1136 1138
1137 1139 void CMSCollector::icms_update_allocation_limits()
1138 1140 {
1139 1141 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
1140 1142 EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
1141 1143
1142 1144 const unsigned int duty_cycle = stats().icms_update_duty_cycle();
1143 1145 if (CMSTraceIncrementalPacing) {
1144 1146 stats().print();
1145 1147 }
1146 1148
1147 1149 assert(duty_cycle <= 100, "invalid duty cycle");
1148 1150 if (duty_cycle != 0) {
1149 1151 // The duty_cycle is a percentage between 0 and 100; convert to words and
1150 1152 // then compute the offset from the endpoints of the space.
1151 1153 size_t free_words = eden->free() / HeapWordSize;
1152 1154 double free_words_dbl = (double)free_words;
1153 1155 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
1154 1156 size_t offset_words = (free_words - duty_cycle_words) / 2;
1155 1157
1156 1158 _icms_start_limit = eden->top() + offset_words;
1157 1159 _icms_stop_limit = eden->end() - offset_words;
1158 1160
1159 1161 // The limits may be adjusted (shifted to the right) by
1160 1162 // CMSIncrementalOffset, to allow the application more mutator time after a
1161 1163 // young gen gc (when all mutators were stopped) and before CMS starts and
1162 1164 // takes away one or more cpus.
1163 1165 if (CMSIncrementalOffset != 0) {
1164 1166 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
1165 1167 size_t adjustment = (size_t)adjustment_dbl;
1166 1168 HeapWord* tmp_stop = _icms_stop_limit + adjustment;
1167 1169 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
1168 1170 _icms_start_limit += adjustment;
1169 1171 _icms_stop_limit = tmp_stop;
1170 1172 }
1171 1173 }
1172 1174 }
1173 1175 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
1174 1176 _icms_start_limit = _icms_stop_limit = eden->end();
1175 1177 }
1176 1178
1177 1179 // Install the new start limit.
1178 1180 eden->set_soft_end(_icms_start_limit);
1179 1181
1180 1182 if (CMSTraceIncrementalMode) {
1181 1183 gclog_or_tty->print(" icms alloc limits: "
1182 1184 PTR_FORMAT "," PTR_FORMAT
1183 1185 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
1184 1186 _icms_start_limit, _icms_stop_limit,
1185 1187 percent_of_space(eden, _icms_start_limit),
1186 1188 percent_of_space(eden, _icms_stop_limit));
1187 1189 if (Verbose) {
1188 1190 gclog_or_tty->print("eden: ");
1189 1191 eden->print_on(gclog_or_tty);
1190 1192 }
1191 1193 }
1192 1194 }
1193 1195
1194 1196 // Any changes here should try to maintain the invariant
1195 1197 // that if this method is called with _icms_start_limit
1196 1198 // and _icms_stop_limit both NULL, then it should return NULL
1197 1199 // and not notify the icms thread.
1198 1200 HeapWord*
1199 1201 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
1200 1202 size_t word_size)
1201 1203 {
1202 1204 // A start_limit equal to end() means the duty cycle is 0, so treat that as a
1203 1205 // nop.
1204 1206 if (CMSIncrementalMode && _icms_start_limit != space->end()) {
1205 1207 if (top <= _icms_start_limit) {
1206 1208 if (CMSTraceIncrementalMode) {
1207 1209 space->print_on(gclog_or_tty);
1208 1210 gclog_or_tty->stamp();
1209 1211 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
1210 1212 ", new limit=" PTR_FORMAT
1211 1213 " (" SIZE_FORMAT "%%)",
1212 1214 top, _icms_stop_limit,
1213 1215 percent_of_space(space, _icms_stop_limit));
1214 1216 }
1215 1217 ConcurrentMarkSweepThread::start_icms();
1216 1218 assert(top < _icms_stop_limit, "Tautology");
1217 1219 if (word_size < pointer_delta(_icms_stop_limit, top)) {
1218 1220 return _icms_stop_limit;
1219 1221 }
1220 1222
1221 1223 // The allocation will cross both the _start and _stop limits, so do the
1222 1224 // stop notification also and return end().
1223 1225 if (CMSTraceIncrementalMode) {
1224 1226 space->print_on(gclog_or_tty);
1225 1227 gclog_or_tty->stamp();
1226 1228 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
1227 1229 ", new limit=" PTR_FORMAT
1228 1230 " (" SIZE_FORMAT "%%)",
1229 1231 top, space->end(),
1230 1232 percent_of_space(space, space->end()));
1231 1233 }
1232 1234 ConcurrentMarkSweepThread::stop_icms();
1233 1235 return space->end();
1234 1236 }
1235 1237
1236 1238 if (top <= _icms_stop_limit) {
1237 1239 if (CMSTraceIncrementalMode) {
1238 1240 space->print_on(gclog_or_tty);
1239 1241 gclog_or_tty->stamp();
1240 1242 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
1241 1243 ", new limit=" PTR_FORMAT
1242 1244 " (" SIZE_FORMAT "%%)",
1243 1245 top, space->end(),
1244 1246 percent_of_space(space, space->end()));
1245 1247 }
1246 1248 ConcurrentMarkSweepThread::stop_icms();
1247 1249 return space->end();
1248 1250 }
1249 1251
1250 1252 if (CMSTraceIncrementalMode) {
1251 1253 space->print_on(gclog_or_tty);
1252 1254 gclog_or_tty->stamp();
1253 1255 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
1254 1256 ", new limit=" PTR_FORMAT,
1255 1257 top, NULL);
1256 1258 }
1257 1259 }
1258 1260
1259 1261 return NULL;
1260 1262 }
1261 1263
1262 1264 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
1263 1265 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1264 1266 // allocate, copy and if necessary update promoinfo --
1265 1267 // delegate to underlying space.
1266 1268 assert_lock_strong(freelistLock());
1267 1269
1268 1270 #ifndef PRODUCT
1269 1271 if (Universe::heap()->promotion_should_fail()) {
1270 1272 return NULL;
1271 1273 }
1272 1274 #endif // #ifndef PRODUCT
1273 1275
1274 1276 oop res = _cmsSpace->promote(obj, obj_size);
1275 1277 if (res == NULL) {
1276 1278 // expand and retry
1277 1279 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
1278 1280 expand(s*HeapWordSize, MinHeapDeltaBytes,
1279 1281 CMSExpansionCause::_satisfy_promotion);
1280 1282 // Since there's currently no next generation, we don't try to promote
1281 1283 // into a more senior generation.
1282 1284 assert(next_gen() == NULL, "assumption, based upon which no attempt "
1283 1285 "is made to pass on a possibly failing "
1284 1286 "promotion to next generation");
1285 1287 res = _cmsSpace->promote(obj, obj_size);
1286 1288 }
1287 1289 if (res != NULL) {
1288 1290 // See comment in allocate() about when objects should
1289 1291 // be allocated live.
1290 1292 assert(obj->is_oop(), "Will dereference klass pointer below");
1291 1293 collector()->promoted(false, // Not parallel
1292 1294 (HeapWord*)res, obj->is_objArray(), obj_size);
1293 1295 // promotion counters
1294 1296 NOT_PRODUCT(
1295 1297 _numObjectsPromoted++;
1296 1298 _numWordsPromoted +=
1297 1299 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1298 1300 )
1299 1301 }
1300 1302 return res;
1301 1303 }
1302 1304
1303 1305
1304 1306 HeapWord*
1305 1307 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
1306 1308 HeapWord* top,
1307 1309 size_t word_sz)
1308 1310 {
1309 1311 return collector()->allocation_limit_reached(space, top, word_sz);
1310 1312 }
1311 1313
1312 1314 // IMPORTANT: Notes on object size recognition in CMS.
1313 1315 // ---------------------------------------------------
1314 1316 // A block of storage in the CMS generation is always in
1315 1317 // one of three states. A free block (FREE), an allocated
1316 1318 // object (OBJECT) whose size() method reports the correct size,
1317 1319 // and an intermediate state (TRANSIENT) in which its size cannot
1318 1320 // be accurately determined.
1319 1321 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1320 1322 // -----------------------------------------------------
1321 1323 // FREE: klass_word & 1 == 1; mark_word holds block size
1322 1324 //
1323 1325 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1324 1326 // obj->size() computes correct size
1325 1327 // [Perm Gen objects needs to be "parsable" before they can be navigated]
1326 1328 //
1327 1329 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1328 1330 //
1329 1331 // STATE IDENTIFICATION: (64 bit+COOPS)
1330 1332 // ------------------------------------
1331 1333 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1332 1334 //
1333 1335 // OBJECT: klass_word installed; klass_word != 0;
1334 1336 // obj->size() computes correct size
1335 1337 // [Perm Gen comment above continues to hold]
1336 1338 //
1337 1339 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1338 1340 //
1339 1341 //
1340 1342 // STATE TRANSITION DIAGRAM
1341 1343 //
1342 1344 // mut / parnew mut / parnew
1343 1345 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1344 1346 // ^ |
1345 1347 // |------------------------ DEAD <------------------------------------|
1346 1348 // sweep mut
1347 1349 //
1348 1350 // While a block is in TRANSIENT state its size cannot be determined
1349 1351 // so readers will either need to come back later or stall until
1350 1352 // the size can be determined. Note that for the case of direct
1351 1353 // allocation, P-bits, when available, may be used to determine the
1352 1354 // size of an object that may not yet have been initialized.
1353 1355
1354 1356 // Things to support parallel young-gen collection.
1355 1357 oop
1356 1358 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1357 1359 oop old, markOop m,
1358 1360 size_t word_sz) {
1359 1361 #ifndef PRODUCT
1360 1362 if (Universe::heap()->promotion_should_fail()) {
1361 1363 return NULL;
1362 1364 }
1363 1365 #endif // #ifndef PRODUCT
1364 1366
1365 1367 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1366 1368 PromotionInfo* promoInfo = &ps->promo;
1367 1369 // if we are tracking promotions, then first ensure space for
1368 1370 // promotion (including spooling space for saving header if necessary).
1369 1371 // then allocate and copy, then track promoted info if needed.
1370 1372 // When tracking (see PromotionInfo::track()), the mark word may
1371 1373 // be displaced and in this case restoration of the mark word
1372 1374 // occurs in the (oop_since_save_marks_)iterate phase.
1373 1375 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1374 1376 // Out of space for allocating spooling buffers;
1375 1377 // try expanding and allocating spooling buffers.
1376 1378 if (!expand_and_ensure_spooling_space(promoInfo)) {
1377 1379 return NULL;
1378 1380 }
1379 1381 }
1380 1382 assert(promoInfo->has_spooling_space(), "Control point invariant");
1381 1383 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1382 1384 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1383 1385 if (obj_ptr == NULL) {
1384 1386 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1385 1387 if (obj_ptr == NULL) {
1386 1388 return NULL;
1387 1389 }
1388 1390 }
1389 1391 oop obj = oop(obj_ptr);
1390 1392 OrderAccess::storestore();
1391 1393 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1392 1394 assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size");
1393 1395 // IMPORTANT: See note on object initialization for CMS above.
1394 1396 // Otherwise, copy the object. Here we must be careful to insert the
1395 1397 // klass pointer last, since this marks the block as an allocated object.
1396 1398 // Except with compressed oops it's the mark word.
1397 1399 HeapWord* old_ptr = (HeapWord*)old;
1398 1400 // Restore the mark word copied above.
1399 1401 obj->set_mark(m);
1400 1402 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1401 1403 assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size");
1402 1404 OrderAccess::storestore();
1403 1405
1404 1406 if (UseCompressedOops) {
1405 1407 // Copy gap missed by (aligned) header size calculation below
1406 1408 obj->set_klass_gap(old->klass_gap());
1407 1409 }
1408 1410 if (word_sz > (size_t)oopDesc::header_size()) {
1409 1411 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1410 1412 obj_ptr + oopDesc::header_size(),
1411 1413 word_sz - oopDesc::header_size());
1412 1414 }
1413 1415
1414 1416 // Now we can track the promoted object, if necessary. We take care
1415 1417 // to delay the transition from uninitialized to full object
1416 1418 // (i.e., insertion of klass pointer) until after, so that it
1417 1419 // atomically becomes a promoted object.
1418 1420 if (promoInfo->tracking()) {
1419 1421 promoInfo->track((PromotedObject*)obj, old->klass());
1420 1422 }
1421 1423 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1422 1424 assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size");
1423 1425 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1424 1426
1425 1427 // Finally, install the klass pointer (this should be volatile).
1426 1428 OrderAccess::storestore();
1427 1429 obj->set_klass(old->klass());
1428 1430 // We should now be able to calculate the right size for this object
1429 1431 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1430 1432
1431 1433 collector()->promoted(true, // parallel
1432 1434 obj_ptr, old->is_objArray(), word_sz);
1433 1435
1434 1436 NOT_PRODUCT(
1435 1437 Atomic::inc_ptr(&_numObjectsPromoted);
1436 1438 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1437 1439 )
1438 1440
1439 1441 return obj;
1440 1442 }
1441 1443
1442 1444 void
1443 1445 ConcurrentMarkSweepGeneration::
1444 1446 par_promote_alloc_undo(int thread_num,
1445 1447 HeapWord* obj, size_t word_sz) {
1446 1448 // CMS does not support promotion undo.
1447 1449 ShouldNotReachHere();
1448 1450 }
1449 1451
1450 1452 void
1451 1453 ConcurrentMarkSweepGeneration::
1452 1454 par_promote_alloc_done(int thread_num) {
1453 1455 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1454 1456 ps->lab.retire(thread_num);
1455 1457 }
1456 1458
1457 1459 void
1458 1460 ConcurrentMarkSweepGeneration::
1459 1461 par_oop_since_save_marks_iterate_done(int thread_num) {
1460 1462 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1461 1463 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1462 1464 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1463 1465 }
1464 1466
1465 1467 // XXXPERM
1466 1468 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1467 1469 size_t size,
1468 1470 bool tlab)
1469 1471 {
1470 1472 // We allow a STW collection only if a full
1471 1473 // collection was requested.
1472 1474 return full || should_allocate(size, tlab); // FIX ME !!!
1473 1475 // This and promotion failure handling are connected at the
1474 1476 // hip and should be fixed by untying them.
1475 1477 }
1476 1478
1477 1479 bool CMSCollector::shouldConcurrentCollect() {
1478 1480 if (_full_gc_requested) {
1479 1481 if (Verbose && PrintGCDetails) {
1480 1482 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1481 1483 " gc request (or gc_locker)");
1482 1484 }
1483 1485 return true;
1484 1486 }
1485 1487
1486 1488 // For debugging purposes, change the type of collection.
1487 1489 // If the rotation is not on the concurrent collection
1488 1490 // type, don't start a concurrent collection.
1489 1491 NOT_PRODUCT(
1490 1492 if (RotateCMSCollectionTypes &&
1491 1493 (_cmsGen->debug_collection_type() !=
1492 1494 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
1493 1495 assert(_cmsGen->debug_collection_type() !=
1494 1496 ConcurrentMarkSweepGeneration::Unknown_collection_type,
1495 1497 "Bad cms collection type");
1496 1498 return false;
1497 1499 }
1498 1500 )
1499 1501
1500 1502 FreelistLocker x(this);
1501 1503 // ------------------------------------------------------------------
1502 1504 // Print out lots of information which affects the initiation of
1503 1505 // a collection.
1504 1506 if (PrintCMSInitiationStatistics && stats().valid()) {
1505 1507 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1506 1508 gclog_or_tty->stamp();
1507 1509 gclog_or_tty->print_cr("");
1508 1510 stats().print_on(gclog_or_tty);
1509 1511 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1510 1512 stats().time_until_cms_gen_full());
1511 1513 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1512 1514 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1513 1515 _cmsGen->contiguous_available());
1514 1516 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1515 1517 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1516 1518 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1517 1519 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1518 1520 gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy());
1519 1521 }
1520 1522 // ------------------------------------------------------------------
1521 1523
1522 1524 // If the estimated time to complete a cms collection (cms_duration())
1523 1525 // is less than the estimated time remaining until the cms generation
1524 1526 // is full, start a collection.
1525 1527 if (!UseCMSInitiatingOccupancyOnly) {
1526 1528 if (stats().valid()) {
1527 1529 if (stats().time_until_cms_start() == 0.0) {
1528 1530 return true;
1529 1531 }
1530 1532 } else {
1531 1533 // We want to conservatively collect somewhat early in order
1532 1534 // to try and "bootstrap" our CMS/promotion statistics;
1533 1535 // this branch will not fire after the first successful CMS
1534 1536 // collection because the stats should then be valid.
1535 1537 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1536 1538 if (Verbose && PrintGCDetails) {
1537 1539 gclog_or_tty->print_cr(
1538 1540 " CMSCollector: collect for bootstrapping statistics:"
1539 1541 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1540 1542 _bootstrap_occupancy);
1541 1543 }
1542 1544 return true;
1543 1545 }
1544 1546 }
1545 1547 }
1546 1548
1547 1549 // Otherwise, we start a collection cycle if either the perm gen or
1548 1550 // old gen want a collection cycle started. Each may use
1549 1551 // an appropriate criterion for making this decision.
1550 1552 // XXX We need to make sure that the gen expansion
1551 1553 // criterion dovetails well with this. XXX NEED TO FIX THIS
1552 1554 if (_cmsGen->should_concurrent_collect()) {
1553 1555 if (Verbose && PrintGCDetails) {
1554 1556 gclog_or_tty->print_cr("CMS old gen initiated");
1555 1557 }
1556 1558 return true;
1557 1559 }
1558 1560
1559 1561 // We start a collection if we believe an incremental collection may fail;
1560 1562 // this is not likely to be productive in practice because it's probably too
1561 1563 // late anyway.
1562 1564 GenCollectedHeap* gch = GenCollectedHeap::heap();
1563 1565 assert(gch->collector_policy()->is_two_generation_policy(),
1564 1566 "You may want to check the correctness of the following");
1565 1567 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1566 1568 if (Verbose && PrintGCDetails) {
1567 1569 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1568 1570 }
1569 1571 return true;
1570 1572 }
1571 1573
1572 1574 if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) {
1573 1575 bool res = update_should_unload_classes();
1574 1576 if (res) {
1575 1577 if (Verbose && PrintGCDetails) {
1576 1578 gclog_or_tty->print_cr("CMS perm gen initiated");
1577 1579 }
1578 1580 return true;
1579 1581 }
1580 1582 }
1581 1583 return false;
1582 1584 }
1583 1585
1584 1586 // Clear _expansion_cause fields of constituent generations
1585 1587 void CMSCollector::clear_expansion_cause() {
1586 1588 _cmsGen->clear_expansion_cause();
1587 1589 _permGen->clear_expansion_cause();
1588 1590 }
1589 1591
1590 1592 // We should be conservative in starting a collection cycle. To
1591 1593 // start too eagerly runs the risk of collecting too often in the
1592 1594 // extreme. To collect too rarely falls back on full collections,
1593 1595 // which works, even if not optimum in terms of concurrent work.
1594 1596 // As a work around for too eagerly collecting, use the flag
1595 1597 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1596 1598 // giving the user an easily understandable way of controlling the
1597 1599 // collections.
1598 1600 // We want to start a new collection cycle if any of the following
1599 1601 // conditions hold:
1600 1602 // . our current occupancy exceeds the configured initiating occupancy
1601 1603 // for this generation, or
1602 1604 // . we recently needed to expand this space and have not, since that
1603 1605 // expansion, done a collection of this generation, or
1604 1606 // . the underlying space believes that it may be a good idea to initiate
1605 1607 // a concurrent collection (this may be based on criteria such as the
1606 1608 // following: the space uses linear allocation and linear allocation is
1607 1609 // going to fail, or there is believed to be excessive fragmentation in
1608 1610 // the generation, etc... or ...
1609 1611 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1610 1612 // the case of the old generation, not the perm generation; see CR 6543076):
1611 1613 // we may be approaching a point at which allocation requests may fail because
1612 1614 // we will be out of sufficient free space given allocation rate estimates.]
1613 1615 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1614 1616
1615 1617 assert_lock_strong(freelistLock());
1616 1618 if (occupancy() > initiating_occupancy()) {
1617 1619 if (PrintGCDetails && Verbose) {
1618 1620 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1619 1621 short_name(), occupancy(), initiating_occupancy());
1620 1622 }
1621 1623 return true;
1622 1624 }
1623 1625 if (UseCMSInitiatingOccupancyOnly) {
1624 1626 return false;
1625 1627 }
1626 1628 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1627 1629 if (PrintGCDetails && Verbose) {
1628 1630 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1629 1631 short_name());
1630 1632 }
1631 1633 return true;
1632 1634 }
1633 1635 if (_cmsSpace->should_concurrent_collect()) {
1634 1636 if (PrintGCDetails && Verbose) {
1635 1637 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1636 1638 short_name());
1637 1639 }
1638 1640 return true;
1639 1641 }
1640 1642 return false;
1641 1643 }
1642 1644
1643 1645 void ConcurrentMarkSweepGeneration::collect(bool full,
1644 1646 bool clear_all_soft_refs,
1645 1647 size_t size,
1646 1648 bool tlab)
1647 1649 {
1648 1650 collector()->collect(full, clear_all_soft_refs, size, tlab);
1649 1651 }
1650 1652
1651 1653 void CMSCollector::collect(bool full,
1652 1654 bool clear_all_soft_refs,
1653 1655 size_t size,
1654 1656 bool tlab)
1655 1657 {
1656 1658 if (!UseCMSCollectionPassing && _collectorState > Idling) {
1657 1659 // For debugging purposes skip the collection if the state
1658 1660 // is not currently idle
1659 1661 if (TraceCMSState) {
1660 1662 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
1661 1663 Thread::current(), full, _collectorState);
1662 1664 }
1663 1665 return;
1664 1666 }
1665 1667
1666 1668 // The following "if" branch is present for defensive reasons.
1667 1669 // In the current uses of this interface, it can be replaced with:
1668 1670 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1669 1671 // But I am not placing that assert here to allow future
1670 1672 // generality in invoking this interface.
1671 1673 if (GC_locker::is_active()) {
1672 1674 // A consistency test for GC_locker
1673 1675 assert(GC_locker::needs_gc(), "Should have been set already");
1674 1676 // Skip this foreground collection, instead
1675 1677 // expanding the heap if necessary.
1676 1678 // Need the free list locks for the call to free() in compute_new_size()
1677 1679 compute_new_size();
1678 1680 return;
1679 1681 }
1680 1682 acquire_control_and_collect(full, clear_all_soft_refs);
1681 1683 _full_gcs_since_conc_gc++;
1682 1684
1683 1685 }
1684 1686
1685 1687 void CMSCollector::request_full_gc(unsigned int full_gc_count) {
1686 1688 GenCollectedHeap* gch = GenCollectedHeap::heap();
1687 1689 unsigned int gc_count = gch->total_full_collections();
1688 1690 if (gc_count == full_gc_count) {
1689 1691 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1690 1692 _full_gc_requested = true;
1691 1693 CGC_lock->notify(); // nudge CMS thread
1692 1694 }
1693 1695 }
1694 1696
1695 1697
1696 1698 // The foreground and background collectors need to coordinate in order
1697 1699 // to make sure that they do not mutually interfere with CMS collections.
1698 1700 // When a background collection is active,
1699 1701 // the foreground collector may need to take over (preempt) and
1700 1702 // synchronously complete an ongoing collection. Depending on the
1701 1703 // frequency of the background collections and the heap usage
1702 1704 // of the application, this preemption can be seldom or frequent.
1703 1705 // There are only certain
1704 1706 // points in the background collection that the "collection-baton"
1705 1707 // can be passed to the foreground collector.
1706 1708 //
1707 1709 // The foreground collector will wait for the baton before
1708 1710 // starting any part of the collection. The foreground collector
1709 1711 // will only wait at one location.
1710 1712 //
1711 1713 // The background collector will yield the baton before starting a new
1712 1714 // phase of the collection (e.g., before initial marking, marking from roots,
1713 1715 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1714 1716 // of the loop which switches the phases. The background collector does some
1715 1717 // of the phases (initial mark, final re-mark) with the world stopped.
1716 1718 // Because of locking involved in stopping the world,
1717 1719 // the foreground collector should not block waiting for the background
1718 1720 // collector when it is doing a stop-the-world phase. The background
1719 1721 // collector will yield the baton at an additional point just before
1720 1722 // it enters a stop-the-world phase. Once the world is stopped, the
1721 1723 // background collector checks the phase of the collection. If the
1722 1724 // phase has not changed, it proceeds with the collection. If the
1723 1725 // phase has changed, it skips that phase of the collection. See
1724 1726 // the comments on the use of the Heap_lock in collect_in_background().
1725 1727 //
1726 1728 // Variable used in baton passing.
1727 1729 // _foregroundGCIsActive - Set to true by the foreground collector when
1728 1730 // it wants the baton. The foreground clears it when it has finished
1729 1731 // the collection.
1730 1732 // _foregroundGCShouldWait - Set to true by the background collector
1731 1733 // when it is running. The foreground collector waits while
1732 1734 // _foregroundGCShouldWait is true.
1733 1735 // CGC_lock - monitor used to protect access to the above variables
1734 1736 // and to notify the foreground and background collectors.
1735 1737 // _collectorState - current state of the CMS collection.
1736 1738 //
1737 1739 // The foreground collector
1738 1740 // acquires the CGC_lock
1739 1741 // sets _foregroundGCIsActive
1740 1742 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1741 1743 // various locks acquired in preparation for the collection
1742 1744 // are released so as not to block the background collector
1743 1745 // that is in the midst of a collection
1744 1746 // proceeds with the collection
1745 1747 // clears _foregroundGCIsActive
1746 1748 // returns
1747 1749 //
1748 1750 // The background collector in a loop iterating on the phases of the
1749 1751 // collection
1750 1752 // acquires the CGC_lock
1751 1753 // sets _foregroundGCShouldWait
1752 1754 // if _foregroundGCIsActive is set
1753 1755 // clears _foregroundGCShouldWait, notifies _CGC_lock
1754 1756 // waits on _CGC_lock for _foregroundGCIsActive to become false
1755 1757 // and exits the loop.
1756 1758 // otherwise
1757 1759 // proceed with that phase of the collection
1758 1760 // if the phase is a stop-the-world phase,
1759 1761 // yield the baton once more just before enqueueing
1760 1762 // the stop-world CMS operation (executed by the VM thread).
1761 1763 // returns after all phases of the collection are done
1762 1764 //
1763 1765
1764 1766 void CMSCollector::acquire_control_and_collect(bool full,
1765 1767 bool clear_all_soft_refs) {
1766 1768 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1767 1769 assert(!Thread::current()->is_ConcurrentGC_thread(),
1768 1770 "shouldn't try to acquire control from self!");
1769 1771
1770 1772 // Start the protocol for acquiring control of the
1771 1773 // collection from the background collector (aka CMS thread).
1772 1774 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1773 1775 "VM thread should have CMS token");
1774 1776 // Remember the possibly interrupted state of an ongoing
1775 1777 // concurrent collection
1776 1778 CollectorState first_state = _collectorState;
1777 1779
1778 1780 // Signal to a possibly ongoing concurrent collection that
1779 1781 // we want to do a foreground collection.
1780 1782 _foregroundGCIsActive = true;
1781 1783
1782 1784 // Disable incremental mode during a foreground collection.
1783 1785 ICMSDisabler icms_disabler;
1784 1786
1785 1787 // release locks and wait for a notify from the background collector
1786 1788 // releasing the locks in only necessary for phases which
1787 1789 // do yields to improve the granularity of the collection.
1788 1790 assert_lock_strong(bitMapLock());
1789 1791 // We need to lock the Free list lock for the space that we are
1790 1792 // currently collecting.
1791 1793 assert(haveFreelistLocks(), "Must be holding free list locks");
1792 1794 bitMapLock()->unlock();
1793 1795 releaseFreelistLocks();
1794 1796 {
1795 1797 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1796 1798 if (_foregroundGCShouldWait) {
1797 1799 // We are going to be waiting for action for the CMS thread;
1798 1800 // it had better not be gone (for instance at shutdown)!
1799 1801 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1800 1802 "CMS thread must be running");
1801 1803 // Wait here until the background collector gives us the go-ahead
1802 1804 ConcurrentMarkSweepThread::clear_CMS_flag(
1803 1805 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1804 1806 // Get a possibly blocked CMS thread going:
1805 1807 // Note that we set _foregroundGCIsActive true above,
1806 1808 // without protection of the CGC_lock.
1807 1809 CGC_lock->notify();
1808 1810 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1809 1811 "Possible deadlock");
1810 1812 while (_foregroundGCShouldWait) {
1811 1813 // wait for notification
1812 1814 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1813 1815 // Possibility of delay/starvation here, since CMS token does
1814 1816 // not know to give priority to VM thread? Actually, i think
1815 1817 // there wouldn't be any delay/starvation, but the proof of
1816 1818 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1817 1819 }
1818 1820 ConcurrentMarkSweepThread::set_CMS_flag(
1819 1821 ConcurrentMarkSweepThread::CMS_vm_has_token);
1820 1822 }
1821 1823 }
1822 1824 // The CMS_token is already held. Get back the other locks.
1823 1825 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1824 1826 "VM thread should have CMS token");
1825 1827 getFreelistLocks();
1826 1828 bitMapLock()->lock_without_safepoint_check();
1827 1829 if (TraceCMSState) {
1828 1830 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1829 1831 INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1830 1832 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1831 1833 }
1832 1834
1833 1835 // Check if we need to do a compaction, or if not, whether
1834 1836 // we need to start the mark-sweep from scratch.
1835 1837 bool should_compact = false;
1836 1838 bool should_start_over = false;
1837 1839 decide_foreground_collection_type(clear_all_soft_refs,
1838 1840 &should_compact, &should_start_over);
1839 1841
1840 1842 NOT_PRODUCT(
1841 1843 if (RotateCMSCollectionTypes) {
1842 1844 if (_cmsGen->debug_collection_type() ==
1843 1845 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1844 1846 should_compact = true;
1845 1847 } else if (_cmsGen->debug_collection_type() ==
1846 1848 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1847 1849 should_compact = false;
1848 1850 }
1849 1851 }
1850 1852 )
1851 1853
1852 1854 if (PrintGCDetails && first_state > Idling) {
1853 1855 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1854 1856 if (GCCause::is_user_requested_gc(cause) ||
1855 1857 GCCause::is_serviceability_requested_gc(cause)) {
1856 1858 gclog_or_tty->print(" (concurrent mode interrupted)");
1857 1859 } else {
1858 1860 gclog_or_tty->print(" (concurrent mode failure)");
1859 1861 }
1860 1862 }
1861 1863
1862 1864 if (should_compact) {
1863 1865 // If the collection is being acquired from the background
1864 1866 // collector, there may be references on the discovered
1865 1867 // references lists that have NULL referents (being those
1866 1868 // that were concurrently cleared by a mutator) or
1867 1869 // that are no longer active (having been enqueued concurrently
1868 1870 // by the mutator).
1869 1871 // Scrub the list of those references because Mark-Sweep-Compact
1870 1872 // code assumes referents are not NULL and that all discovered
1871 1873 // Reference objects are active.
1872 1874 ref_processor()->clean_up_discovered_references();
1873 1875
1874 1876 do_compaction_work(clear_all_soft_refs);
1875 1877
1876 1878 // Has the GC time limit been exceeded?
1877 1879 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration();
1878 1880 size_t max_eden_size = young_gen->max_capacity() -
1879 1881 young_gen->to()->capacity() -
1880 1882 young_gen->from()->capacity();
1881 1883 GenCollectedHeap* gch = GenCollectedHeap::heap();
1882 1884 GCCause::Cause gc_cause = gch->gc_cause();
1883 1885 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1884 1886 young_gen->eden()->used(),
1885 1887 _cmsGen->max_capacity(),
1886 1888 max_eden_size,
1887 1889 full,
1888 1890 gc_cause,
1889 1891 gch->collector_policy());
1890 1892 } else {
1891 1893 do_mark_sweep_work(clear_all_soft_refs, first_state,
1892 1894 should_start_over);
1893 1895 }
1894 1896 // Reset the expansion cause, now that we just completed
1895 1897 // a collection cycle.
1896 1898 clear_expansion_cause();
1897 1899 _foregroundGCIsActive = false;
1898 1900 return;
1899 1901 }
1900 1902
1901 1903 // Resize the perm generation and the tenured generation
1902 1904 // after obtaining the free list locks for the
1903 1905 // two generations.
1904 1906 void CMSCollector::compute_new_size() {
1905 1907 assert_locked_or_safepoint(Heap_lock);
1906 1908 FreelistLocker z(this);
1907 1909 _permGen->compute_new_size();
1908 1910 _cmsGen->compute_new_size();
1909 1911 }
1910 1912
1911 1913 // A work method used by foreground collection to determine
1912 1914 // what type of collection (compacting or not, continuing or fresh)
1913 1915 // it should do.
1914 1916 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1915 1917 // and CMSCompactWhenClearAllSoftRefs the default in the future
1916 1918 // and do away with the flags after a suitable period.
1917 1919 void CMSCollector::decide_foreground_collection_type(
1918 1920 bool clear_all_soft_refs, bool* should_compact,
1919 1921 bool* should_start_over) {
1920 1922 // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1921 1923 // flag is set, and we have either requested a System.gc() or
1922 1924 // the number of full gc's since the last concurrent cycle
1923 1925 // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1924 1926 // or if an incremental collection has failed
1925 1927 GenCollectedHeap* gch = GenCollectedHeap::heap();
1926 1928 assert(gch->collector_policy()->is_two_generation_policy(),
1927 1929 "You may want to check the correctness of the following");
1928 1930 // Inform cms gen if this was due to partial collection failing.
1929 1931 // The CMS gen may use this fact to determine its expansion policy.
1930 1932 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1931 1933 assert(!_cmsGen->incremental_collection_failed(),
1932 1934 "Should have been noticed, reacted to and cleared");
1933 1935 _cmsGen->set_incremental_collection_failed();
1934 1936 }
1935 1937 *should_compact =
1936 1938 UseCMSCompactAtFullCollection &&
1937 1939 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1938 1940 GCCause::is_user_requested_gc(gch->gc_cause()) ||
1939 1941 gch->incremental_collection_will_fail(true /* consult_young */));
1940 1942 *should_start_over = false;
1941 1943 if (clear_all_soft_refs && !*should_compact) {
1942 1944 // We are about to do a last ditch collection attempt
1943 1945 // so it would normally make sense to do a compaction
1944 1946 // to reclaim as much space as possible.
1945 1947 if (CMSCompactWhenClearAllSoftRefs) {
1946 1948 // Default: The rationale is that in this case either
1947 1949 // we are past the final marking phase, in which case
1948 1950 // we'd have to start over, or so little has been done
1949 1951 // that there's little point in saving that work. Compaction
1950 1952 // appears to be the sensible choice in either case.
1951 1953 *should_compact = true;
1952 1954 } else {
1953 1955 // We have been asked to clear all soft refs, but not to
1954 1956 // compact. Make sure that we aren't past the final checkpoint
1955 1957 // phase, for that is where we process soft refs. If we are already
1956 1958 // past that phase, we'll need to redo the refs discovery phase and
1957 1959 // if necessary clear soft refs that weren't previously
1958 1960 // cleared. We do so by remembering the phase in which
1959 1961 // we came in, and if we are past the refs processing
1960 1962 // phase, we'll choose to just redo the mark-sweep
1961 1963 // collection from scratch.
1962 1964 if (_collectorState > FinalMarking) {
1963 1965 // We are past the refs processing phase;
1964 1966 // start over and do a fresh synchronous CMS cycle
1965 1967 _collectorState = Resetting; // skip to reset to start new cycle
1966 1968 reset(false /* == !asynch */);
1967 1969 *should_start_over = true;
1968 1970 } // else we can continue a possibly ongoing current cycle
1969 1971 }
1970 1972 }
1971 1973 }
1972 1974
1973 1975 // A work method used by the foreground collector to do
1974 1976 // a mark-sweep-compact.
1975 1977 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1976 1978 GenCollectedHeap* gch = GenCollectedHeap::heap();
1977 1979 TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty);
1978 1980 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
1979 1981 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
1980 1982 "collections passed to foreground collector", _full_gcs_since_conc_gc);
1981 1983 }
1982 1984
1983 1985 // Sample collection interval time and reset for collection pause.
1984 1986 if (UseAdaptiveSizePolicy) {
1985 1987 size_policy()->msc_collection_begin();
1986 1988 }
1987 1989
1988 1990 // Temporarily widen the span of the weak reference processing to
1989 1991 // the entire heap.
1990 1992 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1991 1993 ReferenceProcessorSpanMutator x(ref_processor(), new_span);
1992 1994
1993 1995 // Temporarily, clear the "is_alive_non_header" field of the
1994 1996 // reference processor.
1995 1997 ReferenceProcessorIsAliveMutator y(ref_processor(), NULL);
1996 1998
1997 1999 // Temporarily make reference _processing_ single threaded (non-MT).
1998 2000 ReferenceProcessorMTProcMutator z(ref_processor(), false);
1999 2001
2000 2002 // Temporarily make refs discovery atomic
2001 2003 ReferenceProcessorAtomicMutator w(ref_processor(), true);
2002 2004
2003 2005 ref_processor()->set_enqueuing_is_done(false);
2004 2006 ref_processor()->enable_discovery();
2005 2007 ref_processor()->setup_policy(clear_all_soft_refs);
2006 2008 // If an asynchronous collection finishes, the _modUnionTable is
2007 2009 // all clear. If we are assuming the collection from an asynchronous
2008 2010 // collection, clear the _modUnionTable.
2009 2011 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
2010 2012 "_modUnionTable should be clear if the baton was not passed");
2011 2013 _modUnionTable.clear_all();
2012 2014
2013 2015 // We must adjust the allocation statistics being maintained
2014 2016 // in the free list space. We do so by reading and clearing
2015 2017 // the sweep timer and updating the block flux rate estimates below.
2016 2018 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
2017 2019 if (_inter_sweep_timer.is_active()) {
2018 2020 _inter_sweep_timer.stop();
2019 2021 // Note that we do not use this sample to update the _inter_sweep_estimate.
2020 2022 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
2021 2023 _inter_sweep_estimate.padded_average(),
2022 2024 _intra_sweep_estimate.padded_average());
2023 2025 }
2024 2026
2025 2027 {
2026 2028 TraceCMSMemoryManagerStats();
2027 2029 }
2028 2030 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
2029 2031 ref_processor(), clear_all_soft_refs);
2030 2032 #ifdef ASSERT
2031 2033 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
2032 2034 size_t free_size = cms_space->free();
2033 2035 assert(free_size ==
2034 2036 pointer_delta(cms_space->end(), cms_space->compaction_top())
2035 2037 * HeapWordSize,
2036 2038 "All the free space should be compacted into one chunk at top");
2037 2039 assert(cms_space->dictionary()->totalChunkSize(
2038 2040 debug_only(cms_space->freelistLock())) == 0 ||
2039 2041 cms_space->totalSizeInIndexedFreeLists() == 0,
2040 2042 "All the free space should be in a single chunk");
2041 2043 size_t num = cms_space->totalCount();
2042 2044 assert((free_size == 0 && num == 0) ||
2043 2045 (free_size > 0 && (num == 1 || num == 2)),
2044 2046 "There should be at most 2 free chunks after compaction");
2045 2047 #endif // ASSERT
2046 2048 _collectorState = Resetting;
2047 2049 assert(_restart_addr == NULL,
2048 2050 "Should have been NULL'd before baton was passed");
2049 2051 reset(false /* == !asynch */);
2050 2052 _cmsGen->reset_after_compaction();
2051 2053 _concurrent_cycles_since_last_unload = 0;
2052 2054
2053 2055 if (verifying() && !should_unload_classes()) {
2054 2056 perm_gen_verify_bit_map()->clear_all();
2055 2057 }
2056 2058
2057 2059 // Clear any data recorded in the PLAB chunk arrays.
2058 2060 if (_survivor_plab_array != NULL) {
2059 2061 reset_survivor_plab_arrays();
2060 2062 }
2061 2063
2062 2064 // Adjust the per-size allocation stats for the next epoch.
2063 2065 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
2064 2066 // Restart the "inter sweep timer" for the next epoch.
2065 2067 _inter_sweep_timer.reset();
2066 2068 _inter_sweep_timer.start();
2067 2069
2068 2070 // Sample collection pause time and reset for collection interval.
2069 2071 if (UseAdaptiveSizePolicy) {
2070 2072 size_policy()->msc_collection_end(gch->gc_cause());
2071 2073 }
2072 2074
2073 2075 // For a mark-sweep-compact, compute_new_size() will be called
2074 2076 // in the heap's do_collection() method.
2075 2077 }
2076 2078
2077 2079 // A work method used by the foreground collector to do
2078 2080 // a mark-sweep, after taking over from a possibly on-going
2079 2081 // concurrent mark-sweep collection.
2080 2082 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2081 2083 CollectorState first_state, bool should_start_over) {
2082 2084 if (PrintGC && Verbose) {
2083 2085 gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2084 2086 "collector with count %d",
2085 2087 _full_gcs_since_conc_gc);
2086 2088 }
2087 2089 switch (_collectorState) {
2088 2090 case Idling:
2089 2091 if (first_state == Idling || should_start_over) {
2090 2092 // The background GC was not active, or should
2091 2093 // restarted from scratch; start the cycle.
2092 2094 _collectorState = InitialMarking;
2093 2095 }
2094 2096 // If first_state was not Idling, then a background GC
2095 2097 // was in progress and has now finished. No need to do it
2096 2098 // again. Leave the state as Idling.
2097 2099 break;
2098 2100 case Precleaning:
2099 2101 // In the foreground case don't do the precleaning since
2100 2102 // it is not done concurrently and there is extra work
2101 2103 // required.
2102 2104 _collectorState = FinalMarking;
2103 2105 }
2104 2106 if (PrintGCDetails &&
2105 2107 (_collectorState > Idling ||
2106 2108 !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) {
2107 2109 gclog_or_tty->print(" (concurrent mode failure)");
2108 2110 }
2109 2111 collect_in_foreground(clear_all_soft_refs);
2110 2112
2111 2113 // For a mark-sweep, compute_new_size() will be called
2112 2114 // in the heap's do_collection() method.
2113 2115 }
2114 2116
2115 2117
2116 2118 void CMSCollector::getFreelistLocks() const {
2117 2119 // Get locks for all free lists in all generations that this
2118 2120 // collector is responsible for
2119 2121 _cmsGen->freelistLock()->lock_without_safepoint_check();
2120 2122 _permGen->freelistLock()->lock_without_safepoint_check();
2121 2123 }
2122 2124
2123 2125 void CMSCollector::releaseFreelistLocks() const {
2124 2126 // Release locks for all free lists in all generations that this
2125 2127 // collector is responsible for
2126 2128 _cmsGen->freelistLock()->unlock();
2127 2129 _permGen->freelistLock()->unlock();
2128 2130 }
2129 2131
2130 2132 bool CMSCollector::haveFreelistLocks() const {
2131 2133 // Check locks for all free lists in all generations that this
2132 2134 // collector is responsible for
2133 2135 assert_lock_strong(_cmsGen->freelistLock());
2134 2136 assert_lock_strong(_permGen->freelistLock());
2135 2137 PRODUCT_ONLY(ShouldNotReachHere());
2136 2138 return true;
2137 2139 }
2138 2140
2139 2141 // A utility class that is used by the CMS collector to
2140 2142 // temporarily "release" the foreground collector from its
2141 2143 // usual obligation to wait for the background collector to
2142 2144 // complete an ongoing phase before proceeding.
2143 2145 class ReleaseForegroundGC: public StackObj {
2144 2146 private:
2145 2147 CMSCollector* _c;
2146 2148 public:
2147 2149 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2148 2150 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2149 2151 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2150 2152 // allow a potentially blocked foreground collector to proceed
2151 2153 _c->_foregroundGCShouldWait = false;
2152 2154 if (_c->_foregroundGCIsActive) {
2153 2155 CGC_lock->notify();
2154 2156 }
2155 2157 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2156 2158 "Possible deadlock");
2157 2159 }
2158 2160
2159 2161 ~ReleaseForegroundGC() {
2160 2162 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2161 2163 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2162 2164 _c->_foregroundGCShouldWait = true;
2163 2165 }
2164 2166 };
2165 2167
2166 2168 // There are separate collect_in_background and collect_in_foreground because of
2167 2169 // the different locking requirements of the background collector and the
2168 2170 // foreground collector. There was originally an attempt to share
2169 2171 // one "collect" method between the background collector and the foreground
2170 2172 // collector but the if-then-else required made it cleaner to have
2171 2173 // separate methods.
2172 2174 void CMSCollector::collect_in_background(bool clear_all_soft_refs) {
2173 2175 assert(Thread::current()->is_ConcurrentGC_thread(),
2174 2176 "A CMS asynchronous collection is only allowed on a CMS thread.");
2175 2177
2176 2178 GenCollectedHeap* gch = GenCollectedHeap::heap();
2177 2179 {
2178 2180 bool safepoint_check = Mutex::_no_safepoint_check_flag;
2179 2181 MutexLockerEx hl(Heap_lock, safepoint_check);
2180 2182 FreelistLocker fll(this);
2181 2183 MutexLockerEx x(CGC_lock, safepoint_check);
2182 2184 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2183 2185 // The foreground collector is active or we're
2184 2186 // not using asynchronous collections. Skip this
2185 2187 // background collection.
2186 2188 assert(!_foregroundGCShouldWait, "Should be clear");
2187 2189 return;
2188 2190 } else {
2189 2191 assert(_collectorState == Idling, "Should be idling before start.");
2190 2192 _collectorState = InitialMarking;
2191 2193 // Reset the expansion cause, now that we are about to begin
2192 2194 // a new cycle.
2193 2195 clear_expansion_cause();
2194 2196 }
2195 2197 // Decide if we want to enable class unloading as part of the
2196 2198 // ensuing concurrent GC cycle.
2197 2199 update_should_unload_classes();
2198 2200 _full_gc_requested = false; // acks all outstanding full gc requests
2199 2201 // Signal that we are about to start a collection
2200 2202 gch->increment_total_full_collections(); // ... starting a collection cycle
2201 2203 _collection_count_start = gch->total_full_collections();
2202 2204 }
2203 2205
2204 2206 // Used for PrintGC
2205 2207 size_t prev_used;
2206 2208 if (PrintGC && Verbose) {
2207 2209 prev_used = _cmsGen->used(); // XXXPERM
2208 2210 }
2209 2211
2210 2212 // The change of the collection state is normally done at this level;
2211 2213 // the exceptions are phases that are executed while the world is
2212 2214 // stopped. For those phases the change of state is done while the
2213 2215 // world is stopped. For baton passing purposes this allows the
2214 2216 // background collector to finish the phase and change state atomically.
2215 2217 // The foreground collector cannot wait on a phase that is done
2216 2218 // while the world is stopped because the foreground collector already
2217 2219 // has the world stopped and would deadlock.
2218 2220 while (_collectorState != Idling) {
2219 2221 if (TraceCMSState) {
2220 2222 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2221 2223 Thread::current(), _collectorState);
2222 2224 }
2223 2225 // The foreground collector
2224 2226 // holds the Heap_lock throughout its collection.
2225 2227 // holds the CMS token (but not the lock)
2226 2228 // except while it is waiting for the background collector to yield.
2227 2229 //
2228 2230 // The foreground collector should be blocked (not for long)
2229 2231 // if the background collector is about to start a phase
2230 2232 // executed with world stopped. If the background
2231 2233 // collector has already started such a phase, the
2232 2234 // foreground collector is blocked waiting for the
2233 2235 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
2234 2236 // are executed in the VM thread.
2235 2237 //
2236 2238 // The locking order is
2237 2239 // PendingListLock (PLL) -- if applicable (FinalMarking)
2238 2240 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
2239 2241 // CMS token (claimed in
2240 2242 // stop_world_and_do() -->
2241 2243 // safepoint_synchronize() -->
2242 2244 // CMSThread::synchronize())
2243 2245
2244 2246 {
2245 2247 // Check if the FG collector wants us to yield.
2246 2248 CMSTokenSync x(true); // is cms thread
2247 2249 if (waitForForegroundGC()) {
2248 2250 // We yielded to a foreground GC, nothing more to be
2249 2251 // done this round.
2250 2252 assert(_foregroundGCShouldWait == false, "We set it to false in "
2251 2253 "waitForForegroundGC()");
2252 2254 if (TraceCMSState) {
2253 2255 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2254 2256 " exiting collection CMS state %d",
2255 2257 Thread::current(), _collectorState);
2256 2258 }
2257 2259 return;
2258 2260 } else {
2259 2261 // The background collector can run but check to see if the
2260 2262 // foreground collector has done a collection while the
2261 2263 // background collector was waiting to get the CGC_lock
2262 2264 // above. If yes, break so that _foregroundGCShouldWait
2263 2265 // is cleared before returning.
2264 2266 if (_collectorState == Idling) {
2265 2267 break;
2266 2268 }
2267 2269 }
2268 2270 }
2269 2271
2270 2272 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2271 2273 "should be waiting");
2272 2274
2273 2275 switch (_collectorState) {
2274 2276 case InitialMarking:
2275 2277 {
2276 2278 ReleaseForegroundGC x(this);
2277 2279 stats().record_cms_begin();
2278 2280
2279 2281 VM_CMS_Initial_Mark initial_mark_op(this);
2280 2282 VMThread::execute(&initial_mark_op);
2281 2283 }
2282 2284 // The collector state may be any legal state at this point
2283 2285 // since the background collector may have yielded to the
2284 2286 // foreground collector.
2285 2287 break;
2286 2288 case Marking:
2287 2289 // initial marking in checkpointRootsInitialWork has been completed
2288 2290 if (markFromRoots(true)) { // we were successful
2289 2291 assert(_collectorState == Precleaning, "Collector state should "
2290 2292 "have changed");
2291 2293 } else {
2292 2294 assert(_foregroundGCIsActive, "Internal state inconsistency");
2293 2295 }
2294 2296 break;
2295 2297 case Precleaning:
2296 2298 if (UseAdaptiveSizePolicy) {
2297 2299 size_policy()->concurrent_precleaning_begin();
2298 2300 }
2299 2301 // marking from roots in markFromRoots has been completed
2300 2302 preclean();
2301 2303 if (UseAdaptiveSizePolicy) {
2302 2304 size_policy()->concurrent_precleaning_end();
2303 2305 }
2304 2306 assert(_collectorState == AbortablePreclean ||
2305 2307 _collectorState == FinalMarking,
2306 2308 "Collector state should have changed");
2307 2309 break;
2308 2310 case AbortablePreclean:
2309 2311 if (UseAdaptiveSizePolicy) {
2310 2312 size_policy()->concurrent_phases_resume();
2311 2313 }
2312 2314 abortable_preclean();
2313 2315 if (UseAdaptiveSizePolicy) {
2314 2316 size_policy()->concurrent_precleaning_end();
2315 2317 }
2316 2318 assert(_collectorState == FinalMarking, "Collector state should "
2317 2319 "have changed");
2318 2320 break;
2319 2321 case FinalMarking:
2320 2322 {
2321 2323 ReleaseForegroundGC x(this);
2322 2324
2323 2325 VM_CMS_Final_Remark final_remark_op(this);
2324 2326 VMThread::execute(&final_remark_op);
2325 2327 }
2326 2328 assert(_foregroundGCShouldWait, "block post-condition");
2327 2329 break;
2328 2330 case Sweeping:
2329 2331 if (UseAdaptiveSizePolicy) {
2330 2332 size_policy()->concurrent_sweeping_begin();
2331 2333 }
2332 2334 // final marking in checkpointRootsFinal has been completed
2333 2335 sweep(true);
2334 2336 assert(_collectorState == Resizing, "Collector state change "
2335 2337 "to Resizing must be done under the free_list_lock");
2336 2338 _full_gcs_since_conc_gc = 0;
2337 2339
2338 2340 // Stop the timers for adaptive size policy for the concurrent phases
2339 2341 if (UseAdaptiveSizePolicy) {
2340 2342 size_policy()->concurrent_sweeping_end();
2341 2343 size_policy()->concurrent_phases_end(gch->gc_cause(),
2342 2344 gch->prev_gen(_cmsGen)->capacity(),
2343 2345 _cmsGen->free());
2344 2346 }
2345 2347
2346 2348 case Resizing: {
2347 2349 // Sweeping has been completed...
2348 2350 // At this point the background collection has completed.
2349 2351 // Don't move the call to compute_new_size() down
2350 2352 // into code that might be executed if the background
2351 2353 // collection was preempted.
2352 2354 {
2353 2355 ReleaseForegroundGC x(this); // unblock FG collection
2354 2356 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
2355 2357 CMSTokenSync z(true); // not strictly needed.
2356 2358 if (_collectorState == Resizing) {
2357 2359 compute_new_size();
2358 2360 _collectorState = Resetting;
2359 2361 } else {
2360 2362 assert(_collectorState == Idling, "The state should only change"
2361 2363 " because the foreground collector has finished the collection");
2362 2364 }
2363 2365 }
2364 2366 break;
2365 2367 }
2366 2368 case Resetting:
2367 2369 // CMS heap resizing has been completed
2368 2370 reset(true);
2369 2371 assert(_collectorState == Idling, "Collector state should "
2370 2372 "have changed");
2371 2373 stats().record_cms_end();
2372 2374 // Don't move the concurrent_phases_end() and compute_new_size()
2373 2375 // calls to here because a preempted background collection
2374 2376 // has it's state set to "Resetting".
2375 2377 break;
2376 2378 case Idling:
2377 2379 default:
2378 2380 ShouldNotReachHere();
2379 2381 break;
2380 2382 }
2381 2383 if (TraceCMSState) {
2382 2384 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2383 2385 Thread::current(), _collectorState);
2384 2386 }
2385 2387 assert(_foregroundGCShouldWait, "block post-condition");
2386 2388 }
2387 2389
2388 2390 // Should this be in gc_epilogue?
2389 2391 collector_policy()->counters()->update_counters();
2390 2392
2391 2393 {
2392 2394 // Clear _foregroundGCShouldWait and, in the event that the
2393 2395 // foreground collector is waiting, notify it, before
2394 2396 // returning.
2395 2397 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2396 2398 _foregroundGCShouldWait = false;
2397 2399 if (_foregroundGCIsActive) {
2398 2400 CGC_lock->notify();
2399 2401 }
2400 2402 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2401 2403 "Possible deadlock");
2402 2404 }
2403 2405 if (TraceCMSState) {
2404 2406 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2405 2407 " exiting collection CMS state %d",
2406 2408 Thread::current(), _collectorState);
2407 2409 }
2408 2410 if (PrintGC && Verbose) {
2409 2411 _cmsGen->print_heap_change(prev_used);
2410 2412 }
2411 2413 }
2412 2414
2413 2415 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) {
2414 2416 assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2415 2417 "Foreground collector should be waiting, not executing");
2416 2418 assert(Thread::current()->is_VM_thread(), "A foreground collection"
2417 2419 "may only be done by the VM Thread with the world stopped");
2418 2420 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2419 2421 "VM thread should have CMS token");
2420 2422
2421 2423 NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2422 2424 true, gclog_or_tty);)
2423 2425 if (UseAdaptiveSizePolicy) {
2424 2426 size_policy()->ms_collection_begin();
2425 2427 }
2426 2428 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2427 2429
2428 2430 HandleMark hm; // Discard invalid handles created during verification
2429 2431
2430 2432 if (VerifyBeforeGC &&
2431 2433 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2432 2434 Universe::verify(true);
2433 2435 }
2434 2436
2435 2437 // Snapshot the soft reference policy to be used in this collection cycle.
2436 2438 ref_processor()->setup_policy(clear_all_soft_refs);
2437 2439
2438 2440 bool init_mark_was_synchronous = false; // until proven otherwise
2439 2441 while (_collectorState != Idling) {
2440 2442 if (TraceCMSState) {
2441 2443 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2442 2444 Thread::current(), _collectorState);
2443 2445 }
2444 2446 switch (_collectorState) {
2445 2447 case InitialMarking:
2446 2448 init_mark_was_synchronous = true; // fact to be exploited in re-mark
2447 2449 checkpointRootsInitial(false);
2448 2450 assert(_collectorState == Marking, "Collector state should have changed"
2449 2451 " within checkpointRootsInitial()");
2450 2452 break;
2451 2453 case Marking:
2452 2454 // initial marking in checkpointRootsInitialWork has been completed
2453 2455 if (VerifyDuringGC &&
2454 2456 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2455 2457 gclog_or_tty->print("Verify before initial mark: ");
2456 2458 Universe::verify(true);
2457 2459 }
2458 2460 {
2459 2461 bool res = markFromRoots(false);
2460 2462 assert(res && _collectorState == FinalMarking, "Collector state should "
2461 2463 "have changed");
2462 2464 break;
2463 2465 }
2464 2466 case FinalMarking:
2465 2467 if (VerifyDuringGC &&
2466 2468 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2467 2469 gclog_or_tty->print("Verify before re-mark: ");
2468 2470 Universe::verify(true);
2469 2471 }
2470 2472 checkpointRootsFinal(false, clear_all_soft_refs,
2471 2473 init_mark_was_synchronous);
2472 2474 assert(_collectorState == Sweeping, "Collector state should not "
2473 2475 "have changed within checkpointRootsFinal()");
2474 2476 break;
2475 2477 case Sweeping:
2476 2478 // final marking in checkpointRootsFinal has been completed
2477 2479 if (VerifyDuringGC &&
2478 2480 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2479 2481 gclog_or_tty->print("Verify before sweep: ");
2480 2482 Universe::verify(true);
2481 2483 }
2482 2484 sweep(false);
2483 2485 assert(_collectorState == Resizing, "Incorrect state");
2484 2486 break;
2485 2487 case Resizing: {
2486 2488 // Sweeping has been completed; the actual resize in this case
2487 2489 // is done separately; nothing to be done in this state.
2488 2490 _collectorState = Resetting;
2489 2491 break;
2490 2492 }
2491 2493 case Resetting:
2492 2494 // The heap has been resized.
2493 2495 if (VerifyDuringGC &&
2494 2496 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2495 2497 gclog_or_tty->print("Verify before reset: ");
2496 2498 Universe::verify(true);
2497 2499 }
2498 2500 reset(false);
2499 2501 assert(_collectorState == Idling, "Collector state should "
2500 2502 "have changed");
2501 2503 break;
2502 2504 case Precleaning:
2503 2505 case AbortablePreclean:
2504 2506 // Elide the preclean phase
2505 2507 _collectorState = FinalMarking;
2506 2508 break;
2507 2509 default:
2508 2510 ShouldNotReachHere();
2509 2511 }
2510 2512 if (TraceCMSState) {
2511 2513 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2512 2514 Thread::current(), _collectorState);
2513 2515 }
2514 2516 }
2515 2517
2516 2518 if (UseAdaptiveSizePolicy) {
2517 2519 GenCollectedHeap* gch = GenCollectedHeap::heap();
2518 2520 size_policy()->ms_collection_end(gch->gc_cause());
2519 2521 }
2520 2522
2521 2523 if (VerifyAfterGC &&
2522 2524 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2523 2525 Universe::verify(true);
2524 2526 }
2525 2527 if (TraceCMSState) {
2526 2528 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2527 2529 " exiting collection CMS state %d",
2528 2530 Thread::current(), _collectorState);
2529 2531 }
2530 2532 }
2531 2533
2532 2534 bool CMSCollector::waitForForegroundGC() {
2533 2535 bool res = false;
2534 2536 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2535 2537 "CMS thread should have CMS token");
2536 2538 // Block the foreground collector until the
2537 2539 // background collectors decides whether to
2538 2540 // yield.
2539 2541 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2540 2542 _foregroundGCShouldWait = true;
2541 2543 if (_foregroundGCIsActive) {
2542 2544 // The background collector yields to the
2543 2545 // foreground collector and returns a value
2544 2546 // indicating that it has yielded. The foreground
2545 2547 // collector can proceed.
2546 2548 res = true;
2547 2549 _foregroundGCShouldWait = false;
2548 2550 ConcurrentMarkSweepThread::clear_CMS_flag(
2549 2551 ConcurrentMarkSweepThread::CMS_cms_has_token);
2550 2552 ConcurrentMarkSweepThread::set_CMS_flag(
2551 2553 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2552 2554 // Get a possibly blocked foreground thread going
2553 2555 CGC_lock->notify();
2554 2556 if (TraceCMSState) {
2555 2557 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2556 2558 Thread::current(), _collectorState);
2557 2559 }
2558 2560 while (_foregroundGCIsActive) {
2559 2561 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2560 2562 }
2561 2563 ConcurrentMarkSweepThread::set_CMS_flag(
2562 2564 ConcurrentMarkSweepThread::CMS_cms_has_token);
2563 2565 ConcurrentMarkSweepThread::clear_CMS_flag(
2564 2566 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2565 2567 }
2566 2568 if (TraceCMSState) {
2567 2569 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2568 2570 Thread::current(), _collectorState);
2569 2571 }
2570 2572 return res;
2571 2573 }
2572 2574
2573 2575 // Because of the need to lock the free lists and other structures in
2574 2576 // the collector, common to all the generations that the collector is
2575 2577 // collecting, we need the gc_prologues of individual CMS generations
2576 2578 // delegate to their collector. It may have been simpler had the
2577 2579 // current infrastructure allowed one to call a prologue on a
2578 2580 // collector. In the absence of that we have the generation's
2579 2581 // prologue delegate to the collector, which delegates back
2580 2582 // some "local" work to a worker method in the individual generations
2581 2583 // that it's responsible for collecting, while itself doing any
2582 2584 // work common to all generations it's responsible for. A similar
2583 2585 // comment applies to the gc_epilogue()'s.
2584 2586 // The role of the varaible _between_prologue_and_epilogue is to
2585 2587 // enforce the invocation protocol.
2586 2588 void CMSCollector::gc_prologue(bool full) {
2587 2589 // Call gc_prologue_work() for each CMSGen and PermGen that
2588 2590 // we are responsible for.
2589 2591
2590 2592 // The following locking discipline assumes that we are only called
2591 2593 // when the world is stopped.
2592 2594 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2593 2595
2594 2596 // The CMSCollector prologue must call the gc_prologues for the
2595 2597 // "generations" (including PermGen if any) that it's responsible
2596 2598 // for.
2597 2599
2598 2600 assert( Thread::current()->is_VM_thread()
2599 2601 || ( CMSScavengeBeforeRemark
2600 2602 && Thread::current()->is_ConcurrentGC_thread()),
2601 2603 "Incorrect thread type for prologue execution");
2602 2604
2603 2605 if (_between_prologue_and_epilogue) {
2604 2606 // We have already been invoked; this is a gc_prologue delegation
2605 2607 // from yet another CMS generation that we are responsible for, just
2606 2608 // ignore it since all relevant work has already been done.
2607 2609 return;
2608 2610 }
2609 2611
2610 2612 // set a bit saying prologue has been called; cleared in epilogue
2611 2613 _between_prologue_and_epilogue = true;
2612 2614 // Claim locks for common data structures, then call gc_prologue_work()
2613 2615 // for each CMSGen and PermGen that we are responsible for.
2614 2616
2615 2617 getFreelistLocks(); // gets free list locks on constituent spaces
2616 2618 bitMapLock()->lock_without_safepoint_check();
2617 2619
2618 2620 // Should call gc_prologue_work() for all cms gens we are responsible for
2619 2621 bool registerClosure = _collectorState >= Marking
2620 2622 && _collectorState < Sweeping;
2621 2623 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ?
2622 2624 &_modUnionClosurePar
2623 2625 : &_modUnionClosure;
2624 2626 _cmsGen->gc_prologue_work(full, registerClosure, muc);
2625 2627 _permGen->gc_prologue_work(full, registerClosure, muc);
2626 2628
2627 2629 if (!full) {
2628 2630 stats().record_gc0_begin();
2629 2631 }
2630 2632 }
2631 2633
2632 2634 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2633 2635 // Delegate to CMScollector which knows how to coordinate between
2634 2636 // this and any other CMS generations that it is responsible for
2635 2637 // collecting.
2636 2638 collector()->gc_prologue(full);
2637 2639 }
2638 2640
2639 2641 // This is a "private" interface for use by this generation's CMSCollector.
2640 2642 // Not to be called directly by any other entity (for instance,
2641 2643 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2642 2644 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2643 2645 bool registerClosure, ModUnionClosure* modUnionClosure) {
2644 2646 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2645 2647 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2646 2648 "Should be NULL");
2647 2649 if (registerClosure) {
2648 2650 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2649 2651 }
2650 2652 cmsSpace()->gc_prologue();
2651 2653 // Clear stat counters
2652 2654 NOT_PRODUCT(
2653 2655 assert(_numObjectsPromoted == 0, "check");
2654 2656 assert(_numWordsPromoted == 0, "check");
2655 2657 if (Verbose && PrintGC) {
2656 2658 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2657 2659 SIZE_FORMAT" bytes concurrently",
2658 2660 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2659 2661 }
2660 2662 _numObjectsAllocated = 0;
2661 2663 _numWordsAllocated = 0;
2662 2664 )
2663 2665 }
2664 2666
2665 2667 void CMSCollector::gc_epilogue(bool full) {
2666 2668 // The following locking discipline assumes that we are only called
2667 2669 // when the world is stopped.
2668 2670 assert(SafepointSynchronize::is_at_safepoint(),
2669 2671 "world is stopped assumption");
2670 2672
2671 2673 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2672 2674 // if linear allocation blocks need to be appropriately marked to allow the
2673 2675 // the blocks to be parsable. We also check here whether we need to nudge the
2674 2676 // CMS collector thread to start a new cycle (if it's not already active).
2675 2677 assert( Thread::current()->is_VM_thread()
2676 2678 || ( CMSScavengeBeforeRemark
2677 2679 && Thread::current()->is_ConcurrentGC_thread()),
2678 2680 "Incorrect thread type for epilogue execution");
2679 2681
2680 2682 if (!_between_prologue_and_epilogue) {
2681 2683 // We have already been invoked; this is a gc_epilogue delegation
2682 2684 // from yet another CMS generation that we are responsible for, just
2683 2685 // ignore it since all relevant work has already been done.
2684 2686 return;
2685 2687 }
2686 2688 assert(haveFreelistLocks(), "must have freelist locks");
2687 2689 assert_lock_strong(bitMapLock());
2688 2690
2689 2691 _cmsGen->gc_epilogue_work(full);
2690 2692 _permGen->gc_epilogue_work(full);
2691 2693
2692 2694 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2693 2695 // in case sampling was not already enabled, enable it
2694 2696 _start_sampling = true;
2695 2697 }
2696 2698 // reset _eden_chunk_array so sampling starts afresh
2697 2699 _eden_chunk_index = 0;
2698 2700
2699 2701 size_t cms_used = _cmsGen->cmsSpace()->used();
2700 2702 size_t perm_used = _permGen->cmsSpace()->used();
2701 2703
2702 2704 // update performance counters - this uses a special version of
2703 2705 // update_counters() that allows the utilization to be passed as a
2704 2706 // parameter, avoiding multiple calls to used().
2705 2707 //
2706 2708 _cmsGen->update_counters(cms_used);
2707 2709 _permGen->update_counters(perm_used);
2708 2710
2709 2711 if (CMSIncrementalMode) {
2710 2712 icms_update_allocation_limits();
2711 2713 }
2712 2714
2713 2715 bitMapLock()->unlock();
2714 2716 releaseFreelistLocks();
2715 2717
2716 2718 _between_prologue_and_epilogue = false; // ready for next cycle
2717 2719 }
2718 2720
2719 2721 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2720 2722 collector()->gc_epilogue(full);
2721 2723
2722 2724 // Also reset promotion tracking in par gc thread states.
2723 2725 if (CollectedHeap::use_parallel_gc_threads()) {
2724 2726 for (uint i = 0; i < ParallelGCThreads; i++) {
2725 2727 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2726 2728 }
2727 2729 }
2728 2730 }
2729 2731
2730 2732 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2731 2733 assert(!incremental_collection_failed(), "Should have been cleared");
2732 2734 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2733 2735 cmsSpace()->gc_epilogue();
2734 2736 // Print stat counters
2735 2737 NOT_PRODUCT(
2736 2738 assert(_numObjectsAllocated == 0, "check");
2737 2739 assert(_numWordsAllocated == 0, "check");
2738 2740 if (Verbose && PrintGC) {
2739 2741 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2740 2742 SIZE_FORMAT" bytes",
2741 2743 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2742 2744 }
2743 2745 _numObjectsPromoted = 0;
2744 2746 _numWordsPromoted = 0;
2745 2747 )
2746 2748
2747 2749 if (PrintGC && Verbose) {
2748 2750 // Call down the chain in contiguous_available needs the freelistLock
2749 2751 // so print this out before releasing the freeListLock.
2750 2752 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2751 2753 contiguous_available());
2752 2754 }
2753 2755 }
2754 2756
2755 2757 #ifndef PRODUCT
2756 2758 bool CMSCollector::have_cms_token() {
2757 2759 Thread* thr = Thread::current();
2758 2760 if (thr->is_VM_thread()) {
2759 2761 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2760 2762 } else if (thr->is_ConcurrentGC_thread()) {
2761 2763 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2762 2764 } else if (thr->is_GC_task_thread()) {
2763 2765 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2764 2766 ParGCRareEvent_lock->owned_by_self();
2765 2767 }
2766 2768 return false;
2767 2769 }
2768 2770 #endif
2769 2771
2770 2772 // Check reachability of the given heap address in CMS generation,
2771 2773 // treating all other generations as roots.
2772 2774 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2773 2775 // We could "guarantee" below, rather than assert, but i'll
2774 2776 // leave these as "asserts" so that an adventurous debugger
2775 2777 // could try this in the product build provided some subset of
2776 2778 // the conditions were met, provided they were intersted in the
2777 2779 // results and knew that the computation below wouldn't interfere
2778 2780 // with other concurrent computations mutating the structures
2779 2781 // being read or written.
2780 2782 assert(SafepointSynchronize::is_at_safepoint(),
2781 2783 "Else mutations in object graph will make answer suspect");
2782 2784 assert(have_cms_token(), "Should hold cms token");
2783 2785 assert(haveFreelistLocks(), "must hold free list locks");
2784 2786 assert_lock_strong(bitMapLock());
2785 2787
2786 2788 // Clear the marking bit map array before starting, but, just
2787 2789 // for kicks, first report if the given address is already marked
2788 2790 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
2789 2791 _markBitMap.isMarked(addr) ? "" : " not");
2790 2792
2791 2793 if (verify_after_remark()) {
2792 2794 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2793 2795 bool result = verification_mark_bm()->isMarked(addr);
2794 2796 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
2795 2797 result ? "IS" : "is NOT");
2796 2798 return result;
2797 2799 } else {
2798 2800 gclog_or_tty->print_cr("Could not compute result");
2799 2801 return false;
2800 2802 }
2801 2803 }
2802 2804
2803 2805 ////////////////////////////////////////////////////////
2804 2806 // CMS Verification Support
2805 2807 ////////////////////////////////////////////////////////
2806 2808 // Following the remark phase, the following invariant
2807 2809 // should hold -- each object in the CMS heap which is
2808 2810 // marked in markBitMap() should be marked in the verification_mark_bm().
2809 2811
2810 2812 class VerifyMarkedClosure: public BitMapClosure {
2811 2813 CMSBitMap* _marks;
2812 2814 bool _failed;
2813 2815
2814 2816 public:
2815 2817 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2816 2818
2817 2819 bool do_bit(size_t offset) {
2818 2820 HeapWord* addr = _marks->offsetToHeapWord(offset);
2819 2821 if (!_marks->isMarked(addr)) {
2820 2822 oop(addr)->print_on(gclog_or_tty);
2821 2823 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2822 2824 _failed = true;
2823 2825 }
2824 2826 return true;
2825 2827 }
2826 2828
2827 2829 bool failed() { return _failed; }
2828 2830 };
2829 2831
2830 2832 bool CMSCollector::verify_after_remark() {
2831 2833 gclog_or_tty->print(" [Verifying CMS Marking... ");
2832 2834 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2833 2835 static bool init = false;
2834 2836
2835 2837 assert(SafepointSynchronize::is_at_safepoint(),
2836 2838 "Else mutations in object graph will make answer suspect");
2837 2839 assert(have_cms_token(),
2838 2840 "Else there may be mutual interference in use of "
2839 2841 " verification data structures");
2840 2842 assert(_collectorState > Marking && _collectorState <= Sweeping,
2841 2843 "Else marking info checked here may be obsolete");
2842 2844 assert(haveFreelistLocks(), "must hold free list locks");
2843 2845 assert_lock_strong(bitMapLock());
2844 2846
2845 2847
2846 2848 // Allocate marking bit map if not already allocated
2847 2849 if (!init) { // first time
2848 2850 if (!verification_mark_bm()->allocate(_span)) {
2849 2851 return false;
2850 2852 }
2851 2853 init = true;
2852 2854 }
2853 2855
2854 2856 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2855 2857
2856 2858 // Turn off refs discovery -- so we will be tracing through refs.
2857 2859 // This is as intended, because by this time
2858 2860 // GC must already have cleared any refs that need to be cleared,
2859 2861 // and traced those that need to be marked; moreover,
2860 2862 // the marking done here is not going to intefere in any
2861 2863 // way with the marking information used by GC.
2862 2864 NoRefDiscovery no_discovery(ref_processor());
2863 2865
2864 2866 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2865 2867
2866 2868 // Clear any marks from a previous round
2867 2869 verification_mark_bm()->clear_all();
2868 2870 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2869 2871 verify_work_stacks_empty();
2870 2872
2871 2873 GenCollectedHeap* gch = GenCollectedHeap::heap();
2872 2874 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2873 2875 // Update the saved marks which may affect the root scans.
2874 2876 gch->save_marks();
2875 2877
2876 2878 if (CMSRemarkVerifyVariant == 1) {
2877 2879 // In this first variant of verification, we complete
2878 2880 // all marking, then check if the new marks-verctor is
2879 2881 // a subset of the CMS marks-vector.
2880 2882 verify_after_remark_work_1();
2881 2883 } else if (CMSRemarkVerifyVariant == 2) {
2882 2884 // In this second variant of verification, we flag an error
2883 2885 // (i.e. an object reachable in the new marks-vector not reachable
2884 2886 // in the CMS marks-vector) immediately, also indicating the
2885 2887 // identify of an object (A) that references the unmarked object (B) --
2886 2888 // presumably, a mutation to A failed to be picked up by preclean/remark?
2887 2889 verify_after_remark_work_2();
2888 2890 } else {
2889 2891 warning("Unrecognized value %d for CMSRemarkVerifyVariant",
2890 2892 CMSRemarkVerifyVariant);
2891 2893 }
2892 2894 gclog_or_tty->print(" done] ");
2893 2895 return true;
2894 2896 }
2895 2897
2896 2898 void CMSCollector::verify_after_remark_work_1() {
2897 2899 ResourceMark rm;
2898 2900 HandleMark hm;
2899 2901 GenCollectedHeap* gch = GenCollectedHeap::heap();
2900 2902
2901 2903 // Mark from roots one level into CMS
2902 2904 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2903 2905 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2904 2906
2905 2907 gch->gen_process_strong_roots(_cmsGen->level(),
2906 2908 true, // younger gens are roots
2907 2909 true, // activate StrongRootsScope
2908 2910 true, // collecting perm gen
2909 2911 SharedHeap::ScanningOption(roots_scanning_options()),
2910 2912 ¬Older,
2911 2913 true, // walk code active on stacks
2912 2914 NULL);
2913 2915
2914 2916 // Now mark from the roots
2915 2917 assert(_revisitStack.isEmpty(), "Should be empty");
2916 2918 MarkFromRootsClosure markFromRootsClosure(this, _span,
2917 2919 verification_mark_bm(), verification_mark_stack(), &_revisitStack,
2918 2920 false /* don't yield */, true /* verifying */);
2919 2921 assert(_restart_addr == NULL, "Expected pre-condition");
2920 2922 verification_mark_bm()->iterate(&markFromRootsClosure);
2921 2923 while (_restart_addr != NULL) {
2922 2924 // Deal with stack overflow: by restarting at the indicated
2923 2925 // address.
2924 2926 HeapWord* ra = _restart_addr;
2925 2927 markFromRootsClosure.reset(ra);
2926 2928 _restart_addr = NULL;
2927 2929 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2928 2930 }
2929 2931 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2930 2932 verify_work_stacks_empty();
2931 2933 // Should reset the revisit stack above, since no class tree
2932 2934 // surgery is forthcoming.
2933 2935 _revisitStack.reset(); // throwing away all contents
2934 2936
2935 2937 // Marking completed -- now verify that each bit marked in
2936 2938 // verification_mark_bm() is also marked in markBitMap(); flag all
2937 2939 // errors by printing corresponding objects.
2938 2940 VerifyMarkedClosure vcl(markBitMap());
2939 2941 verification_mark_bm()->iterate(&vcl);
2940 2942 if (vcl.failed()) {
2941 2943 gclog_or_tty->print("Verification failed");
2942 2944 Universe::heap()->print_on(gclog_or_tty);
2943 2945 fatal("CMS: failed marking verification after remark");
2944 2946 }
2945 2947 }
2946 2948
2947 2949 void CMSCollector::verify_after_remark_work_2() {
2948 2950 ResourceMark rm;
2949 2951 HandleMark hm;
2950 2952 GenCollectedHeap* gch = GenCollectedHeap::heap();
2951 2953
2952 2954 // Mark from roots one level into CMS
2953 2955 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2954 2956 markBitMap());
2955 2957 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2956 2958 gch->gen_process_strong_roots(_cmsGen->level(),
2957 2959 true, // younger gens are roots
2958 2960 true, // activate StrongRootsScope
2959 2961 true, // collecting perm gen
2960 2962 SharedHeap::ScanningOption(roots_scanning_options()),
2961 2963 ¬Older,
2962 2964 true, // walk code active on stacks
2963 2965 NULL);
2964 2966
2965 2967 // Now mark from the roots
2966 2968 assert(_revisitStack.isEmpty(), "Should be empty");
2967 2969 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2968 2970 verification_mark_bm(), markBitMap(), verification_mark_stack());
2969 2971 assert(_restart_addr == NULL, "Expected pre-condition");
2970 2972 verification_mark_bm()->iterate(&markFromRootsClosure);
2971 2973 while (_restart_addr != NULL) {
2972 2974 // Deal with stack overflow: by restarting at the indicated
2973 2975 // address.
2974 2976 HeapWord* ra = _restart_addr;
2975 2977 markFromRootsClosure.reset(ra);
2976 2978 _restart_addr = NULL;
2977 2979 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2978 2980 }
2979 2981 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2980 2982 verify_work_stacks_empty();
2981 2983 // Should reset the revisit stack above, since no class tree
2982 2984 // surgery is forthcoming.
2983 2985 _revisitStack.reset(); // throwing away all contents
2984 2986
2985 2987 // Marking completed -- now verify that each bit marked in
2986 2988 // verification_mark_bm() is also marked in markBitMap(); flag all
2987 2989 // errors by printing corresponding objects.
2988 2990 VerifyMarkedClosure vcl(markBitMap());
2989 2991 verification_mark_bm()->iterate(&vcl);
2990 2992 assert(!vcl.failed(), "Else verification above should not have succeeded");
2991 2993 }
2992 2994
2993 2995 void ConcurrentMarkSweepGeneration::save_marks() {
2994 2996 // delegate to CMS space
2995 2997 cmsSpace()->save_marks();
2996 2998 for (uint i = 0; i < ParallelGCThreads; i++) {
2997 2999 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2998 3000 }
2999 3001 }
3000 3002
3001 3003 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
3002 3004 return cmsSpace()->no_allocs_since_save_marks();
3003 3005 }
3004 3006
3005 3007 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
3006 3008 \
3007 3009 void ConcurrentMarkSweepGeneration:: \
3008 3010 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
3009 3011 cl->set_generation(this); \
3010 3012 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
3011 3013 cl->reset_generation(); \
3012 3014 save_marks(); \
3013 3015 }
3014 3016
3015 3017 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
3016 3018
3017 3019 void
3018 3020 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk)
3019 3021 {
3020 3022 // Not currently implemented; need to do the following. -- ysr.
3021 3023 // dld -- I think that is used for some sort of allocation profiler. So it
3022 3024 // really means the objects allocated by the mutator since the last
3023 3025 // GC. We could potentially implement this cheaply by recording only
3024 3026 // the direct allocations in a side data structure.
3025 3027 //
3026 3028 // I think we probably ought not to be required to support these
3027 3029 // iterations at any arbitrary point; I think there ought to be some
3028 3030 // call to enable/disable allocation profiling in a generation/space,
3029 3031 // and the iterator ought to return the objects allocated in the
3030 3032 // gen/space since the enable call, or the last iterator call (which
3031 3033 // will probably be at a GC.) That way, for gens like CM&S that would
3032 3034 // require some extra data structure to support this, we only pay the
3033 3035 // cost when it's in use...
3034 3036 cmsSpace()->object_iterate_since_last_GC(blk);
3035 3037 }
3036 3038
3037 3039 void
3038 3040 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
3039 3041 cl->set_generation(this);
3040 3042 younger_refs_in_space_iterate(_cmsSpace, cl);
3041 3043 cl->reset_generation();
3042 3044 }
3043 3045
3044 3046 void
3045 3047 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) {
3046 3048 if (freelistLock()->owned_by_self()) {
3047 3049 Generation::oop_iterate(mr, cl);
3048 3050 } else {
3049 3051 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3050 3052 Generation::oop_iterate(mr, cl);
3051 3053 }
3052 3054 }
3053 3055
3054 3056 void
3055 3057 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) {
3056 3058 if (freelistLock()->owned_by_self()) {
3057 3059 Generation::oop_iterate(cl);
3058 3060 } else {
3059 3061 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3060 3062 Generation::oop_iterate(cl);
3061 3063 }
3062 3064 }
3063 3065
3064 3066 void
3065 3067 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
3066 3068 if (freelistLock()->owned_by_self()) {
3067 3069 Generation::object_iterate(cl);
3068 3070 } else {
3069 3071 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3070 3072 Generation::object_iterate(cl);
3071 3073 }
3072 3074 }
3073 3075
3074 3076 void
3075 3077 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
3076 3078 if (freelistLock()->owned_by_self()) {
3077 3079 Generation::safe_object_iterate(cl);
3078 3080 } else {
3079 3081 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3080 3082 Generation::safe_object_iterate(cl);
3081 3083 }
3082 3084 }
3083 3085
3084 3086 void
3085 3087 ConcurrentMarkSweepGeneration::pre_adjust_pointers() {
3086 3088 }
3087 3089
3088 3090 void
3089 3091 ConcurrentMarkSweepGeneration::post_compact() {
3090 3092 }
3091 3093
3092 3094 void
3093 3095 ConcurrentMarkSweepGeneration::prepare_for_verify() {
3094 3096 // Fix the linear allocation blocks to look like free blocks.
3095 3097
3096 3098 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3097 3099 // are not called when the heap is verified during universe initialization and
3098 3100 // at vm shutdown.
3099 3101 if (freelistLock()->owned_by_self()) {
3100 3102 cmsSpace()->prepare_for_verify();
3101 3103 } else {
3102 3104 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3103 3105 cmsSpace()->prepare_for_verify();
3104 3106 }
3105 3107 }
3106 3108
3107 3109 void
3108 3110 ConcurrentMarkSweepGeneration::verify(bool allow_dirty /* ignored */) {
3109 3111 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3110 3112 // are not called when the heap is verified during universe initialization and
3111 3113 // at vm shutdown.
3112 3114 if (freelistLock()->owned_by_self()) {
3113 3115 cmsSpace()->verify(false /* ignored */);
3114 3116 } else {
3115 3117 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3116 3118 cmsSpace()->verify(false /* ignored */);
3117 3119 }
3118 3120 }
3119 3121
3120 3122 void CMSCollector::verify(bool allow_dirty /* ignored */) {
3121 3123 _cmsGen->verify(allow_dirty);
3122 3124 _permGen->verify(allow_dirty);
3123 3125 }
3124 3126
3125 3127 #ifndef PRODUCT
3126 3128 bool CMSCollector::overflow_list_is_empty() const {
3127 3129 assert(_num_par_pushes >= 0, "Inconsistency");
3128 3130 if (_overflow_list == NULL) {
3129 3131 assert(_num_par_pushes == 0, "Inconsistency");
3130 3132 }
3131 3133 return _overflow_list == NULL;
3132 3134 }
3133 3135
3134 3136 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3135 3137 // merely consolidate assertion checks that appear to occur together frequently.
3136 3138 void CMSCollector::verify_work_stacks_empty() const {
3137 3139 assert(_markStack.isEmpty(), "Marking stack should be empty");
3138 3140 assert(overflow_list_is_empty(), "Overflow list should be empty");
3139 3141 }
3140 3142
3141 3143 void CMSCollector::verify_overflow_empty() const {
3142 3144 assert(overflow_list_is_empty(), "Overflow list should be empty");
3143 3145 assert(no_preserved_marks(), "No preserved marks");
3144 3146 }
3145 3147 #endif // PRODUCT
3146 3148
3147 3149 // Decide if we want to enable class unloading as part of the
3148 3150 // ensuing concurrent GC cycle. We will collect the perm gen and
3149 3151 // unload classes if it's the case that:
3150 3152 // (1) an explicit gc request has been made and the flag
3151 3153 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3152 3154 // (2) (a) class unloading is enabled at the command line, and
3153 3155 // (b) (i) perm gen threshold has been crossed, or
3154 3156 // (ii) old gen is getting really full, or
3155 3157 // (iii) the previous N CMS collections did not collect the
3156 3158 // perm gen
3157 3159 // NOTE: Provided there is no change in the state of the heap between
3158 3160 // calls to this method, it should have idempotent results. Moreover,
3159 3161 // its results should be monotonically increasing (i.e. going from 0 to 1,
3160 3162 // but not 1 to 0) between successive calls between which the heap was
3161 3163 // not collected. For the implementation below, it must thus rely on
3162 3164 // the property that concurrent_cycles_since_last_unload()
3163 3165 // will not decrease unless a collection cycle happened and that
3164 3166 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are
3165 3167 // themselves also monotonic in that sense. See check_monotonicity()
3166 3168 // below.
3167 3169 bool CMSCollector::update_should_unload_classes() {
3168 3170 _should_unload_classes = false;
3169 3171 // Condition 1 above
3170 3172 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3171 3173 _should_unload_classes = true;
3172 3174 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3173 3175 // Disjuncts 2.b.(i,ii,iii) above
3174 3176 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3175 3177 CMSClassUnloadingMaxInterval)
3176 3178 || _permGen->should_concurrent_collect()
3177 3179 || _cmsGen->is_too_full();
3178 3180 }
3179 3181 return _should_unload_classes;
3180 3182 }
3181 3183
3182 3184 bool ConcurrentMarkSweepGeneration::is_too_full() const {
3183 3185 bool res = should_concurrent_collect();
3184 3186 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
3185 3187 return res;
3186 3188 }
3187 3189
3188 3190 void CMSCollector::setup_cms_unloading_and_verification_state() {
3189 3191 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3190 3192 || VerifyBeforeExit;
3191 3193 const int rso = SharedHeap::SO_Symbols | SharedHeap::SO_Strings
3192 3194 | SharedHeap::SO_CodeCache;
3193 3195
3194 3196 if (should_unload_classes()) { // Should unload classes this cycle
3195 3197 remove_root_scanning_option(rso); // Shrink the root set appropriately
3196 3198 set_verifying(should_verify); // Set verification state for this cycle
3197 3199 return; // Nothing else needs to be done at this time
3198 3200 }
3199 3201
3200 3202 // Not unloading classes this cycle
3201 3203 assert(!should_unload_classes(), "Inconsitency!");
3202 3204 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3203 3205 // We were not verifying, or we _were_ unloading classes in the last cycle,
3204 3206 // AND some verification options are enabled this cycle; in this case,
3205 3207 // we must make sure that the deadness map is allocated if not already so,
3206 3208 // and cleared (if already allocated previously --
3207 3209 // CMSBitMap::sizeInBits() is used to determine if it's allocated).
3208 3210 if (perm_gen_verify_bit_map()->sizeInBits() == 0) {
3209 3211 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) {
3210 3212 warning("Failed to allocate permanent generation verification CMS Bit Map;\n"
3211 3213 "permanent generation verification disabled");
3212 3214 return; // Note that we leave verification disabled, so we'll retry this
3213 3215 // allocation next cycle. We _could_ remember this failure
3214 3216 // and skip further attempts and permanently disable verification
3215 3217 // attempts if that is considered more desirable.
3216 3218 }
3217 3219 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()),
3218 3220 "_perm_gen_ver_bit_map inconsistency?");
3219 3221 } else {
3220 3222 perm_gen_verify_bit_map()->clear_all();
3221 3223 }
3222 3224 // Include symbols, strings and code cache elements to prevent their resurrection.
3223 3225 add_root_scanning_option(rso);
3224 3226 set_verifying(true);
3225 3227 } else if (verifying() && !should_verify) {
3226 3228 // We were verifying, but some verification flags got disabled.
3227 3229 set_verifying(false);
3228 3230 // Exclude symbols, strings and code cache elements from root scanning to
3229 3231 // reduce IM and RM pauses.
3230 3232 remove_root_scanning_option(rso);
3231 3233 }
3232 3234 }
3233 3235
3234 3236
3235 3237 #ifndef PRODUCT
3236 3238 HeapWord* CMSCollector::block_start(const void* p) const {
3237 3239 const HeapWord* addr = (HeapWord*)p;
3238 3240 if (_span.contains(p)) {
3239 3241 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3240 3242 return _cmsGen->cmsSpace()->block_start(p);
3241 3243 } else {
3242 3244 assert(_permGen->cmsSpace()->is_in_reserved(addr),
3243 3245 "Inconsistent _span?");
3244 3246 return _permGen->cmsSpace()->block_start(p);
3245 3247 }
3246 3248 }
3247 3249 return NULL;
3248 3250 }
3249 3251 #endif
3250 3252
3251 3253 HeapWord*
3252 3254 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3253 3255 bool tlab,
3254 3256 bool parallel) {
3255 3257 CMSSynchronousYieldRequest yr;
3256 3258 assert(!tlab, "Can't deal with TLAB allocation");
3257 3259 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3258 3260 expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3259 3261 CMSExpansionCause::_satisfy_allocation);
3260 3262 if (GCExpandToAllocateDelayMillis > 0) {
3261 3263 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3262 3264 }
3263 3265 return have_lock_and_allocate(word_size, tlab);
3264 3266 }
3265 3267
3266 3268 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3267 3269 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3268 3270 // to CardGeneration and share it...
3269 3271 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) {
3270 3272 return CardGeneration::expand(bytes, expand_bytes);
3271 3273 }
3272 3274
3273 3275 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3274 3276 CMSExpansionCause::Cause cause)
3275 3277 {
3276 3278
3277 3279 bool success = expand(bytes, expand_bytes);
3278 3280
3279 3281 // remember why we expanded; this information is used
3280 3282 // by shouldConcurrentCollect() when making decisions on whether to start
3281 3283 // a new CMS cycle.
3282 3284 if (success) {
3283 3285 set_expansion_cause(cause);
3284 3286 if (PrintGCDetails && Verbose) {
3285 3287 gclog_or_tty->print_cr("Expanded CMS gen for %s",
3286 3288 CMSExpansionCause::to_string(cause));
3287 3289 }
3288 3290 }
3289 3291 }
3290 3292
3291 3293 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3292 3294 HeapWord* res = NULL;
3293 3295 MutexLocker x(ParGCRareEvent_lock);
3294 3296 while (true) {
3295 3297 // Expansion by some other thread might make alloc OK now:
3296 3298 res = ps->lab.alloc(word_sz);
3297 3299 if (res != NULL) return res;
3298 3300 // If there's not enough expansion space available, give up.
3299 3301 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3300 3302 return NULL;
3301 3303 }
3302 3304 // Otherwise, we try expansion.
3303 3305 expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3304 3306 CMSExpansionCause::_allocate_par_lab);
3305 3307 // Now go around the loop and try alloc again;
3306 3308 // A competing par_promote might beat us to the expansion space,
3307 3309 // so we may go around the loop again if promotion fails agaion.
3308 3310 if (GCExpandToAllocateDelayMillis > 0) {
3309 3311 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3310 3312 }
3311 3313 }
3312 3314 }
3313 3315
3314 3316
3315 3317 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3316 3318 PromotionInfo* promo) {
3317 3319 MutexLocker x(ParGCRareEvent_lock);
3318 3320 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3319 3321 while (true) {
3320 3322 // Expansion by some other thread might make alloc OK now:
3321 3323 if (promo->ensure_spooling_space()) {
3322 3324 assert(promo->has_spooling_space(),
3323 3325 "Post-condition of successful ensure_spooling_space()");
3324 3326 return true;
3325 3327 }
3326 3328 // If there's not enough expansion space available, give up.
3327 3329 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3328 3330 return false;
3329 3331 }
3330 3332 // Otherwise, we try expansion.
3331 3333 expand(refill_size_bytes, MinHeapDeltaBytes,
3332 3334 CMSExpansionCause::_allocate_par_spooling_space);
3333 3335 // Now go around the loop and try alloc again;
3334 3336 // A competing allocation might beat us to the expansion space,
3335 3337 // so we may go around the loop again if allocation fails again.
3336 3338 if (GCExpandToAllocateDelayMillis > 0) {
3337 3339 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3338 3340 }
3339 3341 }
3340 3342 }
3341 3343
3342 3344
3343 3345
3344 3346 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3345 3347 assert_locked_or_safepoint(Heap_lock);
3346 3348 size_t size = ReservedSpace::page_align_size_down(bytes);
3347 3349 if (size > 0) {
3348 3350 shrink_by(size);
3349 3351 }
3350 3352 }
3351 3353
3352 3354 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3353 3355 assert_locked_or_safepoint(Heap_lock);
3354 3356 bool result = _virtual_space.expand_by(bytes);
3355 3357 if (result) {
3356 3358 HeapWord* old_end = _cmsSpace->end();
3357 3359 size_t new_word_size =
3358 3360 heap_word_size(_virtual_space.committed_size());
3359 3361 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3360 3362 _bts->resize(new_word_size); // resize the block offset shared array
3361 3363 Universe::heap()->barrier_set()->resize_covered_region(mr);
3362 3364 // Hmmmm... why doesn't CFLS::set_end verify locking?
3363 3365 // This is quite ugly; FIX ME XXX
3364 3366 _cmsSpace->assert_locked(freelistLock());
3365 3367 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3366 3368
3367 3369 // update the space and generation capacity counters
3368 3370 if (UsePerfData) {
3369 3371 _space_counters->update_capacity();
3370 3372 _gen_counters->update_all();
3371 3373 }
3372 3374
3373 3375 if (Verbose && PrintGC) {
3374 3376 size_t new_mem_size = _virtual_space.committed_size();
3375 3377 size_t old_mem_size = new_mem_size - bytes;
3376 3378 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK",
3377 3379 name(), old_mem_size/K, bytes/K, new_mem_size/K);
3378 3380 }
3379 3381 }
3380 3382 return result;
3381 3383 }
3382 3384
3383 3385 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3384 3386 assert_locked_or_safepoint(Heap_lock);
3385 3387 bool success = true;
3386 3388 const size_t remaining_bytes = _virtual_space.uncommitted_size();
3387 3389 if (remaining_bytes > 0) {
3388 3390 success = grow_by(remaining_bytes);
3389 3391 DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3390 3392 }
3391 3393 return success;
3392 3394 }
3393 3395
3394 3396 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3395 3397 assert_locked_or_safepoint(Heap_lock);
3396 3398 assert_lock_strong(freelistLock());
3397 3399 // XXX Fix when compaction is implemented.
3398 3400 warning("Shrinking of CMS not yet implemented");
3399 3401 return;
3400 3402 }
3401 3403
3402 3404
3403 3405 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3404 3406 // phases.
3405 3407 class CMSPhaseAccounting: public StackObj {
3406 3408 public:
3407 3409 CMSPhaseAccounting(CMSCollector *collector,
3408 3410 const char *phase,
3409 3411 bool print_cr = true);
3410 3412 ~CMSPhaseAccounting();
3411 3413
3412 3414 private:
3413 3415 CMSCollector *_collector;
3414 3416 const char *_phase;
3415 3417 elapsedTimer _wallclock;
3416 3418 bool _print_cr;
3417 3419
3418 3420 public:
3419 3421 // Not MT-safe; so do not pass around these StackObj's
3420 3422 // where they may be accessed by other threads.
3421 3423 jlong wallclock_millis() {
3422 3424 assert(_wallclock.is_active(), "Wall clock should not stop");
3423 3425 _wallclock.stop(); // to record time
3424 3426 jlong ret = _wallclock.milliseconds();
3425 3427 _wallclock.start(); // restart
3426 3428 return ret;
3427 3429 }
3428 3430 };
3429 3431
3430 3432 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3431 3433 const char *phase,
3432 3434 bool print_cr) :
3433 3435 _collector(collector), _phase(phase), _print_cr(print_cr) {
3434 3436
3435 3437 if (PrintCMSStatistics != 0) {
3436 3438 _collector->resetYields();
3437 3439 }
3438 3440 if (PrintGCDetails && PrintGCTimeStamps) {
3439 3441 gclog_or_tty->date_stamp(PrintGCDateStamps);
3440 3442 gclog_or_tty->stamp();
3441 3443 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]",
3442 3444 _collector->cmsGen()->short_name(), _phase);
3443 3445 }
3444 3446 _collector->resetTimer();
3445 3447 _wallclock.start();
3446 3448 _collector->startTimer();
3447 3449 }
3448 3450
3449 3451 CMSPhaseAccounting::~CMSPhaseAccounting() {
3450 3452 assert(_wallclock.is_active(), "Wall clock should not have stopped");
3451 3453 _collector->stopTimer();
3452 3454 _wallclock.stop();
3453 3455 if (PrintGCDetails) {
3454 3456 gclog_or_tty->date_stamp(PrintGCDateStamps);
3455 3457 if (PrintGCTimeStamps) {
3456 3458 gclog_or_tty->stamp();
3457 3459 gclog_or_tty->print(": ");
3458 3460 }
3459 3461 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3460 3462 _collector->cmsGen()->short_name(),
3461 3463 _phase, _collector->timerValue(), _wallclock.seconds());
3462 3464 if (_print_cr) {
3463 3465 gclog_or_tty->print_cr("");
3464 3466 }
3465 3467 if (PrintCMSStatistics != 0) {
3466 3468 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3467 3469 _collector->yields());
3468 3470 }
3469 3471 }
3470 3472 }
3471 3473
3472 3474 // CMS work
3473 3475
3474 3476 // Checkpoint the roots into this generation from outside
3475 3477 // this generation. [Note this initial checkpoint need only
3476 3478 // be approximate -- we'll do a catch up phase subsequently.]
3477 3479 void CMSCollector::checkpointRootsInitial(bool asynch) {
3478 3480 assert(_collectorState == InitialMarking, "Wrong collector state");
3479 3481 check_correct_thread_executing();
3480 3482 TraceCMSMemoryManagerStats tms(_collectorState);
3481 3483
3482 3484 ReferenceProcessor* rp = ref_processor();
3483 3485 SpecializationStats::clear();
3484 3486 assert(_restart_addr == NULL, "Control point invariant");
3485 3487 if (asynch) {
3486 3488 // acquire locks for subsequent manipulations
3487 3489 MutexLockerEx x(bitMapLock(),
3488 3490 Mutex::_no_safepoint_check_flag);
3489 3491 checkpointRootsInitialWork(asynch);
3490 3492 rp->verify_no_references_recorded();
3491 3493 rp->enable_discovery(); // enable ("weak") refs discovery
3492 3494 _collectorState = Marking;
3493 3495 } else {
3494 3496 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3495 3497 // which recognizes if we are a CMS generation, and doesn't try to turn on
3496 3498 // discovery; verify that they aren't meddling.
3497 3499 assert(!rp->discovery_is_atomic(),
3498 3500 "incorrect setting of discovery predicate");
3499 3501 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3500 3502 "ref discovery for this generation kind");
3501 3503 // already have locks
3502 3504 checkpointRootsInitialWork(asynch);
3503 3505 rp->enable_discovery(); // now enable ("weak") refs discovery
3504 3506 _collectorState = Marking;
3505 3507 }
3506 3508 SpecializationStats::print();
3507 3509 }
3508 3510
3509 3511 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3510 3512 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3511 3513 assert(_collectorState == InitialMarking, "just checking");
3512 3514
3513 3515 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3514 3516 // precede our marking with a collection of all
3515 3517 // younger generations to keep floating garbage to a minimum.
3516 3518 // XXX: we won't do this for now -- it's an optimization to be done later.
3517 3519
3518 3520 // already have locks
3519 3521 assert_lock_strong(bitMapLock());
3520 3522 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3521 3523
3522 3524 // Setup the verification and class unloading state for this
3523 3525 // CMS collection cycle.
3524 3526 setup_cms_unloading_and_verification_state();
3525 3527
3526 3528 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork",
3527 3529 PrintGCDetails && Verbose, true, gclog_or_tty);)
3528 3530 if (UseAdaptiveSizePolicy) {
3529 3531 size_policy()->checkpoint_roots_initial_begin();
3530 3532 }
3531 3533
3532 3534 // Reset all the PLAB chunk arrays if necessary.
3533 3535 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3534 3536 reset_survivor_plab_arrays();
3535 3537 }
3536 3538
3537 3539 ResourceMark rm;
3538 3540 HandleMark hm;
3539 3541
3540 3542 FalseClosure falseClosure;
3541 3543 // In the case of a synchronous collection, we will elide the
3542 3544 // remark step, so it's important to catch all the nmethod oops
3543 3545 // in this step.
3544 3546 // The final 'true' flag to gen_process_strong_roots will ensure this.
3545 3547 // If 'async' is true, we can relax the nmethod tracing.
3546 3548 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
3547 3549 GenCollectedHeap* gch = GenCollectedHeap::heap();
3548 3550
3549 3551 verify_work_stacks_empty();
3550 3552 verify_overflow_empty();
3551 3553
3552 3554 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3553 3555 // Update the saved marks which may affect the root scans.
3554 3556 gch->save_marks();
3555 3557
3556 3558 // weak reference processing has not started yet.
3557 3559 ref_processor()->set_enqueuing_is_done(false);
3558 3560
3559 3561 {
3560 3562 // This is not needed. DEBUG_ONLY(RememberKlassesChecker imx(true);)
3561 3563 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3562 3564 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3563 3565 gch->gen_process_strong_roots(_cmsGen->level(),
3564 3566 true, // younger gens are roots
3565 3567 true, // activate StrongRootsScope
3566 3568 true, // collecting perm gen
3567 3569 SharedHeap::ScanningOption(roots_scanning_options()),
3568 3570 ¬Older,
3569 3571 true, // walk all of code cache if (so & SO_CodeCache)
3570 3572 NULL);
3571 3573 }
3572 3574
3573 3575 // Clear mod-union table; it will be dirtied in the prologue of
3574 3576 // CMS generation per each younger generation collection.
3575 3577
3576 3578 assert(_modUnionTable.isAllClear(),
3577 3579 "Was cleared in most recent final checkpoint phase"
3578 3580 " or no bits are set in the gc_prologue before the start of the next "
3579 3581 "subsequent marking phase.");
3580 3582
3581 3583 // Temporarily disabled, since pre/post-consumption closures don't
3582 3584 // care about precleaned cards
3583 3585 #if 0
3584 3586 {
3585 3587 MemRegion mr = MemRegion((HeapWord*)_virtual_space.low(),
3586 3588 (HeapWord*)_virtual_space.high());
3587 3589 _ct->ct_bs()->preclean_dirty_cards(mr);
3588 3590 }
3589 3591 #endif
3590 3592
3591 3593 // Save the end of the used_region of the constituent generations
3592 3594 // to be used to limit the extent of sweep in each generation.
3593 3595 save_sweep_limits();
3594 3596 if (UseAdaptiveSizePolicy) {
3595 3597 size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3596 3598 }
3597 3599 verify_overflow_empty();
3598 3600 }
3599 3601
3600 3602 bool CMSCollector::markFromRoots(bool asynch) {
3601 3603 // we might be tempted to assert that:
3602 3604 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3603 3605 // "inconsistent argument?");
3604 3606 // However that wouldn't be right, because it's possible that
3605 3607 // a safepoint is indeed in progress as a younger generation
3606 3608 // stop-the-world GC happens even as we mark in this generation.
3607 3609 assert(_collectorState == Marking, "inconsistent state?");
3608 3610 check_correct_thread_executing();
3609 3611 verify_overflow_empty();
3610 3612
3611 3613 bool res;
3612 3614 if (asynch) {
3613 3615
3614 3616 // Start the timers for adaptive size policy for the concurrent phases
3615 3617 // Do it here so that the foreground MS can use the concurrent
3616 3618 // timer since a foreground MS might has the sweep done concurrently
3617 3619 // or STW.
3618 3620 if (UseAdaptiveSizePolicy) {
3619 3621 size_policy()->concurrent_marking_begin();
3620 3622 }
3621 3623
3622 3624 // Weak ref discovery note: We may be discovering weak
3623 3625 // refs in this generation concurrent (but interleaved) with
3624 3626 // weak ref discovery by a younger generation collector.
3625 3627
3626 3628 CMSTokenSyncWithLocks ts(true, bitMapLock());
3627 3629 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3628 3630 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3629 3631 res = markFromRootsWork(asynch);
3630 3632 if (res) {
3631 3633 _collectorState = Precleaning;
3632 3634 } else { // We failed and a foreground collection wants to take over
3633 3635 assert(_foregroundGCIsActive, "internal state inconsistency");
3634 3636 assert(_restart_addr == NULL, "foreground will restart from scratch");
3635 3637 if (PrintGCDetails) {
3636 3638 gclog_or_tty->print_cr("bailing out to foreground collection");
3637 3639 }
3638 3640 }
3639 3641 if (UseAdaptiveSizePolicy) {
3640 3642 size_policy()->concurrent_marking_end();
3641 3643 }
3642 3644 } else {
3643 3645 assert(SafepointSynchronize::is_at_safepoint(),
3644 3646 "inconsistent with asynch == false");
3645 3647 if (UseAdaptiveSizePolicy) {
3646 3648 size_policy()->ms_collection_marking_begin();
3647 3649 }
3648 3650 // already have locks
3649 3651 res = markFromRootsWork(asynch);
3650 3652 _collectorState = FinalMarking;
3651 3653 if (UseAdaptiveSizePolicy) {
3652 3654 GenCollectedHeap* gch = GenCollectedHeap::heap();
3653 3655 size_policy()->ms_collection_marking_end(gch->gc_cause());
3654 3656 }
3655 3657 }
3656 3658 verify_overflow_empty();
3657 3659 return res;
3658 3660 }
3659 3661
3660 3662 bool CMSCollector::markFromRootsWork(bool asynch) {
3661 3663 // iterate over marked bits in bit map, doing a full scan and mark
3662 3664 // from these roots using the following algorithm:
3663 3665 // . if oop is to the right of the current scan pointer,
3664 3666 // mark corresponding bit (we'll process it later)
3665 3667 // . else (oop is to left of current scan pointer)
3666 3668 // push oop on marking stack
3667 3669 // . drain the marking stack
3668 3670
3669 3671 // Note that when we do a marking step we need to hold the
3670 3672 // bit map lock -- recall that direct allocation (by mutators)
3671 3673 // and promotion (by younger generation collectors) is also
3672 3674 // marking the bit map. [the so-called allocate live policy.]
3673 3675 // Because the implementation of bit map marking is not
3674 3676 // robust wrt simultaneous marking of bits in the same word,
3675 3677 // we need to make sure that there is no such interference
3676 3678 // between concurrent such updates.
3677 3679
3678 3680 // already have locks
3679 3681 assert_lock_strong(bitMapLock());
3680 3682
3681 3683 // Clear the revisit stack, just in case there are any
3682 3684 // obsolete contents from a short-circuited previous CMS cycle.
3683 3685 _revisitStack.reset();
3684 3686 verify_work_stacks_empty();
3685 3687 verify_overflow_empty();
3686 3688 assert(_revisitStack.isEmpty(), "tabula rasa");
3687 3689 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());)
3688 3690 bool result = false;
3689 3691 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3690 3692 result = do_marking_mt(asynch);
3691 3693 } else {
3692 3694 result = do_marking_st(asynch);
3693 3695 }
3694 3696 return result;
3695 3697 }
3696 3698
3697 3699 // Forward decl
3698 3700 class CMSConcMarkingTask;
3699 3701
3700 3702 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3701 3703 CMSCollector* _collector;
3702 3704 CMSConcMarkingTask* _task;
3703 3705 public:
3704 3706 virtual void yield();
3705 3707
3706 3708 // "n_threads" is the number of threads to be terminated.
3707 3709 // "queue_set" is a set of work queues of other threads.
3708 3710 // "collector" is the CMS collector associated with this task terminator.
3709 3711 // "yield" indicates whether we need the gang as a whole to yield.
3710 3712 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3711 3713 ParallelTaskTerminator(n_threads, queue_set),
3712 3714 _collector(collector) { }
3713 3715
3714 3716 void set_task(CMSConcMarkingTask* task) {
3715 3717 _task = task;
3716 3718 }
3717 3719 };
3718 3720
3719 3721 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3720 3722 CMSConcMarkingTask* _task;
3721 3723 public:
3722 3724 bool should_exit_termination();
3723 3725 void set_task(CMSConcMarkingTask* task) {
3724 3726 _task = task;
3725 3727 }
3726 3728 };
3727 3729
3728 3730 // MT Concurrent Marking Task
3729 3731 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3730 3732 CMSCollector* _collector;
3731 3733 int _n_workers; // requested/desired # workers
3732 3734 bool _asynch;
3733 3735 bool _result;
3734 3736 CompactibleFreeListSpace* _cms_space;
3735 3737 CompactibleFreeListSpace* _perm_space;
3736 3738 char _pad_front[64]; // padding to ...
3737 3739 HeapWord* _global_finger; // ... avoid sharing cache line
3738 3740 char _pad_back[64];
3739 3741 HeapWord* _restart_addr;
3740 3742
3741 3743 // Exposed here for yielding support
3742 3744 Mutex* const _bit_map_lock;
3743 3745
3744 3746 // The per thread work queues, available here for stealing
3745 3747 OopTaskQueueSet* _task_queues;
3746 3748
3747 3749 // Termination (and yielding) support
3748 3750 CMSConcMarkingTerminator _term;
3749 3751 CMSConcMarkingTerminatorTerminator _term_term;
3750 3752
3751 3753 public:
3752 3754 CMSConcMarkingTask(CMSCollector* collector,
3753 3755 CompactibleFreeListSpace* cms_space,
3754 3756 CompactibleFreeListSpace* perm_space,
3755 3757 bool asynch,
3756 3758 YieldingFlexibleWorkGang* workers,
3757 3759 OopTaskQueueSet* task_queues):
3758 3760 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3759 3761 _collector(collector),
3760 3762 _cms_space(cms_space),
3761 3763 _perm_space(perm_space),
3762 3764 _asynch(asynch), _n_workers(0), _result(true),
3763 3765 _task_queues(task_queues),
3764 3766 _term(_n_workers, task_queues, _collector),
3765 3767 _bit_map_lock(collector->bitMapLock())
3766 3768 {
3767 3769 _requested_size = _n_workers;
3768 3770 _term.set_task(this);
3769 3771 _term_term.set_task(this);
3770 3772 assert(_cms_space->bottom() < _perm_space->bottom(),
3771 3773 "Finger incorrectly initialized below");
3772 3774 _restart_addr = _global_finger = _cms_space->bottom();
3773 3775 }
3774 3776
3775 3777
3776 3778 OopTaskQueueSet* task_queues() { return _task_queues; }
3777 3779
3778 3780 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3779 3781
3780 3782 HeapWord** global_finger_addr() { return &_global_finger; }
3781 3783
3782 3784 CMSConcMarkingTerminator* terminator() { return &_term; }
3783 3785
3784 3786 virtual void set_for_termination(int active_workers) {
3785 3787 terminator()->reset_for_reuse(active_workers);
3786 3788 }
3787 3789
3788 3790 void work(int i);
3789 3791 bool should_yield() {
3790 3792 return ConcurrentMarkSweepThread::should_yield()
3791 3793 && !_collector->foregroundGCIsActive()
3792 3794 && _asynch;
3793 3795 }
3794 3796
3795 3797 virtual void coordinator_yield(); // stuff done by coordinator
3796 3798 bool result() { return _result; }
3797 3799
3798 3800 void reset(HeapWord* ra) {
3799 3801 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3800 3802 assert(_global_finger >= _perm_space->end(), "Postcondition of ::work(i)");
3801 3803 assert(ra < _perm_space->end(), "ra too large");
3802 3804 _restart_addr = _global_finger = ra;
3803 3805 _term.reset_for_reuse();
3804 3806 }
3805 3807
3806 3808 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3807 3809 OopTaskQueue* work_q);
3808 3810
3809 3811 private:
3810 3812 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3811 3813 void do_work_steal(int i);
3812 3814 void bump_global_finger(HeapWord* f);
3813 3815 };
3814 3816
3815 3817 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3816 3818 assert(_task != NULL, "Error");
3817 3819 return _task->yielding();
3818 3820 // Note that we do not need the disjunct || _task->should_yield() above
3819 3821 // because we want terminating threads to yield only if the task
3820 3822 // is already in the midst of yielding, which happens only after at least one
3821 3823 // thread has yielded.
3822 3824 }
3823 3825
3824 3826 void CMSConcMarkingTerminator::yield() {
3825 3827 if (_task->should_yield()) {
3826 3828 _task->yield();
3827 3829 } else {
3828 3830 ParallelTaskTerminator::yield();
3829 3831 }
3830 3832 }
3831 3833
3832 3834 ////////////////////////////////////////////////////////////////
3833 3835 // Concurrent Marking Algorithm Sketch
3834 3836 ////////////////////////////////////////////////////////////////
3835 3837 // Until all tasks exhausted (both spaces):
3836 3838 // -- claim next available chunk
3837 3839 // -- bump global finger via CAS
3838 3840 // -- find first object that starts in this chunk
3839 3841 // and start scanning bitmap from that position
3840 3842 // -- scan marked objects for oops
3841 3843 // -- CAS-mark target, and if successful:
3842 3844 // . if target oop is above global finger (volatile read)
3843 3845 // nothing to do
3844 3846 // . if target oop is in chunk and above local finger
3845 3847 // then nothing to do
3846 3848 // . else push on work-queue
3847 3849 // -- Deal with possible overflow issues:
3848 3850 // . local work-queue overflow causes stuff to be pushed on
3849 3851 // global (common) overflow queue
3850 3852 // . always first empty local work queue
3851 3853 // . then get a batch of oops from global work queue if any
3852 3854 // . then do work stealing
3853 3855 // -- When all tasks claimed (both spaces)
3854 3856 // and local work queue empty,
3855 3857 // then in a loop do:
3856 3858 // . check global overflow stack; steal a batch of oops and trace
3857 3859 // . try to steal from other threads oif GOS is empty
3858 3860 // . if neither is available, offer termination
3859 3861 // -- Terminate and return result
3860 3862 //
3861 3863 void CMSConcMarkingTask::work(int i) {
3862 3864 elapsedTimer _timer;
3863 3865 ResourceMark rm;
3864 3866 HandleMark hm;
3865 3867
3866 3868 DEBUG_ONLY(_collector->verify_overflow_empty();)
3867 3869
3868 3870 // Before we begin work, our work queue should be empty
3869 3871 assert(work_queue(i)->size() == 0, "Expected to be empty");
3870 3872 // Scan the bitmap covering _cms_space, tracing through grey objects.
3871 3873 _timer.start();
3872 3874 do_scan_and_mark(i, _cms_space);
3873 3875 _timer.stop();
3874 3876 if (PrintCMSStatistics != 0) {
3875 3877 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3876 3878 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3877 3879 }
3878 3880
3879 3881 // ... do the same for the _perm_space
3880 3882 _timer.reset();
3881 3883 _timer.start();
3882 3884 do_scan_and_mark(i, _perm_space);
3883 3885 _timer.stop();
3884 3886 if (PrintCMSStatistics != 0) {
3885 3887 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec",
3886 3888 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3887 3889 }
3888 3890
3889 3891 // ... do work stealing
3890 3892 _timer.reset();
3891 3893 _timer.start();
3892 3894 do_work_steal(i);
3893 3895 _timer.stop();
3894 3896 if (PrintCMSStatistics != 0) {
3895 3897 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3896 3898 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3897 3899 }
3898 3900 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3899 3901 assert(work_queue(i)->size() == 0, "Should have been emptied");
3900 3902 // Note that under the current task protocol, the
3901 3903 // following assertion is true even of the spaces
3902 3904 // expanded since the completion of the concurrent
3903 3905 // marking. XXX This will likely change under a strict
3904 3906 // ABORT semantics.
3905 3907 assert(_global_finger > _cms_space->end() &&
3906 3908 _global_finger >= _perm_space->end(),
3907 3909 "All tasks have been completed");
3908 3910 DEBUG_ONLY(_collector->verify_overflow_empty();)
3909 3911 }
3910 3912
3911 3913 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3912 3914 HeapWord* read = _global_finger;
3913 3915 HeapWord* cur = read;
3914 3916 while (f > read) {
3915 3917 cur = read;
3916 3918 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3917 3919 if (cur == read) {
3918 3920 // our cas succeeded
3919 3921 assert(_global_finger >= f, "protocol consistency");
3920 3922 break;
3921 3923 }
3922 3924 }
3923 3925 }
3924 3926
3925 3927 // This is really inefficient, and should be redone by
3926 3928 // using (not yet available) block-read and -write interfaces to the
3927 3929 // stack and the work_queue. XXX FIX ME !!!
3928 3930 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3929 3931 OopTaskQueue* work_q) {
3930 3932 // Fast lock-free check
3931 3933 if (ovflw_stk->length() == 0) {
3932 3934 return false;
3933 3935 }
3934 3936 assert(work_q->size() == 0, "Shouldn't steal");
3935 3937 MutexLockerEx ml(ovflw_stk->par_lock(),
3936 3938 Mutex::_no_safepoint_check_flag);
3937 3939 // Grab up to 1/4 the size of the work queue
3938 3940 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3939 3941 (size_t)ParGCDesiredObjsFromOverflowList);
3940 3942 num = MIN2(num, ovflw_stk->length());
3941 3943 for (int i = (int) num; i > 0; i--) {
3942 3944 oop cur = ovflw_stk->pop();
3943 3945 assert(cur != NULL, "Counted wrong?");
3944 3946 work_q->push(cur);
3945 3947 }
3946 3948 return num > 0;
3947 3949 }
3948 3950
3949 3951 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3950 3952 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3951 3953 int n_tasks = pst->n_tasks();
3952 3954 // We allow that there may be no tasks to do here because
3953 3955 // we are restarting after a stack overflow.
3954 3956 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3955 3957 int nth_task = 0;
3956 3958
3957 3959 HeapWord* aligned_start = sp->bottom();
3958 3960 if (sp->used_region().contains(_restart_addr)) {
3959 3961 // Align down to a card boundary for the start of 0th task
3960 3962 // for this space.
3961 3963 aligned_start =
3962 3964 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3963 3965 CardTableModRefBS::card_size);
3964 3966 }
3965 3967
3966 3968 size_t chunk_size = sp->marking_task_size();
3967 3969 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3968 3970 // Having claimed the nth task in this space,
3969 3971 // compute the chunk that it corresponds to:
3970 3972 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3971 3973 aligned_start + (nth_task+1)*chunk_size);
3972 3974 // Try and bump the global finger via a CAS;
3973 3975 // note that we need to do the global finger bump
3974 3976 // _before_ taking the intersection below, because
3975 3977 // the task corresponding to that region will be
3976 3978 // deemed done even if the used_region() expands
3977 3979 // because of allocation -- as it almost certainly will
3978 3980 // during start-up while the threads yield in the
3979 3981 // closure below.
3980 3982 HeapWord* finger = span.end();
3981 3983 bump_global_finger(finger); // atomically
3982 3984 // There are null tasks here corresponding to chunks
3983 3985 // beyond the "top" address of the space.
3984 3986 span = span.intersection(sp->used_region());
3985 3987 if (!span.is_empty()) { // Non-null task
3986 3988 HeapWord* prev_obj;
3987 3989 assert(!span.contains(_restart_addr) || nth_task == 0,
3988 3990 "Inconsistency");
3989 3991 if (nth_task == 0) {
3990 3992 // For the 0th task, we'll not need to compute a block_start.
3991 3993 if (span.contains(_restart_addr)) {
3992 3994 // In the case of a restart because of stack overflow,
3993 3995 // we might additionally skip a chunk prefix.
3994 3996 prev_obj = _restart_addr;
3995 3997 } else {
3996 3998 prev_obj = span.start();
3997 3999 }
3998 4000 } else {
3999 4001 // We want to skip the first object because
4000 4002 // the protocol is to scan any object in its entirety
4001 4003 // that _starts_ in this span; a fortiori, any
4002 4004 // object starting in an earlier span is scanned
4003 4005 // as part of an earlier claimed task.
4004 4006 // Below we use the "careful" version of block_start
4005 4007 // so we do not try to navigate uninitialized objects.
4006 4008 prev_obj = sp->block_start_careful(span.start());
4007 4009 // Below we use a variant of block_size that uses the
4008 4010 // Printezis bits to avoid waiting for allocated
4009 4011 // objects to become initialized/parsable.
4010 4012 while (prev_obj < span.start()) {
4011 4013 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
4012 4014 if (sz > 0) {
4013 4015 prev_obj += sz;
4014 4016 } else {
4015 4017 // In this case we may end up doing a bit of redundant
4016 4018 // scanning, but that appears unavoidable, short of
4017 4019 // locking the free list locks; see bug 6324141.
4018 4020 break;
4019 4021 }
4020 4022 }
4021 4023 }
4022 4024 if (prev_obj < span.end()) {
4023 4025 MemRegion my_span = MemRegion(prev_obj, span.end());
4024 4026 // Do the marking work within a non-empty span --
4025 4027 // the last argument to the constructor indicates whether the
4026 4028 // iteration should be incremental with periodic yields.
4027 4029 Par_MarkFromRootsClosure cl(this, _collector, my_span,
4028 4030 &_collector->_markBitMap,
4029 4031 work_queue(i),
4030 4032 &_collector->_markStack,
4031 4033 &_collector->_revisitStack,
4032 4034 _asynch);
4033 4035 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
4034 4036 } // else nothing to do for this task
4035 4037 } // else nothing to do for this task
4036 4038 }
4037 4039 // We'd be tempted to assert here that since there are no
4038 4040 // more tasks left to claim in this space, the global_finger
4039 4041 // must exceed space->top() and a fortiori space->end(). However,
4040 4042 // that would not quite be correct because the bumping of
4041 4043 // global_finger occurs strictly after the claiming of a task,
4042 4044 // so by the time we reach here the global finger may not yet
4043 4045 // have been bumped up by the thread that claimed the last
4044 4046 // task.
4045 4047 pst->all_tasks_completed();
4046 4048 }
4047 4049
4048 4050 class Par_ConcMarkingClosure: public Par_KlassRememberingOopClosure {
4049 4051 private:
4050 4052 CMSConcMarkingTask* _task;
4051 4053 MemRegion _span;
4052 4054 CMSBitMap* _bit_map;
4053 4055 CMSMarkStack* _overflow_stack;
4054 4056 OopTaskQueue* _work_queue;
4055 4057 protected:
4056 4058 DO_OOP_WORK_DEFN
4057 4059 public:
4058 4060 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
4059 4061 CMSBitMap* bit_map, CMSMarkStack* overflow_stack,
4060 4062 CMSMarkStack* revisit_stack):
4061 4063 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
4062 4064 _task(task),
4063 4065 _span(collector->_span),
4064 4066 _work_queue(work_queue),
4065 4067 _bit_map(bit_map),
4066 4068 _overflow_stack(overflow_stack)
4067 4069 { }
4068 4070 virtual void do_oop(oop* p);
4069 4071 virtual void do_oop(narrowOop* p);
4070 4072 void trim_queue(size_t max);
4071 4073 void handle_stack_overflow(HeapWord* lost);
4072 4074 void do_yield_check() {
4073 4075 if (_task->should_yield()) {
4074 4076 _task->yield();
4075 4077 }
4076 4078 }
4077 4079 };
4078 4080
4079 4081 // Grey object scanning during work stealing phase --
4080 4082 // the salient assumption here is that any references
4081 4083 // that are in these stolen objects being scanned must
4082 4084 // already have been initialized (else they would not have
4083 4085 // been published), so we do not need to check for
4084 4086 // uninitialized objects before pushing here.
4085 4087 void Par_ConcMarkingClosure::do_oop(oop obj) {
4086 4088 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
4087 4089 HeapWord* addr = (HeapWord*)obj;
4088 4090 // Check if oop points into the CMS generation
4089 4091 // and is not marked
4090 4092 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
4091 4093 // a white object ...
4092 4094 // If we manage to "claim" the object, by being the
4093 4095 // first thread to mark it, then we push it on our
4094 4096 // marking stack
4095 4097 if (_bit_map->par_mark(addr)) { // ... now grey
4096 4098 // push on work queue (grey set)
4097 4099 bool simulate_overflow = false;
4098 4100 NOT_PRODUCT(
4099 4101 if (CMSMarkStackOverflowALot &&
4100 4102 _collector->simulate_overflow()) {
4101 4103 // simulate a stack overflow
4102 4104 simulate_overflow = true;
4103 4105 }
4104 4106 )
4105 4107 if (simulate_overflow ||
4106 4108 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
4107 4109 // stack overflow
4108 4110 if (PrintCMSStatistics != 0) {
4109 4111 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
4110 4112 SIZE_FORMAT, _overflow_stack->capacity());
4111 4113 }
4112 4114 // We cannot assert that the overflow stack is full because
4113 4115 // it may have been emptied since.
4114 4116 assert(simulate_overflow ||
4115 4117 _work_queue->size() == _work_queue->max_elems(),
4116 4118 "Else push should have succeeded");
4117 4119 handle_stack_overflow(addr);
4118 4120 }
4119 4121 } // Else, some other thread got there first
4120 4122 do_yield_check();
4121 4123 }
4122 4124 }
4123 4125
4124 4126 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4125 4127 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4126 4128
4127 4129 void Par_ConcMarkingClosure::trim_queue(size_t max) {
4128 4130 while (_work_queue->size() > max) {
4129 4131 oop new_oop;
4130 4132 if (_work_queue->pop_local(new_oop)) {
4131 4133 assert(new_oop->is_oop(), "Should be an oop");
4132 4134 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
4133 4135 assert(_span.contains((HeapWord*)new_oop), "Not in span");
4134 4136 assert(new_oop->is_parsable(), "Should be parsable");
4135 4137 new_oop->oop_iterate(this); // do_oop() above
4136 4138 do_yield_check();
4137 4139 }
4138 4140 }
4139 4141 }
4140 4142
4141 4143 // Upon stack overflow, we discard (part of) the stack,
4142 4144 // remembering the least address amongst those discarded
4143 4145 // in CMSCollector's _restart_address.
4144 4146 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
4145 4147 // We need to do this under a mutex to prevent other
4146 4148 // workers from interfering with the work done below.
4147 4149 MutexLockerEx ml(_overflow_stack->par_lock(),
4148 4150 Mutex::_no_safepoint_check_flag);
4149 4151 // Remember the least grey address discarded
4150 4152 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
4151 4153 _collector->lower_restart_addr(ra);
4152 4154 _overflow_stack->reset(); // discard stack contents
4153 4155 _overflow_stack->expand(); // expand the stack if possible
4154 4156 }
4155 4157
4156 4158
4157 4159 void CMSConcMarkingTask::do_work_steal(int i) {
4158 4160 OopTaskQueue* work_q = work_queue(i);
4159 4161 oop obj_to_scan;
4160 4162 CMSBitMap* bm = &(_collector->_markBitMap);
4161 4163 CMSMarkStack* ovflw = &(_collector->_markStack);
4162 4164 CMSMarkStack* revisit = &(_collector->_revisitStack);
4163 4165 int* seed = _collector->hash_seed(i);
4164 4166 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw, revisit);
4165 4167 while (true) {
4166 4168 cl.trim_queue(0);
4167 4169 assert(work_q->size() == 0, "Should have been emptied above");
4168 4170 if (get_work_from_overflow_stack(ovflw, work_q)) {
4169 4171 // Can't assert below because the work obtained from the
4170 4172 // overflow stack may already have been stolen from us.
4171 4173 // assert(work_q->size() > 0, "Work from overflow stack");
4172 4174 continue;
4173 4175 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4174 4176 assert(obj_to_scan->is_oop(), "Should be an oop");
4175 4177 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
4176 4178 obj_to_scan->oop_iterate(&cl);
4177 4179 } else if (terminator()->offer_termination(&_term_term)) {
4178 4180 assert(work_q->size() == 0, "Impossible!");
4179 4181 break;
4180 4182 } else if (yielding() || should_yield()) {
4181 4183 yield();
4182 4184 }
4183 4185 }
4184 4186 }
4185 4187
4186 4188 // This is run by the CMS (coordinator) thread.
4187 4189 void CMSConcMarkingTask::coordinator_yield() {
4188 4190 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4189 4191 "CMS thread should hold CMS token");
4190 4192 DEBUG_ONLY(RememberKlassesChecker mux(false);)
4191 4193 // First give up the locks, then yield, then re-lock
4192 4194 // We should probably use a constructor/destructor idiom to
4193 4195 // do this unlock/lock or modify the MutexUnlocker class to
4194 4196 // serve our purpose. XXX
4195 4197 assert_lock_strong(_bit_map_lock);
4196 4198 _bit_map_lock->unlock();
4197 4199 ConcurrentMarkSweepThread::desynchronize(true);
4198 4200 ConcurrentMarkSweepThread::acknowledge_yield_request();
4199 4201 _collector->stopTimer();
4200 4202 if (PrintCMSStatistics != 0) {
4201 4203 _collector->incrementYields();
4202 4204 }
4203 4205 _collector->icms_wait();
4204 4206
4205 4207 // It is possible for whichever thread initiated the yield request
4206 4208 // not to get a chance to wake up and take the bitmap lock between
4207 4209 // this thread releasing it and reacquiring it. So, while the
4208 4210 // should_yield() flag is on, let's sleep for a bit to give the
4209 4211 // other thread a chance to wake up. The limit imposed on the number
4210 4212 // of iterations is defensive, to avoid any unforseen circumstances
4211 4213 // putting us into an infinite loop. Since it's always been this
4212 4214 // (coordinator_yield()) method that was observed to cause the
4213 4215 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4214 4216 // which is by default non-zero. For the other seven methods that
4215 4217 // also perform the yield operation, as are using a different
4216 4218 // parameter (CMSYieldSleepCount) which is by default zero. This way we
4217 4219 // can enable the sleeping for those methods too, if necessary.
4218 4220 // See 6442774.
4219 4221 //
4220 4222 // We really need to reconsider the synchronization between the GC
4221 4223 // thread and the yield-requesting threads in the future and we
4222 4224 // should really use wait/notify, which is the recommended
4223 4225 // way of doing this type of interaction. Additionally, we should
4224 4226 // consolidate the eight methods that do the yield operation and they
4225 4227 // are almost identical into one for better maintenability and
4226 4228 // readability. See 6445193.
4227 4229 //
4228 4230 // Tony 2006.06.29
4229 4231 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4230 4232 ConcurrentMarkSweepThread::should_yield() &&
4231 4233 !CMSCollector::foregroundGCIsActive(); ++i) {
4232 4234 os::sleep(Thread::current(), 1, false);
4233 4235 ConcurrentMarkSweepThread::acknowledge_yield_request();
4234 4236 }
4235 4237
4236 4238 ConcurrentMarkSweepThread::synchronize(true);
4237 4239 _bit_map_lock->lock_without_safepoint_check();
4238 4240 _collector->startTimer();
4239 4241 }
4240 4242
4241 4243 bool CMSCollector::do_marking_mt(bool asynch) {
4242 4244 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
4243 4245 // In the future this would be determined ergonomically, based
4244 4246 // on #cpu's, # active mutator threads (and load), and mutation rate.
4245 4247 int num_workers = ConcGCThreads;
4246 4248
4247 4249 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4248 4250 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
4249 4251
4250 4252 CMSConcMarkingTask tsk(this,
4251 4253 cms_space,
4252 4254 perm_space,
4253 4255 asynch,
4254 4256 conc_workers(),
4255 4257 task_queues());
4256 4258
4257 4259 // Since the actual number of workers we get may be different
4258 4260 // from the number we requested above, do we need to do anything different
4259 4261 // below? In particular, may be we need to subclass the SequantialSubTasksDone
4260 4262 // class?? XXX
4261 4263 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4262 4264 perm_space->initialize_sequential_subtasks_for_marking(num_workers);
4263 4265
4264 4266 // Refs discovery is already non-atomic.
4265 4267 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4266 4268 // Mutate the Refs discovery so it is MT during the
4267 4269 // multi-threaded marking phase.
4268 4270 ReferenceProcessorMTMutator mt(ref_processor(), num_workers > 1);
4269 4271 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());)
4270 4272 conc_workers()->start_task(&tsk);
4271 4273 while (tsk.yielded()) {
4272 4274 tsk.coordinator_yield();
4273 4275 conc_workers()->continue_task(&tsk);
4274 4276 }
4275 4277 // If the task was aborted, _restart_addr will be non-NULL
4276 4278 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4277 4279 while (_restart_addr != NULL) {
4278 4280 // XXX For now we do not make use of ABORTED state and have not
4279 4281 // yet implemented the right abort semantics (even in the original
4280 4282 // single-threaded CMS case). That needs some more investigation
4281 4283 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4282 4284 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4283 4285 // If _restart_addr is non-NULL, a marking stack overflow
4284 4286 // occurred; we need to do a fresh marking iteration from the
4285 4287 // indicated restart address.
4286 4288 if (_foregroundGCIsActive && asynch) {
4287 4289 // We may be running into repeated stack overflows, having
4288 4290 // reached the limit of the stack size, while making very
4289 4291 // slow forward progress. It may be best to bail out and
4290 4292 // let the foreground collector do its job.
4291 4293 // Clear _restart_addr, so that foreground GC
4292 4294 // works from scratch. This avoids the headache of
4293 4295 // a "rescan" which would otherwise be needed because
4294 4296 // of the dirty mod union table & card table.
4295 4297 _restart_addr = NULL;
4296 4298 return false;
4297 4299 }
4298 4300 // Adjust the task to restart from _restart_addr
4299 4301 tsk.reset(_restart_addr);
4300 4302 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4301 4303 _restart_addr);
4302 4304 perm_space->initialize_sequential_subtasks_for_marking(num_workers,
4303 4305 _restart_addr);
4304 4306 _restart_addr = NULL;
4305 4307 // Get the workers going again
4306 4308 conc_workers()->start_task(&tsk);
4307 4309 while (tsk.yielded()) {
4308 4310 tsk.coordinator_yield();
4309 4311 conc_workers()->continue_task(&tsk);
4310 4312 }
4311 4313 }
4312 4314 assert(tsk.completed(), "Inconsistency");
4313 4315 assert(tsk.result() == true, "Inconsistency");
4314 4316 return true;
4315 4317 }
4316 4318
4317 4319 bool CMSCollector::do_marking_st(bool asynch) {
4318 4320 ResourceMark rm;
4319 4321 HandleMark hm;
4320 4322
4321 4323 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4322 4324 &_markStack, &_revisitStack, CMSYield && asynch);
4323 4325 // the last argument to iterate indicates whether the iteration
4324 4326 // should be incremental with periodic yields.
4325 4327 _markBitMap.iterate(&markFromRootsClosure);
4326 4328 // If _restart_addr is non-NULL, a marking stack overflow
4327 4329 // occurred; we need to do a fresh iteration from the
4328 4330 // indicated restart address.
4329 4331 while (_restart_addr != NULL) {
4330 4332 if (_foregroundGCIsActive && asynch) {
4331 4333 // We may be running into repeated stack overflows, having
4332 4334 // reached the limit of the stack size, while making very
4333 4335 // slow forward progress. It may be best to bail out and
4334 4336 // let the foreground collector do its job.
4335 4337 // Clear _restart_addr, so that foreground GC
4336 4338 // works from scratch. This avoids the headache of
4337 4339 // a "rescan" which would otherwise be needed because
4338 4340 // of the dirty mod union table & card table.
4339 4341 _restart_addr = NULL;
4340 4342 return false; // indicating failure to complete marking
4341 4343 }
4342 4344 // Deal with stack overflow:
4343 4345 // we restart marking from _restart_addr
4344 4346 HeapWord* ra = _restart_addr;
4345 4347 markFromRootsClosure.reset(ra);
4346 4348 _restart_addr = NULL;
4347 4349 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4348 4350 }
4349 4351 return true;
4350 4352 }
4351 4353
4352 4354 void CMSCollector::preclean() {
4353 4355 check_correct_thread_executing();
4354 4356 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4355 4357 verify_work_stacks_empty();
4356 4358 verify_overflow_empty();
4357 4359 _abort_preclean = false;
4358 4360 if (CMSPrecleaningEnabled) {
4359 4361 // Precleaning is currently not MT but the reference processor
4360 4362 // may be set for MT. Disable it temporarily here.
4361 4363 ReferenceProcessor* rp = ref_processor();
4362 4364 ReferenceProcessorMTProcMutator z(rp, false);
4363 4365 _eden_chunk_index = 0;
4364 4366 size_t used = get_eden_used();
4365 4367 size_t capacity = get_eden_capacity();
4366 4368 // Don't start sampling unless we will get sufficiently
4367 4369 // many samples.
4368 4370 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4369 4371 * CMSScheduleRemarkEdenPenetration)) {
4370 4372 _start_sampling = true;
4371 4373 } else {
4372 4374 _start_sampling = false;
4373 4375 }
4374 4376 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4375 4377 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
4376 4378 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4377 4379 }
4378 4380 CMSTokenSync x(true); // is cms thread
4379 4381 if (CMSPrecleaningEnabled) {
4380 4382 sample_eden();
4381 4383 _collectorState = AbortablePreclean;
4382 4384 } else {
4383 4385 _collectorState = FinalMarking;
4384 4386 }
4385 4387 verify_work_stacks_empty();
4386 4388 verify_overflow_empty();
4387 4389 }
4388 4390
4389 4391 // Try and schedule the remark such that young gen
4390 4392 // occupancy is CMSScheduleRemarkEdenPenetration %.
4391 4393 void CMSCollector::abortable_preclean() {
4392 4394 check_correct_thread_executing();
4393 4395 assert(CMSPrecleaningEnabled, "Inconsistent control state");
4394 4396 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4395 4397
4396 4398 // If Eden's current occupancy is below this threshold,
4397 4399 // immediately schedule the remark; else preclean
4398 4400 // past the next scavenge in an effort to
4399 4401 // schedule the pause as described avove. By choosing
4400 4402 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4401 4403 // we will never do an actual abortable preclean cycle.
4402 4404 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4403 4405 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4404 4406 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
4405 4407 // We need more smarts in the abortable preclean
4406 4408 // loop below to deal with cases where allocation
4407 4409 // in young gen is very very slow, and our precleaning
4408 4410 // is running a losing race against a horde of
4409 4411 // mutators intent on flooding us with CMS updates
4410 4412 // (dirty cards).
4411 4413 // One, admittedly dumb, strategy is to give up
4412 4414 // after a certain number of abortable precleaning loops
4413 4415 // or after a certain maximum time. We want to make
4414 4416 // this smarter in the next iteration.
4415 4417 // XXX FIX ME!!! YSR
4416 4418 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4417 4419 while (!(should_abort_preclean() ||
4418 4420 ConcurrentMarkSweepThread::should_terminate())) {
4419 4421 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4420 4422 cumworkdone += workdone;
4421 4423 loops++;
4422 4424 // Voluntarily terminate abortable preclean phase if we have
4423 4425 // been at it for too long.
4424 4426 if ((CMSMaxAbortablePrecleanLoops != 0) &&
4425 4427 loops >= CMSMaxAbortablePrecleanLoops) {
4426 4428 if (PrintGCDetails) {
4427 4429 gclog_or_tty->print(" CMS: abort preclean due to loops ");
4428 4430 }
4429 4431 break;
4430 4432 }
4431 4433 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4432 4434 if (PrintGCDetails) {
4433 4435 gclog_or_tty->print(" CMS: abort preclean due to time ");
4434 4436 }
4435 4437 break;
4436 4438 }
4437 4439 // If we are doing little work each iteration, we should
4438 4440 // take a short break.
4439 4441 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4440 4442 // Sleep for some time, waiting for work to accumulate
4441 4443 stopTimer();
4442 4444 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4443 4445 startTimer();
4444 4446 waited++;
4445 4447 }
4446 4448 }
4447 4449 if (PrintCMSStatistics > 0) {
4448 4450 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4449 4451 loops, waited, cumworkdone);
4450 4452 }
4451 4453 }
4452 4454 CMSTokenSync x(true); // is cms thread
4453 4455 if (_collectorState != Idling) {
4454 4456 assert(_collectorState == AbortablePreclean,
4455 4457 "Spontaneous state transition?");
4456 4458 _collectorState = FinalMarking;
4457 4459 } // Else, a foreground collection completed this CMS cycle.
4458 4460 return;
4459 4461 }
4460 4462
4461 4463 // Respond to an Eden sampling opportunity
4462 4464 void CMSCollector::sample_eden() {
4463 4465 // Make sure a young gc cannot sneak in between our
4464 4466 // reading and recording of a sample.
4465 4467 assert(Thread::current()->is_ConcurrentGC_thread(),
4466 4468 "Only the cms thread may collect Eden samples");
4467 4469 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4468 4470 "Should collect samples while holding CMS token");
4469 4471 if (!_start_sampling) {
4470 4472 return;
4471 4473 }
4472 4474 if (_eden_chunk_array) {
4473 4475 if (_eden_chunk_index < _eden_chunk_capacity) {
4474 4476 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
4475 4477 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4476 4478 "Unexpected state of Eden");
4477 4479 // We'd like to check that what we just sampled is an oop-start address;
4478 4480 // however, we cannot do that here since the object may not yet have been
4479 4481 // initialized. So we'll instead do the check when we _use_ this sample
4480 4482 // later.
4481 4483 if (_eden_chunk_index == 0 ||
4482 4484 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4483 4485 _eden_chunk_array[_eden_chunk_index-1])
4484 4486 >= CMSSamplingGrain)) {
4485 4487 _eden_chunk_index++; // commit sample
4486 4488 }
4487 4489 }
4488 4490 }
4489 4491 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4490 4492 size_t used = get_eden_used();
4491 4493 size_t capacity = get_eden_capacity();
4492 4494 assert(used <= capacity, "Unexpected state of Eden");
4493 4495 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4494 4496 _abort_preclean = true;
4495 4497 }
4496 4498 }
4497 4499 }
4498 4500
4499 4501
4500 4502 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4501 4503 assert(_collectorState == Precleaning ||
4502 4504 _collectorState == AbortablePreclean, "incorrect state");
4503 4505 ResourceMark rm;
4504 4506 HandleMark hm;
4505 4507 // Do one pass of scrubbing the discovered reference lists
4506 4508 // to remove any reference objects with strongly-reachable
4507 4509 // referents.
4508 4510 if (clean_refs) {
4509 4511 ReferenceProcessor* rp = ref_processor();
4510 4512 CMSPrecleanRefsYieldClosure yield_cl(this);
4511 4513 assert(rp->span().equals(_span), "Spans should be equal");
4512 4514 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4513 4515 &_markStack, &_revisitStack,
4514 4516 true /* preclean */);
4515 4517 CMSDrainMarkingStackClosure complete_trace(this,
4516 4518 _span, &_markBitMap, &_markStack,
4517 4519 &keep_alive, true /* preclean */);
4518 4520
4519 4521 // We don't want this step to interfere with a young
4520 4522 // collection because we don't want to take CPU
4521 4523 // or memory bandwidth away from the young GC threads
4522 4524 // (which may be as many as there are CPUs).
4523 4525 // Note that we don't need to protect ourselves from
4524 4526 // interference with mutators because they can't
4525 4527 // manipulate the discovered reference lists nor affect
4526 4528 // the computed reachability of the referents, the
4527 4529 // only properties manipulated by the precleaning
4528 4530 // of these reference lists.
4529 4531 stopTimer();
4530 4532 CMSTokenSyncWithLocks x(true /* is cms thread */,
4531 4533 bitMapLock());
4532 4534 startTimer();
4533 4535 sample_eden();
4534 4536
4535 4537 // The following will yield to allow foreground
4536 4538 // collection to proceed promptly. XXX YSR:
4537 4539 // The code in this method may need further
4538 4540 // tweaking for better performance and some restructuring
4539 4541 // for cleaner interfaces.
4540 4542 rp->preclean_discovered_references(
4541 4543 rp->is_alive_non_header(), &keep_alive, &complete_trace,
4542 4544 &yield_cl, should_unload_classes());
4543 4545 }
4544 4546
4545 4547 if (clean_survivor) { // preclean the active survivor space(s)
4546 4548 assert(_young_gen->kind() == Generation::DefNew ||
4547 4549 _young_gen->kind() == Generation::ParNew ||
4548 4550 _young_gen->kind() == Generation::ASParNew,
4549 4551 "incorrect type for cast");
4550 4552 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4551 4553 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4552 4554 &_markBitMap, &_modUnionTable,
4553 4555 &_markStack, &_revisitStack,
4554 4556 true /* precleaning phase */);
4555 4557 stopTimer();
4556 4558 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4557 4559 bitMapLock());
4558 4560 startTimer();
4559 4561 unsigned int before_count =
4560 4562 GenCollectedHeap::heap()->total_collections();
4561 4563 SurvivorSpacePrecleanClosure
4562 4564 sss_cl(this, _span, &_markBitMap, &_markStack,
4563 4565 &pam_cl, before_count, CMSYield);
4564 4566 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4565 4567 dng->from()->object_iterate_careful(&sss_cl);
4566 4568 dng->to()->object_iterate_careful(&sss_cl);
4567 4569 }
4568 4570 MarkRefsIntoAndScanClosure
4569 4571 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4570 4572 &_markStack, &_revisitStack, this, CMSYield,
4571 4573 true /* precleaning phase */);
4572 4574 // CAUTION: The following closure has persistent state that may need to
4573 4575 // be reset upon a decrease in the sequence of addresses it
4574 4576 // processes.
4575 4577 ScanMarkedObjectsAgainCarefullyClosure
4576 4578 smoac_cl(this, _span,
4577 4579 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield);
4578 4580
4579 4581 // Preclean dirty cards in ModUnionTable and CardTable using
4580 4582 // appropriate convergence criterion;
4581 4583 // repeat CMSPrecleanIter times unless we find that
4582 4584 // we are losing.
4583 4585 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4584 4586 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4585 4587 "Bad convergence multiplier");
4586 4588 assert(CMSPrecleanThreshold >= 100,
4587 4589 "Unreasonably low CMSPrecleanThreshold");
4588 4590
4589 4591 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4590 4592 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4591 4593 numIter < CMSPrecleanIter;
4592 4594 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4593 4595 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4594 4596 if (CMSPermGenPrecleaningEnabled) {
4595 4597 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl);
4596 4598 }
4597 4599 if (Verbose && PrintGCDetails) {
4598 4600 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4599 4601 }
4600 4602 // Either there are very few dirty cards, so re-mark
4601 4603 // pause will be small anyway, or our pre-cleaning isn't
4602 4604 // that much faster than the rate at which cards are being
4603 4605 // dirtied, so we might as well stop and re-mark since
4604 4606 // precleaning won't improve our re-mark time by much.
4605 4607 if (curNumCards <= CMSPrecleanThreshold ||
4606 4608 (numIter > 0 &&
4607 4609 (curNumCards * CMSPrecleanDenominator >
4608 4610 lastNumCards * CMSPrecleanNumerator))) {
4609 4611 numIter++;
4610 4612 cumNumCards += curNumCards;
4611 4613 break;
4612 4614 }
4613 4615 }
4614 4616 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4615 4617 if (CMSPermGenPrecleaningEnabled) {
4616 4618 curNumCards += preclean_card_table(_permGen, &smoac_cl);
4617 4619 }
4618 4620 cumNumCards += curNumCards;
4619 4621 if (PrintGCDetails && PrintCMSStatistics != 0) {
4620 4622 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4621 4623 curNumCards, cumNumCards, numIter);
4622 4624 }
4623 4625 return cumNumCards; // as a measure of useful work done
4624 4626 }
4625 4627
4626 4628 // PRECLEANING NOTES:
4627 4629 // Precleaning involves:
4628 4630 // . reading the bits of the modUnionTable and clearing the set bits.
4629 4631 // . For the cards corresponding to the set bits, we scan the
4630 4632 // objects on those cards. This means we need the free_list_lock
4631 4633 // so that we can safely iterate over the CMS space when scanning
4632 4634 // for oops.
4633 4635 // . When we scan the objects, we'll be both reading and setting
4634 4636 // marks in the marking bit map, so we'll need the marking bit map.
4635 4637 // . For protecting _collector_state transitions, we take the CGC_lock.
4636 4638 // Note that any races in the reading of of card table entries by the
4637 4639 // CMS thread on the one hand and the clearing of those entries by the
4638 4640 // VM thread or the setting of those entries by the mutator threads on the
4639 4641 // other are quite benign. However, for efficiency it makes sense to keep
4640 4642 // the VM thread from racing with the CMS thread while the latter is
4641 4643 // dirty card info to the modUnionTable. We therefore also use the
4642 4644 // CGC_lock to protect the reading of the card table and the mod union
4643 4645 // table by the CM thread.
4644 4646 // . We run concurrently with mutator updates, so scanning
4645 4647 // needs to be done carefully -- we should not try to scan
4646 4648 // potentially uninitialized objects.
4647 4649 //
4648 4650 // Locking strategy: While holding the CGC_lock, we scan over and
4649 4651 // reset a maximal dirty range of the mod union / card tables, then lock
4650 4652 // the free_list_lock and bitmap lock to do a full marking, then
4651 4653 // release these locks; and repeat the cycle. This allows for a
4652 4654 // certain amount of fairness in the sharing of these locks between
4653 4655 // the CMS collector on the one hand, and the VM thread and the
4654 4656 // mutators on the other.
4655 4657
4656 4658 // NOTE: preclean_mod_union_table() and preclean_card_table()
4657 4659 // further below are largely identical; if you need to modify
4658 4660 // one of these methods, please check the other method too.
4659 4661
4660 4662 size_t CMSCollector::preclean_mod_union_table(
4661 4663 ConcurrentMarkSweepGeneration* gen,
4662 4664 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4663 4665 verify_work_stacks_empty();
4664 4666 verify_overflow_empty();
4665 4667
4666 4668 // Turn off checking for this method but turn it back on
4667 4669 // selectively. There are yield points in this method
4668 4670 // but it is difficult to turn the checking off just around
4669 4671 // the yield points. It is simpler to selectively turn
4670 4672 // it on.
4671 4673 DEBUG_ONLY(RememberKlassesChecker mux(false);)
4672 4674
4673 4675 // strategy: starting with the first card, accumulate contiguous
4674 4676 // ranges of dirty cards; clear these cards, then scan the region
4675 4677 // covered by these cards.
4676 4678
4677 4679 // Since all of the MUT is committed ahead, we can just use
4678 4680 // that, in case the generations expand while we are precleaning.
4679 4681 // It might also be fine to just use the committed part of the
4680 4682 // generation, but we might potentially miss cards when the
4681 4683 // generation is rapidly expanding while we are in the midst
4682 4684 // of precleaning.
4683 4685 HeapWord* startAddr = gen->reserved().start();
4684 4686 HeapWord* endAddr = gen->reserved().end();
4685 4687
4686 4688 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4687 4689
4688 4690 size_t numDirtyCards, cumNumDirtyCards;
4689 4691 HeapWord *nextAddr, *lastAddr;
4690 4692 for (cumNumDirtyCards = numDirtyCards = 0,
4691 4693 nextAddr = lastAddr = startAddr;
4692 4694 nextAddr < endAddr;
4693 4695 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4694 4696
4695 4697 ResourceMark rm;
4696 4698 HandleMark hm;
4697 4699
4698 4700 MemRegion dirtyRegion;
4699 4701 {
4700 4702 stopTimer();
4701 4703 // Potential yield point
4702 4704 CMSTokenSync ts(true);
4703 4705 startTimer();
4704 4706 sample_eden();
4705 4707 // Get dirty region starting at nextOffset (inclusive),
4706 4708 // simultaneously clearing it.
4707 4709 dirtyRegion =
4708 4710 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4709 4711 assert(dirtyRegion.start() >= nextAddr,
4710 4712 "returned region inconsistent?");
4711 4713 }
4712 4714 // Remember where the next search should begin.
4713 4715 // The returned region (if non-empty) is a right open interval,
4714 4716 // so lastOffset is obtained from the right end of that
4715 4717 // interval.
4716 4718 lastAddr = dirtyRegion.end();
4717 4719 // Should do something more transparent and less hacky XXX
4718 4720 numDirtyCards =
4719 4721 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4720 4722
4721 4723 // We'll scan the cards in the dirty region (with periodic
4722 4724 // yields for foreground GC as needed).
4723 4725 if (!dirtyRegion.is_empty()) {
4724 4726 assert(numDirtyCards > 0, "consistency check");
4725 4727 HeapWord* stop_point = NULL;
4726 4728 stopTimer();
4727 4729 // Potential yield point
4728 4730 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4729 4731 bitMapLock());
4730 4732 startTimer();
4731 4733 {
4732 4734 verify_work_stacks_empty();
4733 4735 verify_overflow_empty();
4734 4736 sample_eden();
4735 4737 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4736 4738 stop_point =
4737 4739 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4738 4740 }
4739 4741 if (stop_point != NULL) {
4740 4742 // The careful iteration stopped early either because it found an
4741 4743 // uninitialized object, or because we were in the midst of an
4742 4744 // "abortable preclean", which should now be aborted. Redirty
4743 4745 // the bits corresponding to the partially-scanned or unscanned
4744 4746 // cards. We'll either restart at the next block boundary or
4745 4747 // abort the preclean.
4746 4748 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) ||
4747 4749 (_collectorState == AbortablePreclean && should_abort_preclean()),
4748 4750 "Unparsable objects should only be in perm gen.");
4749 4751 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4750 4752 if (should_abort_preclean()) {
4751 4753 break; // out of preclean loop
4752 4754 } else {
4753 4755 // Compute the next address at which preclean should pick up;
4754 4756 // might need bitMapLock in order to read P-bits.
4755 4757 lastAddr = next_card_start_after_block(stop_point);
4756 4758 }
4757 4759 }
4758 4760 } else {
4759 4761 assert(lastAddr == endAddr, "consistency check");
4760 4762 assert(numDirtyCards == 0, "consistency check");
4761 4763 break;
4762 4764 }
4763 4765 }
4764 4766 verify_work_stacks_empty();
4765 4767 verify_overflow_empty();
4766 4768 return cumNumDirtyCards;
4767 4769 }
4768 4770
4769 4771 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4770 4772 // below are largely identical; if you need to modify
4771 4773 // one of these methods, please check the other method too.
4772 4774
4773 4775 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4774 4776 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4775 4777 // strategy: it's similar to precleamModUnionTable above, in that
4776 4778 // we accumulate contiguous ranges of dirty cards, mark these cards
4777 4779 // precleaned, then scan the region covered by these cards.
4778 4780 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4779 4781 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4780 4782
4781 4783 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4782 4784
4783 4785 size_t numDirtyCards, cumNumDirtyCards;
4784 4786 HeapWord *lastAddr, *nextAddr;
4785 4787
4786 4788 for (cumNumDirtyCards = numDirtyCards = 0,
4787 4789 nextAddr = lastAddr = startAddr;
4788 4790 nextAddr < endAddr;
4789 4791 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4790 4792
4791 4793 ResourceMark rm;
4792 4794 HandleMark hm;
4793 4795
4794 4796 MemRegion dirtyRegion;
4795 4797 {
4796 4798 // See comments in "Precleaning notes" above on why we
4797 4799 // do this locking. XXX Could the locking overheads be
4798 4800 // too high when dirty cards are sparse? [I don't think so.]
4799 4801 stopTimer();
4800 4802 CMSTokenSync x(true); // is cms thread
4801 4803 startTimer();
4802 4804 sample_eden();
4803 4805 // Get and clear dirty region from card table
4804 4806 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4805 4807 MemRegion(nextAddr, endAddr),
4806 4808 true,
4807 4809 CardTableModRefBS::precleaned_card_val());
4808 4810
4809 4811 assert(dirtyRegion.start() >= nextAddr,
4810 4812 "returned region inconsistent?");
4811 4813 }
4812 4814 lastAddr = dirtyRegion.end();
4813 4815 numDirtyCards =
4814 4816 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4815 4817
4816 4818 if (!dirtyRegion.is_empty()) {
4817 4819 stopTimer();
4818 4820 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4819 4821 startTimer();
4820 4822 sample_eden();
4821 4823 verify_work_stacks_empty();
4822 4824 verify_overflow_empty();
4823 4825 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4824 4826 HeapWord* stop_point =
4825 4827 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4826 4828 if (stop_point != NULL) {
4827 4829 // The careful iteration stopped early because it found an
4828 4830 // uninitialized object. Redirty the bits corresponding to the
4829 4831 // partially-scanned or unscanned cards, and start again at the
4830 4832 // next block boundary.
4831 4833 assert(CMSPermGenPrecleaningEnabled ||
4832 4834 (_collectorState == AbortablePreclean && should_abort_preclean()),
4833 4835 "Unparsable objects should only be in perm gen.");
4834 4836 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4835 4837 if (should_abort_preclean()) {
4836 4838 break; // out of preclean loop
4837 4839 } else {
4838 4840 // Compute the next address at which preclean should pick up.
4839 4841 lastAddr = next_card_start_after_block(stop_point);
4840 4842 }
4841 4843 }
4842 4844 } else {
4843 4845 break;
4844 4846 }
4845 4847 }
4846 4848 verify_work_stacks_empty();
4847 4849 verify_overflow_empty();
4848 4850 return cumNumDirtyCards;
4849 4851 }
4850 4852
4851 4853 void CMSCollector::checkpointRootsFinal(bool asynch,
4852 4854 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4853 4855 assert(_collectorState == FinalMarking, "incorrect state transition?");
4854 4856 check_correct_thread_executing();
4855 4857 // world is stopped at this checkpoint
4856 4858 assert(SafepointSynchronize::is_at_safepoint(),
4857 4859 "world should be stopped");
4858 4860 TraceCMSMemoryManagerStats tms(_collectorState);
4859 4861 verify_work_stacks_empty();
4860 4862 verify_overflow_empty();
4861 4863
4862 4864 SpecializationStats::clear();
4863 4865 if (PrintGCDetails) {
4864 4866 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4865 4867 _young_gen->used() / K,
4866 4868 _young_gen->capacity() / K);
4867 4869 }
4868 4870 if (asynch) {
4869 4871 if (CMSScavengeBeforeRemark) {
4870 4872 GenCollectedHeap* gch = GenCollectedHeap::heap();
4871 4873 // Temporarily set flag to false, GCH->do_collection will
4872 4874 // expect it to be false and set to true
4873 4875 FlagSetting fl(gch->_is_gc_active, false);
4874 4876 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark",
4875 4877 PrintGCDetails && Verbose, true, gclog_or_tty);)
4876 4878 int level = _cmsGen->level() - 1;
4877 4879 if (level >= 0) {
4878 4880 gch->do_collection(true, // full (i.e. force, see below)
4879 4881 false, // !clear_all_soft_refs
4880 4882 0, // size
4881 4883 false, // is_tlab
4882 4884 level // max_level
4883 4885 );
4884 4886 }
4885 4887 }
4886 4888 FreelistLocker x(this);
4887 4889 MutexLockerEx y(bitMapLock(),
4888 4890 Mutex::_no_safepoint_check_flag);
4889 4891 assert(!init_mark_was_synchronous, "but that's impossible!");
4890 4892 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
4891 4893 } else {
4892 4894 // already have all the locks
4893 4895 checkpointRootsFinalWork(asynch, clear_all_soft_refs,
4894 4896 init_mark_was_synchronous);
4895 4897 }
4896 4898 verify_work_stacks_empty();
4897 4899 verify_overflow_empty();
4898 4900 SpecializationStats::print();
4899 4901 }
4900 4902
4901 4903 void CMSCollector::checkpointRootsFinalWork(bool asynch,
4902 4904 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4903 4905
4904 4906 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);)
4905 4907
4906 4908 assert(haveFreelistLocks(), "must have free list locks");
4907 4909 assert_lock_strong(bitMapLock());
4908 4910
4909 4911 if (UseAdaptiveSizePolicy) {
4910 4912 size_policy()->checkpoint_roots_final_begin();
4911 4913 }
4912 4914
4913 4915 ResourceMark rm;
4914 4916 HandleMark hm;
4915 4917
4916 4918 GenCollectedHeap* gch = GenCollectedHeap::heap();
4917 4919
4918 4920 if (should_unload_classes()) {
4919 4921 CodeCache::gc_prologue();
4920 4922 }
4921 4923 assert(haveFreelistLocks(), "must have free list locks");
4922 4924 assert_lock_strong(bitMapLock());
4923 4925
4924 4926 DEBUG_ONLY(RememberKlassesChecker fmx(should_unload_classes());)
4925 4927 if (!init_mark_was_synchronous) {
4926 4928 // We might assume that we need not fill TLAB's when
4927 4929 // CMSScavengeBeforeRemark is set, because we may have just done
4928 4930 // a scavenge which would have filled all TLAB's -- and besides
4929 4931 // Eden would be empty. This however may not always be the case --
4930 4932 // for instance although we asked for a scavenge, it may not have
4931 4933 // happened because of a JNI critical section. We probably need
4932 4934 // a policy for deciding whether we can in that case wait until
4933 4935 // the critical section releases and then do the remark following
4934 4936 // the scavenge, and skip it here. In the absence of that policy,
4935 4937 // or of an indication of whether the scavenge did indeed occur,
4936 4938 // we cannot rely on TLAB's having been filled and must do
4937 4939 // so here just in case a scavenge did not happen.
4938 4940 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4939 4941 // Update the saved marks which may affect the root scans.
4940 4942 gch->save_marks();
4941 4943
4942 4944 {
4943 4945 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4944 4946
4945 4947 // Note on the role of the mod union table:
4946 4948 // Since the marker in "markFromRoots" marks concurrently with
4947 4949 // mutators, it is possible for some reachable objects not to have been
4948 4950 // scanned. For instance, an only reference to an object A was
4949 4951 // placed in object B after the marker scanned B. Unless B is rescanned,
4950 4952 // A would be collected. Such updates to references in marked objects
4951 4953 // are detected via the mod union table which is the set of all cards
4952 4954 // dirtied since the first checkpoint in this GC cycle and prior to
4953 4955 // the most recent young generation GC, minus those cleaned up by the
4954 4956 // concurrent precleaning.
4955 4957 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
4956 4958 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty);
4957 4959 do_remark_parallel();
4958 4960 } else {
4959 4961 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4960 4962 gclog_or_tty);
4961 4963 do_remark_non_parallel();
4962 4964 }
4963 4965 }
4964 4966 } else {
4965 4967 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
4966 4968 // The initial mark was stop-world, so there's no rescanning to
4967 4969 // do; go straight on to the next step below.
4968 4970 }
4969 4971 verify_work_stacks_empty();
4970 4972 verify_overflow_empty();
4971 4973
4972 4974 {
4973 4975 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);)
4974 4976 refProcessingWork(asynch, clear_all_soft_refs);
4975 4977 }
4976 4978 verify_work_stacks_empty();
4977 4979 verify_overflow_empty();
4978 4980
4979 4981 if (should_unload_classes()) {
4980 4982 CodeCache::gc_epilogue();
4981 4983 }
4982 4984 JvmtiExport::gc_epilogue();
4983 4985
4984 4986 // If we encountered any (marking stack / work queue) overflow
4985 4987 // events during the current CMS cycle, take appropriate
4986 4988 // remedial measures, where possible, so as to try and avoid
4987 4989 // recurrence of that condition.
4988 4990 assert(_markStack.isEmpty(), "No grey objects");
4989 4991 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4990 4992 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4991 4993 if (ser_ovflw > 0) {
4992 4994 if (PrintCMSStatistics != 0) {
4993 4995 gclog_or_tty->print_cr("Marking stack overflow (benign) "
4994 4996 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
4995 4997 ", kac_preclean="SIZE_FORMAT")",
4996 4998 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4997 4999 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4998 5000 }
4999 5001 _markStack.expand();
5000 5002 _ser_pmc_remark_ovflw = 0;
5001 5003 _ser_pmc_preclean_ovflw = 0;
5002 5004 _ser_kac_preclean_ovflw = 0;
5003 5005 _ser_kac_ovflw = 0;
5004 5006 }
5005 5007 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
5006 5008 if (PrintCMSStatistics != 0) {
5007 5009 gclog_or_tty->print_cr("Work queue overflow (benign) "
5008 5010 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
5009 5011 _par_pmc_remark_ovflw, _par_kac_ovflw);
5010 5012 }
5011 5013 _par_pmc_remark_ovflw = 0;
5012 5014 _par_kac_ovflw = 0;
5013 5015 }
5014 5016 if (PrintCMSStatistics != 0) {
5015 5017 if (_markStack._hit_limit > 0) {
5016 5018 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
5017 5019 _markStack._hit_limit);
5018 5020 }
5019 5021 if (_markStack._failed_double > 0) {
5020 5022 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
5021 5023 " current capacity "SIZE_FORMAT,
5022 5024 _markStack._failed_double,
5023 5025 _markStack.capacity());
5024 5026 }
5025 5027 }
5026 5028 _markStack._hit_limit = 0;
5027 5029 _markStack._failed_double = 0;
5028 5030
5029 5031 // Check that all the klasses have been checked
5030 5032 assert(_revisitStack.isEmpty(), "Not all klasses revisited");
5031 5033
5032 5034 if ((VerifyAfterGC || VerifyDuringGC) &&
5033 5035 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5034 5036 verify_after_remark();
5035 5037 }
5036 5038
5037 5039 // Change under the freelistLocks.
5038 5040 _collectorState = Sweeping;
5039 5041 // Call isAllClear() under bitMapLock
5040 5042 assert(_modUnionTable.isAllClear(), "Should be clear by end of the"
5041 5043 " final marking");
5042 5044 if (UseAdaptiveSizePolicy) {
5043 5045 size_policy()->checkpoint_roots_final_end(gch->gc_cause());
5044 5046 }
5045 5047 }
5046 5048
5047 5049 // Parallel remark task
5048 5050 class CMSParRemarkTask: public AbstractGangTask {
5049 5051 CMSCollector* _collector;
5050 5052 int _n_workers;
5051 5053 CompactibleFreeListSpace* _cms_space;
5052 5054 CompactibleFreeListSpace* _perm_space;
5053 5055
5054 5056 // The per-thread work queues, available here for stealing.
5055 5057 OopTaskQueueSet* _task_queues;
5056 5058 ParallelTaskTerminator _term;
5057 5059
5058 5060 public:
5059 5061 CMSParRemarkTask(CMSCollector* collector,
5060 5062 CompactibleFreeListSpace* cms_space,
5061 5063 CompactibleFreeListSpace* perm_space,
5062 5064 int n_workers, FlexibleWorkGang* workers,
5063 5065 OopTaskQueueSet* task_queues):
5064 5066 AbstractGangTask("Rescan roots and grey objects in parallel"),
5065 5067 _collector(collector),
5066 5068 _cms_space(cms_space), _perm_space(perm_space),
5067 5069 _n_workers(n_workers),
5068 5070 _task_queues(task_queues),
5069 5071 _term(n_workers, task_queues) { }
5070 5072
5071 5073 OopTaskQueueSet* task_queues() { return _task_queues; }
5072 5074
5073 5075 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5074 5076
5075 5077 ParallelTaskTerminator* terminator() { return &_term; }
5076 5078 int n_workers() { return _n_workers; }
5077 5079
5078 5080 void work(int i);
5079 5081
5080 5082 private:
5081 5083 // Work method in support of parallel rescan ... of young gen spaces
5082 5084 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl,
5083 5085 ContiguousSpace* space,
5084 5086 HeapWord** chunk_array, size_t chunk_top);
5085 5087
5086 5088 // ... of dirty cards in old space
5087 5089 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
5088 5090 Par_MarkRefsIntoAndScanClosure* cl);
5089 5091
5090 5092 // ... work stealing for the above
5091 5093 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
5092 5094 };
5093 5095
5094 5096 // work_queue(i) is passed to the closure
5095 5097 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
5096 5098 // also is passed to do_dirty_card_rescan_tasks() and to
5097 5099 // do_work_steal() to select the i-th task_queue.
5098 5100
5099 5101 void CMSParRemarkTask::work(int i) {
5100 5102 elapsedTimer _timer;
5101 5103 ResourceMark rm;
5102 5104 HandleMark hm;
5103 5105
5104 5106 // ---------- rescan from roots --------------
5105 5107 _timer.start();
5106 5108 GenCollectedHeap* gch = GenCollectedHeap::heap();
5107 5109 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
5108 5110 _collector->_span, _collector->ref_processor(),
5109 5111 &(_collector->_markBitMap),
5110 5112 work_queue(i), &(_collector->_revisitStack));
5111 5113
5112 5114 // Rescan young gen roots first since these are likely
5113 5115 // coarsely partitioned and may, on that account, constitute
5114 5116 // the critical path; thus, it's best to start off that
5115 5117 // work first.
5116 5118 // ---------- young gen roots --------------
5117 5119 {
5118 5120 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
5119 5121 EdenSpace* eden_space = dng->eden();
5120 5122 ContiguousSpace* from_space = dng->from();
5121 5123 ContiguousSpace* to_space = dng->to();
5122 5124
5123 5125 HeapWord** eca = _collector->_eden_chunk_array;
5124 5126 size_t ect = _collector->_eden_chunk_index;
5125 5127 HeapWord** sca = _collector->_survivor_chunk_array;
5126 5128 size_t sct = _collector->_survivor_chunk_index;
5127 5129
5128 5130 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
5129 5131 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
5130 5132
5131 5133 do_young_space_rescan(i, &par_mrias_cl, to_space, NULL, 0);
5132 5134 do_young_space_rescan(i, &par_mrias_cl, from_space, sca, sct);
5133 5135 do_young_space_rescan(i, &par_mrias_cl, eden_space, eca, ect);
5134 5136
5135 5137 _timer.stop();
5136 5138 if (PrintCMSStatistics != 0) {
5137 5139 gclog_or_tty->print_cr(
5138 5140 "Finished young gen rescan work in %dth thread: %3.3f sec",
5139 5141 i, _timer.seconds());
5140 5142 }
5141 5143 }
5142 5144
5143 5145 // ---------- remaining roots --------------
5144 5146 _timer.reset();
5145 5147 _timer.start();
5146 5148 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
5147 5149 false, // yg was scanned above
5148 5150 false, // this is parallel code
5149 5151 true, // collecting perm gen
5150 5152 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5151 5153 &par_mrias_cl,
5152 5154 true, // walk all of code cache if (so & SO_CodeCache)
5153 5155 NULL);
5154 5156 assert(_collector->should_unload_classes()
5155 5157 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache),
5156 5158 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5157 5159 _timer.stop();
5158 5160 if (PrintCMSStatistics != 0) {
5159 5161 gclog_or_tty->print_cr(
5160 5162 "Finished remaining root rescan work in %dth thread: %3.3f sec",
5161 5163 i, _timer.seconds());
5162 5164 }
5163 5165
5164 5166 // ---------- rescan dirty cards ------------
5165 5167 _timer.reset();
5166 5168 _timer.start();
5167 5169
5168 5170 // Do the rescan tasks for each of the two spaces
5169 5171 // (cms_space and perm_space) in turn.
5170 5172 // "i" is passed to select the "i-th" task_queue
5171 5173 do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl);
5172 5174 do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl);
5173 5175 _timer.stop();
5174 5176 if (PrintCMSStatistics != 0) {
5175 5177 gclog_or_tty->print_cr(
5176 5178 "Finished dirty card rescan work in %dth thread: %3.3f sec",
5177 5179 i, _timer.seconds());
5178 5180 }
5179 5181
5180 5182 // ---------- steal work from other threads ...
5181 5183 // ---------- ... and drain overflow list.
5182 5184 _timer.reset();
5183 5185 _timer.start();
5184 5186 do_work_steal(i, &par_mrias_cl, _collector->hash_seed(i));
5185 5187 _timer.stop();
5186 5188 if (PrintCMSStatistics != 0) {
5187 5189 gclog_or_tty->print_cr(
5188 5190 "Finished work stealing in %dth thread: %3.3f sec",
5189 5191 i, _timer.seconds());
5190 5192 }
5191 5193 }
5192 5194
5193 5195 // Note that parameter "i" is not used.
5194 5196 void
5195 5197 CMSParRemarkTask::do_young_space_rescan(int i,
5196 5198 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space,
5197 5199 HeapWord** chunk_array, size_t chunk_top) {
5198 5200 // Until all tasks completed:
5199 5201 // . claim an unclaimed task
5200 5202 // . compute region boundaries corresponding to task claimed
5201 5203 // using chunk_array
5202 5204 // . par_oop_iterate(cl) over that region
5203 5205
5204 5206 ResourceMark rm;
5205 5207 HandleMark hm;
5206 5208
5207 5209 SequentialSubTasksDone* pst = space->par_seq_tasks();
5208 5210 assert(pst->valid(), "Uninitialized use?");
5209 5211
5210 5212 int nth_task = 0;
5211 5213 int n_tasks = pst->n_tasks();
5212 5214
5213 5215 HeapWord *start, *end;
5214 5216 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5215 5217 // We claimed task # nth_task; compute its boundaries.
5216 5218 if (chunk_top == 0) { // no samples were taken
5217 5219 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
5218 5220 start = space->bottom();
5219 5221 end = space->top();
5220 5222 } else if (nth_task == 0) {
5221 5223 start = space->bottom();
5222 5224 end = chunk_array[nth_task];
5223 5225 } else if (nth_task < (jint)chunk_top) {
5224 5226 assert(nth_task >= 1, "Control point invariant");
5225 5227 start = chunk_array[nth_task - 1];
5226 5228 end = chunk_array[nth_task];
5227 5229 } else {
5228 5230 assert(nth_task == (jint)chunk_top, "Control point invariant");
5229 5231 start = chunk_array[chunk_top - 1];
5230 5232 end = space->top();
5231 5233 }
5232 5234 MemRegion mr(start, end);
5233 5235 // Verify that mr is in space
5234 5236 assert(mr.is_empty() || space->used_region().contains(mr),
5235 5237 "Should be in space");
5236 5238 // Verify that "start" is an object boundary
5237 5239 assert(mr.is_empty() || oop(mr.start())->is_oop(),
5238 5240 "Should be an oop");
5239 5241 space->par_oop_iterate(mr, cl);
5240 5242 }
5241 5243 pst->all_tasks_completed();
5242 5244 }
5243 5245
5244 5246 void
5245 5247 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5246 5248 CompactibleFreeListSpace* sp, int i,
5247 5249 Par_MarkRefsIntoAndScanClosure* cl) {
5248 5250 // Until all tasks completed:
5249 5251 // . claim an unclaimed task
5250 5252 // . compute region boundaries corresponding to task claimed
5251 5253 // . transfer dirty bits ct->mut for that region
5252 5254 // . apply rescanclosure to dirty mut bits for that region
5253 5255
5254 5256 ResourceMark rm;
5255 5257 HandleMark hm;
5256 5258
5257 5259 OopTaskQueue* work_q = work_queue(i);
5258 5260 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5259 5261 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5260 5262 // CAUTION: This closure has state that persists across calls to
5261 5263 // the work method dirty_range_iterate_clear() in that it has
5262 5264 // imbedded in it a (subtype of) UpwardsObjectClosure. The
5263 5265 // use of that state in the imbedded UpwardsObjectClosure instance
5264 5266 // assumes that the cards are always iterated (even if in parallel
5265 5267 // by several threads) in monotonically increasing order per each
5266 5268 // thread. This is true of the implementation below which picks
5267 5269 // card ranges (chunks) in monotonically increasing order globally
5268 5270 // and, a-fortiori, in monotonically increasing order per thread
5269 5271 // (the latter order being a subsequence of the former).
5270 5272 // If the work code below is ever reorganized into a more chaotic
5271 5273 // work-partitioning form than the current "sequential tasks"
5272 5274 // paradigm, the use of that persistent state will have to be
5273 5275 // revisited and modified appropriately. See also related
5274 5276 // bug 4756801 work on which should examine this code to make
5275 5277 // sure that the changes there do not run counter to the
5276 5278 // assumptions made here and necessary for correctness and
5277 5279 // efficiency. Note also that this code might yield inefficient
5278 5280 // behaviour in the case of very large objects that span one or
5279 5281 // more work chunks. Such objects would potentially be scanned
5280 5282 // several times redundantly. Work on 4756801 should try and
5281 5283 // address that performance anomaly if at all possible. XXX
5282 5284 MemRegion full_span = _collector->_span;
5283 5285 CMSBitMap* bm = &(_collector->_markBitMap); // shared
5284 5286 CMSMarkStack* rs = &(_collector->_revisitStack); // shared
5285 5287 MarkFromDirtyCardsClosure
5286 5288 greyRescanClosure(_collector, full_span, // entire span of interest
5287 5289 sp, bm, work_q, rs, cl);
5288 5290
5289 5291 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5290 5292 assert(pst->valid(), "Uninitialized use?");
5291 5293 int nth_task = 0;
5292 5294 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5293 5295 MemRegion span = sp->used_region();
5294 5296 HeapWord* start_addr = span.start();
5295 5297 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5296 5298 alignment);
5297 5299 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5298 5300 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5299 5301 start_addr, "Check alignment");
5300 5302 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5301 5303 chunk_size, "Check alignment");
5302 5304
5303 5305 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5304 5306 // Having claimed the nth_task, compute corresponding mem-region,
5305 5307 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
5306 5308 // The alignment restriction ensures that we do not need any
5307 5309 // synchronization with other gang-workers while setting or
5308 5310 // clearing bits in thus chunk of the MUT.
5309 5311 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5310 5312 start_addr + (nth_task+1)*chunk_size);
5311 5313 // The last chunk's end might be way beyond end of the
5312 5314 // used region. In that case pull back appropriately.
5313 5315 if (this_span.end() > end_addr) {
5314 5316 this_span.set_end(end_addr);
5315 5317 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5316 5318 }
5317 5319 // Iterate over the dirty cards covering this chunk, marking them
5318 5320 // precleaned, and setting the corresponding bits in the mod union
5319 5321 // table. Since we have been careful to partition at Card and MUT-word
5320 5322 // boundaries no synchronization is needed between parallel threads.
5321 5323 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5322 5324 &modUnionClosure);
5323 5325
5324 5326 // Having transferred these marks into the modUnionTable,
5325 5327 // rescan the marked objects on the dirty cards in the modUnionTable.
5326 5328 // Even if this is at a synchronous collection, the initial marking
5327 5329 // may have been done during an asynchronous collection so there
5328 5330 // may be dirty bits in the mod-union table.
5329 5331 _collector->_modUnionTable.dirty_range_iterate_clear(
5330 5332 this_span, &greyRescanClosure);
5331 5333 _collector->_modUnionTable.verifyNoOneBitsInRange(
5332 5334 this_span.start(),
5333 5335 this_span.end());
5334 5336 }
5335 5337 pst->all_tasks_completed(); // declare that i am done
5336 5338 }
5337 5339
5338 5340 // . see if we can share work_queues with ParNew? XXX
5339 5341 void
5340 5342 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5341 5343 int* seed) {
5342 5344 OopTaskQueue* work_q = work_queue(i);
5343 5345 NOT_PRODUCT(int num_steals = 0;)
5344 5346 oop obj_to_scan;
5345 5347 CMSBitMap* bm = &(_collector->_markBitMap);
5346 5348
5347 5349 while (true) {
5348 5350 // Completely finish any left over work from (an) earlier round(s)
5349 5351 cl->trim_queue(0);
5350 5352 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5351 5353 (size_t)ParGCDesiredObjsFromOverflowList);
5352 5354 // Now check if there's any work in the overflow list
5353 5355 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5354 5356 // only affects the number of attempts made to get work from the
5355 5357 // overflow list and does not affect the number of workers. Just
5356 5358 // pass ParallelGCThreads so this behavior is unchanged.
5357 5359 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5358 5360 work_q,
5359 5361 ParallelGCThreads)) {
5360 5362 // found something in global overflow list;
5361 5363 // not yet ready to go stealing work from others.
5362 5364 // We'd like to assert(work_q->size() != 0, ...)
5363 5365 // because we just took work from the overflow list,
5364 5366 // but of course we can't since all of that could have
5365 5367 // been already stolen from us.
5366 5368 // "He giveth and He taketh away."
5367 5369 continue;
5368 5370 }
5369 5371 // Verify that we have no work before we resort to stealing
5370 5372 assert(work_q->size() == 0, "Have work, shouldn't steal");
5371 5373 // Try to steal from other queues that have work
5372 5374 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5373 5375 NOT_PRODUCT(num_steals++;)
5374 5376 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5375 5377 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5376 5378 // Do scanning work
5377 5379 obj_to_scan->oop_iterate(cl);
5378 5380 // Loop around, finish this work, and try to steal some more
5379 5381 } else if (terminator()->offer_termination()) {
5380 5382 break; // nirvana from the infinite cycle
5381 5383 }
5382 5384 }
5383 5385 NOT_PRODUCT(
5384 5386 if (PrintCMSStatistics != 0) {
5385 5387 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5386 5388 }
5387 5389 )
5388 5390 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5389 5391 "Else our work is not yet done");
5390 5392 }
5391 5393
5392 5394 // Return a thread-local PLAB recording array, as appropriate.
5393 5395 void* CMSCollector::get_data_recorder(int thr_num) {
5394 5396 if (_survivor_plab_array != NULL &&
5395 5397 (CMSPLABRecordAlways ||
5396 5398 (_collectorState > Marking && _collectorState < FinalMarking))) {
5397 5399 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5398 5400 ChunkArray* ca = &_survivor_plab_array[thr_num];
5399 5401 ca->reset(); // clear it so that fresh data is recorded
5400 5402 return (void*) ca;
5401 5403 } else {
5402 5404 return NULL;
5403 5405 }
5404 5406 }
5405 5407
5406 5408 // Reset all the thread-local PLAB recording arrays
5407 5409 void CMSCollector::reset_survivor_plab_arrays() {
5408 5410 for (uint i = 0; i < ParallelGCThreads; i++) {
5409 5411 _survivor_plab_array[i].reset();
5410 5412 }
5411 5413 }
5412 5414
5413 5415 // Merge the per-thread plab arrays into the global survivor chunk
5414 5416 // array which will provide the partitioning of the survivor space
5415 5417 // for CMS rescan.
5416 5418 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
5417 5419 int no_of_gc_threads) {
5418 5420 assert(_survivor_plab_array != NULL, "Error");
5419 5421 assert(_survivor_chunk_array != NULL, "Error");
5420 5422 assert(_collectorState == FinalMarking, "Error");
5421 5423 for (int j = 0; j < no_of_gc_threads; j++) {
5422 5424 _cursor[j] = 0;
5423 5425 }
5424 5426 HeapWord* top = surv->top();
5425 5427 size_t i;
5426 5428 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
5427 5429 HeapWord* min_val = top; // Higher than any PLAB address
5428 5430 uint min_tid = 0; // position of min_val this round
5429 5431 for (int j = 0; j < no_of_gc_threads; j++) {
5430 5432 ChunkArray* cur_sca = &_survivor_plab_array[j];
5431 5433 if (_cursor[j] == cur_sca->end()) {
5432 5434 continue;
5433 5435 }
5434 5436 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5435 5437 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5436 5438 assert(surv->used_region().contains(cur_val), "Out of bounds value");
5437 5439 if (cur_val < min_val) {
5438 5440 min_tid = j;
5439 5441 min_val = cur_val;
5440 5442 } else {
5441 5443 assert(cur_val < top, "All recorded addresses should be less");
5442 5444 }
5443 5445 }
5444 5446 // At this point min_val and min_tid are respectively
5445 5447 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5446 5448 // and the thread (j) that witnesses that address.
5447 5449 // We record this address in the _survivor_chunk_array[i]
5448 5450 // and increment _cursor[min_tid] prior to the next round i.
5449 5451 if (min_val == top) {
5450 5452 break;
5451 5453 }
5452 5454 _survivor_chunk_array[i] = min_val;
5453 5455 _cursor[min_tid]++;
5454 5456 }
5455 5457 // We are all done; record the size of the _survivor_chunk_array
5456 5458 _survivor_chunk_index = i; // exclusive: [0, i)
5457 5459 if (PrintCMSStatistics > 0) {
5458 5460 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5459 5461 }
5460 5462 // Verify that we used up all the recorded entries
5461 5463 #ifdef ASSERT
5462 5464 size_t total = 0;
5463 5465 for (int j = 0; j < no_of_gc_threads; j++) {
5464 5466 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5465 5467 total += _cursor[j];
5466 5468 }
5467 5469 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5468 5470 // Check that the merged array is in sorted order
5469 5471 if (total > 0) {
5470 5472 for (size_t i = 0; i < total - 1; i++) {
5471 5473 if (PrintCMSStatistics > 0) {
5472 5474 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5473 5475 i, _survivor_chunk_array[i]);
5474 5476 }
5475 5477 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5476 5478 "Not sorted");
5477 5479 }
5478 5480 }
5479 5481 #endif // ASSERT
5480 5482 }
5481 5483
5482 5484 // Set up the space's par_seq_tasks structure for work claiming
5483 5485 // for parallel rescan of young gen.
5484 5486 // See ParRescanTask where this is currently used.
5485 5487 void
5486 5488 CMSCollector::
5487 5489 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5488 5490 assert(n_threads > 0, "Unexpected n_threads argument");
5489 5491 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5490 5492
5491 5493 // Eden space
5492 5494 {
5493 5495 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5494 5496 assert(!pst->valid(), "Clobbering existing data?");
5495 5497 // Each valid entry in [0, _eden_chunk_index) represents a task.
5496 5498 size_t n_tasks = _eden_chunk_index + 1;
5497 5499 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5498 5500 // Sets the condition for completion of the subtask (how many threads
5499 5501 // need to finish in order to be done).
5500 5502 pst->set_n_threads(n_threads);
5501 5503 pst->set_n_tasks((int)n_tasks);
5502 5504 }
5503 5505
5504 5506 // Merge the survivor plab arrays into _survivor_chunk_array
5505 5507 if (_survivor_plab_array != NULL) {
5506 5508 merge_survivor_plab_arrays(dng->from(), n_threads);
5507 5509 } else {
5508 5510 assert(_survivor_chunk_index == 0, "Error");
5509 5511 }
5510 5512
5511 5513 // To space
5512 5514 {
5513 5515 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5514 5516 assert(!pst->valid(), "Clobbering existing data?");
5515 5517 // Sets the condition for completion of the subtask (how many threads
5516 5518 // need to finish in order to be done).
5517 5519 pst->set_n_threads(n_threads);
5518 5520 pst->set_n_tasks(1);
5519 5521 assert(pst->valid(), "Error");
5520 5522 }
5521 5523
5522 5524 // From space
5523 5525 {
5524 5526 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5525 5527 assert(!pst->valid(), "Clobbering existing data?");
5526 5528 size_t n_tasks = _survivor_chunk_index + 1;
5527 5529 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5528 5530 // Sets the condition for completion of the subtask (how many threads
5529 5531 // need to finish in order to be done).
5530 5532 pst->set_n_threads(n_threads);
5531 5533 pst->set_n_tasks((int)n_tasks);
5532 5534 assert(pst->valid(), "Error");
5533 5535 }
5534 5536 }
5535 5537
5536 5538 // Parallel version of remark
5537 5539 void CMSCollector::do_remark_parallel() {
5538 5540 GenCollectedHeap* gch = GenCollectedHeap::heap();
5539 5541 FlexibleWorkGang* workers = gch->workers();
5540 5542 assert(workers != NULL, "Need parallel worker threads.");
5541 5543 int n_workers = workers->total_workers();
5542 5544 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5543 5545 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
5544 5546
5545 5547 CMSParRemarkTask tsk(this,
5546 5548 cms_space, perm_space,
5547 5549 n_workers, workers, task_queues());
5548 5550
5549 5551 // Set up for parallel process_strong_roots work.
5550 5552 gch->set_par_threads(n_workers);
5551 5553 // We won't be iterating over the cards in the card table updating
5552 5554 // the younger_gen cards, so we shouldn't call the following else
5553 5555 // the verification code as well as subsequent younger_refs_iterate
5554 5556 // code would get confused. XXX
5555 5557 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5556 5558
5557 5559 // The young gen rescan work will not be done as part of
5558 5560 // process_strong_roots (which currently doesn't knw how to
5559 5561 // parallelize such a scan), but rather will be broken up into
5560 5562 // a set of parallel tasks (via the sampling that the [abortable]
5561 5563 // preclean phase did of EdenSpace, plus the [two] tasks of
5562 5564 // scanning the [two] survivor spaces. Further fine-grain
5563 5565 // parallelization of the scanning of the survivor spaces
5564 5566 // themselves, and of precleaning of the younger gen itself
5565 5567 // is deferred to the future.
5566 5568 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5567 5569
5568 5570 // The dirty card rescan work is broken up into a "sequence"
5569 5571 // of parallel tasks (per constituent space) that are dynamically
5570 5572 // claimed by the parallel threads.
5571 5573 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5572 5574 perm_space->initialize_sequential_subtasks_for_rescan(n_workers);
5573 5575
5574 5576 // It turns out that even when we're using 1 thread, doing the work in a
5575 5577 // separate thread causes wide variance in run times. We can't help this
5576 5578 // in the multi-threaded case, but we special-case n=1 here to get
5577 5579 // repeatable measurements of the 1-thread overhead of the parallel code.
5578 5580 if (n_workers > 1) {
5579 5581 // Make refs discovery MT-safe
5580 5582 ReferenceProcessorMTMutator mt(ref_processor(), true);
5581 5583 GenCollectedHeap::StrongRootsScope srs(gch);
5582 5584 workers->run_task(&tsk);
5583 5585 } else {
5584 5586 GenCollectedHeap::StrongRootsScope srs(gch);
5585 5587 tsk.work(0);
5586 5588 }
5587 5589 gch->set_par_threads(0); // 0 ==> non-parallel.
5588 5590 // restore, single-threaded for now, any preserved marks
5589 5591 // as a result of work_q overflow
5590 5592 restore_preserved_marks_if_any();
5591 5593 }
5592 5594
5593 5595 // Non-parallel version of remark
5594 5596 void CMSCollector::do_remark_non_parallel() {
5595 5597 ResourceMark rm;
5596 5598 HandleMark hm;
5597 5599 GenCollectedHeap* gch = GenCollectedHeap::heap();
5598 5600 MarkRefsIntoAndScanClosure
5599 5601 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
5600 5602 &_markStack, &_revisitStack, this,
5601 5603 false /* should_yield */, false /* not precleaning */);
5602 5604 MarkFromDirtyCardsClosure
5603 5605 markFromDirtyCardsClosure(this, _span,
5604 5606 NULL, // space is set further below
5605 5607 &_markBitMap, &_markStack, &_revisitStack,
5606 5608 &mrias_cl);
5607 5609 {
5608 5610 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty);
5609 5611 // Iterate over the dirty cards, setting the corresponding bits in the
5610 5612 // mod union table.
5611 5613 {
5612 5614 ModUnionClosure modUnionClosure(&_modUnionTable);
5613 5615 _ct->ct_bs()->dirty_card_iterate(
5614 5616 _cmsGen->used_region(),
5615 5617 &modUnionClosure);
5616 5618 _ct->ct_bs()->dirty_card_iterate(
5617 5619 _permGen->used_region(),
5618 5620 &modUnionClosure);
5619 5621 }
5620 5622 // Having transferred these marks into the modUnionTable, we just need
5621 5623 // to rescan the marked objects on the dirty cards in the modUnionTable.
5622 5624 // The initial marking may have been done during an asynchronous
5623 5625 // collection so there may be dirty bits in the mod-union table.
5624 5626 const int alignment =
5625 5627 CardTableModRefBS::card_size * BitsPerWord;
5626 5628 {
5627 5629 // ... First handle dirty cards in CMS gen
5628 5630 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5629 5631 MemRegion ur = _cmsGen->used_region();
5630 5632 HeapWord* lb = ur.start();
5631 5633 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5632 5634 MemRegion cms_span(lb, ub);
5633 5635 _modUnionTable.dirty_range_iterate_clear(cms_span,
5634 5636 &markFromDirtyCardsClosure);
5635 5637 verify_work_stacks_empty();
5636 5638 if (PrintCMSStatistics != 0) {
5637 5639 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5638 5640 markFromDirtyCardsClosure.num_dirty_cards());
5639 5641 }
5640 5642 }
5641 5643 {
5642 5644 // .. and then repeat for dirty cards in perm gen
5643 5645 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace());
5644 5646 MemRegion ur = _permGen->used_region();
5645 5647 HeapWord* lb = ur.start();
5646 5648 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5647 5649 MemRegion perm_span(lb, ub);
5648 5650 _modUnionTable.dirty_range_iterate_clear(perm_span,
5649 5651 &markFromDirtyCardsClosure);
5650 5652 verify_work_stacks_empty();
5651 5653 if (PrintCMSStatistics != 0) {
5652 5654 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ",
5653 5655 markFromDirtyCardsClosure.num_dirty_cards());
5654 5656 }
5655 5657 }
5656 5658 }
5657 5659 if (VerifyDuringGC &&
5658 5660 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5659 5661 HandleMark hm; // Discard invalid handles created during verification
5660 5662 Universe::verify(true);
5661 5663 }
5662 5664 {
5663 5665 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty);
5664 5666
5665 5667 verify_work_stacks_empty();
5666 5668
5667 5669 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5668 5670 GenCollectedHeap::StrongRootsScope srs(gch);
5669 5671 gch->gen_process_strong_roots(_cmsGen->level(),
5670 5672 true, // younger gens as roots
5671 5673 false, // use the local StrongRootsScope
5672 5674 true, // collecting perm gen
5673 5675 SharedHeap::ScanningOption(roots_scanning_options()),
5674 5676 &mrias_cl,
5675 5677 true, // walk code active on stacks
5676 5678 NULL);
5677 5679 assert(should_unload_classes()
5678 5680 || (roots_scanning_options() & SharedHeap::SO_CodeCache),
5679 5681 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5680 5682 }
5681 5683 verify_work_stacks_empty();
5682 5684 // Restore evacuated mark words, if any, used for overflow list links
5683 5685 if (!CMSOverflowEarlyRestoration) {
5684 5686 restore_preserved_marks_if_any();
5685 5687 }
5686 5688 verify_overflow_empty();
5687 5689 }
5688 5690
5689 5691 ////////////////////////////////////////////////////////
5690 5692 // Parallel Reference Processing Task Proxy Class
5691 5693 ////////////////////////////////////////////////////////
5692 5694 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5693 5695 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5694 5696 CMSCollector* _collector;
5695 5697 CMSBitMap* _mark_bit_map;
|
↓ open down ↓ |
5384 lines elided |
↑ open up ↑ |
5696 5698 const MemRegion _span;
5697 5699 ProcessTask& _task;
5698 5700
5699 5701 public:
5700 5702 CMSRefProcTaskProxy(ProcessTask& task,
5701 5703 CMSCollector* collector,
5702 5704 const MemRegion& span,
5703 5705 CMSBitMap* mark_bit_map,
5704 5706 AbstractWorkGang* workers,
5705 5707 OopTaskQueueSet* task_queues):
5708 + // XXX Should superclass AGTWOQ also know about AWG since it knows
5709 + // about the task_queues used by the AWG? Then it could initialize
5710 + // the terminator() object. See 6984287. The set_for_termination()
5711 + // below is a temporary band-aid for the regression in 6984287.
5706 5712 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5707 5713 task_queues),
5708 5714 _task(task),
5709 5715 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5710 - {
5711 - assert(_collector->_span.equals(_span) && !_span.is_empty(),
5712 - "Inconsistency in _span");
5713 - }
5716 + {
5717 + assert(_collector->_span.equals(_span) && !_span.is_empty(),
5718 + "Inconsistency in _span");
5719 + set_for_termination(workers->active_workers());
5720 + }
5714 5721
5715 5722 OopTaskQueueSet* task_queues() { return queues(); }
5716 5723
5717 5724 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5718 5725
5719 5726 void do_work_steal(int i,
5720 5727 CMSParDrainMarkingStackClosure* drain,
5721 5728 CMSParKeepAliveClosure* keep_alive,
5722 5729 int* seed);
5723 5730
5724 5731 virtual void work(int i);
5725 5732 };
5726 5733
5727 5734 void CMSRefProcTaskProxy::work(int i) {
5728 5735 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5729 5736 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5730 5737 _mark_bit_map,
5731 5738 &_collector->_revisitStack,
5732 5739 work_queue(i));
5733 5740 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5734 5741 _mark_bit_map,
5735 5742 &_collector->_revisitStack,
5736 5743 work_queue(i));
5737 5744 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5738 5745 _task.work(i, is_alive_closure, par_keep_alive, par_drain_stack);
5739 5746 if (_task.marks_oops_alive()) {
5740 5747 do_work_steal(i, &par_drain_stack, &par_keep_alive,
5741 5748 _collector->hash_seed(i));
5742 5749 }
5743 5750 assert(work_queue(i)->size() == 0, "work_queue should be empty");
5744 5751 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5745 5752 }
5746 5753
5747 5754 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5748 5755 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5749 5756 EnqueueTask& _task;
5750 5757
5751 5758 public:
5752 5759 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5753 5760 : AbstractGangTask("Enqueue reference objects in parallel"),
5754 5761 _task(task)
5755 5762 { }
5756 5763
5757 5764 virtual void work(int i)
5758 5765 {
5759 5766 _task.work(i);
5760 5767 }
5761 5768 };
5762 5769
5763 5770 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5764 5771 MemRegion span, CMSBitMap* bit_map, CMSMarkStack* revisit_stack,
5765 5772 OopTaskQueue* work_queue):
5766 5773 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
5767 5774 _span(span),
5768 5775 _bit_map(bit_map),
5769 5776 _work_queue(work_queue),
5770 5777 _mark_and_push(collector, span, bit_map, revisit_stack, work_queue),
5771 5778 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5772 5779 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5773 5780 { }
5774 5781
5775 5782 // . see if we can share work_queues with ParNew? XXX
5776 5783 void CMSRefProcTaskProxy::do_work_steal(int i,
5777 5784 CMSParDrainMarkingStackClosure* drain,
5778 5785 CMSParKeepAliveClosure* keep_alive,
5779 5786 int* seed) {
5780 5787 OopTaskQueue* work_q = work_queue(i);
5781 5788 NOT_PRODUCT(int num_steals = 0;)
5782 5789 oop obj_to_scan;
5783 5790
5784 5791 while (true) {
5785 5792 // Completely finish any left over work from (an) earlier round(s)
5786 5793 drain->trim_queue(0);
5787 5794 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5788 5795 (size_t)ParGCDesiredObjsFromOverflowList);
5789 5796 // Now check if there's any work in the overflow list
5790 5797 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5791 5798 // only affects the number of attempts made to get work from the
5792 5799 // overflow list and does not affect the number of workers. Just
5793 5800 // pass ParallelGCThreads so this behavior is unchanged.
5794 5801 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5795 5802 work_q,
5796 5803 ParallelGCThreads)) {
5797 5804 // Found something in global overflow list;
5798 5805 // not yet ready to go stealing work from others.
5799 5806 // We'd like to assert(work_q->size() != 0, ...)
5800 5807 // because we just took work from the overflow list,
5801 5808 // but of course we can't, since all of that might have
5802 5809 // been already stolen from us.
5803 5810 continue;
5804 5811 }
5805 5812 // Verify that we have no work before we resort to stealing
5806 5813 assert(work_q->size() == 0, "Have work, shouldn't steal");
5807 5814 // Try to steal from other queues that have work
5808 5815 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5809 5816 NOT_PRODUCT(num_steals++;)
5810 5817 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5811 5818 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5812 5819 // Do scanning work
5813 5820 obj_to_scan->oop_iterate(keep_alive);
5814 5821 // Loop around, finish this work, and try to steal some more
5815 5822 } else if (terminator()->offer_termination()) {
5816 5823 break; // nirvana from the infinite cycle
5817 5824 }
5818 5825 }
5819 5826 NOT_PRODUCT(
5820 5827 if (PrintCMSStatistics != 0) {
5821 5828 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5822 5829 }
5823 5830 )
5824 5831 }
5825 5832
5826 5833 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5827 5834 {
5828 5835 GenCollectedHeap* gch = GenCollectedHeap::heap();
5829 5836 FlexibleWorkGang* workers = gch->workers();
5830 5837 assert(workers != NULL, "Need parallel worker threads.");
5831 5838 CMSRefProcTaskProxy rp_task(task, &_collector,
5832 5839 _collector.ref_processor()->span(),
5833 5840 _collector.markBitMap(),
5834 5841 workers, _collector.task_queues());
5835 5842 workers->run_task(&rp_task);
5836 5843 }
5837 5844
5838 5845 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5839 5846 {
5840 5847
5841 5848 GenCollectedHeap* gch = GenCollectedHeap::heap();
5842 5849 FlexibleWorkGang* workers = gch->workers();
5843 5850 assert(workers != NULL, "Need parallel worker threads.");
5844 5851 CMSRefEnqueueTaskProxy enq_task(task);
5845 5852 workers->run_task(&enq_task);
5846 5853 }
5847 5854
5848 5855 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
5849 5856
5850 5857 ResourceMark rm;
5851 5858 HandleMark hm;
5852 5859
5853 5860 ReferenceProcessor* rp = ref_processor();
5854 5861 assert(rp->span().equals(_span), "Spans should be equal");
5855 5862 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5856 5863 // Process weak references.
5857 5864 rp->setup_policy(clear_all_soft_refs);
5858 5865 verify_work_stacks_empty();
5859 5866
5860 5867 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5861 5868 &_markStack, &_revisitStack,
5862 5869 false /* !preclean */);
5863 5870 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5864 5871 _span, &_markBitMap, &_markStack,
5865 5872 &cmsKeepAliveClosure, false /* !preclean */);
5866 5873 {
5867 5874 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty);
5868 5875 if (rp->processing_is_mt()) {
5869 5876 // Set the degree of MT here. If the discovery is done MT, there
5870 5877 // may have been a different number of threads doing the discovery
5871 5878 // and a different number of discovered lists may have Ref objects.
5872 5879 // That is OK as long as the Reference lists are balanced (see
5873 5880 // balance_all_queues() and balance_queues()).
5874 5881
5875 5882
5876 5883 rp->set_mt_degree(ParallelGCThreads);
5877 5884 CMSRefProcTaskExecutor task_executor(*this);
5878 5885 rp->process_discovered_references(&_is_alive_closure,
5879 5886 &cmsKeepAliveClosure,
5880 5887 &cmsDrainMarkingStackClosure,
5881 5888 &task_executor);
5882 5889 } else {
5883 5890 rp->process_discovered_references(&_is_alive_closure,
5884 5891 &cmsKeepAliveClosure,
5885 5892 &cmsDrainMarkingStackClosure,
5886 5893 NULL);
5887 5894 }
5888 5895 verify_work_stacks_empty();
5889 5896 }
5890 5897
5891 5898 if (should_unload_classes()) {
5892 5899 {
5893 5900 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty);
5894 5901
5895 5902 // Follow SystemDictionary roots and unload classes
5896 5903 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5897 5904
5898 5905 // Follow CodeCache roots and unload any methods marked for unloading
5899 5906 CodeCache::do_unloading(&_is_alive_closure,
5900 5907 &cmsKeepAliveClosure,
5901 5908 purged_class);
5902 5909
5903 5910 cmsDrainMarkingStackClosure.do_void();
5904 5911 verify_work_stacks_empty();
5905 5912
5906 5913 // Update subklass/sibling/implementor links in KlassKlass descendants
5907 5914 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty");
5908 5915 oop k;
5909 5916 while ((k = _revisitStack.pop()) != NULL) {
5910 5917 ((Klass*)(oopDesc*)k)->follow_weak_klass_links(
5911 5918 &_is_alive_closure,
5912 5919 &cmsKeepAliveClosure);
5913 5920 }
5914 5921 assert(!ClassUnloading ||
5915 5922 (_markStack.isEmpty() && overflow_list_is_empty()),
5916 5923 "Should not have found new reachable objects");
5917 5924 assert(_revisitStack.isEmpty(), "revisit stack should have been drained");
5918 5925 cmsDrainMarkingStackClosure.do_void();
5919 5926 verify_work_stacks_empty();
5920 5927 }
5921 5928
5922 5929 {
5923 5930 TraceTime t("scrub symbol & string tables", PrintGCDetails, false, gclog_or_tty);
5924 5931 // Now clean up stale oops in StringTable
5925 5932 StringTable::unlink(&_is_alive_closure);
5926 5933 // Clean up unreferenced symbols in symbol table.
5927 5934 SymbolTable::unlink();
5928 5935 }
5929 5936 }
5930 5937
5931 5938 verify_work_stacks_empty();
5932 5939 // Restore any preserved marks as a result of mark stack or
5933 5940 // work queue overflow
5934 5941 restore_preserved_marks_if_any(); // done single-threaded for now
5935 5942
5936 5943 rp->set_enqueuing_is_done(true);
5937 5944 if (rp->processing_is_mt()) {
5938 5945 rp->balance_all_queues();
5939 5946 CMSRefProcTaskExecutor task_executor(*this);
5940 5947 rp->enqueue_discovered_references(&task_executor);
5941 5948 } else {
5942 5949 rp->enqueue_discovered_references(NULL);
5943 5950 }
5944 5951 rp->verify_no_references_recorded();
5945 5952 assert(!rp->discovery_enabled(), "should have been disabled");
5946 5953 }
5947 5954
5948 5955 #ifndef PRODUCT
5949 5956 void CMSCollector::check_correct_thread_executing() {
5950 5957 Thread* t = Thread::current();
5951 5958 // Only the VM thread or the CMS thread should be here.
5952 5959 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5953 5960 "Unexpected thread type");
5954 5961 // If this is the vm thread, the foreground process
5955 5962 // should not be waiting. Note that _foregroundGCIsActive is
5956 5963 // true while the foreground collector is waiting.
5957 5964 if (_foregroundGCShouldWait) {
5958 5965 // We cannot be the VM thread
5959 5966 assert(t->is_ConcurrentGC_thread(),
5960 5967 "Should be CMS thread");
5961 5968 } else {
5962 5969 // We can be the CMS thread only if we are in a stop-world
5963 5970 // phase of CMS collection.
5964 5971 if (t->is_ConcurrentGC_thread()) {
5965 5972 assert(_collectorState == InitialMarking ||
5966 5973 _collectorState == FinalMarking,
5967 5974 "Should be a stop-world phase");
5968 5975 // The CMS thread should be holding the CMS_token.
5969 5976 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5970 5977 "Potential interference with concurrently "
5971 5978 "executing VM thread");
5972 5979 }
5973 5980 }
5974 5981 }
5975 5982 #endif
5976 5983
5977 5984 void CMSCollector::sweep(bool asynch) {
5978 5985 assert(_collectorState == Sweeping, "just checking");
5979 5986 check_correct_thread_executing();
5980 5987 verify_work_stacks_empty();
5981 5988 verify_overflow_empty();
5982 5989 increment_sweep_count();
5983 5990 TraceCMSMemoryManagerStats tms(_collectorState);
5984 5991
5985 5992 _inter_sweep_timer.stop();
5986 5993 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5987 5994 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
5988 5995
5989 5996 // PermGen verification support: If perm gen sweeping is disabled in
5990 5997 // this cycle, we preserve the perm gen object "deadness" information
5991 5998 // in the perm_gen_verify_bit_map. In order to do that we traverse
5992 5999 // all blocks in perm gen and mark all dead objects.
5993 6000 if (verifying() && !should_unload_classes()) {
5994 6001 assert(perm_gen_verify_bit_map()->sizeInBits() != 0,
5995 6002 "Should have already been allocated");
5996 6003 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(),
5997 6004 markBitMap(), perm_gen_verify_bit_map());
5998 6005 if (asynch) {
5999 6006 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
6000 6007 bitMapLock());
6001 6008 _permGen->cmsSpace()->blk_iterate(&mdo);
6002 6009 } else {
6003 6010 // In the case of synchronous sweep, we already have
6004 6011 // the requisite locks/tokens.
6005 6012 _permGen->cmsSpace()->blk_iterate(&mdo);
6006 6013 }
6007 6014 }
6008 6015
6009 6016 assert(!_intra_sweep_timer.is_active(), "Should not be active");
6010 6017 _intra_sweep_timer.reset();
6011 6018 _intra_sweep_timer.start();
6012 6019 if (asynch) {
6013 6020 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6014 6021 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
6015 6022 // First sweep the old gen then the perm gen
6016 6023 {
6017 6024 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6018 6025 bitMapLock());
6019 6026 sweepWork(_cmsGen, asynch);
6020 6027 }
6021 6028
6022 6029 // Now repeat for perm gen
6023 6030 if (should_unload_classes()) {
6024 6031 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
6025 6032 bitMapLock());
6026 6033 sweepWork(_permGen, asynch);
6027 6034 }
6028 6035
6029 6036 // Update Universe::_heap_*_at_gc figures.
6030 6037 // We need all the free list locks to make the abstract state
6031 6038 // transition from Sweeping to Resetting. See detailed note
6032 6039 // further below.
6033 6040 {
6034 6041 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6035 6042 _permGen->freelistLock());
6036 6043 // Update heap occupancy information which is used as
6037 6044 // input to soft ref clearing policy at the next gc.
6038 6045 Universe::update_heap_info_at_gc();
6039 6046 _collectorState = Resizing;
6040 6047 }
6041 6048 } else {
6042 6049 // already have needed locks
6043 6050 sweepWork(_cmsGen, asynch);
6044 6051
6045 6052 if (should_unload_classes()) {
6046 6053 sweepWork(_permGen, asynch);
6047 6054 }
6048 6055 // Update heap occupancy information which is used as
6049 6056 // input to soft ref clearing policy at the next gc.
6050 6057 Universe::update_heap_info_at_gc();
6051 6058 _collectorState = Resizing;
6052 6059 }
6053 6060 verify_work_stacks_empty();
6054 6061 verify_overflow_empty();
6055 6062
6056 6063 _intra_sweep_timer.stop();
6057 6064 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
6058 6065
6059 6066 _inter_sweep_timer.reset();
6060 6067 _inter_sweep_timer.start();
6061 6068
6062 6069 update_time_of_last_gc(os::javaTimeMillis());
6063 6070
6064 6071 // NOTE on abstract state transitions:
6065 6072 // Mutators allocate-live and/or mark the mod-union table dirty
6066 6073 // based on the state of the collection. The former is done in
6067 6074 // the interval [Marking, Sweeping] and the latter in the interval
6068 6075 // [Marking, Sweeping). Thus the transitions into the Marking state
6069 6076 // and out of the Sweeping state must be synchronously visible
6070 6077 // globally to the mutators.
6071 6078 // The transition into the Marking state happens with the world
6072 6079 // stopped so the mutators will globally see it. Sweeping is
6073 6080 // done asynchronously by the background collector so the transition
6074 6081 // from the Sweeping state to the Resizing state must be done
6075 6082 // under the freelistLock (as is the check for whether to
6076 6083 // allocate-live and whether to dirty the mod-union table).
6077 6084 assert(_collectorState == Resizing, "Change of collector state to"
6078 6085 " Resizing must be done under the freelistLocks (plural)");
6079 6086
6080 6087 // Now that sweeping has been completed, we clear
6081 6088 // the incremental_collection_failed flag,
6082 6089 // thus inviting a younger gen collection to promote into
6083 6090 // this generation. If such a promotion may still fail,
6084 6091 // the flag will be set again when a young collection is
6085 6092 // attempted.
6086 6093 GenCollectedHeap* gch = GenCollectedHeap::heap();
6087 6094 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
6088 6095 gch->update_full_collections_completed(_collection_count_start);
6089 6096 }
6090 6097
6091 6098 // FIX ME!!! Looks like this belongs in CFLSpace, with
6092 6099 // CMSGen merely delegating to it.
6093 6100 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
6094 6101 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
6095 6102 HeapWord* minAddr = _cmsSpace->bottom();
6096 6103 HeapWord* largestAddr =
6097 6104 (HeapWord*) _cmsSpace->dictionary()->findLargestDict();
6098 6105 if (largestAddr == NULL) {
6099 6106 // The dictionary appears to be empty. In this case
6100 6107 // try to coalesce at the end of the heap.
6101 6108 largestAddr = _cmsSpace->end();
6102 6109 }
6103 6110 size_t largestOffset = pointer_delta(largestAddr, minAddr);
6104 6111 size_t nearLargestOffset =
6105 6112 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
6106 6113 if (PrintFLSStatistics != 0) {
6107 6114 gclog_or_tty->print_cr(
6108 6115 "CMS: Large Block: " PTR_FORMAT ";"
6109 6116 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
6110 6117 largestAddr,
6111 6118 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset);
6112 6119 }
6113 6120 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
6114 6121 }
6115 6122
6116 6123 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
6117 6124 return addr >= _cmsSpace->nearLargestChunk();
6118 6125 }
6119 6126
6120 6127 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
6121 6128 return _cmsSpace->find_chunk_at_end();
6122 6129 }
6123 6130
6124 6131 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
6125 6132 bool full) {
6126 6133 // The next lower level has been collected. Gather any statistics
6127 6134 // that are of interest at this point.
6128 6135 if (!full && (current_level + 1) == level()) {
6129 6136 // Gather statistics on the young generation collection.
6130 6137 collector()->stats().record_gc0_end(used());
6131 6138 }
6132 6139 }
6133 6140
6134 6141 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
6135 6142 GenCollectedHeap* gch = GenCollectedHeap::heap();
6136 6143 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
6137 6144 "Wrong type of heap");
6138 6145 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
6139 6146 gch->gen_policy()->size_policy();
6140 6147 assert(sp->is_gc_cms_adaptive_size_policy(),
6141 6148 "Wrong type of size policy");
6142 6149 return sp;
6143 6150 }
6144 6151
6145 6152 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
6146 6153 if (PrintGCDetails && Verbose) {
6147 6154 gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
6148 6155 }
6149 6156 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
6150 6157 _debug_collection_type =
6151 6158 (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
6152 6159 if (PrintGCDetails && Verbose) {
6153 6160 gclog_or_tty->print_cr("to %d ", _debug_collection_type);
6154 6161 }
6155 6162 }
6156 6163
6157 6164 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
6158 6165 bool asynch) {
6159 6166 // We iterate over the space(s) underlying this generation,
6160 6167 // checking the mark bit map to see if the bits corresponding
6161 6168 // to specific blocks are marked or not. Blocks that are
6162 6169 // marked are live and are not swept up. All remaining blocks
6163 6170 // are swept up, with coalescing on-the-fly as we sweep up
6164 6171 // contiguous free and/or garbage blocks:
6165 6172 // We need to ensure that the sweeper synchronizes with allocators
6166 6173 // and stop-the-world collectors. In particular, the following
6167 6174 // locks are used:
6168 6175 // . CMS token: if this is held, a stop the world collection cannot occur
6169 6176 // . freelistLock: if this is held no allocation can occur from this
6170 6177 // generation by another thread
6171 6178 // . bitMapLock: if this is held, no other thread can access or update
6172 6179 //
6173 6180
6174 6181 // Note that we need to hold the freelistLock if we use
6175 6182 // block iterate below; else the iterator might go awry if
6176 6183 // a mutator (or promotion) causes block contents to change
6177 6184 // (for instance if the allocator divvies up a block).
6178 6185 // If we hold the free list lock, for all practical purposes
6179 6186 // young generation GC's can't occur (they'll usually need to
6180 6187 // promote), so we might as well prevent all young generation
6181 6188 // GC's while we do a sweeping step. For the same reason, we might
6182 6189 // as well take the bit map lock for the entire duration
6183 6190
6184 6191 // check that we hold the requisite locks
6185 6192 assert(have_cms_token(), "Should hold cms token");
6186 6193 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
6187 6194 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
6188 6195 "Should possess CMS token to sweep");
6189 6196 assert_lock_strong(gen->freelistLock());
6190 6197 assert_lock_strong(bitMapLock());
6191 6198
6192 6199 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
6193 6200 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
6194 6201 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
6195 6202 _inter_sweep_estimate.padded_average(),
6196 6203 _intra_sweep_estimate.padded_average());
6197 6204 gen->setNearLargestChunk();
6198 6205
6199 6206 {
6200 6207 SweepClosure sweepClosure(this, gen, &_markBitMap,
6201 6208 CMSYield && asynch);
6202 6209 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
6203 6210 // We need to free-up/coalesce garbage/blocks from a
6204 6211 // co-terminal free run. This is done in the SweepClosure
6205 6212 // destructor; so, do not remove this scope, else the
6206 6213 // end-of-sweep-census below will be off by a little bit.
6207 6214 }
6208 6215 gen->cmsSpace()->sweep_completed();
6209 6216 gen->cmsSpace()->endSweepFLCensus(sweep_count());
6210 6217 if (should_unload_classes()) { // unloaded classes this cycle,
6211 6218 _concurrent_cycles_since_last_unload = 0; // ... reset count
6212 6219 } else { // did not unload classes,
6213 6220 _concurrent_cycles_since_last_unload++; // ... increment count
6214 6221 }
6215 6222 }
6216 6223
6217 6224 // Reset CMS data structures (for now just the marking bit map)
6218 6225 // preparatory for the next cycle.
6219 6226 void CMSCollector::reset(bool asynch) {
6220 6227 GenCollectedHeap* gch = GenCollectedHeap::heap();
6221 6228 CMSAdaptiveSizePolicy* sp = size_policy();
6222 6229 AdaptiveSizePolicyOutput(sp, gch->total_collections());
6223 6230 if (asynch) {
6224 6231 CMSTokenSyncWithLocks ts(true, bitMapLock());
6225 6232
6226 6233 // If the state is not "Resetting", the foreground thread
6227 6234 // has done a collection and the resetting.
6228 6235 if (_collectorState != Resetting) {
6229 6236 assert(_collectorState == Idling, "The state should only change"
6230 6237 " because the foreground collector has finished the collection");
6231 6238 return;
6232 6239 }
6233 6240
6234 6241 // Clear the mark bitmap (no grey objects to start with)
6235 6242 // for the next cycle.
6236 6243 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6237 6244 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
6238 6245
6239 6246 HeapWord* curAddr = _markBitMap.startWord();
6240 6247 while (curAddr < _markBitMap.endWord()) {
6241 6248 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
6242 6249 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
6243 6250 _markBitMap.clear_large_range(chunk);
6244 6251 if (ConcurrentMarkSweepThread::should_yield() &&
6245 6252 !foregroundGCIsActive() &&
6246 6253 CMSYield) {
6247 6254 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6248 6255 "CMS thread should hold CMS token");
6249 6256 assert_lock_strong(bitMapLock());
6250 6257 bitMapLock()->unlock();
6251 6258 ConcurrentMarkSweepThread::desynchronize(true);
6252 6259 ConcurrentMarkSweepThread::acknowledge_yield_request();
6253 6260 stopTimer();
6254 6261 if (PrintCMSStatistics != 0) {
6255 6262 incrementYields();
6256 6263 }
6257 6264 icms_wait();
6258 6265
6259 6266 // See the comment in coordinator_yield()
6260 6267 for (unsigned i = 0; i < CMSYieldSleepCount &&
6261 6268 ConcurrentMarkSweepThread::should_yield() &&
6262 6269 !CMSCollector::foregroundGCIsActive(); ++i) {
6263 6270 os::sleep(Thread::current(), 1, false);
6264 6271 ConcurrentMarkSweepThread::acknowledge_yield_request();
6265 6272 }
6266 6273
6267 6274 ConcurrentMarkSweepThread::synchronize(true);
6268 6275 bitMapLock()->lock_without_safepoint_check();
6269 6276 startTimer();
6270 6277 }
6271 6278 curAddr = chunk.end();
6272 6279 }
6273 6280 // A successful mostly concurrent collection has been done.
6274 6281 // Because only the full (i.e., concurrent mode failure) collections
6275 6282 // are being measured for gc overhead limits, clean the "near" flag
6276 6283 // and count.
6277 6284 sp->reset_gc_overhead_limit_count();
6278 6285 _collectorState = Idling;
6279 6286 } else {
6280 6287 // already have the lock
6281 6288 assert(_collectorState == Resetting, "just checking");
6282 6289 assert_lock_strong(bitMapLock());
6283 6290 _markBitMap.clear_all();
6284 6291 _collectorState = Idling;
6285 6292 }
6286 6293
6287 6294 // Stop incremental mode after a cycle completes, so that any future cycles
6288 6295 // are triggered by allocation.
6289 6296 stop_icms();
6290 6297
6291 6298 NOT_PRODUCT(
6292 6299 if (RotateCMSCollectionTypes) {
6293 6300 _cmsGen->rotate_debug_collection_type();
6294 6301 }
6295 6302 )
6296 6303 }
6297 6304
6298 6305 void CMSCollector::do_CMS_operation(CMS_op_type op) {
6299 6306 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
6300 6307 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6301 6308 TraceTime t("GC", PrintGC, !PrintGCDetails, gclog_or_tty);
6302 6309 TraceCollectorStats tcs(counters());
6303 6310
6304 6311 switch (op) {
6305 6312 case CMS_op_checkpointRootsInitial: {
6306 6313 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6307 6314 checkpointRootsInitial(true); // asynch
6308 6315 if (PrintGC) {
6309 6316 _cmsGen->printOccupancy("initial-mark");
6310 6317 }
6311 6318 break;
6312 6319 }
6313 6320 case CMS_op_checkpointRootsFinal: {
6314 6321 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6315 6322 checkpointRootsFinal(true, // asynch
6316 6323 false, // !clear_all_soft_refs
6317 6324 false); // !init_mark_was_synchronous
6318 6325 if (PrintGC) {
6319 6326 _cmsGen->printOccupancy("remark");
6320 6327 }
6321 6328 break;
6322 6329 }
6323 6330 default:
6324 6331 fatal("No such CMS_op");
6325 6332 }
6326 6333 }
6327 6334
6328 6335 #ifndef PRODUCT
6329 6336 size_t const CMSCollector::skip_header_HeapWords() {
6330 6337 return FreeChunk::header_size();
6331 6338 }
6332 6339
6333 6340 // Try and collect here conditions that should hold when
6334 6341 // CMS thread is exiting. The idea is that the foreground GC
6335 6342 // thread should not be blocked if it wants to terminate
6336 6343 // the CMS thread and yet continue to run the VM for a while
6337 6344 // after that.
6338 6345 void CMSCollector::verify_ok_to_terminate() const {
6339 6346 assert(Thread::current()->is_ConcurrentGC_thread(),
6340 6347 "should be called by CMS thread");
6341 6348 assert(!_foregroundGCShouldWait, "should be false");
6342 6349 // We could check here that all the various low-level locks
6343 6350 // are not held by the CMS thread, but that is overkill; see
6344 6351 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6345 6352 // is checked.
6346 6353 }
6347 6354 #endif
6348 6355
6349 6356 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6350 6357 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6351 6358 "missing Printezis mark?");
6352 6359 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6353 6360 size_t size = pointer_delta(nextOneAddr + 1, addr);
6354 6361 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6355 6362 "alignment problem");
6356 6363 assert(size >= 3, "Necessary for Printezis marks to work");
6357 6364 return size;
6358 6365 }
6359 6366
6360 6367 // A variant of the above (block_size_using_printezis_bits()) except
6361 6368 // that we return 0 if the P-bits are not yet set.
6362 6369 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6363 6370 if (_markBitMap.isMarked(addr + 1)) {
6364 6371 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
6365 6372 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6366 6373 size_t size = pointer_delta(nextOneAddr + 1, addr);
6367 6374 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6368 6375 "alignment problem");
6369 6376 assert(size >= 3, "Necessary for Printezis marks to work");
6370 6377 return size;
6371 6378 }
6372 6379 return 0;
6373 6380 }
6374 6381
6375 6382 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6376 6383 size_t sz = 0;
6377 6384 oop p = (oop)addr;
6378 6385 if (p->klass_or_null() != NULL && p->is_parsable()) {
6379 6386 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6380 6387 } else {
6381 6388 sz = block_size_using_printezis_bits(addr);
6382 6389 }
6383 6390 assert(sz > 0, "size must be nonzero");
6384 6391 HeapWord* next_block = addr + sz;
6385 6392 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
6386 6393 CardTableModRefBS::card_size);
6387 6394 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
6388 6395 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6389 6396 "must be different cards");
6390 6397 return next_card;
6391 6398 }
6392 6399
6393 6400
6394 6401 // CMS Bit Map Wrapper /////////////////////////////////////////
6395 6402
6396 6403 // Construct a CMS bit map infrastructure, but don't create the
6397 6404 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6398 6405 // further below.
6399 6406 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6400 6407 _bm(),
6401 6408 _shifter(shifter),
6402 6409 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6403 6410 {
6404 6411 _bmStartWord = 0;
6405 6412 _bmWordSize = 0;
6406 6413 }
6407 6414
6408 6415 bool CMSBitMap::allocate(MemRegion mr) {
6409 6416 _bmStartWord = mr.start();
6410 6417 _bmWordSize = mr.word_size();
6411 6418 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6412 6419 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6413 6420 if (!brs.is_reserved()) {
6414 6421 warning("CMS bit map allocation failure");
6415 6422 return false;
6416 6423 }
6417 6424 // For now we'll just commit all of the bit map up fromt.
6418 6425 // Later on we'll try to be more parsimonious with swap.
6419 6426 if (!_virtual_space.initialize(brs, brs.size())) {
6420 6427 warning("CMS bit map backing store failure");
6421 6428 return false;
6422 6429 }
6423 6430 assert(_virtual_space.committed_size() == brs.size(),
6424 6431 "didn't reserve backing store for all of CMS bit map?");
6425 6432 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
6426 6433 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6427 6434 _bmWordSize, "inconsistency in bit map sizing");
6428 6435 _bm.set_size(_bmWordSize >> _shifter);
6429 6436
6430 6437 // bm.clear(); // can we rely on getting zero'd memory? verify below
6431 6438 assert(isAllClear(),
6432 6439 "Expected zero'd memory from ReservedSpace constructor");
6433 6440 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6434 6441 "consistency check");
6435 6442 return true;
6436 6443 }
6437 6444
6438 6445 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6439 6446 HeapWord *next_addr, *end_addr, *last_addr;
6440 6447 assert_locked();
6441 6448 assert(covers(mr), "out-of-range error");
6442 6449 // XXX assert that start and end are appropriately aligned
6443 6450 for (next_addr = mr.start(), end_addr = mr.end();
6444 6451 next_addr < end_addr; next_addr = last_addr) {
6445 6452 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6446 6453 last_addr = dirty_region.end();
6447 6454 if (!dirty_region.is_empty()) {
6448 6455 cl->do_MemRegion(dirty_region);
6449 6456 } else {
6450 6457 assert(last_addr == end_addr, "program logic");
6451 6458 return;
6452 6459 }
6453 6460 }
6454 6461 }
6455 6462
6456 6463 #ifndef PRODUCT
6457 6464 void CMSBitMap::assert_locked() const {
6458 6465 CMSLockVerifier::assert_locked(lock());
6459 6466 }
6460 6467
6461 6468 bool CMSBitMap::covers(MemRegion mr) const {
6462 6469 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6463 6470 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6464 6471 "size inconsistency");
6465 6472 return (mr.start() >= _bmStartWord) &&
6466 6473 (mr.end() <= endWord());
6467 6474 }
6468 6475
6469 6476 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6470 6477 return (start >= _bmStartWord && (start + size) <= endWord());
6471 6478 }
6472 6479
6473 6480 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6474 6481 // verify that there are no 1 bits in the interval [left, right)
6475 6482 FalseBitMapClosure falseBitMapClosure;
6476 6483 iterate(&falseBitMapClosure, left, right);
6477 6484 }
6478 6485
6479 6486 void CMSBitMap::region_invariant(MemRegion mr)
6480 6487 {
6481 6488 assert_locked();
6482 6489 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6483 6490 assert(!mr.is_empty(), "unexpected empty region");
6484 6491 assert(covers(mr), "mr should be covered by bit map");
6485 6492 // convert address range into offset range
6486 6493 size_t start_ofs = heapWordToOffset(mr.start());
6487 6494 // Make sure that end() is appropriately aligned
6488 6495 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6489 6496 (1 << (_shifter+LogHeapWordSize))),
6490 6497 "Misaligned mr.end()");
6491 6498 size_t end_ofs = heapWordToOffset(mr.end());
6492 6499 assert(end_ofs > start_ofs, "Should mark at least one bit");
6493 6500 }
6494 6501
6495 6502 #endif
6496 6503
6497 6504 bool CMSMarkStack::allocate(size_t size) {
6498 6505 // allocate a stack of the requisite depth
6499 6506 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6500 6507 size * sizeof(oop)));
6501 6508 if (!rs.is_reserved()) {
6502 6509 warning("CMSMarkStack allocation failure");
6503 6510 return false;
6504 6511 }
6505 6512 if (!_virtual_space.initialize(rs, rs.size())) {
6506 6513 warning("CMSMarkStack backing store failure");
6507 6514 return false;
6508 6515 }
6509 6516 assert(_virtual_space.committed_size() == rs.size(),
6510 6517 "didn't reserve backing store for all of CMS stack?");
6511 6518 _base = (oop*)(_virtual_space.low());
6512 6519 _index = 0;
6513 6520 _capacity = size;
6514 6521 NOT_PRODUCT(_max_depth = 0);
6515 6522 return true;
6516 6523 }
6517 6524
6518 6525 // XXX FIX ME !!! In the MT case we come in here holding a
6519 6526 // leaf lock. For printing we need to take a further lock
6520 6527 // which has lower rank. We need to recallibrate the two
6521 6528 // lock-ranks involved in order to be able to rpint the
6522 6529 // messages below. (Or defer the printing to the caller.
6523 6530 // For now we take the expedient path of just disabling the
6524 6531 // messages for the problematic case.)
6525 6532 void CMSMarkStack::expand() {
6526 6533 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6527 6534 if (_capacity == MarkStackSizeMax) {
6528 6535 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6529 6536 // We print a warning message only once per CMS cycle.
6530 6537 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6531 6538 }
6532 6539 return;
6533 6540 }
6534 6541 // Double capacity if possible
6535 6542 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6536 6543 // Do not give up existing stack until we have managed to
6537 6544 // get the double capacity that we desired.
6538 6545 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6539 6546 new_capacity * sizeof(oop)));
6540 6547 if (rs.is_reserved()) {
6541 6548 // Release the backing store associated with old stack
6542 6549 _virtual_space.release();
6543 6550 // Reinitialize virtual space for new stack
6544 6551 if (!_virtual_space.initialize(rs, rs.size())) {
6545 6552 fatal("Not enough swap for expanded marking stack");
6546 6553 }
6547 6554 _base = (oop*)(_virtual_space.low());
6548 6555 _index = 0;
6549 6556 _capacity = new_capacity;
6550 6557 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6551 6558 // Failed to double capacity, continue;
6552 6559 // we print a detail message only once per CMS cycle.
6553 6560 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6554 6561 SIZE_FORMAT"K",
6555 6562 _capacity / K, new_capacity / K);
6556 6563 }
6557 6564 }
6558 6565
6559 6566
6560 6567 // Closures
6561 6568 // XXX: there seems to be a lot of code duplication here;
6562 6569 // should refactor and consolidate common code.
6563 6570
6564 6571 // This closure is used to mark refs into the CMS generation in
6565 6572 // the CMS bit map. Called at the first checkpoint. This closure
6566 6573 // assumes that we do not need to re-mark dirty cards; if the CMS
6567 6574 // generation on which this is used is not an oldest (modulo perm gen)
6568 6575 // generation then this will lose younger_gen cards!
6569 6576
6570 6577 MarkRefsIntoClosure::MarkRefsIntoClosure(
6571 6578 MemRegion span, CMSBitMap* bitMap):
6572 6579 _span(span),
6573 6580 _bitMap(bitMap)
6574 6581 {
6575 6582 assert(_ref_processor == NULL, "deliberately left NULL");
6576 6583 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6577 6584 }
6578 6585
6579 6586 void MarkRefsIntoClosure::do_oop(oop obj) {
6580 6587 // if p points into _span, then mark corresponding bit in _markBitMap
6581 6588 assert(obj->is_oop(), "expected an oop");
6582 6589 HeapWord* addr = (HeapWord*)obj;
6583 6590 if (_span.contains(addr)) {
6584 6591 // this should be made more efficient
6585 6592 _bitMap->mark(addr);
6586 6593 }
6587 6594 }
6588 6595
6589 6596 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6590 6597 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6591 6598
6592 6599 // A variant of the above, used for CMS marking verification.
6593 6600 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6594 6601 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6595 6602 _span(span),
6596 6603 _verification_bm(verification_bm),
6597 6604 _cms_bm(cms_bm)
6598 6605 {
6599 6606 assert(_ref_processor == NULL, "deliberately left NULL");
6600 6607 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6601 6608 }
6602 6609
6603 6610 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6604 6611 // if p points into _span, then mark corresponding bit in _markBitMap
6605 6612 assert(obj->is_oop(), "expected an oop");
6606 6613 HeapWord* addr = (HeapWord*)obj;
6607 6614 if (_span.contains(addr)) {
6608 6615 _verification_bm->mark(addr);
6609 6616 if (!_cms_bm->isMarked(addr)) {
6610 6617 oop(addr)->print();
6611 6618 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6612 6619 fatal("... aborting");
6613 6620 }
6614 6621 }
6615 6622 }
6616 6623
6617 6624 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6618 6625 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6619 6626
6620 6627 //////////////////////////////////////////////////
6621 6628 // MarkRefsIntoAndScanClosure
6622 6629 //////////////////////////////////////////////////
6623 6630
6624 6631 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6625 6632 ReferenceProcessor* rp,
6626 6633 CMSBitMap* bit_map,
6627 6634 CMSBitMap* mod_union_table,
6628 6635 CMSMarkStack* mark_stack,
6629 6636 CMSMarkStack* revisit_stack,
6630 6637 CMSCollector* collector,
6631 6638 bool should_yield,
6632 6639 bool concurrent_precleaning):
6633 6640 _collector(collector),
6634 6641 _span(span),
6635 6642 _bit_map(bit_map),
6636 6643 _mark_stack(mark_stack),
6637 6644 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6638 6645 mark_stack, revisit_stack, concurrent_precleaning),
6639 6646 _yield(should_yield),
6640 6647 _concurrent_precleaning(concurrent_precleaning),
6641 6648 _freelistLock(NULL)
6642 6649 {
6643 6650 _ref_processor = rp;
6644 6651 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6645 6652 }
6646 6653
6647 6654 // This closure is used to mark refs into the CMS generation at the
6648 6655 // second (final) checkpoint, and to scan and transitively follow
6649 6656 // the unmarked oops. It is also used during the concurrent precleaning
6650 6657 // phase while scanning objects on dirty cards in the CMS generation.
6651 6658 // The marks are made in the marking bit map and the marking stack is
6652 6659 // used for keeping the (newly) grey objects during the scan.
6653 6660 // The parallel version (Par_...) appears further below.
6654 6661 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6655 6662 if (obj != NULL) {
6656 6663 assert(obj->is_oop(), "expected an oop");
6657 6664 HeapWord* addr = (HeapWord*)obj;
6658 6665 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6659 6666 assert(_collector->overflow_list_is_empty(),
6660 6667 "overflow list should be empty");
6661 6668 if (_span.contains(addr) &&
6662 6669 !_bit_map->isMarked(addr)) {
6663 6670 // mark bit map (object is now grey)
6664 6671 _bit_map->mark(addr);
6665 6672 // push on marking stack (stack should be empty), and drain the
6666 6673 // stack by applying this closure to the oops in the oops popped
6667 6674 // from the stack (i.e. blacken the grey objects)
6668 6675 bool res = _mark_stack->push(obj);
6669 6676 assert(res, "Should have space to push on empty stack");
6670 6677 do {
6671 6678 oop new_oop = _mark_stack->pop();
6672 6679 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6673 6680 assert(new_oop->is_parsable(), "Found unparsable oop");
6674 6681 assert(_bit_map->isMarked((HeapWord*)new_oop),
6675 6682 "only grey objects on this stack");
6676 6683 // iterate over the oops in this oop, marking and pushing
6677 6684 // the ones in CMS heap (i.e. in _span).
6678 6685 new_oop->oop_iterate(&_pushAndMarkClosure);
6679 6686 // check if it's time to yield
6680 6687 do_yield_check();
6681 6688 } while (!_mark_stack->isEmpty() ||
6682 6689 (!_concurrent_precleaning && take_from_overflow_list()));
6683 6690 // if marking stack is empty, and we are not doing this
6684 6691 // during precleaning, then check the overflow list
6685 6692 }
6686 6693 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6687 6694 assert(_collector->overflow_list_is_empty(),
6688 6695 "overflow list was drained above");
6689 6696 // We could restore evacuated mark words, if any, used for
6690 6697 // overflow list links here because the overflow list is
6691 6698 // provably empty here. That would reduce the maximum
6692 6699 // size requirements for preserved_{oop,mark}_stack.
6693 6700 // But we'll just postpone it until we are all done
6694 6701 // so we can just stream through.
6695 6702 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6696 6703 _collector->restore_preserved_marks_if_any();
6697 6704 assert(_collector->no_preserved_marks(), "No preserved marks");
6698 6705 }
6699 6706 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6700 6707 "All preserved marks should have been restored above");
6701 6708 }
6702 6709 }
6703 6710
6704 6711 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6705 6712 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6706 6713
6707 6714 void MarkRefsIntoAndScanClosure::do_yield_work() {
6708 6715 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6709 6716 "CMS thread should hold CMS token");
6710 6717 assert_lock_strong(_freelistLock);
6711 6718 assert_lock_strong(_bit_map->lock());
6712 6719 // relinquish the free_list_lock and bitMaplock()
6713 6720 DEBUG_ONLY(RememberKlassesChecker mux(false);)
6714 6721 _bit_map->lock()->unlock();
6715 6722 _freelistLock->unlock();
6716 6723 ConcurrentMarkSweepThread::desynchronize(true);
6717 6724 ConcurrentMarkSweepThread::acknowledge_yield_request();
6718 6725 _collector->stopTimer();
6719 6726 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6720 6727 if (PrintCMSStatistics != 0) {
6721 6728 _collector->incrementYields();
6722 6729 }
6723 6730 _collector->icms_wait();
6724 6731
6725 6732 // See the comment in coordinator_yield()
6726 6733 for (unsigned i = 0;
6727 6734 i < CMSYieldSleepCount &&
6728 6735 ConcurrentMarkSweepThread::should_yield() &&
6729 6736 !CMSCollector::foregroundGCIsActive();
6730 6737 ++i) {
6731 6738 os::sleep(Thread::current(), 1, false);
6732 6739 ConcurrentMarkSweepThread::acknowledge_yield_request();
6733 6740 }
6734 6741
6735 6742 ConcurrentMarkSweepThread::synchronize(true);
6736 6743 _freelistLock->lock_without_safepoint_check();
6737 6744 _bit_map->lock()->lock_without_safepoint_check();
6738 6745 _collector->startTimer();
6739 6746 }
6740 6747
6741 6748 ///////////////////////////////////////////////////////////
6742 6749 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6743 6750 // MarkRefsIntoAndScanClosure
6744 6751 ///////////////////////////////////////////////////////////
6745 6752 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6746 6753 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6747 6754 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack):
6748 6755 _span(span),
6749 6756 _bit_map(bit_map),
6750 6757 _work_queue(work_queue),
6751 6758 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6752 6759 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6753 6760 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue,
6754 6761 revisit_stack)
6755 6762 {
6756 6763 _ref_processor = rp;
6757 6764 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6758 6765 }
6759 6766
6760 6767 // This closure is used to mark refs into the CMS generation at the
6761 6768 // second (final) checkpoint, and to scan and transitively follow
6762 6769 // the unmarked oops. The marks are made in the marking bit map and
6763 6770 // the work_queue is used for keeping the (newly) grey objects during
6764 6771 // the scan phase whence they are also available for stealing by parallel
6765 6772 // threads. Since the marking bit map is shared, updates are
6766 6773 // synchronized (via CAS).
6767 6774 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6768 6775 if (obj != NULL) {
6769 6776 // Ignore mark word because this could be an already marked oop
6770 6777 // that may be chained at the end of the overflow list.
6771 6778 assert(obj->is_oop(true), "expected an oop");
6772 6779 HeapWord* addr = (HeapWord*)obj;
6773 6780 if (_span.contains(addr) &&
6774 6781 !_bit_map->isMarked(addr)) {
6775 6782 // mark bit map (object will become grey):
6776 6783 // It is possible for several threads to be
6777 6784 // trying to "claim" this object concurrently;
6778 6785 // the unique thread that succeeds in marking the
6779 6786 // object first will do the subsequent push on
6780 6787 // to the work queue (or overflow list).
6781 6788 if (_bit_map->par_mark(addr)) {
6782 6789 // push on work_queue (which may not be empty), and trim the
6783 6790 // queue to an appropriate length by applying this closure to
6784 6791 // the oops in the oops popped from the stack (i.e. blacken the
6785 6792 // grey objects)
6786 6793 bool res = _work_queue->push(obj);
6787 6794 assert(res, "Low water mark should be less than capacity?");
6788 6795 trim_queue(_low_water_mark);
6789 6796 } // Else, another thread claimed the object
6790 6797 }
6791 6798 }
6792 6799 }
6793 6800
6794 6801 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6795 6802 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6796 6803
6797 6804 // This closure is used to rescan the marked objects on the dirty cards
6798 6805 // in the mod union table and the card table proper.
6799 6806 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6800 6807 oop p, MemRegion mr) {
6801 6808
6802 6809 size_t size = 0;
6803 6810 HeapWord* addr = (HeapWord*)p;
6804 6811 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6805 6812 assert(_span.contains(addr), "we are scanning the CMS generation");
6806 6813 // check if it's time to yield
6807 6814 if (do_yield_check()) {
6808 6815 // We yielded for some foreground stop-world work,
6809 6816 // and we have been asked to abort this ongoing preclean cycle.
6810 6817 return 0;
6811 6818 }
6812 6819 if (_bitMap->isMarked(addr)) {
6813 6820 // it's marked; is it potentially uninitialized?
6814 6821 if (p->klass_or_null() != NULL) {
6815 6822 // If is_conc_safe is false, the object may be undergoing
6816 6823 // change by the VM outside a safepoint. Don't try to
6817 6824 // scan it, but rather leave it for the remark phase.
6818 6825 if (CMSPermGenPrecleaningEnabled &&
6819 6826 (!p->is_conc_safe() || !p->is_parsable())) {
6820 6827 // Signal precleaning to redirty the card since
6821 6828 // the klass pointer is already installed.
6822 6829 assert(size == 0, "Initial value");
6823 6830 } else {
6824 6831 assert(p->is_parsable(), "must be parsable.");
6825 6832 // an initialized object; ignore mark word in verification below
6826 6833 // since we are running concurrent with mutators
6827 6834 assert(p->is_oop(true), "should be an oop");
6828 6835 if (p->is_objArray()) {
6829 6836 // objArrays are precisely marked; restrict scanning
6830 6837 // to dirty cards only.
6831 6838 size = CompactibleFreeListSpace::adjustObjectSize(
6832 6839 p->oop_iterate(_scanningClosure, mr));
6833 6840 } else {
6834 6841 // A non-array may have been imprecisely marked; we need
6835 6842 // to scan object in its entirety.
6836 6843 size = CompactibleFreeListSpace::adjustObjectSize(
6837 6844 p->oop_iterate(_scanningClosure));
6838 6845 }
6839 6846 #ifdef DEBUG
6840 6847 size_t direct_size =
6841 6848 CompactibleFreeListSpace::adjustObjectSize(p->size());
6842 6849 assert(size == direct_size, "Inconsistency in size");
6843 6850 assert(size >= 3, "Necessary for Printezis marks to work");
6844 6851 if (!_bitMap->isMarked(addr+1)) {
6845 6852 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6846 6853 } else {
6847 6854 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6848 6855 assert(_bitMap->isMarked(addr+size-1),
6849 6856 "inconsistent Printezis mark");
6850 6857 }
6851 6858 #endif // DEBUG
6852 6859 }
6853 6860 } else {
6854 6861 // an unitialized object
6855 6862 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6856 6863 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6857 6864 size = pointer_delta(nextOneAddr + 1, addr);
6858 6865 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6859 6866 "alignment problem");
6860 6867 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6861 6868 // will dirty the card when the klass pointer is installed in the
6862 6869 // object (signalling the completion of initialization).
6863 6870 }
6864 6871 } else {
6865 6872 // Either a not yet marked object or an uninitialized object
6866 6873 if (p->klass_or_null() == NULL || !p->is_parsable()) {
6867 6874 // An uninitialized object, skip to the next card, since
6868 6875 // we may not be able to read its P-bits yet.
6869 6876 assert(size == 0, "Initial value");
6870 6877 } else {
6871 6878 // An object not (yet) reached by marking: we merely need to
6872 6879 // compute its size so as to go look at the next block.
6873 6880 assert(p->is_oop(true), "should be an oop");
6874 6881 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6875 6882 }
6876 6883 }
6877 6884 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6878 6885 return size;
6879 6886 }
6880 6887
6881 6888 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6882 6889 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6883 6890 "CMS thread should hold CMS token");
6884 6891 assert_lock_strong(_freelistLock);
6885 6892 assert_lock_strong(_bitMap->lock());
6886 6893 DEBUG_ONLY(RememberKlassesChecker mux(false);)
6887 6894 // relinquish the free_list_lock and bitMaplock()
6888 6895 _bitMap->lock()->unlock();
6889 6896 _freelistLock->unlock();
6890 6897 ConcurrentMarkSweepThread::desynchronize(true);
6891 6898 ConcurrentMarkSweepThread::acknowledge_yield_request();
6892 6899 _collector->stopTimer();
6893 6900 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6894 6901 if (PrintCMSStatistics != 0) {
6895 6902 _collector->incrementYields();
6896 6903 }
6897 6904 _collector->icms_wait();
6898 6905
6899 6906 // See the comment in coordinator_yield()
6900 6907 for (unsigned i = 0; i < CMSYieldSleepCount &&
6901 6908 ConcurrentMarkSweepThread::should_yield() &&
6902 6909 !CMSCollector::foregroundGCIsActive(); ++i) {
6903 6910 os::sleep(Thread::current(), 1, false);
6904 6911 ConcurrentMarkSweepThread::acknowledge_yield_request();
6905 6912 }
6906 6913
6907 6914 ConcurrentMarkSweepThread::synchronize(true);
6908 6915 _freelistLock->lock_without_safepoint_check();
6909 6916 _bitMap->lock()->lock_without_safepoint_check();
6910 6917 _collector->startTimer();
6911 6918 }
6912 6919
6913 6920
6914 6921 //////////////////////////////////////////////////////////////////
6915 6922 // SurvivorSpacePrecleanClosure
6916 6923 //////////////////////////////////////////////////////////////////
6917 6924 // This (single-threaded) closure is used to preclean the oops in
6918 6925 // the survivor spaces.
6919 6926 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6920 6927
6921 6928 HeapWord* addr = (HeapWord*)p;
6922 6929 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6923 6930 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6924 6931 assert(p->klass_or_null() != NULL, "object should be initializd");
6925 6932 assert(p->is_parsable(), "must be parsable.");
6926 6933 // an initialized object; ignore mark word in verification below
6927 6934 // since we are running concurrent with mutators
6928 6935 assert(p->is_oop(true), "should be an oop");
6929 6936 // Note that we do not yield while we iterate over
6930 6937 // the interior oops of p, pushing the relevant ones
6931 6938 // on our marking stack.
6932 6939 size_t size = p->oop_iterate(_scanning_closure);
6933 6940 do_yield_check();
6934 6941 // Observe that below, we do not abandon the preclean
6935 6942 // phase as soon as we should; rather we empty the
6936 6943 // marking stack before returning. This is to satisfy
6937 6944 // some existing assertions. In general, it may be a
6938 6945 // good idea to abort immediately and complete the marking
6939 6946 // from the grey objects at a later time.
6940 6947 while (!_mark_stack->isEmpty()) {
6941 6948 oop new_oop = _mark_stack->pop();
6942 6949 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6943 6950 assert(new_oop->is_parsable(), "Found unparsable oop");
6944 6951 assert(_bit_map->isMarked((HeapWord*)new_oop),
6945 6952 "only grey objects on this stack");
6946 6953 // iterate over the oops in this oop, marking and pushing
6947 6954 // the ones in CMS heap (i.e. in _span).
6948 6955 new_oop->oop_iterate(_scanning_closure);
6949 6956 // check if it's time to yield
6950 6957 do_yield_check();
6951 6958 }
6952 6959 unsigned int after_count =
6953 6960 GenCollectedHeap::heap()->total_collections();
6954 6961 bool abort = (_before_count != after_count) ||
6955 6962 _collector->should_abort_preclean();
6956 6963 return abort ? 0 : size;
6957 6964 }
6958 6965
6959 6966 void SurvivorSpacePrecleanClosure::do_yield_work() {
6960 6967 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6961 6968 "CMS thread should hold CMS token");
6962 6969 assert_lock_strong(_bit_map->lock());
6963 6970 DEBUG_ONLY(RememberKlassesChecker smx(false);)
6964 6971 // Relinquish the bit map lock
6965 6972 _bit_map->lock()->unlock();
6966 6973 ConcurrentMarkSweepThread::desynchronize(true);
6967 6974 ConcurrentMarkSweepThread::acknowledge_yield_request();
6968 6975 _collector->stopTimer();
6969 6976 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6970 6977 if (PrintCMSStatistics != 0) {
6971 6978 _collector->incrementYields();
6972 6979 }
6973 6980 _collector->icms_wait();
6974 6981
6975 6982 // See the comment in coordinator_yield()
6976 6983 for (unsigned i = 0; i < CMSYieldSleepCount &&
6977 6984 ConcurrentMarkSweepThread::should_yield() &&
6978 6985 !CMSCollector::foregroundGCIsActive(); ++i) {
6979 6986 os::sleep(Thread::current(), 1, false);
6980 6987 ConcurrentMarkSweepThread::acknowledge_yield_request();
6981 6988 }
6982 6989
6983 6990 ConcurrentMarkSweepThread::synchronize(true);
6984 6991 _bit_map->lock()->lock_without_safepoint_check();
6985 6992 _collector->startTimer();
6986 6993 }
6987 6994
6988 6995 // This closure is used to rescan the marked objects on the dirty cards
6989 6996 // in the mod union table and the card table proper. In the parallel
6990 6997 // case, although the bitMap is shared, we do a single read so the
6991 6998 // isMarked() query is "safe".
6992 6999 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6993 7000 // Ignore mark word because we are running concurrent with mutators
6994 7001 assert(p->is_oop_or_null(true), "expected an oop or null");
6995 7002 HeapWord* addr = (HeapWord*)p;
6996 7003 assert(_span.contains(addr), "we are scanning the CMS generation");
6997 7004 bool is_obj_array = false;
6998 7005 #ifdef DEBUG
6999 7006 if (!_parallel) {
7000 7007 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
7001 7008 assert(_collector->overflow_list_is_empty(),
7002 7009 "overflow list should be empty");
7003 7010
7004 7011 }
7005 7012 #endif // DEBUG
7006 7013 if (_bit_map->isMarked(addr)) {
7007 7014 // Obj arrays are precisely marked, non-arrays are not;
7008 7015 // so we scan objArrays precisely and non-arrays in their
7009 7016 // entirety.
7010 7017 if (p->is_objArray()) {
7011 7018 is_obj_array = true;
7012 7019 if (_parallel) {
7013 7020 p->oop_iterate(_par_scan_closure, mr);
7014 7021 } else {
7015 7022 p->oop_iterate(_scan_closure, mr);
7016 7023 }
7017 7024 } else {
7018 7025 if (_parallel) {
7019 7026 p->oop_iterate(_par_scan_closure);
7020 7027 } else {
7021 7028 p->oop_iterate(_scan_closure);
7022 7029 }
7023 7030 }
7024 7031 }
7025 7032 #ifdef DEBUG
7026 7033 if (!_parallel) {
7027 7034 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
7028 7035 assert(_collector->overflow_list_is_empty(),
7029 7036 "overflow list should be empty");
7030 7037
7031 7038 }
7032 7039 #endif // DEBUG
7033 7040 return is_obj_array;
7034 7041 }
7035 7042
7036 7043 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
7037 7044 MemRegion span,
7038 7045 CMSBitMap* bitMap, CMSMarkStack* markStack,
7039 7046 CMSMarkStack* revisitStack,
7040 7047 bool should_yield, bool verifying):
7041 7048 _collector(collector),
7042 7049 _span(span),
7043 7050 _bitMap(bitMap),
7044 7051 _mut(&collector->_modUnionTable),
7045 7052 _markStack(markStack),
7046 7053 _revisitStack(revisitStack),
7047 7054 _yield(should_yield),
7048 7055 _skipBits(0)
7049 7056 {
7050 7057 assert(_markStack->isEmpty(), "stack should be empty");
7051 7058 _finger = _bitMap->startWord();
7052 7059 _threshold = _finger;
7053 7060 assert(_collector->_restart_addr == NULL, "Sanity check");
7054 7061 assert(_span.contains(_finger), "Out of bounds _finger?");
7055 7062 DEBUG_ONLY(_verifying = verifying;)
7056 7063 }
7057 7064
7058 7065 void MarkFromRootsClosure::reset(HeapWord* addr) {
7059 7066 assert(_markStack->isEmpty(), "would cause duplicates on stack");
7060 7067 assert(_span.contains(addr), "Out of bounds _finger?");
7061 7068 _finger = addr;
7062 7069 _threshold = (HeapWord*)round_to(
7063 7070 (intptr_t)_finger, CardTableModRefBS::card_size);
7064 7071 }
7065 7072
7066 7073 // Should revisit to see if this should be restructured for
7067 7074 // greater efficiency.
7068 7075 bool MarkFromRootsClosure::do_bit(size_t offset) {
7069 7076 if (_skipBits > 0) {
7070 7077 _skipBits--;
7071 7078 return true;
7072 7079 }
7073 7080 // convert offset into a HeapWord*
7074 7081 HeapWord* addr = _bitMap->startWord() + offset;
7075 7082 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
7076 7083 "address out of range");
7077 7084 assert(_bitMap->isMarked(addr), "tautology");
7078 7085 if (_bitMap->isMarked(addr+1)) {
7079 7086 // this is an allocated but not yet initialized object
7080 7087 assert(_skipBits == 0, "tautology");
7081 7088 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
7082 7089 oop p = oop(addr);
7083 7090 if (p->klass_or_null() == NULL || !p->is_parsable()) {
7084 7091 DEBUG_ONLY(if (!_verifying) {)
7085 7092 // We re-dirty the cards on which this object lies and increase
7086 7093 // the _threshold so that we'll come back to scan this object
7087 7094 // during the preclean or remark phase. (CMSCleanOnEnter)
7088 7095 if (CMSCleanOnEnter) {
7089 7096 size_t sz = _collector->block_size_using_printezis_bits(addr);
7090 7097 HeapWord* end_card_addr = (HeapWord*)round_to(
7091 7098 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7092 7099 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7093 7100 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7094 7101 // Bump _threshold to end_card_addr; note that
7095 7102 // _threshold cannot possibly exceed end_card_addr, anyhow.
7096 7103 // This prevents future clearing of the card as the scan proceeds
7097 7104 // to the right.
7098 7105 assert(_threshold <= end_card_addr,
7099 7106 "Because we are just scanning into this object");
7100 7107 if (_threshold < end_card_addr) {
7101 7108 _threshold = end_card_addr;
7102 7109 }
7103 7110 if (p->klass_or_null() != NULL) {
7104 7111 // Redirty the range of cards...
7105 7112 _mut->mark_range(redirty_range);
7106 7113 } // ...else the setting of klass will dirty the card anyway.
7107 7114 }
7108 7115 DEBUG_ONLY(})
7109 7116 return true;
7110 7117 }
7111 7118 }
7112 7119 scanOopsInOop(addr);
7113 7120 return true;
7114 7121 }
7115 7122
7116 7123 // We take a break if we've been at this for a while,
7117 7124 // so as to avoid monopolizing the locks involved.
7118 7125 void MarkFromRootsClosure::do_yield_work() {
7119 7126 // First give up the locks, then yield, then re-lock
7120 7127 // We should probably use a constructor/destructor idiom to
7121 7128 // do this unlock/lock or modify the MutexUnlocker class to
7122 7129 // serve our purpose. XXX
7123 7130 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7124 7131 "CMS thread should hold CMS token");
7125 7132 assert_lock_strong(_bitMap->lock());
7126 7133 DEBUG_ONLY(RememberKlassesChecker mux(false);)
7127 7134 _bitMap->lock()->unlock();
7128 7135 ConcurrentMarkSweepThread::desynchronize(true);
7129 7136 ConcurrentMarkSweepThread::acknowledge_yield_request();
7130 7137 _collector->stopTimer();
7131 7138 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7132 7139 if (PrintCMSStatistics != 0) {
7133 7140 _collector->incrementYields();
7134 7141 }
7135 7142 _collector->icms_wait();
7136 7143
7137 7144 // See the comment in coordinator_yield()
7138 7145 for (unsigned i = 0; i < CMSYieldSleepCount &&
7139 7146 ConcurrentMarkSweepThread::should_yield() &&
7140 7147 !CMSCollector::foregroundGCIsActive(); ++i) {
7141 7148 os::sleep(Thread::current(), 1, false);
7142 7149 ConcurrentMarkSweepThread::acknowledge_yield_request();
7143 7150 }
7144 7151
7145 7152 ConcurrentMarkSweepThread::synchronize(true);
7146 7153 _bitMap->lock()->lock_without_safepoint_check();
7147 7154 _collector->startTimer();
7148 7155 }
7149 7156
7150 7157 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
7151 7158 assert(_bitMap->isMarked(ptr), "expected bit to be set");
7152 7159 assert(_markStack->isEmpty(),
7153 7160 "should drain stack to limit stack usage");
7154 7161 // convert ptr to an oop preparatory to scanning
7155 7162 oop obj = oop(ptr);
7156 7163 // Ignore mark word in verification below, since we
7157 7164 // may be running concurrent with mutators.
7158 7165 assert(obj->is_oop(true), "should be an oop");
7159 7166 assert(_finger <= ptr, "_finger runneth ahead");
7160 7167 // advance the finger to right end of this object
7161 7168 _finger = ptr + obj->size();
7162 7169 assert(_finger > ptr, "we just incremented it above");
7163 7170 // On large heaps, it may take us some time to get through
7164 7171 // the marking phase (especially if running iCMS). During
7165 7172 // this time it's possible that a lot of mutations have
7166 7173 // accumulated in the card table and the mod union table --
7167 7174 // these mutation records are redundant until we have
7168 7175 // actually traced into the corresponding card.
7169 7176 // Here, we check whether advancing the finger would make
7170 7177 // us cross into a new card, and if so clear corresponding
7171 7178 // cards in the MUT (preclean them in the card-table in the
7172 7179 // future).
7173 7180
7174 7181 DEBUG_ONLY(if (!_verifying) {)
7175 7182 // The clean-on-enter optimization is disabled by default,
7176 7183 // until we fix 6178663.
7177 7184 if (CMSCleanOnEnter && (_finger > _threshold)) {
7178 7185 // [_threshold, _finger) represents the interval
7179 7186 // of cards to be cleared in MUT (or precleaned in card table).
7180 7187 // The set of cards to be cleared is all those that overlap
7181 7188 // with the interval [_threshold, _finger); note that
7182 7189 // _threshold is always kept card-aligned but _finger isn't
7183 7190 // always card-aligned.
7184 7191 HeapWord* old_threshold = _threshold;
7185 7192 assert(old_threshold == (HeapWord*)round_to(
7186 7193 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7187 7194 "_threshold should always be card-aligned");
7188 7195 _threshold = (HeapWord*)round_to(
7189 7196 (intptr_t)_finger, CardTableModRefBS::card_size);
7190 7197 MemRegion mr(old_threshold, _threshold);
7191 7198 assert(!mr.is_empty(), "Control point invariant");
7192 7199 assert(_span.contains(mr), "Should clear within span");
7193 7200 // XXX When _finger crosses from old gen into perm gen
7194 7201 // we may be doing unnecessary cleaning; do better in the
7195 7202 // future by detecting that condition and clearing fewer
7196 7203 // MUT/CT entries.
7197 7204 _mut->clear_range(mr);
7198 7205 }
7199 7206 DEBUG_ONLY(})
7200 7207 // Note: the finger doesn't advance while we drain
7201 7208 // the stack below.
7202 7209 PushOrMarkClosure pushOrMarkClosure(_collector,
7203 7210 _span, _bitMap, _markStack,
7204 7211 _revisitStack,
7205 7212 _finger, this);
7206 7213 bool res = _markStack->push(obj);
7207 7214 assert(res, "Empty non-zero size stack should have space for single push");
7208 7215 while (!_markStack->isEmpty()) {
7209 7216 oop new_oop = _markStack->pop();
7210 7217 // Skip verifying header mark word below because we are
7211 7218 // running concurrent with mutators.
7212 7219 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7213 7220 // now scan this oop's oops
7214 7221 new_oop->oop_iterate(&pushOrMarkClosure);
7215 7222 do_yield_check();
7216 7223 }
7217 7224 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
7218 7225 }
7219 7226
7220 7227 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
7221 7228 CMSCollector* collector, MemRegion span,
7222 7229 CMSBitMap* bit_map,
7223 7230 OopTaskQueue* work_queue,
7224 7231 CMSMarkStack* overflow_stack,
7225 7232 CMSMarkStack* revisit_stack,
7226 7233 bool should_yield):
7227 7234 _collector(collector),
7228 7235 _whole_span(collector->_span),
7229 7236 _span(span),
7230 7237 _bit_map(bit_map),
7231 7238 _mut(&collector->_modUnionTable),
7232 7239 _work_queue(work_queue),
7233 7240 _overflow_stack(overflow_stack),
7234 7241 _revisit_stack(revisit_stack),
7235 7242 _yield(should_yield),
7236 7243 _skip_bits(0),
7237 7244 _task(task)
7238 7245 {
7239 7246 assert(_work_queue->size() == 0, "work_queue should be empty");
7240 7247 _finger = span.start();
7241 7248 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
7242 7249 assert(_span.contains(_finger), "Out of bounds _finger?");
7243 7250 }
7244 7251
7245 7252 // Should revisit to see if this should be restructured for
7246 7253 // greater efficiency.
7247 7254 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
7248 7255 if (_skip_bits > 0) {
7249 7256 _skip_bits--;
7250 7257 return true;
7251 7258 }
7252 7259 // convert offset into a HeapWord*
7253 7260 HeapWord* addr = _bit_map->startWord() + offset;
7254 7261 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
7255 7262 "address out of range");
7256 7263 assert(_bit_map->isMarked(addr), "tautology");
7257 7264 if (_bit_map->isMarked(addr+1)) {
7258 7265 // this is an allocated object that might not yet be initialized
7259 7266 assert(_skip_bits == 0, "tautology");
7260 7267 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
7261 7268 oop p = oop(addr);
7262 7269 if (p->klass_or_null() == NULL || !p->is_parsable()) {
7263 7270 // in the case of Clean-on-Enter optimization, redirty card
7264 7271 // and avoid clearing card by increasing the threshold.
7265 7272 return true;
7266 7273 }
7267 7274 }
7268 7275 scan_oops_in_oop(addr);
7269 7276 return true;
7270 7277 }
7271 7278
7272 7279 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
7273 7280 assert(_bit_map->isMarked(ptr), "expected bit to be set");
7274 7281 // Should we assert that our work queue is empty or
7275 7282 // below some drain limit?
7276 7283 assert(_work_queue->size() == 0,
7277 7284 "should drain stack to limit stack usage");
7278 7285 // convert ptr to an oop preparatory to scanning
7279 7286 oop obj = oop(ptr);
7280 7287 // Ignore mark word in verification below, since we
7281 7288 // may be running concurrent with mutators.
7282 7289 assert(obj->is_oop(true), "should be an oop");
7283 7290 assert(_finger <= ptr, "_finger runneth ahead");
7284 7291 // advance the finger to right end of this object
7285 7292 _finger = ptr + obj->size();
7286 7293 assert(_finger > ptr, "we just incremented it above");
7287 7294 // On large heaps, it may take us some time to get through
7288 7295 // the marking phase (especially if running iCMS). During
7289 7296 // this time it's possible that a lot of mutations have
7290 7297 // accumulated in the card table and the mod union table --
7291 7298 // these mutation records are redundant until we have
7292 7299 // actually traced into the corresponding card.
7293 7300 // Here, we check whether advancing the finger would make
7294 7301 // us cross into a new card, and if so clear corresponding
7295 7302 // cards in the MUT (preclean them in the card-table in the
7296 7303 // future).
7297 7304
7298 7305 // The clean-on-enter optimization is disabled by default,
7299 7306 // until we fix 6178663.
7300 7307 if (CMSCleanOnEnter && (_finger > _threshold)) {
7301 7308 // [_threshold, _finger) represents the interval
7302 7309 // of cards to be cleared in MUT (or precleaned in card table).
7303 7310 // The set of cards to be cleared is all those that overlap
7304 7311 // with the interval [_threshold, _finger); note that
7305 7312 // _threshold is always kept card-aligned but _finger isn't
7306 7313 // always card-aligned.
7307 7314 HeapWord* old_threshold = _threshold;
7308 7315 assert(old_threshold == (HeapWord*)round_to(
7309 7316 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7310 7317 "_threshold should always be card-aligned");
7311 7318 _threshold = (HeapWord*)round_to(
7312 7319 (intptr_t)_finger, CardTableModRefBS::card_size);
7313 7320 MemRegion mr(old_threshold, _threshold);
7314 7321 assert(!mr.is_empty(), "Control point invariant");
7315 7322 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7316 7323 // XXX When _finger crosses from old gen into perm gen
7317 7324 // we may be doing unnecessary cleaning; do better in the
7318 7325 // future by detecting that condition and clearing fewer
7319 7326 // MUT/CT entries.
7320 7327 _mut->clear_range(mr);
7321 7328 }
7322 7329
7323 7330 // Note: the local finger doesn't advance while we drain
7324 7331 // the stack below, but the global finger sure can and will.
7325 7332 HeapWord** gfa = _task->global_finger_addr();
7326 7333 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7327 7334 _span, _bit_map,
7328 7335 _work_queue,
7329 7336 _overflow_stack,
7330 7337 _revisit_stack,
7331 7338 _finger,
7332 7339 gfa, this);
7333 7340 bool res = _work_queue->push(obj); // overflow could occur here
7334 7341 assert(res, "Will hold once we use workqueues");
7335 7342 while (true) {
7336 7343 oop new_oop;
7337 7344 if (!_work_queue->pop_local(new_oop)) {
7338 7345 // We emptied our work_queue; check if there's stuff that can
7339 7346 // be gotten from the overflow stack.
7340 7347 if (CMSConcMarkingTask::get_work_from_overflow_stack(
7341 7348 _overflow_stack, _work_queue)) {
7342 7349 do_yield_check();
7343 7350 continue;
7344 7351 } else { // done
7345 7352 break;
7346 7353 }
7347 7354 }
7348 7355 // Skip verifying header mark word below because we are
7349 7356 // running concurrent with mutators.
7350 7357 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7351 7358 // now scan this oop's oops
7352 7359 new_oop->oop_iterate(&pushOrMarkClosure);
7353 7360 do_yield_check();
7354 7361 }
7355 7362 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7356 7363 }
7357 7364
7358 7365 // Yield in response to a request from VM Thread or
7359 7366 // from mutators.
7360 7367 void Par_MarkFromRootsClosure::do_yield_work() {
7361 7368 assert(_task != NULL, "sanity");
7362 7369 _task->yield();
7363 7370 }
7364 7371
7365 7372 // A variant of the above used for verifying CMS marking work.
7366 7373 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7367 7374 MemRegion span,
7368 7375 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7369 7376 CMSMarkStack* mark_stack):
7370 7377 _collector(collector),
7371 7378 _span(span),
7372 7379 _verification_bm(verification_bm),
7373 7380 _cms_bm(cms_bm),
7374 7381 _mark_stack(mark_stack),
7375 7382 _pam_verify_closure(collector, span, verification_bm, cms_bm,
7376 7383 mark_stack)
7377 7384 {
7378 7385 assert(_mark_stack->isEmpty(), "stack should be empty");
7379 7386 _finger = _verification_bm->startWord();
7380 7387 assert(_collector->_restart_addr == NULL, "Sanity check");
7381 7388 assert(_span.contains(_finger), "Out of bounds _finger?");
7382 7389 }
7383 7390
7384 7391 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7385 7392 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7386 7393 assert(_span.contains(addr), "Out of bounds _finger?");
7387 7394 _finger = addr;
7388 7395 }
7389 7396
7390 7397 // Should revisit to see if this should be restructured for
7391 7398 // greater efficiency.
7392 7399 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7393 7400 // convert offset into a HeapWord*
7394 7401 HeapWord* addr = _verification_bm->startWord() + offset;
7395 7402 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7396 7403 "address out of range");
7397 7404 assert(_verification_bm->isMarked(addr), "tautology");
7398 7405 assert(_cms_bm->isMarked(addr), "tautology");
7399 7406
7400 7407 assert(_mark_stack->isEmpty(),
7401 7408 "should drain stack to limit stack usage");
7402 7409 // convert addr to an oop preparatory to scanning
7403 7410 oop obj = oop(addr);
7404 7411 assert(obj->is_oop(), "should be an oop");
7405 7412 assert(_finger <= addr, "_finger runneth ahead");
7406 7413 // advance the finger to right end of this object
7407 7414 _finger = addr + obj->size();
7408 7415 assert(_finger > addr, "we just incremented it above");
7409 7416 // Note: the finger doesn't advance while we drain
7410 7417 // the stack below.
7411 7418 bool res = _mark_stack->push(obj);
7412 7419 assert(res, "Empty non-zero size stack should have space for single push");
7413 7420 while (!_mark_stack->isEmpty()) {
7414 7421 oop new_oop = _mark_stack->pop();
7415 7422 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7416 7423 // now scan this oop's oops
7417 7424 new_oop->oop_iterate(&_pam_verify_closure);
7418 7425 }
7419 7426 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7420 7427 return true;
7421 7428 }
7422 7429
7423 7430 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7424 7431 CMSCollector* collector, MemRegion span,
7425 7432 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7426 7433 CMSMarkStack* mark_stack):
7427 7434 OopClosure(collector->ref_processor()),
7428 7435 _collector(collector),
7429 7436 _span(span),
7430 7437 _verification_bm(verification_bm),
7431 7438 _cms_bm(cms_bm),
7432 7439 _mark_stack(mark_stack)
7433 7440 { }
7434 7441
7435 7442 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7436 7443 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7437 7444
7438 7445 // Upon stack overflow, we discard (part of) the stack,
7439 7446 // remembering the least address amongst those discarded
7440 7447 // in CMSCollector's _restart_address.
7441 7448 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7442 7449 // Remember the least grey address discarded
7443 7450 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7444 7451 _collector->lower_restart_addr(ra);
7445 7452 _mark_stack->reset(); // discard stack contents
7446 7453 _mark_stack->expand(); // expand the stack if possible
7447 7454 }
7448 7455
7449 7456 void PushAndMarkVerifyClosure::do_oop(oop obj) {
7450 7457 assert(obj->is_oop_or_null(), "expected an oop or NULL");
7451 7458 HeapWord* addr = (HeapWord*)obj;
7452 7459 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7453 7460 // Oop lies in _span and isn't yet grey or black
7454 7461 _verification_bm->mark(addr); // now grey
7455 7462 if (!_cms_bm->isMarked(addr)) {
7456 7463 oop(addr)->print();
7457 7464 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
7458 7465 addr);
7459 7466 fatal("... aborting");
7460 7467 }
7461 7468
7462 7469 if (!_mark_stack->push(obj)) { // stack overflow
7463 7470 if (PrintCMSStatistics != 0) {
7464 7471 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7465 7472 SIZE_FORMAT, _mark_stack->capacity());
7466 7473 }
7467 7474 assert(_mark_stack->isFull(), "Else push should have succeeded");
7468 7475 handle_stack_overflow(addr);
7469 7476 }
7470 7477 // anything including and to the right of _finger
7471 7478 // will be scanned as we iterate over the remainder of the
7472 7479 // bit map
7473 7480 }
7474 7481 }
7475 7482
7476 7483 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7477 7484 MemRegion span,
7478 7485 CMSBitMap* bitMap, CMSMarkStack* markStack,
7479 7486 CMSMarkStack* revisitStack,
7480 7487 HeapWord* finger, MarkFromRootsClosure* parent) :
7481 7488 KlassRememberingOopClosure(collector, collector->ref_processor(), revisitStack),
7482 7489 _span(span),
7483 7490 _bitMap(bitMap),
7484 7491 _markStack(markStack),
7485 7492 _finger(finger),
7486 7493 _parent(parent)
7487 7494 { }
7488 7495
7489 7496 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7490 7497 MemRegion span,
7491 7498 CMSBitMap* bit_map,
7492 7499 OopTaskQueue* work_queue,
7493 7500 CMSMarkStack* overflow_stack,
7494 7501 CMSMarkStack* revisit_stack,
7495 7502 HeapWord* finger,
7496 7503 HeapWord** global_finger_addr,
7497 7504 Par_MarkFromRootsClosure* parent) :
7498 7505 Par_KlassRememberingOopClosure(collector,
7499 7506 collector->ref_processor(),
7500 7507 revisit_stack),
7501 7508 _whole_span(collector->_span),
7502 7509 _span(span),
7503 7510 _bit_map(bit_map),
7504 7511 _work_queue(work_queue),
7505 7512 _overflow_stack(overflow_stack),
7506 7513 _finger(finger),
7507 7514 _global_finger_addr(global_finger_addr),
7508 7515 _parent(parent)
7509 7516 { }
7510 7517
7511 7518 // Assumes thread-safe access by callers, who are
7512 7519 // responsible for mutual exclusion.
7513 7520 void CMSCollector::lower_restart_addr(HeapWord* low) {
7514 7521 assert(_span.contains(low), "Out of bounds addr");
7515 7522 if (_restart_addr == NULL) {
7516 7523 _restart_addr = low;
7517 7524 } else {
7518 7525 _restart_addr = MIN2(_restart_addr, low);
7519 7526 }
7520 7527 }
7521 7528
7522 7529 // Upon stack overflow, we discard (part of) the stack,
7523 7530 // remembering the least address amongst those discarded
7524 7531 // in CMSCollector's _restart_address.
7525 7532 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7526 7533 // Remember the least grey address discarded
7527 7534 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7528 7535 _collector->lower_restart_addr(ra);
7529 7536 _markStack->reset(); // discard stack contents
7530 7537 _markStack->expand(); // expand the stack if possible
7531 7538 }
7532 7539
7533 7540 // Upon stack overflow, we discard (part of) the stack,
7534 7541 // remembering the least address amongst those discarded
7535 7542 // in CMSCollector's _restart_address.
7536 7543 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7537 7544 // We need to do this under a mutex to prevent other
7538 7545 // workers from interfering with the work done below.
7539 7546 MutexLockerEx ml(_overflow_stack->par_lock(),
7540 7547 Mutex::_no_safepoint_check_flag);
7541 7548 // Remember the least grey address discarded
7542 7549 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7543 7550 _collector->lower_restart_addr(ra);
7544 7551 _overflow_stack->reset(); // discard stack contents
7545 7552 _overflow_stack->expand(); // expand the stack if possible
7546 7553 }
7547 7554
7548 7555 void PushOrMarkClosure::do_oop(oop obj) {
7549 7556 // Ignore mark word because we are running concurrent with mutators.
7550 7557 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7551 7558 HeapWord* addr = (HeapWord*)obj;
7552 7559 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7553 7560 // Oop lies in _span and isn't yet grey or black
7554 7561 _bitMap->mark(addr); // now grey
7555 7562 if (addr < _finger) {
7556 7563 // the bit map iteration has already either passed, or
7557 7564 // sampled, this bit in the bit map; we'll need to
7558 7565 // use the marking stack to scan this oop's oops.
7559 7566 bool simulate_overflow = false;
7560 7567 NOT_PRODUCT(
7561 7568 if (CMSMarkStackOverflowALot &&
7562 7569 _collector->simulate_overflow()) {
7563 7570 // simulate a stack overflow
7564 7571 simulate_overflow = true;
7565 7572 }
7566 7573 )
7567 7574 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7568 7575 if (PrintCMSStatistics != 0) {
7569 7576 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7570 7577 SIZE_FORMAT, _markStack->capacity());
7571 7578 }
7572 7579 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7573 7580 handle_stack_overflow(addr);
7574 7581 }
7575 7582 }
7576 7583 // anything including and to the right of _finger
7577 7584 // will be scanned as we iterate over the remainder of the
7578 7585 // bit map
7579 7586 do_yield_check();
7580 7587 }
7581 7588 }
7582 7589
7583 7590 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7584 7591 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7585 7592
7586 7593 void Par_PushOrMarkClosure::do_oop(oop obj) {
7587 7594 // Ignore mark word because we are running concurrent with mutators.
7588 7595 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7589 7596 HeapWord* addr = (HeapWord*)obj;
7590 7597 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7591 7598 // Oop lies in _span and isn't yet grey or black
7592 7599 // We read the global_finger (volatile read) strictly after marking oop
7593 7600 bool res = _bit_map->par_mark(addr); // now grey
7594 7601 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7595 7602 // Should we push this marked oop on our stack?
7596 7603 // -- if someone else marked it, nothing to do
7597 7604 // -- if target oop is above global finger nothing to do
7598 7605 // -- if target oop is in chunk and above local finger
7599 7606 // then nothing to do
7600 7607 // -- else push on work queue
7601 7608 if ( !res // someone else marked it, they will deal with it
7602 7609 || (addr >= *gfa) // will be scanned in a later task
7603 7610 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7604 7611 return;
7605 7612 }
7606 7613 // the bit map iteration has already either passed, or
7607 7614 // sampled, this bit in the bit map; we'll need to
7608 7615 // use the marking stack to scan this oop's oops.
7609 7616 bool simulate_overflow = false;
7610 7617 NOT_PRODUCT(
7611 7618 if (CMSMarkStackOverflowALot &&
7612 7619 _collector->simulate_overflow()) {
7613 7620 // simulate a stack overflow
7614 7621 simulate_overflow = true;
7615 7622 }
7616 7623 )
7617 7624 if (simulate_overflow ||
7618 7625 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7619 7626 // stack overflow
7620 7627 if (PrintCMSStatistics != 0) {
7621 7628 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7622 7629 SIZE_FORMAT, _overflow_stack->capacity());
7623 7630 }
7624 7631 // We cannot assert that the overflow stack is full because
7625 7632 // it may have been emptied since.
7626 7633 assert(simulate_overflow ||
7627 7634 _work_queue->size() == _work_queue->max_elems(),
7628 7635 "Else push should have succeeded");
7629 7636 handle_stack_overflow(addr);
7630 7637 }
7631 7638 do_yield_check();
7632 7639 }
7633 7640 }
7634 7641
7635 7642 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7636 7643 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7637 7644
7638 7645 KlassRememberingOopClosure::KlassRememberingOopClosure(CMSCollector* collector,
7639 7646 ReferenceProcessor* rp,
7640 7647 CMSMarkStack* revisit_stack) :
7641 7648 OopClosure(rp),
7642 7649 _collector(collector),
7643 7650 _revisit_stack(revisit_stack),
7644 7651 _should_remember_klasses(collector->should_unload_classes()) {}
7645 7652
7646 7653 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7647 7654 MemRegion span,
7648 7655 ReferenceProcessor* rp,
7649 7656 CMSBitMap* bit_map,
7650 7657 CMSBitMap* mod_union_table,
7651 7658 CMSMarkStack* mark_stack,
7652 7659 CMSMarkStack* revisit_stack,
7653 7660 bool concurrent_precleaning):
7654 7661 KlassRememberingOopClosure(collector, rp, revisit_stack),
7655 7662 _span(span),
7656 7663 _bit_map(bit_map),
7657 7664 _mod_union_table(mod_union_table),
7658 7665 _mark_stack(mark_stack),
7659 7666 _concurrent_precleaning(concurrent_precleaning)
7660 7667 {
7661 7668 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7662 7669 }
7663 7670
7664 7671 // Grey object rescan during pre-cleaning and second checkpoint phases --
7665 7672 // the non-parallel version (the parallel version appears further below.)
7666 7673 void PushAndMarkClosure::do_oop(oop obj) {
7667 7674 // Ignore mark word verification. If during concurrent precleaning,
7668 7675 // the object monitor may be locked. If during the checkpoint
7669 7676 // phases, the object may already have been reached by a different
7670 7677 // path and may be at the end of the global overflow list (so
7671 7678 // the mark word may be NULL).
7672 7679 assert(obj->is_oop_or_null(true /* ignore mark word */),
7673 7680 "expected an oop or NULL");
7674 7681 HeapWord* addr = (HeapWord*)obj;
7675 7682 // Check if oop points into the CMS generation
7676 7683 // and is not marked
7677 7684 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7678 7685 // a white object ...
7679 7686 _bit_map->mark(addr); // ... now grey
7680 7687 // push on the marking stack (grey set)
7681 7688 bool simulate_overflow = false;
7682 7689 NOT_PRODUCT(
7683 7690 if (CMSMarkStackOverflowALot &&
7684 7691 _collector->simulate_overflow()) {
7685 7692 // simulate a stack overflow
7686 7693 simulate_overflow = true;
7687 7694 }
7688 7695 )
7689 7696 if (simulate_overflow || !_mark_stack->push(obj)) {
7690 7697 if (_concurrent_precleaning) {
7691 7698 // During precleaning we can just dirty the appropriate card(s)
7692 7699 // in the mod union table, thus ensuring that the object remains
7693 7700 // in the grey set and continue. In the case of object arrays
7694 7701 // we need to dirty all of the cards that the object spans,
7695 7702 // since the rescan of object arrays will be limited to the
7696 7703 // dirty cards.
7697 7704 // Note that no one can be intefering with us in this action
7698 7705 // of dirtying the mod union table, so no locking or atomics
7699 7706 // are required.
7700 7707 if (obj->is_objArray()) {
7701 7708 size_t sz = obj->size();
7702 7709 HeapWord* end_card_addr = (HeapWord*)round_to(
7703 7710 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7704 7711 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7705 7712 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7706 7713 _mod_union_table->mark_range(redirty_range);
7707 7714 } else {
7708 7715 _mod_union_table->mark(addr);
7709 7716 }
7710 7717 _collector->_ser_pmc_preclean_ovflw++;
7711 7718 } else {
7712 7719 // During the remark phase, we need to remember this oop
7713 7720 // in the overflow list.
7714 7721 _collector->push_on_overflow_list(obj);
7715 7722 _collector->_ser_pmc_remark_ovflw++;
7716 7723 }
7717 7724 }
7718 7725 }
7719 7726 }
7720 7727
7721 7728 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7722 7729 MemRegion span,
7723 7730 ReferenceProcessor* rp,
7724 7731 CMSBitMap* bit_map,
7725 7732 OopTaskQueue* work_queue,
7726 7733 CMSMarkStack* revisit_stack):
7727 7734 Par_KlassRememberingOopClosure(collector, rp, revisit_stack),
7728 7735 _span(span),
7729 7736 _bit_map(bit_map),
7730 7737 _work_queue(work_queue)
7731 7738 {
7732 7739 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7733 7740 }
7734 7741
7735 7742 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7736 7743 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7737 7744
7738 7745 // Grey object rescan during second checkpoint phase --
7739 7746 // the parallel version.
7740 7747 void Par_PushAndMarkClosure::do_oop(oop obj) {
7741 7748 // In the assert below, we ignore the mark word because
7742 7749 // this oop may point to an already visited object that is
7743 7750 // on the overflow stack (in which case the mark word has
7744 7751 // been hijacked for chaining into the overflow stack --
7745 7752 // if this is the last object in the overflow stack then
7746 7753 // its mark word will be NULL). Because this object may
7747 7754 // have been subsequently popped off the global overflow
7748 7755 // stack, and the mark word possibly restored to the prototypical
7749 7756 // value, by the time we get to examined this failing assert in
7750 7757 // the debugger, is_oop_or_null(false) may subsequently start
7751 7758 // to hold.
7752 7759 assert(obj->is_oop_or_null(true),
7753 7760 "expected an oop or NULL");
7754 7761 HeapWord* addr = (HeapWord*)obj;
7755 7762 // Check if oop points into the CMS generation
7756 7763 // and is not marked
7757 7764 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7758 7765 // a white object ...
7759 7766 // If we manage to "claim" the object, by being the
7760 7767 // first thread to mark it, then we push it on our
7761 7768 // marking stack
7762 7769 if (_bit_map->par_mark(addr)) { // ... now grey
7763 7770 // push on work queue (grey set)
7764 7771 bool simulate_overflow = false;
7765 7772 NOT_PRODUCT(
7766 7773 if (CMSMarkStackOverflowALot &&
7767 7774 _collector->par_simulate_overflow()) {
7768 7775 // simulate a stack overflow
7769 7776 simulate_overflow = true;
7770 7777 }
7771 7778 )
7772 7779 if (simulate_overflow || !_work_queue->push(obj)) {
7773 7780 _collector->par_push_on_overflow_list(obj);
7774 7781 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7775 7782 }
7776 7783 } // Else, some other thread got there first
7777 7784 }
7778 7785 }
7779 7786
7780 7787 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7781 7788 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7782 7789
7783 7790 void PushAndMarkClosure::remember_mdo(DataLayout* v) {
7784 7791 // TBD
7785 7792 }
7786 7793
7787 7794 void Par_PushAndMarkClosure::remember_mdo(DataLayout* v) {
7788 7795 // TBD
7789 7796 }
7790 7797
7791 7798 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7792 7799 DEBUG_ONLY(RememberKlassesChecker mux(false);)
7793 7800 Mutex* bml = _collector->bitMapLock();
7794 7801 assert_lock_strong(bml);
7795 7802 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7796 7803 "CMS thread should hold CMS token");
7797 7804
7798 7805 bml->unlock();
7799 7806 ConcurrentMarkSweepThread::desynchronize(true);
7800 7807
7801 7808 ConcurrentMarkSweepThread::acknowledge_yield_request();
7802 7809
7803 7810 _collector->stopTimer();
7804 7811 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7805 7812 if (PrintCMSStatistics != 0) {
7806 7813 _collector->incrementYields();
7807 7814 }
7808 7815 _collector->icms_wait();
7809 7816
7810 7817 // See the comment in coordinator_yield()
7811 7818 for (unsigned i = 0; i < CMSYieldSleepCount &&
7812 7819 ConcurrentMarkSweepThread::should_yield() &&
7813 7820 !CMSCollector::foregroundGCIsActive(); ++i) {
7814 7821 os::sleep(Thread::current(), 1, false);
7815 7822 ConcurrentMarkSweepThread::acknowledge_yield_request();
7816 7823 }
7817 7824
7818 7825 ConcurrentMarkSweepThread::synchronize(true);
7819 7826 bml->lock();
7820 7827
7821 7828 _collector->startTimer();
7822 7829 }
7823 7830
7824 7831 bool CMSPrecleanRefsYieldClosure::should_return() {
7825 7832 if (ConcurrentMarkSweepThread::should_yield()) {
7826 7833 do_yield_work();
7827 7834 }
7828 7835 return _collector->foregroundGCIsActive();
7829 7836 }
7830 7837
7831 7838 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7832 7839 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7833 7840 "mr should be aligned to start at a card boundary");
7834 7841 // We'd like to assert:
7835 7842 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7836 7843 // "mr should be a range of cards");
7837 7844 // However, that would be too strong in one case -- the last
7838 7845 // partition ends at _unallocated_block which, in general, can be
7839 7846 // an arbitrary boundary, not necessarily card aligned.
7840 7847 if (PrintCMSStatistics != 0) {
7841 7848 _num_dirty_cards +=
7842 7849 mr.word_size()/CardTableModRefBS::card_size_in_words;
7843 7850 }
7844 7851 _space->object_iterate_mem(mr, &_scan_cl);
7845 7852 }
7846 7853
7847 7854 SweepClosure::SweepClosure(CMSCollector* collector,
7848 7855 ConcurrentMarkSweepGeneration* g,
7849 7856 CMSBitMap* bitMap, bool should_yield) :
7850 7857 _collector(collector),
7851 7858 _g(g),
7852 7859 _sp(g->cmsSpace()),
7853 7860 _limit(_sp->sweep_limit()),
7854 7861 _freelistLock(_sp->freelistLock()),
7855 7862 _bitMap(bitMap),
7856 7863 _yield(should_yield),
7857 7864 _inFreeRange(false), // No free range at beginning of sweep
7858 7865 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7859 7866 _lastFreeRangeCoalesced(false),
7860 7867 _freeFinger(g->used_region().start())
7861 7868 {
7862 7869 NOT_PRODUCT(
7863 7870 _numObjectsFreed = 0;
7864 7871 _numWordsFreed = 0;
7865 7872 _numObjectsLive = 0;
7866 7873 _numWordsLive = 0;
7867 7874 _numObjectsAlreadyFree = 0;
7868 7875 _numWordsAlreadyFree = 0;
7869 7876 _last_fc = NULL;
7870 7877
7871 7878 _sp->initializeIndexedFreeListArrayReturnedBytes();
7872 7879 _sp->dictionary()->initializeDictReturnedBytes();
7873 7880 )
7874 7881 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7875 7882 "sweep _limit out of bounds");
7876 7883 if (CMSTraceSweeper) {
7877 7884 gclog_or_tty->print("\n====================\nStarting new sweep\n");
7878 7885 }
7879 7886 }
7880 7887
7881 7888 // We need this destructor to reclaim any space at the end
7882 7889 // of the space, which do_blk below may not yet have added back to
7883 7890 // the free lists.
7884 7891 SweepClosure::~SweepClosure() {
7885 7892 assert_lock_strong(_freelistLock);
7886 7893 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7887 7894 "sweep _limit out of bounds");
7888 7895 // Flush any remaining coterminal free run as a single
7889 7896 // coalesced chunk to the appropriate free list.
7890 7897 if (inFreeRange()) {
7891 7898 assert(freeFinger() < _limit, "freeFinger points too high");
7892 7899 flush_cur_free_chunk(freeFinger(), pointer_delta(_limit, freeFinger()));
7893 7900 if (CMSTraceSweeper) {
7894 7901 gclog_or_tty->print("Sweep: last chunk: ");
7895 7902 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") [coalesced:"SIZE_FORMAT"]\n",
7896 7903 freeFinger(), pointer_delta(_limit, freeFinger()), lastFreeRangeCoalesced());
7897 7904 }
7898 7905 } // else nothing to flush
7899 7906 NOT_PRODUCT(
7900 7907 if (Verbose && PrintGC) {
7901 7908 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, "
7902 7909 SIZE_FORMAT " bytes",
7903 7910 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7904 7911 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
7905 7912 SIZE_FORMAT" bytes "
7906 7913 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
7907 7914 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7908 7915 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7909 7916 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) *
7910 7917 sizeof(HeapWord);
7911 7918 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
7912 7919
7913 7920 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7914 7921 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7915 7922 size_t dictReturnedBytes = _sp->dictionary()->sumDictReturnedBytes();
7916 7923 size_t returnedBytes = indexListReturnedBytes + dictReturnedBytes;
7917 7924 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returnedBytes);
7918 7925 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
7919 7926 indexListReturnedBytes);
7920 7927 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
7921 7928 dictReturnedBytes);
7922 7929 }
7923 7930 }
7924 7931 )
7925 7932 // Now, in debug mode, just null out the sweep_limit
7926 7933 NOT_PRODUCT(_sp->clear_sweep_limit();)
7927 7934 if (CMSTraceSweeper) {
7928 7935 gclog_or_tty->print("end of sweep\n================\n");
7929 7936 }
7930 7937 }
7931 7938
7932 7939 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7933 7940 bool freeRangeInFreeLists) {
7934 7941 if (CMSTraceSweeper) {
7935 7942 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n",
7936 7943 freeFinger, freeRangeInFreeLists);
7937 7944 }
7938 7945 assert(!inFreeRange(), "Trampling existing free range");
7939 7946 set_inFreeRange(true);
7940 7947 set_lastFreeRangeCoalesced(false);
7941 7948
7942 7949 set_freeFinger(freeFinger);
7943 7950 set_freeRangeInFreeLists(freeRangeInFreeLists);
7944 7951 if (CMSTestInFreeList) {
7945 7952 if (freeRangeInFreeLists) {
7946 7953 FreeChunk* fc = (FreeChunk*) freeFinger;
7947 7954 assert(fc->isFree(), "A chunk on the free list should be free.");
7948 7955 assert(fc->size() > 0, "Free range should have a size");
7949 7956 assert(_sp->verifyChunkInFreeLists(fc), "Chunk is not in free lists");
7950 7957 }
7951 7958 }
7952 7959 }
7953 7960
7954 7961 // Note that the sweeper runs concurrently with mutators. Thus,
7955 7962 // it is possible for direct allocation in this generation to happen
7956 7963 // in the middle of the sweep. Note that the sweeper also coalesces
7957 7964 // contiguous free blocks. Thus, unless the sweeper and the allocator
7958 7965 // synchronize appropriately freshly allocated blocks may get swept up.
7959 7966 // This is accomplished by the sweeper locking the free lists while
7960 7967 // it is sweeping. Thus blocks that are determined to be free are
7961 7968 // indeed free. There is however one additional complication:
7962 7969 // blocks that have been allocated since the final checkpoint and
7963 7970 // mark, will not have been marked and so would be treated as
7964 7971 // unreachable and swept up. To prevent this, the allocator marks
7965 7972 // the bit map when allocating during the sweep phase. This leads,
7966 7973 // however, to a further complication -- objects may have been allocated
7967 7974 // but not yet initialized -- in the sense that the header isn't yet
7968 7975 // installed. The sweeper can not then determine the size of the block
7969 7976 // in order to skip over it. To deal with this case, we use a technique
7970 7977 // (due to Printezis) to encode such uninitialized block sizes in the
7971 7978 // bit map. Since the bit map uses a bit per every HeapWord, but the
7972 7979 // CMS generation has a minimum object size of 3 HeapWords, it follows
7973 7980 // that "normal marks" won't be adjacent in the bit map (there will
7974 7981 // always be at least two 0 bits between successive 1 bits). We make use
7975 7982 // of these "unused" bits to represent uninitialized blocks -- the bit
7976 7983 // corresponding to the start of the uninitialized object and the next
7977 7984 // bit are both set. Finally, a 1 bit marks the end of the object that
7978 7985 // started with the two consecutive 1 bits to indicate its potentially
7979 7986 // uninitialized state.
7980 7987
7981 7988 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7982 7989 FreeChunk* fc = (FreeChunk*)addr;
7983 7990 size_t res;
7984 7991
7985 7992 // Check if we are done sweeping. Below we check "addr >= _limit" rather
7986 7993 // than "addr == _limit" because although _limit was a block boundary when
7987 7994 // we started the sweep, it may no longer be one because heap expansion
7988 7995 // may have caused us to coalesce the block ending at the address _limit
7989 7996 // with a newly expanded chunk (this happens when _limit was set to the
7990 7997 // previous _end of the space), so we may have stepped past _limit; see CR 6977970.
7991 7998 if (addr >= _limit) { // we have swept up to or past the limit: finish up
7992 7999 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7993 8000 "sweep _limit out of bounds");
7994 8001 assert(addr < _sp->end(), "addr out of bounds");
7995 8002 // Flush any remaining coterminal free run as a single
7996 8003 // coalesced chunk to the appropriate free list.
7997 8004 if (inFreeRange()) {
7998 8005 assert(freeFinger() < _limit, "finger points too high");
7999 8006 flush_cur_free_chunk(freeFinger(),
8000 8007 pointer_delta(addr, freeFinger()));
8001 8008 if (CMSTraceSweeper) {
8002 8009 gclog_or_tty->print("Sweep: last chunk: ");
8003 8010 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") "
8004 8011 "[coalesced:"SIZE_FORMAT"]\n",
8005 8012 freeFinger(), pointer_delta(addr, freeFinger()),
8006 8013 lastFreeRangeCoalesced());
8007 8014 }
8008 8015 }
8009 8016
8010 8017 // help the iterator loop finish
8011 8018 return pointer_delta(_sp->end(), addr);
8012 8019 }
8013 8020
8014 8021 assert(addr < _limit, "sweep invariant");
8015 8022 // check if we should yield
8016 8023 do_yield_check(addr);
8017 8024 if (fc->isFree()) {
8018 8025 // Chunk that is already free
8019 8026 res = fc->size();
8020 8027 do_already_free_chunk(fc);
8021 8028 debug_only(_sp->verifyFreeLists());
8022 8029 assert(res == fc->size(), "Don't expect the size to change");
8023 8030 NOT_PRODUCT(
8024 8031 _numObjectsAlreadyFree++;
8025 8032 _numWordsAlreadyFree += res;
8026 8033 )
8027 8034 NOT_PRODUCT(_last_fc = fc;)
8028 8035 } else if (!_bitMap->isMarked(addr)) {
8029 8036 // Chunk is fresh garbage
8030 8037 res = do_garbage_chunk(fc);
8031 8038 debug_only(_sp->verifyFreeLists());
8032 8039 NOT_PRODUCT(
8033 8040 _numObjectsFreed++;
8034 8041 _numWordsFreed += res;
8035 8042 )
8036 8043 } else {
8037 8044 // Chunk that is alive.
8038 8045 res = do_live_chunk(fc);
8039 8046 debug_only(_sp->verifyFreeLists());
8040 8047 NOT_PRODUCT(
8041 8048 _numObjectsLive++;
8042 8049 _numWordsLive += res;
8043 8050 )
8044 8051 }
8045 8052 return res;
8046 8053 }
8047 8054
8048 8055 // For the smart allocation, record following
8049 8056 // split deaths - a free chunk is removed from its free list because
8050 8057 // it is being split into two or more chunks.
8051 8058 // split birth - a free chunk is being added to its free list because
8052 8059 // a larger free chunk has been split and resulted in this free chunk.
8053 8060 // coal death - a free chunk is being removed from its free list because
8054 8061 // it is being coalesced into a large free chunk.
8055 8062 // coal birth - a free chunk is being added to its free list because
8056 8063 // it was created when two or more free chunks where coalesced into
8057 8064 // this free chunk.
8058 8065 //
8059 8066 // These statistics are used to determine the desired number of free
8060 8067 // chunks of a given size. The desired number is chosen to be relative
8061 8068 // to the end of a CMS sweep. The desired number at the end of a sweep
8062 8069 // is the
8063 8070 // count-at-end-of-previous-sweep (an amount that was enough)
8064 8071 // - count-at-beginning-of-current-sweep (the excess)
8065 8072 // + split-births (gains in this size during interval)
8066 8073 // - split-deaths (demands on this size during interval)
8067 8074 // where the interval is from the end of one sweep to the end of the
8068 8075 // next.
8069 8076 //
8070 8077 // When sweeping the sweeper maintains an accumulated chunk which is
8071 8078 // the chunk that is made up of chunks that have been coalesced. That
8072 8079 // will be termed the left-hand chunk. A new chunk of garbage that
8073 8080 // is being considered for coalescing will be referred to as the
8074 8081 // right-hand chunk.
8075 8082 //
8076 8083 // When making a decision on whether to coalesce a right-hand chunk with
8077 8084 // the current left-hand chunk, the current count vs. the desired count
8078 8085 // of the left-hand chunk is considered. Also if the right-hand chunk
8079 8086 // is near the large chunk at the end of the heap (see
8080 8087 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
8081 8088 // left-hand chunk is coalesced.
8082 8089 //
8083 8090 // When making a decision about whether to split a chunk, the desired count
8084 8091 // vs. the current count of the candidate to be split is also considered.
8085 8092 // If the candidate is underpopulated (currently fewer chunks than desired)
8086 8093 // a chunk of an overpopulated (currently more chunks than desired) size may
8087 8094 // be chosen. The "hint" associated with a free list, if non-null, points
8088 8095 // to a free list which may be overpopulated.
8089 8096 //
8090 8097
8091 8098 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
8092 8099 size_t size = fc->size();
8093 8100 // Chunks that cannot be coalesced are not in the
8094 8101 // free lists.
8095 8102 if (CMSTestInFreeList && !fc->cantCoalesce()) {
8096 8103 assert(_sp->verifyChunkInFreeLists(fc),
8097 8104 "free chunk should be in free lists");
8098 8105 }
8099 8106 // a chunk that is already free, should not have been
8100 8107 // marked in the bit map
8101 8108 HeapWord* addr = (HeapWord*) fc;
8102 8109 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
8103 8110 // Verify that the bit map has no bits marked between
8104 8111 // addr and purported end of this block.
8105 8112 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8106 8113
8107 8114 // Some chunks cannot be coalesced under any circumstances.
8108 8115 // See the definition of cantCoalesce().
8109 8116 if (!fc->cantCoalesce()) {
8110 8117 // This chunk can potentially be coalesced.
8111 8118 if (_sp->adaptive_freelists()) {
8112 8119 // All the work is done in
8113 8120 do_post_free_or_garbage_chunk(fc, size);
8114 8121 } else { // Not adaptive free lists
8115 8122 // this is a free chunk that can potentially be coalesced by the sweeper;
8116 8123 if (!inFreeRange()) {
8117 8124 // if the next chunk is a free block that can't be coalesced
8118 8125 // it doesn't make sense to remove this chunk from the free lists
8119 8126 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
8120 8127 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
8121 8128 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
8122 8129 nextChunk->isFree() && // ... which is free...
8123 8130 nextChunk->cantCoalesce()) { // ... but can't be coalesced
8124 8131 // nothing to do
8125 8132 } else {
8126 8133 // Potentially the start of a new free range:
8127 8134 // Don't eagerly remove it from the free lists.
8128 8135 // No need to remove it if it will just be put
8129 8136 // back again. (Also from a pragmatic point of view
8130 8137 // if it is a free block in a region that is beyond
8131 8138 // any allocated blocks, an assertion will fail)
8132 8139 // Remember the start of a free run.
8133 8140 initialize_free_range(addr, true);
8134 8141 // end - can coalesce with next chunk
8135 8142 }
8136 8143 } else {
8137 8144 // the midst of a free range, we are coalescing
8138 8145 debug_only(record_free_block_coalesced(fc);)
8139 8146 if (CMSTraceSweeper) {
8140 8147 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size);
8141 8148 }
8142 8149 // remove it from the free lists
8143 8150 _sp->removeFreeChunkFromFreeLists(fc);
8144 8151 set_lastFreeRangeCoalesced(true);
8145 8152 // If the chunk is being coalesced and the current free range is
8146 8153 // in the free lists, remove the current free range so that it
8147 8154 // will be returned to the free lists in its entirety - all
8148 8155 // the coalesced pieces included.
8149 8156 if (freeRangeInFreeLists()) {
8150 8157 FreeChunk* ffc = (FreeChunk*) freeFinger();
8151 8158 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8152 8159 "Size of free range is inconsistent with chunk size.");
8153 8160 if (CMSTestInFreeList) {
8154 8161 assert(_sp->verifyChunkInFreeLists(ffc),
8155 8162 "free range is not in free lists");
8156 8163 }
8157 8164 _sp->removeFreeChunkFromFreeLists(ffc);
8158 8165 set_freeRangeInFreeLists(false);
8159 8166 }
8160 8167 }
8161 8168 }
8162 8169 } else {
8163 8170 // Code path common to both original and adaptive free lists.
8164 8171
8165 8172 // cant coalesce with previous block; this should be treated
8166 8173 // as the end of a free run if any
8167 8174 if (inFreeRange()) {
8168 8175 // we kicked some butt; time to pick up the garbage
8169 8176 assert(freeFinger() < addr, "freeFinger points too high");
8170 8177 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8171 8178 }
8172 8179 // else, nothing to do, just continue
8173 8180 }
8174 8181 }
8175 8182
8176 8183 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
8177 8184 // This is a chunk of garbage. It is not in any free list.
8178 8185 // Add it to a free list or let it possibly be coalesced into
8179 8186 // a larger chunk.
8180 8187 HeapWord* addr = (HeapWord*) fc;
8181 8188 size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8182 8189
8183 8190 if (_sp->adaptive_freelists()) {
8184 8191 // Verify that the bit map has no bits marked between
8185 8192 // addr and purported end of just dead object.
8186 8193 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8187 8194
8188 8195 do_post_free_or_garbage_chunk(fc, size);
8189 8196 } else {
8190 8197 if (!inFreeRange()) {
8191 8198 // start of a new free range
8192 8199 assert(size > 0, "A free range should have a size");
8193 8200 initialize_free_range(addr, false);
8194 8201
8195 8202 } else {
8196 8203 // this will be swept up when we hit the end of the
8197 8204 // free range
8198 8205 if (CMSTraceSweeper) {
8199 8206 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size);
8200 8207 }
8201 8208 // If the chunk is being coalesced and the current free range is
8202 8209 // in the free lists, remove the current free range so that it
8203 8210 // will be returned to the free lists in its entirety - all
8204 8211 // the coalesced pieces included.
8205 8212 if (freeRangeInFreeLists()) {
8206 8213 FreeChunk* ffc = (FreeChunk*)freeFinger();
8207 8214 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8208 8215 "Size of free range is inconsistent with chunk size.");
8209 8216 if (CMSTestInFreeList) {
8210 8217 assert(_sp->verifyChunkInFreeLists(ffc),
8211 8218 "free range is not in free lists");
8212 8219 }
8213 8220 _sp->removeFreeChunkFromFreeLists(ffc);
8214 8221 set_freeRangeInFreeLists(false);
8215 8222 }
8216 8223 set_lastFreeRangeCoalesced(true);
8217 8224 }
8218 8225 // this will be swept up when we hit the end of the free range
8219 8226
8220 8227 // Verify that the bit map has no bits marked between
8221 8228 // addr and purported end of just dead object.
8222 8229 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8223 8230 }
8224 8231 return size;
8225 8232 }
8226 8233
8227 8234 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
8228 8235 HeapWord* addr = (HeapWord*) fc;
8229 8236 // The sweeper has just found a live object. Return any accumulated
8230 8237 // left hand chunk to the free lists.
8231 8238 if (inFreeRange()) {
8232 8239 assert(freeFinger() < addr, "freeFinger points too high");
8233 8240 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8234 8241 }
8235 8242
8236 8243 // This object is live: we'd normally expect this to be
8237 8244 // an oop, and like to assert the following:
8238 8245 // assert(oop(addr)->is_oop(), "live block should be an oop");
8239 8246 // However, as we commented above, this may be an object whose
8240 8247 // header hasn't yet been initialized.
8241 8248 size_t size;
8242 8249 assert(_bitMap->isMarked(addr), "Tautology for this control point");
8243 8250 if (_bitMap->isMarked(addr + 1)) {
8244 8251 // Determine the size from the bit map, rather than trying to
8245 8252 // compute it from the object header.
8246 8253 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
8247 8254 size = pointer_delta(nextOneAddr + 1, addr);
8248 8255 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
8249 8256 "alignment problem");
8250 8257
8251 8258 #ifdef DEBUG
8252 8259 if (oop(addr)->klass_or_null() != NULL &&
8253 8260 ( !_collector->should_unload_classes()
8254 8261 || (oop(addr)->is_parsable()) &&
8255 8262 oop(addr)->is_conc_safe())) {
8256 8263 // Ignore mark word because we are running concurrent with mutators
8257 8264 assert(oop(addr)->is_oop(true), "live block should be an oop");
8258 8265 // is_conc_safe is checked before performing this assertion
8259 8266 // because an object that is not is_conc_safe may yet have
8260 8267 // the return from size() correct.
8261 8268 assert(size ==
8262 8269 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
8263 8270 "P-mark and computed size do not agree");
8264 8271 }
8265 8272 #endif
8266 8273
8267 8274 } else {
8268 8275 // This should be an initialized object that's alive.
8269 8276 assert(oop(addr)->klass_or_null() != NULL &&
8270 8277 (!_collector->should_unload_classes()
8271 8278 || oop(addr)->is_parsable()),
8272 8279 "Should be an initialized object");
8273 8280 // Note that there are objects used during class redefinition
8274 8281 // (e.g., merge_cp in VM_RedefineClasses::merge_cp_and_rewrite()
8275 8282 // which are discarded with their is_conc_safe state still
8276 8283 // false. These object may be floating garbage so may be
8277 8284 // seen here. If they are floating garbage their size
8278 8285 // should be attainable from their klass. Do not that
8279 8286 // is_conc_safe() is true for oop(addr).
8280 8287 // Ignore mark word because we are running concurrent with mutators
8281 8288 assert(oop(addr)->is_oop(true), "live block should be an oop");
8282 8289 // Verify that the bit map has no bits marked between
8283 8290 // addr and purported end of this block.
8284 8291 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8285 8292 assert(size >= 3, "Necessary for Printezis marks to work");
8286 8293 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
8287 8294 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
8288 8295 }
8289 8296 return size;
8290 8297 }
8291 8298
8292 8299 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
8293 8300 size_t chunkSize) {
8294 8301 // do_post_free_or_garbage_chunk() should only be called in the case
8295 8302 // of the adaptive free list allocator.
8296 8303 bool fcInFreeLists = fc->isFree();
8297 8304 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
8298 8305 assert((HeapWord*)fc <= _limit, "sweep invariant");
8299 8306 if (CMSTestInFreeList && fcInFreeLists) {
8300 8307 assert(_sp->verifyChunkInFreeLists(fc), "free chunk is not in free lists");
8301 8308 }
8302 8309
8303 8310 if (CMSTraceSweeper) {
8304 8311 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
8305 8312 }
8306 8313
8307 8314 HeapWord* addr = (HeapWord*) fc;
8308 8315
8309 8316 bool coalesce;
8310 8317 size_t left = pointer_delta(addr, freeFinger());
8311 8318 size_t right = chunkSize;
8312 8319 switch (FLSCoalescePolicy) {
8313 8320 // numeric value forms a coalition aggressiveness metric
8314 8321 case 0: { // never coalesce
8315 8322 coalesce = false;
8316 8323 break;
8317 8324 }
8318 8325 case 1: { // coalesce if left & right chunks on overpopulated lists
8319 8326 coalesce = _sp->coalOverPopulated(left) &&
8320 8327 _sp->coalOverPopulated(right);
8321 8328 break;
8322 8329 }
8323 8330 case 2: { // coalesce if left chunk on overpopulated list (default)
8324 8331 coalesce = _sp->coalOverPopulated(left);
8325 8332 break;
8326 8333 }
8327 8334 case 3: { // coalesce if left OR right chunk on overpopulated list
8328 8335 coalesce = _sp->coalOverPopulated(left) ||
8329 8336 _sp->coalOverPopulated(right);
8330 8337 break;
8331 8338 }
8332 8339 case 4: { // always coalesce
8333 8340 coalesce = true;
8334 8341 break;
8335 8342 }
8336 8343 default:
8337 8344 ShouldNotReachHere();
8338 8345 }
8339 8346
8340 8347 // Should the current free range be coalesced?
8341 8348 // If the chunk is in a free range and either we decided to coalesce above
8342 8349 // or the chunk is near the large block at the end of the heap
8343 8350 // (isNearLargestChunk() returns true), then coalesce this chunk.
8344 8351 bool doCoalesce = inFreeRange() &&
8345 8352 (coalesce || _g->isNearLargestChunk((HeapWord*)fc));
8346 8353 if (doCoalesce) {
8347 8354 // Coalesce the current free range on the left with the new
8348 8355 // chunk on the right. If either is on a free list,
8349 8356 // it must be removed from the list and stashed in the closure.
8350 8357 if (freeRangeInFreeLists()) {
8351 8358 FreeChunk* ffc = (FreeChunk*)freeFinger();
8352 8359 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8353 8360 "Size of free range is inconsistent with chunk size.");
8354 8361 if (CMSTestInFreeList) {
8355 8362 assert(_sp->verifyChunkInFreeLists(ffc),
8356 8363 "Chunk is not in free lists");
8357 8364 }
8358 8365 _sp->coalDeath(ffc->size());
8359 8366 _sp->removeFreeChunkFromFreeLists(ffc);
8360 8367 set_freeRangeInFreeLists(false);
8361 8368 }
8362 8369 if (fcInFreeLists) {
8363 8370 _sp->coalDeath(chunkSize);
8364 8371 assert(fc->size() == chunkSize,
8365 8372 "The chunk has the wrong size or is not in the free lists");
8366 8373 _sp->removeFreeChunkFromFreeLists(fc);
8367 8374 }
8368 8375 set_lastFreeRangeCoalesced(true);
8369 8376 } else { // not in a free range and/or should not coalesce
8370 8377 // Return the current free range and start a new one.
8371 8378 if (inFreeRange()) {
8372 8379 // In a free range but cannot coalesce with the right hand chunk.
8373 8380 // Put the current free range into the free lists.
8374 8381 flush_cur_free_chunk(freeFinger(),
8375 8382 pointer_delta(addr, freeFinger()));
8376 8383 }
8377 8384 // Set up for new free range. Pass along whether the right hand
8378 8385 // chunk is in the free lists.
8379 8386 initialize_free_range((HeapWord*)fc, fcInFreeLists);
8380 8387 }
8381 8388 }
8382 8389
8383 8390 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
8384 8391 assert(inFreeRange(), "Should only be called if currently in a free range.");
8385 8392 assert(size > 0,
8386 8393 "A zero sized chunk cannot be added to the free lists.");
8387 8394 if (!freeRangeInFreeLists()) {
8388 8395 if (CMSTestInFreeList) {
8389 8396 FreeChunk* fc = (FreeChunk*) chunk;
8390 8397 fc->setSize(size);
8391 8398 assert(!_sp->verifyChunkInFreeLists(fc),
8392 8399 "chunk should not be in free lists yet");
8393 8400 }
8394 8401 if (CMSTraceSweeper) {
8395 8402 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8396 8403 chunk, size);
8397 8404 }
8398 8405 // A new free range is going to be starting. The current
8399 8406 // free range has not been added to the free lists yet or
8400 8407 // was removed so add it back.
8401 8408 // If the current free range was coalesced, then the death
8402 8409 // of the free range was recorded. Record a birth now.
8403 8410 if (lastFreeRangeCoalesced()) {
8404 8411 _sp->coalBirth(size);
8405 8412 }
8406 8413 _sp->addChunkAndRepairOffsetTable(chunk, size,
8407 8414 lastFreeRangeCoalesced());
8408 8415 }
8409 8416 set_inFreeRange(false);
8410 8417 set_freeRangeInFreeLists(false);
8411 8418 }
8412 8419
8413 8420 // We take a break if we've been at this for a while,
8414 8421 // so as to avoid monopolizing the locks involved.
8415 8422 void SweepClosure::do_yield_work(HeapWord* addr) {
8416 8423 // Return current free chunk being used for coalescing (if any)
8417 8424 // to the appropriate freelist. After yielding, the next
8418 8425 // free block encountered will start a coalescing range of
8419 8426 // free blocks. If the next free block is adjacent to the
8420 8427 // chunk just flushed, they will need to wait for the next
8421 8428 // sweep to be coalesced.
8422 8429 if (inFreeRange()) {
8423 8430 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8424 8431 }
8425 8432
8426 8433 // First give up the locks, then yield, then re-lock.
8427 8434 // We should probably use a constructor/destructor idiom to
8428 8435 // do this unlock/lock or modify the MutexUnlocker class to
8429 8436 // serve our purpose. XXX
8430 8437 assert_lock_strong(_bitMap->lock());
8431 8438 assert_lock_strong(_freelistLock);
8432 8439 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8433 8440 "CMS thread should hold CMS token");
8434 8441 _bitMap->lock()->unlock();
8435 8442 _freelistLock->unlock();
8436 8443 ConcurrentMarkSweepThread::desynchronize(true);
8437 8444 ConcurrentMarkSweepThread::acknowledge_yield_request();
8438 8445 _collector->stopTimer();
8439 8446 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8440 8447 if (PrintCMSStatistics != 0) {
8441 8448 _collector->incrementYields();
8442 8449 }
8443 8450 _collector->icms_wait();
8444 8451
8445 8452 // See the comment in coordinator_yield()
8446 8453 for (unsigned i = 0; i < CMSYieldSleepCount &&
8447 8454 ConcurrentMarkSweepThread::should_yield() &&
8448 8455 !CMSCollector::foregroundGCIsActive(); ++i) {
8449 8456 os::sleep(Thread::current(), 1, false);
8450 8457 ConcurrentMarkSweepThread::acknowledge_yield_request();
8451 8458 }
8452 8459
8453 8460 ConcurrentMarkSweepThread::synchronize(true);
8454 8461 _freelistLock->lock();
8455 8462 _bitMap->lock()->lock_without_safepoint_check();
8456 8463 _collector->startTimer();
8457 8464 }
8458 8465
8459 8466 #ifndef PRODUCT
8460 8467 // This is actually very useful in a product build if it can
8461 8468 // be called from the debugger. Compile it into the product
8462 8469 // as needed.
8463 8470 bool debug_verifyChunkInFreeLists(FreeChunk* fc) {
8464 8471 return debug_cms_space->verifyChunkInFreeLists(fc);
8465 8472 }
8466 8473
8467 8474 void SweepClosure::record_free_block_coalesced(FreeChunk* fc) const {
8468 8475 if (CMSTraceSweeper) {
8469 8476 gclog_or_tty->print("Sweep:coal_free_blk 0x%x (%d)\n", fc, fc->size());
8470 8477 }
8471 8478 }
8472 8479 #endif
8473 8480
8474 8481 // CMSIsAliveClosure
8475 8482 bool CMSIsAliveClosure::do_object_b(oop obj) {
8476 8483 HeapWord* addr = (HeapWord*)obj;
8477 8484 return addr != NULL &&
8478 8485 (!_span.contains(addr) || _bit_map->isMarked(addr));
8479 8486 }
8480 8487
8481 8488 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
8482 8489 MemRegion span,
8483 8490 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
8484 8491 CMSMarkStack* revisit_stack, bool cpc):
8485 8492 KlassRememberingOopClosure(collector, NULL, revisit_stack),
8486 8493 _span(span),
8487 8494 _bit_map(bit_map),
8488 8495 _mark_stack(mark_stack),
8489 8496 _concurrent_precleaning(cpc) {
8490 8497 assert(!_span.is_empty(), "Empty span could spell trouble");
8491 8498 }
8492 8499
8493 8500
8494 8501 // CMSKeepAliveClosure: the serial version
8495 8502 void CMSKeepAliveClosure::do_oop(oop obj) {
8496 8503 HeapWord* addr = (HeapWord*)obj;
8497 8504 if (_span.contains(addr) &&
8498 8505 !_bit_map->isMarked(addr)) {
8499 8506 _bit_map->mark(addr);
8500 8507 bool simulate_overflow = false;
8501 8508 NOT_PRODUCT(
8502 8509 if (CMSMarkStackOverflowALot &&
8503 8510 _collector->simulate_overflow()) {
8504 8511 // simulate a stack overflow
8505 8512 simulate_overflow = true;
8506 8513 }
8507 8514 )
8508 8515 if (simulate_overflow || !_mark_stack->push(obj)) {
8509 8516 if (_concurrent_precleaning) {
8510 8517 // We dirty the overflown object and let the remark
8511 8518 // phase deal with it.
8512 8519 assert(_collector->overflow_list_is_empty(), "Error");
8513 8520 // In the case of object arrays, we need to dirty all of
8514 8521 // the cards that the object spans. No locking or atomics
8515 8522 // are needed since no one else can be mutating the mod union
8516 8523 // table.
8517 8524 if (obj->is_objArray()) {
8518 8525 size_t sz = obj->size();
8519 8526 HeapWord* end_card_addr =
8520 8527 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
8521 8528 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8522 8529 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8523 8530 _collector->_modUnionTable.mark_range(redirty_range);
8524 8531 } else {
8525 8532 _collector->_modUnionTable.mark(addr);
8526 8533 }
8527 8534 _collector->_ser_kac_preclean_ovflw++;
8528 8535 } else {
8529 8536 _collector->push_on_overflow_list(obj);
8530 8537 _collector->_ser_kac_ovflw++;
8531 8538 }
8532 8539 }
8533 8540 }
8534 8541 }
8535 8542
8536 8543 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8537 8544 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8538 8545
8539 8546 // CMSParKeepAliveClosure: a parallel version of the above.
8540 8547 // The work queues are private to each closure (thread),
8541 8548 // but (may be) available for stealing by other threads.
8542 8549 void CMSParKeepAliveClosure::do_oop(oop obj) {
8543 8550 HeapWord* addr = (HeapWord*)obj;
8544 8551 if (_span.contains(addr) &&
8545 8552 !_bit_map->isMarked(addr)) {
8546 8553 // In general, during recursive tracing, several threads
8547 8554 // may be concurrently getting here; the first one to
8548 8555 // "tag" it, claims it.
8549 8556 if (_bit_map->par_mark(addr)) {
8550 8557 bool res = _work_queue->push(obj);
8551 8558 assert(res, "Low water mark should be much less than capacity");
8552 8559 // Do a recursive trim in the hope that this will keep
8553 8560 // stack usage lower, but leave some oops for potential stealers
8554 8561 trim_queue(_low_water_mark);
8555 8562 } // Else, another thread got there first
8556 8563 }
8557 8564 }
8558 8565
8559 8566 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8560 8567 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8561 8568
8562 8569 void CMSParKeepAliveClosure::trim_queue(uint max) {
8563 8570 while (_work_queue->size() > max) {
8564 8571 oop new_oop;
8565 8572 if (_work_queue->pop_local(new_oop)) {
8566 8573 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8567 8574 assert(_bit_map->isMarked((HeapWord*)new_oop),
8568 8575 "no white objects on this stack!");
8569 8576 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8570 8577 // iterate over the oops in this oop, marking and pushing
8571 8578 // the ones in CMS heap (i.e. in _span).
8572 8579 new_oop->oop_iterate(&_mark_and_push);
8573 8580 }
8574 8581 }
8575 8582 }
8576 8583
8577 8584 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8578 8585 CMSCollector* collector,
8579 8586 MemRegion span, CMSBitMap* bit_map,
8580 8587 CMSMarkStack* revisit_stack,
8581 8588 OopTaskQueue* work_queue):
8582 8589 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
8583 8590 _span(span),
8584 8591 _bit_map(bit_map),
8585 8592 _work_queue(work_queue) { }
8586 8593
8587 8594 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8588 8595 HeapWord* addr = (HeapWord*)obj;
8589 8596 if (_span.contains(addr) &&
8590 8597 !_bit_map->isMarked(addr)) {
8591 8598 if (_bit_map->par_mark(addr)) {
8592 8599 bool simulate_overflow = false;
8593 8600 NOT_PRODUCT(
8594 8601 if (CMSMarkStackOverflowALot &&
8595 8602 _collector->par_simulate_overflow()) {
8596 8603 // simulate a stack overflow
8597 8604 simulate_overflow = true;
8598 8605 }
8599 8606 )
8600 8607 if (simulate_overflow || !_work_queue->push(obj)) {
8601 8608 _collector->par_push_on_overflow_list(obj);
8602 8609 _collector->_par_kac_ovflw++;
8603 8610 }
8604 8611 } // Else another thread got there already
8605 8612 }
8606 8613 }
8607 8614
8608 8615 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8609 8616 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8610 8617
8611 8618 //////////////////////////////////////////////////////////////////
8612 8619 // CMSExpansionCause /////////////////////////////
8613 8620 //////////////////////////////////////////////////////////////////
8614 8621 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8615 8622 switch (cause) {
8616 8623 case _no_expansion:
8617 8624 return "No expansion";
8618 8625 case _satisfy_free_ratio:
8619 8626 return "Free ratio";
8620 8627 case _satisfy_promotion:
8621 8628 return "Satisfy promotion";
8622 8629 case _satisfy_allocation:
8623 8630 return "allocation";
8624 8631 case _allocate_par_lab:
8625 8632 return "Par LAB";
8626 8633 case _allocate_par_spooling_space:
8627 8634 return "Par Spooling Space";
8628 8635 case _adaptive_size_policy:
8629 8636 return "Ergonomics";
8630 8637 default:
8631 8638 return "unknown";
8632 8639 }
8633 8640 }
8634 8641
8635 8642 void CMSDrainMarkingStackClosure::do_void() {
8636 8643 // the max number to take from overflow list at a time
8637 8644 const size_t num = _mark_stack->capacity()/4;
8638 8645 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8639 8646 "Overflow list should be NULL during concurrent phases");
8640 8647 while (!_mark_stack->isEmpty() ||
8641 8648 // if stack is empty, check the overflow list
8642 8649 _collector->take_from_overflow_list(num, _mark_stack)) {
8643 8650 oop obj = _mark_stack->pop();
8644 8651 HeapWord* addr = (HeapWord*)obj;
8645 8652 assert(_span.contains(addr), "Should be within span");
8646 8653 assert(_bit_map->isMarked(addr), "Should be marked");
8647 8654 assert(obj->is_oop(), "Should be an oop");
8648 8655 obj->oop_iterate(_keep_alive);
8649 8656 }
8650 8657 }
8651 8658
8652 8659 void CMSParDrainMarkingStackClosure::do_void() {
8653 8660 // drain queue
8654 8661 trim_queue(0);
8655 8662 }
8656 8663
8657 8664 // Trim our work_queue so its length is below max at return
8658 8665 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8659 8666 while (_work_queue->size() > max) {
8660 8667 oop new_oop;
8661 8668 if (_work_queue->pop_local(new_oop)) {
8662 8669 assert(new_oop->is_oop(), "Expected an oop");
8663 8670 assert(_bit_map->isMarked((HeapWord*)new_oop),
8664 8671 "no white objects on this stack!");
8665 8672 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8666 8673 // iterate over the oops in this oop, marking and pushing
8667 8674 // the ones in CMS heap (i.e. in _span).
8668 8675 new_oop->oop_iterate(&_mark_and_push);
8669 8676 }
8670 8677 }
8671 8678 }
8672 8679
8673 8680 ////////////////////////////////////////////////////////////////////
8674 8681 // Support for Marking Stack Overflow list handling and related code
8675 8682 ////////////////////////////////////////////////////////////////////
8676 8683 // Much of the following code is similar in shape and spirit to the
8677 8684 // code used in ParNewGC. We should try and share that code
8678 8685 // as much as possible in the future.
8679 8686
8680 8687 #ifndef PRODUCT
8681 8688 // Debugging support for CMSStackOverflowALot
8682 8689
8683 8690 // It's OK to call this multi-threaded; the worst thing
8684 8691 // that can happen is that we'll get a bunch of closely
8685 8692 // spaced simulated oveflows, but that's OK, in fact
8686 8693 // probably good as it would exercise the overflow code
8687 8694 // under contention.
8688 8695 bool CMSCollector::simulate_overflow() {
8689 8696 if (_overflow_counter-- <= 0) { // just being defensive
8690 8697 _overflow_counter = CMSMarkStackOverflowInterval;
8691 8698 return true;
8692 8699 } else {
8693 8700 return false;
8694 8701 }
8695 8702 }
8696 8703
8697 8704 bool CMSCollector::par_simulate_overflow() {
8698 8705 return simulate_overflow();
8699 8706 }
8700 8707 #endif
8701 8708
8702 8709 // Single-threaded
8703 8710 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8704 8711 assert(stack->isEmpty(), "Expected precondition");
8705 8712 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8706 8713 size_t i = num;
8707 8714 oop cur = _overflow_list;
8708 8715 const markOop proto = markOopDesc::prototype();
8709 8716 NOT_PRODUCT(ssize_t n = 0;)
8710 8717 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8711 8718 next = oop(cur->mark());
8712 8719 cur->set_mark(proto); // until proven otherwise
8713 8720 assert(cur->is_oop(), "Should be an oop");
8714 8721 bool res = stack->push(cur);
8715 8722 assert(res, "Bit off more than can chew?");
8716 8723 NOT_PRODUCT(n++;)
8717 8724 }
8718 8725 _overflow_list = cur;
8719 8726 #ifndef PRODUCT
8720 8727 assert(_num_par_pushes >= n, "Too many pops?");
8721 8728 _num_par_pushes -=n;
8722 8729 #endif
8723 8730 return !stack->isEmpty();
8724 8731 }
8725 8732
8726 8733 #define BUSY (oop(0x1aff1aff))
8727 8734 // (MT-safe) Get a prefix of at most "num" from the list.
8728 8735 // The overflow list is chained through the mark word of
8729 8736 // each object in the list. We fetch the entire list,
8730 8737 // break off a prefix of the right size and return the
8731 8738 // remainder. If other threads try to take objects from
8732 8739 // the overflow list at that time, they will wait for
8733 8740 // some time to see if data becomes available. If (and
8734 8741 // only if) another thread places one or more object(s)
8735 8742 // on the global list before we have returned the suffix
8736 8743 // to the global list, we will walk down our local list
8737 8744 // to find its end and append the global list to
8738 8745 // our suffix before returning it. This suffix walk can
8739 8746 // prove to be expensive (quadratic in the amount of traffic)
8740 8747 // when there are many objects in the overflow list and
8741 8748 // there is much producer-consumer contention on the list.
8742 8749 // *NOTE*: The overflow list manipulation code here and
8743 8750 // in ParNewGeneration:: are very similar in shape,
8744 8751 // except that in the ParNew case we use the old (from/eden)
8745 8752 // copy of the object to thread the list via its klass word.
8746 8753 // Because of the common code, if you make any changes in
8747 8754 // the code below, please check the ParNew version to see if
8748 8755 // similar changes might be needed.
8749 8756 // CR 6797058 has been filed to consolidate the common code.
8750 8757 bool CMSCollector::par_take_from_overflow_list(size_t num,
8751 8758 OopTaskQueue* work_q,
8752 8759 int no_of_gc_threads) {
8753 8760 assert(work_q->size() == 0, "First empty local work queue");
8754 8761 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8755 8762 if (_overflow_list == NULL) {
8756 8763 return false;
8757 8764 }
8758 8765 // Grab the entire list; we'll put back a suffix
8759 8766 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
8760 8767 Thread* tid = Thread::current();
8761 8768 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
8762 8769 // set to ParallelGCThreads.
8763 8770 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
8764 8771 size_t sleep_time_millis = MAX2((size_t)1, num/100);
8765 8772 // If the list is busy, we spin for a short while,
8766 8773 // sleeping between attempts to get the list.
8767 8774 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
8768 8775 os::sleep(tid, sleep_time_millis, false);
8769 8776 if (_overflow_list == NULL) {
8770 8777 // Nothing left to take
8771 8778 return false;
8772 8779 } else if (_overflow_list != BUSY) {
8773 8780 // Try and grab the prefix
8774 8781 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
8775 8782 }
8776 8783 }
8777 8784 // If the list was found to be empty, or we spun long
8778 8785 // enough, we give up and return empty-handed. If we leave
8779 8786 // the list in the BUSY state below, it must be the case that
8780 8787 // some other thread holds the overflow list and will set it
8781 8788 // to a non-BUSY state in the future.
8782 8789 if (prefix == NULL || prefix == BUSY) {
8783 8790 // Nothing to take or waited long enough
8784 8791 if (prefix == NULL) {
8785 8792 // Write back the NULL in case we overwrote it with BUSY above
8786 8793 // and it is still the same value.
8787 8794 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8788 8795 }
8789 8796 return false;
8790 8797 }
8791 8798 assert(prefix != NULL && prefix != BUSY, "Error");
8792 8799 size_t i = num;
8793 8800 oop cur = prefix;
8794 8801 // Walk down the first "num" objects, unless we reach the end.
8795 8802 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8796 8803 if (cur->mark() == NULL) {
8797 8804 // We have "num" or fewer elements in the list, so there
8798 8805 // is nothing to return to the global list.
8799 8806 // Write back the NULL in lieu of the BUSY we wrote
8800 8807 // above, if it is still the same value.
8801 8808 if (_overflow_list == BUSY) {
8802 8809 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8803 8810 }
8804 8811 } else {
8805 8812 // Chop off the suffix and rerturn it to the global list.
8806 8813 assert(cur->mark() != BUSY, "Error");
8807 8814 oop suffix_head = cur->mark(); // suffix will be put back on global list
8808 8815 cur->set_mark(NULL); // break off suffix
8809 8816 // It's possible that the list is still in the empty(busy) state
8810 8817 // we left it in a short while ago; in that case we may be
8811 8818 // able to place back the suffix without incurring the cost
8812 8819 // of a walk down the list.
8813 8820 oop observed_overflow_list = _overflow_list;
8814 8821 oop cur_overflow_list = observed_overflow_list;
8815 8822 bool attached = false;
8816 8823 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
8817 8824 observed_overflow_list =
8818 8825 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8819 8826 if (cur_overflow_list == observed_overflow_list) {
8820 8827 attached = true;
8821 8828 break;
8822 8829 } else cur_overflow_list = observed_overflow_list;
8823 8830 }
8824 8831 if (!attached) {
8825 8832 // Too bad, someone else sneaked in (at least) an element; we'll need
8826 8833 // to do a splice. Find tail of suffix so we can prepend suffix to global
8827 8834 // list.
8828 8835 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8829 8836 oop suffix_tail = cur;
8830 8837 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8831 8838 "Tautology");
8832 8839 observed_overflow_list = _overflow_list;
8833 8840 do {
8834 8841 cur_overflow_list = observed_overflow_list;
8835 8842 if (cur_overflow_list != BUSY) {
8836 8843 // Do the splice ...
8837 8844 suffix_tail->set_mark(markOop(cur_overflow_list));
8838 8845 } else { // cur_overflow_list == BUSY
8839 8846 suffix_tail->set_mark(NULL);
8840 8847 }
8841 8848 // ... and try to place spliced list back on overflow_list ...
8842 8849 observed_overflow_list =
8843 8850 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8844 8851 } while (cur_overflow_list != observed_overflow_list);
8845 8852 // ... until we have succeeded in doing so.
8846 8853 }
8847 8854 }
8848 8855
8849 8856 // Push the prefix elements on work_q
8850 8857 assert(prefix != NULL, "control point invariant");
8851 8858 const markOop proto = markOopDesc::prototype();
8852 8859 oop next;
8853 8860 NOT_PRODUCT(ssize_t n = 0;)
8854 8861 for (cur = prefix; cur != NULL; cur = next) {
8855 8862 next = oop(cur->mark());
8856 8863 cur->set_mark(proto); // until proven otherwise
8857 8864 assert(cur->is_oop(), "Should be an oop");
8858 8865 bool res = work_q->push(cur);
8859 8866 assert(res, "Bit off more than we can chew?");
8860 8867 NOT_PRODUCT(n++;)
8861 8868 }
8862 8869 #ifndef PRODUCT
8863 8870 assert(_num_par_pushes >= n, "Too many pops?");
8864 8871 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8865 8872 #endif
8866 8873 return true;
8867 8874 }
8868 8875
8869 8876 // Single-threaded
8870 8877 void CMSCollector::push_on_overflow_list(oop p) {
8871 8878 NOT_PRODUCT(_num_par_pushes++;)
8872 8879 assert(p->is_oop(), "Not an oop");
8873 8880 preserve_mark_if_necessary(p);
8874 8881 p->set_mark((markOop)_overflow_list);
8875 8882 _overflow_list = p;
8876 8883 }
8877 8884
8878 8885 // Multi-threaded; use CAS to prepend to overflow list
8879 8886 void CMSCollector::par_push_on_overflow_list(oop p) {
8880 8887 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8881 8888 assert(p->is_oop(), "Not an oop");
8882 8889 par_preserve_mark_if_necessary(p);
8883 8890 oop observed_overflow_list = _overflow_list;
8884 8891 oop cur_overflow_list;
8885 8892 do {
8886 8893 cur_overflow_list = observed_overflow_list;
8887 8894 if (cur_overflow_list != BUSY) {
8888 8895 p->set_mark(markOop(cur_overflow_list));
8889 8896 } else {
8890 8897 p->set_mark(NULL);
8891 8898 }
8892 8899 observed_overflow_list =
8893 8900 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8894 8901 } while (cur_overflow_list != observed_overflow_list);
8895 8902 }
8896 8903 #undef BUSY
8897 8904
8898 8905 // Single threaded
8899 8906 // General Note on GrowableArray: pushes may silently fail
8900 8907 // because we are (temporarily) out of C-heap for expanding
8901 8908 // the stack. The problem is quite ubiquitous and affects
8902 8909 // a lot of code in the JVM. The prudent thing for GrowableArray
8903 8910 // to do (for now) is to exit with an error. However, that may
8904 8911 // be too draconian in some cases because the caller may be
8905 8912 // able to recover without much harm. For such cases, we
8906 8913 // should probably introduce a "soft_push" method which returns
8907 8914 // an indication of success or failure with the assumption that
8908 8915 // the caller may be able to recover from a failure; code in
8909 8916 // the VM can then be changed, incrementally, to deal with such
8910 8917 // failures where possible, thus, incrementally hardening the VM
8911 8918 // in such low resource situations.
8912 8919 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8913 8920 _preserved_oop_stack.push(p);
8914 8921 _preserved_mark_stack.push(m);
8915 8922 assert(m == p->mark(), "Mark word changed");
8916 8923 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8917 8924 "bijection");
8918 8925 }
8919 8926
8920 8927 // Single threaded
8921 8928 void CMSCollector::preserve_mark_if_necessary(oop p) {
8922 8929 markOop m = p->mark();
8923 8930 if (m->must_be_preserved(p)) {
8924 8931 preserve_mark_work(p, m);
8925 8932 }
8926 8933 }
8927 8934
8928 8935 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8929 8936 markOop m = p->mark();
8930 8937 if (m->must_be_preserved(p)) {
8931 8938 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8932 8939 // Even though we read the mark word without holding
8933 8940 // the lock, we are assured that it will not change
8934 8941 // because we "own" this oop, so no other thread can
8935 8942 // be trying to push it on the overflow list; see
8936 8943 // the assertion in preserve_mark_work() that checks
8937 8944 // that m == p->mark().
8938 8945 preserve_mark_work(p, m);
8939 8946 }
8940 8947 }
8941 8948
8942 8949 // We should be able to do this multi-threaded,
8943 8950 // a chunk of stack being a task (this is
8944 8951 // correct because each oop only ever appears
8945 8952 // once in the overflow list. However, it's
8946 8953 // not very easy to completely overlap this with
8947 8954 // other operations, so will generally not be done
8948 8955 // until all work's been completed. Because we
8949 8956 // expect the preserved oop stack (set) to be small,
8950 8957 // it's probably fine to do this single-threaded.
8951 8958 // We can explore cleverer concurrent/overlapped/parallel
8952 8959 // processing of preserved marks if we feel the
8953 8960 // need for this in the future. Stack overflow should
8954 8961 // be so rare in practice and, when it happens, its
8955 8962 // effect on performance so great that this will
8956 8963 // likely just be in the noise anyway.
8957 8964 void CMSCollector::restore_preserved_marks_if_any() {
8958 8965 assert(SafepointSynchronize::is_at_safepoint(),
8959 8966 "world should be stopped");
8960 8967 assert(Thread::current()->is_ConcurrentGC_thread() ||
8961 8968 Thread::current()->is_VM_thread(),
8962 8969 "should be single-threaded");
8963 8970 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8964 8971 "bijection");
8965 8972
8966 8973 while (!_preserved_oop_stack.is_empty()) {
8967 8974 oop p = _preserved_oop_stack.pop();
8968 8975 assert(p->is_oop(), "Should be an oop");
8969 8976 assert(_span.contains(p), "oop should be in _span");
8970 8977 assert(p->mark() == markOopDesc::prototype(),
8971 8978 "Set when taken from overflow list");
8972 8979 markOop m = _preserved_mark_stack.pop();
8973 8980 p->set_mark(m);
8974 8981 }
8975 8982 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8976 8983 "stacks were cleared above");
8977 8984 }
8978 8985
8979 8986 #ifndef PRODUCT
8980 8987 bool CMSCollector::no_preserved_marks() const {
8981 8988 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8982 8989 }
8983 8990 #endif
8984 8991
8985 8992 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
8986 8993 {
8987 8994 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
8988 8995 CMSAdaptiveSizePolicy* size_policy =
8989 8996 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
8990 8997 assert(size_policy->is_gc_cms_adaptive_size_policy(),
8991 8998 "Wrong type for size policy");
8992 8999 return size_policy;
8993 9000 }
8994 9001
8995 9002 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
8996 9003 size_t desired_promo_size) {
8997 9004 if (cur_promo_size < desired_promo_size) {
8998 9005 size_t expand_bytes = desired_promo_size - cur_promo_size;
8999 9006 if (PrintAdaptiveSizePolicy && Verbose) {
9000 9007 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9001 9008 "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
9002 9009 expand_bytes);
9003 9010 }
9004 9011 expand(expand_bytes,
9005 9012 MinHeapDeltaBytes,
9006 9013 CMSExpansionCause::_adaptive_size_policy);
9007 9014 } else if (desired_promo_size < cur_promo_size) {
9008 9015 size_t shrink_bytes = cur_promo_size - desired_promo_size;
9009 9016 if (PrintAdaptiveSizePolicy && Verbose) {
9010 9017 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9011 9018 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
9012 9019 shrink_bytes);
9013 9020 }
9014 9021 shrink(shrink_bytes);
9015 9022 }
9016 9023 }
9017 9024
9018 9025 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
9019 9026 GenCollectedHeap* gch = GenCollectedHeap::heap();
9020 9027 CMSGCAdaptivePolicyCounters* counters =
9021 9028 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
9022 9029 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
9023 9030 "Wrong kind of counters");
9024 9031 return counters;
9025 9032 }
9026 9033
9027 9034
9028 9035 void ASConcurrentMarkSweepGeneration::update_counters() {
9029 9036 if (UsePerfData) {
9030 9037 _space_counters->update_all();
9031 9038 _gen_counters->update_all();
9032 9039 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9033 9040 GenCollectedHeap* gch = GenCollectedHeap::heap();
9034 9041 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9035 9042 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9036 9043 "Wrong gc statistics type");
9037 9044 counters->update_counters(gc_stats_l);
9038 9045 }
9039 9046 }
9040 9047
9041 9048 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
9042 9049 if (UsePerfData) {
9043 9050 _space_counters->update_used(used);
9044 9051 _space_counters->update_capacity();
9045 9052 _gen_counters->update_all();
9046 9053
9047 9054 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9048 9055 GenCollectedHeap* gch = GenCollectedHeap::heap();
9049 9056 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9050 9057 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9051 9058 "Wrong gc statistics type");
9052 9059 counters->update_counters(gc_stats_l);
9053 9060 }
9054 9061 }
9055 9062
9056 9063 // The desired expansion delta is computed so that:
9057 9064 // . desired free percentage or greater is used
9058 9065 void ASConcurrentMarkSweepGeneration::compute_new_size() {
9059 9066 assert_locked_or_safepoint(Heap_lock);
9060 9067
9061 9068 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
9062 9069
9063 9070 // If incremental collection failed, we just want to expand
9064 9071 // to the limit.
9065 9072 if (incremental_collection_failed()) {
9066 9073 clear_incremental_collection_failed();
9067 9074 grow_to_reserved();
9068 9075 return;
9069 9076 }
9070 9077
9071 9078 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing");
9072 9079
9073 9080 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
9074 9081 "Wrong type of heap");
9075 9082 int prev_level = level() - 1;
9076 9083 assert(prev_level >= 0, "The cms generation is the lowest generation");
9077 9084 Generation* prev_gen = gch->get_gen(prev_level);
9078 9085 assert(prev_gen->kind() == Generation::ASParNew,
9079 9086 "Wrong type of young generation");
9080 9087 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen;
9081 9088 size_t cur_eden = younger_gen->eden()->capacity();
9082 9089 CMSAdaptiveSizePolicy* size_policy = cms_size_policy();
9083 9090 size_t cur_promo = free();
9084 9091 size_policy->compute_tenured_generation_free_space(cur_promo,
9085 9092 max_available(),
9086 9093 cur_eden);
9087 9094 resize(cur_promo, size_policy->promo_size());
9088 9095
9089 9096 // Record the new size of the space in the cms generation
9090 9097 // that is available for promotions. This is temporary.
9091 9098 // It should be the desired promo size.
9092 9099 size_policy->avg_cms_promo()->sample(free());
9093 9100 size_policy->avg_old_live()->sample(used());
9094 9101
9095 9102 if (UsePerfData) {
9096 9103 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9097 9104 counters->update_cms_capacity_counter(capacity());
9098 9105 }
9099 9106 }
9100 9107
9101 9108 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
9102 9109 assert_locked_or_safepoint(Heap_lock);
9103 9110 assert_lock_strong(freelistLock());
9104 9111 HeapWord* old_end = _cmsSpace->end();
9105 9112 HeapWord* unallocated_start = _cmsSpace->unallocated_block();
9106 9113 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
9107 9114 FreeChunk* chunk_at_end = find_chunk_at_end();
9108 9115 if (chunk_at_end == NULL) {
9109 9116 // No room to shrink
9110 9117 if (PrintGCDetails && Verbose) {
9111 9118 gclog_or_tty->print_cr("No room to shrink: old_end "
9112 9119 PTR_FORMAT " unallocated_start " PTR_FORMAT
9113 9120 " chunk_at_end " PTR_FORMAT,
9114 9121 old_end, unallocated_start, chunk_at_end);
9115 9122 }
9116 9123 return;
9117 9124 } else {
9118 9125
9119 9126 // Find the chunk at the end of the space and determine
9120 9127 // how much it can be shrunk.
9121 9128 size_t shrinkable_size_in_bytes = chunk_at_end->size();
9122 9129 size_t aligned_shrinkable_size_in_bytes =
9123 9130 align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
9124 9131 assert(unallocated_start <= chunk_at_end->end(),
9125 9132 "Inconsistent chunk at end of space");
9126 9133 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
9127 9134 size_t word_size_before = heap_word_size(_virtual_space.committed_size());
9128 9135
9129 9136 // Shrink the underlying space
9130 9137 _virtual_space.shrink_by(bytes);
9131 9138 if (PrintGCDetails && Verbose) {
9132 9139 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
9133 9140 " desired_bytes " SIZE_FORMAT
9134 9141 " shrinkable_size_in_bytes " SIZE_FORMAT
9135 9142 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
9136 9143 " bytes " SIZE_FORMAT,
9137 9144 desired_bytes, shrinkable_size_in_bytes,
9138 9145 aligned_shrinkable_size_in_bytes, bytes);
9139 9146 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT
9140 9147 " unallocated_start " SIZE_FORMAT,
9141 9148 old_end, unallocated_start);
9142 9149 }
9143 9150
9144 9151 // If the space did shrink (shrinking is not guaranteed),
9145 9152 // shrink the chunk at the end by the appropriate amount.
9146 9153 if (((HeapWord*)_virtual_space.high()) < old_end) {
9147 9154 size_t new_word_size =
9148 9155 heap_word_size(_virtual_space.committed_size());
9149 9156
9150 9157 // Have to remove the chunk from the dictionary because it is changing
9151 9158 // size and might be someplace elsewhere in the dictionary.
9152 9159
9153 9160 // Get the chunk at end, shrink it, and put it
9154 9161 // back.
9155 9162 _cmsSpace->removeChunkFromDictionary(chunk_at_end);
9156 9163 size_t word_size_change = word_size_before - new_word_size;
9157 9164 size_t chunk_at_end_old_size = chunk_at_end->size();
9158 9165 assert(chunk_at_end_old_size >= word_size_change,
9159 9166 "Shrink is too large");
9160 9167 chunk_at_end->setSize(chunk_at_end_old_size -
9161 9168 word_size_change);
9162 9169 _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
9163 9170 word_size_change);
9164 9171
9165 9172 _cmsSpace->returnChunkToDictionary(chunk_at_end);
9166 9173
9167 9174 MemRegion mr(_cmsSpace->bottom(), new_word_size);
9168 9175 _bts->resize(new_word_size); // resize the block offset shared array
9169 9176 Universe::heap()->barrier_set()->resize_covered_region(mr);
9170 9177 _cmsSpace->assert_locked();
9171 9178 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
9172 9179
9173 9180 NOT_PRODUCT(_cmsSpace->dictionary()->verify());
9174 9181
9175 9182 // update the space and generation capacity counters
9176 9183 if (UsePerfData) {
9177 9184 _space_counters->update_capacity();
9178 9185 _gen_counters->update_all();
9179 9186 }
9180 9187
9181 9188 if (Verbose && PrintGCDetails) {
9182 9189 size_t new_mem_size = _virtual_space.committed_size();
9183 9190 size_t old_mem_size = new_mem_size + bytes;
9184 9191 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK",
9185 9192 name(), old_mem_size/K, bytes/K, new_mem_size/K);
9186 9193 }
9187 9194 }
9188 9195
9189 9196 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
9190 9197 "Inconsistency at end of space");
9191 9198 assert(chunk_at_end->end() == _cmsSpace->end(),
9192 9199 "Shrinking is inconsistent");
9193 9200 return;
9194 9201 }
9195 9202 }
9196 9203
9197 9204 // Transfer some number of overflown objects to usual marking
9198 9205 // stack. Return true if some objects were transferred.
9199 9206 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
9200 9207 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
9201 9208 (size_t)ParGCDesiredObjsFromOverflowList);
9202 9209
9203 9210 bool res = _collector->take_from_overflow_list(num, _mark_stack);
9204 9211 assert(_collector->overflow_list_is_empty() || res,
9205 9212 "If list is not empty, we should have taken something");
9206 9213 assert(!res || !_mark_stack->isEmpty(),
9207 9214 "If we took something, it should now be on our stack");
9208 9215 return res;
9209 9216 }
9210 9217
9211 9218 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
9212 9219 size_t res = _sp->block_size_no_stall(addr, _collector);
9213 9220 if (_sp->block_is_obj(addr)) {
9214 9221 if (_live_bit_map->isMarked(addr)) {
9215 9222 // It can't have been dead in a previous cycle
9216 9223 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
9217 9224 } else {
9218 9225 _dead_bit_map->mark(addr); // mark the dead object
9219 9226 }
9220 9227 }
9221 9228 // Could be 0, if the block size could not be computed without stalling.
9222 9229 return res;
9223 9230 }
9224 9231
9225 9232 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase): TraceMemoryManagerStats() {
9226 9233
9227 9234 switch (phase) {
9228 9235 case CMSCollector::InitialMarking:
9229 9236 initialize(true /* fullGC */ ,
9230 9237 true /* recordGCBeginTime */,
9231 9238 true /* recordPreGCUsage */,
9232 9239 false /* recordPeakUsage */,
9233 9240 false /* recordPostGCusage */,
9234 9241 true /* recordAccumulatedGCTime */,
9235 9242 false /* recordGCEndTime */,
9236 9243 false /* countCollection */ );
9237 9244 break;
9238 9245
9239 9246 case CMSCollector::FinalMarking:
9240 9247 initialize(true /* fullGC */ ,
9241 9248 false /* recordGCBeginTime */,
9242 9249 false /* recordPreGCUsage */,
9243 9250 false /* recordPeakUsage */,
9244 9251 false /* recordPostGCusage */,
9245 9252 true /* recordAccumulatedGCTime */,
9246 9253 false /* recordGCEndTime */,
9247 9254 false /* countCollection */ );
9248 9255 break;
9249 9256
9250 9257 case CMSCollector::Sweeping:
9251 9258 initialize(true /* fullGC */ ,
9252 9259 false /* recordGCBeginTime */,
9253 9260 false /* recordPreGCUsage */,
9254 9261 true /* recordPeakUsage */,
9255 9262 true /* recordPostGCusage */,
9256 9263 false /* recordAccumulatedGCTime */,
9257 9264 true /* recordGCEndTime */,
9258 9265 true /* countCollection */ );
9259 9266 break;
9260 9267
9261 9268 default:
9262 9269 ShouldNotReachHere();
9263 9270 }
9264 9271 }
9265 9272
9266 9273 // when bailing out of cms in concurrent mode failure
9267 9274 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(): TraceMemoryManagerStats() {
9268 9275 initialize(true /* fullGC */ ,
9269 9276 true /* recordGCBeginTime */,
9270 9277 true /* recordPreGCUsage */,
9271 9278 true /* recordPeakUsage */,
9272 9279 true /* recordPostGCusage */,
9273 9280 true /* recordAccumulatedGCTime */,
9274 9281 true /* recordGCEndTime */,
9275 9282 true /* countCollection */ );
9276 9283 }
|
↓ open down ↓ |
3553 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX