Print this page
rev 2691 : [mq]: g1-reference-processing
| 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, 2011, 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 }
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
295 295 _ref_processor =
296 296 new ReferenceProcessor(_span, // span
297 297 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
298 298 (int) ParallelGCThreads, // mt processing degree
299 299 _cmsGen->refs_discovery_is_mt(), // mt discovery
300 300 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
301 301 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
302 302 &_is_alive_closure, // closure for liveness info
303 303 false); // next field updates do not need write barrier
304 304 // Initialize the _ref_processor field of CMSGen
305 305 _cmsGen->set_ref_processor(_ref_processor);
306 306
307 307 // Allocate a dummy ref processor for perm gen.
308 308 ReferenceProcessor* rp2 = new ReferenceProcessor();
309 309 if (rp2 == NULL) {
310 310 vm_exit_during_initialization("Could not allocate ReferenceProcessor object");
311 311 }
312 312 _permGen->set_ref_processor(rp2);
313 313 }
314 314 }
315 315
316 316 CMSAdaptiveSizePolicy* CMSCollector::size_policy() {
317 317 GenCollectedHeap* gch = GenCollectedHeap::heap();
318 318 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
319 319 "Wrong type of heap");
320 320 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
321 321 gch->gen_policy()->size_policy();
322 322 assert(sp->is_gc_cms_adaptive_size_policy(),
323 323 "Wrong type of size policy");
324 324 return sp;
325 325 }
326 326
327 327 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() {
328 328 CMSGCAdaptivePolicyCounters* results =
329 329 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters();
330 330 assert(
331 331 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
332 332 "Wrong gc policy counter kind");
333 333 return results;
334 334 }
335 335
336 336
337 337 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
338 338
339 339 const char* gen_name = "old";
340 340
341 341 // Generation Counters - generation 1, 1 subspace
342 342 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space);
343 343
344 344 _space_counters = new GSpaceCounters(gen_name, 0,
345 345 _virtual_space.reserved_size(),
346 346 this, _gen_counters);
347 347 }
348 348
349 349 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
350 350 _cms_gen(cms_gen)
351 351 {
352 352 assert(alpha <= 100, "bad value");
353 353 _saved_alpha = alpha;
354 354
355 355 // Initialize the alphas to the bootstrap value of 100.
356 356 _gc0_alpha = _cms_alpha = 100;
357 357
358 358 _cms_begin_time.update();
359 359 _cms_end_time.update();
360 360
361 361 _gc0_duration = 0.0;
362 362 _gc0_period = 0.0;
363 363 _gc0_promoted = 0;
364 364
365 365 _cms_duration = 0.0;
366 366 _cms_period = 0.0;
367 367 _cms_allocated = 0;
368 368
369 369 _cms_used_at_gc0_begin = 0;
370 370 _cms_used_at_gc0_end = 0;
371 371 _allow_duty_cycle_reduction = false;
372 372 _valid_bits = 0;
373 373 _icms_duty_cycle = CMSIncrementalDutyCycle;
374 374 }
375 375
376 376 double CMSStats::cms_free_adjustment_factor(size_t free) const {
377 377 // TBD: CR 6909490
378 378 return 1.0;
379 379 }
380 380
381 381 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
382 382 }
383 383
384 384 // If promotion failure handling is on use
385 385 // the padded average size of the promotion for each
386 386 // young generation collection.
387 387 double CMSStats::time_until_cms_gen_full() const {
388 388 size_t cms_free = _cms_gen->cmsSpace()->free();
389 389 GenCollectedHeap* gch = GenCollectedHeap::heap();
390 390 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(),
391 391 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
392 392 if (cms_free > expected_promotion) {
393 393 // Start a cms collection if there isn't enough space to promote
394 394 // for the next minor collection. Use the padded average as
395 395 // a safety factor.
396 396 cms_free -= expected_promotion;
397 397
398 398 // Adjust by the safety factor.
399 399 double cms_free_dbl = (double)cms_free;
400 400 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
401 401 // Apply a further correction factor which tries to adjust
402 402 // for recent occurance of concurrent mode failures.
403 403 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
404 404 cms_free_dbl = cms_free_dbl * cms_adjustment;
405 405
406 406 if (PrintGCDetails && Verbose) {
407 407 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
408 408 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
409 409 cms_free, expected_promotion);
410 410 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
411 411 cms_free_dbl, cms_consumption_rate() + 1.0);
412 412 }
413 413 // Add 1 in case the consumption rate goes to zero.
414 414 return cms_free_dbl / (cms_consumption_rate() + 1.0);
415 415 }
416 416 return 0.0;
417 417 }
418 418
419 419 // Compare the duration of the cms collection to the
420 420 // time remaining before the cms generation is empty.
421 421 // Note that the time from the start of the cms collection
422 422 // to the start of the cms sweep (less than the total
423 423 // duration of the cms collection) can be used. This
424 424 // has been tried and some applications experienced
425 425 // promotion failures early in execution. This was
426 426 // possibly because the averages were not accurate
427 427 // enough at the beginning.
428 428 double CMSStats::time_until_cms_start() const {
429 429 // We add "gc0_period" to the "work" calculation
430 430 // below because this query is done (mostly) at the
431 431 // end of a scavenge, so we need to conservatively
432 432 // account for that much possible delay
433 433 // in the query so as to avoid concurrent mode failures
434 434 // due to starting the collection just a wee bit too
435 435 // late.
436 436 double work = cms_duration() + gc0_period();
437 437 double deadline = time_until_cms_gen_full();
438 438 // If a concurrent mode failure occurred recently, we want to be
439 439 // more conservative and halve our expected time_until_cms_gen_full()
440 440 if (work > deadline) {
441 441 if (Verbose && PrintGCDetails) {
442 442 gclog_or_tty->print(
443 443 " CMSCollector: collect because of anticipated promotion "
444 444 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
445 445 gc0_period(), time_until_cms_gen_full());
446 446 }
447 447 return 0.0;
448 448 }
449 449 return work - deadline;
450 450 }
451 451
452 452 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the
453 453 // amount of change to prevent wild oscillation.
454 454 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle,
455 455 unsigned int new_duty_cycle) {
456 456 assert(old_duty_cycle <= 100, "bad input value");
457 457 assert(new_duty_cycle <= 100, "bad input value");
458 458
459 459 // Note: use subtraction with caution since it may underflow (values are
460 460 // unsigned). Addition is safe since we're in the range 0-100.
461 461 unsigned int damped_duty_cycle = new_duty_cycle;
462 462 if (new_duty_cycle < old_duty_cycle) {
463 463 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U);
464 464 if (new_duty_cycle + largest_delta < old_duty_cycle) {
465 465 damped_duty_cycle = old_duty_cycle - largest_delta;
466 466 }
467 467 } else if (new_duty_cycle > old_duty_cycle) {
468 468 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U);
469 469 if (new_duty_cycle > old_duty_cycle + largest_delta) {
470 470 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U);
471 471 }
472 472 }
473 473 assert(damped_duty_cycle <= 100, "invalid duty cycle computed");
474 474
475 475 if (CMSTraceIncrementalPacing) {
476 476 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ",
477 477 old_duty_cycle, new_duty_cycle, damped_duty_cycle);
478 478 }
479 479 return damped_duty_cycle;
480 480 }
481 481
482 482 unsigned int CMSStats::icms_update_duty_cycle_impl() {
483 483 assert(CMSIncrementalPacing && valid(),
484 484 "should be handled in icms_update_duty_cycle()");
485 485
486 486 double cms_time_so_far = cms_timer().seconds();
487 487 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M;
488 488 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far);
489 489
490 490 // Avoid division by 0.
491 491 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01);
492 492 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full;
493 493
494 494 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U);
495 495 if (new_duty_cycle > _icms_duty_cycle) {
496 496 // Avoid very small duty cycles (1 or 2); 0 is allowed.
497 497 if (new_duty_cycle > 2) {
498 498 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle,
499 499 new_duty_cycle);
500 500 }
501 501 } else if (_allow_duty_cycle_reduction) {
502 502 // The duty cycle is reduced only once per cms cycle (see record_cms_end()).
503 503 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle);
504 504 // Respect the minimum duty cycle.
505 505 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin;
506 506 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle);
507 507 }
508 508
509 509 if (PrintGCDetails || CMSTraceIncrementalPacing) {
510 510 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle);
511 511 }
512 512
513 513 _allow_duty_cycle_reduction = false;
514 514 return _icms_duty_cycle;
515 515 }
516 516
517 517 #ifndef PRODUCT
518 518 void CMSStats::print_on(outputStream *st) const {
519 519 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
520 520 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
521 521 gc0_duration(), gc0_period(), gc0_promoted());
522 522 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
523 523 cms_duration(), cms_duration_per_mb(),
524 524 cms_period(), cms_allocated());
525 525 st->print(",cms_since_beg=%g,cms_since_end=%g",
526 526 cms_time_since_begin(), cms_time_since_end());
527 527 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
528 528 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
529 529 if (CMSIncrementalMode) {
530 530 st->print(",dc=%d", icms_duty_cycle());
531 531 }
532 532
533 533 if (valid()) {
534 534 st->print(",promo_rate=%g,cms_alloc_rate=%g",
535 535 promotion_rate(), cms_allocation_rate());
536 536 st->print(",cms_consumption_rate=%g,time_until_full=%g",
537 537 cms_consumption_rate(), time_until_cms_gen_full());
538 538 }
539 539 st->print(" ");
540 540 }
541 541 #endif // #ifndef PRODUCT
542 542
543 543 CMSCollector::CollectorState CMSCollector::_collectorState =
544 544 CMSCollector::Idling;
545 545 bool CMSCollector::_foregroundGCIsActive = false;
546 546 bool CMSCollector::_foregroundGCShouldWait = false;
547 547
548 548 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
549 549 ConcurrentMarkSweepGeneration* permGen,
550 550 CardTableRS* ct,
551 551 ConcurrentMarkSweepPolicy* cp):
552 552 _cmsGen(cmsGen),
553 553 _permGen(permGen),
554 554 _ct(ct),
555 555 _ref_processor(NULL), // will be set later
556 556 _conc_workers(NULL), // may be set later
557 557 _abort_preclean(false),
558 558 _start_sampling(false),
559 559 _between_prologue_and_epilogue(false),
560 560 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
561 561 _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"),
562 562 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
563 563 -1 /* lock-free */, "No_lock" /* dummy */),
564 564 _modUnionClosure(&_modUnionTable),
565 565 _modUnionClosurePar(&_modUnionTable),
566 566 // Adjust my span to cover old (cms) gen and perm gen
567 567 _span(cmsGen->reserved()._union(permGen->reserved())),
568 568 // Construct the is_alive_closure with _span & markBitMap
569 569 _is_alive_closure(_span, &_markBitMap),
570 570 _restart_addr(NULL),
571 571 _overflow_list(NULL),
572 572 _stats(cmsGen),
573 573 _eden_chunk_array(NULL), // may be set in ctor body
574 574 _eden_chunk_capacity(0), // -- ditto --
575 575 _eden_chunk_index(0), // -- ditto --
576 576 _survivor_plab_array(NULL), // -- ditto --
577 577 _survivor_chunk_array(NULL), // -- ditto --
578 578 _survivor_chunk_capacity(0), // -- ditto --
579 579 _survivor_chunk_index(0), // -- ditto --
580 580 _ser_pmc_preclean_ovflw(0),
581 581 _ser_kac_preclean_ovflw(0),
582 582 _ser_pmc_remark_ovflw(0),
583 583 _par_pmc_remark_ovflw(0),
584 584 _ser_kac_ovflw(0),
585 585 _par_kac_ovflw(0),
586 586 #ifndef PRODUCT
587 587 _num_par_pushes(0),
588 588 #endif
589 589 _collection_count_start(0),
590 590 _verifying(false),
591 591 _icms_start_limit(NULL),
592 592 _icms_stop_limit(NULL),
593 593 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
594 594 _completed_initialization(false),
595 595 _collector_policy(cp),
596 596 _should_unload_classes(false),
597 597 _concurrent_cycles_since_last_unload(0),
598 598 _roots_scanning_options(0),
599 599 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
600 600 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding)
601 601 {
602 602 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
603 603 ExplicitGCInvokesConcurrent = true;
604 604 }
605 605 // Now expand the span and allocate the collection support structures
606 606 // (MUT, marking bit map etc.) to cover both generations subject to
607 607 // collection.
608 608
609 609 // First check that _permGen is adjacent to _cmsGen and above it.
610 610 assert( _cmsGen->reserved().word_size() > 0
611 611 && _permGen->reserved().word_size() > 0,
612 612 "generations should not be of zero size");
613 613 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(),
614 614 "_cmsGen and _permGen should not overlap");
615 615 assert(_cmsGen->reserved().end() == _permGen->reserved().start(),
616 616 "_cmsGen->end() different from _permGen->start()");
617 617
618 618 // For use by dirty card to oop closures.
619 619 _cmsGen->cmsSpace()->set_collector(this);
620 620 _permGen->cmsSpace()->set_collector(this);
621 621
622 622 // Allocate MUT and marking bit map
623 623 {
624 624 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
625 625 if (!_markBitMap.allocate(_span)) {
626 626 warning("Failed to allocate CMS Bit Map");
627 627 return;
628 628 }
629 629 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
630 630 }
631 631 {
632 632 _modUnionTable.allocate(_span);
633 633 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
634 634 }
635 635
636 636 if (!_markStack.allocate(MarkStackSize)) {
637 637 warning("Failed to allocate CMS Marking Stack");
638 638 return;
639 639 }
640 640 if (!_revisitStack.allocate(CMSRevisitStackSize)) {
641 641 warning("Failed to allocate CMS Revisit Stack");
642 642 return;
643 643 }
644 644
645 645 // Support for multi-threaded concurrent phases
646 646 if (CMSConcurrentMTEnabled) {
647 647 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
648 648 // just for now
649 649 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
650 650 }
651 651 if (ConcGCThreads > 1) {
652 652 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
653 653 ConcGCThreads, true);
654 654 if (_conc_workers == NULL) {
655 655 warning("GC/CMS: _conc_workers allocation failure: "
656 656 "forcing -CMSConcurrentMTEnabled");
657 657 CMSConcurrentMTEnabled = false;
658 658 } else {
659 659 _conc_workers->initialize_workers();
660 660 }
661 661 } else {
662 662 CMSConcurrentMTEnabled = false;
663 663 }
664 664 }
665 665 if (!CMSConcurrentMTEnabled) {
666 666 ConcGCThreads = 0;
667 667 } else {
668 668 // Turn off CMSCleanOnEnter optimization temporarily for
669 669 // the MT case where it's not fixed yet; see 6178663.
670 670 CMSCleanOnEnter = false;
671 671 }
672 672 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
673 673 "Inconsistency");
674 674
675 675 // Parallel task queues; these are shared for the
676 676 // concurrent and stop-world phases of CMS, but
677 677 // are not shared with parallel scavenge (ParNew).
678 678 {
679 679 uint i;
680 680 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
681 681
682 682 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
683 683 || ParallelRefProcEnabled)
684 684 && num_queues > 0) {
685 685 _task_queues = new OopTaskQueueSet(num_queues);
686 686 if (_task_queues == NULL) {
687 687 warning("task_queues allocation failure.");
688 688 return;
689 689 }
690 690 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues);
691 691 if (_hash_seed == NULL) {
692 692 warning("_hash_seed array allocation failure");
693 693 return;
694 694 }
695 695
696 696 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
697 697 for (i = 0; i < num_queues; i++) {
698 698 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
699 699 if (q == NULL) {
700 700 warning("work_queue allocation failure.");
701 701 return;
702 702 }
703 703 _task_queues->register_queue(i, q);
704 704 }
705 705 for (i = 0; i < num_queues; i++) {
706 706 _task_queues->queue(i)->initialize();
707 707 _hash_seed[i] = 17; // copied from ParNew
708 708 }
709 709 }
710 710 }
711 711
712 712 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
713 713 _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio);
714 714
715 715 // Clip CMSBootstrapOccupancy between 0 and 100.
716 716 _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy)))
717 717 /(double)100;
718 718
719 719 _full_gcs_since_conc_gc = 0;
720 720
721 721 // Now tell CMS generations the identity of their collector
722 722 ConcurrentMarkSweepGeneration::set_collector(this);
723 723
724 724 // Create & start a CMS thread for this CMS collector
725 725 _cmsThread = ConcurrentMarkSweepThread::start(this);
726 726 assert(cmsThread() != NULL, "CMS Thread should have been created");
727 727 assert(cmsThread()->collector() == this,
728 728 "CMS Thread should refer to this gen");
729 729 assert(CGC_lock != NULL, "Where's the CGC_lock?");
730 730
731 731 // Support for parallelizing young gen rescan
732 732 GenCollectedHeap* gch = GenCollectedHeap::heap();
733 733 _young_gen = gch->prev_gen(_cmsGen);
734 734 if (gch->supports_inline_contig_alloc()) {
735 735 _top_addr = gch->top_addr();
736 736 _end_addr = gch->end_addr();
737 737 assert(_young_gen != NULL, "no _young_gen");
738 738 _eden_chunk_index = 0;
739 739 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
740 740 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity);
741 741 if (_eden_chunk_array == NULL) {
742 742 _eden_chunk_capacity = 0;
743 743 warning("GC/CMS: _eden_chunk_array allocation failure");
744 744 }
745 745 }
746 746 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
747 747
748 748 // Support for parallelizing survivor space rescan
749 749 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) {
750 750 const size_t max_plab_samples =
751 751 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize;
752 752
753 753 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads);
754 754 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples);
755 755 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads);
756 756 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
757 757 || _cursor == NULL) {
758 758 warning("Failed to allocate survivor plab/chunk array");
759 759 if (_survivor_plab_array != NULL) {
760 760 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
761 761 _survivor_plab_array = NULL;
762 762 }
763 763 if (_survivor_chunk_array != NULL) {
764 764 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
765 765 _survivor_chunk_array = NULL;
766 766 }
767 767 if (_cursor != NULL) {
768 768 FREE_C_HEAP_ARRAY(size_t, _cursor);
769 769 _cursor = NULL;
770 770 }
771 771 } else {
772 772 _survivor_chunk_capacity = 2*max_plab_samples;
773 773 for (uint i = 0; i < ParallelGCThreads; i++) {
774 774 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples);
775 775 if (vec == NULL) {
776 776 warning("Failed to allocate survivor plab array");
777 777 for (int j = i; j > 0; j--) {
778 778 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array());
779 779 }
780 780 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
781 781 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
782 782 _survivor_plab_array = NULL;
783 783 _survivor_chunk_array = NULL;
784 784 _survivor_chunk_capacity = 0;
785 785 break;
786 786 } else {
787 787 ChunkArray* cur =
788 788 ::new (&_survivor_plab_array[i]) ChunkArray(vec,
789 789 max_plab_samples);
790 790 assert(cur->end() == 0, "Should be 0");
791 791 assert(cur->array() == vec, "Should be vec");
792 792 assert(cur->capacity() == max_plab_samples, "Error");
793 793 }
794 794 }
795 795 }
796 796 }
797 797 assert( ( _survivor_plab_array != NULL
798 798 && _survivor_chunk_array != NULL)
799 799 || ( _survivor_chunk_capacity == 0
800 800 && _survivor_chunk_index == 0),
801 801 "Error");
802 802
803 803 // Choose what strong roots should be scanned depending on verification options
804 804 // and perm gen collection mode.
805 805 if (!CMSClassUnloadingEnabled) {
806 806 // If class unloading is disabled we want to include all classes into the root set.
807 807 add_root_scanning_option(SharedHeap::SO_AllClasses);
808 808 } else {
809 809 add_root_scanning_option(SharedHeap::SO_SystemClasses);
810 810 }
811 811
812 812 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
813 813 _gc_counters = new CollectorCounters("CMS", 1);
814 814 _completed_initialization = true;
815 815 _inter_sweep_timer.start(); // start of time
816 816 #ifdef SPARC
817 817 // Issue a stern warning, but allow use for experimentation and debugging.
818 818 if (VM_Version::is_sun4v() && UseMemSetInBOT) {
819 819 assert(!FLAG_IS_DEFAULT(UseMemSetInBOT), "Error");
820 820 warning("Experimental flag -XX:+UseMemSetInBOT is known to cause instability"
821 821 " on sun4v; please understand that you are using at your own risk!");
822 822 }
823 823 #endif
824 824 }
825 825
826 826 const char* ConcurrentMarkSweepGeneration::name() const {
827 827 return "concurrent mark-sweep generation";
828 828 }
829 829 void ConcurrentMarkSweepGeneration::update_counters() {
830 830 if (UsePerfData) {
831 831 _space_counters->update_all();
832 832 _gen_counters->update_all();
833 833 }
834 834 }
835 835
836 836 // this is an optimized version of update_counters(). it takes the
837 837 // used value as a parameter rather than computing it.
838 838 //
839 839 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
840 840 if (UsePerfData) {
841 841 _space_counters->update_used(used);
842 842 _space_counters->update_capacity();
843 843 _gen_counters->update_all();
844 844 }
845 845 }
846 846
847 847 void ConcurrentMarkSweepGeneration::print() const {
848 848 Generation::print();
849 849 cmsSpace()->print();
850 850 }
851 851
852 852 #ifndef PRODUCT
853 853 void ConcurrentMarkSweepGeneration::print_statistics() {
854 854 cmsSpace()->printFLCensus(0);
855 855 }
856 856 #endif
857 857
858 858 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
859 859 GenCollectedHeap* gch = GenCollectedHeap::heap();
860 860 if (PrintGCDetails) {
861 861 if (Verbose) {
862 862 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
863 863 level(), short_name(), s, used(), capacity());
864 864 } else {
865 865 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
866 866 level(), short_name(), s, used() / K, capacity() / K);
867 867 }
868 868 }
869 869 if (Verbose) {
870 870 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
871 871 gch->used(), gch->capacity());
872 872 } else {
873 873 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
874 874 gch->used() / K, gch->capacity() / K);
875 875 }
876 876 }
877 877
878 878 size_t
879 879 ConcurrentMarkSweepGeneration::contiguous_available() const {
880 880 // dld proposes an improvement in precision here. If the committed
881 881 // part of the space ends in a free block we should add that to
882 882 // uncommitted size in the calculation below. Will make this
883 883 // change later, staying with the approximation below for the
884 884 // time being. -- ysr.
885 885 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
886 886 }
887 887
888 888 size_t
889 889 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
890 890 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
891 891 }
892 892
893 893 size_t ConcurrentMarkSweepGeneration::max_available() const {
894 894 return free() + _virtual_space.uncommitted_size();
895 895 }
896 896
897 897 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
898 898 size_t available = max_available();
899 899 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
900 900 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
901 901 if (Verbose && PrintGCDetails) {
902 902 gclog_or_tty->print_cr(
903 903 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
904 904 "max_promo("SIZE_FORMAT")",
905 905 res? "":" not", available, res? ">=":"<",
906 906 av_promo, max_promotion_in_bytes);
907 907 }
908 908 return res;
909 909 }
910 910
911 911 // At a promotion failure dump information on block layout in heap
912 912 // (cms old generation).
913 913 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
914 914 if (CMSDumpAtPromotionFailure) {
915 915 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
916 916 }
917 917 }
918 918
919 919 CompactibleSpace*
920 920 ConcurrentMarkSweepGeneration::first_compaction_space() const {
921 921 return _cmsSpace;
922 922 }
923 923
924 924 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
925 925 // Clear the promotion information. These pointers can be adjusted
926 926 // along with all the other pointers into the heap but
927 927 // compaction is expected to be a rare event with
928 928 // a heap using cms so don't do it without seeing the need.
929 929 if (CollectedHeap::use_parallel_gc_threads()) {
930 930 for (uint i = 0; i < ParallelGCThreads; i++) {
931 931 _par_gc_thread_states[i]->promo.reset();
932 932 }
933 933 }
934 934 }
935 935
936 936 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
937 937 blk->do_space(_cmsSpace);
938 938 }
939 939
940 940 void ConcurrentMarkSweepGeneration::compute_new_size() {
941 941 assert_locked_or_safepoint(Heap_lock);
942 942
943 943 // If incremental collection failed, we just want to expand
944 944 // to the limit.
945 945 if (incremental_collection_failed()) {
946 946 clear_incremental_collection_failed();
947 947 grow_to_reserved();
948 948 return;
949 949 }
950 950
951 951 size_t expand_bytes = 0;
952 952 double free_percentage = ((double) free()) / capacity();
953 953 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
954 954 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
955 955
956 956 // compute expansion delta needed for reaching desired free percentage
957 957 if (free_percentage < desired_free_percentage) {
958 958 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
959 959 assert(desired_capacity >= capacity(), "invalid expansion size");
960 960 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
961 961 }
962 962 if (expand_bytes > 0) {
963 963 if (PrintGCDetails && Verbose) {
964 964 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
965 965 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
966 966 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
967 967 gclog_or_tty->print_cr(" Desired free fraction %f",
968 968 desired_free_percentage);
969 969 gclog_or_tty->print_cr(" Maximum free fraction %f",
970 970 maximum_free_percentage);
971 971 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000);
972 972 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
973 973 desired_capacity/1000);
974 974 int prev_level = level() - 1;
975 975 if (prev_level >= 0) {
976 976 size_t prev_size = 0;
977 977 GenCollectedHeap* gch = GenCollectedHeap::heap();
978 978 Generation* prev_gen = gch->_gens[prev_level];
979 979 prev_size = prev_gen->capacity();
980 980 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
981 981 prev_size/1000);
982 982 }
983 983 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
984 984 unsafe_max_alloc_nogc()/1000);
985 985 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
986 986 contiguous_available()/1000);
987 987 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
988 988 expand_bytes);
989 989 }
990 990 // safe if expansion fails
991 991 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
992 992 if (PrintGCDetails && Verbose) {
993 993 gclog_or_tty->print_cr(" Expanded free fraction %f",
994 994 ((double) free()) / capacity());
995 995 }
996 996 }
997 997 }
998 998
999 999 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
1000 1000 return cmsSpace()->freelistLock();
1001 1001 }
1002 1002
1003 1003 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
1004 1004 bool tlab) {
1005 1005 CMSSynchronousYieldRequest yr;
1006 1006 MutexLockerEx x(freelistLock(),
1007 1007 Mutex::_no_safepoint_check_flag);
1008 1008 return have_lock_and_allocate(size, tlab);
1009 1009 }
1010 1010
1011 1011 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
1012 1012 bool tlab /* ignored */) {
1013 1013 assert_lock_strong(freelistLock());
1014 1014 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
1015 1015 HeapWord* res = cmsSpace()->allocate(adjustedSize);
1016 1016 // Allocate the object live (grey) if the background collector has
1017 1017 // started marking. This is necessary because the marker may
1018 1018 // have passed this address and consequently this object will
1019 1019 // not otherwise be greyed and would be incorrectly swept up.
1020 1020 // Note that if this object contains references, the writing
1021 1021 // of those references will dirty the card containing this object
1022 1022 // allowing the object to be blackened (and its references scanned)
1023 1023 // either during a preclean phase or at the final checkpoint.
1024 1024 if (res != NULL) {
1025 1025 // We may block here with an uninitialized object with
1026 1026 // its mark-bit or P-bits not yet set. Such objects need
1027 1027 // to be safely navigable by block_start().
1028 1028 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
1029 1029 assert(!((FreeChunk*)res)->isFree(), "Error, block will look free but show wrong size");
1030 1030 collector()->direct_allocated(res, adjustedSize);
1031 1031 _direct_allocated_words += adjustedSize;
1032 1032 // allocation counters
1033 1033 NOT_PRODUCT(
1034 1034 _numObjectsAllocated++;
1035 1035 _numWordsAllocated += (int)adjustedSize;
1036 1036 )
1037 1037 }
1038 1038 return res;
1039 1039 }
1040 1040
1041 1041 // In the case of direct allocation by mutators in a generation that
1042 1042 // is being concurrently collected, the object must be allocated
1043 1043 // live (grey) if the background collector has started marking.
1044 1044 // This is necessary because the marker may
1045 1045 // have passed this address and consequently this object will
1046 1046 // not otherwise be greyed and would be incorrectly swept up.
1047 1047 // Note that if this object contains references, the writing
1048 1048 // of those references will dirty the card containing this object
1049 1049 // allowing the object to be blackened (and its references scanned)
1050 1050 // either during a preclean phase or at the final checkpoint.
1051 1051 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
1052 1052 assert(_markBitMap.covers(start, size), "Out of bounds");
1053 1053 if (_collectorState >= Marking) {
1054 1054 MutexLockerEx y(_markBitMap.lock(),
1055 1055 Mutex::_no_safepoint_check_flag);
1056 1056 // [see comments preceding SweepClosure::do_blk() below for details]
1057 1057 // 1. need to mark the object as live so it isn't collected
1058 1058 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
1059 1059 // 3. need to mark the end of the object so marking, precleaning or sweeping
1060 1060 // can skip over uninitialized or unparsable objects. An allocated
1061 1061 // object is considered uninitialized for our purposes as long as
1062 1062 // its klass word is NULL. (Unparsable objects are those which are
1063 1063 // initialized in the sense just described, but whose sizes can still
1064 1064 // not be correctly determined. Note that the class of unparsable objects
1065 1065 // can only occur in the perm gen. All old gen objects are parsable
1066 1066 // as soon as they are initialized.)
1067 1067 _markBitMap.mark(start); // object is live
1068 1068 _markBitMap.mark(start + 1); // object is potentially uninitialized?
1069 1069 _markBitMap.mark(start + size - 1);
1070 1070 // mark end of object
1071 1071 }
1072 1072 // check that oop looks uninitialized
1073 1073 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
1074 1074 }
1075 1075
1076 1076 void CMSCollector::promoted(bool par, HeapWord* start,
1077 1077 bool is_obj_array, size_t obj_size) {
1078 1078 assert(_markBitMap.covers(start), "Out of bounds");
1079 1079 // See comment in direct_allocated() about when objects should
1080 1080 // be allocated live.
1081 1081 if (_collectorState >= Marking) {
1082 1082 // we already hold the marking bit map lock, taken in
1083 1083 // the prologue
1084 1084 if (par) {
1085 1085 _markBitMap.par_mark(start);
1086 1086 } else {
1087 1087 _markBitMap.mark(start);
1088 1088 }
1089 1089 // We don't need to mark the object as uninitialized (as
1090 1090 // in direct_allocated above) because this is being done with the
1091 1091 // world stopped and the object will be initialized by the
1092 1092 // time the marking, precleaning or sweeping get to look at it.
1093 1093 // But see the code for copying objects into the CMS generation,
1094 1094 // where we need to ensure that concurrent readers of the
1095 1095 // block offset table are able to safely navigate a block that
1096 1096 // is in flux from being free to being allocated (and in
1097 1097 // transition while being copied into) and subsequently
1098 1098 // becoming a bona-fide object when the copy/promotion is complete.
1099 1099 assert(SafepointSynchronize::is_at_safepoint(),
1100 1100 "expect promotion only at safepoints");
1101 1101
1102 1102 if (_collectorState < Sweeping) {
1103 1103 // Mark the appropriate cards in the modUnionTable, so that
1104 1104 // this object gets scanned before the sweep. If this is
1105 1105 // not done, CMS generation references in the object might
1106 1106 // not get marked.
1107 1107 // For the case of arrays, which are otherwise precisely
1108 1108 // marked, we need to dirty the entire array, not just its head.
1109 1109 if (is_obj_array) {
1110 1110 // The [par_]mark_range() method expects mr.end() below to
1111 1111 // be aligned to the granularity of a bit's representation
1112 1112 // in the heap. In the case of the MUT below, that's a
1113 1113 // card size.
1114 1114 MemRegion mr(start,
1115 1115 (HeapWord*)round_to((intptr_t)(start + obj_size),
1116 1116 CardTableModRefBS::card_size /* bytes */));
1117 1117 if (par) {
1118 1118 _modUnionTable.par_mark_range(mr);
1119 1119 } else {
1120 1120 _modUnionTable.mark_range(mr);
1121 1121 }
1122 1122 } else { // not an obj array; we can just mark the head
1123 1123 if (par) {
1124 1124 _modUnionTable.par_mark(start);
1125 1125 } else {
1126 1126 _modUnionTable.mark(start);
1127 1127 }
1128 1128 }
1129 1129 }
1130 1130 }
1131 1131 }
1132 1132
1133 1133 static inline size_t percent_of_space(Space* space, HeapWord* addr)
1134 1134 {
1135 1135 size_t delta = pointer_delta(addr, space->bottom());
1136 1136 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
1137 1137 }
1138 1138
1139 1139 void CMSCollector::icms_update_allocation_limits()
1140 1140 {
1141 1141 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
1142 1142 EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
1143 1143
1144 1144 const unsigned int duty_cycle = stats().icms_update_duty_cycle();
1145 1145 if (CMSTraceIncrementalPacing) {
1146 1146 stats().print();
1147 1147 }
1148 1148
1149 1149 assert(duty_cycle <= 100, "invalid duty cycle");
1150 1150 if (duty_cycle != 0) {
1151 1151 // The duty_cycle is a percentage between 0 and 100; convert to words and
1152 1152 // then compute the offset from the endpoints of the space.
1153 1153 size_t free_words = eden->free() / HeapWordSize;
1154 1154 double free_words_dbl = (double)free_words;
1155 1155 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
1156 1156 size_t offset_words = (free_words - duty_cycle_words) / 2;
1157 1157
1158 1158 _icms_start_limit = eden->top() + offset_words;
1159 1159 _icms_stop_limit = eden->end() - offset_words;
1160 1160
1161 1161 // The limits may be adjusted (shifted to the right) by
1162 1162 // CMSIncrementalOffset, to allow the application more mutator time after a
1163 1163 // young gen gc (when all mutators were stopped) and before CMS starts and
1164 1164 // takes away one or more cpus.
1165 1165 if (CMSIncrementalOffset != 0) {
1166 1166 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
1167 1167 size_t adjustment = (size_t)adjustment_dbl;
1168 1168 HeapWord* tmp_stop = _icms_stop_limit + adjustment;
1169 1169 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
1170 1170 _icms_start_limit += adjustment;
1171 1171 _icms_stop_limit = tmp_stop;
1172 1172 }
1173 1173 }
1174 1174 }
1175 1175 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
1176 1176 _icms_start_limit = _icms_stop_limit = eden->end();
1177 1177 }
1178 1178
1179 1179 // Install the new start limit.
1180 1180 eden->set_soft_end(_icms_start_limit);
1181 1181
1182 1182 if (CMSTraceIncrementalMode) {
1183 1183 gclog_or_tty->print(" icms alloc limits: "
1184 1184 PTR_FORMAT "," PTR_FORMAT
1185 1185 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
1186 1186 _icms_start_limit, _icms_stop_limit,
1187 1187 percent_of_space(eden, _icms_start_limit),
1188 1188 percent_of_space(eden, _icms_stop_limit));
1189 1189 if (Verbose) {
1190 1190 gclog_or_tty->print("eden: ");
1191 1191 eden->print_on(gclog_or_tty);
1192 1192 }
1193 1193 }
1194 1194 }
1195 1195
1196 1196 // Any changes here should try to maintain the invariant
1197 1197 // that if this method is called with _icms_start_limit
1198 1198 // and _icms_stop_limit both NULL, then it should return NULL
1199 1199 // and not notify the icms thread.
1200 1200 HeapWord*
1201 1201 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
1202 1202 size_t word_size)
1203 1203 {
1204 1204 // A start_limit equal to end() means the duty cycle is 0, so treat that as a
1205 1205 // nop.
1206 1206 if (CMSIncrementalMode && _icms_start_limit != space->end()) {
1207 1207 if (top <= _icms_start_limit) {
1208 1208 if (CMSTraceIncrementalMode) {
1209 1209 space->print_on(gclog_or_tty);
1210 1210 gclog_or_tty->stamp();
1211 1211 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
1212 1212 ", new limit=" PTR_FORMAT
1213 1213 " (" SIZE_FORMAT "%%)",
1214 1214 top, _icms_stop_limit,
1215 1215 percent_of_space(space, _icms_stop_limit));
1216 1216 }
1217 1217 ConcurrentMarkSweepThread::start_icms();
1218 1218 assert(top < _icms_stop_limit, "Tautology");
1219 1219 if (word_size < pointer_delta(_icms_stop_limit, top)) {
1220 1220 return _icms_stop_limit;
1221 1221 }
1222 1222
1223 1223 // The allocation will cross both the _start and _stop limits, so do the
1224 1224 // stop notification also and return end().
1225 1225 if (CMSTraceIncrementalMode) {
1226 1226 space->print_on(gclog_or_tty);
1227 1227 gclog_or_tty->stamp();
1228 1228 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
1229 1229 ", new limit=" PTR_FORMAT
1230 1230 " (" SIZE_FORMAT "%%)",
1231 1231 top, space->end(),
1232 1232 percent_of_space(space, space->end()));
1233 1233 }
1234 1234 ConcurrentMarkSweepThread::stop_icms();
1235 1235 return space->end();
1236 1236 }
1237 1237
1238 1238 if (top <= _icms_stop_limit) {
1239 1239 if (CMSTraceIncrementalMode) {
1240 1240 space->print_on(gclog_or_tty);
1241 1241 gclog_or_tty->stamp();
1242 1242 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
1243 1243 ", new limit=" PTR_FORMAT
1244 1244 " (" SIZE_FORMAT "%%)",
1245 1245 top, space->end(),
1246 1246 percent_of_space(space, space->end()));
1247 1247 }
1248 1248 ConcurrentMarkSweepThread::stop_icms();
1249 1249 return space->end();
1250 1250 }
1251 1251
1252 1252 if (CMSTraceIncrementalMode) {
1253 1253 space->print_on(gclog_or_tty);
1254 1254 gclog_or_tty->stamp();
1255 1255 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
1256 1256 ", new limit=" PTR_FORMAT,
1257 1257 top, NULL);
1258 1258 }
1259 1259 }
1260 1260
1261 1261 return NULL;
1262 1262 }
1263 1263
1264 1264 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
1265 1265 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1266 1266 // allocate, copy and if necessary update promoinfo --
1267 1267 // delegate to underlying space.
1268 1268 assert_lock_strong(freelistLock());
1269 1269
1270 1270 #ifndef PRODUCT
1271 1271 if (Universe::heap()->promotion_should_fail()) {
1272 1272 return NULL;
1273 1273 }
1274 1274 #endif // #ifndef PRODUCT
1275 1275
1276 1276 oop res = _cmsSpace->promote(obj, obj_size);
1277 1277 if (res == NULL) {
1278 1278 // expand and retry
1279 1279 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
1280 1280 expand(s*HeapWordSize, MinHeapDeltaBytes,
1281 1281 CMSExpansionCause::_satisfy_promotion);
1282 1282 // Since there's currently no next generation, we don't try to promote
1283 1283 // into a more senior generation.
1284 1284 assert(next_gen() == NULL, "assumption, based upon which no attempt "
1285 1285 "is made to pass on a possibly failing "
1286 1286 "promotion to next generation");
1287 1287 res = _cmsSpace->promote(obj, obj_size);
1288 1288 }
1289 1289 if (res != NULL) {
1290 1290 // See comment in allocate() about when objects should
1291 1291 // be allocated live.
1292 1292 assert(obj->is_oop(), "Will dereference klass pointer below");
1293 1293 collector()->promoted(false, // Not parallel
1294 1294 (HeapWord*)res, obj->is_objArray(), obj_size);
1295 1295 // promotion counters
1296 1296 NOT_PRODUCT(
1297 1297 _numObjectsPromoted++;
1298 1298 _numWordsPromoted +=
1299 1299 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1300 1300 )
1301 1301 }
1302 1302 return res;
1303 1303 }
1304 1304
1305 1305
1306 1306 HeapWord*
1307 1307 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
1308 1308 HeapWord* top,
1309 1309 size_t word_sz)
1310 1310 {
1311 1311 return collector()->allocation_limit_reached(space, top, word_sz);
1312 1312 }
1313 1313
1314 1314 // IMPORTANT: Notes on object size recognition in CMS.
1315 1315 // ---------------------------------------------------
1316 1316 // A block of storage in the CMS generation is always in
1317 1317 // one of three states. A free block (FREE), an allocated
1318 1318 // object (OBJECT) whose size() method reports the correct size,
1319 1319 // and an intermediate state (TRANSIENT) in which its size cannot
1320 1320 // be accurately determined.
1321 1321 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1322 1322 // -----------------------------------------------------
1323 1323 // FREE: klass_word & 1 == 1; mark_word holds block size
1324 1324 //
1325 1325 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1326 1326 // obj->size() computes correct size
1327 1327 // [Perm Gen objects needs to be "parsable" before they can be navigated]
1328 1328 //
1329 1329 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1330 1330 //
1331 1331 // STATE IDENTIFICATION: (64 bit+COOPS)
1332 1332 // ------------------------------------
1333 1333 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1334 1334 //
1335 1335 // OBJECT: klass_word installed; klass_word != 0;
1336 1336 // obj->size() computes correct size
1337 1337 // [Perm Gen comment above continues to hold]
1338 1338 //
1339 1339 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1340 1340 //
1341 1341 //
1342 1342 // STATE TRANSITION DIAGRAM
1343 1343 //
1344 1344 // mut / parnew mut / parnew
1345 1345 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1346 1346 // ^ |
1347 1347 // |------------------------ DEAD <------------------------------------|
1348 1348 // sweep mut
1349 1349 //
1350 1350 // While a block is in TRANSIENT state its size cannot be determined
1351 1351 // so readers will either need to come back later or stall until
1352 1352 // the size can be determined. Note that for the case of direct
1353 1353 // allocation, P-bits, when available, may be used to determine the
1354 1354 // size of an object that may not yet have been initialized.
1355 1355
1356 1356 // Things to support parallel young-gen collection.
1357 1357 oop
1358 1358 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1359 1359 oop old, markOop m,
1360 1360 size_t word_sz) {
1361 1361 #ifndef PRODUCT
1362 1362 if (Universe::heap()->promotion_should_fail()) {
1363 1363 return NULL;
1364 1364 }
1365 1365 #endif // #ifndef PRODUCT
1366 1366
1367 1367 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1368 1368 PromotionInfo* promoInfo = &ps->promo;
1369 1369 // if we are tracking promotions, then first ensure space for
1370 1370 // promotion (including spooling space for saving header if necessary).
1371 1371 // then allocate and copy, then track promoted info if needed.
1372 1372 // When tracking (see PromotionInfo::track()), the mark word may
1373 1373 // be displaced and in this case restoration of the mark word
1374 1374 // occurs in the (oop_since_save_marks_)iterate phase.
1375 1375 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1376 1376 // Out of space for allocating spooling buffers;
1377 1377 // try expanding and allocating spooling buffers.
1378 1378 if (!expand_and_ensure_spooling_space(promoInfo)) {
1379 1379 return NULL;
1380 1380 }
1381 1381 }
1382 1382 assert(promoInfo->has_spooling_space(), "Control point invariant");
1383 1383 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1384 1384 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1385 1385 if (obj_ptr == NULL) {
1386 1386 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1387 1387 if (obj_ptr == NULL) {
1388 1388 return NULL;
1389 1389 }
1390 1390 }
1391 1391 oop obj = oop(obj_ptr);
1392 1392 OrderAccess::storestore();
1393 1393 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1394 1394 assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size");
1395 1395 // IMPORTANT: See note on object initialization for CMS above.
1396 1396 // Otherwise, copy the object. Here we must be careful to insert the
1397 1397 // klass pointer last, since this marks the block as an allocated object.
1398 1398 // Except with compressed oops it's the mark word.
1399 1399 HeapWord* old_ptr = (HeapWord*)old;
1400 1400 // Restore the mark word copied above.
1401 1401 obj->set_mark(m);
1402 1402 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1403 1403 assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size");
1404 1404 OrderAccess::storestore();
1405 1405
1406 1406 if (UseCompressedOops) {
1407 1407 // Copy gap missed by (aligned) header size calculation below
1408 1408 obj->set_klass_gap(old->klass_gap());
1409 1409 }
1410 1410 if (word_sz > (size_t)oopDesc::header_size()) {
1411 1411 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1412 1412 obj_ptr + oopDesc::header_size(),
1413 1413 word_sz - oopDesc::header_size());
1414 1414 }
1415 1415
1416 1416 // Now we can track the promoted object, if necessary. We take care
1417 1417 // to delay the transition from uninitialized to full object
1418 1418 // (i.e., insertion of klass pointer) until after, so that it
1419 1419 // atomically becomes a promoted object.
1420 1420 if (promoInfo->tracking()) {
1421 1421 promoInfo->track((PromotedObject*)obj, old->klass());
1422 1422 }
1423 1423 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1424 1424 assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size");
1425 1425 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1426 1426
1427 1427 // Finally, install the klass pointer (this should be volatile).
1428 1428 OrderAccess::storestore();
1429 1429 obj->set_klass(old->klass());
1430 1430 // We should now be able to calculate the right size for this object
1431 1431 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1432 1432
1433 1433 collector()->promoted(true, // parallel
1434 1434 obj_ptr, old->is_objArray(), word_sz);
1435 1435
1436 1436 NOT_PRODUCT(
1437 1437 Atomic::inc_ptr(&_numObjectsPromoted);
1438 1438 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1439 1439 )
1440 1440
1441 1441 return obj;
1442 1442 }
1443 1443
1444 1444 void
1445 1445 ConcurrentMarkSweepGeneration::
1446 1446 par_promote_alloc_undo(int thread_num,
1447 1447 HeapWord* obj, size_t word_sz) {
1448 1448 // CMS does not support promotion undo.
1449 1449 ShouldNotReachHere();
1450 1450 }
1451 1451
1452 1452 void
1453 1453 ConcurrentMarkSweepGeneration::
1454 1454 par_promote_alloc_done(int thread_num) {
1455 1455 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1456 1456 ps->lab.retire(thread_num);
1457 1457 }
1458 1458
1459 1459 void
1460 1460 ConcurrentMarkSweepGeneration::
1461 1461 par_oop_since_save_marks_iterate_done(int thread_num) {
1462 1462 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1463 1463 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1464 1464 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1465 1465 }
1466 1466
1467 1467 // XXXPERM
1468 1468 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1469 1469 size_t size,
1470 1470 bool tlab)
1471 1471 {
1472 1472 // We allow a STW collection only if a full
1473 1473 // collection was requested.
1474 1474 return full || should_allocate(size, tlab); // FIX ME !!!
1475 1475 // This and promotion failure handling are connected at the
1476 1476 // hip and should be fixed by untying them.
1477 1477 }
1478 1478
1479 1479 bool CMSCollector::shouldConcurrentCollect() {
1480 1480 if (_full_gc_requested) {
1481 1481 if (Verbose && PrintGCDetails) {
1482 1482 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1483 1483 " gc request (or gc_locker)");
1484 1484 }
1485 1485 return true;
1486 1486 }
1487 1487
1488 1488 // For debugging purposes, change the type of collection.
1489 1489 // If the rotation is not on the concurrent collection
1490 1490 // type, don't start a concurrent collection.
1491 1491 NOT_PRODUCT(
1492 1492 if (RotateCMSCollectionTypes &&
1493 1493 (_cmsGen->debug_collection_type() !=
1494 1494 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
1495 1495 assert(_cmsGen->debug_collection_type() !=
1496 1496 ConcurrentMarkSweepGeneration::Unknown_collection_type,
1497 1497 "Bad cms collection type");
1498 1498 return false;
1499 1499 }
1500 1500 )
1501 1501
1502 1502 FreelistLocker x(this);
1503 1503 // ------------------------------------------------------------------
1504 1504 // Print out lots of information which affects the initiation of
1505 1505 // a collection.
1506 1506 if (PrintCMSInitiationStatistics && stats().valid()) {
1507 1507 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1508 1508 gclog_or_tty->stamp();
1509 1509 gclog_or_tty->print_cr("");
1510 1510 stats().print_on(gclog_or_tty);
1511 1511 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1512 1512 stats().time_until_cms_gen_full());
1513 1513 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1514 1514 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1515 1515 _cmsGen->contiguous_available());
1516 1516 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1517 1517 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1518 1518 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1519 1519 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1520 1520 gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy());
1521 1521 }
1522 1522 // ------------------------------------------------------------------
1523 1523
1524 1524 // If the estimated time to complete a cms collection (cms_duration())
1525 1525 // is less than the estimated time remaining until the cms generation
1526 1526 // is full, start a collection.
1527 1527 if (!UseCMSInitiatingOccupancyOnly) {
1528 1528 if (stats().valid()) {
1529 1529 if (stats().time_until_cms_start() == 0.0) {
1530 1530 return true;
1531 1531 }
1532 1532 } else {
1533 1533 // We want to conservatively collect somewhat early in order
1534 1534 // to try and "bootstrap" our CMS/promotion statistics;
1535 1535 // this branch will not fire after the first successful CMS
1536 1536 // collection because the stats should then be valid.
1537 1537 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1538 1538 if (Verbose && PrintGCDetails) {
1539 1539 gclog_or_tty->print_cr(
1540 1540 " CMSCollector: collect for bootstrapping statistics:"
1541 1541 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1542 1542 _bootstrap_occupancy);
1543 1543 }
1544 1544 return true;
1545 1545 }
1546 1546 }
1547 1547 }
1548 1548
1549 1549 // Otherwise, we start a collection cycle if either the perm gen or
1550 1550 // old gen want a collection cycle started. Each may use
1551 1551 // an appropriate criterion for making this decision.
1552 1552 // XXX We need to make sure that the gen expansion
1553 1553 // criterion dovetails well with this. XXX NEED TO FIX THIS
1554 1554 if (_cmsGen->should_concurrent_collect()) {
1555 1555 if (Verbose && PrintGCDetails) {
1556 1556 gclog_or_tty->print_cr("CMS old gen initiated");
1557 1557 }
1558 1558 return true;
1559 1559 }
1560 1560
1561 1561 // We start a collection if we believe an incremental collection may fail;
1562 1562 // this is not likely to be productive in practice because it's probably too
1563 1563 // late anyway.
1564 1564 GenCollectedHeap* gch = GenCollectedHeap::heap();
1565 1565 assert(gch->collector_policy()->is_two_generation_policy(),
1566 1566 "You may want to check the correctness of the following");
1567 1567 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1568 1568 if (Verbose && PrintGCDetails) {
1569 1569 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1570 1570 }
1571 1571 return true;
1572 1572 }
1573 1573
1574 1574 if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) {
1575 1575 bool res = update_should_unload_classes();
1576 1576 if (res) {
1577 1577 if (Verbose && PrintGCDetails) {
1578 1578 gclog_or_tty->print_cr("CMS perm gen initiated");
1579 1579 }
1580 1580 return true;
1581 1581 }
1582 1582 }
1583 1583 return false;
1584 1584 }
1585 1585
1586 1586 // Clear _expansion_cause fields of constituent generations
1587 1587 void CMSCollector::clear_expansion_cause() {
1588 1588 _cmsGen->clear_expansion_cause();
1589 1589 _permGen->clear_expansion_cause();
1590 1590 }
1591 1591
1592 1592 // We should be conservative in starting a collection cycle. To
1593 1593 // start too eagerly runs the risk of collecting too often in the
1594 1594 // extreme. To collect too rarely falls back on full collections,
1595 1595 // which works, even if not optimum in terms of concurrent work.
1596 1596 // As a work around for too eagerly collecting, use the flag
1597 1597 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1598 1598 // giving the user an easily understandable way of controlling the
1599 1599 // collections.
1600 1600 // We want to start a new collection cycle if any of the following
1601 1601 // conditions hold:
1602 1602 // . our current occupancy exceeds the configured initiating occupancy
1603 1603 // for this generation, or
1604 1604 // . we recently needed to expand this space and have not, since that
1605 1605 // expansion, done a collection of this generation, or
1606 1606 // . the underlying space believes that it may be a good idea to initiate
1607 1607 // a concurrent collection (this may be based on criteria such as the
1608 1608 // following: the space uses linear allocation and linear allocation is
1609 1609 // going to fail, or there is believed to be excessive fragmentation in
1610 1610 // the generation, etc... or ...
1611 1611 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1612 1612 // the case of the old generation, not the perm generation; see CR 6543076):
1613 1613 // we may be approaching a point at which allocation requests may fail because
1614 1614 // we will be out of sufficient free space given allocation rate estimates.]
1615 1615 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1616 1616
1617 1617 assert_lock_strong(freelistLock());
1618 1618 if (occupancy() > initiating_occupancy()) {
1619 1619 if (PrintGCDetails && Verbose) {
1620 1620 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1621 1621 short_name(), occupancy(), initiating_occupancy());
1622 1622 }
1623 1623 return true;
1624 1624 }
1625 1625 if (UseCMSInitiatingOccupancyOnly) {
1626 1626 return false;
1627 1627 }
1628 1628 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1629 1629 if (PrintGCDetails && Verbose) {
1630 1630 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1631 1631 short_name());
1632 1632 }
1633 1633 return true;
1634 1634 }
1635 1635 if (_cmsSpace->should_concurrent_collect()) {
1636 1636 if (PrintGCDetails && Verbose) {
1637 1637 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1638 1638 short_name());
1639 1639 }
1640 1640 return true;
1641 1641 }
1642 1642 return false;
1643 1643 }
1644 1644
1645 1645 void ConcurrentMarkSweepGeneration::collect(bool full,
1646 1646 bool clear_all_soft_refs,
1647 1647 size_t size,
1648 1648 bool tlab)
1649 1649 {
1650 1650 collector()->collect(full, clear_all_soft_refs, size, tlab);
1651 1651 }
1652 1652
1653 1653 void CMSCollector::collect(bool full,
1654 1654 bool clear_all_soft_refs,
1655 1655 size_t size,
1656 1656 bool tlab)
1657 1657 {
1658 1658 if (!UseCMSCollectionPassing && _collectorState > Idling) {
1659 1659 // For debugging purposes skip the collection if the state
1660 1660 // is not currently idle
1661 1661 if (TraceCMSState) {
1662 1662 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
1663 1663 Thread::current(), full, _collectorState);
1664 1664 }
1665 1665 return;
1666 1666 }
1667 1667
1668 1668 // The following "if" branch is present for defensive reasons.
1669 1669 // In the current uses of this interface, it can be replaced with:
1670 1670 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1671 1671 // But I am not placing that assert here to allow future
1672 1672 // generality in invoking this interface.
1673 1673 if (GC_locker::is_active()) {
1674 1674 // A consistency test for GC_locker
1675 1675 assert(GC_locker::needs_gc(), "Should have been set already");
1676 1676 // Skip this foreground collection, instead
1677 1677 // expanding the heap if necessary.
1678 1678 // Need the free list locks for the call to free() in compute_new_size()
1679 1679 compute_new_size();
1680 1680 return;
1681 1681 }
1682 1682 acquire_control_and_collect(full, clear_all_soft_refs);
1683 1683 _full_gcs_since_conc_gc++;
1684 1684
1685 1685 }
1686 1686
1687 1687 void CMSCollector::request_full_gc(unsigned int full_gc_count) {
1688 1688 GenCollectedHeap* gch = GenCollectedHeap::heap();
1689 1689 unsigned int gc_count = gch->total_full_collections();
1690 1690 if (gc_count == full_gc_count) {
1691 1691 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1692 1692 _full_gc_requested = true;
1693 1693 CGC_lock->notify(); // nudge CMS thread
1694 1694 } else {
1695 1695 assert(gc_count > full_gc_count, "Error: causal loop");
1696 1696 }
1697 1697 }
1698 1698
1699 1699
1700 1700 // The foreground and background collectors need to coordinate in order
1701 1701 // to make sure that they do not mutually interfere with CMS collections.
1702 1702 // When a background collection is active,
1703 1703 // the foreground collector may need to take over (preempt) and
1704 1704 // synchronously complete an ongoing collection. Depending on the
1705 1705 // frequency of the background collections and the heap usage
1706 1706 // of the application, this preemption can be seldom or frequent.
1707 1707 // There are only certain
1708 1708 // points in the background collection that the "collection-baton"
1709 1709 // can be passed to the foreground collector.
1710 1710 //
1711 1711 // The foreground collector will wait for the baton before
1712 1712 // starting any part of the collection. The foreground collector
1713 1713 // will only wait at one location.
1714 1714 //
1715 1715 // The background collector will yield the baton before starting a new
1716 1716 // phase of the collection (e.g., before initial marking, marking from roots,
1717 1717 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1718 1718 // of the loop which switches the phases. The background collector does some
1719 1719 // of the phases (initial mark, final re-mark) with the world stopped.
1720 1720 // Because of locking involved in stopping the world,
1721 1721 // the foreground collector should not block waiting for the background
1722 1722 // collector when it is doing a stop-the-world phase. The background
1723 1723 // collector will yield the baton at an additional point just before
1724 1724 // it enters a stop-the-world phase. Once the world is stopped, the
1725 1725 // background collector checks the phase of the collection. If the
1726 1726 // phase has not changed, it proceeds with the collection. If the
1727 1727 // phase has changed, it skips that phase of the collection. See
1728 1728 // the comments on the use of the Heap_lock in collect_in_background().
1729 1729 //
1730 1730 // Variable used in baton passing.
1731 1731 // _foregroundGCIsActive - Set to true by the foreground collector when
1732 1732 // it wants the baton. The foreground clears it when it has finished
1733 1733 // the collection.
1734 1734 // _foregroundGCShouldWait - Set to true by the background collector
1735 1735 // when it is running. The foreground collector waits while
1736 1736 // _foregroundGCShouldWait is true.
1737 1737 // CGC_lock - monitor used to protect access to the above variables
1738 1738 // and to notify the foreground and background collectors.
1739 1739 // _collectorState - current state of the CMS collection.
1740 1740 //
1741 1741 // The foreground collector
1742 1742 // acquires the CGC_lock
1743 1743 // sets _foregroundGCIsActive
1744 1744 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1745 1745 // various locks acquired in preparation for the collection
1746 1746 // are released so as not to block the background collector
1747 1747 // that is in the midst of a collection
1748 1748 // proceeds with the collection
1749 1749 // clears _foregroundGCIsActive
1750 1750 // returns
1751 1751 //
1752 1752 // The background collector in a loop iterating on the phases of the
1753 1753 // collection
1754 1754 // acquires the CGC_lock
1755 1755 // sets _foregroundGCShouldWait
1756 1756 // if _foregroundGCIsActive is set
1757 1757 // clears _foregroundGCShouldWait, notifies _CGC_lock
1758 1758 // waits on _CGC_lock for _foregroundGCIsActive to become false
1759 1759 // and exits the loop.
1760 1760 // otherwise
1761 1761 // proceed with that phase of the collection
1762 1762 // if the phase is a stop-the-world phase,
1763 1763 // yield the baton once more just before enqueueing
1764 1764 // the stop-world CMS operation (executed by the VM thread).
1765 1765 // returns after all phases of the collection are done
1766 1766 //
1767 1767
1768 1768 void CMSCollector::acquire_control_and_collect(bool full,
1769 1769 bool clear_all_soft_refs) {
1770 1770 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1771 1771 assert(!Thread::current()->is_ConcurrentGC_thread(),
1772 1772 "shouldn't try to acquire control from self!");
1773 1773
1774 1774 // Start the protocol for acquiring control of the
1775 1775 // collection from the background collector (aka CMS thread).
1776 1776 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1777 1777 "VM thread should have CMS token");
1778 1778 // Remember the possibly interrupted state of an ongoing
1779 1779 // concurrent collection
1780 1780 CollectorState first_state = _collectorState;
1781 1781
1782 1782 // Signal to a possibly ongoing concurrent collection that
1783 1783 // we want to do a foreground collection.
1784 1784 _foregroundGCIsActive = true;
1785 1785
1786 1786 // Disable incremental mode during a foreground collection.
1787 1787 ICMSDisabler icms_disabler;
1788 1788
1789 1789 // release locks and wait for a notify from the background collector
1790 1790 // releasing the locks in only necessary for phases which
1791 1791 // do yields to improve the granularity of the collection.
1792 1792 assert_lock_strong(bitMapLock());
1793 1793 // We need to lock the Free list lock for the space that we are
1794 1794 // currently collecting.
1795 1795 assert(haveFreelistLocks(), "Must be holding free list locks");
1796 1796 bitMapLock()->unlock();
1797 1797 releaseFreelistLocks();
1798 1798 {
1799 1799 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1800 1800 if (_foregroundGCShouldWait) {
1801 1801 // We are going to be waiting for action for the CMS thread;
1802 1802 // it had better not be gone (for instance at shutdown)!
1803 1803 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1804 1804 "CMS thread must be running");
1805 1805 // Wait here until the background collector gives us the go-ahead
1806 1806 ConcurrentMarkSweepThread::clear_CMS_flag(
1807 1807 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1808 1808 // Get a possibly blocked CMS thread going:
1809 1809 // Note that we set _foregroundGCIsActive true above,
1810 1810 // without protection of the CGC_lock.
1811 1811 CGC_lock->notify();
1812 1812 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1813 1813 "Possible deadlock");
1814 1814 while (_foregroundGCShouldWait) {
1815 1815 // wait for notification
1816 1816 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1817 1817 // Possibility of delay/starvation here, since CMS token does
1818 1818 // not know to give priority to VM thread? Actually, i think
1819 1819 // there wouldn't be any delay/starvation, but the proof of
1820 1820 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1821 1821 }
1822 1822 ConcurrentMarkSweepThread::set_CMS_flag(
1823 1823 ConcurrentMarkSweepThread::CMS_vm_has_token);
1824 1824 }
1825 1825 }
1826 1826 // The CMS_token is already held. Get back the other locks.
1827 1827 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1828 1828 "VM thread should have CMS token");
1829 1829 getFreelistLocks();
1830 1830 bitMapLock()->lock_without_safepoint_check();
1831 1831 if (TraceCMSState) {
1832 1832 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1833 1833 INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1834 1834 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1835 1835 }
1836 1836
1837 1837 // Check if we need to do a compaction, or if not, whether
1838 1838 // we need to start the mark-sweep from scratch.
1839 1839 bool should_compact = false;
1840 1840 bool should_start_over = false;
1841 1841 decide_foreground_collection_type(clear_all_soft_refs,
1842 1842 &should_compact, &should_start_over);
1843 1843
1844 1844 NOT_PRODUCT(
1845 1845 if (RotateCMSCollectionTypes) {
1846 1846 if (_cmsGen->debug_collection_type() ==
1847 1847 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1848 1848 should_compact = true;
1849 1849 } else if (_cmsGen->debug_collection_type() ==
1850 1850 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1851 1851 should_compact = false;
1852 1852 }
1853 1853 }
1854 1854 )
1855 1855
1856 1856 if (PrintGCDetails && first_state > Idling) {
1857 1857 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1858 1858 if (GCCause::is_user_requested_gc(cause) ||
1859 1859 GCCause::is_serviceability_requested_gc(cause)) {
1860 1860 gclog_or_tty->print(" (concurrent mode interrupted)");
1861 1861 } else {
1862 1862 gclog_or_tty->print(" (concurrent mode failure)");
1863 1863 }
1864 1864 }
1865 1865
1866 1866 if (should_compact) {
1867 1867 // If the collection is being acquired from the background
1868 1868 // collector, there may be references on the discovered
1869 1869 // references lists that have NULL referents (being those
1870 1870 // that were concurrently cleared by a mutator) or
1871 1871 // that are no longer active (having been enqueued concurrently
1872 1872 // by the mutator).
1873 1873 // Scrub the list of those references because Mark-Sweep-Compact
1874 1874 // code assumes referents are not NULL and that all discovered
1875 1875 // Reference objects are active.
1876 1876 ref_processor()->clean_up_discovered_references();
1877 1877
1878 1878 do_compaction_work(clear_all_soft_refs);
1879 1879
1880 1880 // Has the GC time limit been exceeded?
1881 1881 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration();
1882 1882 size_t max_eden_size = young_gen->max_capacity() -
1883 1883 young_gen->to()->capacity() -
1884 1884 young_gen->from()->capacity();
1885 1885 GenCollectedHeap* gch = GenCollectedHeap::heap();
1886 1886 GCCause::Cause gc_cause = gch->gc_cause();
1887 1887 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1888 1888 young_gen->eden()->used(),
1889 1889 _cmsGen->max_capacity(),
1890 1890 max_eden_size,
1891 1891 full,
1892 1892 gc_cause,
1893 1893 gch->collector_policy());
1894 1894 } else {
1895 1895 do_mark_sweep_work(clear_all_soft_refs, first_state,
1896 1896 should_start_over);
1897 1897 }
1898 1898 // Reset the expansion cause, now that we just completed
1899 1899 // a collection cycle.
1900 1900 clear_expansion_cause();
1901 1901 _foregroundGCIsActive = false;
1902 1902 return;
1903 1903 }
1904 1904
1905 1905 // Resize the perm generation and the tenured generation
1906 1906 // after obtaining the free list locks for the
1907 1907 // two generations.
1908 1908 void CMSCollector::compute_new_size() {
1909 1909 assert_locked_or_safepoint(Heap_lock);
1910 1910 FreelistLocker z(this);
1911 1911 _permGen->compute_new_size();
1912 1912 _cmsGen->compute_new_size();
1913 1913 }
1914 1914
1915 1915 // A work method used by foreground collection to determine
1916 1916 // what type of collection (compacting or not, continuing or fresh)
1917 1917 // it should do.
1918 1918 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1919 1919 // and CMSCompactWhenClearAllSoftRefs the default in the future
1920 1920 // and do away with the flags after a suitable period.
1921 1921 void CMSCollector::decide_foreground_collection_type(
1922 1922 bool clear_all_soft_refs, bool* should_compact,
1923 1923 bool* should_start_over) {
1924 1924 // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1925 1925 // flag is set, and we have either requested a System.gc() or
1926 1926 // the number of full gc's since the last concurrent cycle
1927 1927 // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1928 1928 // or if an incremental collection has failed
1929 1929 GenCollectedHeap* gch = GenCollectedHeap::heap();
1930 1930 assert(gch->collector_policy()->is_two_generation_policy(),
1931 1931 "You may want to check the correctness of the following");
1932 1932 // Inform cms gen if this was due to partial collection failing.
1933 1933 // The CMS gen may use this fact to determine its expansion policy.
1934 1934 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1935 1935 assert(!_cmsGen->incremental_collection_failed(),
1936 1936 "Should have been noticed, reacted to and cleared");
1937 1937 _cmsGen->set_incremental_collection_failed();
1938 1938 }
1939 1939 *should_compact =
1940 1940 UseCMSCompactAtFullCollection &&
1941 1941 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1942 1942 GCCause::is_user_requested_gc(gch->gc_cause()) ||
1943 1943 gch->incremental_collection_will_fail(true /* consult_young */));
1944 1944 *should_start_over = false;
1945 1945 if (clear_all_soft_refs && !*should_compact) {
1946 1946 // We are about to do a last ditch collection attempt
1947 1947 // so it would normally make sense to do a compaction
1948 1948 // to reclaim as much space as possible.
1949 1949 if (CMSCompactWhenClearAllSoftRefs) {
1950 1950 // Default: The rationale is that in this case either
1951 1951 // we are past the final marking phase, in which case
1952 1952 // we'd have to start over, or so little has been done
1953 1953 // that there's little point in saving that work. Compaction
1954 1954 // appears to be the sensible choice in either case.
1955 1955 *should_compact = true;
1956 1956 } else {
1957 1957 // We have been asked to clear all soft refs, but not to
1958 1958 // compact. Make sure that we aren't past the final checkpoint
1959 1959 // phase, for that is where we process soft refs. If we are already
1960 1960 // past that phase, we'll need to redo the refs discovery phase and
1961 1961 // if necessary clear soft refs that weren't previously
1962 1962 // cleared. We do so by remembering the phase in which
1963 1963 // we came in, and if we are past the refs processing
1964 1964 // phase, we'll choose to just redo the mark-sweep
1965 1965 // collection from scratch.
1966 1966 if (_collectorState > FinalMarking) {
1967 1967 // We are past the refs processing phase;
1968 1968 // start over and do a fresh synchronous CMS cycle
1969 1969 _collectorState = Resetting; // skip to reset to start new cycle
1970 1970 reset(false /* == !asynch */);
1971 1971 *should_start_over = true;
1972 1972 } // else we can continue a possibly ongoing current cycle
1973 1973 }
1974 1974 }
1975 1975 }
1976 1976
1977 1977 // A work method used by the foreground collector to do
1978 1978 // a mark-sweep-compact.
1979 1979 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1980 1980 GenCollectedHeap* gch = GenCollectedHeap::heap();
1981 1981 TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty);
1982 1982 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
1983 1983 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
1984 1984 "collections passed to foreground collector", _full_gcs_since_conc_gc);
1985 1985 }
1986 1986
1987 1987 // Sample collection interval time and reset for collection pause.
1988 1988 if (UseAdaptiveSizePolicy) {
1989 1989 size_policy()->msc_collection_begin();
1990 1990 }
1991 1991
1992 1992 // Temporarily widen the span of the weak reference processing to
1993 1993 // the entire heap.
1994 1994 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1995 1995 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1996 1996 // Temporarily, clear the "is_alive_non_header" field of the
|
↓ open down ↓ |
1996 lines elided |
↑ open up ↑ |
1997 1997 // reference processor.
1998 1998 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1999 1999 // Temporarily make reference _processing_ single threaded (non-MT).
2000 2000 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
2001 2001 // Temporarily make refs discovery atomic
2002 2002 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
2003 2003 // Temporarily make reference _discovery_ single threaded (non-MT)
2004 2004 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
2005 2005
2006 2006 ref_processor()->set_enqueuing_is_done(false);
2007 - ref_processor()->enable_discovery();
2007 + ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/);
2008 2008 ref_processor()->setup_policy(clear_all_soft_refs);
2009 2009 // If an asynchronous collection finishes, the _modUnionTable is
2010 2010 // all clear. If we are assuming the collection from an asynchronous
2011 2011 // collection, clear the _modUnionTable.
2012 2012 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
2013 2013 "_modUnionTable should be clear if the baton was not passed");
2014 2014 _modUnionTable.clear_all();
2015 2015
2016 2016 // We must adjust the allocation statistics being maintained
2017 2017 // in the free list space. We do so by reading and clearing
2018 2018 // the sweep timer and updating the block flux rate estimates below.
2019 2019 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
2020 2020 if (_inter_sweep_timer.is_active()) {
2021 2021 _inter_sweep_timer.stop();
2022 2022 // Note that we do not use this sample to update the _inter_sweep_estimate.
2023 2023 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
2024 2024 _inter_sweep_estimate.padded_average(),
2025 2025 _intra_sweep_estimate.padded_average());
2026 2026 }
2027 2027
2028 2028 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
2029 2029 ref_processor(), clear_all_soft_refs);
2030 2030 #ifdef ASSERT
2031 2031 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
2032 2032 size_t free_size = cms_space->free();
2033 2033 assert(free_size ==
2034 2034 pointer_delta(cms_space->end(), cms_space->compaction_top())
2035 2035 * HeapWordSize,
2036 2036 "All the free space should be compacted into one chunk at top");
2037 2037 assert(cms_space->dictionary()->totalChunkSize(
2038 2038 debug_only(cms_space->freelistLock())) == 0 ||
2039 2039 cms_space->totalSizeInIndexedFreeLists() == 0,
2040 2040 "All the free space should be in a single chunk");
2041 2041 size_t num = cms_space->totalCount();
2042 2042 assert((free_size == 0 && num == 0) ||
2043 2043 (free_size > 0 && (num == 1 || num == 2)),
2044 2044 "There should be at most 2 free chunks after compaction");
2045 2045 #endif // ASSERT
2046 2046 _collectorState = Resetting;
2047 2047 assert(_restart_addr == NULL,
2048 2048 "Should have been NULL'd before baton was passed");
2049 2049 reset(false /* == !asynch */);
2050 2050 _cmsGen->reset_after_compaction();
2051 2051 _concurrent_cycles_since_last_unload = 0;
2052 2052
2053 2053 if (verifying() && !should_unload_classes()) {
2054 2054 perm_gen_verify_bit_map()->clear_all();
2055 2055 }
2056 2056
2057 2057 // Clear any data recorded in the PLAB chunk arrays.
2058 2058 if (_survivor_plab_array != NULL) {
2059 2059 reset_survivor_plab_arrays();
2060 2060 }
2061 2061
2062 2062 // Adjust the per-size allocation stats for the next epoch.
2063 2063 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
2064 2064 // Restart the "inter sweep timer" for the next epoch.
2065 2065 _inter_sweep_timer.reset();
2066 2066 _inter_sweep_timer.start();
2067 2067
2068 2068 // Sample collection pause time and reset for collection interval.
2069 2069 if (UseAdaptiveSizePolicy) {
2070 2070 size_policy()->msc_collection_end(gch->gc_cause());
2071 2071 }
2072 2072
2073 2073 // For a mark-sweep-compact, compute_new_size() will be called
2074 2074 // in the heap's do_collection() method.
2075 2075 }
2076 2076
2077 2077 // A work method used by the foreground collector to do
2078 2078 // a mark-sweep, after taking over from a possibly on-going
2079 2079 // concurrent mark-sweep collection.
2080 2080 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2081 2081 CollectorState first_state, bool should_start_over) {
2082 2082 if (PrintGC && Verbose) {
2083 2083 gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2084 2084 "collector with count %d",
2085 2085 _full_gcs_since_conc_gc);
2086 2086 }
2087 2087 switch (_collectorState) {
2088 2088 case Idling:
2089 2089 if (first_state == Idling || should_start_over) {
2090 2090 // The background GC was not active, or should
2091 2091 // restarted from scratch; start the cycle.
2092 2092 _collectorState = InitialMarking;
2093 2093 }
2094 2094 // If first_state was not Idling, then a background GC
2095 2095 // was in progress and has now finished. No need to do it
2096 2096 // again. Leave the state as Idling.
2097 2097 break;
2098 2098 case Precleaning:
2099 2099 // In the foreground case don't do the precleaning since
2100 2100 // it is not done concurrently and there is extra work
2101 2101 // required.
2102 2102 _collectorState = FinalMarking;
2103 2103 }
2104 2104 if (PrintGCDetails &&
2105 2105 (_collectorState > Idling ||
2106 2106 !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) {
2107 2107 gclog_or_tty->print(" (concurrent mode failure)");
2108 2108 }
2109 2109 collect_in_foreground(clear_all_soft_refs);
2110 2110
2111 2111 // For a mark-sweep, compute_new_size() will be called
2112 2112 // in the heap's do_collection() method.
2113 2113 }
2114 2114
2115 2115
2116 2116 void CMSCollector::getFreelistLocks() const {
2117 2117 // Get locks for all free lists in all generations that this
2118 2118 // collector is responsible for
2119 2119 _cmsGen->freelistLock()->lock_without_safepoint_check();
2120 2120 _permGen->freelistLock()->lock_without_safepoint_check();
2121 2121 }
2122 2122
2123 2123 void CMSCollector::releaseFreelistLocks() const {
2124 2124 // Release locks for all free lists in all generations that this
2125 2125 // collector is responsible for
2126 2126 _cmsGen->freelistLock()->unlock();
2127 2127 _permGen->freelistLock()->unlock();
2128 2128 }
2129 2129
2130 2130 bool CMSCollector::haveFreelistLocks() const {
2131 2131 // Check locks for all free lists in all generations that this
2132 2132 // collector is responsible for
2133 2133 assert_lock_strong(_cmsGen->freelistLock());
2134 2134 assert_lock_strong(_permGen->freelistLock());
2135 2135 PRODUCT_ONLY(ShouldNotReachHere());
2136 2136 return true;
2137 2137 }
2138 2138
2139 2139 // A utility class that is used by the CMS collector to
2140 2140 // temporarily "release" the foreground collector from its
2141 2141 // usual obligation to wait for the background collector to
2142 2142 // complete an ongoing phase before proceeding.
2143 2143 class ReleaseForegroundGC: public StackObj {
2144 2144 private:
2145 2145 CMSCollector* _c;
2146 2146 public:
2147 2147 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2148 2148 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2149 2149 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2150 2150 // allow a potentially blocked foreground collector to proceed
2151 2151 _c->_foregroundGCShouldWait = false;
2152 2152 if (_c->_foregroundGCIsActive) {
2153 2153 CGC_lock->notify();
2154 2154 }
2155 2155 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2156 2156 "Possible deadlock");
2157 2157 }
2158 2158
2159 2159 ~ReleaseForegroundGC() {
2160 2160 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2161 2161 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2162 2162 _c->_foregroundGCShouldWait = true;
2163 2163 }
2164 2164 };
2165 2165
2166 2166 // There are separate collect_in_background and collect_in_foreground because of
2167 2167 // the different locking requirements of the background collector and the
2168 2168 // foreground collector. There was originally an attempt to share
2169 2169 // one "collect" method between the background collector and the foreground
2170 2170 // collector but the if-then-else required made it cleaner to have
2171 2171 // separate methods.
2172 2172 void CMSCollector::collect_in_background(bool clear_all_soft_refs) {
2173 2173 assert(Thread::current()->is_ConcurrentGC_thread(),
2174 2174 "A CMS asynchronous collection is only allowed on a CMS thread.");
2175 2175
2176 2176 GenCollectedHeap* gch = GenCollectedHeap::heap();
2177 2177 {
2178 2178 bool safepoint_check = Mutex::_no_safepoint_check_flag;
2179 2179 MutexLockerEx hl(Heap_lock, safepoint_check);
2180 2180 FreelistLocker fll(this);
2181 2181 MutexLockerEx x(CGC_lock, safepoint_check);
2182 2182 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2183 2183 // The foreground collector is active or we're
2184 2184 // not using asynchronous collections. Skip this
2185 2185 // background collection.
2186 2186 assert(!_foregroundGCShouldWait, "Should be clear");
2187 2187 return;
2188 2188 } else {
2189 2189 assert(_collectorState == Idling, "Should be idling before start.");
2190 2190 _collectorState = InitialMarking;
2191 2191 // Reset the expansion cause, now that we are about to begin
2192 2192 // a new cycle.
2193 2193 clear_expansion_cause();
2194 2194 }
2195 2195 // Decide if we want to enable class unloading as part of the
2196 2196 // ensuing concurrent GC cycle.
2197 2197 update_should_unload_classes();
2198 2198 _full_gc_requested = false; // acks all outstanding full gc requests
2199 2199 // Signal that we are about to start a collection
2200 2200 gch->increment_total_full_collections(); // ... starting a collection cycle
2201 2201 _collection_count_start = gch->total_full_collections();
2202 2202 }
2203 2203
2204 2204 // Used for PrintGC
2205 2205 size_t prev_used;
2206 2206 if (PrintGC && Verbose) {
2207 2207 prev_used = _cmsGen->used(); // XXXPERM
2208 2208 }
2209 2209
2210 2210 // The change of the collection state is normally done at this level;
2211 2211 // the exceptions are phases that are executed while the world is
2212 2212 // stopped. For those phases the change of state is done while the
2213 2213 // world is stopped. For baton passing purposes this allows the
2214 2214 // background collector to finish the phase and change state atomically.
2215 2215 // The foreground collector cannot wait on a phase that is done
2216 2216 // while the world is stopped because the foreground collector already
2217 2217 // has the world stopped and would deadlock.
2218 2218 while (_collectorState != Idling) {
2219 2219 if (TraceCMSState) {
2220 2220 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2221 2221 Thread::current(), _collectorState);
2222 2222 }
2223 2223 // The foreground collector
2224 2224 // holds the Heap_lock throughout its collection.
2225 2225 // holds the CMS token (but not the lock)
2226 2226 // except while it is waiting for the background collector to yield.
2227 2227 //
2228 2228 // The foreground collector should be blocked (not for long)
2229 2229 // if the background collector is about to start a phase
2230 2230 // executed with world stopped. If the background
2231 2231 // collector has already started such a phase, the
2232 2232 // foreground collector is blocked waiting for the
2233 2233 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
2234 2234 // are executed in the VM thread.
2235 2235 //
2236 2236 // The locking order is
2237 2237 // PendingListLock (PLL) -- if applicable (FinalMarking)
2238 2238 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
2239 2239 // CMS token (claimed in
2240 2240 // stop_world_and_do() -->
2241 2241 // safepoint_synchronize() -->
2242 2242 // CMSThread::synchronize())
2243 2243
2244 2244 {
2245 2245 // Check if the FG collector wants us to yield.
2246 2246 CMSTokenSync x(true); // is cms thread
2247 2247 if (waitForForegroundGC()) {
2248 2248 // We yielded to a foreground GC, nothing more to be
2249 2249 // done this round.
2250 2250 assert(_foregroundGCShouldWait == false, "We set it to false in "
2251 2251 "waitForForegroundGC()");
2252 2252 if (TraceCMSState) {
2253 2253 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2254 2254 " exiting collection CMS state %d",
2255 2255 Thread::current(), _collectorState);
2256 2256 }
2257 2257 return;
2258 2258 } else {
2259 2259 // The background collector can run but check to see if the
2260 2260 // foreground collector has done a collection while the
2261 2261 // background collector was waiting to get the CGC_lock
2262 2262 // above. If yes, break so that _foregroundGCShouldWait
2263 2263 // is cleared before returning.
2264 2264 if (_collectorState == Idling) {
2265 2265 break;
2266 2266 }
2267 2267 }
2268 2268 }
2269 2269
2270 2270 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2271 2271 "should be waiting");
2272 2272
2273 2273 switch (_collectorState) {
2274 2274 case InitialMarking:
2275 2275 {
2276 2276 ReleaseForegroundGC x(this);
2277 2277 stats().record_cms_begin();
2278 2278
2279 2279 VM_CMS_Initial_Mark initial_mark_op(this);
2280 2280 VMThread::execute(&initial_mark_op);
2281 2281 }
2282 2282 // The collector state may be any legal state at this point
2283 2283 // since the background collector may have yielded to the
2284 2284 // foreground collector.
2285 2285 break;
2286 2286 case Marking:
2287 2287 // initial marking in checkpointRootsInitialWork has been completed
2288 2288 if (markFromRoots(true)) { // we were successful
2289 2289 assert(_collectorState == Precleaning, "Collector state should "
2290 2290 "have changed");
2291 2291 } else {
2292 2292 assert(_foregroundGCIsActive, "Internal state inconsistency");
2293 2293 }
2294 2294 break;
2295 2295 case Precleaning:
2296 2296 if (UseAdaptiveSizePolicy) {
2297 2297 size_policy()->concurrent_precleaning_begin();
2298 2298 }
2299 2299 // marking from roots in markFromRoots has been completed
2300 2300 preclean();
2301 2301 if (UseAdaptiveSizePolicy) {
2302 2302 size_policy()->concurrent_precleaning_end();
2303 2303 }
2304 2304 assert(_collectorState == AbortablePreclean ||
2305 2305 _collectorState == FinalMarking,
2306 2306 "Collector state should have changed");
2307 2307 break;
2308 2308 case AbortablePreclean:
2309 2309 if (UseAdaptiveSizePolicy) {
2310 2310 size_policy()->concurrent_phases_resume();
2311 2311 }
2312 2312 abortable_preclean();
2313 2313 if (UseAdaptiveSizePolicy) {
2314 2314 size_policy()->concurrent_precleaning_end();
2315 2315 }
2316 2316 assert(_collectorState == FinalMarking, "Collector state should "
2317 2317 "have changed");
2318 2318 break;
2319 2319 case FinalMarking:
2320 2320 {
2321 2321 ReleaseForegroundGC x(this);
2322 2322
2323 2323 VM_CMS_Final_Remark final_remark_op(this);
2324 2324 VMThread::execute(&final_remark_op);
2325 2325 }
2326 2326 assert(_foregroundGCShouldWait, "block post-condition");
2327 2327 break;
2328 2328 case Sweeping:
2329 2329 if (UseAdaptiveSizePolicy) {
2330 2330 size_policy()->concurrent_sweeping_begin();
2331 2331 }
2332 2332 // final marking in checkpointRootsFinal has been completed
2333 2333 sweep(true);
2334 2334 assert(_collectorState == Resizing, "Collector state change "
2335 2335 "to Resizing must be done under the free_list_lock");
2336 2336 _full_gcs_since_conc_gc = 0;
2337 2337
2338 2338 // Stop the timers for adaptive size policy for the concurrent phases
2339 2339 if (UseAdaptiveSizePolicy) {
2340 2340 size_policy()->concurrent_sweeping_end();
2341 2341 size_policy()->concurrent_phases_end(gch->gc_cause(),
2342 2342 gch->prev_gen(_cmsGen)->capacity(),
2343 2343 _cmsGen->free());
2344 2344 }
2345 2345
2346 2346 case Resizing: {
2347 2347 // Sweeping has been completed...
2348 2348 // At this point the background collection has completed.
2349 2349 // Don't move the call to compute_new_size() down
2350 2350 // into code that might be executed if the background
2351 2351 // collection was preempted.
2352 2352 {
2353 2353 ReleaseForegroundGC x(this); // unblock FG collection
2354 2354 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
2355 2355 CMSTokenSync z(true); // not strictly needed.
2356 2356 if (_collectorState == Resizing) {
2357 2357 compute_new_size();
2358 2358 _collectorState = Resetting;
2359 2359 } else {
2360 2360 assert(_collectorState == Idling, "The state should only change"
2361 2361 " because the foreground collector has finished the collection");
2362 2362 }
2363 2363 }
2364 2364 break;
2365 2365 }
2366 2366 case Resetting:
2367 2367 // CMS heap resizing has been completed
2368 2368 reset(true);
2369 2369 assert(_collectorState == Idling, "Collector state should "
2370 2370 "have changed");
2371 2371 stats().record_cms_end();
2372 2372 // Don't move the concurrent_phases_end() and compute_new_size()
2373 2373 // calls to here because a preempted background collection
2374 2374 // has it's state set to "Resetting".
2375 2375 break;
2376 2376 case Idling:
2377 2377 default:
2378 2378 ShouldNotReachHere();
2379 2379 break;
2380 2380 }
2381 2381 if (TraceCMSState) {
2382 2382 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2383 2383 Thread::current(), _collectorState);
2384 2384 }
2385 2385 assert(_foregroundGCShouldWait, "block post-condition");
2386 2386 }
2387 2387
2388 2388 // Should this be in gc_epilogue?
2389 2389 collector_policy()->counters()->update_counters();
2390 2390
2391 2391 {
2392 2392 // Clear _foregroundGCShouldWait and, in the event that the
2393 2393 // foreground collector is waiting, notify it, before
2394 2394 // returning.
2395 2395 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2396 2396 _foregroundGCShouldWait = false;
2397 2397 if (_foregroundGCIsActive) {
2398 2398 CGC_lock->notify();
2399 2399 }
2400 2400 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2401 2401 "Possible deadlock");
2402 2402 }
2403 2403 if (TraceCMSState) {
2404 2404 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2405 2405 " exiting collection CMS state %d",
2406 2406 Thread::current(), _collectorState);
2407 2407 }
2408 2408 if (PrintGC && Verbose) {
2409 2409 _cmsGen->print_heap_change(prev_used);
2410 2410 }
2411 2411 }
2412 2412
2413 2413 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) {
2414 2414 assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2415 2415 "Foreground collector should be waiting, not executing");
2416 2416 assert(Thread::current()->is_VM_thread(), "A foreground collection"
2417 2417 "may only be done by the VM Thread with the world stopped");
2418 2418 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2419 2419 "VM thread should have CMS token");
2420 2420
2421 2421 NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2422 2422 true, gclog_or_tty);)
2423 2423 if (UseAdaptiveSizePolicy) {
2424 2424 size_policy()->ms_collection_begin();
2425 2425 }
2426 2426 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2427 2427
2428 2428 HandleMark hm; // Discard invalid handles created during verification
2429 2429
2430 2430 if (VerifyBeforeGC &&
2431 2431 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2432 2432 Universe::verify(true);
2433 2433 }
2434 2434
2435 2435 // Snapshot the soft reference policy to be used in this collection cycle.
2436 2436 ref_processor()->setup_policy(clear_all_soft_refs);
2437 2437
2438 2438 bool init_mark_was_synchronous = false; // until proven otherwise
2439 2439 while (_collectorState != Idling) {
2440 2440 if (TraceCMSState) {
2441 2441 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2442 2442 Thread::current(), _collectorState);
2443 2443 }
2444 2444 switch (_collectorState) {
2445 2445 case InitialMarking:
2446 2446 init_mark_was_synchronous = true; // fact to be exploited in re-mark
2447 2447 checkpointRootsInitial(false);
2448 2448 assert(_collectorState == Marking, "Collector state should have changed"
2449 2449 " within checkpointRootsInitial()");
2450 2450 break;
2451 2451 case Marking:
2452 2452 // initial marking in checkpointRootsInitialWork has been completed
2453 2453 if (VerifyDuringGC &&
2454 2454 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2455 2455 gclog_or_tty->print("Verify before initial mark: ");
2456 2456 Universe::verify(true);
2457 2457 }
2458 2458 {
2459 2459 bool res = markFromRoots(false);
2460 2460 assert(res && _collectorState == FinalMarking, "Collector state should "
2461 2461 "have changed");
2462 2462 break;
2463 2463 }
2464 2464 case FinalMarking:
2465 2465 if (VerifyDuringGC &&
2466 2466 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2467 2467 gclog_or_tty->print("Verify before re-mark: ");
2468 2468 Universe::verify(true);
2469 2469 }
2470 2470 checkpointRootsFinal(false, clear_all_soft_refs,
2471 2471 init_mark_was_synchronous);
2472 2472 assert(_collectorState == Sweeping, "Collector state should not "
2473 2473 "have changed within checkpointRootsFinal()");
2474 2474 break;
2475 2475 case Sweeping:
2476 2476 // final marking in checkpointRootsFinal has been completed
2477 2477 if (VerifyDuringGC &&
2478 2478 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2479 2479 gclog_or_tty->print("Verify before sweep: ");
2480 2480 Universe::verify(true);
2481 2481 }
2482 2482 sweep(false);
2483 2483 assert(_collectorState == Resizing, "Incorrect state");
2484 2484 break;
2485 2485 case Resizing: {
2486 2486 // Sweeping has been completed; the actual resize in this case
2487 2487 // is done separately; nothing to be done in this state.
2488 2488 _collectorState = Resetting;
2489 2489 break;
2490 2490 }
2491 2491 case Resetting:
2492 2492 // The heap has been resized.
2493 2493 if (VerifyDuringGC &&
2494 2494 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2495 2495 gclog_or_tty->print("Verify before reset: ");
2496 2496 Universe::verify(true);
2497 2497 }
2498 2498 reset(false);
2499 2499 assert(_collectorState == Idling, "Collector state should "
2500 2500 "have changed");
2501 2501 break;
2502 2502 case Precleaning:
2503 2503 case AbortablePreclean:
2504 2504 // Elide the preclean phase
2505 2505 _collectorState = FinalMarking;
2506 2506 break;
2507 2507 default:
2508 2508 ShouldNotReachHere();
2509 2509 }
2510 2510 if (TraceCMSState) {
2511 2511 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2512 2512 Thread::current(), _collectorState);
2513 2513 }
2514 2514 }
2515 2515
2516 2516 if (UseAdaptiveSizePolicy) {
2517 2517 GenCollectedHeap* gch = GenCollectedHeap::heap();
2518 2518 size_policy()->ms_collection_end(gch->gc_cause());
2519 2519 }
2520 2520
2521 2521 if (VerifyAfterGC &&
2522 2522 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2523 2523 Universe::verify(true);
2524 2524 }
2525 2525 if (TraceCMSState) {
2526 2526 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2527 2527 " exiting collection CMS state %d",
2528 2528 Thread::current(), _collectorState);
2529 2529 }
2530 2530 }
2531 2531
2532 2532 bool CMSCollector::waitForForegroundGC() {
2533 2533 bool res = false;
2534 2534 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2535 2535 "CMS thread should have CMS token");
2536 2536 // Block the foreground collector until the
2537 2537 // background collectors decides whether to
2538 2538 // yield.
2539 2539 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2540 2540 _foregroundGCShouldWait = true;
2541 2541 if (_foregroundGCIsActive) {
2542 2542 // The background collector yields to the
2543 2543 // foreground collector and returns a value
2544 2544 // indicating that it has yielded. The foreground
2545 2545 // collector can proceed.
2546 2546 res = true;
2547 2547 _foregroundGCShouldWait = false;
2548 2548 ConcurrentMarkSweepThread::clear_CMS_flag(
2549 2549 ConcurrentMarkSweepThread::CMS_cms_has_token);
2550 2550 ConcurrentMarkSweepThread::set_CMS_flag(
2551 2551 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2552 2552 // Get a possibly blocked foreground thread going
2553 2553 CGC_lock->notify();
2554 2554 if (TraceCMSState) {
2555 2555 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2556 2556 Thread::current(), _collectorState);
2557 2557 }
2558 2558 while (_foregroundGCIsActive) {
2559 2559 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2560 2560 }
2561 2561 ConcurrentMarkSweepThread::set_CMS_flag(
2562 2562 ConcurrentMarkSweepThread::CMS_cms_has_token);
2563 2563 ConcurrentMarkSweepThread::clear_CMS_flag(
2564 2564 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2565 2565 }
2566 2566 if (TraceCMSState) {
2567 2567 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2568 2568 Thread::current(), _collectorState);
2569 2569 }
2570 2570 return res;
2571 2571 }
2572 2572
2573 2573 // Because of the need to lock the free lists and other structures in
2574 2574 // the collector, common to all the generations that the collector is
2575 2575 // collecting, we need the gc_prologues of individual CMS generations
2576 2576 // delegate to their collector. It may have been simpler had the
2577 2577 // current infrastructure allowed one to call a prologue on a
2578 2578 // collector. In the absence of that we have the generation's
2579 2579 // prologue delegate to the collector, which delegates back
2580 2580 // some "local" work to a worker method in the individual generations
2581 2581 // that it's responsible for collecting, while itself doing any
2582 2582 // work common to all generations it's responsible for. A similar
2583 2583 // comment applies to the gc_epilogue()'s.
2584 2584 // The role of the varaible _between_prologue_and_epilogue is to
2585 2585 // enforce the invocation protocol.
2586 2586 void CMSCollector::gc_prologue(bool full) {
2587 2587 // Call gc_prologue_work() for each CMSGen and PermGen that
2588 2588 // we are responsible for.
2589 2589
2590 2590 // The following locking discipline assumes that we are only called
2591 2591 // when the world is stopped.
2592 2592 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2593 2593
2594 2594 // The CMSCollector prologue must call the gc_prologues for the
2595 2595 // "generations" (including PermGen if any) that it's responsible
2596 2596 // for.
2597 2597
2598 2598 assert( Thread::current()->is_VM_thread()
2599 2599 || ( CMSScavengeBeforeRemark
2600 2600 && Thread::current()->is_ConcurrentGC_thread()),
2601 2601 "Incorrect thread type for prologue execution");
2602 2602
2603 2603 if (_between_prologue_and_epilogue) {
2604 2604 // We have already been invoked; this is a gc_prologue delegation
2605 2605 // from yet another CMS generation that we are responsible for, just
2606 2606 // ignore it since all relevant work has already been done.
2607 2607 return;
2608 2608 }
2609 2609
2610 2610 // set a bit saying prologue has been called; cleared in epilogue
2611 2611 _between_prologue_and_epilogue = true;
2612 2612 // Claim locks for common data structures, then call gc_prologue_work()
2613 2613 // for each CMSGen and PermGen that we are responsible for.
2614 2614
2615 2615 getFreelistLocks(); // gets free list locks on constituent spaces
2616 2616 bitMapLock()->lock_without_safepoint_check();
2617 2617
2618 2618 // Should call gc_prologue_work() for all cms gens we are responsible for
2619 2619 bool registerClosure = _collectorState >= Marking
2620 2620 && _collectorState < Sweeping;
2621 2621 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ?
2622 2622 &_modUnionClosurePar
2623 2623 : &_modUnionClosure;
2624 2624 _cmsGen->gc_prologue_work(full, registerClosure, muc);
2625 2625 _permGen->gc_prologue_work(full, registerClosure, muc);
2626 2626
2627 2627 if (!full) {
2628 2628 stats().record_gc0_begin();
2629 2629 }
2630 2630 }
2631 2631
2632 2632 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2633 2633 // Delegate to CMScollector which knows how to coordinate between
2634 2634 // this and any other CMS generations that it is responsible for
2635 2635 // collecting.
2636 2636 collector()->gc_prologue(full);
2637 2637 }
2638 2638
2639 2639 // This is a "private" interface for use by this generation's CMSCollector.
2640 2640 // Not to be called directly by any other entity (for instance,
2641 2641 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2642 2642 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2643 2643 bool registerClosure, ModUnionClosure* modUnionClosure) {
2644 2644 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2645 2645 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2646 2646 "Should be NULL");
2647 2647 if (registerClosure) {
2648 2648 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2649 2649 }
2650 2650 cmsSpace()->gc_prologue();
2651 2651 // Clear stat counters
2652 2652 NOT_PRODUCT(
2653 2653 assert(_numObjectsPromoted == 0, "check");
2654 2654 assert(_numWordsPromoted == 0, "check");
2655 2655 if (Verbose && PrintGC) {
2656 2656 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2657 2657 SIZE_FORMAT" bytes concurrently",
2658 2658 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2659 2659 }
2660 2660 _numObjectsAllocated = 0;
2661 2661 _numWordsAllocated = 0;
2662 2662 )
2663 2663 }
2664 2664
2665 2665 void CMSCollector::gc_epilogue(bool full) {
2666 2666 // The following locking discipline assumes that we are only called
2667 2667 // when the world is stopped.
2668 2668 assert(SafepointSynchronize::is_at_safepoint(),
2669 2669 "world is stopped assumption");
2670 2670
2671 2671 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2672 2672 // if linear allocation blocks need to be appropriately marked to allow the
2673 2673 // the blocks to be parsable. We also check here whether we need to nudge the
2674 2674 // CMS collector thread to start a new cycle (if it's not already active).
2675 2675 assert( Thread::current()->is_VM_thread()
2676 2676 || ( CMSScavengeBeforeRemark
2677 2677 && Thread::current()->is_ConcurrentGC_thread()),
2678 2678 "Incorrect thread type for epilogue execution");
2679 2679
2680 2680 if (!_between_prologue_and_epilogue) {
2681 2681 // We have already been invoked; this is a gc_epilogue delegation
2682 2682 // from yet another CMS generation that we are responsible for, just
2683 2683 // ignore it since all relevant work has already been done.
2684 2684 return;
2685 2685 }
2686 2686 assert(haveFreelistLocks(), "must have freelist locks");
2687 2687 assert_lock_strong(bitMapLock());
2688 2688
2689 2689 _cmsGen->gc_epilogue_work(full);
2690 2690 _permGen->gc_epilogue_work(full);
2691 2691
2692 2692 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2693 2693 // in case sampling was not already enabled, enable it
2694 2694 _start_sampling = true;
2695 2695 }
2696 2696 // reset _eden_chunk_array so sampling starts afresh
2697 2697 _eden_chunk_index = 0;
2698 2698
2699 2699 size_t cms_used = _cmsGen->cmsSpace()->used();
2700 2700 size_t perm_used = _permGen->cmsSpace()->used();
2701 2701
2702 2702 // update performance counters - this uses a special version of
2703 2703 // update_counters() that allows the utilization to be passed as a
2704 2704 // parameter, avoiding multiple calls to used().
2705 2705 //
2706 2706 _cmsGen->update_counters(cms_used);
2707 2707 _permGen->update_counters(perm_used);
2708 2708
2709 2709 if (CMSIncrementalMode) {
2710 2710 icms_update_allocation_limits();
2711 2711 }
2712 2712
2713 2713 bitMapLock()->unlock();
2714 2714 releaseFreelistLocks();
2715 2715
2716 2716 if (!CleanChunkPoolAsync) {
2717 2717 Chunk::clean_chunk_pool();
2718 2718 }
2719 2719
2720 2720 _between_prologue_and_epilogue = false; // ready for next cycle
2721 2721 }
2722 2722
2723 2723 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2724 2724 collector()->gc_epilogue(full);
2725 2725
2726 2726 // Also reset promotion tracking in par gc thread states.
2727 2727 if (CollectedHeap::use_parallel_gc_threads()) {
2728 2728 for (uint i = 0; i < ParallelGCThreads; i++) {
2729 2729 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2730 2730 }
2731 2731 }
2732 2732 }
2733 2733
2734 2734 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2735 2735 assert(!incremental_collection_failed(), "Should have been cleared");
2736 2736 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2737 2737 cmsSpace()->gc_epilogue();
2738 2738 // Print stat counters
2739 2739 NOT_PRODUCT(
2740 2740 assert(_numObjectsAllocated == 0, "check");
2741 2741 assert(_numWordsAllocated == 0, "check");
2742 2742 if (Verbose && PrintGC) {
2743 2743 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2744 2744 SIZE_FORMAT" bytes",
2745 2745 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2746 2746 }
2747 2747 _numObjectsPromoted = 0;
2748 2748 _numWordsPromoted = 0;
2749 2749 )
2750 2750
2751 2751 if (PrintGC && Verbose) {
2752 2752 // Call down the chain in contiguous_available needs the freelistLock
2753 2753 // so print this out before releasing the freeListLock.
2754 2754 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2755 2755 contiguous_available());
2756 2756 }
2757 2757 }
2758 2758
2759 2759 #ifndef PRODUCT
2760 2760 bool CMSCollector::have_cms_token() {
2761 2761 Thread* thr = Thread::current();
2762 2762 if (thr->is_VM_thread()) {
2763 2763 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2764 2764 } else if (thr->is_ConcurrentGC_thread()) {
2765 2765 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2766 2766 } else if (thr->is_GC_task_thread()) {
2767 2767 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2768 2768 ParGCRareEvent_lock->owned_by_self();
2769 2769 }
2770 2770 return false;
2771 2771 }
2772 2772 #endif
2773 2773
2774 2774 // Check reachability of the given heap address in CMS generation,
2775 2775 // treating all other generations as roots.
2776 2776 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2777 2777 // We could "guarantee" below, rather than assert, but i'll
2778 2778 // leave these as "asserts" so that an adventurous debugger
2779 2779 // could try this in the product build provided some subset of
2780 2780 // the conditions were met, provided they were intersted in the
2781 2781 // results and knew that the computation below wouldn't interfere
2782 2782 // with other concurrent computations mutating the structures
2783 2783 // being read or written.
2784 2784 assert(SafepointSynchronize::is_at_safepoint(),
2785 2785 "Else mutations in object graph will make answer suspect");
2786 2786 assert(have_cms_token(), "Should hold cms token");
2787 2787 assert(haveFreelistLocks(), "must hold free list locks");
2788 2788 assert_lock_strong(bitMapLock());
2789 2789
2790 2790 // Clear the marking bit map array before starting, but, just
2791 2791 // for kicks, first report if the given address is already marked
2792 2792 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
2793 2793 _markBitMap.isMarked(addr) ? "" : " not");
2794 2794
2795 2795 if (verify_after_remark()) {
2796 2796 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2797 2797 bool result = verification_mark_bm()->isMarked(addr);
2798 2798 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
2799 2799 result ? "IS" : "is NOT");
2800 2800 return result;
2801 2801 } else {
2802 2802 gclog_or_tty->print_cr("Could not compute result");
2803 2803 return false;
2804 2804 }
2805 2805 }
2806 2806
2807 2807 ////////////////////////////////////////////////////////
2808 2808 // CMS Verification Support
2809 2809 ////////////////////////////////////////////////////////
2810 2810 // Following the remark phase, the following invariant
2811 2811 // should hold -- each object in the CMS heap which is
2812 2812 // marked in markBitMap() should be marked in the verification_mark_bm().
2813 2813
2814 2814 class VerifyMarkedClosure: public BitMapClosure {
2815 2815 CMSBitMap* _marks;
2816 2816 bool _failed;
2817 2817
2818 2818 public:
2819 2819 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2820 2820
2821 2821 bool do_bit(size_t offset) {
2822 2822 HeapWord* addr = _marks->offsetToHeapWord(offset);
2823 2823 if (!_marks->isMarked(addr)) {
2824 2824 oop(addr)->print_on(gclog_or_tty);
2825 2825 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2826 2826 _failed = true;
2827 2827 }
2828 2828 return true;
2829 2829 }
2830 2830
2831 2831 bool failed() { return _failed; }
2832 2832 };
2833 2833
2834 2834 bool CMSCollector::verify_after_remark() {
2835 2835 gclog_or_tty->print(" [Verifying CMS Marking... ");
2836 2836 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2837 2837 static bool init = false;
2838 2838
2839 2839 assert(SafepointSynchronize::is_at_safepoint(),
2840 2840 "Else mutations in object graph will make answer suspect");
2841 2841 assert(have_cms_token(),
2842 2842 "Else there may be mutual interference in use of "
2843 2843 " verification data structures");
2844 2844 assert(_collectorState > Marking && _collectorState <= Sweeping,
2845 2845 "Else marking info checked here may be obsolete");
2846 2846 assert(haveFreelistLocks(), "must hold free list locks");
2847 2847 assert_lock_strong(bitMapLock());
2848 2848
2849 2849
2850 2850 // Allocate marking bit map if not already allocated
2851 2851 if (!init) { // first time
2852 2852 if (!verification_mark_bm()->allocate(_span)) {
2853 2853 return false;
2854 2854 }
2855 2855 init = true;
2856 2856 }
2857 2857
2858 2858 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2859 2859
2860 2860 // Turn off refs discovery -- so we will be tracing through refs.
2861 2861 // This is as intended, because by this time
2862 2862 // GC must already have cleared any refs that need to be cleared,
2863 2863 // and traced those that need to be marked; moreover,
2864 2864 // the marking done here is not going to intefere in any
2865 2865 // way with the marking information used by GC.
2866 2866 NoRefDiscovery no_discovery(ref_processor());
2867 2867
2868 2868 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2869 2869
2870 2870 // Clear any marks from a previous round
2871 2871 verification_mark_bm()->clear_all();
2872 2872 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2873 2873 verify_work_stacks_empty();
2874 2874
2875 2875 GenCollectedHeap* gch = GenCollectedHeap::heap();
2876 2876 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2877 2877 // Update the saved marks which may affect the root scans.
2878 2878 gch->save_marks();
2879 2879
2880 2880 if (CMSRemarkVerifyVariant == 1) {
2881 2881 // In this first variant of verification, we complete
2882 2882 // all marking, then check if the new marks-verctor is
2883 2883 // a subset of the CMS marks-vector.
2884 2884 verify_after_remark_work_1();
2885 2885 } else if (CMSRemarkVerifyVariant == 2) {
2886 2886 // In this second variant of verification, we flag an error
2887 2887 // (i.e. an object reachable in the new marks-vector not reachable
2888 2888 // in the CMS marks-vector) immediately, also indicating the
2889 2889 // identify of an object (A) that references the unmarked object (B) --
2890 2890 // presumably, a mutation to A failed to be picked up by preclean/remark?
2891 2891 verify_after_remark_work_2();
2892 2892 } else {
2893 2893 warning("Unrecognized value %d for CMSRemarkVerifyVariant",
2894 2894 CMSRemarkVerifyVariant);
2895 2895 }
2896 2896 gclog_or_tty->print(" done] ");
2897 2897 return true;
2898 2898 }
2899 2899
2900 2900 void CMSCollector::verify_after_remark_work_1() {
2901 2901 ResourceMark rm;
2902 2902 HandleMark hm;
2903 2903 GenCollectedHeap* gch = GenCollectedHeap::heap();
2904 2904
2905 2905 // Mark from roots one level into CMS
2906 2906 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2907 2907 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2908 2908
2909 2909 gch->gen_process_strong_roots(_cmsGen->level(),
2910 2910 true, // younger gens are roots
2911 2911 true, // activate StrongRootsScope
2912 2912 true, // collecting perm gen
2913 2913 SharedHeap::ScanningOption(roots_scanning_options()),
2914 2914 ¬Older,
2915 2915 true, // walk code active on stacks
2916 2916 NULL);
2917 2917
2918 2918 // Now mark from the roots
2919 2919 assert(_revisitStack.isEmpty(), "Should be empty");
2920 2920 MarkFromRootsClosure markFromRootsClosure(this, _span,
2921 2921 verification_mark_bm(), verification_mark_stack(), &_revisitStack,
2922 2922 false /* don't yield */, true /* verifying */);
2923 2923 assert(_restart_addr == NULL, "Expected pre-condition");
2924 2924 verification_mark_bm()->iterate(&markFromRootsClosure);
2925 2925 while (_restart_addr != NULL) {
2926 2926 // Deal with stack overflow: by restarting at the indicated
2927 2927 // address.
2928 2928 HeapWord* ra = _restart_addr;
2929 2929 markFromRootsClosure.reset(ra);
2930 2930 _restart_addr = NULL;
2931 2931 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2932 2932 }
2933 2933 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2934 2934 verify_work_stacks_empty();
2935 2935 // Should reset the revisit stack above, since no class tree
2936 2936 // surgery is forthcoming.
2937 2937 _revisitStack.reset(); // throwing away all contents
2938 2938
2939 2939 // Marking completed -- now verify that each bit marked in
2940 2940 // verification_mark_bm() is also marked in markBitMap(); flag all
2941 2941 // errors by printing corresponding objects.
2942 2942 VerifyMarkedClosure vcl(markBitMap());
2943 2943 verification_mark_bm()->iterate(&vcl);
2944 2944 if (vcl.failed()) {
2945 2945 gclog_or_tty->print("Verification failed");
2946 2946 Universe::heap()->print_on(gclog_or_tty);
2947 2947 fatal("CMS: failed marking verification after remark");
2948 2948 }
2949 2949 }
2950 2950
2951 2951 void CMSCollector::verify_after_remark_work_2() {
2952 2952 ResourceMark rm;
2953 2953 HandleMark hm;
2954 2954 GenCollectedHeap* gch = GenCollectedHeap::heap();
2955 2955
2956 2956 // Mark from roots one level into CMS
2957 2957 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2958 2958 markBitMap());
2959 2959 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2960 2960 gch->gen_process_strong_roots(_cmsGen->level(),
2961 2961 true, // younger gens are roots
2962 2962 true, // activate StrongRootsScope
2963 2963 true, // collecting perm gen
2964 2964 SharedHeap::ScanningOption(roots_scanning_options()),
2965 2965 ¬Older,
2966 2966 true, // walk code active on stacks
2967 2967 NULL);
2968 2968
2969 2969 // Now mark from the roots
2970 2970 assert(_revisitStack.isEmpty(), "Should be empty");
2971 2971 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2972 2972 verification_mark_bm(), markBitMap(), verification_mark_stack());
2973 2973 assert(_restart_addr == NULL, "Expected pre-condition");
2974 2974 verification_mark_bm()->iterate(&markFromRootsClosure);
2975 2975 while (_restart_addr != NULL) {
2976 2976 // Deal with stack overflow: by restarting at the indicated
2977 2977 // address.
2978 2978 HeapWord* ra = _restart_addr;
2979 2979 markFromRootsClosure.reset(ra);
2980 2980 _restart_addr = NULL;
2981 2981 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2982 2982 }
2983 2983 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2984 2984 verify_work_stacks_empty();
2985 2985 // Should reset the revisit stack above, since no class tree
2986 2986 // surgery is forthcoming.
2987 2987 _revisitStack.reset(); // throwing away all contents
2988 2988
2989 2989 // Marking completed -- now verify that each bit marked in
2990 2990 // verification_mark_bm() is also marked in markBitMap(); flag all
2991 2991 // errors by printing corresponding objects.
2992 2992 VerifyMarkedClosure vcl(markBitMap());
2993 2993 verification_mark_bm()->iterate(&vcl);
2994 2994 assert(!vcl.failed(), "Else verification above should not have succeeded");
2995 2995 }
2996 2996
2997 2997 void ConcurrentMarkSweepGeneration::save_marks() {
2998 2998 // delegate to CMS space
2999 2999 cmsSpace()->save_marks();
3000 3000 for (uint i = 0; i < ParallelGCThreads; i++) {
3001 3001 _par_gc_thread_states[i]->promo.startTrackingPromotions();
3002 3002 }
3003 3003 }
3004 3004
3005 3005 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
3006 3006 return cmsSpace()->no_allocs_since_save_marks();
3007 3007 }
3008 3008
3009 3009 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
3010 3010 \
3011 3011 void ConcurrentMarkSweepGeneration:: \
3012 3012 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
3013 3013 cl->set_generation(this); \
3014 3014 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
3015 3015 cl->reset_generation(); \
3016 3016 save_marks(); \
3017 3017 }
3018 3018
3019 3019 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
3020 3020
3021 3021 void
3022 3022 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk)
3023 3023 {
3024 3024 // Not currently implemented; need to do the following. -- ysr.
3025 3025 // dld -- I think that is used for some sort of allocation profiler. So it
3026 3026 // really means the objects allocated by the mutator since the last
3027 3027 // GC. We could potentially implement this cheaply by recording only
3028 3028 // the direct allocations in a side data structure.
3029 3029 //
3030 3030 // I think we probably ought not to be required to support these
3031 3031 // iterations at any arbitrary point; I think there ought to be some
3032 3032 // call to enable/disable allocation profiling in a generation/space,
3033 3033 // and the iterator ought to return the objects allocated in the
3034 3034 // gen/space since the enable call, or the last iterator call (which
3035 3035 // will probably be at a GC.) That way, for gens like CM&S that would
3036 3036 // require some extra data structure to support this, we only pay the
3037 3037 // cost when it's in use...
3038 3038 cmsSpace()->object_iterate_since_last_GC(blk);
3039 3039 }
3040 3040
3041 3041 void
3042 3042 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
3043 3043 cl->set_generation(this);
3044 3044 younger_refs_in_space_iterate(_cmsSpace, cl);
3045 3045 cl->reset_generation();
3046 3046 }
3047 3047
3048 3048 void
3049 3049 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) {
3050 3050 if (freelistLock()->owned_by_self()) {
3051 3051 Generation::oop_iterate(mr, cl);
3052 3052 } else {
3053 3053 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3054 3054 Generation::oop_iterate(mr, cl);
3055 3055 }
3056 3056 }
3057 3057
3058 3058 void
3059 3059 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) {
3060 3060 if (freelistLock()->owned_by_self()) {
3061 3061 Generation::oop_iterate(cl);
3062 3062 } else {
3063 3063 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3064 3064 Generation::oop_iterate(cl);
3065 3065 }
3066 3066 }
3067 3067
3068 3068 void
3069 3069 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
3070 3070 if (freelistLock()->owned_by_self()) {
3071 3071 Generation::object_iterate(cl);
3072 3072 } else {
3073 3073 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3074 3074 Generation::object_iterate(cl);
3075 3075 }
3076 3076 }
3077 3077
3078 3078 void
3079 3079 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
3080 3080 if (freelistLock()->owned_by_self()) {
3081 3081 Generation::safe_object_iterate(cl);
3082 3082 } else {
3083 3083 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3084 3084 Generation::safe_object_iterate(cl);
3085 3085 }
3086 3086 }
3087 3087
3088 3088 void
3089 3089 ConcurrentMarkSweepGeneration::pre_adjust_pointers() {
3090 3090 }
3091 3091
3092 3092 void
3093 3093 ConcurrentMarkSweepGeneration::post_compact() {
3094 3094 }
3095 3095
3096 3096 void
3097 3097 ConcurrentMarkSweepGeneration::prepare_for_verify() {
3098 3098 // Fix the linear allocation blocks to look like free blocks.
3099 3099
3100 3100 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3101 3101 // are not called when the heap is verified during universe initialization and
3102 3102 // at vm shutdown.
3103 3103 if (freelistLock()->owned_by_self()) {
3104 3104 cmsSpace()->prepare_for_verify();
3105 3105 } else {
3106 3106 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3107 3107 cmsSpace()->prepare_for_verify();
3108 3108 }
3109 3109 }
3110 3110
3111 3111 void
3112 3112 ConcurrentMarkSweepGeneration::verify(bool allow_dirty /* ignored */) {
3113 3113 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3114 3114 // are not called when the heap is verified during universe initialization and
3115 3115 // at vm shutdown.
3116 3116 if (freelistLock()->owned_by_self()) {
3117 3117 cmsSpace()->verify(false /* ignored */);
3118 3118 } else {
3119 3119 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3120 3120 cmsSpace()->verify(false /* ignored */);
3121 3121 }
3122 3122 }
3123 3123
3124 3124 void CMSCollector::verify(bool allow_dirty /* ignored */) {
3125 3125 _cmsGen->verify(allow_dirty);
3126 3126 _permGen->verify(allow_dirty);
3127 3127 }
3128 3128
3129 3129 #ifndef PRODUCT
3130 3130 bool CMSCollector::overflow_list_is_empty() const {
3131 3131 assert(_num_par_pushes >= 0, "Inconsistency");
3132 3132 if (_overflow_list == NULL) {
3133 3133 assert(_num_par_pushes == 0, "Inconsistency");
3134 3134 }
3135 3135 return _overflow_list == NULL;
3136 3136 }
3137 3137
3138 3138 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3139 3139 // merely consolidate assertion checks that appear to occur together frequently.
3140 3140 void CMSCollector::verify_work_stacks_empty() const {
3141 3141 assert(_markStack.isEmpty(), "Marking stack should be empty");
3142 3142 assert(overflow_list_is_empty(), "Overflow list should be empty");
3143 3143 }
3144 3144
3145 3145 void CMSCollector::verify_overflow_empty() const {
3146 3146 assert(overflow_list_is_empty(), "Overflow list should be empty");
3147 3147 assert(no_preserved_marks(), "No preserved marks");
3148 3148 }
3149 3149 #endif // PRODUCT
3150 3150
3151 3151 // Decide if we want to enable class unloading as part of the
3152 3152 // ensuing concurrent GC cycle. We will collect the perm gen and
3153 3153 // unload classes if it's the case that:
3154 3154 // (1) an explicit gc request has been made and the flag
3155 3155 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3156 3156 // (2) (a) class unloading is enabled at the command line, and
3157 3157 // (b) (i) perm gen threshold has been crossed, or
3158 3158 // (ii) old gen is getting really full, or
3159 3159 // (iii) the previous N CMS collections did not collect the
3160 3160 // perm gen
3161 3161 // NOTE: Provided there is no change in the state of the heap between
3162 3162 // calls to this method, it should have idempotent results. Moreover,
3163 3163 // its results should be monotonically increasing (i.e. going from 0 to 1,
3164 3164 // but not 1 to 0) between successive calls between which the heap was
3165 3165 // not collected. For the implementation below, it must thus rely on
3166 3166 // the property that concurrent_cycles_since_last_unload()
3167 3167 // will not decrease unless a collection cycle happened and that
3168 3168 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are
3169 3169 // themselves also monotonic in that sense. See check_monotonicity()
3170 3170 // below.
3171 3171 bool CMSCollector::update_should_unload_classes() {
3172 3172 _should_unload_classes = false;
3173 3173 // Condition 1 above
3174 3174 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3175 3175 _should_unload_classes = true;
3176 3176 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3177 3177 // Disjuncts 2.b.(i,ii,iii) above
3178 3178 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3179 3179 CMSClassUnloadingMaxInterval)
3180 3180 || _permGen->should_concurrent_collect()
3181 3181 || _cmsGen->is_too_full();
3182 3182 }
3183 3183 return _should_unload_classes;
3184 3184 }
3185 3185
3186 3186 bool ConcurrentMarkSweepGeneration::is_too_full() const {
3187 3187 bool res = should_concurrent_collect();
3188 3188 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
3189 3189 return res;
3190 3190 }
3191 3191
3192 3192 void CMSCollector::setup_cms_unloading_and_verification_state() {
3193 3193 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3194 3194 || VerifyBeforeExit;
3195 3195 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
3196 3196
3197 3197 if (should_unload_classes()) { // Should unload classes this cycle
3198 3198 remove_root_scanning_option(rso); // Shrink the root set appropriately
3199 3199 set_verifying(should_verify); // Set verification state for this cycle
3200 3200 return; // Nothing else needs to be done at this time
3201 3201 }
3202 3202
3203 3203 // Not unloading classes this cycle
3204 3204 assert(!should_unload_classes(), "Inconsitency!");
3205 3205 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3206 3206 // We were not verifying, or we _were_ unloading classes in the last cycle,
3207 3207 // AND some verification options are enabled this cycle; in this case,
3208 3208 // we must make sure that the deadness map is allocated if not already so,
3209 3209 // and cleared (if already allocated previously --
3210 3210 // CMSBitMap::sizeInBits() is used to determine if it's allocated).
3211 3211 if (perm_gen_verify_bit_map()->sizeInBits() == 0) {
3212 3212 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) {
3213 3213 warning("Failed to allocate permanent generation verification CMS Bit Map;\n"
3214 3214 "permanent generation verification disabled");
3215 3215 return; // Note that we leave verification disabled, so we'll retry this
3216 3216 // allocation next cycle. We _could_ remember this failure
3217 3217 // and skip further attempts and permanently disable verification
3218 3218 // attempts if that is considered more desirable.
3219 3219 }
3220 3220 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()),
3221 3221 "_perm_gen_ver_bit_map inconsistency?");
3222 3222 } else {
3223 3223 perm_gen_verify_bit_map()->clear_all();
3224 3224 }
3225 3225 // Include symbols, strings and code cache elements to prevent their resurrection.
3226 3226 add_root_scanning_option(rso);
3227 3227 set_verifying(true);
3228 3228 } else if (verifying() && !should_verify) {
3229 3229 // We were verifying, but some verification flags got disabled.
3230 3230 set_verifying(false);
3231 3231 // Exclude symbols, strings and code cache elements from root scanning to
3232 3232 // reduce IM and RM pauses.
3233 3233 remove_root_scanning_option(rso);
3234 3234 }
3235 3235 }
3236 3236
3237 3237
3238 3238 #ifndef PRODUCT
3239 3239 HeapWord* CMSCollector::block_start(const void* p) const {
3240 3240 const HeapWord* addr = (HeapWord*)p;
3241 3241 if (_span.contains(p)) {
3242 3242 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3243 3243 return _cmsGen->cmsSpace()->block_start(p);
3244 3244 } else {
3245 3245 assert(_permGen->cmsSpace()->is_in_reserved(addr),
3246 3246 "Inconsistent _span?");
3247 3247 return _permGen->cmsSpace()->block_start(p);
3248 3248 }
3249 3249 }
3250 3250 return NULL;
3251 3251 }
3252 3252 #endif
3253 3253
3254 3254 HeapWord*
3255 3255 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3256 3256 bool tlab,
3257 3257 bool parallel) {
3258 3258 CMSSynchronousYieldRequest yr;
3259 3259 assert(!tlab, "Can't deal with TLAB allocation");
3260 3260 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3261 3261 expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3262 3262 CMSExpansionCause::_satisfy_allocation);
3263 3263 if (GCExpandToAllocateDelayMillis > 0) {
3264 3264 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3265 3265 }
3266 3266 return have_lock_and_allocate(word_size, tlab);
3267 3267 }
3268 3268
3269 3269 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3270 3270 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3271 3271 // to CardGeneration and share it...
3272 3272 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) {
3273 3273 return CardGeneration::expand(bytes, expand_bytes);
3274 3274 }
3275 3275
3276 3276 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3277 3277 CMSExpansionCause::Cause cause)
3278 3278 {
3279 3279
3280 3280 bool success = expand(bytes, expand_bytes);
3281 3281
3282 3282 // remember why we expanded; this information is used
3283 3283 // by shouldConcurrentCollect() when making decisions on whether to start
3284 3284 // a new CMS cycle.
3285 3285 if (success) {
3286 3286 set_expansion_cause(cause);
3287 3287 if (PrintGCDetails && Verbose) {
3288 3288 gclog_or_tty->print_cr("Expanded CMS gen for %s",
3289 3289 CMSExpansionCause::to_string(cause));
3290 3290 }
3291 3291 }
3292 3292 }
3293 3293
3294 3294 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3295 3295 HeapWord* res = NULL;
3296 3296 MutexLocker x(ParGCRareEvent_lock);
3297 3297 while (true) {
3298 3298 // Expansion by some other thread might make alloc OK now:
3299 3299 res = ps->lab.alloc(word_sz);
3300 3300 if (res != NULL) return res;
3301 3301 // If there's not enough expansion space available, give up.
3302 3302 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3303 3303 return NULL;
3304 3304 }
3305 3305 // Otherwise, we try expansion.
3306 3306 expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3307 3307 CMSExpansionCause::_allocate_par_lab);
3308 3308 // Now go around the loop and try alloc again;
3309 3309 // A competing par_promote might beat us to the expansion space,
3310 3310 // so we may go around the loop again if promotion fails agaion.
3311 3311 if (GCExpandToAllocateDelayMillis > 0) {
3312 3312 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3313 3313 }
3314 3314 }
3315 3315 }
3316 3316
3317 3317
3318 3318 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3319 3319 PromotionInfo* promo) {
3320 3320 MutexLocker x(ParGCRareEvent_lock);
3321 3321 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3322 3322 while (true) {
3323 3323 // Expansion by some other thread might make alloc OK now:
3324 3324 if (promo->ensure_spooling_space()) {
3325 3325 assert(promo->has_spooling_space(),
3326 3326 "Post-condition of successful ensure_spooling_space()");
3327 3327 return true;
3328 3328 }
3329 3329 // If there's not enough expansion space available, give up.
3330 3330 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3331 3331 return false;
3332 3332 }
3333 3333 // Otherwise, we try expansion.
3334 3334 expand(refill_size_bytes, MinHeapDeltaBytes,
3335 3335 CMSExpansionCause::_allocate_par_spooling_space);
3336 3336 // Now go around the loop and try alloc again;
3337 3337 // A competing allocation might beat us to the expansion space,
3338 3338 // so we may go around the loop again if allocation fails again.
3339 3339 if (GCExpandToAllocateDelayMillis > 0) {
3340 3340 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3341 3341 }
3342 3342 }
3343 3343 }
3344 3344
3345 3345
3346 3346
3347 3347 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3348 3348 assert_locked_or_safepoint(Heap_lock);
3349 3349 size_t size = ReservedSpace::page_align_size_down(bytes);
3350 3350 if (size > 0) {
3351 3351 shrink_by(size);
3352 3352 }
3353 3353 }
3354 3354
3355 3355 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3356 3356 assert_locked_or_safepoint(Heap_lock);
3357 3357 bool result = _virtual_space.expand_by(bytes);
3358 3358 if (result) {
3359 3359 HeapWord* old_end = _cmsSpace->end();
3360 3360 size_t new_word_size =
3361 3361 heap_word_size(_virtual_space.committed_size());
3362 3362 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3363 3363 _bts->resize(new_word_size); // resize the block offset shared array
3364 3364 Universe::heap()->barrier_set()->resize_covered_region(mr);
3365 3365 // Hmmmm... why doesn't CFLS::set_end verify locking?
3366 3366 // This is quite ugly; FIX ME XXX
3367 3367 _cmsSpace->assert_locked(freelistLock());
3368 3368 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3369 3369
3370 3370 // update the space and generation capacity counters
3371 3371 if (UsePerfData) {
3372 3372 _space_counters->update_capacity();
3373 3373 _gen_counters->update_all();
3374 3374 }
3375 3375
3376 3376 if (Verbose && PrintGC) {
3377 3377 size_t new_mem_size = _virtual_space.committed_size();
3378 3378 size_t old_mem_size = new_mem_size - bytes;
3379 3379 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK",
3380 3380 name(), old_mem_size/K, bytes/K, new_mem_size/K);
3381 3381 }
3382 3382 }
3383 3383 return result;
3384 3384 }
3385 3385
3386 3386 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3387 3387 assert_locked_or_safepoint(Heap_lock);
3388 3388 bool success = true;
3389 3389 const size_t remaining_bytes = _virtual_space.uncommitted_size();
3390 3390 if (remaining_bytes > 0) {
3391 3391 success = grow_by(remaining_bytes);
3392 3392 DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3393 3393 }
3394 3394 return success;
3395 3395 }
3396 3396
3397 3397 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3398 3398 assert_locked_or_safepoint(Heap_lock);
3399 3399 assert_lock_strong(freelistLock());
3400 3400 // XXX Fix when compaction is implemented.
3401 3401 warning("Shrinking of CMS not yet implemented");
3402 3402 return;
3403 3403 }
3404 3404
3405 3405
3406 3406 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3407 3407 // phases.
3408 3408 class CMSPhaseAccounting: public StackObj {
3409 3409 public:
3410 3410 CMSPhaseAccounting(CMSCollector *collector,
3411 3411 const char *phase,
3412 3412 bool print_cr = true);
3413 3413 ~CMSPhaseAccounting();
3414 3414
3415 3415 private:
3416 3416 CMSCollector *_collector;
3417 3417 const char *_phase;
3418 3418 elapsedTimer _wallclock;
3419 3419 bool _print_cr;
3420 3420
3421 3421 public:
3422 3422 // Not MT-safe; so do not pass around these StackObj's
3423 3423 // where they may be accessed by other threads.
3424 3424 jlong wallclock_millis() {
3425 3425 assert(_wallclock.is_active(), "Wall clock should not stop");
3426 3426 _wallclock.stop(); // to record time
3427 3427 jlong ret = _wallclock.milliseconds();
3428 3428 _wallclock.start(); // restart
3429 3429 return ret;
3430 3430 }
3431 3431 };
3432 3432
3433 3433 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3434 3434 const char *phase,
3435 3435 bool print_cr) :
3436 3436 _collector(collector), _phase(phase), _print_cr(print_cr) {
3437 3437
3438 3438 if (PrintCMSStatistics != 0) {
3439 3439 _collector->resetYields();
3440 3440 }
3441 3441 if (PrintGCDetails && PrintGCTimeStamps) {
3442 3442 gclog_or_tty->date_stamp(PrintGCDateStamps);
3443 3443 gclog_or_tty->stamp();
3444 3444 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]",
3445 3445 _collector->cmsGen()->short_name(), _phase);
3446 3446 }
3447 3447 _collector->resetTimer();
3448 3448 _wallclock.start();
3449 3449 _collector->startTimer();
3450 3450 }
3451 3451
3452 3452 CMSPhaseAccounting::~CMSPhaseAccounting() {
3453 3453 assert(_wallclock.is_active(), "Wall clock should not have stopped");
3454 3454 _collector->stopTimer();
3455 3455 _wallclock.stop();
3456 3456 if (PrintGCDetails) {
3457 3457 gclog_or_tty->date_stamp(PrintGCDateStamps);
3458 3458 if (PrintGCTimeStamps) {
3459 3459 gclog_or_tty->stamp();
3460 3460 gclog_or_tty->print(": ");
3461 3461 }
3462 3462 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3463 3463 _collector->cmsGen()->short_name(),
3464 3464 _phase, _collector->timerValue(), _wallclock.seconds());
3465 3465 if (_print_cr) {
3466 3466 gclog_or_tty->print_cr("");
3467 3467 }
3468 3468 if (PrintCMSStatistics != 0) {
3469 3469 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3470 3470 _collector->yields());
3471 3471 }
3472 3472 }
3473 3473 }
3474 3474
3475 3475 // CMS work
3476 3476
3477 3477 // Checkpoint the roots into this generation from outside
3478 3478 // this generation. [Note this initial checkpoint need only
3479 3479 // be approximate -- we'll do a catch up phase subsequently.]
3480 3480 void CMSCollector::checkpointRootsInitial(bool asynch) {
3481 3481 assert(_collectorState == InitialMarking, "Wrong collector state");
3482 3482 check_correct_thread_executing();
|
↓ open down ↓ |
1465 lines elided |
↑ open up ↑ |
3483 3483 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
3484 3484
3485 3485 ReferenceProcessor* rp = ref_processor();
3486 3486 SpecializationStats::clear();
3487 3487 assert(_restart_addr == NULL, "Control point invariant");
3488 3488 if (asynch) {
3489 3489 // acquire locks for subsequent manipulations
3490 3490 MutexLockerEx x(bitMapLock(),
3491 3491 Mutex::_no_safepoint_check_flag);
3492 3492 checkpointRootsInitialWork(asynch);
3493 - rp->verify_no_references_recorded();
3494 - rp->enable_discovery(); // enable ("weak") refs discovery
3493 + // enable ("weak") refs discovery
3494 + rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/);
3495 3495 _collectorState = Marking;
3496 3496 } else {
3497 3497 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3498 3498 // which recognizes if we are a CMS generation, and doesn't try to turn on
3499 3499 // discovery; verify that they aren't meddling.
3500 3500 assert(!rp->discovery_is_atomic(),
3501 3501 "incorrect setting of discovery predicate");
3502 3502 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3503 3503 "ref discovery for this generation kind");
3504 3504 // already have locks
3505 3505 checkpointRootsInitialWork(asynch);
3506 - rp->enable_discovery(); // now enable ("weak") refs discovery
3506 + // now enable ("weak") refs discovery
3507 + rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/);
3507 3508 _collectorState = Marking;
3508 3509 }
3509 3510 SpecializationStats::print();
3510 3511 }
3511 3512
3512 3513 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3513 3514 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3514 3515 assert(_collectorState == InitialMarking, "just checking");
3515 3516
3516 3517 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3517 3518 // precede our marking with a collection of all
3518 3519 // younger generations to keep floating garbage to a minimum.
3519 3520 // XXX: we won't do this for now -- it's an optimization to be done later.
3520 3521
3521 3522 // already have locks
3522 3523 assert_lock_strong(bitMapLock());
3523 3524 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3524 3525
3525 3526 // Setup the verification and class unloading state for this
3526 3527 // CMS collection cycle.
3527 3528 setup_cms_unloading_and_verification_state();
3528 3529
3529 3530 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork",
3530 3531 PrintGCDetails && Verbose, true, gclog_or_tty);)
3531 3532 if (UseAdaptiveSizePolicy) {
3532 3533 size_policy()->checkpoint_roots_initial_begin();
3533 3534 }
3534 3535
3535 3536 // Reset all the PLAB chunk arrays if necessary.
3536 3537 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3537 3538 reset_survivor_plab_arrays();
3538 3539 }
3539 3540
3540 3541 ResourceMark rm;
3541 3542 HandleMark hm;
3542 3543
3543 3544 FalseClosure falseClosure;
3544 3545 // In the case of a synchronous collection, we will elide the
3545 3546 // remark step, so it's important to catch all the nmethod oops
3546 3547 // in this step.
3547 3548 // The final 'true' flag to gen_process_strong_roots will ensure this.
3548 3549 // If 'async' is true, we can relax the nmethod tracing.
3549 3550 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
3550 3551 GenCollectedHeap* gch = GenCollectedHeap::heap();
3551 3552
3552 3553 verify_work_stacks_empty();
3553 3554 verify_overflow_empty();
3554 3555
3555 3556 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3556 3557 // Update the saved marks which may affect the root scans.
3557 3558 gch->save_marks();
3558 3559
3559 3560 // weak reference processing has not started yet.
3560 3561 ref_processor()->set_enqueuing_is_done(false);
3561 3562
3562 3563 {
3563 3564 // This is not needed. DEBUG_ONLY(RememberKlassesChecker imx(true);)
3564 3565 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3565 3566 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3566 3567 gch->gen_process_strong_roots(_cmsGen->level(),
3567 3568 true, // younger gens are roots
3568 3569 true, // activate StrongRootsScope
3569 3570 true, // collecting perm gen
3570 3571 SharedHeap::ScanningOption(roots_scanning_options()),
3571 3572 ¬Older,
3572 3573 true, // walk all of code cache if (so & SO_CodeCache)
3573 3574 NULL);
3574 3575 }
3575 3576
3576 3577 // Clear mod-union table; it will be dirtied in the prologue of
3577 3578 // CMS generation per each younger generation collection.
3578 3579
3579 3580 assert(_modUnionTable.isAllClear(),
3580 3581 "Was cleared in most recent final checkpoint phase"
3581 3582 " or no bits are set in the gc_prologue before the start of the next "
3582 3583 "subsequent marking phase.");
3583 3584
3584 3585 // Temporarily disabled, since pre/post-consumption closures don't
3585 3586 // care about precleaned cards
3586 3587 #if 0
3587 3588 {
3588 3589 MemRegion mr = MemRegion((HeapWord*)_virtual_space.low(),
3589 3590 (HeapWord*)_virtual_space.high());
3590 3591 _ct->ct_bs()->preclean_dirty_cards(mr);
3591 3592 }
3592 3593 #endif
3593 3594
3594 3595 // Save the end of the used_region of the constituent generations
3595 3596 // to be used to limit the extent of sweep in each generation.
3596 3597 save_sweep_limits();
3597 3598 if (UseAdaptiveSizePolicy) {
3598 3599 size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3599 3600 }
3600 3601 verify_overflow_empty();
3601 3602 }
3602 3603
3603 3604 bool CMSCollector::markFromRoots(bool asynch) {
3604 3605 // we might be tempted to assert that:
3605 3606 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3606 3607 // "inconsistent argument?");
3607 3608 // However that wouldn't be right, because it's possible that
3608 3609 // a safepoint is indeed in progress as a younger generation
3609 3610 // stop-the-world GC happens even as we mark in this generation.
3610 3611 assert(_collectorState == Marking, "inconsistent state?");
3611 3612 check_correct_thread_executing();
3612 3613 verify_overflow_empty();
3613 3614
3614 3615 bool res;
3615 3616 if (asynch) {
3616 3617
3617 3618 // Start the timers for adaptive size policy for the concurrent phases
3618 3619 // Do it here so that the foreground MS can use the concurrent
3619 3620 // timer since a foreground MS might has the sweep done concurrently
3620 3621 // or STW.
3621 3622 if (UseAdaptiveSizePolicy) {
3622 3623 size_policy()->concurrent_marking_begin();
3623 3624 }
3624 3625
3625 3626 // Weak ref discovery note: We may be discovering weak
3626 3627 // refs in this generation concurrent (but interleaved) with
3627 3628 // weak ref discovery by a younger generation collector.
3628 3629
3629 3630 CMSTokenSyncWithLocks ts(true, bitMapLock());
3630 3631 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3631 3632 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3632 3633 res = markFromRootsWork(asynch);
3633 3634 if (res) {
3634 3635 _collectorState = Precleaning;
3635 3636 } else { // We failed and a foreground collection wants to take over
3636 3637 assert(_foregroundGCIsActive, "internal state inconsistency");
3637 3638 assert(_restart_addr == NULL, "foreground will restart from scratch");
3638 3639 if (PrintGCDetails) {
3639 3640 gclog_or_tty->print_cr("bailing out to foreground collection");
3640 3641 }
3641 3642 }
3642 3643 if (UseAdaptiveSizePolicy) {
3643 3644 size_policy()->concurrent_marking_end();
3644 3645 }
3645 3646 } else {
3646 3647 assert(SafepointSynchronize::is_at_safepoint(),
3647 3648 "inconsistent with asynch == false");
3648 3649 if (UseAdaptiveSizePolicy) {
3649 3650 size_policy()->ms_collection_marking_begin();
3650 3651 }
3651 3652 // already have locks
3652 3653 res = markFromRootsWork(asynch);
3653 3654 _collectorState = FinalMarking;
3654 3655 if (UseAdaptiveSizePolicy) {
3655 3656 GenCollectedHeap* gch = GenCollectedHeap::heap();
3656 3657 size_policy()->ms_collection_marking_end(gch->gc_cause());
3657 3658 }
3658 3659 }
3659 3660 verify_overflow_empty();
3660 3661 return res;
3661 3662 }
3662 3663
3663 3664 bool CMSCollector::markFromRootsWork(bool asynch) {
3664 3665 // iterate over marked bits in bit map, doing a full scan and mark
3665 3666 // from these roots using the following algorithm:
3666 3667 // . if oop is to the right of the current scan pointer,
3667 3668 // mark corresponding bit (we'll process it later)
3668 3669 // . else (oop is to left of current scan pointer)
3669 3670 // push oop on marking stack
3670 3671 // . drain the marking stack
3671 3672
3672 3673 // Note that when we do a marking step we need to hold the
3673 3674 // bit map lock -- recall that direct allocation (by mutators)
3674 3675 // and promotion (by younger generation collectors) is also
3675 3676 // marking the bit map. [the so-called allocate live policy.]
3676 3677 // Because the implementation of bit map marking is not
3677 3678 // robust wrt simultaneous marking of bits in the same word,
3678 3679 // we need to make sure that there is no such interference
3679 3680 // between concurrent such updates.
3680 3681
3681 3682 // already have locks
3682 3683 assert_lock_strong(bitMapLock());
3683 3684
3684 3685 // Clear the revisit stack, just in case there are any
3685 3686 // obsolete contents from a short-circuited previous CMS cycle.
3686 3687 _revisitStack.reset();
3687 3688 verify_work_stacks_empty();
3688 3689 verify_overflow_empty();
3689 3690 assert(_revisitStack.isEmpty(), "tabula rasa");
3690 3691 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());)
3691 3692 bool result = false;
3692 3693 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3693 3694 result = do_marking_mt(asynch);
3694 3695 } else {
3695 3696 result = do_marking_st(asynch);
3696 3697 }
3697 3698 return result;
3698 3699 }
3699 3700
3700 3701 // Forward decl
3701 3702 class CMSConcMarkingTask;
3702 3703
3703 3704 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3704 3705 CMSCollector* _collector;
3705 3706 CMSConcMarkingTask* _task;
3706 3707 public:
3707 3708 virtual void yield();
3708 3709
3709 3710 // "n_threads" is the number of threads to be terminated.
3710 3711 // "queue_set" is a set of work queues of other threads.
3711 3712 // "collector" is the CMS collector associated with this task terminator.
3712 3713 // "yield" indicates whether we need the gang as a whole to yield.
3713 3714 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3714 3715 ParallelTaskTerminator(n_threads, queue_set),
3715 3716 _collector(collector) { }
3716 3717
3717 3718 void set_task(CMSConcMarkingTask* task) {
3718 3719 _task = task;
3719 3720 }
3720 3721 };
3721 3722
3722 3723 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3723 3724 CMSConcMarkingTask* _task;
3724 3725 public:
3725 3726 bool should_exit_termination();
3726 3727 void set_task(CMSConcMarkingTask* task) {
3727 3728 _task = task;
3728 3729 }
3729 3730 };
3730 3731
3731 3732 // MT Concurrent Marking Task
3732 3733 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3733 3734 CMSCollector* _collector;
3734 3735 int _n_workers; // requested/desired # workers
3735 3736 bool _asynch;
3736 3737 bool _result;
3737 3738 CompactibleFreeListSpace* _cms_space;
3738 3739 CompactibleFreeListSpace* _perm_space;
3739 3740 char _pad_front[64]; // padding to ...
3740 3741 HeapWord* _global_finger; // ... avoid sharing cache line
3741 3742 char _pad_back[64];
3742 3743 HeapWord* _restart_addr;
3743 3744
3744 3745 // Exposed here for yielding support
3745 3746 Mutex* const _bit_map_lock;
3746 3747
3747 3748 // The per thread work queues, available here for stealing
3748 3749 OopTaskQueueSet* _task_queues;
3749 3750
3750 3751 // Termination (and yielding) support
3751 3752 CMSConcMarkingTerminator _term;
3752 3753 CMSConcMarkingTerminatorTerminator _term_term;
3753 3754
3754 3755 public:
3755 3756 CMSConcMarkingTask(CMSCollector* collector,
3756 3757 CompactibleFreeListSpace* cms_space,
3757 3758 CompactibleFreeListSpace* perm_space,
3758 3759 bool asynch,
3759 3760 YieldingFlexibleWorkGang* workers,
3760 3761 OopTaskQueueSet* task_queues):
3761 3762 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3762 3763 _collector(collector),
3763 3764 _cms_space(cms_space),
3764 3765 _perm_space(perm_space),
3765 3766 _asynch(asynch), _n_workers(0), _result(true),
3766 3767 _task_queues(task_queues),
3767 3768 _term(_n_workers, task_queues, _collector),
3768 3769 _bit_map_lock(collector->bitMapLock())
3769 3770 {
3770 3771 _requested_size = _n_workers;
3771 3772 _term.set_task(this);
3772 3773 _term_term.set_task(this);
3773 3774 assert(_cms_space->bottom() < _perm_space->bottom(),
3774 3775 "Finger incorrectly initialized below");
3775 3776 _restart_addr = _global_finger = _cms_space->bottom();
3776 3777 }
3777 3778
3778 3779
3779 3780 OopTaskQueueSet* task_queues() { return _task_queues; }
3780 3781
3781 3782 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3782 3783
3783 3784 HeapWord** global_finger_addr() { return &_global_finger; }
3784 3785
3785 3786 CMSConcMarkingTerminator* terminator() { return &_term; }
3786 3787
3787 3788 virtual void set_for_termination(int active_workers) {
3788 3789 terminator()->reset_for_reuse(active_workers);
3789 3790 }
3790 3791
3791 3792 void work(int i);
3792 3793 bool should_yield() {
3793 3794 return ConcurrentMarkSweepThread::should_yield()
3794 3795 && !_collector->foregroundGCIsActive()
3795 3796 && _asynch;
3796 3797 }
3797 3798
3798 3799 virtual void coordinator_yield(); // stuff done by coordinator
3799 3800 bool result() { return _result; }
3800 3801
3801 3802 void reset(HeapWord* ra) {
3802 3803 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3803 3804 assert(_global_finger >= _perm_space->end(), "Postcondition of ::work(i)");
3804 3805 assert(ra < _perm_space->end(), "ra too large");
3805 3806 _restart_addr = _global_finger = ra;
3806 3807 _term.reset_for_reuse();
3807 3808 }
3808 3809
3809 3810 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3810 3811 OopTaskQueue* work_q);
3811 3812
3812 3813 private:
3813 3814 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3814 3815 void do_work_steal(int i);
3815 3816 void bump_global_finger(HeapWord* f);
3816 3817 };
3817 3818
3818 3819 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3819 3820 assert(_task != NULL, "Error");
3820 3821 return _task->yielding();
3821 3822 // Note that we do not need the disjunct || _task->should_yield() above
3822 3823 // because we want terminating threads to yield only if the task
3823 3824 // is already in the midst of yielding, which happens only after at least one
3824 3825 // thread has yielded.
3825 3826 }
3826 3827
3827 3828 void CMSConcMarkingTerminator::yield() {
3828 3829 if (_task->should_yield()) {
3829 3830 _task->yield();
3830 3831 } else {
3831 3832 ParallelTaskTerminator::yield();
3832 3833 }
3833 3834 }
3834 3835
3835 3836 ////////////////////////////////////////////////////////////////
3836 3837 // Concurrent Marking Algorithm Sketch
3837 3838 ////////////////////////////////////////////////////////////////
3838 3839 // Until all tasks exhausted (both spaces):
3839 3840 // -- claim next available chunk
3840 3841 // -- bump global finger via CAS
3841 3842 // -- find first object that starts in this chunk
3842 3843 // and start scanning bitmap from that position
3843 3844 // -- scan marked objects for oops
3844 3845 // -- CAS-mark target, and if successful:
3845 3846 // . if target oop is above global finger (volatile read)
3846 3847 // nothing to do
3847 3848 // . if target oop is in chunk and above local finger
3848 3849 // then nothing to do
3849 3850 // . else push on work-queue
3850 3851 // -- Deal with possible overflow issues:
3851 3852 // . local work-queue overflow causes stuff to be pushed on
3852 3853 // global (common) overflow queue
3853 3854 // . always first empty local work queue
3854 3855 // . then get a batch of oops from global work queue if any
3855 3856 // . then do work stealing
3856 3857 // -- When all tasks claimed (both spaces)
3857 3858 // and local work queue empty,
3858 3859 // then in a loop do:
3859 3860 // . check global overflow stack; steal a batch of oops and trace
3860 3861 // . try to steal from other threads oif GOS is empty
3861 3862 // . if neither is available, offer termination
3862 3863 // -- Terminate and return result
3863 3864 //
3864 3865 void CMSConcMarkingTask::work(int i) {
3865 3866 elapsedTimer _timer;
3866 3867 ResourceMark rm;
3867 3868 HandleMark hm;
3868 3869
3869 3870 DEBUG_ONLY(_collector->verify_overflow_empty();)
3870 3871
3871 3872 // Before we begin work, our work queue should be empty
3872 3873 assert(work_queue(i)->size() == 0, "Expected to be empty");
3873 3874 // Scan the bitmap covering _cms_space, tracing through grey objects.
3874 3875 _timer.start();
3875 3876 do_scan_and_mark(i, _cms_space);
3876 3877 _timer.stop();
3877 3878 if (PrintCMSStatistics != 0) {
3878 3879 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3879 3880 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3880 3881 }
3881 3882
3882 3883 // ... do the same for the _perm_space
3883 3884 _timer.reset();
3884 3885 _timer.start();
3885 3886 do_scan_and_mark(i, _perm_space);
3886 3887 _timer.stop();
3887 3888 if (PrintCMSStatistics != 0) {
3888 3889 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec",
3889 3890 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3890 3891 }
3891 3892
3892 3893 // ... do work stealing
3893 3894 _timer.reset();
3894 3895 _timer.start();
3895 3896 do_work_steal(i);
3896 3897 _timer.stop();
3897 3898 if (PrintCMSStatistics != 0) {
3898 3899 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3899 3900 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3900 3901 }
3901 3902 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3902 3903 assert(work_queue(i)->size() == 0, "Should have been emptied");
3903 3904 // Note that under the current task protocol, the
3904 3905 // following assertion is true even of the spaces
3905 3906 // expanded since the completion of the concurrent
3906 3907 // marking. XXX This will likely change under a strict
3907 3908 // ABORT semantics.
3908 3909 assert(_global_finger > _cms_space->end() &&
3909 3910 _global_finger >= _perm_space->end(),
3910 3911 "All tasks have been completed");
3911 3912 DEBUG_ONLY(_collector->verify_overflow_empty();)
3912 3913 }
3913 3914
3914 3915 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3915 3916 HeapWord* read = _global_finger;
3916 3917 HeapWord* cur = read;
3917 3918 while (f > read) {
3918 3919 cur = read;
3919 3920 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3920 3921 if (cur == read) {
3921 3922 // our cas succeeded
3922 3923 assert(_global_finger >= f, "protocol consistency");
3923 3924 break;
3924 3925 }
3925 3926 }
3926 3927 }
3927 3928
3928 3929 // This is really inefficient, and should be redone by
3929 3930 // using (not yet available) block-read and -write interfaces to the
3930 3931 // stack and the work_queue. XXX FIX ME !!!
3931 3932 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3932 3933 OopTaskQueue* work_q) {
3933 3934 // Fast lock-free check
3934 3935 if (ovflw_stk->length() == 0) {
3935 3936 return false;
3936 3937 }
3937 3938 assert(work_q->size() == 0, "Shouldn't steal");
3938 3939 MutexLockerEx ml(ovflw_stk->par_lock(),
3939 3940 Mutex::_no_safepoint_check_flag);
3940 3941 // Grab up to 1/4 the size of the work queue
3941 3942 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3942 3943 (size_t)ParGCDesiredObjsFromOverflowList);
3943 3944 num = MIN2(num, ovflw_stk->length());
3944 3945 for (int i = (int) num; i > 0; i--) {
3945 3946 oop cur = ovflw_stk->pop();
3946 3947 assert(cur != NULL, "Counted wrong?");
3947 3948 work_q->push(cur);
3948 3949 }
3949 3950 return num > 0;
3950 3951 }
3951 3952
3952 3953 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3953 3954 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3954 3955 int n_tasks = pst->n_tasks();
3955 3956 // We allow that there may be no tasks to do here because
3956 3957 // we are restarting after a stack overflow.
3957 3958 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3958 3959 int nth_task = 0;
3959 3960
3960 3961 HeapWord* aligned_start = sp->bottom();
3961 3962 if (sp->used_region().contains(_restart_addr)) {
3962 3963 // Align down to a card boundary for the start of 0th task
3963 3964 // for this space.
3964 3965 aligned_start =
3965 3966 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3966 3967 CardTableModRefBS::card_size);
3967 3968 }
3968 3969
3969 3970 size_t chunk_size = sp->marking_task_size();
3970 3971 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3971 3972 // Having claimed the nth task in this space,
3972 3973 // compute the chunk that it corresponds to:
3973 3974 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3974 3975 aligned_start + (nth_task+1)*chunk_size);
3975 3976 // Try and bump the global finger via a CAS;
3976 3977 // note that we need to do the global finger bump
3977 3978 // _before_ taking the intersection below, because
3978 3979 // the task corresponding to that region will be
3979 3980 // deemed done even if the used_region() expands
3980 3981 // because of allocation -- as it almost certainly will
3981 3982 // during start-up while the threads yield in the
3982 3983 // closure below.
3983 3984 HeapWord* finger = span.end();
3984 3985 bump_global_finger(finger); // atomically
3985 3986 // There are null tasks here corresponding to chunks
3986 3987 // beyond the "top" address of the space.
3987 3988 span = span.intersection(sp->used_region());
3988 3989 if (!span.is_empty()) { // Non-null task
3989 3990 HeapWord* prev_obj;
3990 3991 assert(!span.contains(_restart_addr) || nth_task == 0,
3991 3992 "Inconsistency");
3992 3993 if (nth_task == 0) {
3993 3994 // For the 0th task, we'll not need to compute a block_start.
3994 3995 if (span.contains(_restart_addr)) {
3995 3996 // In the case of a restart because of stack overflow,
3996 3997 // we might additionally skip a chunk prefix.
3997 3998 prev_obj = _restart_addr;
3998 3999 } else {
3999 4000 prev_obj = span.start();
4000 4001 }
4001 4002 } else {
4002 4003 // We want to skip the first object because
4003 4004 // the protocol is to scan any object in its entirety
4004 4005 // that _starts_ in this span; a fortiori, any
4005 4006 // object starting in an earlier span is scanned
4006 4007 // as part of an earlier claimed task.
4007 4008 // Below we use the "careful" version of block_start
4008 4009 // so we do not try to navigate uninitialized objects.
4009 4010 prev_obj = sp->block_start_careful(span.start());
4010 4011 // Below we use a variant of block_size that uses the
4011 4012 // Printezis bits to avoid waiting for allocated
4012 4013 // objects to become initialized/parsable.
4013 4014 while (prev_obj < span.start()) {
4014 4015 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
4015 4016 if (sz > 0) {
4016 4017 prev_obj += sz;
4017 4018 } else {
4018 4019 // In this case we may end up doing a bit of redundant
4019 4020 // scanning, but that appears unavoidable, short of
4020 4021 // locking the free list locks; see bug 6324141.
4021 4022 break;
4022 4023 }
4023 4024 }
4024 4025 }
4025 4026 if (prev_obj < span.end()) {
4026 4027 MemRegion my_span = MemRegion(prev_obj, span.end());
4027 4028 // Do the marking work within a non-empty span --
4028 4029 // the last argument to the constructor indicates whether the
4029 4030 // iteration should be incremental with periodic yields.
4030 4031 Par_MarkFromRootsClosure cl(this, _collector, my_span,
4031 4032 &_collector->_markBitMap,
4032 4033 work_queue(i),
4033 4034 &_collector->_markStack,
4034 4035 &_collector->_revisitStack,
4035 4036 _asynch);
4036 4037 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
4037 4038 } // else nothing to do for this task
4038 4039 } // else nothing to do for this task
4039 4040 }
4040 4041 // We'd be tempted to assert here that since there are no
4041 4042 // more tasks left to claim in this space, the global_finger
4042 4043 // must exceed space->top() and a fortiori space->end(). However,
4043 4044 // that would not quite be correct because the bumping of
4044 4045 // global_finger occurs strictly after the claiming of a task,
4045 4046 // so by the time we reach here the global finger may not yet
4046 4047 // have been bumped up by the thread that claimed the last
4047 4048 // task.
4048 4049 pst->all_tasks_completed();
4049 4050 }
4050 4051
4051 4052 class Par_ConcMarkingClosure: public Par_KlassRememberingOopClosure {
4052 4053 private:
4053 4054 CMSConcMarkingTask* _task;
4054 4055 MemRegion _span;
4055 4056 CMSBitMap* _bit_map;
4056 4057 CMSMarkStack* _overflow_stack;
4057 4058 OopTaskQueue* _work_queue;
4058 4059 protected:
4059 4060 DO_OOP_WORK_DEFN
4060 4061 public:
4061 4062 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
4062 4063 CMSBitMap* bit_map, CMSMarkStack* overflow_stack,
4063 4064 CMSMarkStack* revisit_stack):
4064 4065 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
4065 4066 _task(task),
4066 4067 _span(collector->_span),
4067 4068 _work_queue(work_queue),
4068 4069 _bit_map(bit_map),
4069 4070 _overflow_stack(overflow_stack)
4070 4071 { }
4071 4072 virtual void do_oop(oop* p);
4072 4073 virtual void do_oop(narrowOop* p);
4073 4074 void trim_queue(size_t max);
4074 4075 void handle_stack_overflow(HeapWord* lost);
4075 4076 void do_yield_check() {
4076 4077 if (_task->should_yield()) {
4077 4078 _task->yield();
4078 4079 }
4079 4080 }
4080 4081 };
4081 4082
4082 4083 // Grey object scanning during work stealing phase --
4083 4084 // the salient assumption here is that any references
4084 4085 // that are in these stolen objects being scanned must
4085 4086 // already have been initialized (else they would not have
4086 4087 // been published), so we do not need to check for
4087 4088 // uninitialized objects before pushing here.
4088 4089 void Par_ConcMarkingClosure::do_oop(oop obj) {
4089 4090 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
4090 4091 HeapWord* addr = (HeapWord*)obj;
4091 4092 // Check if oop points into the CMS generation
4092 4093 // and is not marked
4093 4094 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
4094 4095 // a white object ...
4095 4096 // If we manage to "claim" the object, by being the
4096 4097 // first thread to mark it, then we push it on our
4097 4098 // marking stack
4098 4099 if (_bit_map->par_mark(addr)) { // ... now grey
4099 4100 // push on work queue (grey set)
4100 4101 bool simulate_overflow = false;
4101 4102 NOT_PRODUCT(
4102 4103 if (CMSMarkStackOverflowALot &&
4103 4104 _collector->simulate_overflow()) {
4104 4105 // simulate a stack overflow
4105 4106 simulate_overflow = true;
4106 4107 }
4107 4108 )
4108 4109 if (simulate_overflow ||
4109 4110 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
4110 4111 // stack overflow
4111 4112 if (PrintCMSStatistics != 0) {
4112 4113 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
4113 4114 SIZE_FORMAT, _overflow_stack->capacity());
4114 4115 }
4115 4116 // We cannot assert that the overflow stack is full because
4116 4117 // it may have been emptied since.
4117 4118 assert(simulate_overflow ||
4118 4119 _work_queue->size() == _work_queue->max_elems(),
4119 4120 "Else push should have succeeded");
4120 4121 handle_stack_overflow(addr);
4121 4122 }
4122 4123 } // Else, some other thread got there first
4123 4124 do_yield_check();
4124 4125 }
4125 4126 }
4126 4127
4127 4128 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4128 4129 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4129 4130
4130 4131 void Par_ConcMarkingClosure::trim_queue(size_t max) {
4131 4132 while (_work_queue->size() > max) {
4132 4133 oop new_oop;
4133 4134 if (_work_queue->pop_local(new_oop)) {
4134 4135 assert(new_oop->is_oop(), "Should be an oop");
4135 4136 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
4136 4137 assert(_span.contains((HeapWord*)new_oop), "Not in span");
4137 4138 assert(new_oop->is_parsable(), "Should be parsable");
4138 4139 new_oop->oop_iterate(this); // do_oop() above
4139 4140 do_yield_check();
4140 4141 }
4141 4142 }
4142 4143 }
4143 4144
4144 4145 // Upon stack overflow, we discard (part of) the stack,
4145 4146 // remembering the least address amongst those discarded
4146 4147 // in CMSCollector's _restart_address.
4147 4148 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
4148 4149 // We need to do this under a mutex to prevent other
4149 4150 // workers from interfering with the work done below.
4150 4151 MutexLockerEx ml(_overflow_stack->par_lock(),
4151 4152 Mutex::_no_safepoint_check_flag);
4152 4153 // Remember the least grey address discarded
4153 4154 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
4154 4155 _collector->lower_restart_addr(ra);
4155 4156 _overflow_stack->reset(); // discard stack contents
4156 4157 _overflow_stack->expand(); // expand the stack if possible
4157 4158 }
4158 4159
4159 4160
4160 4161 void CMSConcMarkingTask::do_work_steal(int i) {
4161 4162 OopTaskQueue* work_q = work_queue(i);
4162 4163 oop obj_to_scan;
4163 4164 CMSBitMap* bm = &(_collector->_markBitMap);
4164 4165 CMSMarkStack* ovflw = &(_collector->_markStack);
4165 4166 CMSMarkStack* revisit = &(_collector->_revisitStack);
4166 4167 int* seed = _collector->hash_seed(i);
4167 4168 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw, revisit);
4168 4169 while (true) {
4169 4170 cl.trim_queue(0);
4170 4171 assert(work_q->size() == 0, "Should have been emptied above");
4171 4172 if (get_work_from_overflow_stack(ovflw, work_q)) {
4172 4173 // Can't assert below because the work obtained from the
4173 4174 // overflow stack may already have been stolen from us.
4174 4175 // assert(work_q->size() > 0, "Work from overflow stack");
4175 4176 continue;
4176 4177 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4177 4178 assert(obj_to_scan->is_oop(), "Should be an oop");
4178 4179 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
4179 4180 obj_to_scan->oop_iterate(&cl);
4180 4181 } else if (terminator()->offer_termination(&_term_term)) {
4181 4182 assert(work_q->size() == 0, "Impossible!");
4182 4183 break;
4183 4184 } else if (yielding() || should_yield()) {
4184 4185 yield();
4185 4186 }
4186 4187 }
4187 4188 }
4188 4189
4189 4190 // This is run by the CMS (coordinator) thread.
4190 4191 void CMSConcMarkingTask::coordinator_yield() {
4191 4192 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4192 4193 "CMS thread should hold CMS token");
4193 4194 DEBUG_ONLY(RememberKlassesChecker mux(false);)
4194 4195 // First give up the locks, then yield, then re-lock
4195 4196 // We should probably use a constructor/destructor idiom to
4196 4197 // do this unlock/lock or modify the MutexUnlocker class to
4197 4198 // serve our purpose. XXX
4198 4199 assert_lock_strong(_bit_map_lock);
4199 4200 _bit_map_lock->unlock();
4200 4201 ConcurrentMarkSweepThread::desynchronize(true);
4201 4202 ConcurrentMarkSweepThread::acknowledge_yield_request();
4202 4203 _collector->stopTimer();
4203 4204 if (PrintCMSStatistics != 0) {
4204 4205 _collector->incrementYields();
4205 4206 }
4206 4207 _collector->icms_wait();
4207 4208
4208 4209 // It is possible for whichever thread initiated the yield request
4209 4210 // not to get a chance to wake up and take the bitmap lock between
4210 4211 // this thread releasing it and reacquiring it. So, while the
4211 4212 // should_yield() flag is on, let's sleep for a bit to give the
4212 4213 // other thread a chance to wake up. The limit imposed on the number
4213 4214 // of iterations is defensive, to avoid any unforseen circumstances
4214 4215 // putting us into an infinite loop. Since it's always been this
4215 4216 // (coordinator_yield()) method that was observed to cause the
4216 4217 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4217 4218 // which is by default non-zero. For the other seven methods that
4218 4219 // also perform the yield operation, as are using a different
4219 4220 // parameter (CMSYieldSleepCount) which is by default zero. This way we
4220 4221 // can enable the sleeping for those methods too, if necessary.
4221 4222 // See 6442774.
4222 4223 //
4223 4224 // We really need to reconsider the synchronization between the GC
4224 4225 // thread and the yield-requesting threads in the future and we
4225 4226 // should really use wait/notify, which is the recommended
4226 4227 // way of doing this type of interaction. Additionally, we should
4227 4228 // consolidate the eight methods that do the yield operation and they
4228 4229 // are almost identical into one for better maintenability and
4229 4230 // readability. See 6445193.
4230 4231 //
4231 4232 // Tony 2006.06.29
4232 4233 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4233 4234 ConcurrentMarkSweepThread::should_yield() &&
4234 4235 !CMSCollector::foregroundGCIsActive(); ++i) {
4235 4236 os::sleep(Thread::current(), 1, false);
4236 4237 ConcurrentMarkSweepThread::acknowledge_yield_request();
4237 4238 }
4238 4239
4239 4240 ConcurrentMarkSweepThread::synchronize(true);
4240 4241 _bit_map_lock->lock_without_safepoint_check();
4241 4242 _collector->startTimer();
4242 4243 }
4243 4244
4244 4245 bool CMSCollector::do_marking_mt(bool asynch) {
4245 4246 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
4246 4247 // In the future this would be determined ergonomically, based
4247 4248 // on #cpu's, # active mutator threads (and load), and mutation rate.
4248 4249 int num_workers = ConcGCThreads;
4249 4250
4250 4251 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4251 4252 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
4252 4253
4253 4254 CMSConcMarkingTask tsk(this,
4254 4255 cms_space,
4255 4256 perm_space,
4256 4257 asynch,
4257 4258 conc_workers(),
4258 4259 task_queues());
4259 4260
4260 4261 // Since the actual number of workers we get may be different
4261 4262 // from the number we requested above, do we need to do anything different
4262 4263 // below? In particular, may be we need to subclass the SequantialSubTasksDone
4263 4264 // class?? XXX
4264 4265 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4265 4266 perm_space->initialize_sequential_subtasks_for_marking(num_workers);
4266 4267
4267 4268 // Refs discovery is already non-atomic.
4268 4269 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4269 4270 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
4270 4271 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());)
4271 4272 conc_workers()->start_task(&tsk);
4272 4273 while (tsk.yielded()) {
4273 4274 tsk.coordinator_yield();
4274 4275 conc_workers()->continue_task(&tsk);
4275 4276 }
4276 4277 // If the task was aborted, _restart_addr will be non-NULL
4277 4278 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4278 4279 while (_restart_addr != NULL) {
4279 4280 // XXX For now we do not make use of ABORTED state and have not
4280 4281 // yet implemented the right abort semantics (even in the original
4281 4282 // single-threaded CMS case). That needs some more investigation
4282 4283 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4283 4284 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4284 4285 // If _restart_addr is non-NULL, a marking stack overflow
4285 4286 // occurred; we need to do a fresh marking iteration from the
4286 4287 // indicated restart address.
4287 4288 if (_foregroundGCIsActive && asynch) {
4288 4289 // We may be running into repeated stack overflows, having
4289 4290 // reached the limit of the stack size, while making very
4290 4291 // slow forward progress. It may be best to bail out and
4291 4292 // let the foreground collector do its job.
4292 4293 // Clear _restart_addr, so that foreground GC
4293 4294 // works from scratch. This avoids the headache of
4294 4295 // a "rescan" which would otherwise be needed because
4295 4296 // of the dirty mod union table & card table.
4296 4297 _restart_addr = NULL;
4297 4298 return false;
4298 4299 }
4299 4300 // Adjust the task to restart from _restart_addr
4300 4301 tsk.reset(_restart_addr);
4301 4302 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4302 4303 _restart_addr);
4303 4304 perm_space->initialize_sequential_subtasks_for_marking(num_workers,
4304 4305 _restart_addr);
4305 4306 _restart_addr = NULL;
4306 4307 // Get the workers going again
4307 4308 conc_workers()->start_task(&tsk);
4308 4309 while (tsk.yielded()) {
4309 4310 tsk.coordinator_yield();
4310 4311 conc_workers()->continue_task(&tsk);
4311 4312 }
4312 4313 }
4313 4314 assert(tsk.completed(), "Inconsistency");
4314 4315 assert(tsk.result() == true, "Inconsistency");
4315 4316 return true;
4316 4317 }
4317 4318
4318 4319 bool CMSCollector::do_marking_st(bool asynch) {
4319 4320 ResourceMark rm;
4320 4321 HandleMark hm;
4321 4322
4322 4323 // Temporarily make refs discovery single threaded (non-MT)
4323 4324 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
4324 4325 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4325 4326 &_markStack, &_revisitStack, CMSYield && asynch);
4326 4327 // the last argument to iterate indicates whether the iteration
4327 4328 // should be incremental with periodic yields.
4328 4329 _markBitMap.iterate(&markFromRootsClosure);
4329 4330 // If _restart_addr is non-NULL, a marking stack overflow
4330 4331 // occurred; we need to do a fresh iteration from the
4331 4332 // indicated restart address.
4332 4333 while (_restart_addr != NULL) {
4333 4334 if (_foregroundGCIsActive && asynch) {
4334 4335 // We may be running into repeated stack overflows, having
4335 4336 // reached the limit of the stack size, while making very
4336 4337 // slow forward progress. It may be best to bail out and
4337 4338 // let the foreground collector do its job.
4338 4339 // Clear _restart_addr, so that foreground GC
4339 4340 // works from scratch. This avoids the headache of
4340 4341 // a "rescan" which would otherwise be needed because
4341 4342 // of the dirty mod union table & card table.
4342 4343 _restart_addr = NULL;
4343 4344 return false; // indicating failure to complete marking
4344 4345 }
4345 4346 // Deal with stack overflow:
4346 4347 // we restart marking from _restart_addr
4347 4348 HeapWord* ra = _restart_addr;
4348 4349 markFromRootsClosure.reset(ra);
4349 4350 _restart_addr = NULL;
4350 4351 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4351 4352 }
4352 4353 return true;
4353 4354 }
4354 4355
4355 4356 void CMSCollector::preclean() {
4356 4357 check_correct_thread_executing();
4357 4358 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4358 4359 verify_work_stacks_empty();
4359 4360 verify_overflow_empty();
4360 4361 _abort_preclean = false;
4361 4362 if (CMSPrecleaningEnabled) {
4362 4363 _eden_chunk_index = 0;
4363 4364 size_t used = get_eden_used();
4364 4365 size_t capacity = get_eden_capacity();
4365 4366 // Don't start sampling unless we will get sufficiently
4366 4367 // many samples.
4367 4368 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4368 4369 * CMSScheduleRemarkEdenPenetration)) {
4369 4370 _start_sampling = true;
4370 4371 } else {
4371 4372 _start_sampling = false;
4372 4373 }
4373 4374 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4374 4375 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
4375 4376 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4376 4377 }
4377 4378 CMSTokenSync x(true); // is cms thread
4378 4379 if (CMSPrecleaningEnabled) {
4379 4380 sample_eden();
4380 4381 _collectorState = AbortablePreclean;
4381 4382 } else {
4382 4383 _collectorState = FinalMarking;
4383 4384 }
4384 4385 verify_work_stacks_empty();
4385 4386 verify_overflow_empty();
4386 4387 }
4387 4388
4388 4389 // Try and schedule the remark such that young gen
4389 4390 // occupancy is CMSScheduleRemarkEdenPenetration %.
4390 4391 void CMSCollector::abortable_preclean() {
4391 4392 check_correct_thread_executing();
4392 4393 assert(CMSPrecleaningEnabled, "Inconsistent control state");
4393 4394 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4394 4395
4395 4396 // If Eden's current occupancy is below this threshold,
4396 4397 // immediately schedule the remark; else preclean
4397 4398 // past the next scavenge in an effort to
4398 4399 // schedule the pause as described avove. By choosing
4399 4400 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4400 4401 // we will never do an actual abortable preclean cycle.
4401 4402 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4402 4403 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4403 4404 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
4404 4405 // We need more smarts in the abortable preclean
4405 4406 // loop below to deal with cases where allocation
4406 4407 // in young gen is very very slow, and our precleaning
4407 4408 // is running a losing race against a horde of
4408 4409 // mutators intent on flooding us with CMS updates
4409 4410 // (dirty cards).
4410 4411 // One, admittedly dumb, strategy is to give up
4411 4412 // after a certain number of abortable precleaning loops
4412 4413 // or after a certain maximum time. We want to make
4413 4414 // this smarter in the next iteration.
4414 4415 // XXX FIX ME!!! YSR
4415 4416 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4416 4417 while (!(should_abort_preclean() ||
4417 4418 ConcurrentMarkSweepThread::should_terminate())) {
4418 4419 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4419 4420 cumworkdone += workdone;
4420 4421 loops++;
4421 4422 // Voluntarily terminate abortable preclean phase if we have
4422 4423 // been at it for too long.
4423 4424 if ((CMSMaxAbortablePrecleanLoops != 0) &&
4424 4425 loops >= CMSMaxAbortablePrecleanLoops) {
4425 4426 if (PrintGCDetails) {
4426 4427 gclog_or_tty->print(" CMS: abort preclean due to loops ");
4427 4428 }
4428 4429 break;
4429 4430 }
4430 4431 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4431 4432 if (PrintGCDetails) {
4432 4433 gclog_or_tty->print(" CMS: abort preclean due to time ");
4433 4434 }
4434 4435 break;
4435 4436 }
4436 4437 // If we are doing little work each iteration, we should
4437 4438 // take a short break.
4438 4439 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4439 4440 // Sleep for some time, waiting for work to accumulate
4440 4441 stopTimer();
4441 4442 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4442 4443 startTimer();
4443 4444 waited++;
4444 4445 }
4445 4446 }
4446 4447 if (PrintCMSStatistics > 0) {
4447 4448 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4448 4449 loops, waited, cumworkdone);
4449 4450 }
4450 4451 }
4451 4452 CMSTokenSync x(true); // is cms thread
4452 4453 if (_collectorState != Idling) {
4453 4454 assert(_collectorState == AbortablePreclean,
4454 4455 "Spontaneous state transition?");
4455 4456 _collectorState = FinalMarking;
4456 4457 } // Else, a foreground collection completed this CMS cycle.
4457 4458 return;
4458 4459 }
4459 4460
4460 4461 // Respond to an Eden sampling opportunity
4461 4462 void CMSCollector::sample_eden() {
4462 4463 // Make sure a young gc cannot sneak in between our
4463 4464 // reading and recording of a sample.
4464 4465 assert(Thread::current()->is_ConcurrentGC_thread(),
4465 4466 "Only the cms thread may collect Eden samples");
4466 4467 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4467 4468 "Should collect samples while holding CMS token");
4468 4469 if (!_start_sampling) {
4469 4470 return;
4470 4471 }
4471 4472 if (_eden_chunk_array) {
4472 4473 if (_eden_chunk_index < _eden_chunk_capacity) {
4473 4474 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
4474 4475 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4475 4476 "Unexpected state of Eden");
4476 4477 // We'd like to check that what we just sampled is an oop-start address;
4477 4478 // however, we cannot do that here since the object may not yet have been
4478 4479 // initialized. So we'll instead do the check when we _use_ this sample
4479 4480 // later.
4480 4481 if (_eden_chunk_index == 0 ||
4481 4482 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4482 4483 _eden_chunk_array[_eden_chunk_index-1])
4483 4484 >= CMSSamplingGrain)) {
4484 4485 _eden_chunk_index++; // commit sample
4485 4486 }
4486 4487 }
4487 4488 }
4488 4489 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4489 4490 size_t used = get_eden_used();
4490 4491 size_t capacity = get_eden_capacity();
4491 4492 assert(used <= capacity, "Unexpected state of Eden");
4492 4493 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4493 4494 _abort_preclean = true;
4494 4495 }
4495 4496 }
4496 4497 }
4497 4498
4498 4499
4499 4500 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4500 4501 assert(_collectorState == Precleaning ||
4501 4502 _collectorState == AbortablePreclean, "incorrect state");
4502 4503 ResourceMark rm;
4503 4504 HandleMark hm;
4504 4505
4505 4506 // Precleaning is currently not MT but the reference processor
4506 4507 // may be set for MT. Disable it temporarily here.
4507 4508 ReferenceProcessor* rp = ref_processor();
4508 4509 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
4509 4510
4510 4511 // Do one pass of scrubbing the discovered reference lists
4511 4512 // to remove any reference objects with strongly-reachable
4512 4513 // referents.
4513 4514 if (clean_refs) {
4514 4515 CMSPrecleanRefsYieldClosure yield_cl(this);
4515 4516 assert(rp->span().equals(_span), "Spans should be equal");
4516 4517 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4517 4518 &_markStack, &_revisitStack,
4518 4519 true /* preclean */);
4519 4520 CMSDrainMarkingStackClosure complete_trace(this,
4520 4521 _span, &_markBitMap, &_markStack,
4521 4522 &keep_alive, true /* preclean */);
4522 4523
4523 4524 // We don't want this step to interfere with a young
4524 4525 // collection because we don't want to take CPU
4525 4526 // or memory bandwidth away from the young GC threads
4526 4527 // (which may be as many as there are CPUs).
4527 4528 // Note that we don't need to protect ourselves from
4528 4529 // interference with mutators because they can't
4529 4530 // manipulate the discovered reference lists nor affect
4530 4531 // the computed reachability of the referents, the
4531 4532 // only properties manipulated by the precleaning
4532 4533 // of these reference lists.
4533 4534 stopTimer();
4534 4535 CMSTokenSyncWithLocks x(true /* is cms thread */,
4535 4536 bitMapLock());
4536 4537 startTimer();
4537 4538 sample_eden();
4538 4539
4539 4540 // The following will yield to allow foreground
4540 4541 // collection to proceed promptly. XXX YSR:
4541 4542 // The code in this method may need further
4542 4543 // tweaking for better performance and some restructuring
4543 4544 // for cleaner interfaces.
4544 4545 rp->preclean_discovered_references(
4545 4546 rp->is_alive_non_header(), &keep_alive, &complete_trace,
4546 4547 &yield_cl, should_unload_classes());
4547 4548 }
4548 4549
4549 4550 if (clean_survivor) { // preclean the active survivor space(s)
4550 4551 assert(_young_gen->kind() == Generation::DefNew ||
4551 4552 _young_gen->kind() == Generation::ParNew ||
4552 4553 _young_gen->kind() == Generation::ASParNew,
4553 4554 "incorrect type for cast");
4554 4555 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4555 4556 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4556 4557 &_markBitMap, &_modUnionTable,
4557 4558 &_markStack, &_revisitStack,
4558 4559 true /* precleaning phase */);
4559 4560 stopTimer();
4560 4561 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4561 4562 bitMapLock());
4562 4563 startTimer();
4563 4564 unsigned int before_count =
4564 4565 GenCollectedHeap::heap()->total_collections();
4565 4566 SurvivorSpacePrecleanClosure
4566 4567 sss_cl(this, _span, &_markBitMap, &_markStack,
4567 4568 &pam_cl, before_count, CMSYield);
4568 4569 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4569 4570 dng->from()->object_iterate_careful(&sss_cl);
4570 4571 dng->to()->object_iterate_careful(&sss_cl);
4571 4572 }
4572 4573 MarkRefsIntoAndScanClosure
4573 4574 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4574 4575 &_markStack, &_revisitStack, this, CMSYield,
4575 4576 true /* precleaning phase */);
4576 4577 // CAUTION: The following closure has persistent state that may need to
4577 4578 // be reset upon a decrease in the sequence of addresses it
4578 4579 // processes.
4579 4580 ScanMarkedObjectsAgainCarefullyClosure
4580 4581 smoac_cl(this, _span,
4581 4582 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield);
4582 4583
4583 4584 // Preclean dirty cards in ModUnionTable and CardTable using
4584 4585 // appropriate convergence criterion;
4585 4586 // repeat CMSPrecleanIter times unless we find that
4586 4587 // we are losing.
4587 4588 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4588 4589 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4589 4590 "Bad convergence multiplier");
4590 4591 assert(CMSPrecleanThreshold >= 100,
4591 4592 "Unreasonably low CMSPrecleanThreshold");
4592 4593
4593 4594 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4594 4595 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4595 4596 numIter < CMSPrecleanIter;
4596 4597 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4597 4598 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4598 4599 if (CMSPermGenPrecleaningEnabled) {
4599 4600 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl);
4600 4601 }
4601 4602 if (Verbose && PrintGCDetails) {
4602 4603 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4603 4604 }
4604 4605 // Either there are very few dirty cards, so re-mark
4605 4606 // pause will be small anyway, or our pre-cleaning isn't
4606 4607 // that much faster than the rate at which cards are being
4607 4608 // dirtied, so we might as well stop and re-mark since
4608 4609 // precleaning won't improve our re-mark time by much.
4609 4610 if (curNumCards <= CMSPrecleanThreshold ||
4610 4611 (numIter > 0 &&
4611 4612 (curNumCards * CMSPrecleanDenominator >
4612 4613 lastNumCards * CMSPrecleanNumerator))) {
4613 4614 numIter++;
4614 4615 cumNumCards += curNumCards;
4615 4616 break;
4616 4617 }
4617 4618 }
4618 4619 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4619 4620 if (CMSPermGenPrecleaningEnabled) {
4620 4621 curNumCards += preclean_card_table(_permGen, &smoac_cl);
4621 4622 }
4622 4623 cumNumCards += curNumCards;
4623 4624 if (PrintGCDetails && PrintCMSStatistics != 0) {
4624 4625 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4625 4626 curNumCards, cumNumCards, numIter);
4626 4627 }
4627 4628 return cumNumCards; // as a measure of useful work done
4628 4629 }
4629 4630
4630 4631 // PRECLEANING NOTES:
4631 4632 // Precleaning involves:
4632 4633 // . reading the bits of the modUnionTable and clearing the set bits.
4633 4634 // . For the cards corresponding to the set bits, we scan the
4634 4635 // objects on those cards. This means we need the free_list_lock
4635 4636 // so that we can safely iterate over the CMS space when scanning
4636 4637 // for oops.
4637 4638 // . When we scan the objects, we'll be both reading and setting
4638 4639 // marks in the marking bit map, so we'll need the marking bit map.
4639 4640 // . For protecting _collector_state transitions, we take the CGC_lock.
4640 4641 // Note that any races in the reading of of card table entries by the
4641 4642 // CMS thread on the one hand and the clearing of those entries by the
4642 4643 // VM thread or the setting of those entries by the mutator threads on the
4643 4644 // other are quite benign. However, for efficiency it makes sense to keep
4644 4645 // the VM thread from racing with the CMS thread while the latter is
4645 4646 // dirty card info to the modUnionTable. We therefore also use the
4646 4647 // CGC_lock to protect the reading of the card table and the mod union
4647 4648 // table by the CM thread.
4648 4649 // . We run concurrently with mutator updates, so scanning
4649 4650 // needs to be done carefully -- we should not try to scan
4650 4651 // potentially uninitialized objects.
4651 4652 //
4652 4653 // Locking strategy: While holding the CGC_lock, we scan over and
4653 4654 // reset a maximal dirty range of the mod union / card tables, then lock
4654 4655 // the free_list_lock and bitmap lock to do a full marking, then
4655 4656 // release these locks; and repeat the cycle. This allows for a
4656 4657 // certain amount of fairness in the sharing of these locks between
4657 4658 // the CMS collector on the one hand, and the VM thread and the
4658 4659 // mutators on the other.
4659 4660
4660 4661 // NOTE: preclean_mod_union_table() and preclean_card_table()
4661 4662 // further below are largely identical; if you need to modify
4662 4663 // one of these methods, please check the other method too.
4663 4664
4664 4665 size_t CMSCollector::preclean_mod_union_table(
4665 4666 ConcurrentMarkSweepGeneration* gen,
4666 4667 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4667 4668 verify_work_stacks_empty();
4668 4669 verify_overflow_empty();
4669 4670
4670 4671 // Turn off checking for this method but turn it back on
4671 4672 // selectively. There are yield points in this method
4672 4673 // but it is difficult to turn the checking off just around
4673 4674 // the yield points. It is simpler to selectively turn
4674 4675 // it on.
4675 4676 DEBUG_ONLY(RememberKlassesChecker mux(false);)
4676 4677
4677 4678 // strategy: starting with the first card, accumulate contiguous
4678 4679 // ranges of dirty cards; clear these cards, then scan the region
4679 4680 // covered by these cards.
4680 4681
4681 4682 // Since all of the MUT is committed ahead, we can just use
4682 4683 // that, in case the generations expand while we are precleaning.
4683 4684 // It might also be fine to just use the committed part of the
4684 4685 // generation, but we might potentially miss cards when the
4685 4686 // generation is rapidly expanding while we are in the midst
4686 4687 // of precleaning.
4687 4688 HeapWord* startAddr = gen->reserved().start();
4688 4689 HeapWord* endAddr = gen->reserved().end();
4689 4690
4690 4691 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4691 4692
4692 4693 size_t numDirtyCards, cumNumDirtyCards;
4693 4694 HeapWord *nextAddr, *lastAddr;
4694 4695 for (cumNumDirtyCards = numDirtyCards = 0,
4695 4696 nextAddr = lastAddr = startAddr;
4696 4697 nextAddr < endAddr;
4697 4698 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4698 4699
4699 4700 ResourceMark rm;
4700 4701 HandleMark hm;
4701 4702
4702 4703 MemRegion dirtyRegion;
4703 4704 {
4704 4705 stopTimer();
4705 4706 // Potential yield point
4706 4707 CMSTokenSync ts(true);
4707 4708 startTimer();
4708 4709 sample_eden();
4709 4710 // Get dirty region starting at nextOffset (inclusive),
4710 4711 // simultaneously clearing it.
4711 4712 dirtyRegion =
4712 4713 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4713 4714 assert(dirtyRegion.start() >= nextAddr,
4714 4715 "returned region inconsistent?");
4715 4716 }
4716 4717 // Remember where the next search should begin.
4717 4718 // The returned region (if non-empty) is a right open interval,
4718 4719 // so lastOffset is obtained from the right end of that
4719 4720 // interval.
4720 4721 lastAddr = dirtyRegion.end();
4721 4722 // Should do something more transparent and less hacky XXX
4722 4723 numDirtyCards =
4723 4724 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4724 4725
4725 4726 // We'll scan the cards in the dirty region (with periodic
4726 4727 // yields for foreground GC as needed).
4727 4728 if (!dirtyRegion.is_empty()) {
4728 4729 assert(numDirtyCards > 0, "consistency check");
4729 4730 HeapWord* stop_point = NULL;
4730 4731 stopTimer();
4731 4732 // Potential yield point
4732 4733 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4733 4734 bitMapLock());
4734 4735 startTimer();
4735 4736 {
4736 4737 verify_work_stacks_empty();
4737 4738 verify_overflow_empty();
4738 4739 sample_eden();
4739 4740 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4740 4741 stop_point =
4741 4742 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4742 4743 }
4743 4744 if (stop_point != NULL) {
4744 4745 // The careful iteration stopped early either because it found an
4745 4746 // uninitialized object, or because we were in the midst of an
4746 4747 // "abortable preclean", which should now be aborted. Redirty
4747 4748 // the bits corresponding to the partially-scanned or unscanned
4748 4749 // cards. We'll either restart at the next block boundary or
4749 4750 // abort the preclean.
4750 4751 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) ||
4751 4752 (_collectorState == AbortablePreclean && should_abort_preclean()),
4752 4753 "Unparsable objects should only be in perm gen.");
4753 4754 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4754 4755 if (should_abort_preclean()) {
4755 4756 break; // out of preclean loop
4756 4757 } else {
4757 4758 // Compute the next address at which preclean should pick up;
4758 4759 // might need bitMapLock in order to read P-bits.
4759 4760 lastAddr = next_card_start_after_block(stop_point);
4760 4761 }
4761 4762 }
4762 4763 } else {
4763 4764 assert(lastAddr == endAddr, "consistency check");
4764 4765 assert(numDirtyCards == 0, "consistency check");
4765 4766 break;
4766 4767 }
4767 4768 }
4768 4769 verify_work_stacks_empty();
4769 4770 verify_overflow_empty();
4770 4771 return cumNumDirtyCards;
4771 4772 }
4772 4773
4773 4774 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4774 4775 // below are largely identical; if you need to modify
4775 4776 // one of these methods, please check the other method too.
4776 4777
4777 4778 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4778 4779 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4779 4780 // strategy: it's similar to precleamModUnionTable above, in that
4780 4781 // we accumulate contiguous ranges of dirty cards, mark these cards
4781 4782 // precleaned, then scan the region covered by these cards.
4782 4783 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4783 4784 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4784 4785
4785 4786 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4786 4787
4787 4788 size_t numDirtyCards, cumNumDirtyCards;
4788 4789 HeapWord *lastAddr, *nextAddr;
4789 4790
4790 4791 for (cumNumDirtyCards = numDirtyCards = 0,
4791 4792 nextAddr = lastAddr = startAddr;
4792 4793 nextAddr < endAddr;
4793 4794 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4794 4795
4795 4796 ResourceMark rm;
4796 4797 HandleMark hm;
4797 4798
4798 4799 MemRegion dirtyRegion;
4799 4800 {
4800 4801 // See comments in "Precleaning notes" above on why we
4801 4802 // do this locking. XXX Could the locking overheads be
4802 4803 // too high when dirty cards are sparse? [I don't think so.]
4803 4804 stopTimer();
4804 4805 CMSTokenSync x(true); // is cms thread
4805 4806 startTimer();
4806 4807 sample_eden();
4807 4808 // Get and clear dirty region from card table
4808 4809 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4809 4810 MemRegion(nextAddr, endAddr),
4810 4811 true,
4811 4812 CardTableModRefBS::precleaned_card_val());
4812 4813
4813 4814 assert(dirtyRegion.start() >= nextAddr,
4814 4815 "returned region inconsistent?");
4815 4816 }
4816 4817 lastAddr = dirtyRegion.end();
4817 4818 numDirtyCards =
4818 4819 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4819 4820
4820 4821 if (!dirtyRegion.is_empty()) {
4821 4822 stopTimer();
4822 4823 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4823 4824 startTimer();
4824 4825 sample_eden();
4825 4826 verify_work_stacks_empty();
4826 4827 verify_overflow_empty();
4827 4828 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4828 4829 HeapWord* stop_point =
4829 4830 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4830 4831 if (stop_point != NULL) {
4831 4832 // The careful iteration stopped early because it found an
4832 4833 // uninitialized object. Redirty the bits corresponding to the
4833 4834 // partially-scanned or unscanned cards, and start again at the
4834 4835 // next block boundary.
4835 4836 assert(CMSPermGenPrecleaningEnabled ||
4836 4837 (_collectorState == AbortablePreclean && should_abort_preclean()),
4837 4838 "Unparsable objects should only be in perm gen.");
4838 4839 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4839 4840 if (should_abort_preclean()) {
4840 4841 break; // out of preclean loop
4841 4842 } else {
4842 4843 // Compute the next address at which preclean should pick up.
4843 4844 lastAddr = next_card_start_after_block(stop_point);
4844 4845 }
4845 4846 }
4846 4847 } else {
4847 4848 break;
4848 4849 }
4849 4850 }
4850 4851 verify_work_stacks_empty();
4851 4852 verify_overflow_empty();
4852 4853 return cumNumDirtyCards;
4853 4854 }
4854 4855
4855 4856 void CMSCollector::checkpointRootsFinal(bool asynch,
4856 4857 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4857 4858 assert(_collectorState == FinalMarking, "incorrect state transition?");
4858 4859 check_correct_thread_executing();
4859 4860 // world is stopped at this checkpoint
4860 4861 assert(SafepointSynchronize::is_at_safepoint(),
4861 4862 "world should be stopped");
4862 4863 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4863 4864
4864 4865 verify_work_stacks_empty();
4865 4866 verify_overflow_empty();
4866 4867
4867 4868 SpecializationStats::clear();
4868 4869 if (PrintGCDetails) {
4869 4870 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4870 4871 _young_gen->used() / K,
4871 4872 _young_gen->capacity() / K);
4872 4873 }
4873 4874 if (asynch) {
4874 4875 if (CMSScavengeBeforeRemark) {
4875 4876 GenCollectedHeap* gch = GenCollectedHeap::heap();
4876 4877 // Temporarily set flag to false, GCH->do_collection will
4877 4878 // expect it to be false and set to true
4878 4879 FlagSetting fl(gch->_is_gc_active, false);
4879 4880 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark",
4880 4881 PrintGCDetails && Verbose, true, gclog_or_tty);)
4881 4882 int level = _cmsGen->level() - 1;
4882 4883 if (level >= 0) {
4883 4884 gch->do_collection(true, // full (i.e. force, see below)
4884 4885 false, // !clear_all_soft_refs
4885 4886 0, // size
4886 4887 false, // is_tlab
4887 4888 level // max_level
4888 4889 );
4889 4890 }
4890 4891 }
4891 4892 FreelistLocker x(this);
4892 4893 MutexLockerEx y(bitMapLock(),
4893 4894 Mutex::_no_safepoint_check_flag);
4894 4895 assert(!init_mark_was_synchronous, "but that's impossible!");
4895 4896 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
4896 4897 } else {
4897 4898 // already have all the locks
4898 4899 checkpointRootsFinalWork(asynch, clear_all_soft_refs,
4899 4900 init_mark_was_synchronous);
4900 4901 }
4901 4902 verify_work_stacks_empty();
4902 4903 verify_overflow_empty();
4903 4904 SpecializationStats::print();
4904 4905 }
4905 4906
4906 4907 void CMSCollector::checkpointRootsFinalWork(bool asynch,
4907 4908 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4908 4909
4909 4910 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);)
4910 4911
4911 4912 assert(haveFreelistLocks(), "must have free list locks");
4912 4913 assert_lock_strong(bitMapLock());
4913 4914
4914 4915 if (UseAdaptiveSizePolicy) {
4915 4916 size_policy()->checkpoint_roots_final_begin();
4916 4917 }
4917 4918
4918 4919 ResourceMark rm;
4919 4920 HandleMark hm;
4920 4921
4921 4922 GenCollectedHeap* gch = GenCollectedHeap::heap();
4922 4923
4923 4924 if (should_unload_classes()) {
4924 4925 CodeCache::gc_prologue();
4925 4926 }
4926 4927 assert(haveFreelistLocks(), "must have free list locks");
4927 4928 assert_lock_strong(bitMapLock());
4928 4929
4929 4930 DEBUG_ONLY(RememberKlassesChecker fmx(should_unload_classes());)
4930 4931 if (!init_mark_was_synchronous) {
4931 4932 // We might assume that we need not fill TLAB's when
4932 4933 // CMSScavengeBeforeRemark is set, because we may have just done
4933 4934 // a scavenge which would have filled all TLAB's -- and besides
4934 4935 // Eden would be empty. This however may not always be the case --
4935 4936 // for instance although we asked for a scavenge, it may not have
4936 4937 // happened because of a JNI critical section. We probably need
4937 4938 // a policy for deciding whether we can in that case wait until
4938 4939 // the critical section releases and then do the remark following
4939 4940 // the scavenge, and skip it here. In the absence of that policy,
4940 4941 // or of an indication of whether the scavenge did indeed occur,
4941 4942 // we cannot rely on TLAB's having been filled and must do
4942 4943 // so here just in case a scavenge did not happen.
4943 4944 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4944 4945 // Update the saved marks which may affect the root scans.
4945 4946 gch->save_marks();
4946 4947
4947 4948 {
4948 4949 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4949 4950
4950 4951 // Note on the role of the mod union table:
4951 4952 // Since the marker in "markFromRoots" marks concurrently with
4952 4953 // mutators, it is possible for some reachable objects not to have been
4953 4954 // scanned. For instance, an only reference to an object A was
4954 4955 // placed in object B after the marker scanned B. Unless B is rescanned,
4955 4956 // A would be collected. Such updates to references in marked objects
4956 4957 // are detected via the mod union table which is the set of all cards
4957 4958 // dirtied since the first checkpoint in this GC cycle and prior to
4958 4959 // the most recent young generation GC, minus those cleaned up by the
4959 4960 // concurrent precleaning.
4960 4961 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
4961 4962 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty);
4962 4963 do_remark_parallel();
4963 4964 } else {
4964 4965 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4965 4966 gclog_or_tty);
4966 4967 do_remark_non_parallel();
4967 4968 }
4968 4969 }
4969 4970 } else {
4970 4971 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
4971 4972 // The initial mark was stop-world, so there's no rescanning to
4972 4973 // do; go straight on to the next step below.
4973 4974 }
4974 4975 verify_work_stacks_empty();
4975 4976 verify_overflow_empty();
4976 4977
4977 4978 {
4978 4979 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);)
4979 4980 refProcessingWork(asynch, clear_all_soft_refs);
4980 4981 }
4981 4982 verify_work_stacks_empty();
4982 4983 verify_overflow_empty();
4983 4984
4984 4985 if (should_unload_classes()) {
4985 4986 CodeCache::gc_epilogue();
4986 4987 }
4987 4988 JvmtiExport::gc_epilogue();
4988 4989
4989 4990 // If we encountered any (marking stack / work queue) overflow
4990 4991 // events during the current CMS cycle, take appropriate
4991 4992 // remedial measures, where possible, so as to try and avoid
4992 4993 // recurrence of that condition.
4993 4994 assert(_markStack.isEmpty(), "No grey objects");
4994 4995 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4995 4996 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4996 4997 if (ser_ovflw > 0) {
4997 4998 if (PrintCMSStatistics != 0) {
4998 4999 gclog_or_tty->print_cr("Marking stack overflow (benign) "
4999 5000 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
5000 5001 ", kac_preclean="SIZE_FORMAT")",
5001 5002 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
5002 5003 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
5003 5004 }
5004 5005 _markStack.expand();
5005 5006 _ser_pmc_remark_ovflw = 0;
5006 5007 _ser_pmc_preclean_ovflw = 0;
5007 5008 _ser_kac_preclean_ovflw = 0;
5008 5009 _ser_kac_ovflw = 0;
5009 5010 }
5010 5011 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
5011 5012 if (PrintCMSStatistics != 0) {
5012 5013 gclog_or_tty->print_cr("Work queue overflow (benign) "
5013 5014 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
5014 5015 _par_pmc_remark_ovflw, _par_kac_ovflw);
5015 5016 }
5016 5017 _par_pmc_remark_ovflw = 0;
5017 5018 _par_kac_ovflw = 0;
5018 5019 }
5019 5020 if (PrintCMSStatistics != 0) {
5020 5021 if (_markStack._hit_limit > 0) {
5021 5022 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
5022 5023 _markStack._hit_limit);
5023 5024 }
5024 5025 if (_markStack._failed_double > 0) {
5025 5026 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
5026 5027 " current capacity "SIZE_FORMAT,
5027 5028 _markStack._failed_double,
5028 5029 _markStack.capacity());
5029 5030 }
5030 5031 }
5031 5032 _markStack._hit_limit = 0;
5032 5033 _markStack._failed_double = 0;
5033 5034
5034 5035 // Check that all the klasses have been checked
5035 5036 assert(_revisitStack.isEmpty(), "Not all klasses revisited");
5036 5037
5037 5038 if ((VerifyAfterGC || VerifyDuringGC) &&
5038 5039 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5039 5040 verify_after_remark();
5040 5041 }
5041 5042
5042 5043 // Change under the freelistLocks.
5043 5044 _collectorState = Sweeping;
5044 5045 // Call isAllClear() under bitMapLock
5045 5046 assert(_modUnionTable.isAllClear(), "Should be clear by end of the"
5046 5047 " final marking");
5047 5048 if (UseAdaptiveSizePolicy) {
5048 5049 size_policy()->checkpoint_roots_final_end(gch->gc_cause());
5049 5050 }
5050 5051 }
5051 5052
5052 5053 // Parallel remark task
5053 5054 class CMSParRemarkTask: public AbstractGangTask {
5054 5055 CMSCollector* _collector;
5055 5056 int _n_workers;
5056 5057 CompactibleFreeListSpace* _cms_space;
5057 5058 CompactibleFreeListSpace* _perm_space;
5058 5059
5059 5060 // The per-thread work queues, available here for stealing.
5060 5061 OopTaskQueueSet* _task_queues;
5061 5062 ParallelTaskTerminator _term;
5062 5063
5063 5064 public:
5064 5065 CMSParRemarkTask(CMSCollector* collector,
5065 5066 CompactibleFreeListSpace* cms_space,
5066 5067 CompactibleFreeListSpace* perm_space,
5067 5068 int n_workers, FlexibleWorkGang* workers,
5068 5069 OopTaskQueueSet* task_queues):
5069 5070 AbstractGangTask("Rescan roots and grey objects in parallel"),
5070 5071 _collector(collector),
5071 5072 _cms_space(cms_space), _perm_space(perm_space),
5072 5073 _n_workers(n_workers),
5073 5074 _task_queues(task_queues),
5074 5075 _term(n_workers, task_queues) { }
5075 5076
5076 5077 OopTaskQueueSet* task_queues() { return _task_queues; }
5077 5078
5078 5079 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5079 5080
5080 5081 ParallelTaskTerminator* terminator() { return &_term; }
5081 5082 int n_workers() { return _n_workers; }
5082 5083
5083 5084 void work(int i);
5084 5085
5085 5086 private:
5086 5087 // Work method in support of parallel rescan ... of young gen spaces
5087 5088 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl,
5088 5089 ContiguousSpace* space,
5089 5090 HeapWord** chunk_array, size_t chunk_top);
5090 5091
5091 5092 // ... of dirty cards in old space
5092 5093 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
5093 5094 Par_MarkRefsIntoAndScanClosure* cl);
5094 5095
5095 5096 // ... work stealing for the above
5096 5097 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
5097 5098 };
5098 5099
5099 5100 // work_queue(i) is passed to the closure
5100 5101 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
5101 5102 // also is passed to do_dirty_card_rescan_tasks() and to
5102 5103 // do_work_steal() to select the i-th task_queue.
5103 5104
5104 5105 void CMSParRemarkTask::work(int i) {
5105 5106 elapsedTimer _timer;
5106 5107 ResourceMark rm;
5107 5108 HandleMark hm;
5108 5109
5109 5110 // ---------- rescan from roots --------------
5110 5111 _timer.start();
5111 5112 GenCollectedHeap* gch = GenCollectedHeap::heap();
5112 5113 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
5113 5114 _collector->_span, _collector->ref_processor(),
5114 5115 &(_collector->_markBitMap),
5115 5116 work_queue(i), &(_collector->_revisitStack));
5116 5117
5117 5118 // Rescan young gen roots first since these are likely
5118 5119 // coarsely partitioned and may, on that account, constitute
5119 5120 // the critical path; thus, it's best to start off that
5120 5121 // work first.
5121 5122 // ---------- young gen roots --------------
5122 5123 {
5123 5124 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
5124 5125 EdenSpace* eden_space = dng->eden();
5125 5126 ContiguousSpace* from_space = dng->from();
5126 5127 ContiguousSpace* to_space = dng->to();
5127 5128
5128 5129 HeapWord** eca = _collector->_eden_chunk_array;
5129 5130 size_t ect = _collector->_eden_chunk_index;
5130 5131 HeapWord** sca = _collector->_survivor_chunk_array;
5131 5132 size_t sct = _collector->_survivor_chunk_index;
5132 5133
5133 5134 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
5134 5135 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
5135 5136
5136 5137 do_young_space_rescan(i, &par_mrias_cl, to_space, NULL, 0);
5137 5138 do_young_space_rescan(i, &par_mrias_cl, from_space, sca, sct);
5138 5139 do_young_space_rescan(i, &par_mrias_cl, eden_space, eca, ect);
5139 5140
5140 5141 _timer.stop();
5141 5142 if (PrintCMSStatistics != 0) {
5142 5143 gclog_or_tty->print_cr(
5143 5144 "Finished young gen rescan work in %dth thread: %3.3f sec",
5144 5145 i, _timer.seconds());
5145 5146 }
5146 5147 }
5147 5148
5148 5149 // ---------- remaining roots --------------
5149 5150 _timer.reset();
5150 5151 _timer.start();
5151 5152 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
5152 5153 false, // yg was scanned above
5153 5154 false, // this is parallel code
5154 5155 true, // collecting perm gen
5155 5156 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5156 5157 &par_mrias_cl,
5157 5158 true, // walk all of code cache if (so & SO_CodeCache)
5158 5159 NULL);
5159 5160 assert(_collector->should_unload_classes()
5160 5161 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache),
5161 5162 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5162 5163 _timer.stop();
5163 5164 if (PrintCMSStatistics != 0) {
5164 5165 gclog_or_tty->print_cr(
5165 5166 "Finished remaining root rescan work in %dth thread: %3.3f sec",
5166 5167 i, _timer.seconds());
5167 5168 }
5168 5169
5169 5170 // ---------- rescan dirty cards ------------
5170 5171 _timer.reset();
5171 5172 _timer.start();
5172 5173
5173 5174 // Do the rescan tasks for each of the two spaces
5174 5175 // (cms_space and perm_space) in turn.
5175 5176 // "i" is passed to select the "i-th" task_queue
5176 5177 do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl);
5177 5178 do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl);
5178 5179 _timer.stop();
5179 5180 if (PrintCMSStatistics != 0) {
5180 5181 gclog_or_tty->print_cr(
5181 5182 "Finished dirty card rescan work in %dth thread: %3.3f sec",
5182 5183 i, _timer.seconds());
5183 5184 }
5184 5185
5185 5186 // ---------- steal work from other threads ...
5186 5187 // ---------- ... and drain overflow list.
5187 5188 _timer.reset();
5188 5189 _timer.start();
5189 5190 do_work_steal(i, &par_mrias_cl, _collector->hash_seed(i));
5190 5191 _timer.stop();
5191 5192 if (PrintCMSStatistics != 0) {
5192 5193 gclog_or_tty->print_cr(
5193 5194 "Finished work stealing in %dth thread: %3.3f sec",
5194 5195 i, _timer.seconds());
5195 5196 }
5196 5197 }
5197 5198
5198 5199 // Note that parameter "i" is not used.
5199 5200 void
5200 5201 CMSParRemarkTask::do_young_space_rescan(int i,
5201 5202 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space,
5202 5203 HeapWord** chunk_array, size_t chunk_top) {
5203 5204 // Until all tasks completed:
5204 5205 // . claim an unclaimed task
5205 5206 // . compute region boundaries corresponding to task claimed
5206 5207 // using chunk_array
5207 5208 // . par_oop_iterate(cl) over that region
5208 5209
5209 5210 ResourceMark rm;
5210 5211 HandleMark hm;
5211 5212
5212 5213 SequentialSubTasksDone* pst = space->par_seq_tasks();
5213 5214 assert(pst->valid(), "Uninitialized use?");
5214 5215
5215 5216 int nth_task = 0;
5216 5217 int n_tasks = pst->n_tasks();
5217 5218
5218 5219 HeapWord *start, *end;
5219 5220 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5220 5221 // We claimed task # nth_task; compute its boundaries.
5221 5222 if (chunk_top == 0) { // no samples were taken
5222 5223 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
5223 5224 start = space->bottom();
5224 5225 end = space->top();
5225 5226 } else if (nth_task == 0) {
5226 5227 start = space->bottom();
5227 5228 end = chunk_array[nth_task];
5228 5229 } else if (nth_task < (jint)chunk_top) {
5229 5230 assert(nth_task >= 1, "Control point invariant");
5230 5231 start = chunk_array[nth_task - 1];
5231 5232 end = chunk_array[nth_task];
5232 5233 } else {
5233 5234 assert(nth_task == (jint)chunk_top, "Control point invariant");
5234 5235 start = chunk_array[chunk_top - 1];
5235 5236 end = space->top();
5236 5237 }
5237 5238 MemRegion mr(start, end);
5238 5239 // Verify that mr is in space
5239 5240 assert(mr.is_empty() || space->used_region().contains(mr),
5240 5241 "Should be in space");
5241 5242 // Verify that "start" is an object boundary
5242 5243 assert(mr.is_empty() || oop(mr.start())->is_oop(),
5243 5244 "Should be an oop");
5244 5245 space->par_oop_iterate(mr, cl);
5245 5246 }
5246 5247 pst->all_tasks_completed();
5247 5248 }
5248 5249
5249 5250 void
5250 5251 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5251 5252 CompactibleFreeListSpace* sp, int i,
5252 5253 Par_MarkRefsIntoAndScanClosure* cl) {
5253 5254 // Until all tasks completed:
5254 5255 // . claim an unclaimed task
5255 5256 // . compute region boundaries corresponding to task claimed
5256 5257 // . transfer dirty bits ct->mut for that region
5257 5258 // . apply rescanclosure to dirty mut bits for that region
5258 5259
5259 5260 ResourceMark rm;
5260 5261 HandleMark hm;
5261 5262
5262 5263 OopTaskQueue* work_q = work_queue(i);
5263 5264 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5264 5265 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5265 5266 // CAUTION: This closure has state that persists across calls to
5266 5267 // the work method dirty_range_iterate_clear() in that it has
5267 5268 // imbedded in it a (subtype of) UpwardsObjectClosure. The
5268 5269 // use of that state in the imbedded UpwardsObjectClosure instance
5269 5270 // assumes that the cards are always iterated (even if in parallel
5270 5271 // by several threads) in monotonically increasing order per each
5271 5272 // thread. This is true of the implementation below which picks
5272 5273 // card ranges (chunks) in monotonically increasing order globally
5273 5274 // and, a-fortiori, in monotonically increasing order per thread
5274 5275 // (the latter order being a subsequence of the former).
5275 5276 // If the work code below is ever reorganized into a more chaotic
5276 5277 // work-partitioning form than the current "sequential tasks"
5277 5278 // paradigm, the use of that persistent state will have to be
5278 5279 // revisited and modified appropriately. See also related
5279 5280 // bug 4756801 work on which should examine this code to make
5280 5281 // sure that the changes there do not run counter to the
5281 5282 // assumptions made here and necessary for correctness and
5282 5283 // efficiency. Note also that this code might yield inefficient
5283 5284 // behaviour in the case of very large objects that span one or
5284 5285 // more work chunks. Such objects would potentially be scanned
5285 5286 // several times redundantly. Work on 4756801 should try and
5286 5287 // address that performance anomaly if at all possible. XXX
5287 5288 MemRegion full_span = _collector->_span;
5288 5289 CMSBitMap* bm = &(_collector->_markBitMap); // shared
5289 5290 CMSMarkStack* rs = &(_collector->_revisitStack); // shared
5290 5291 MarkFromDirtyCardsClosure
5291 5292 greyRescanClosure(_collector, full_span, // entire span of interest
5292 5293 sp, bm, work_q, rs, cl);
5293 5294
5294 5295 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5295 5296 assert(pst->valid(), "Uninitialized use?");
5296 5297 int nth_task = 0;
5297 5298 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5298 5299 MemRegion span = sp->used_region();
5299 5300 HeapWord* start_addr = span.start();
5300 5301 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5301 5302 alignment);
5302 5303 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5303 5304 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5304 5305 start_addr, "Check alignment");
5305 5306 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5306 5307 chunk_size, "Check alignment");
5307 5308
5308 5309 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5309 5310 // Having claimed the nth_task, compute corresponding mem-region,
5310 5311 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
5311 5312 // The alignment restriction ensures that we do not need any
5312 5313 // synchronization with other gang-workers while setting or
5313 5314 // clearing bits in thus chunk of the MUT.
5314 5315 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5315 5316 start_addr + (nth_task+1)*chunk_size);
5316 5317 // The last chunk's end might be way beyond end of the
5317 5318 // used region. In that case pull back appropriately.
5318 5319 if (this_span.end() > end_addr) {
5319 5320 this_span.set_end(end_addr);
5320 5321 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5321 5322 }
5322 5323 // Iterate over the dirty cards covering this chunk, marking them
5323 5324 // precleaned, and setting the corresponding bits in the mod union
5324 5325 // table. Since we have been careful to partition at Card and MUT-word
5325 5326 // boundaries no synchronization is needed between parallel threads.
5326 5327 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5327 5328 &modUnionClosure);
5328 5329
5329 5330 // Having transferred these marks into the modUnionTable,
5330 5331 // rescan the marked objects on the dirty cards in the modUnionTable.
5331 5332 // Even if this is at a synchronous collection, the initial marking
5332 5333 // may have been done during an asynchronous collection so there
5333 5334 // may be dirty bits in the mod-union table.
5334 5335 _collector->_modUnionTable.dirty_range_iterate_clear(
5335 5336 this_span, &greyRescanClosure);
5336 5337 _collector->_modUnionTable.verifyNoOneBitsInRange(
5337 5338 this_span.start(),
5338 5339 this_span.end());
5339 5340 }
5340 5341 pst->all_tasks_completed(); // declare that i am done
5341 5342 }
5342 5343
5343 5344 // . see if we can share work_queues with ParNew? XXX
5344 5345 void
5345 5346 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5346 5347 int* seed) {
5347 5348 OopTaskQueue* work_q = work_queue(i);
5348 5349 NOT_PRODUCT(int num_steals = 0;)
5349 5350 oop obj_to_scan;
5350 5351 CMSBitMap* bm = &(_collector->_markBitMap);
5351 5352
5352 5353 while (true) {
5353 5354 // Completely finish any left over work from (an) earlier round(s)
5354 5355 cl->trim_queue(0);
5355 5356 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5356 5357 (size_t)ParGCDesiredObjsFromOverflowList);
5357 5358 // Now check if there's any work in the overflow list
5358 5359 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5359 5360 // only affects the number of attempts made to get work from the
5360 5361 // overflow list and does not affect the number of workers. Just
5361 5362 // pass ParallelGCThreads so this behavior is unchanged.
5362 5363 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5363 5364 work_q,
5364 5365 ParallelGCThreads)) {
5365 5366 // found something in global overflow list;
5366 5367 // not yet ready to go stealing work from others.
5367 5368 // We'd like to assert(work_q->size() != 0, ...)
5368 5369 // because we just took work from the overflow list,
5369 5370 // but of course we can't since all of that could have
5370 5371 // been already stolen from us.
5371 5372 // "He giveth and He taketh away."
5372 5373 continue;
5373 5374 }
5374 5375 // Verify that we have no work before we resort to stealing
5375 5376 assert(work_q->size() == 0, "Have work, shouldn't steal");
5376 5377 // Try to steal from other queues that have work
5377 5378 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5378 5379 NOT_PRODUCT(num_steals++;)
5379 5380 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5380 5381 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5381 5382 // Do scanning work
5382 5383 obj_to_scan->oop_iterate(cl);
5383 5384 // Loop around, finish this work, and try to steal some more
5384 5385 } else if (terminator()->offer_termination()) {
5385 5386 break; // nirvana from the infinite cycle
5386 5387 }
5387 5388 }
5388 5389 NOT_PRODUCT(
5389 5390 if (PrintCMSStatistics != 0) {
5390 5391 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5391 5392 }
5392 5393 )
5393 5394 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5394 5395 "Else our work is not yet done");
5395 5396 }
5396 5397
5397 5398 // Return a thread-local PLAB recording array, as appropriate.
5398 5399 void* CMSCollector::get_data_recorder(int thr_num) {
5399 5400 if (_survivor_plab_array != NULL &&
5400 5401 (CMSPLABRecordAlways ||
5401 5402 (_collectorState > Marking && _collectorState < FinalMarking))) {
5402 5403 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5403 5404 ChunkArray* ca = &_survivor_plab_array[thr_num];
5404 5405 ca->reset(); // clear it so that fresh data is recorded
5405 5406 return (void*) ca;
5406 5407 } else {
5407 5408 return NULL;
5408 5409 }
5409 5410 }
5410 5411
5411 5412 // Reset all the thread-local PLAB recording arrays
5412 5413 void CMSCollector::reset_survivor_plab_arrays() {
5413 5414 for (uint i = 0; i < ParallelGCThreads; i++) {
5414 5415 _survivor_plab_array[i].reset();
5415 5416 }
5416 5417 }
5417 5418
5418 5419 // Merge the per-thread plab arrays into the global survivor chunk
5419 5420 // array which will provide the partitioning of the survivor space
5420 5421 // for CMS rescan.
5421 5422 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
5422 5423 int no_of_gc_threads) {
5423 5424 assert(_survivor_plab_array != NULL, "Error");
5424 5425 assert(_survivor_chunk_array != NULL, "Error");
5425 5426 assert(_collectorState == FinalMarking, "Error");
5426 5427 for (int j = 0; j < no_of_gc_threads; j++) {
5427 5428 _cursor[j] = 0;
5428 5429 }
5429 5430 HeapWord* top = surv->top();
5430 5431 size_t i;
5431 5432 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
5432 5433 HeapWord* min_val = top; // Higher than any PLAB address
5433 5434 uint min_tid = 0; // position of min_val this round
5434 5435 for (int j = 0; j < no_of_gc_threads; j++) {
5435 5436 ChunkArray* cur_sca = &_survivor_plab_array[j];
5436 5437 if (_cursor[j] == cur_sca->end()) {
5437 5438 continue;
5438 5439 }
5439 5440 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5440 5441 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5441 5442 assert(surv->used_region().contains(cur_val), "Out of bounds value");
5442 5443 if (cur_val < min_val) {
5443 5444 min_tid = j;
5444 5445 min_val = cur_val;
5445 5446 } else {
5446 5447 assert(cur_val < top, "All recorded addresses should be less");
5447 5448 }
5448 5449 }
5449 5450 // At this point min_val and min_tid are respectively
5450 5451 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5451 5452 // and the thread (j) that witnesses that address.
5452 5453 // We record this address in the _survivor_chunk_array[i]
5453 5454 // and increment _cursor[min_tid] prior to the next round i.
5454 5455 if (min_val == top) {
5455 5456 break;
5456 5457 }
5457 5458 _survivor_chunk_array[i] = min_val;
5458 5459 _cursor[min_tid]++;
5459 5460 }
5460 5461 // We are all done; record the size of the _survivor_chunk_array
5461 5462 _survivor_chunk_index = i; // exclusive: [0, i)
5462 5463 if (PrintCMSStatistics > 0) {
5463 5464 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5464 5465 }
5465 5466 // Verify that we used up all the recorded entries
5466 5467 #ifdef ASSERT
5467 5468 size_t total = 0;
5468 5469 for (int j = 0; j < no_of_gc_threads; j++) {
5469 5470 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5470 5471 total += _cursor[j];
5471 5472 }
5472 5473 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5473 5474 // Check that the merged array is in sorted order
5474 5475 if (total > 0) {
5475 5476 for (size_t i = 0; i < total - 1; i++) {
5476 5477 if (PrintCMSStatistics > 0) {
5477 5478 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5478 5479 i, _survivor_chunk_array[i]);
5479 5480 }
5480 5481 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5481 5482 "Not sorted");
5482 5483 }
5483 5484 }
5484 5485 #endif // ASSERT
5485 5486 }
5486 5487
5487 5488 // Set up the space's par_seq_tasks structure for work claiming
5488 5489 // for parallel rescan of young gen.
5489 5490 // See ParRescanTask where this is currently used.
5490 5491 void
5491 5492 CMSCollector::
5492 5493 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5493 5494 assert(n_threads > 0, "Unexpected n_threads argument");
5494 5495 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5495 5496
5496 5497 // Eden space
5497 5498 {
5498 5499 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5499 5500 assert(!pst->valid(), "Clobbering existing data?");
5500 5501 // Each valid entry in [0, _eden_chunk_index) represents a task.
5501 5502 size_t n_tasks = _eden_chunk_index + 1;
5502 5503 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5503 5504 // Sets the condition for completion of the subtask (how many threads
5504 5505 // need to finish in order to be done).
5505 5506 pst->set_n_threads(n_threads);
5506 5507 pst->set_n_tasks((int)n_tasks);
5507 5508 }
5508 5509
5509 5510 // Merge the survivor plab arrays into _survivor_chunk_array
5510 5511 if (_survivor_plab_array != NULL) {
5511 5512 merge_survivor_plab_arrays(dng->from(), n_threads);
5512 5513 } else {
5513 5514 assert(_survivor_chunk_index == 0, "Error");
5514 5515 }
5515 5516
5516 5517 // To space
5517 5518 {
5518 5519 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5519 5520 assert(!pst->valid(), "Clobbering existing data?");
5520 5521 // Sets the condition for completion of the subtask (how many threads
5521 5522 // need to finish in order to be done).
5522 5523 pst->set_n_threads(n_threads);
5523 5524 pst->set_n_tasks(1);
5524 5525 assert(pst->valid(), "Error");
5525 5526 }
5526 5527
5527 5528 // From space
5528 5529 {
5529 5530 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5530 5531 assert(!pst->valid(), "Clobbering existing data?");
5531 5532 size_t n_tasks = _survivor_chunk_index + 1;
5532 5533 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5533 5534 // Sets the condition for completion of the subtask (how many threads
5534 5535 // need to finish in order to be done).
5535 5536 pst->set_n_threads(n_threads);
5536 5537 pst->set_n_tasks((int)n_tasks);
5537 5538 assert(pst->valid(), "Error");
5538 5539 }
5539 5540 }
5540 5541
5541 5542 // Parallel version of remark
5542 5543 void CMSCollector::do_remark_parallel() {
5543 5544 GenCollectedHeap* gch = GenCollectedHeap::heap();
5544 5545 FlexibleWorkGang* workers = gch->workers();
5545 5546 assert(workers != NULL, "Need parallel worker threads.");
5546 5547 int n_workers = workers->total_workers();
5547 5548 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5548 5549 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
5549 5550
5550 5551 CMSParRemarkTask tsk(this,
5551 5552 cms_space, perm_space,
5552 5553 n_workers, workers, task_queues());
5553 5554
5554 5555 // Set up for parallel process_strong_roots work.
5555 5556 gch->set_par_threads(n_workers);
5556 5557 // We won't be iterating over the cards in the card table updating
5557 5558 // the younger_gen cards, so we shouldn't call the following else
5558 5559 // the verification code as well as subsequent younger_refs_iterate
5559 5560 // code would get confused. XXX
5560 5561 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5561 5562
5562 5563 // The young gen rescan work will not be done as part of
5563 5564 // process_strong_roots (which currently doesn't knw how to
5564 5565 // parallelize such a scan), but rather will be broken up into
5565 5566 // a set of parallel tasks (via the sampling that the [abortable]
5566 5567 // preclean phase did of EdenSpace, plus the [two] tasks of
5567 5568 // scanning the [two] survivor spaces. Further fine-grain
5568 5569 // parallelization of the scanning of the survivor spaces
5569 5570 // themselves, and of precleaning of the younger gen itself
5570 5571 // is deferred to the future.
5571 5572 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5572 5573
5573 5574 // The dirty card rescan work is broken up into a "sequence"
5574 5575 // of parallel tasks (per constituent space) that are dynamically
5575 5576 // claimed by the parallel threads.
5576 5577 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5577 5578 perm_space->initialize_sequential_subtasks_for_rescan(n_workers);
5578 5579
5579 5580 // It turns out that even when we're using 1 thread, doing the work in a
5580 5581 // separate thread causes wide variance in run times. We can't help this
5581 5582 // in the multi-threaded case, but we special-case n=1 here to get
5582 5583 // repeatable measurements of the 1-thread overhead of the parallel code.
5583 5584 if (n_workers > 1) {
5584 5585 // Make refs discovery MT-safe, if it isn't already: it may not
5585 5586 // necessarily be so, since it's possible that we are doing
5586 5587 // ST marking.
5587 5588 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5588 5589 GenCollectedHeap::StrongRootsScope srs(gch);
5589 5590 workers->run_task(&tsk);
5590 5591 } else {
5591 5592 GenCollectedHeap::StrongRootsScope srs(gch);
5592 5593 tsk.work(0);
5593 5594 }
5594 5595 gch->set_par_threads(0); // 0 ==> non-parallel.
5595 5596 // restore, single-threaded for now, any preserved marks
5596 5597 // as a result of work_q overflow
5597 5598 restore_preserved_marks_if_any();
5598 5599 }
5599 5600
5600 5601 // Non-parallel version of remark
5601 5602 void CMSCollector::do_remark_non_parallel() {
5602 5603 ResourceMark rm;
5603 5604 HandleMark hm;
5604 5605 GenCollectedHeap* gch = GenCollectedHeap::heap();
5605 5606 MarkRefsIntoAndScanClosure
5606 5607 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
5607 5608 &_markStack, &_revisitStack, this,
5608 5609 false /* should_yield */, false /* not precleaning */);
5609 5610 MarkFromDirtyCardsClosure
5610 5611 markFromDirtyCardsClosure(this, _span,
5611 5612 NULL, // space is set further below
5612 5613 &_markBitMap, &_markStack, &_revisitStack,
5613 5614 &mrias_cl);
5614 5615 {
5615 5616 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty);
5616 5617 // Iterate over the dirty cards, setting the corresponding bits in the
5617 5618 // mod union table.
5618 5619 {
5619 5620 ModUnionClosure modUnionClosure(&_modUnionTable);
5620 5621 _ct->ct_bs()->dirty_card_iterate(
5621 5622 _cmsGen->used_region(),
5622 5623 &modUnionClosure);
5623 5624 _ct->ct_bs()->dirty_card_iterate(
5624 5625 _permGen->used_region(),
5625 5626 &modUnionClosure);
5626 5627 }
5627 5628 // Having transferred these marks into the modUnionTable, we just need
5628 5629 // to rescan the marked objects on the dirty cards in the modUnionTable.
5629 5630 // The initial marking may have been done during an asynchronous
5630 5631 // collection so there may be dirty bits in the mod-union table.
5631 5632 const int alignment =
5632 5633 CardTableModRefBS::card_size * BitsPerWord;
5633 5634 {
5634 5635 // ... First handle dirty cards in CMS gen
5635 5636 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5636 5637 MemRegion ur = _cmsGen->used_region();
5637 5638 HeapWord* lb = ur.start();
5638 5639 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5639 5640 MemRegion cms_span(lb, ub);
5640 5641 _modUnionTable.dirty_range_iterate_clear(cms_span,
5641 5642 &markFromDirtyCardsClosure);
5642 5643 verify_work_stacks_empty();
5643 5644 if (PrintCMSStatistics != 0) {
5644 5645 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5645 5646 markFromDirtyCardsClosure.num_dirty_cards());
5646 5647 }
5647 5648 }
5648 5649 {
5649 5650 // .. and then repeat for dirty cards in perm gen
5650 5651 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace());
5651 5652 MemRegion ur = _permGen->used_region();
5652 5653 HeapWord* lb = ur.start();
5653 5654 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5654 5655 MemRegion perm_span(lb, ub);
5655 5656 _modUnionTable.dirty_range_iterate_clear(perm_span,
5656 5657 &markFromDirtyCardsClosure);
5657 5658 verify_work_stacks_empty();
5658 5659 if (PrintCMSStatistics != 0) {
5659 5660 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ",
5660 5661 markFromDirtyCardsClosure.num_dirty_cards());
5661 5662 }
5662 5663 }
5663 5664 }
5664 5665 if (VerifyDuringGC &&
5665 5666 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5666 5667 HandleMark hm; // Discard invalid handles created during verification
5667 5668 Universe::verify(true);
5668 5669 }
5669 5670 {
5670 5671 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty);
5671 5672
5672 5673 verify_work_stacks_empty();
5673 5674
5674 5675 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5675 5676 GenCollectedHeap::StrongRootsScope srs(gch);
5676 5677 gch->gen_process_strong_roots(_cmsGen->level(),
5677 5678 true, // younger gens as roots
5678 5679 false, // use the local StrongRootsScope
5679 5680 true, // collecting perm gen
5680 5681 SharedHeap::ScanningOption(roots_scanning_options()),
5681 5682 &mrias_cl,
5682 5683 true, // walk code active on stacks
5683 5684 NULL);
5684 5685 assert(should_unload_classes()
5685 5686 || (roots_scanning_options() & SharedHeap::SO_CodeCache),
5686 5687 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5687 5688 }
5688 5689 verify_work_stacks_empty();
5689 5690 // Restore evacuated mark words, if any, used for overflow list links
5690 5691 if (!CMSOverflowEarlyRestoration) {
5691 5692 restore_preserved_marks_if_any();
5692 5693 }
5693 5694 verify_overflow_empty();
5694 5695 }
5695 5696
5696 5697 ////////////////////////////////////////////////////////
5697 5698 // Parallel Reference Processing Task Proxy Class
5698 5699 ////////////////////////////////////////////////////////
5699 5700 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5700 5701 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5701 5702 CMSCollector* _collector;
5702 5703 CMSBitMap* _mark_bit_map;
5703 5704 const MemRegion _span;
5704 5705 ProcessTask& _task;
5705 5706
5706 5707 public:
5707 5708 CMSRefProcTaskProxy(ProcessTask& task,
5708 5709 CMSCollector* collector,
5709 5710 const MemRegion& span,
5710 5711 CMSBitMap* mark_bit_map,
5711 5712 AbstractWorkGang* workers,
5712 5713 OopTaskQueueSet* task_queues):
5713 5714 // XXX Should superclass AGTWOQ also know about AWG since it knows
5714 5715 // about the task_queues used by the AWG? Then it could initialize
5715 5716 // the terminator() object. See 6984287. The set_for_termination()
5716 5717 // below is a temporary band-aid for the regression in 6984287.
5717 5718 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5718 5719 task_queues),
5719 5720 _task(task),
5720 5721 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5721 5722 {
5722 5723 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5723 5724 "Inconsistency in _span");
5724 5725 set_for_termination(workers->active_workers());
5725 5726 }
5726 5727
5727 5728 OopTaskQueueSet* task_queues() { return queues(); }
5728 5729
5729 5730 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5730 5731
5731 5732 void do_work_steal(int i,
5732 5733 CMSParDrainMarkingStackClosure* drain,
5733 5734 CMSParKeepAliveClosure* keep_alive,
5734 5735 int* seed);
5735 5736
5736 5737 virtual void work(int i);
5737 5738 };
5738 5739
5739 5740 void CMSRefProcTaskProxy::work(int i) {
5740 5741 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5741 5742 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5742 5743 _mark_bit_map,
5743 5744 &_collector->_revisitStack,
5744 5745 work_queue(i));
5745 5746 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5746 5747 _mark_bit_map,
5747 5748 &_collector->_revisitStack,
5748 5749 work_queue(i));
5749 5750 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5750 5751 _task.work(i, is_alive_closure, par_keep_alive, par_drain_stack);
5751 5752 if (_task.marks_oops_alive()) {
5752 5753 do_work_steal(i, &par_drain_stack, &par_keep_alive,
5753 5754 _collector->hash_seed(i));
5754 5755 }
5755 5756 assert(work_queue(i)->size() == 0, "work_queue should be empty");
5756 5757 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5757 5758 }
5758 5759
5759 5760 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5760 5761 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5761 5762 EnqueueTask& _task;
5762 5763
5763 5764 public:
5764 5765 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5765 5766 : AbstractGangTask("Enqueue reference objects in parallel"),
5766 5767 _task(task)
5767 5768 { }
5768 5769
5769 5770 virtual void work(int i)
5770 5771 {
5771 5772 _task.work(i);
5772 5773 }
5773 5774 };
5774 5775
5775 5776 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5776 5777 MemRegion span, CMSBitMap* bit_map, CMSMarkStack* revisit_stack,
5777 5778 OopTaskQueue* work_queue):
5778 5779 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
5779 5780 _span(span),
5780 5781 _bit_map(bit_map),
5781 5782 _work_queue(work_queue),
5782 5783 _mark_and_push(collector, span, bit_map, revisit_stack, work_queue),
5783 5784 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5784 5785 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5785 5786 { }
5786 5787
5787 5788 // . see if we can share work_queues with ParNew? XXX
5788 5789 void CMSRefProcTaskProxy::do_work_steal(int i,
5789 5790 CMSParDrainMarkingStackClosure* drain,
5790 5791 CMSParKeepAliveClosure* keep_alive,
5791 5792 int* seed) {
5792 5793 OopTaskQueue* work_q = work_queue(i);
5793 5794 NOT_PRODUCT(int num_steals = 0;)
5794 5795 oop obj_to_scan;
5795 5796
5796 5797 while (true) {
5797 5798 // Completely finish any left over work from (an) earlier round(s)
5798 5799 drain->trim_queue(0);
5799 5800 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5800 5801 (size_t)ParGCDesiredObjsFromOverflowList);
5801 5802 // Now check if there's any work in the overflow list
5802 5803 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5803 5804 // only affects the number of attempts made to get work from the
5804 5805 // overflow list and does not affect the number of workers. Just
5805 5806 // pass ParallelGCThreads so this behavior is unchanged.
5806 5807 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5807 5808 work_q,
5808 5809 ParallelGCThreads)) {
5809 5810 // Found something in global overflow list;
5810 5811 // not yet ready to go stealing work from others.
5811 5812 // We'd like to assert(work_q->size() != 0, ...)
5812 5813 // because we just took work from the overflow list,
5813 5814 // but of course we can't, since all of that might have
5814 5815 // been already stolen from us.
5815 5816 continue;
5816 5817 }
5817 5818 // Verify that we have no work before we resort to stealing
5818 5819 assert(work_q->size() == 0, "Have work, shouldn't steal");
5819 5820 // Try to steal from other queues that have work
5820 5821 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5821 5822 NOT_PRODUCT(num_steals++;)
5822 5823 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5823 5824 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5824 5825 // Do scanning work
5825 5826 obj_to_scan->oop_iterate(keep_alive);
5826 5827 // Loop around, finish this work, and try to steal some more
5827 5828 } else if (terminator()->offer_termination()) {
5828 5829 break; // nirvana from the infinite cycle
5829 5830 }
5830 5831 }
5831 5832 NOT_PRODUCT(
5832 5833 if (PrintCMSStatistics != 0) {
5833 5834 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5834 5835 }
5835 5836 )
5836 5837 }
5837 5838
5838 5839 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5839 5840 {
5840 5841 GenCollectedHeap* gch = GenCollectedHeap::heap();
5841 5842 FlexibleWorkGang* workers = gch->workers();
5842 5843 assert(workers != NULL, "Need parallel worker threads.");
5843 5844 CMSRefProcTaskProxy rp_task(task, &_collector,
5844 5845 _collector.ref_processor()->span(),
5845 5846 _collector.markBitMap(),
5846 5847 workers, _collector.task_queues());
5847 5848 workers->run_task(&rp_task);
5848 5849 }
5849 5850
5850 5851 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5851 5852 {
5852 5853
5853 5854 GenCollectedHeap* gch = GenCollectedHeap::heap();
5854 5855 FlexibleWorkGang* workers = gch->workers();
5855 5856 assert(workers != NULL, "Need parallel worker threads.");
5856 5857 CMSRefEnqueueTaskProxy enq_task(task);
5857 5858 workers->run_task(&enq_task);
5858 5859 }
5859 5860
5860 5861 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
5861 5862
5862 5863 ResourceMark rm;
5863 5864 HandleMark hm;
5864 5865
5865 5866 ReferenceProcessor* rp = ref_processor();
5866 5867 assert(rp->span().equals(_span), "Spans should be equal");
5867 5868 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5868 5869 // Process weak references.
5869 5870 rp->setup_policy(clear_all_soft_refs);
5870 5871 verify_work_stacks_empty();
5871 5872
5872 5873 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5873 5874 &_markStack, &_revisitStack,
5874 5875 false /* !preclean */);
5875 5876 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5876 5877 _span, &_markBitMap, &_markStack,
5877 5878 &cmsKeepAliveClosure, false /* !preclean */);
5878 5879 {
5879 5880 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty);
5880 5881 if (rp->processing_is_mt()) {
5881 5882 // Set the degree of MT here. If the discovery is done MT, there
5882 5883 // may have been a different number of threads doing the discovery
5883 5884 // and a different number of discovered lists may have Ref objects.
5884 5885 // That is OK as long as the Reference lists are balanced (see
5885 5886 // balance_all_queues() and balance_queues()).
5886 5887
5887 5888 rp->set_active_mt_degree(ParallelGCThreads);
5888 5889 CMSRefProcTaskExecutor task_executor(*this);
5889 5890 rp->process_discovered_references(&_is_alive_closure,
5890 5891 &cmsKeepAliveClosure,
5891 5892 &cmsDrainMarkingStackClosure,
5892 5893 &task_executor);
5893 5894 } else {
5894 5895 rp->process_discovered_references(&_is_alive_closure,
5895 5896 &cmsKeepAliveClosure,
5896 5897 &cmsDrainMarkingStackClosure,
5897 5898 NULL);
5898 5899 }
5899 5900 verify_work_stacks_empty();
5900 5901 }
5901 5902
5902 5903 if (should_unload_classes()) {
5903 5904 {
5904 5905 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty);
5905 5906
5906 5907 // Follow SystemDictionary roots and unload classes
5907 5908 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5908 5909
5909 5910 // Follow CodeCache roots and unload any methods marked for unloading
5910 5911 CodeCache::do_unloading(&_is_alive_closure,
5911 5912 &cmsKeepAliveClosure,
5912 5913 purged_class);
5913 5914
5914 5915 cmsDrainMarkingStackClosure.do_void();
5915 5916 verify_work_stacks_empty();
5916 5917
5917 5918 // Update subklass/sibling/implementor links in KlassKlass descendants
5918 5919 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty");
5919 5920 oop k;
5920 5921 while ((k = _revisitStack.pop()) != NULL) {
5921 5922 ((Klass*)(oopDesc*)k)->follow_weak_klass_links(
5922 5923 &_is_alive_closure,
5923 5924 &cmsKeepAliveClosure);
5924 5925 }
5925 5926 assert(!ClassUnloading ||
5926 5927 (_markStack.isEmpty() && overflow_list_is_empty()),
5927 5928 "Should not have found new reachable objects");
5928 5929 assert(_revisitStack.isEmpty(), "revisit stack should have been drained");
5929 5930 cmsDrainMarkingStackClosure.do_void();
5930 5931 verify_work_stacks_empty();
5931 5932 }
5932 5933
5933 5934 {
5934 5935 TraceTime t("scrub symbol table", PrintGCDetails, false, gclog_or_tty);
5935 5936 // Clean up unreferenced symbols in symbol table.
5936 5937 SymbolTable::unlink();
5937 5938 }
5938 5939 }
5939 5940
5940 5941 if (should_unload_classes() || !JavaObjectsInPerm) {
5941 5942 TraceTime t("scrub string table", PrintGCDetails, false, gclog_or_tty);
5942 5943 // Now clean up stale oops in StringTable
5943 5944 StringTable::unlink(&_is_alive_closure);
5944 5945 }
5945 5946
5946 5947 verify_work_stacks_empty();
5947 5948 // Restore any preserved marks as a result of mark stack or
5948 5949 // work queue overflow
5949 5950 restore_preserved_marks_if_any(); // done single-threaded for now
5950 5951
5951 5952 rp->set_enqueuing_is_done(true);
5952 5953 if (rp->processing_is_mt()) {
5953 5954 rp->balance_all_queues();
5954 5955 CMSRefProcTaskExecutor task_executor(*this);
5955 5956 rp->enqueue_discovered_references(&task_executor);
5956 5957 } else {
5957 5958 rp->enqueue_discovered_references(NULL);
5958 5959 }
5959 5960 rp->verify_no_references_recorded();
5960 5961 assert(!rp->discovery_enabled(), "should have been disabled");
5961 5962 }
5962 5963
5963 5964 #ifndef PRODUCT
5964 5965 void CMSCollector::check_correct_thread_executing() {
5965 5966 Thread* t = Thread::current();
5966 5967 // Only the VM thread or the CMS thread should be here.
5967 5968 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5968 5969 "Unexpected thread type");
5969 5970 // If this is the vm thread, the foreground process
5970 5971 // should not be waiting. Note that _foregroundGCIsActive is
5971 5972 // true while the foreground collector is waiting.
5972 5973 if (_foregroundGCShouldWait) {
5973 5974 // We cannot be the VM thread
5974 5975 assert(t->is_ConcurrentGC_thread(),
5975 5976 "Should be CMS thread");
5976 5977 } else {
5977 5978 // We can be the CMS thread only if we are in a stop-world
5978 5979 // phase of CMS collection.
5979 5980 if (t->is_ConcurrentGC_thread()) {
5980 5981 assert(_collectorState == InitialMarking ||
5981 5982 _collectorState == FinalMarking,
5982 5983 "Should be a stop-world phase");
5983 5984 // The CMS thread should be holding the CMS_token.
5984 5985 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5985 5986 "Potential interference with concurrently "
5986 5987 "executing VM thread");
5987 5988 }
5988 5989 }
5989 5990 }
5990 5991 #endif
5991 5992
5992 5993 void CMSCollector::sweep(bool asynch) {
5993 5994 assert(_collectorState == Sweeping, "just checking");
5994 5995 check_correct_thread_executing();
5995 5996 verify_work_stacks_empty();
5996 5997 verify_overflow_empty();
5997 5998 increment_sweep_count();
5998 5999 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5999 6000
6000 6001 _inter_sweep_timer.stop();
6001 6002 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
6002 6003 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
6003 6004
6004 6005 // PermGen verification support: If perm gen sweeping is disabled in
6005 6006 // this cycle, we preserve the perm gen object "deadness" information
6006 6007 // in the perm_gen_verify_bit_map. In order to do that we traverse
6007 6008 // all blocks in perm gen and mark all dead objects.
6008 6009 if (verifying() && !should_unload_classes()) {
6009 6010 assert(perm_gen_verify_bit_map()->sizeInBits() != 0,
6010 6011 "Should have already been allocated");
6011 6012 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(),
6012 6013 markBitMap(), perm_gen_verify_bit_map());
6013 6014 if (asynch) {
6014 6015 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
6015 6016 bitMapLock());
6016 6017 _permGen->cmsSpace()->blk_iterate(&mdo);
6017 6018 } else {
6018 6019 // In the case of synchronous sweep, we already have
6019 6020 // the requisite locks/tokens.
6020 6021 _permGen->cmsSpace()->blk_iterate(&mdo);
6021 6022 }
6022 6023 }
6023 6024
6024 6025 assert(!_intra_sweep_timer.is_active(), "Should not be active");
6025 6026 _intra_sweep_timer.reset();
6026 6027 _intra_sweep_timer.start();
6027 6028 if (asynch) {
6028 6029 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6029 6030 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
6030 6031 // First sweep the old gen then the perm gen
6031 6032 {
6032 6033 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6033 6034 bitMapLock());
6034 6035 sweepWork(_cmsGen, asynch);
6035 6036 }
6036 6037
6037 6038 // Now repeat for perm gen
6038 6039 if (should_unload_classes()) {
6039 6040 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
6040 6041 bitMapLock());
6041 6042 sweepWork(_permGen, asynch);
6042 6043 }
6043 6044
6044 6045 // Update Universe::_heap_*_at_gc figures.
6045 6046 // We need all the free list locks to make the abstract state
6046 6047 // transition from Sweeping to Resetting. See detailed note
6047 6048 // further below.
6048 6049 {
6049 6050 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6050 6051 _permGen->freelistLock());
6051 6052 // Update heap occupancy information which is used as
6052 6053 // input to soft ref clearing policy at the next gc.
6053 6054 Universe::update_heap_info_at_gc();
6054 6055 _collectorState = Resizing;
6055 6056 }
6056 6057 } else {
6057 6058 // already have needed locks
6058 6059 sweepWork(_cmsGen, asynch);
6059 6060
6060 6061 if (should_unload_classes()) {
6061 6062 sweepWork(_permGen, asynch);
6062 6063 }
6063 6064 // Update heap occupancy information which is used as
6064 6065 // input to soft ref clearing policy at the next gc.
6065 6066 Universe::update_heap_info_at_gc();
6066 6067 _collectorState = Resizing;
6067 6068 }
6068 6069 verify_work_stacks_empty();
6069 6070 verify_overflow_empty();
6070 6071
6071 6072 _intra_sweep_timer.stop();
6072 6073 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
6073 6074
6074 6075 _inter_sweep_timer.reset();
6075 6076 _inter_sweep_timer.start();
6076 6077
6077 6078 update_time_of_last_gc(os::javaTimeMillis());
6078 6079
6079 6080 // NOTE on abstract state transitions:
6080 6081 // Mutators allocate-live and/or mark the mod-union table dirty
6081 6082 // based on the state of the collection. The former is done in
6082 6083 // the interval [Marking, Sweeping] and the latter in the interval
6083 6084 // [Marking, Sweeping). Thus the transitions into the Marking state
6084 6085 // and out of the Sweeping state must be synchronously visible
6085 6086 // globally to the mutators.
6086 6087 // The transition into the Marking state happens with the world
6087 6088 // stopped so the mutators will globally see it. Sweeping is
6088 6089 // done asynchronously by the background collector so the transition
6089 6090 // from the Sweeping state to the Resizing state must be done
6090 6091 // under the freelistLock (as is the check for whether to
6091 6092 // allocate-live and whether to dirty the mod-union table).
6092 6093 assert(_collectorState == Resizing, "Change of collector state to"
6093 6094 " Resizing must be done under the freelistLocks (plural)");
6094 6095
6095 6096 // Now that sweeping has been completed, we clear
6096 6097 // the incremental_collection_failed flag,
6097 6098 // thus inviting a younger gen collection to promote into
6098 6099 // this generation. If such a promotion may still fail,
6099 6100 // the flag will be set again when a young collection is
6100 6101 // attempted.
6101 6102 GenCollectedHeap* gch = GenCollectedHeap::heap();
6102 6103 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
6103 6104 gch->update_full_collections_completed(_collection_count_start);
6104 6105 }
6105 6106
6106 6107 // FIX ME!!! Looks like this belongs in CFLSpace, with
6107 6108 // CMSGen merely delegating to it.
6108 6109 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
6109 6110 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
6110 6111 HeapWord* minAddr = _cmsSpace->bottom();
6111 6112 HeapWord* largestAddr =
6112 6113 (HeapWord*) _cmsSpace->dictionary()->findLargestDict();
6113 6114 if (largestAddr == NULL) {
6114 6115 // The dictionary appears to be empty. In this case
6115 6116 // try to coalesce at the end of the heap.
6116 6117 largestAddr = _cmsSpace->end();
6117 6118 }
6118 6119 size_t largestOffset = pointer_delta(largestAddr, minAddr);
6119 6120 size_t nearLargestOffset =
6120 6121 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
6121 6122 if (PrintFLSStatistics != 0) {
6122 6123 gclog_or_tty->print_cr(
6123 6124 "CMS: Large Block: " PTR_FORMAT ";"
6124 6125 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
6125 6126 largestAddr,
6126 6127 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset);
6127 6128 }
6128 6129 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
6129 6130 }
6130 6131
6131 6132 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
6132 6133 return addr >= _cmsSpace->nearLargestChunk();
6133 6134 }
6134 6135
6135 6136 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
6136 6137 return _cmsSpace->find_chunk_at_end();
6137 6138 }
6138 6139
6139 6140 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
6140 6141 bool full) {
6141 6142 // The next lower level has been collected. Gather any statistics
6142 6143 // that are of interest at this point.
6143 6144 if (!full && (current_level + 1) == level()) {
6144 6145 // Gather statistics on the young generation collection.
6145 6146 collector()->stats().record_gc0_end(used());
6146 6147 }
6147 6148 }
6148 6149
6149 6150 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
6150 6151 GenCollectedHeap* gch = GenCollectedHeap::heap();
6151 6152 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
6152 6153 "Wrong type of heap");
6153 6154 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
6154 6155 gch->gen_policy()->size_policy();
6155 6156 assert(sp->is_gc_cms_adaptive_size_policy(),
6156 6157 "Wrong type of size policy");
6157 6158 return sp;
6158 6159 }
6159 6160
6160 6161 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
6161 6162 if (PrintGCDetails && Verbose) {
6162 6163 gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
6163 6164 }
6164 6165 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
6165 6166 _debug_collection_type =
6166 6167 (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
6167 6168 if (PrintGCDetails && Verbose) {
6168 6169 gclog_or_tty->print_cr("to %d ", _debug_collection_type);
6169 6170 }
6170 6171 }
6171 6172
6172 6173 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
6173 6174 bool asynch) {
6174 6175 // We iterate over the space(s) underlying this generation,
6175 6176 // checking the mark bit map to see if the bits corresponding
6176 6177 // to specific blocks are marked or not. Blocks that are
6177 6178 // marked are live and are not swept up. All remaining blocks
6178 6179 // are swept up, with coalescing on-the-fly as we sweep up
6179 6180 // contiguous free and/or garbage blocks:
6180 6181 // We need to ensure that the sweeper synchronizes with allocators
6181 6182 // and stop-the-world collectors. In particular, the following
6182 6183 // locks are used:
6183 6184 // . CMS token: if this is held, a stop the world collection cannot occur
6184 6185 // . freelistLock: if this is held no allocation can occur from this
6185 6186 // generation by another thread
6186 6187 // . bitMapLock: if this is held, no other thread can access or update
6187 6188 //
6188 6189
6189 6190 // Note that we need to hold the freelistLock if we use
6190 6191 // block iterate below; else the iterator might go awry if
6191 6192 // a mutator (or promotion) causes block contents to change
6192 6193 // (for instance if the allocator divvies up a block).
6193 6194 // If we hold the free list lock, for all practical purposes
6194 6195 // young generation GC's can't occur (they'll usually need to
6195 6196 // promote), so we might as well prevent all young generation
6196 6197 // GC's while we do a sweeping step. For the same reason, we might
6197 6198 // as well take the bit map lock for the entire duration
6198 6199
6199 6200 // check that we hold the requisite locks
6200 6201 assert(have_cms_token(), "Should hold cms token");
6201 6202 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
6202 6203 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
6203 6204 "Should possess CMS token to sweep");
6204 6205 assert_lock_strong(gen->freelistLock());
6205 6206 assert_lock_strong(bitMapLock());
6206 6207
6207 6208 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
6208 6209 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
6209 6210 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
6210 6211 _inter_sweep_estimate.padded_average(),
6211 6212 _intra_sweep_estimate.padded_average());
6212 6213 gen->setNearLargestChunk();
6213 6214
6214 6215 {
6215 6216 SweepClosure sweepClosure(this, gen, &_markBitMap,
6216 6217 CMSYield && asynch);
6217 6218 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
6218 6219 // We need to free-up/coalesce garbage/blocks from a
6219 6220 // co-terminal free run. This is done in the SweepClosure
6220 6221 // destructor; so, do not remove this scope, else the
6221 6222 // end-of-sweep-census below will be off by a little bit.
6222 6223 }
6223 6224 gen->cmsSpace()->sweep_completed();
6224 6225 gen->cmsSpace()->endSweepFLCensus(sweep_count());
6225 6226 if (should_unload_classes()) { // unloaded classes this cycle,
6226 6227 _concurrent_cycles_since_last_unload = 0; // ... reset count
6227 6228 } else { // did not unload classes,
6228 6229 _concurrent_cycles_since_last_unload++; // ... increment count
6229 6230 }
6230 6231 }
6231 6232
6232 6233 // Reset CMS data structures (for now just the marking bit map)
6233 6234 // preparatory for the next cycle.
6234 6235 void CMSCollector::reset(bool asynch) {
6235 6236 GenCollectedHeap* gch = GenCollectedHeap::heap();
6236 6237 CMSAdaptiveSizePolicy* sp = size_policy();
6237 6238 AdaptiveSizePolicyOutput(sp, gch->total_collections());
6238 6239 if (asynch) {
6239 6240 CMSTokenSyncWithLocks ts(true, bitMapLock());
6240 6241
6241 6242 // If the state is not "Resetting", the foreground thread
6242 6243 // has done a collection and the resetting.
6243 6244 if (_collectorState != Resetting) {
6244 6245 assert(_collectorState == Idling, "The state should only change"
6245 6246 " because the foreground collector has finished the collection");
6246 6247 return;
6247 6248 }
6248 6249
6249 6250 // Clear the mark bitmap (no grey objects to start with)
6250 6251 // for the next cycle.
6251 6252 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6252 6253 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
6253 6254
6254 6255 HeapWord* curAddr = _markBitMap.startWord();
6255 6256 while (curAddr < _markBitMap.endWord()) {
6256 6257 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
6257 6258 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
6258 6259 _markBitMap.clear_large_range(chunk);
6259 6260 if (ConcurrentMarkSweepThread::should_yield() &&
6260 6261 !foregroundGCIsActive() &&
6261 6262 CMSYield) {
6262 6263 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6263 6264 "CMS thread should hold CMS token");
6264 6265 assert_lock_strong(bitMapLock());
6265 6266 bitMapLock()->unlock();
6266 6267 ConcurrentMarkSweepThread::desynchronize(true);
6267 6268 ConcurrentMarkSweepThread::acknowledge_yield_request();
6268 6269 stopTimer();
6269 6270 if (PrintCMSStatistics != 0) {
6270 6271 incrementYields();
6271 6272 }
6272 6273 icms_wait();
6273 6274
6274 6275 // See the comment in coordinator_yield()
6275 6276 for (unsigned i = 0; i < CMSYieldSleepCount &&
6276 6277 ConcurrentMarkSweepThread::should_yield() &&
6277 6278 !CMSCollector::foregroundGCIsActive(); ++i) {
6278 6279 os::sleep(Thread::current(), 1, false);
6279 6280 ConcurrentMarkSweepThread::acknowledge_yield_request();
6280 6281 }
6281 6282
6282 6283 ConcurrentMarkSweepThread::synchronize(true);
6283 6284 bitMapLock()->lock_without_safepoint_check();
6284 6285 startTimer();
6285 6286 }
6286 6287 curAddr = chunk.end();
6287 6288 }
6288 6289 // A successful mostly concurrent collection has been done.
6289 6290 // Because only the full (i.e., concurrent mode failure) collections
6290 6291 // are being measured for gc overhead limits, clean the "near" flag
6291 6292 // and count.
6292 6293 sp->reset_gc_overhead_limit_count();
6293 6294 _collectorState = Idling;
6294 6295 } else {
6295 6296 // already have the lock
6296 6297 assert(_collectorState == Resetting, "just checking");
6297 6298 assert_lock_strong(bitMapLock());
6298 6299 _markBitMap.clear_all();
6299 6300 _collectorState = Idling;
6300 6301 }
6301 6302
6302 6303 // Stop incremental mode after a cycle completes, so that any future cycles
6303 6304 // are triggered by allocation.
6304 6305 stop_icms();
6305 6306
6306 6307 NOT_PRODUCT(
6307 6308 if (RotateCMSCollectionTypes) {
6308 6309 _cmsGen->rotate_debug_collection_type();
6309 6310 }
6310 6311 )
6311 6312 }
6312 6313
6313 6314 void CMSCollector::do_CMS_operation(CMS_op_type op) {
6314 6315 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
6315 6316 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6316 6317 TraceTime t("GC", PrintGC, !PrintGCDetails, gclog_or_tty);
6317 6318 TraceCollectorStats tcs(counters());
6318 6319
6319 6320 switch (op) {
6320 6321 case CMS_op_checkpointRootsInitial: {
6321 6322 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6322 6323 checkpointRootsInitial(true); // asynch
6323 6324 if (PrintGC) {
6324 6325 _cmsGen->printOccupancy("initial-mark");
6325 6326 }
6326 6327 break;
6327 6328 }
6328 6329 case CMS_op_checkpointRootsFinal: {
6329 6330 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6330 6331 checkpointRootsFinal(true, // asynch
6331 6332 false, // !clear_all_soft_refs
6332 6333 false); // !init_mark_was_synchronous
6333 6334 if (PrintGC) {
6334 6335 _cmsGen->printOccupancy("remark");
6335 6336 }
6336 6337 break;
6337 6338 }
6338 6339 default:
6339 6340 fatal("No such CMS_op");
6340 6341 }
6341 6342 }
6342 6343
6343 6344 #ifndef PRODUCT
6344 6345 size_t const CMSCollector::skip_header_HeapWords() {
6345 6346 return FreeChunk::header_size();
6346 6347 }
6347 6348
6348 6349 // Try and collect here conditions that should hold when
6349 6350 // CMS thread is exiting. The idea is that the foreground GC
6350 6351 // thread should not be blocked if it wants to terminate
6351 6352 // the CMS thread and yet continue to run the VM for a while
6352 6353 // after that.
6353 6354 void CMSCollector::verify_ok_to_terminate() const {
6354 6355 assert(Thread::current()->is_ConcurrentGC_thread(),
6355 6356 "should be called by CMS thread");
6356 6357 assert(!_foregroundGCShouldWait, "should be false");
6357 6358 // We could check here that all the various low-level locks
6358 6359 // are not held by the CMS thread, but that is overkill; see
6359 6360 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6360 6361 // is checked.
6361 6362 }
6362 6363 #endif
6363 6364
6364 6365 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6365 6366 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6366 6367 "missing Printezis mark?");
6367 6368 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6368 6369 size_t size = pointer_delta(nextOneAddr + 1, addr);
6369 6370 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6370 6371 "alignment problem");
6371 6372 assert(size >= 3, "Necessary for Printezis marks to work");
6372 6373 return size;
6373 6374 }
6374 6375
6375 6376 // A variant of the above (block_size_using_printezis_bits()) except
6376 6377 // that we return 0 if the P-bits are not yet set.
6377 6378 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6378 6379 if (_markBitMap.isMarked(addr + 1)) {
6379 6380 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
6380 6381 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6381 6382 size_t size = pointer_delta(nextOneAddr + 1, addr);
6382 6383 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6383 6384 "alignment problem");
6384 6385 assert(size >= 3, "Necessary for Printezis marks to work");
6385 6386 return size;
6386 6387 }
6387 6388 return 0;
6388 6389 }
6389 6390
6390 6391 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6391 6392 size_t sz = 0;
6392 6393 oop p = (oop)addr;
6393 6394 if (p->klass_or_null() != NULL && p->is_parsable()) {
6394 6395 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6395 6396 } else {
6396 6397 sz = block_size_using_printezis_bits(addr);
6397 6398 }
6398 6399 assert(sz > 0, "size must be nonzero");
6399 6400 HeapWord* next_block = addr + sz;
6400 6401 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
6401 6402 CardTableModRefBS::card_size);
6402 6403 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
6403 6404 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6404 6405 "must be different cards");
6405 6406 return next_card;
6406 6407 }
6407 6408
6408 6409
6409 6410 // CMS Bit Map Wrapper /////////////////////////////////////////
6410 6411
6411 6412 // Construct a CMS bit map infrastructure, but don't create the
6412 6413 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6413 6414 // further below.
6414 6415 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6415 6416 _bm(),
6416 6417 _shifter(shifter),
6417 6418 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6418 6419 {
6419 6420 _bmStartWord = 0;
6420 6421 _bmWordSize = 0;
6421 6422 }
6422 6423
6423 6424 bool CMSBitMap::allocate(MemRegion mr) {
6424 6425 _bmStartWord = mr.start();
6425 6426 _bmWordSize = mr.word_size();
6426 6427 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6427 6428 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6428 6429 if (!brs.is_reserved()) {
6429 6430 warning("CMS bit map allocation failure");
6430 6431 return false;
6431 6432 }
6432 6433 // For now we'll just commit all of the bit map up fromt.
6433 6434 // Later on we'll try to be more parsimonious with swap.
6434 6435 if (!_virtual_space.initialize(brs, brs.size())) {
6435 6436 warning("CMS bit map backing store failure");
6436 6437 return false;
6437 6438 }
6438 6439 assert(_virtual_space.committed_size() == brs.size(),
6439 6440 "didn't reserve backing store for all of CMS bit map?");
6440 6441 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
6441 6442 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6442 6443 _bmWordSize, "inconsistency in bit map sizing");
6443 6444 _bm.set_size(_bmWordSize >> _shifter);
6444 6445
6445 6446 // bm.clear(); // can we rely on getting zero'd memory? verify below
6446 6447 assert(isAllClear(),
6447 6448 "Expected zero'd memory from ReservedSpace constructor");
6448 6449 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6449 6450 "consistency check");
6450 6451 return true;
6451 6452 }
6452 6453
6453 6454 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6454 6455 HeapWord *next_addr, *end_addr, *last_addr;
6455 6456 assert_locked();
6456 6457 assert(covers(mr), "out-of-range error");
6457 6458 // XXX assert that start and end are appropriately aligned
6458 6459 for (next_addr = mr.start(), end_addr = mr.end();
6459 6460 next_addr < end_addr; next_addr = last_addr) {
6460 6461 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6461 6462 last_addr = dirty_region.end();
6462 6463 if (!dirty_region.is_empty()) {
6463 6464 cl->do_MemRegion(dirty_region);
6464 6465 } else {
6465 6466 assert(last_addr == end_addr, "program logic");
6466 6467 return;
6467 6468 }
6468 6469 }
6469 6470 }
6470 6471
6471 6472 #ifndef PRODUCT
6472 6473 void CMSBitMap::assert_locked() const {
6473 6474 CMSLockVerifier::assert_locked(lock());
6474 6475 }
6475 6476
6476 6477 bool CMSBitMap::covers(MemRegion mr) const {
6477 6478 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6478 6479 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6479 6480 "size inconsistency");
6480 6481 return (mr.start() >= _bmStartWord) &&
6481 6482 (mr.end() <= endWord());
6482 6483 }
6483 6484
6484 6485 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6485 6486 return (start >= _bmStartWord && (start + size) <= endWord());
6486 6487 }
6487 6488
6488 6489 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6489 6490 // verify that there are no 1 bits in the interval [left, right)
6490 6491 FalseBitMapClosure falseBitMapClosure;
6491 6492 iterate(&falseBitMapClosure, left, right);
6492 6493 }
6493 6494
6494 6495 void CMSBitMap::region_invariant(MemRegion mr)
6495 6496 {
6496 6497 assert_locked();
6497 6498 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6498 6499 assert(!mr.is_empty(), "unexpected empty region");
6499 6500 assert(covers(mr), "mr should be covered by bit map");
6500 6501 // convert address range into offset range
6501 6502 size_t start_ofs = heapWordToOffset(mr.start());
6502 6503 // Make sure that end() is appropriately aligned
6503 6504 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6504 6505 (1 << (_shifter+LogHeapWordSize))),
6505 6506 "Misaligned mr.end()");
6506 6507 size_t end_ofs = heapWordToOffset(mr.end());
6507 6508 assert(end_ofs > start_ofs, "Should mark at least one bit");
6508 6509 }
6509 6510
6510 6511 #endif
6511 6512
6512 6513 bool CMSMarkStack::allocate(size_t size) {
6513 6514 // allocate a stack of the requisite depth
6514 6515 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6515 6516 size * sizeof(oop)));
6516 6517 if (!rs.is_reserved()) {
6517 6518 warning("CMSMarkStack allocation failure");
6518 6519 return false;
6519 6520 }
6520 6521 if (!_virtual_space.initialize(rs, rs.size())) {
6521 6522 warning("CMSMarkStack backing store failure");
6522 6523 return false;
6523 6524 }
6524 6525 assert(_virtual_space.committed_size() == rs.size(),
6525 6526 "didn't reserve backing store for all of CMS stack?");
6526 6527 _base = (oop*)(_virtual_space.low());
6527 6528 _index = 0;
6528 6529 _capacity = size;
6529 6530 NOT_PRODUCT(_max_depth = 0);
6530 6531 return true;
6531 6532 }
6532 6533
6533 6534 // XXX FIX ME !!! In the MT case we come in here holding a
6534 6535 // leaf lock. For printing we need to take a further lock
6535 6536 // which has lower rank. We need to recallibrate the two
6536 6537 // lock-ranks involved in order to be able to rpint the
6537 6538 // messages below. (Or defer the printing to the caller.
6538 6539 // For now we take the expedient path of just disabling the
6539 6540 // messages for the problematic case.)
6540 6541 void CMSMarkStack::expand() {
6541 6542 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6542 6543 if (_capacity == MarkStackSizeMax) {
6543 6544 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6544 6545 // We print a warning message only once per CMS cycle.
6545 6546 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6546 6547 }
6547 6548 return;
6548 6549 }
6549 6550 // Double capacity if possible
6550 6551 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6551 6552 // Do not give up existing stack until we have managed to
6552 6553 // get the double capacity that we desired.
6553 6554 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6554 6555 new_capacity * sizeof(oop)));
6555 6556 if (rs.is_reserved()) {
6556 6557 // Release the backing store associated with old stack
6557 6558 _virtual_space.release();
6558 6559 // Reinitialize virtual space for new stack
6559 6560 if (!_virtual_space.initialize(rs, rs.size())) {
6560 6561 fatal("Not enough swap for expanded marking stack");
6561 6562 }
6562 6563 _base = (oop*)(_virtual_space.low());
6563 6564 _index = 0;
6564 6565 _capacity = new_capacity;
6565 6566 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6566 6567 // Failed to double capacity, continue;
6567 6568 // we print a detail message only once per CMS cycle.
6568 6569 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6569 6570 SIZE_FORMAT"K",
6570 6571 _capacity / K, new_capacity / K);
6571 6572 }
6572 6573 }
6573 6574
6574 6575
6575 6576 // Closures
6576 6577 // XXX: there seems to be a lot of code duplication here;
6577 6578 // should refactor and consolidate common code.
6578 6579
6579 6580 // This closure is used to mark refs into the CMS generation in
6580 6581 // the CMS bit map. Called at the first checkpoint. This closure
6581 6582 // assumes that we do not need to re-mark dirty cards; if the CMS
6582 6583 // generation on which this is used is not an oldest (modulo perm gen)
6583 6584 // generation then this will lose younger_gen cards!
6584 6585
6585 6586 MarkRefsIntoClosure::MarkRefsIntoClosure(
6586 6587 MemRegion span, CMSBitMap* bitMap):
6587 6588 _span(span),
6588 6589 _bitMap(bitMap)
6589 6590 {
6590 6591 assert(_ref_processor == NULL, "deliberately left NULL");
6591 6592 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6592 6593 }
6593 6594
6594 6595 void MarkRefsIntoClosure::do_oop(oop obj) {
6595 6596 // if p points into _span, then mark corresponding bit in _markBitMap
6596 6597 assert(obj->is_oop(), "expected an oop");
6597 6598 HeapWord* addr = (HeapWord*)obj;
6598 6599 if (_span.contains(addr)) {
6599 6600 // this should be made more efficient
6600 6601 _bitMap->mark(addr);
6601 6602 }
6602 6603 }
6603 6604
6604 6605 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6605 6606 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6606 6607
6607 6608 // A variant of the above, used for CMS marking verification.
6608 6609 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6609 6610 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6610 6611 _span(span),
6611 6612 _verification_bm(verification_bm),
6612 6613 _cms_bm(cms_bm)
6613 6614 {
6614 6615 assert(_ref_processor == NULL, "deliberately left NULL");
6615 6616 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6616 6617 }
6617 6618
6618 6619 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6619 6620 // if p points into _span, then mark corresponding bit in _markBitMap
6620 6621 assert(obj->is_oop(), "expected an oop");
6621 6622 HeapWord* addr = (HeapWord*)obj;
6622 6623 if (_span.contains(addr)) {
6623 6624 _verification_bm->mark(addr);
6624 6625 if (!_cms_bm->isMarked(addr)) {
6625 6626 oop(addr)->print();
6626 6627 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6627 6628 fatal("... aborting");
6628 6629 }
6629 6630 }
6630 6631 }
6631 6632
6632 6633 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6633 6634 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6634 6635
6635 6636 //////////////////////////////////////////////////
6636 6637 // MarkRefsIntoAndScanClosure
6637 6638 //////////////////////////////////////////////////
6638 6639
6639 6640 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6640 6641 ReferenceProcessor* rp,
6641 6642 CMSBitMap* bit_map,
6642 6643 CMSBitMap* mod_union_table,
6643 6644 CMSMarkStack* mark_stack,
6644 6645 CMSMarkStack* revisit_stack,
6645 6646 CMSCollector* collector,
6646 6647 bool should_yield,
6647 6648 bool concurrent_precleaning):
6648 6649 _collector(collector),
6649 6650 _span(span),
6650 6651 _bit_map(bit_map),
6651 6652 _mark_stack(mark_stack),
6652 6653 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6653 6654 mark_stack, revisit_stack, concurrent_precleaning),
6654 6655 _yield(should_yield),
6655 6656 _concurrent_precleaning(concurrent_precleaning),
6656 6657 _freelistLock(NULL)
6657 6658 {
6658 6659 _ref_processor = rp;
6659 6660 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6660 6661 }
6661 6662
6662 6663 // This closure is used to mark refs into the CMS generation at the
6663 6664 // second (final) checkpoint, and to scan and transitively follow
6664 6665 // the unmarked oops. It is also used during the concurrent precleaning
6665 6666 // phase while scanning objects on dirty cards in the CMS generation.
6666 6667 // The marks are made in the marking bit map and the marking stack is
6667 6668 // used for keeping the (newly) grey objects during the scan.
6668 6669 // The parallel version (Par_...) appears further below.
6669 6670 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6670 6671 if (obj != NULL) {
6671 6672 assert(obj->is_oop(), "expected an oop");
6672 6673 HeapWord* addr = (HeapWord*)obj;
6673 6674 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6674 6675 assert(_collector->overflow_list_is_empty(),
6675 6676 "overflow list should be empty");
6676 6677 if (_span.contains(addr) &&
6677 6678 !_bit_map->isMarked(addr)) {
6678 6679 // mark bit map (object is now grey)
6679 6680 _bit_map->mark(addr);
6680 6681 // push on marking stack (stack should be empty), and drain the
6681 6682 // stack by applying this closure to the oops in the oops popped
6682 6683 // from the stack (i.e. blacken the grey objects)
6683 6684 bool res = _mark_stack->push(obj);
6684 6685 assert(res, "Should have space to push on empty stack");
6685 6686 do {
6686 6687 oop new_oop = _mark_stack->pop();
6687 6688 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6688 6689 assert(new_oop->is_parsable(), "Found unparsable oop");
6689 6690 assert(_bit_map->isMarked((HeapWord*)new_oop),
6690 6691 "only grey objects on this stack");
6691 6692 // iterate over the oops in this oop, marking and pushing
6692 6693 // the ones in CMS heap (i.e. in _span).
6693 6694 new_oop->oop_iterate(&_pushAndMarkClosure);
6694 6695 // check if it's time to yield
6695 6696 do_yield_check();
6696 6697 } while (!_mark_stack->isEmpty() ||
6697 6698 (!_concurrent_precleaning && take_from_overflow_list()));
6698 6699 // if marking stack is empty, and we are not doing this
6699 6700 // during precleaning, then check the overflow list
6700 6701 }
6701 6702 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6702 6703 assert(_collector->overflow_list_is_empty(),
6703 6704 "overflow list was drained above");
6704 6705 // We could restore evacuated mark words, if any, used for
6705 6706 // overflow list links here because the overflow list is
6706 6707 // provably empty here. That would reduce the maximum
6707 6708 // size requirements for preserved_{oop,mark}_stack.
6708 6709 // But we'll just postpone it until we are all done
6709 6710 // so we can just stream through.
6710 6711 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6711 6712 _collector->restore_preserved_marks_if_any();
6712 6713 assert(_collector->no_preserved_marks(), "No preserved marks");
6713 6714 }
6714 6715 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6715 6716 "All preserved marks should have been restored above");
6716 6717 }
6717 6718 }
6718 6719
6719 6720 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6720 6721 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6721 6722
6722 6723 void MarkRefsIntoAndScanClosure::do_yield_work() {
6723 6724 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6724 6725 "CMS thread should hold CMS token");
6725 6726 assert_lock_strong(_freelistLock);
6726 6727 assert_lock_strong(_bit_map->lock());
6727 6728 // relinquish the free_list_lock and bitMaplock()
6728 6729 DEBUG_ONLY(RememberKlassesChecker mux(false);)
6729 6730 _bit_map->lock()->unlock();
6730 6731 _freelistLock->unlock();
6731 6732 ConcurrentMarkSweepThread::desynchronize(true);
6732 6733 ConcurrentMarkSweepThread::acknowledge_yield_request();
6733 6734 _collector->stopTimer();
6734 6735 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6735 6736 if (PrintCMSStatistics != 0) {
6736 6737 _collector->incrementYields();
6737 6738 }
6738 6739 _collector->icms_wait();
6739 6740
6740 6741 // See the comment in coordinator_yield()
6741 6742 for (unsigned i = 0;
6742 6743 i < CMSYieldSleepCount &&
6743 6744 ConcurrentMarkSweepThread::should_yield() &&
6744 6745 !CMSCollector::foregroundGCIsActive();
6745 6746 ++i) {
6746 6747 os::sleep(Thread::current(), 1, false);
6747 6748 ConcurrentMarkSweepThread::acknowledge_yield_request();
6748 6749 }
6749 6750
6750 6751 ConcurrentMarkSweepThread::synchronize(true);
6751 6752 _freelistLock->lock_without_safepoint_check();
6752 6753 _bit_map->lock()->lock_without_safepoint_check();
6753 6754 _collector->startTimer();
6754 6755 }
6755 6756
6756 6757 ///////////////////////////////////////////////////////////
6757 6758 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6758 6759 // MarkRefsIntoAndScanClosure
6759 6760 ///////////////////////////////////////////////////////////
6760 6761 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6761 6762 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6762 6763 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack):
6763 6764 _span(span),
6764 6765 _bit_map(bit_map),
6765 6766 _work_queue(work_queue),
6766 6767 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6767 6768 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6768 6769 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue,
6769 6770 revisit_stack)
6770 6771 {
6771 6772 _ref_processor = rp;
6772 6773 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6773 6774 }
6774 6775
6775 6776 // This closure is used to mark refs into the CMS generation at the
6776 6777 // second (final) checkpoint, and to scan and transitively follow
6777 6778 // the unmarked oops. The marks are made in the marking bit map and
6778 6779 // the work_queue is used for keeping the (newly) grey objects during
6779 6780 // the scan phase whence they are also available for stealing by parallel
6780 6781 // threads. Since the marking bit map is shared, updates are
6781 6782 // synchronized (via CAS).
6782 6783 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6783 6784 if (obj != NULL) {
6784 6785 // Ignore mark word because this could be an already marked oop
6785 6786 // that may be chained at the end of the overflow list.
6786 6787 assert(obj->is_oop(true), "expected an oop");
6787 6788 HeapWord* addr = (HeapWord*)obj;
6788 6789 if (_span.contains(addr) &&
6789 6790 !_bit_map->isMarked(addr)) {
6790 6791 // mark bit map (object will become grey):
6791 6792 // It is possible for several threads to be
6792 6793 // trying to "claim" this object concurrently;
6793 6794 // the unique thread that succeeds in marking the
6794 6795 // object first will do the subsequent push on
6795 6796 // to the work queue (or overflow list).
6796 6797 if (_bit_map->par_mark(addr)) {
6797 6798 // push on work_queue (which may not be empty), and trim the
6798 6799 // queue to an appropriate length by applying this closure to
6799 6800 // the oops in the oops popped from the stack (i.e. blacken the
6800 6801 // grey objects)
6801 6802 bool res = _work_queue->push(obj);
6802 6803 assert(res, "Low water mark should be less than capacity?");
6803 6804 trim_queue(_low_water_mark);
6804 6805 } // Else, another thread claimed the object
6805 6806 }
6806 6807 }
6807 6808 }
6808 6809
6809 6810 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6810 6811 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6811 6812
6812 6813 // This closure is used to rescan the marked objects on the dirty cards
6813 6814 // in the mod union table and the card table proper.
6814 6815 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6815 6816 oop p, MemRegion mr) {
6816 6817
6817 6818 size_t size = 0;
6818 6819 HeapWord* addr = (HeapWord*)p;
6819 6820 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6820 6821 assert(_span.contains(addr), "we are scanning the CMS generation");
6821 6822 // check if it's time to yield
6822 6823 if (do_yield_check()) {
6823 6824 // We yielded for some foreground stop-world work,
6824 6825 // and we have been asked to abort this ongoing preclean cycle.
6825 6826 return 0;
6826 6827 }
6827 6828 if (_bitMap->isMarked(addr)) {
6828 6829 // it's marked; is it potentially uninitialized?
6829 6830 if (p->klass_or_null() != NULL) {
6830 6831 // If is_conc_safe is false, the object may be undergoing
6831 6832 // change by the VM outside a safepoint. Don't try to
6832 6833 // scan it, but rather leave it for the remark phase.
6833 6834 if (CMSPermGenPrecleaningEnabled &&
6834 6835 (!p->is_conc_safe() || !p->is_parsable())) {
6835 6836 // Signal precleaning to redirty the card since
6836 6837 // the klass pointer is already installed.
6837 6838 assert(size == 0, "Initial value");
6838 6839 } else {
6839 6840 assert(p->is_parsable(), "must be parsable.");
6840 6841 // an initialized object; ignore mark word in verification below
6841 6842 // since we are running concurrent with mutators
6842 6843 assert(p->is_oop(true), "should be an oop");
6843 6844 if (p->is_objArray()) {
6844 6845 // objArrays are precisely marked; restrict scanning
6845 6846 // to dirty cards only.
6846 6847 size = CompactibleFreeListSpace::adjustObjectSize(
6847 6848 p->oop_iterate(_scanningClosure, mr));
6848 6849 } else {
6849 6850 // A non-array may have been imprecisely marked; we need
6850 6851 // to scan object in its entirety.
6851 6852 size = CompactibleFreeListSpace::adjustObjectSize(
6852 6853 p->oop_iterate(_scanningClosure));
6853 6854 }
6854 6855 #ifdef DEBUG
6855 6856 size_t direct_size =
6856 6857 CompactibleFreeListSpace::adjustObjectSize(p->size());
6857 6858 assert(size == direct_size, "Inconsistency in size");
6858 6859 assert(size >= 3, "Necessary for Printezis marks to work");
6859 6860 if (!_bitMap->isMarked(addr+1)) {
6860 6861 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6861 6862 } else {
6862 6863 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6863 6864 assert(_bitMap->isMarked(addr+size-1),
6864 6865 "inconsistent Printezis mark");
6865 6866 }
6866 6867 #endif // DEBUG
6867 6868 }
6868 6869 } else {
6869 6870 // an unitialized object
6870 6871 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6871 6872 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6872 6873 size = pointer_delta(nextOneAddr + 1, addr);
6873 6874 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6874 6875 "alignment problem");
6875 6876 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6876 6877 // will dirty the card when the klass pointer is installed in the
6877 6878 // object (signalling the completion of initialization).
6878 6879 }
6879 6880 } else {
6880 6881 // Either a not yet marked object or an uninitialized object
6881 6882 if (p->klass_or_null() == NULL || !p->is_parsable()) {
6882 6883 // An uninitialized object, skip to the next card, since
6883 6884 // we may not be able to read its P-bits yet.
6884 6885 assert(size == 0, "Initial value");
6885 6886 } else {
6886 6887 // An object not (yet) reached by marking: we merely need to
6887 6888 // compute its size so as to go look at the next block.
6888 6889 assert(p->is_oop(true), "should be an oop");
6889 6890 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6890 6891 }
6891 6892 }
6892 6893 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6893 6894 return size;
6894 6895 }
6895 6896
6896 6897 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6897 6898 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6898 6899 "CMS thread should hold CMS token");
6899 6900 assert_lock_strong(_freelistLock);
6900 6901 assert_lock_strong(_bitMap->lock());
6901 6902 DEBUG_ONLY(RememberKlassesChecker mux(false);)
6902 6903 // relinquish the free_list_lock and bitMaplock()
6903 6904 _bitMap->lock()->unlock();
6904 6905 _freelistLock->unlock();
6905 6906 ConcurrentMarkSweepThread::desynchronize(true);
6906 6907 ConcurrentMarkSweepThread::acknowledge_yield_request();
6907 6908 _collector->stopTimer();
6908 6909 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6909 6910 if (PrintCMSStatistics != 0) {
6910 6911 _collector->incrementYields();
6911 6912 }
6912 6913 _collector->icms_wait();
6913 6914
6914 6915 // See the comment in coordinator_yield()
6915 6916 for (unsigned i = 0; i < CMSYieldSleepCount &&
6916 6917 ConcurrentMarkSweepThread::should_yield() &&
6917 6918 !CMSCollector::foregroundGCIsActive(); ++i) {
6918 6919 os::sleep(Thread::current(), 1, false);
6919 6920 ConcurrentMarkSweepThread::acknowledge_yield_request();
6920 6921 }
6921 6922
6922 6923 ConcurrentMarkSweepThread::synchronize(true);
6923 6924 _freelistLock->lock_without_safepoint_check();
6924 6925 _bitMap->lock()->lock_without_safepoint_check();
6925 6926 _collector->startTimer();
6926 6927 }
6927 6928
6928 6929
6929 6930 //////////////////////////////////////////////////////////////////
6930 6931 // SurvivorSpacePrecleanClosure
6931 6932 //////////////////////////////////////////////////////////////////
6932 6933 // This (single-threaded) closure is used to preclean the oops in
6933 6934 // the survivor spaces.
6934 6935 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6935 6936
6936 6937 HeapWord* addr = (HeapWord*)p;
6937 6938 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6938 6939 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6939 6940 assert(p->klass_or_null() != NULL, "object should be initializd");
6940 6941 assert(p->is_parsable(), "must be parsable.");
6941 6942 // an initialized object; ignore mark word in verification below
6942 6943 // since we are running concurrent with mutators
6943 6944 assert(p->is_oop(true), "should be an oop");
6944 6945 // Note that we do not yield while we iterate over
6945 6946 // the interior oops of p, pushing the relevant ones
6946 6947 // on our marking stack.
6947 6948 size_t size = p->oop_iterate(_scanning_closure);
6948 6949 do_yield_check();
6949 6950 // Observe that below, we do not abandon the preclean
6950 6951 // phase as soon as we should; rather we empty the
6951 6952 // marking stack before returning. This is to satisfy
6952 6953 // some existing assertions. In general, it may be a
6953 6954 // good idea to abort immediately and complete the marking
6954 6955 // from the grey objects at a later time.
6955 6956 while (!_mark_stack->isEmpty()) {
6956 6957 oop new_oop = _mark_stack->pop();
6957 6958 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6958 6959 assert(new_oop->is_parsable(), "Found unparsable oop");
6959 6960 assert(_bit_map->isMarked((HeapWord*)new_oop),
6960 6961 "only grey objects on this stack");
6961 6962 // iterate over the oops in this oop, marking and pushing
6962 6963 // the ones in CMS heap (i.e. in _span).
6963 6964 new_oop->oop_iterate(_scanning_closure);
6964 6965 // check if it's time to yield
6965 6966 do_yield_check();
6966 6967 }
6967 6968 unsigned int after_count =
6968 6969 GenCollectedHeap::heap()->total_collections();
6969 6970 bool abort = (_before_count != after_count) ||
6970 6971 _collector->should_abort_preclean();
6971 6972 return abort ? 0 : size;
6972 6973 }
6973 6974
6974 6975 void SurvivorSpacePrecleanClosure::do_yield_work() {
6975 6976 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6976 6977 "CMS thread should hold CMS token");
6977 6978 assert_lock_strong(_bit_map->lock());
6978 6979 DEBUG_ONLY(RememberKlassesChecker smx(false);)
6979 6980 // Relinquish the bit map lock
6980 6981 _bit_map->lock()->unlock();
6981 6982 ConcurrentMarkSweepThread::desynchronize(true);
6982 6983 ConcurrentMarkSweepThread::acknowledge_yield_request();
6983 6984 _collector->stopTimer();
6984 6985 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6985 6986 if (PrintCMSStatistics != 0) {
6986 6987 _collector->incrementYields();
6987 6988 }
6988 6989 _collector->icms_wait();
6989 6990
6990 6991 // See the comment in coordinator_yield()
6991 6992 for (unsigned i = 0; i < CMSYieldSleepCount &&
6992 6993 ConcurrentMarkSweepThread::should_yield() &&
6993 6994 !CMSCollector::foregroundGCIsActive(); ++i) {
6994 6995 os::sleep(Thread::current(), 1, false);
6995 6996 ConcurrentMarkSweepThread::acknowledge_yield_request();
6996 6997 }
6997 6998
6998 6999 ConcurrentMarkSweepThread::synchronize(true);
6999 7000 _bit_map->lock()->lock_without_safepoint_check();
7000 7001 _collector->startTimer();
7001 7002 }
7002 7003
7003 7004 // This closure is used to rescan the marked objects on the dirty cards
7004 7005 // in the mod union table and the card table proper. In the parallel
7005 7006 // case, although the bitMap is shared, we do a single read so the
7006 7007 // isMarked() query is "safe".
7007 7008 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
7008 7009 // Ignore mark word because we are running concurrent with mutators
7009 7010 assert(p->is_oop_or_null(true), "expected an oop or null");
7010 7011 HeapWord* addr = (HeapWord*)p;
7011 7012 assert(_span.contains(addr), "we are scanning the CMS generation");
7012 7013 bool is_obj_array = false;
7013 7014 #ifdef DEBUG
7014 7015 if (!_parallel) {
7015 7016 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
7016 7017 assert(_collector->overflow_list_is_empty(),
7017 7018 "overflow list should be empty");
7018 7019
7019 7020 }
7020 7021 #endif // DEBUG
7021 7022 if (_bit_map->isMarked(addr)) {
7022 7023 // Obj arrays are precisely marked, non-arrays are not;
7023 7024 // so we scan objArrays precisely and non-arrays in their
7024 7025 // entirety.
7025 7026 if (p->is_objArray()) {
7026 7027 is_obj_array = true;
7027 7028 if (_parallel) {
7028 7029 p->oop_iterate(_par_scan_closure, mr);
7029 7030 } else {
7030 7031 p->oop_iterate(_scan_closure, mr);
7031 7032 }
7032 7033 } else {
7033 7034 if (_parallel) {
7034 7035 p->oop_iterate(_par_scan_closure);
7035 7036 } else {
7036 7037 p->oop_iterate(_scan_closure);
7037 7038 }
7038 7039 }
7039 7040 }
7040 7041 #ifdef DEBUG
7041 7042 if (!_parallel) {
7042 7043 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
7043 7044 assert(_collector->overflow_list_is_empty(),
7044 7045 "overflow list should be empty");
7045 7046
7046 7047 }
7047 7048 #endif // DEBUG
7048 7049 return is_obj_array;
7049 7050 }
7050 7051
7051 7052 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
7052 7053 MemRegion span,
7053 7054 CMSBitMap* bitMap, CMSMarkStack* markStack,
7054 7055 CMSMarkStack* revisitStack,
7055 7056 bool should_yield, bool verifying):
7056 7057 _collector(collector),
7057 7058 _span(span),
7058 7059 _bitMap(bitMap),
7059 7060 _mut(&collector->_modUnionTable),
7060 7061 _markStack(markStack),
7061 7062 _revisitStack(revisitStack),
7062 7063 _yield(should_yield),
7063 7064 _skipBits(0)
7064 7065 {
7065 7066 assert(_markStack->isEmpty(), "stack should be empty");
7066 7067 _finger = _bitMap->startWord();
7067 7068 _threshold = _finger;
7068 7069 assert(_collector->_restart_addr == NULL, "Sanity check");
7069 7070 assert(_span.contains(_finger), "Out of bounds _finger?");
7070 7071 DEBUG_ONLY(_verifying = verifying;)
7071 7072 }
7072 7073
7073 7074 void MarkFromRootsClosure::reset(HeapWord* addr) {
7074 7075 assert(_markStack->isEmpty(), "would cause duplicates on stack");
7075 7076 assert(_span.contains(addr), "Out of bounds _finger?");
7076 7077 _finger = addr;
7077 7078 _threshold = (HeapWord*)round_to(
7078 7079 (intptr_t)_finger, CardTableModRefBS::card_size);
7079 7080 }
7080 7081
7081 7082 // Should revisit to see if this should be restructured for
7082 7083 // greater efficiency.
7083 7084 bool MarkFromRootsClosure::do_bit(size_t offset) {
7084 7085 if (_skipBits > 0) {
7085 7086 _skipBits--;
7086 7087 return true;
7087 7088 }
7088 7089 // convert offset into a HeapWord*
7089 7090 HeapWord* addr = _bitMap->startWord() + offset;
7090 7091 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
7091 7092 "address out of range");
7092 7093 assert(_bitMap->isMarked(addr), "tautology");
7093 7094 if (_bitMap->isMarked(addr+1)) {
7094 7095 // this is an allocated but not yet initialized object
7095 7096 assert(_skipBits == 0, "tautology");
7096 7097 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
7097 7098 oop p = oop(addr);
7098 7099 if (p->klass_or_null() == NULL || !p->is_parsable()) {
7099 7100 DEBUG_ONLY(if (!_verifying) {)
7100 7101 // We re-dirty the cards on which this object lies and increase
7101 7102 // the _threshold so that we'll come back to scan this object
7102 7103 // during the preclean or remark phase. (CMSCleanOnEnter)
7103 7104 if (CMSCleanOnEnter) {
7104 7105 size_t sz = _collector->block_size_using_printezis_bits(addr);
7105 7106 HeapWord* end_card_addr = (HeapWord*)round_to(
7106 7107 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7107 7108 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7108 7109 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7109 7110 // Bump _threshold to end_card_addr; note that
7110 7111 // _threshold cannot possibly exceed end_card_addr, anyhow.
7111 7112 // This prevents future clearing of the card as the scan proceeds
7112 7113 // to the right.
7113 7114 assert(_threshold <= end_card_addr,
7114 7115 "Because we are just scanning into this object");
7115 7116 if (_threshold < end_card_addr) {
7116 7117 _threshold = end_card_addr;
7117 7118 }
7118 7119 if (p->klass_or_null() != NULL) {
7119 7120 // Redirty the range of cards...
7120 7121 _mut->mark_range(redirty_range);
7121 7122 } // ...else the setting of klass will dirty the card anyway.
7122 7123 }
7123 7124 DEBUG_ONLY(})
7124 7125 return true;
7125 7126 }
7126 7127 }
7127 7128 scanOopsInOop(addr);
7128 7129 return true;
7129 7130 }
7130 7131
7131 7132 // We take a break if we've been at this for a while,
7132 7133 // so as to avoid monopolizing the locks involved.
7133 7134 void MarkFromRootsClosure::do_yield_work() {
7134 7135 // First give up the locks, then yield, then re-lock
7135 7136 // We should probably use a constructor/destructor idiom to
7136 7137 // do this unlock/lock or modify the MutexUnlocker class to
7137 7138 // serve our purpose. XXX
7138 7139 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7139 7140 "CMS thread should hold CMS token");
7140 7141 assert_lock_strong(_bitMap->lock());
7141 7142 DEBUG_ONLY(RememberKlassesChecker mux(false);)
7142 7143 _bitMap->lock()->unlock();
7143 7144 ConcurrentMarkSweepThread::desynchronize(true);
7144 7145 ConcurrentMarkSweepThread::acknowledge_yield_request();
7145 7146 _collector->stopTimer();
7146 7147 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7147 7148 if (PrintCMSStatistics != 0) {
7148 7149 _collector->incrementYields();
7149 7150 }
7150 7151 _collector->icms_wait();
7151 7152
7152 7153 // See the comment in coordinator_yield()
7153 7154 for (unsigned i = 0; i < CMSYieldSleepCount &&
7154 7155 ConcurrentMarkSweepThread::should_yield() &&
7155 7156 !CMSCollector::foregroundGCIsActive(); ++i) {
7156 7157 os::sleep(Thread::current(), 1, false);
7157 7158 ConcurrentMarkSweepThread::acknowledge_yield_request();
7158 7159 }
7159 7160
7160 7161 ConcurrentMarkSweepThread::synchronize(true);
7161 7162 _bitMap->lock()->lock_without_safepoint_check();
7162 7163 _collector->startTimer();
7163 7164 }
7164 7165
7165 7166 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
7166 7167 assert(_bitMap->isMarked(ptr), "expected bit to be set");
7167 7168 assert(_markStack->isEmpty(),
7168 7169 "should drain stack to limit stack usage");
7169 7170 // convert ptr to an oop preparatory to scanning
7170 7171 oop obj = oop(ptr);
7171 7172 // Ignore mark word in verification below, since we
7172 7173 // may be running concurrent with mutators.
7173 7174 assert(obj->is_oop(true), "should be an oop");
7174 7175 assert(_finger <= ptr, "_finger runneth ahead");
7175 7176 // advance the finger to right end of this object
7176 7177 _finger = ptr + obj->size();
7177 7178 assert(_finger > ptr, "we just incremented it above");
7178 7179 // On large heaps, it may take us some time to get through
7179 7180 // the marking phase (especially if running iCMS). During
7180 7181 // this time it's possible that a lot of mutations have
7181 7182 // accumulated in the card table and the mod union table --
7182 7183 // these mutation records are redundant until we have
7183 7184 // actually traced into the corresponding card.
7184 7185 // Here, we check whether advancing the finger would make
7185 7186 // us cross into a new card, and if so clear corresponding
7186 7187 // cards in the MUT (preclean them in the card-table in the
7187 7188 // future).
7188 7189
7189 7190 DEBUG_ONLY(if (!_verifying) {)
7190 7191 // The clean-on-enter optimization is disabled by default,
7191 7192 // until we fix 6178663.
7192 7193 if (CMSCleanOnEnter && (_finger > _threshold)) {
7193 7194 // [_threshold, _finger) represents the interval
7194 7195 // of cards to be cleared in MUT (or precleaned in card table).
7195 7196 // The set of cards to be cleared is all those that overlap
7196 7197 // with the interval [_threshold, _finger); note that
7197 7198 // _threshold is always kept card-aligned but _finger isn't
7198 7199 // always card-aligned.
7199 7200 HeapWord* old_threshold = _threshold;
7200 7201 assert(old_threshold == (HeapWord*)round_to(
7201 7202 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7202 7203 "_threshold should always be card-aligned");
7203 7204 _threshold = (HeapWord*)round_to(
7204 7205 (intptr_t)_finger, CardTableModRefBS::card_size);
7205 7206 MemRegion mr(old_threshold, _threshold);
7206 7207 assert(!mr.is_empty(), "Control point invariant");
7207 7208 assert(_span.contains(mr), "Should clear within span");
7208 7209 // XXX When _finger crosses from old gen into perm gen
7209 7210 // we may be doing unnecessary cleaning; do better in the
7210 7211 // future by detecting that condition and clearing fewer
7211 7212 // MUT/CT entries.
7212 7213 _mut->clear_range(mr);
7213 7214 }
7214 7215 DEBUG_ONLY(})
7215 7216 // Note: the finger doesn't advance while we drain
7216 7217 // the stack below.
7217 7218 PushOrMarkClosure pushOrMarkClosure(_collector,
7218 7219 _span, _bitMap, _markStack,
7219 7220 _revisitStack,
7220 7221 _finger, this);
7221 7222 bool res = _markStack->push(obj);
7222 7223 assert(res, "Empty non-zero size stack should have space for single push");
7223 7224 while (!_markStack->isEmpty()) {
7224 7225 oop new_oop = _markStack->pop();
7225 7226 // Skip verifying header mark word below because we are
7226 7227 // running concurrent with mutators.
7227 7228 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7228 7229 // now scan this oop's oops
7229 7230 new_oop->oop_iterate(&pushOrMarkClosure);
7230 7231 do_yield_check();
7231 7232 }
7232 7233 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
7233 7234 }
7234 7235
7235 7236 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
7236 7237 CMSCollector* collector, MemRegion span,
7237 7238 CMSBitMap* bit_map,
7238 7239 OopTaskQueue* work_queue,
7239 7240 CMSMarkStack* overflow_stack,
7240 7241 CMSMarkStack* revisit_stack,
7241 7242 bool should_yield):
7242 7243 _collector(collector),
7243 7244 _whole_span(collector->_span),
7244 7245 _span(span),
7245 7246 _bit_map(bit_map),
7246 7247 _mut(&collector->_modUnionTable),
7247 7248 _work_queue(work_queue),
7248 7249 _overflow_stack(overflow_stack),
7249 7250 _revisit_stack(revisit_stack),
7250 7251 _yield(should_yield),
7251 7252 _skip_bits(0),
7252 7253 _task(task)
7253 7254 {
7254 7255 assert(_work_queue->size() == 0, "work_queue should be empty");
7255 7256 _finger = span.start();
7256 7257 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
7257 7258 assert(_span.contains(_finger), "Out of bounds _finger?");
7258 7259 }
7259 7260
7260 7261 // Should revisit to see if this should be restructured for
7261 7262 // greater efficiency.
7262 7263 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
7263 7264 if (_skip_bits > 0) {
7264 7265 _skip_bits--;
7265 7266 return true;
7266 7267 }
7267 7268 // convert offset into a HeapWord*
7268 7269 HeapWord* addr = _bit_map->startWord() + offset;
7269 7270 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
7270 7271 "address out of range");
7271 7272 assert(_bit_map->isMarked(addr), "tautology");
7272 7273 if (_bit_map->isMarked(addr+1)) {
7273 7274 // this is an allocated object that might not yet be initialized
7274 7275 assert(_skip_bits == 0, "tautology");
7275 7276 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
7276 7277 oop p = oop(addr);
7277 7278 if (p->klass_or_null() == NULL || !p->is_parsable()) {
7278 7279 // in the case of Clean-on-Enter optimization, redirty card
7279 7280 // and avoid clearing card by increasing the threshold.
7280 7281 return true;
7281 7282 }
7282 7283 }
7283 7284 scan_oops_in_oop(addr);
7284 7285 return true;
7285 7286 }
7286 7287
7287 7288 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
7288 7289 assert(_bit_map->isMarked(ptr), "expected bit to be set");
7289 7290 // Should we assert that our work queue is empty or
7290 7291 // below some drain limit?
7291 7292 assert(_work_queue->size() == 0,
7292 7293 "should drain stack to limit stack usage");
7293 7294 // convert ptr to an oop preparatory to scanning
7294 7295 oop obj = oop(ptr);
7295 7296 // Ignore mark word in verification below, since we
7296 7297 // may be running concurrent with mutators.
7297 7298 assert(obj->is_oop(true), "should be an oop");
7298 7299 assert(_finger <= ptr, "_finger runneth ahead");
7299 7300 // advance the finger to right end of this object
7300 7301 _finger = ptr + obj->size();
7301 7302 assert(_finger > ptr, "we just incremented it above");
7302 7303 // On large heaps, it may take us some time to get through
7303 7304 // the marking phase (especially if running iCMS). During
7304 7305 // this time it's possible that a lot of mutations have
7305 7306 // accumulated in the card table and the mod union table --
7306 7307 // these mutation records are redundant until we have
7307 7308 // actually traced into the corresponding card.
7308 7309 // Here, we check whether advancing the finger would make
7309 7310 // us cross into a new card, and if so clear corresponding
7310 7311 // cards in the MUT (preclean them in the card-table in the
7311 7312 // future).
7312 7313
7313 7314 // The clean-on-enter optimization is disabled by default,
7314 7315 // until we fix 6178663.
7315 7316 if (CMSCleanOnEnter && (_finger > _threshold)) {
7316 7317 // [_threshold, _finger) represents the interval
7317 7318 // of cards to be cleared in MUT (or precleaned in card table).
7318 7319 // The set of cards to be cleared is all those that overlap
7319 7320 // with the interval [_threshold, _finger); note that
7320 7321 // _threshold is always kept card-aligned but _finger isn't
7321 7322 // always card-aligned.
7322 7323 HeapWord* old_threshold = _threshold;
7323 7324 assert(old_threshold == (HeapWord*)round_to(
7324 7325 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7325 7326 "_threshold should always be card-aligned");
7326 7327 _threshold = (HeapWord*)round_to(
7327 7328 (intptr_t)_finger, CardTableModRefBS::card_size);
7328 7329 MemRegion mr(old_threshold, _threshold);
7329 7330 assert(!mr.is_empty(), "Control point invariant");
7330 7331 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7331 7332 // XXX When _finger crosses from old gen into perm gen
7332 7333 // we may be doing unnecessary cleaning; do better in the
7333 7334 // future by detecting that condition and clearing fewer
7334 7335 // MUT/CT entries.
7335 7336 _mut->clear_range(mr);
7336 7337 }
7337 7338
7338 7339 // Note: the local finger doesn't advance while we drain
7339 7340 // the stack below, but the global finger sure can and will.
7340 7341 HeapWord** gfa = _task->global_finger_addr();
7341 7342 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7342 7343 _span, _bit_map,
7343 7344 _work_queue,
7344 7345 _overflow_stack,
7345 7346 _revisit_stack,
7346 7347 _finger,
7347 7348 gfa, this);
7348 7349 bool res = _work_queue->push(obj); // overflow could occur here
7349 7350 assert(res, "Will hold once we use workqueues");
7350 7351 while (true) {
7351 7352 oop new_oop;
7352 7353 if (!_work_queue->pop_local(new_oop)) {
7353 7354 // We emptied our work_queue; check if there's stuff that can
7354 7355 // be gotten from the overflow stack.
7355 7356 if (CMSConcMarkingTask::get_work_from_overflow_stack(
7356 7357 _overflow_stack, _work_queue)) {
7357 7358 do_yield_check();
7358 7359 continue;
7359 7360 } else { // done
7360 7361 break;
7361 7362 }
7362 7363 }
7363 7364 // Skip verifying header mark word below because we are
7364 7365 // running concurrent with mutators.
7365 7366 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7366 7367 // now scan this oop's oops
7367 7368 new_oop->oop_iterate(&pushOrMarkClosure);
7368 7369 do_yield_check();
7369 7370 }
7370 7371 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7371 7372 }
7372 7373
7373 7374 // Yield in response to a request from VM Thread or
7374 7375 // from mutators.
7375 7376 void Par_MarkFromRootsClosure::do_yield_work() {
7376 7377 assert(_task != NULL, "sanity");
7377 7378 _task->yield();
7378 7379 }
7379 7380
7380 7381 // A variant of the above used for verifying CMS marking work.
7381 7382 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7382 7383 MemRegion span,
7383 7384 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7384 7385 CMSMarkStack* mark_stack):
7385 7386 _collector(collector),
7386 7387 _span(span),
7387 7388 _verification_bm(verification_bm),
7388 7389 _cms_bm(cms_bm),
7389 7390 _mark_stack(mark_stack),
7390 7391 _pam_verify_closure(collector, span, verification_bm, cms_bm,
7391 7392 mark_stack)
7392 7393 {
7393 7394 assert(_mark_stack->isEmpty(), "stack should be empty");
7394 7395 _finger = _verification_bm->startWord();
7395 7396 assert(_collector->_restart_addr == NULL, "Sanity check");
7396 7397 assert(_span.contains(_finger), "Out of bounds _finger?");
7397 7398 }
7398 7399
7399 7400 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7400 7401 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7401 7402 assert(_span.contains(addr), "Out of bounds _finger?");
7402 7403 _finger = addr;
7403 7404 }
7404 7405
7405 7406 // Should revisit to see if this should be restructured for
7406 7407 // greater efficiency.
7407 7408 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7408 7409 // convert offset into a HeapWord*
7409 7410 HeapWord* addr = _verification_bm->startWord() + offset;
7410 7411 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7411 7412 "address out of range");
7412 7413 assert(_verification_bm->isMarked(addr), "tautology");
7413 7414 assert(_cms_bm->isMarked(addr), "tautology");
7414 7415
7415 7416 assert(_mark_stack->isEmpty(),
7416 7417 "should drain stack to limit stack usage");
7417 7418 // convert addr to an oop preparatory to scanning
7418 7419 oop obj = oop(addr);
7419 7420 assert(obj->is_oop(), "should be an oop");
7420 7421 assert(_finger <= addr, "_finger runneth ahead");
7421 7422 // advance the finger to right end of this object
7422 7423 _finger = addr + obj->size();
7423 7424 assert(_finger > addr, "we just incremented it above");
7424 7425 // Note: the finger doesn't advance while we drain
7425 7426 // the stack below.
7426 7427 bool res = _mark_stack->push(obj);
7427 7428 assert(res, "Empty non-zero size stack should have space for single push");
7428 7429 while (!_mark_stack->isEmpty()) {
7429 7430 oop new_oop = _mark_stack->pop();
7430 7431 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7431 7432 // now scan this oop's oops
7432 7433 new_oop->oop_iterate(&_pam_verify_closure);
7433 7434 }
7434 7435 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7435 7436 return true;
7436 7437 }
7437 7438
7438 7439 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7439 7440 CMSCollector* collector, MemRegion span,
7440 7441 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7441 7442 CMSMarkStack* mark_stack):
7442 7443 OopClosure(collector->ref_processor()),
7443 7444 _collector(collector),
7444 7445 _span(span),
7445 7446 _verification_bm(verification_bm),
7446 7447 _cms_bm(cms_bm),
7447 7448 _mark_stack(mark_stack)
7448 7449 { }
7449 7450
7450 7451 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7451 7452 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7452 7453
7453 7454 // Upon stack overflow, we discard (part of) the stack,
7454 7455 // remembering the least address amongst those discarded
7455 7456 // in CMSCollector's _restart_address.
7456 7457 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7457 7458 // Remember the least grey address discarded
7458 7459 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7459 7460 _collector->lower_restart_addr(ra);
7460 7461 _mark_stack->reset(); // discard stack contents
7461 7462 _mark_stack->expand(); // expand the stack if possible
7462 7463 }
7463 7464
7464 7465 void PushAndMarkVerifyClosure::do_oop(oop obj) {
7465 7466 assert(obj->is_oop_or_null(), "expected an oop or NULL");
7466 7467 HeapWord* addr = (HeapWord*)obj;
7467 7468 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7468 7469 // Oop lies in _span and isn't yet grey or black
7469 7470 _verification_bm->mark(addr); // now grey
7470 7471 if (!_cms_bm->isMarked(addr)) {
7471 7472 oop(addr)->print();
7472 7473 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
7473 7474 addr);
7474 7475 fatal("... aborting");
7475 7476 }
7476 7477
7477 7478 if (!_mark_stack->push(obj)) { // stack overflow
7478 7479 if (PrintCMSStatistics != 0) {
7479 7480 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7480 7481 SIZE_FORMAT, _mark_stack->capacity());
7481 7482 }
7482 7483 assert(_mark_stack->isFull(), "Else push should have succeeded");
7483 7484 handle_stack_overflow(addr);
7484 7485 }
7485 7486 // anything including and to the right of _finger
7486 7487 // will be scanned as we iterate over the remainder of the
7487 7488 // bit map
7488 7489 }
7489 7490 }
7490 7491
7491 7492 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7492 7493 MemRegion span,
7493 7494 CMSBitMap* bitMap, CMSMarkStack* markStack,
7494 7495 CMSMarkStack* revisitStack,
7495 7496 HeapWord* finger, MarkFromRootsClosure* parent) :
7496 7497 KlassRememberingOopClosure(collector, collector->ref_processor(), revisitStack),
7497 7498 _span(span),
7498 7499 _bitMap(bitMap),
7499 7500 _markStack(markStack),
7500 7501 _finger(finger),
7501 7502 _parent(parent)
7502 7503 { }
7503 7504
7504 7505 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7505 7506 MemRegion span,
7506 7507 CMSBitMap* bit_map,
7507 7508 OopTaskQueue* work_queue,
7508 7509 CMSMarkStack* overflow_stack,
7509 7510 CMSMarkStack* revisit_stack,
7510 7511 HeapWord* finger,
7511 7512 HeapWord** global_finger_addr,
7512 7513 Par_MarkFromRootsClosure* parent) :
7513 7514 Par_KlassRememberingOopClosure(collector,
7514 7515 collector->ref_processor(),
7515 7516 revisit_stack),
7516 7517 _whole_span(collector->_span),
7517 7518 _span(span),
7518 7519 _bit_map(bit_map),
7519 7520 _work_queue(work_queue),
7520 7521 _overflow_stack(overflow_stack),
7521 7522 _finger(finger),
7522 7523 _global_finger_addr(global_finger_addr),
7523 7524 _parent(parent)
7524 7525 { }
7525 7526
7526 7527 // Assumes thread-safe access by callers, who are
7527 7528 // responsible for mutual exclusion.
7528 7529 void CMSCollector::lower_restart_addr(HeapWord* low) {
7529 7530 assert(_span.contains(low), "Out of bounds addr");
7530 7531 if (_restart_addr == NULL) {
7531 7532 _restart_addr = low;
7532 7533 } else {
7533 7534 _restart_addr = MIN2(_restart_addr, low);
7534 7535 }
7535 7536 }
7536 7537
7537 7538 // Upon stack overflow, we discard (part of) the stack,
7538 7539 // remembering the least address amongst those discarded
7539 7540 // in CMSCollector's _restart_address.
7540 7541 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7541 7542 // Remember the least grey address discarded
7542 7543 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7543 7544 _collector->lower_restart_addr(ra);
7544 7545 _markStack->reset(); // discard stack contents
7545 7546 _markStack->expand(); // expand the stack if possible
7546 7547 }
7547 7548
7548 7549 // Upon stack overflow, we discard (part of) the stack,
7549 7550 // remembering the least address amongst those discarded
7550 7551 // in CMSCollector's _restart_address.
7551 7552 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7552 7553 // We need to do this under a mutex to prevent other
7553 7554 // workers from interfering with the work done below.
7554 7555 MutexLockerEx ml(_overflow_stack->par_lock(),
7555 7556 Mutex::_no_safepoint_check_flag);
7556 7557 // Remember the least grey address discarded
7557 7558 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7558 7559 _collector->lower_restart_addr(ra);
7559 7560 _overflow_stack->reset(); // discard stack contents
7560 7561 _overflow_stack->expand(); // expand the stack if possible
7561 7562 }
7562 7563
7563 7564 void PushOrMarkClosure::do_oop(oop obj) {
7564 7565 // Ignore mark word because we are running concurrent with mutators.
7565 7566 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7566 7567 HeapWord* addr = (HeapWord*)obj;
7567 7568 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7568 7569 // Oop lies in _span and isn't yet grey or black
7569 7570 _bitMap->mark(addr); // now grey
7570 7571 if (addr < _finger) {
7571 7572 // the bit map iteration has already either passed, or
7572 7573 // sampled, this bit in the bit map; we'll need to
7573 7574 // use the marking stack to scan this oop's oops.
7574 7575 bool simulate_overflow = false;
7575 7576 NOT_PRODUCT(
7576 7577 if (CMSMarkStackOverflowALot &&
7577 7578 _collector->simulate_overflow()) {
7578 7579 // simulate a stack overflow
7579 7580 simulate_overflow = true;
7580 7581 }
7581 7582 )
7582 7583 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7583 7584 if (PrintCMSStatistics != 0) {
7584 7585 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7585 7586 SIZE_FORMAT, _markStack->capacity());
7586 7587 }
7587 7588 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7588 7589 handle_stack_overflow(addr);
7589 7590 }
7590 7591 }
7591 7592 // anything including and to the right of _finger
7592 7593 // will be scanned as we iterate over the remainder of the
7593 7594 // bit map
7594 7595 do_yield_check();
7595 7596 }
7596 7597 }
7597 7598
7598 7599 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7599 7600 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7600 7601
7601 7602 void Par_PushOrMarkClosure::do_oop(oop obj) {
7602 7603 // Ignore mark word because we are running concurrent with mutators.
7603 7604 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7604 7605 HeapWord* addr = (HeapWord*)obj;
7605 7606 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7606 7607 // Oop lies in _span and isn't yet grey or black
7607 7608 // We read the global_finger (volatile read) strictly after marking oop
7608 7609 bool res = _bit_map->par_mark(addr); // now grey
7609 7610 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7610 7611 // Should we push this marked oop on our stack?
7611 7612 // -- if someone else marked it, nothing to do
7612 7613 // -- if target oop is above global finger nothing to do
7613 7614 // -- if target oop is in chunk and above local finger
7614 7615 // then nothing to do
7615 7616 // -- else push on work queue
7616 7617 if ( !res // someone else marked it, they will deal with it
7617 7618 || (addr >= *gfa) // will be scanned in a later task
7618 7619 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7619 7620 return;
7620 7621 }
7621 7622 // the bit map iteration has already either passed, or
7622 7623 // sampled, this bit in the bit map; we'll need to
7623 7624 // use the marking stack to scan this oop's oops.
7624 7625 bool simulate_overflow = false;
7625 7626 NOT_PRODUCT(
7626 7627 if (CMSMarkStackOverflowALot &&
7627 7628 _collector->simulate_overflow()) {
7628 7629 // simulate a stack overflow
7629 7630 simulate_overflow = true;
7630 7631 }
7631 7632 )
7632 7633 if (simulate_overflow ||
7633 7634 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7634 7635 // stack overflow
7635 7636 if (PrintCMSStatistics != 0) {
7636 7637 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7637 7638 SIZE_FORMAT, _overflow_stack->capacity());
7638 7639 }
7639 7640 // We cannot assert that the overflow stack is full because
7640 7641 // it may have been emptied since.
7641 7642 assert(simulate_overflow ||
7642 7643 _work_queue->size() == _work_queue->max_elems(),
7643 7644 "Else push should have succeeded");
7644 7645 handle_stack_overflow(addr);
7645 7646 }
7646 7647 do_yield_check();
7647 7648 }
7648 7649 }
7649 7650
7650 7651 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7651 7652 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7652 7653
7653 7654 KlassRememberingOopClosure::KlassRememberingOopClosure(CMSCollector* collector,
7654 7655 ReferenceProcessor* rp,
7655 7656 CMSMarkStack* revisit_stack) :
7656 7657 OopClosure(rp),
7657 7658 _collector(collector),
7658 7659 _revisit_stack(revisit_stack),
7659 7660 _should_remember_klasses(collector->should_unload_classes()) {}
7660 7661
7661 7662 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7662 7663 MemRegion span,
7663 7664 ReferenceProcessor* rp,
7664 7665 CMSBitMap* bit_map,
7665 7666 CMSBitMap* mod_union_table,
7666 7667 CMSMarkStack* mark_stack,
7667 7668 CMSMarkStack* revisit_stack,
7668 7669 bool concurrent_precleaning):
7669 7670 KlassRememberingOopClosure(collector, rp, revisit_stack),
7670 7671 _span(span),
7671 7672 _bit_map(bit_map),
7672 7673 _mod_union_table(mod_union_table),
7673 7674 _mark_stack(mark_stack),
7674 7675 _concurrent_precleaning(concurrent_precleaning)
7675 7676 {
7676 7677 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7677 7678 }
7678 7679
7679 7680 // Grey object rescan during pre-cleaning and second checkpoint phases --
7680 7681 // the non-parallel version (the parallel version appears further below.)
7681 7682 void PushAndMarkClosure::do_oop(oop obj) {
7682 7683 // Ignore mark word verification. If during concurrent precleaning,
7683 7684 // the object monitor may be locked. If during the checkpoint
7684 7685 // phases, the object may already have been reached by a different
7685 7686 // path and may be at the end of the global overflow list (so
7686 7687 // the mark word may be NULL).
7687 7688 assert(obj->is_oop_or_null(true /* ignore mark word */),
7688 7689 "expected an oop or NULL");
7689 7690 HeapWord* addr = (HeapWord*)obj;
7690 7691 // Check if oop points into the CMS generation
7691 7692 // and is not marked
7692 7693 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7693 7694 // a white object ...
7694 7695 _bit_map->mark(addr); // ... now grey
7695 7696 // push on the marking stack (grey set)
7696 7697 bool simulate_overflow = false;
7697 7698 NOT_PRODUCT(
7698 7699 if (CMSMarkStackOverflowALot &&
7699 7700 _collector->simulate_overflow()) {
7700 7701 // simulate a stack overflow
7701 7702 simulate_overflow = true;
7702 7703 }
7703 7704 )
7704 7705 if (simulate_overflow || !_mark_stack->push(obj)) {
7705 7706 if (_concurrent_precleaning) {
7706 7707 // During precleaning we can just dirty the appropriate card(s)
7707 7708 // in the mod union table, thus ensuring that the object remains
7708 7709 // in the grey set and continue. In the case of object arrays
7709 7710 // we need to dirty all of the cards that the object spans,
7710 7711 // since the rescan of object arrays will be limited to the
7711 7712 // dirty cards.
7712 7713 // Note that no one can be intefering with us in this action
7713 7714 // of dirtying the mod union table, so no locking or atomics
7714 7715 // are required.
7715 7716 if (obj->is_objArray()) {
7716 7717 size_t sz = obj->size();
7717 7718 HeapWord* end_card_addr = (HeapWord*)round_to(
7718 7719 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7719 7720 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7720 7721 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7721 7722 _mod_union_table->mark_range(redirty_range);
7722 7723 } else {
7723 7724 _mod_union_table->mark(addr);
7724 7725 }
7725 7726 _collector->_ser_pmc_preclean_ovflw++;
7726 7727 } else {
7727 7728 // During the remark phase, we need to remember this oop
7728 7729 // in the overflow list.
7729 7730 _collector->push_on_overflow_list(obj);
7730 7731 _collector->_ser_pmc_remark_ovflw++;
7731 7732 }
7732 7733 }
7733 7734 }
7734 7735 }
7735 7736
7736 7737 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7737 7738 MemRegion span,
7738 7739 ReferenceProcessor* rp,
7739 7740 CMSBitMap* bit_map,
7740 7741 OopTaskQueue* work_queue,
7741 7742 CMSMarkStack* revisit_stack):
7742 7743 Par_KlassRememberingOopClosure(collector, rp, revisit_stack),
7743 7744 _span(span),
7744 7745 _bit_map(bit_map),
7745 7746 _work_queue(work_queue)
7746 7747 {
7747 7748 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7748 7749 }
7749 7750
7750 7751 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7751 7752 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7752 7753
7753 7754 // Grey object rescan during second checkpoint phase --
7754 7755 // the parallel version.
7755 7756 void Par_PushAndMarkClosure::do_oop(oop obj) {
7756 7757 // In the assert below, we ignore the mark word because
7757 7758 // this oop may point to an already visited object that is
7758 7759 // on the overflow stack (in which case the mark word has
7759 7760 // been hijacked for chaining into the overflow stack --
7760 7761 // if this is the last object in the overflow stack then
7761 7762 // its mark word will be NULL). Because this object may
7762 7763 // have been subsequently popped off the global overflow
7763 7764 // stack, and the mark word possibly restored to the prototypical
7764 7765 // value, by the time we get to examined this failing assert in
7765 7766 // the debugger, is_oop_or_null(false) may subsequently start
7766 7767 // to hold.
7767 7768 assert(obj->is_oop_or_null(true),
7768 7769 "expected an oop or NULL");
7769 7770 HeapWord* addr = (HeapWord*)obj;
7770 7771 // Check if oop points into the CMS generation
7771 7772 // and is not marked
7772 7773 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7773 7774 // a white object ...
7774 7775 // If we manage to "claim" the object, by being the
7775 7776 // first thread to mark it, then we push it on our
7776 7777 // marking stack
7777 7778 if (_bit_map->par_mark(addr)) { // ... now grey
7778 7779 // push on work queue (grey set)
7779 7780 bool simulate_overflow = false;
7780 7781 NOT_PRODUCT(
7781 7782 if (CMSMarkStackOverflowALot &&
7782 7783 _collector->par_simulate_overflow()) {
7783 7784 // simulate a stack overflow
7784 7785 simulate_overflow = true;
7785 7786 }
7786 7787 )
7787 7788 if (simulate_overflow || !_work_queue->push(obj)) {
7788 7789 _collector->par_push_on_overflow_list(obj);
7789 7790 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7790 7791 }
7791 7792 } // Else, some other thread got there first
7792 7793 }
7793 7794 }
7794 7795
7795 7796 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7796 7797 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7797 7798
7798 7799 void PushAndMarkClosure::remember_mdo(DataLayout* v) {
7799 7800 // TBD
7800 7801 }
7801 7802
7802 7803 void Par_PushAndMarkClosure::remember_mdo(DataLayout* v) {
7803 7804 // TBD
7804 7805 }
7805 7806
7806 7807 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7807 7808 DEBUG_ONLY(RememberKlassesChecker mux(false);)
7808 7809 Mutex* bml = _collector->bitMapLock();
7809 7810 assert_lock_strong(bml);
7810 7811 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7811 7812 "CMS thread should hold CMS token");
7812 7813
7813 7814 bml->unlock();
7814 7815 ConcurrentMarkSweepThread::desynchronize(true);
7815 7816
7816 7817 ConcurrentMarkSweepThread::acknowledge_yield_request();
7817 7818
7818 7819 _collector->stopTimer();
7819 7820 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7820 7821 if (PrintCMSStatistics != 0) {
7821 7822 _collector->incrementYields();
7822 7823 }
7823 7824 _collector->icms_wait();
7824 7825
7825 7826 // See the comment in coordinator_yield()
7826 7827 for (unsigned i = 0; i < CMSYieldSleepCount &&
7827 7828 ConcurrentMarkSweepThread::should_yield() &&
7828 7829 !CMSCollector::foregroundGCIsActive(); ++i) {
7829 7830 os::sleep(Thread::current(), 1, false);
7830 7831 ConcurrentMarkSweepThread::acknowledge_yield_request();
7831 7832 }
7832 7833
7833 7834 ConcurrentMarkSweepThread::synchronize(true);
7834 7835 bml->lock();
7835 7836
7836 7837 _collector->startTimer();
7837 7838 }
7838 7839
7839 7840 bool CMSPrecleanRefsYieldClosure::should_return() {
7840 7841 if (ConcurrentMarkSweepThread::should_yield()) {
7841 7842 do_yield_work();
7842 7843 }
7843 7844 return _collector->foregroundGCIsActive();
7844 7845 }
7845 7846
7846 7847 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7847 7848 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7848 7849 "mr should be aligned to start at a card boundary");
7849 7850 // We'd like to assert:
7850 7851 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7851 7852 // "mr should be a range of cards");
7852 7853 // However, that would be too strong in one case -- the last
7853 7854 // partition ends at _unallocated_block which, in general, can be
7854 7855 // an arbitrary boundary, not necessarily card aligned.
7855 7856 if (PrintCMSStatistics != 0) {
7856 7857 _num_dirty_cards +=
7857 7858 mr.word_size()/CardTableModRefBS::card_size_in_words;
7858 7859 }
7859 7860 _space->object_iterate_mem(mr, &_scan_cl);
7860 7861 }
7861 7862
7862 7863 SweepClosure::SweepClosure(CMSCollector* collector,
7863 7864 ConcurrentMarkSweepGeneration* g,
7864 7865 CMSBitMap* bitMap, bool should_yield) :
7865 7866 _collector(collector),
7866 7867 _g(g),
7867 7868 _sp(g->cmsSpace()),
7868 7869 _limit(_sp->sweep_limit()),
7869 7870 _freelistLock(_sp->freelistLock()),
7870 7871 _bitMap(bitMap),
7871 7872 _yield(should_yield),
7872 7873 _inFreeRange(false), // No free range at beginning of sweep
7873 7874 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7874 7875 _lastFreeRangeCoalesced(false),
7875 7876 _freeFinger(g->used_region().start())
7876 7877 {
7877 7878 NOT_PRODUCT(
7878 7879 _numObjectsFreed = 0;
7879 7880 _numWordsFreed = 0;
7880 7881 _numObjectsLive = 0;
7881 7882 _numWordsLive = 0;
7882 7883 _numObjectsAlreadyFree = 0;
7883 7884 _numWordsAlreadyFree = 0;
7884 7885 _last_fc = NULL;
7885 7886
7886 7887 _sp->initializeIndexedFreeListArrayReturnedBytes();
7887 7888 _sp->dictionary()->initializeDictReturnedBytes();
7888 7889 )
7889 7890 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7890 7891 "sweep _limit out of bounds");
7891 7892 if (CMSTraceSweeper) {
7892 7893 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
7893 7894 _limit);
7894 7895 }
7895 7896 }
7896 7897
7897 7898 void SweepClosure::print_on(outputStream* st) const {
7898 7899 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7899 7900 _sp->bottom(), _sp->end());
7900 7901 tty->print_cr("_limit = " PTR_FORMAT, _limit);
7901 7902 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger);
7902 7903 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);)
7903 7904 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7904 7905 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7905 7906 }
7906 7907
7907 7908 #ifndef PRODUCT
7908 7909 // Assertion checking only: no useful work in product mode --
7909 7910 // however, if any of the flags below become product flags,
7910 7911 // you may need to review this code to see if it needs to be
7911 7912 // enabled in product mode.
7912 7913 SweepClosure::~SweepClosure() {
7913 7914 assert_lock_strong(_freelistLock);
7914 7915 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7915 7916 "sweep _limit out of bounds");
7916 7917 if (inFreeRange()) {
7917 7918 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
7918 7919 print();
7919 7920 ShouldNotReachHere();
7920 7921 }
7921 7922 if (Verbose && PrintGC) {
7922 7923 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes",
7923 7924 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7924 7925 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
7925 7926 SIZE_FORMAT" bytes "
7926 7927 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
7927 7928 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7928 7929 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7929 7930 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
7930 7931 * sizeof(HeapWord);
7931 7932 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
7932 7933
7933 7934 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7934 7935 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7935 7936 size_t dictReturnedBytes = _sp->dictionary()->sumDictReturnedBytes();
7936 7937 size_t returnedBytes = indexListReturnedBytes + dictReturnedBytes;
7937 7938 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returnedBytes);
7938 7939 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
7939 7940 indexListReturnedBytes);
7940 7941 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
7941 7942 dictReturnedBytes);
7942 7943 }
7943 7944 }
7944 7945 if (CMSTraceSweeper) {
7945 7946 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
7946 7947 _limit);
7947 7948 }
7948 7949 }
7949 7950 #endif // PRODUCT
7950 7951
7951 7952 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7952 7953 bool freeRangeInFreeLists) {
7953 7954 if (CMSTraceSweeper) {
7954 7955 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n",
7955 7956 freeFinger, freeRangeInFreeLists);
7956 7957 }
7957 7958 assert(!inFreeRange(), "Trampling existing free range");
7958 7959 set_inFreeRange(true);
7959 7960 set_lastFreeRangeCoalesced(false);
7960 7961
7961 7962 set_freeFinger(freeFinger);
7962 7963 set_freeRangeInFreeLists(freeRangeInFreeLists);
7963 7964 if (CMSTestInFreeList) {
7964 7965 if (freeRangeInFreeLists) {
7965 7966 FreeChunk* fc = (FreeChunk*) freeFinger;
7966 7967 assert(fc->isFree(), "A chunk on the free list should be free.");
7967 7968 assert(fc->size() > 0, "Free range should have a size");
7968 7969 assert(_sp->verifyChunkInFreeLists(fc), "Chunk is not in free lists");
7969 7970 }
7970 7971 }
7971 7972 }
7972 7973
7973 7974 // Note that the sweeper runs concurrently with mutators. Thus,
7974 7975 // it is possible for direct allocation in this generation to happen
7975 7976 // in the middle of the sweep. Note that the sweeper also coalesces
7976 7977 // contiguous free blocks. Thus, unless the sweeper and the allocator
7977 7978 // synchronize appropriately freshly allocated blocks may get swept up.
7978 7979 // This is accomplished by the sweeper locking the free lists while
7979 7980 // it is sweeping. Thus blocks that are determined to be free are
7980 7981 // indeed free. There is however one additional complication:
7981 7982 // blocks that have been allocated since the final checkpoint and
7982 7983 // mark, will not have been marked and so would be treated as
7983 7984 // unreachable and swept up. To prevent this, the allocator marks
7984 7985 // the bit map when allocating during the sweep phase. This leads,
7985 7986 // however, to a further complication -- objects may have been allocated
7986 7987 // but not yet initialized -- in the sense that the header isn't yet
7987 7988 // installed. The sweeper can not then determine the size of the block
7988 7989 // in order to skip over it. To deal with this case, we use a technique
7989 7990 // (due to Printezis) to encode such uninitialized block sizes in the
7990 7991 // bit map. Since the bit map uses a bit per every HeapWord, but the
7991 7992 // CMS generation has a minimum object size of 3 HeapWords, it follows
7992 7993 // that "normal marks" won't be adjacent in the bit map (there will
7993 7994 // always be at least two 0 bits between successive 1 bits). We make use
7994 7995 // of these "unused" bits to represent uninitialized blocks -- the bit
7995 7996 // corresponding to the start of the uninitialized object and the next
7996 7997 // bit are both set. Finally, a 1 bit marks the end of the object that
7997 7998 // started with the two consecutive 1 bits to indicate its potentially
7998 7999 // uninitialized state.
7999 8000
8000 8001 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
8001 8002 FreeChunk* fc = (FreeChunk*)addr;
8002 8003 size_t res;
8003 8004
8004 8005 // Check if we are done sweeping. Below we check "addr >= _limit" rather
8005 8006 // than "addr == _limit" because although _limit was a block boundary when
8006 8007 // we started the sweep, it may no longer be one because heap expansion
8007 8008 // may have caused us to coalesce the block ending at the address _limit
8008 8009 // with a newly expanded chunk (this happens when _limit was set to the
8009 8010 // previous _end of the space), so we may have stepped past _limit:
8010 8011 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
8011 8012 if (addr >= _limit) { // we have swept up to or past the limit: finish up
8012 8013 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8013 8014 "sweep _limit out of bounds");
8014 8015 assert(addr < _sp->end(), "addr out of bounds");
8015 8016 // Flush any free range we might be holding as a single
8016 8017 // coalesced chunk to the appropriate free list.
8017 8018 if (inFreeRange()) {
8018 8019 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
8019 8020 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger()));
8020 8021 flush_cur_free_chunk(freeFinger(),
8021 8022 pointer_delta(addr, freeFinger()));
8022 8023 if (CMSTraceSweeper) {
8023 8024 gclog_or_tty->print("Sweep: last chunk: ");
8024 8025 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") "
8025 8026 "[coalesced:"SIZE_FORMAT"]\n",
8026 8027 freeFinger(), pointer_delta(addr, freeFinger()),
8027 8028 lastFreeRangeCoalesced());
8028 8029 }
8029 8030 }
8030 8031
8031 8032 // help the iterator loop finish
8032 8033 return pointer_delta(_sp->end(), addr);
8033 8034 }
8034 8035
8035 8036 assert(addr < _limit, "sweep invariant");
8036 8037 // check if we should yield
8037 8038 do_yield_check(addr);
8038 8039 if (fc->isFree()) {
8039 8040 // Chunk that is already free
8040 8041 res = fc->size();
8041 8042 do_already_free_chunk(fc);
8042 8043 debug_only(_sp->verifyFreeLists());
8043 8044 // If we flush the chunk at hand in lookahead_and_flush()
8044 8045 // and it's coalesced with a preceding chunk, then the
8045 8046 // process of "mangling" the payload of the coalesced block
8046 8047 // will cause erasure of the size information from the
8047 8048 // (erstwhile) header of all the coalesced blocks but the
8048 8049 // first, so the first disjunct in the assert will not hold
8049 8050 // in that specific case (in which case the second disjunct
8050 8051 // will hold).
8051 8052 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
8052 8053 "Otherwise the size info doesn't change at this step");
8053 8054 NOT_PRODUCT(
8054 8055 _numObjectsAlreadyFree++;
8055 8056 _numWordsAlreadyFree += res;
8056 8057 )
8057 8058 NOT_PRODUCT(_last_fc = fc;)
8058 8059 } else if (!_bitMap->isMarked(addr)) {
8059 8060 // Chunk is fresh garbage
8060 8061 res = do_garbage_chunk(fc);
8061 8062 debug_only(_sp->verifyFreeLists());
8062 8063 NOT_PRODUCT(
8063 8064 _numObjectsFreed++;
8064 8065 _numWordsFreed += res;
8065 8066 )
8066 8067 } else {
8067 8068 // Chunk that is alive.
8068 8069 res = do_live_chunk(fc);
8069 8070 debug_only(_sp->verifyFreeLists());
8070 8071 NOT_PRODUCT(
8071 8072 _numObjectsLive++;
8072 8073 _numWordsLive += res;
8073 8074 )
8074 8075 }
8075 8076 return res;
8076 8077 }
8077 8078
8078 8079 // For the smart allocation, record following
8079 8080 // split deaths - a free chunk is removed from its free list because
8080 8081 // it is being split into two or more chunks.
8081 8082 // split birth - a free chunk is being added to its free list because
8082 8083 // a larger free chunk has been split and resulted in this free chunk.
8083 8084 // coal death - a free chunk is being removed from its free list because
8084 8085 // it is being coalesced into a large free chunk.
8085 8086 // coal birth - a free chunk is being added to its free list because
8086 8087 // it was created when two or more free chunks where coalesced into
8087 8088 // this free chunk.
8088 8089 //
8089 8090 // These statistics are used to determine the desired number of free
8090 8091 // chunks of a given size. The desired number is chosen to be relative
8091 8092 // to the end of a CMS sweep. The desired number at the end of a sweep
8092 8093 // is the
8093 8094 // count-at-end-of-previous-sweep (an amount that was enough)
8094 8095 // - count-at-beginning-of-current-sweep (the excess)
8095 8096 // + split-births (gains in this size during interval)
8096 8097 // - split-deaths (demands on this size during interval)
8097 8098 // where the interval is from the end of one sweep to the end of the
8098 8099 // next.
8099 8100 //
8100 8101 // When sweeping the sweeper maintains an accumulated chunk which is
8101 8102 // the chunk that is made up of chunks that have been coalesced. That
8102 8103 // will be termed the left-hand chunk. A new chunk of garbage that
8103 8104 // is being considered for coalescing will be referred to as the
8104 8105 // right-hand chunk.
8105 8106 //
8106 8107 // When making a decision on whether to coalesce a right-hand chunk with
8107 8108 // the current left-hand chunk, the current count vs. the desired count
8108 8109 // of the left-hand chunk is considered. Also if the right-hand chunk
8109 8110 // is near the large chunk at the end of the heap (see
8110 8111 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
8111 8112 // left-hand chunk is coalesced.
8112 8113 //
8113 8114 // When making a decision about whether to split a chunk, the desired count
8114 8115 // vs. the current count of the candidate to be split is also considered.
8115 8116 // If the candidate is underpopulated (currently fewer chunks than desired)
8116 8117 // a chunk of an overpopulated (currently more chunks than desired) size may
8117 8118 // be chosen. The "hint" associated with a free list, if non-null, points
8118 8119 // to a free list which may be overpopulated.
8119 8120 //
8120 8121
8121 8122 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
8122 8123 const size_t size = fc->size();
8123 8124 // Chunks that cannot be coalesced are not in the
8124 8125 // free lists.
8125 8126 if (CMSTestInFreeList && !fc->cantCoalesce()) {
8126 8127 assert(_sp->verifyChunkInFreeLists(fc),
8127 8128 "free chunk should be in free lists");
8128 8129 }
8129 8130 // a chunk that is already free, should not have been
8130 8131 // marked in the bit map
8131 8132 HeapWord* const addr = (HeapWord*) fc;
8132 8133 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
8133 8134 // Verify that the bit map has no bits marked between
8134 8135 // addr and purported end of this block.
8135 8136 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8136 8137
8137 8138 // Some chunks cannot be coalesced under any circumstances.
8138 8139 // See the definition of cantCoalesce().
8139 8140 if (!fc->cantCoalesce()) {
8140 8141 // This chunk can potentially be coalesced.
8141 8142 if (_sp->adaptive_freelists()) {
8142 8143 // All the work is done in
8143 8144 do_post_free_or_garbage_chunk(fc, size);
8144 8145 } else { // Not adaptive free lists
8145 8146 // this is a free chunk that can potentially be coalesced by the sweeper;
8146 8147 if (!inFreeRange()) {
8147 8148 // if the next chunk is a free block that can't be coalesced
8148 8149 // it doesn't make sense to remove this chunk from the free lists
8149 8150 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
8150 8151 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
8151 8152 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
8152 8153 nextChunk->isFree() && // ... which is free...
8153 8154 nextChunk->cantCoalesce()) { // ... but can't be coalesced
8154 8155 // nothing to do
8155 8156 } else {
8156 8157 // Potentially the start of a new free range:
8157 8158 // Don't eagerly remove it from the free lists.
8158 8159 // No need to remove it if it will just be put
8159 8160 // back again. (Also from a pragmatic point of view
8160 8161 // if it is a free block in a region that is beyond
8161 8162 // any allocated blocks, an assertion will fail)
8162 8163 // Remember the start of a free run.
8163 8164 initialize_free_range(addr, true);
8164 8165 // end - can coalesce with next chunk
8165 8166 }
8166 8167 } else {
8167 8168 // the midst of a free range, we are coalescing
8168 8169 print_free_block_coalesced(fc);
8169 8170 if (CMSTraceSweeper) {
8170 8171 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size);
8171 8172 }
8172 8173 // remove it from the free lists
8173 8174 _sp->removeFreeChunkFromFreeLists(fc);
8174 8175 set_lastFreeRangeCoalesced(true);
8175 8176 // If the chunk is being coalesced and the current free range is
8176 8177 // in the free lists, remove the current free range so that it
8177 8178 // will be returned to the free lists in its entirety - all
8178 8179 // the coalesced pieces included.
8179 8180 if (freeRangeInFreeLists()) {
8180 8181 FreeChunk* ffc = (FreeChunk*) freeFinger();
8181 8182 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8182 8183 "Size of free range is inconsistent with chunk size.");
8183 8184 if (CMSTestInFreeList) {
8184 8185 assert(_sp->verifyChunkInFreeLists(ffc),
8185 8186 "free range is not in free lists");
8186 8187 }
8187 8188 _sp->removeFreeChunkFromFreeLists(ffc);
8188 8189 set_freeRangeInFreeLists(false);
8189 8190 }
8190 8191 }
8191 8192 }
8192 8193 // Note that if the chunk is not coalescable (the else arm
8193 8194 // below), we unconditionally flush, without needing to do
8194 8195 // a "lookahead," as we do below.
8195 8196 if (inFreeRange()) lookahead_and_flush(fc, size);
8196 8197 } else {
8197 8198 // Code path common to both original and adaptive free lists.
8198 8199
8199 8200 // cant coalesce with previous block; this should be treated
8200 8201 // as the end of a free run if any
8201 8202 if (inFreeRange()) {
8202 8203 // we kicked some butt; time to pick up the garbage
8203 8204 assert(freeFinger() < addr, "freeFinger points too high");
8204 8205 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8205 8206 }
8206 8207 // else, nothing to do, just continue
8207 8208 }
8208 8209 }
8209 8210
8210 8211 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
8211 8212 // This is a chunk of garbage. It is not in any free list.
8212 8213 // Add it to a free list or let it possibly be coalesced into
8213 8214 // a larger chunk.
8214 8215 HeapWord* const addr = (HeapWord*) fc;
8215 8216 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8216 8217
8217 8218 if (_sp->adaptive_freelists()) {
8218 8219 // Verify that the bit map has no bits marked between
8219 8220 // addr and purported end of just dead object.
8220 8221 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8221 8222
8222 8223 do_post_free_or_garbage_chunk(fc, size);
8223 8224 } else {
8224 8225 if (!inFreeRange()) {
8225 8226 // start of a new free range
8226 8227 assert(size > 0, "A free range should have a size");
8227 8228 initialize_free_range(addr, false);
8228 8229 } else {
8229 8230 // this will be swept up when we hit the end of the
8230 8231 // free range
8231 8232 if (CMSTraceSweeper) {
8232 8233 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size);
8233 8234 }
8234 8235 // If the chunk is being coalesced and the current free range is
8235 8236 // in the free lists, remove the current free range so that it
8236 8237 // will be returned to the free lists in its entirety - all
8237 8238 // the coalesced pieces included.
8238 8239 if (freeRangeInFreeLists()) {
8239 8240 FreeChunk* ffc = (FreeChunk*)freeFinger();
8240 8241 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8241 8242 "Size of free range is inconsistent with chunk size.");
8242 8243 if (CMSTestInFreeList) {
8243 8244 assert(_sp->verifyChunkInFreeLists(ffc),
8244 8245 "free range is not in free lists");
8245 8246 }
8246 8247 _sp->removeFreeChunkFromFreeLists(ffc);
8247 8248 set_freeRangeInFreeLists(false);
8248 8249 }
8249 8250 set_lastFreeRangeCoalesced(true);
8250 8251 }
8251 8252 // this will be swept up when we hit the end of the free range
8252 8253
8253 8254 // Verify that the bit map has no bits marked between
8254 8255 // addr and purported end of just dead object.
8255 8256 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8256 8257 }
8257 8258 assert(_limit >= addr + size,
8258 8259 "A freshly garbage chunk can't possibly straddle over _limit");
8259 8260 if (inFreeRange()) lookahead_and_flush(fc, size);
8260 8261 return size;
8261 8262 }
8262 8263
8263 8264 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
8264 8265 HeapWord* addr = (HeapWord*) fc;
8265 8266 // The sweeper has just found a live object. Return any accumulated
8266 8267 // left hand chunk to the free lists.
8267 8268 if (inFreeRange()) {
8268 8269 assert(freeFinger() < addr, "freeFinger points too high");
8269 8270 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8270 8271 }
8271 8272
8272 8273 // This object is live: we'd normally expect this to be
8273 8274 // an oop, and like to assert the following:
8274 8275 // assert(oop(addr)->is_oop(), "live block should be an oop");
8275 8276 // However, as we commented above, this may be an object whose
8276 8277 // header hasn't yet been initialized.
8277 8278 size_t size;
8278 8279 assert(_bitMap->isMarked(addr), "Tautology for this control point");
8279 8280 if (_bitMap->isMarked(addr + 1)) {
8280 8281 // Determine the size from the bit map, rather than trying to
8281 8282 // compute it from the object header.
8282 8283 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
8283 8284 size = pointer_delta(nextOneAddr + 1, addr);
8284 8285 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
8285 8286 "alignment problem");
8286 8287
8287 8288 #ifdef DEBUG
8288 8289 if (oop(addr)->klass_or_null() != NULL &&
8289 8290 ( !_collector->should_unload_classes()
8290 8291 || (oop(addr)->is_parsable()) &&
8291 8292 oop(addr)->is_conc_safe())) {
8292 8293 // Ignore mark word because we are running concurrent with mutators
8293 8294 assert(oop(addr)->is_oop(true), "live block should be an oop");
8294 8295 // is_conc_safe is checked before performing this assertion
8295 8296 // because an object that is not is_conc_safe may yet have
8296 8297 // the return from size() correct.
8297 8298 assert(size ==
8298 8299 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
8299 8300 "P-mark and computed size do not agree");
8300 8301 }
8301 8302 #endif
8302 8303
8303 8304 } else {
8304 8305 // This should be an initialized object that's alive.
8305 8306 assert(oop(addr)->klass_or_null() != NULL &&
8306 8307 (!_collector->should_unload_classes()
8307 8308 || oop(addr)->is_parsable()),
8308 8309 "Should be an initialized object");
8309 8310 // Note that there are objects used during class redefinition,
8310 8311 // e.g. merge_cp in VM_RedefineClasses::merge_cp_and_rewrite(),
8311 8312 // which are discarded with their is_conc_safe state still
8312 8313 // false. These object may be floating garbage so may be
8313 8314 // seen here. If they are floating garbage their size
8314 8315 // should be attainable from their klass. Do not that
8315 8316 // is_conc_safe() is true for oop(addr).
8316 8317 // Ignore mark word because we are running concurrent with mutators
8317 8318 assert(oop(addr)->is_oop(true), "live block should be an oop");
8318 8319 // Verify that the bit map has no bits marked between
8319 8320 // addr and purported end of this block.
8320 8321 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8321 8322 assert(size >= 3, "Necessary for Printezis marks to work");
8322 8323 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
8323 8324 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
8324 8325 }
8325 8326 return size;
8326 8327 }
8327 8328
8328 8329 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
8329 8330 size_t chunkSize) {
8330 8331 // do_post_free_or_garbage_chunk() should only be called in the case
8331 8332 // of the adaptive free list allocator.
8332 8333 const bool fcInFreeLists = fc->isFree();
8333 8334 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
8334 8335 assert((HeapWord*)fc <= _limit, "sweep invariant");
8335 8336 if (CMSTestInFreeList && fcInFreeLists) {
8336 8337 assert(_sp->verifyChunkInFreeLists(fc), "free chunk is not in free lists");
8337 8338 }
8338 8339
8339 8340 if (CMSTraceSweeper) {
8340 8341 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
8341 8342 }
8342 8343
8343 8344 HeapWord* const fc_addr = (HeapWord*) fc;
8344 8345
8345 8346 bool coalesce;
8346 8347 const size_t left = pointer_delta(fc_addr, freeFinger());
8347 8348 const size_t right = chunkSize;
8348 8349 switch (FLSCoalescePolicy) {
8349 8350 // numeric value forms a coalition aggressiveness metric
8350 8351 case 0: { // never coalesce
8351 8352 coalesce = false;
8352 8353 break;
8353 8354 }
8354 8355 case 1: { // coalesce if left & right chunks on overpopulated lists
8355 8356 coalesce = _sp->coalOverPopulated(left) &&
8356 8357 _sp->coalOverPopulated(right);
8357 8358 break;
8358 8359 }
8359 8360 case 2: { // coalesce if left chunk on overpopulated list (default)
8360 8361 coalesce = _sp->coalOverPopulated(left);
8361 8362 break;
8362 8363 }
8363 8364 case 3: { // coalesce if left OR right chunk on overpopulated list
8364 8365 coalesce = _sp->coalOverPopulated(left) ||
8365 8366 _sp->coalOverPopulated(right);
8366 8367 break;
8367 8368 }
8368 8369 case 4: { // always coalesce
8369 8370 coalesce = true;
8370 8371 break;
8371 8372 }
8372 8373 default:
8373 8374 ShouldNotReachHere();
8374 8375 }
8375 8376
8376 8377 // Should the current free range be coalesced?
8377 8378 // If the chunk is in a free range and either we decided to coalesce above
8378 8379 // or the chunk is near the large block at the end of the heap
8379 8380 // (isNearLargestChunk() returns true), then coalesce this chunk.
8380 8381 const bool doCoalesce = inFreeRange()
8381 8382 && (coalesce || _g->isNearLargestChunk(fc_addr));
8382 8383 if (doCoalesce) {
8383 8384 // Coalesce the current free range on the left with the new
8384 8385 // chunk on the right. If either is on a free list,
8385 8386 // it must be removed from the list and stashed in the closure.
8386 8387 if (freeRangeInFreeLists()) {
8387 8388 FreeChunk* const ffc = (FreeChunk*)freeFinger();
8388 8389 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
8389 8390 "Size of free range is inconsistent with chunk size.");
8390 8391 if (CMSTestInFreeList) {
8391 8392 assert(_sp->verifyChunkInFreeLists(ffc),
8392 8393 "Chunk is not in free lists");
8393 8394 }
8394 8395 _sp->coalDeath(ffc->size());
8395 8396 _sp->removeFreeChunkFromFreeLists(ffc);
8396 8397 set_freeRangeInFreeLists(false);
8397 8398 }
8398 8399 if (fcInFreeLists) {
8399 8400 _sp->coalDeath(chunkSize);
8400 8401 assert(fc->size() == chunkSize,
8401 8402 "The chunk has the wrong size or is not in the free lists");
8402 8403 _sp->removeFreeChunkFromFreeLists(fc);
8403 8404 }
8404 8405 set_lastFreeRangeCoalesced(true);
8405 8406 print_free_block_coalesced(fc);
8406 8407 } else { // not in a free range and/or should not coalesce
8407 8408 // Return the current free range and start a new one.
8408 8409 if (inFreeRange()) {
8409 8410 // In a free range but cannot coalesce with the right hand chunk.
8410 8411 // Put the current free range into the free lists.
8411 8412 flush_cur_free_chunk(freeFinger(),
8412 8413 pointer_delta(fc_addr, freeFinger()));
8413 8414 }
8414 8415 // Set up for new free range. Pass along whether the right hand
8415 8416 // chunk is in the free lists.
8416 8417 initialize_free_range((HeapWord*)fc, fcInFreeLists);
8417 8418 }
8418 8419 }
8419 8420
8420 8421 // Lookahead flush:
8421 8422 // If we are tracking a free range, and this is the last chunk that
8422 8423 // we'll look at because its end crosses past _limit, we'll preemptively
8423 8424 // flush it along with any free range we may be holding on to. Note that
8424 8425 // this can be the case only for an already free or freshly garbage
8425 8426 // chunk. If this block is an object, it can never straddle
8426 8427 // over _limit. The "straddling" occurs when _limit is set at
8427 8428 // the previous end of the space when this cycle started, and
8428 8429 // a subsequent heap expansion caused the previously co-terminal
8429 8430 // free block to be coalesced with the newly expanded portion,
8430 8431 // thus rendering _limit a non-block-boundary making it dangerous
8431 8432 // for the sweeper to step over and examine.
8432 8433 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
8433 8434 assert(inFreeRange(), "Should only be called if currently in a free range.");
8434 8435 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
8435 8436 assert(_sp->used_region().contains(eob - 1),
8436 8437 err_msg("eob = " PTR_FORMAT " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
8437 8438 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
8438 8439 _limit, _sp->bottom(), _sp->end(), fc, chunk_size));
8439 8440 if (eob >= _limit) {
8440 8441 assert(eob == _limit || fc->isFree(), "Only a free chunk should allow us to cross over the limit");
8441 8442 if (CMSTraceSweeper) {
8442 8443 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
8443 8444 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
8444 8445 "[" PTR_FORMAT "," PTR_FORMAT ")",
8445 8446 _limit, fc, eob, _sp->bottom(), _sp->end());
8446 8447 }
8447 8448 // Return the storage we are tracking back into the free lists.
8448 8449 if (CMSTraceSweeper) {
8449 8450 gclog_or_tty->print_cr("Flushing ... ");
8450 8451 }
8451 8452 assert(freeFinger() < eob, "Error");
8452 8453 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
8453 8454 }
8454 8455 }
8455 8456
8456 8457 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
8457 8458 assert(inFreeRange(), "Should only be called if currently in a free range.");
8458 8459 assert(size > 0,
8459 8460 "A zero sized chunk cannot be added to the free lists.");
8460 8461 if (!freeRangeInFreeLists()) {
8461 8462 if (CMSTestInFreeList) {
8462 8463 FreeChunk* fc = (FreeChunk*) chunk;
8463 8464 fc->setSize(size);
8464 8465 assert(!_sp->verifyChunkInFreeLists(fc),
8465 8466 "chunk should not be in free lists yet");
8466 8467 }
8467 8468 if (CMSTraceSweeper) {
8468 8469 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8469 8470 chunk, size);
8470 8471 }
8471 8472 // A new free range is going to be starting. The current
8472 8473 // free range has not been added to the free lists yet or
8473 8474 // was removed so add it back.
8474 8475 // If the current free range was coalesced, then the death
8475 8476 // of the free range was recorded. Record a birth now.
8476 8477 if (lastFreeRangeCoalesced()) {
8477 8478 _sp->coalBirth(size);
8478 8479 }
8479 8480 _sp->addChunkAndRepairOffsetTable(chunk, size,
8480 8481 lastFreeRangeCoalesced());
8481 8482 } else if (CMSTraceSweeper) {
8482 8483 gclog_or_tty->print_cr("Already in free list: nothing to flush");
8483 8484 }
8484 8485 set_inFreeRange(false);
8485 8486 set_freeRangeInFreeLists(false);
8486 8487 }
8487 8488
8488 8489 // We take a break if we've been at this for a while,
8489 8490 // so as to avoid monopolizing the locks involved.
8490 8491 void SweepClosure::do_yield_work(HeapWord* addr) {
8491 8492 // Return current free chunk being used for coalescing (if any)
8492 8493 // to the appropriate freelist. After yielding, the next
8493 8494 // free block encountered will start a coalescing range of
8494 8495 // free blocks. If the next free block is adjacent to the
8495 8496 // chunk just flushed, they will need to wait for the next
8496 8497 // sweep to be coalesced.
8497 8498 if (inFreeRange()) {
8498 8499 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8499 8500 }
8500 8501
8501 8502 // First give up the locks, then yield, then re-lock.
8502 8503 // We should probably use a constructor/destructor idiom to
8503 8504 // do this unlock/lock or modify the MutexUnlocker class to
8504 8505 // serve our purpose. XXX
8505 8506 assert_lock_strong(_bitMap->lock());
8506 8507 assert_lock_strong(_freelistLock);
8507 8508 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8508 8509 "CMS thread should hold CMS token");
8509 8510 _bitMap->lock()->unlock();
8510 8511 _freelistLock->unlock();
8511 8512 ConcurrentMarkSweepThread::desynchronize(true);
8512 8513 ConcurrentMarkSweepThread::acknowledge_yield_request();
8513 8514 _collector->stopTimer();
8514 8515 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8515 8516 if (PrintCMSStatistics != 0) {
8516 8517 _collector->incrementYields();
8517 8518 }
8518 8519 _collector->icms_wait();
8519 8520
8520 8521 // See the comment in coordinator_yield()
8521 8522 for (unsigned i = 0; i < CMSYieldSleepCount &&
8522 8523 ConcurrentMarkSweepThread::should_yield() &&
8523 8524 !CMSCollector::foregroundGCIsActive(); ++i) {
8524 8525 os::sleep(Thread::current(), 1, false);
8525 8526 ConcurrentMarkSweepThread::acknowledge_yield_request();
8526 8527 }
8527 8528
8528 8529 ConcurrentMarkSweepThread::synchronize(true);
8529 8530 _freelistLock->lock();
8530 8531 _bitMap->lock()->lock_without_safepoint_check();
8531 8532 _collector->startTimer();
8532 8533 }
8533 8534
8534 8535 #ifndef PRODUCT
8535 8536 // This is actually very useful in a product build if it can
8536 8537 // be called from the debugger. Compile it into the product
8537 8538 // as needed.
8538 8539 bool debug_verifyChunkInFreeLists(FreeChunk* fc) {
8539 8540 return debug_cms_space->verifyChunkInFreeLists(fc);
8540 8541 }
8541 8542 #endif
8542 8543
8543 8544 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
8544 8545 if (CMSTraceSweeper) {
8545 8546 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
8546 8547 fc, fc->size());
8547 8548 }
8548 8549 }
8549 8550
8550 8551 // CMSIsAliveClosure
8551 8552 bool CMSIsAliveClosure::do_object_b(oop obj) {
8552 8553 HeapWord* addr = (HeapWord*)obj;
8553 8554 return addr != NULL &&
8554 8555 (!_span.contains(addr) || _bit_map->isMarked(addr));
8555 8556 }
8556 8557
8557 8558 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
8558 8559 MemRegion span,
8559 8560 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
8560 8561 CMSMarkStack* revisit_stack, bool cpc):
8561 8562 KlassRememberingOopClosure(collector, NULL, revisit_stack),
8562 8563 _span(span),
8563 8564 _bit_map(bit_map),
8564 8565 _mark_stack(mark_stack),
8565 8566 _concurrent_precleaning(cpc) {
8566 8567 assert(!_span.is_empty(), "Empty span could spell trouble");
8567 8568 }
8568 8569
8569 8570
8570 8571 // CMSKeepAliveClosure: the serial version
8571 8572 void CMSKeepAliveClosure::do_oop(oop obj) {
8572 8573 HeapWord* addr = (HeapWord*)obj;
8573 8574 if (_span.contains(addr) &&
8574 8575 !_bit_map->isMarked(addr)) {
8575 8576 _bit_map->mark(addr);
8576 8577 bool simulate_overflow = false;
8577 8578 NOT_PRODUCT(
8578 8579 if (CMSMarkStackOverflowALot &&
8579 8580 _collector->simulate_overflow()) {
8580 8581 // simulate a stack overflow
8581 8582 simulate_overflow = true;
8582 8583 }
8583 8584 )
8584 8585 if (simulate_overflow || !_mark_stack->push(obj)) {
8585 8586 if (_concurrent_precleaning) {
8586 8587 // We dirty the overflown object and let the remark
8587 8588 // phase deal with it.
8588 8589 assert(_collector->overflow_list_is_empty(), "Error");
8589 8590 // In the case of object arrays, we need to dirty all of
8590 8591 // the cards that the object spans. No locking or atomics
8591 8592 // are needed since no one else can be mutating the mod union
8592 8593 // table.
8593 8594 if (obj->is_objArray()) {
8594 8595 size_t sz = obj->size();
8595 8596 HeapWord* end_card_addr =
8596 8597 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
8597 8598 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8598 8599 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8599 8600 _collector->_modUnionTable.mark_range(redirty_range);
8600 8601 } else {
8601 8602 _collector->_modUnionTable.mark(addr);
8602 8603 }
8603 8604 _collector->_ser_kac_preclean_ovflw++;
8604 8605 } else {
8605 8606 _collector->push_on_overflow_list(obj);
8606 8607 _collector->_ser_kac_ovflw++;
8607 8608 }
8608 8609 }
8609 8610 }
8610 8611 }
8611 8612
8612 8613 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8613 8614 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8614 8615
8615 8616 // CMSParKeepAliveClosure: a parallel version of the above.
8616 8617 // The work queues are private to each closure (thread),
8617 8618 // but (may be) available for stealing by other threads.
8618 8619 void CMSParKeepAliveClosure::do_oop(oop obj) {
8619 8620 HeapWord* addr = (HeapWord*)obj;
8620 8621 if (_span.contains(addr) &&
8621 8622 !_bit_map->isMarked(addr)) {
8622 8623 // In general, during recursive tracing, several threads
8623 8624 // may be concurrently getting here; the first one to
8624 8625 // "tag" it, claims it.
8625 8626 if (_bit_map->par_mark(addr)) {
8626 8627 bool res = _work_queue->push(obj);
8627 8628 assert(res, "Low water mark should be much less than capacity");
8628 8629 // Do a recursive trim in the hope that this will keep
8629 8630 // stack usage lower, but leave some oops for potential stealers
8630 8631 trim_queue(_low_water_mark);
8631 8632 } // Else, another thread got there first
8632 8633 }
8633 8634 }
8634 8635
8635 8636 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8636 8637 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8637 8638
8638 8639 void CMSParKeepAliveClosure::trim_queue(uint max) {
8639 8640 while (_work_queue->size() > max) {
8640 8641 oop new_oop;
8641 8642 if (_work_queue->pop_local(new_oop)) {
8642 8643 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8643 8644 assert(_bit_map->isMarked((HeapWord*)new_oop),
8644 8645 "no white objects on this stack!");
8645 8646 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8646 8647 // iterate over the oops in this oop, marking and pushing
8647 8648 // the ones in CMS heap (i.e. in _span).
8648 8649 new_oop->oop_iterate(&_mark_and_push);
8649 8650 }
8650 8651 }
8651 8652 }
8652 8653
8653 8654 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8654 8655 CMSCollector* collector,
8655 8656 MemRegion span, CMSBitMap* bit_map,
8656 8657 CMSMarkStack* revisit_stack,
8657 8658 OopTaskQueue* work_queue):
8658 8659 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
8659 8660 _span(span),
8660 8661 _bit_map(bit_map),
8661 8662 _work_queue(work_queue) { }
8662 8663
8663 8664 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8664 8665 HeapWord* addr = (HeapWord*)obj;
8665 8666 if (_span.contains(addr) &&
8666 8667 !_bit_map->isMarked(addr)) {
8667 8668 if (_bit_map->par_mark(addr)) {
8668 8669 bool simulate_overflow = false;
8669 8670 NOT_PRODUCT(
8670 8671 if (CMSMarkStackOverflowALot &&
8671 8672 _collector->par_simulate_overflow()) {
8672 8673 // simulate a stack overflow
8673 8674 simulate_overflow = true;
8674 8675 }
8675 8676 )
8676 8677 if (simulate_overflow || !_work_queue->push(obj)) {
8677 8678 _collector->par_push_on_overflow_list(obj);
8678 8679 _collector->_par_kac_ovflw++;
8679 8680 }
8680 8681 } // Else another thread got there already
8681 8682 }
8682 8683 }
8683 8684
8684 8685 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8685 8686 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8686 8687
8687 8688 //////////////////////////////////////////////////////////////////
8688 8689 // CMSExpansionCause /////////////////////////////
8689 8690 //////////////////////////////////////////////////////////////////
8690 8691 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8691 8692 switch (cause) {
8692 8693 case _no_expansion:
8693 8694 return "No expansion";
8694 8695 case _satisfy_free_ratio:
8695 8696 return "Free ratio";
8696 8697 case _satisfy_promotion:
8697 8698 return "Satisfy promotion";
8698 8699 case _satisfy_allocation:
8699 8700 return "allocation";
8700 8701 case _allocate_par_lab:
8701 8702 return "Par LAB";
8702 8703 case _allocate_par_spooling_space:
8703 8704 return "Par Spooling Space";
8704 8705 case _adaptive_size_policy:
8705 8706 return "Ergonomics";
8706 8707 default:
8707 8708 return "unknown";
8708 8709 }
8709 8710 }
8710 8711
8711 8712 void CMSDrainMarkingStackClosure::do_void() {
8712 8713 // the max number to take from overflow list at a time
8713 8714 const size_t num = _mark_stack->capacity()/4;
8714 8715 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8715 8716 "Overflow list should be NULL during concurrent phases");
8716 8717 while (!_mark_stack->isEmpty() ||
8717 8718 // if stack is empty, check the overflow list
8718 8719 _collector->take_from_overflow_list(num, _mark_stack)) {
8719 8720 oop obj = _mark_stack->pop();
8720 8721 HeapWord* addr = (HeapWord*)obj;
8721 8722 assert(_span.contains(addr), "Should be within span");
8722 8723 assert(_bit_map->isMarked(addr), "Should be marked");
8723 8724 assert(obj->is_oop(), "Should be an oop");
8724 8725 obj->oop_iterate(_keep_alive);
8725 8726 }
8726 8727 }
8727 8728
8728 8729 void CMSParDrainMarkingStackClosure::do_void() {
8729 8730 // drain queue
8730 8731 trim_queue(0);
8731 8732 }
8732 8733
8733 8734 // Trim our work_queue so its length is below max at return
8734 8735 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8735 8736 while (_work_queue->size() > max) {
8736 8737 oop new_oop;
8737 8738 if (_work_queue->pop_local(new_oop)) {
8738 8739 assert(new_oop->is_oop(), "Expected an oop");
8739 8740 assert(_bit_map->isMarked((HeapWord*)new_oop),
8740 8741 "no white objects on this stack!");
8741 8742 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8742 8743 // iterate over the oops in this oop, marking and pushing
8743 8744 // the ones in CMS heap (i.e. in _span).
8744 8745 new_oop->oop_iterate(&_mark_and_push);
8745 8746 }
8746 8747 }
8747 8748 }
8748 8749
8749 8750 ////////////////////////////////////////////////////////////////////
8750 8751 // Support for Marking Stack Overflow list handling and related code
8751 8752 ////////////////////////////////////////////////////////////////////
8752 8753 // Much of the following code is similar in shape and spirit to the
8753 8754 // code used in ParNewGC. We should try and share that code
8754 8755 // as much as possible in the future.
8755 8756
8756 8757 #ifndef PRODUCT
8757 8758 // Debugging support for CMSStackOverflowALot
8758 8759
8759 8760 // It's OK to call this multi-threaded; the worst thing
8760 8761 // that can happen is that we'll get a bunch of closely
8761 8762 // spaced simulated oveflows, but that's OK, in fact
8762 8763 // probably good as it would exercise the overflow code
8763 8764 // under contention.
8764 8765 bool CMSCollector::simulate_overflow() {
8765 8766 if (_overflow_counter-- <= 0) { // just being defensive
8766 8767 _overflow_counter = CMSMarkStackOverflowInterval;
8767 8768 return true;
8768 8769 } else {
8769 8770 return false;
8770 8771 }
8771 8772 }
8772 8773
8773 8774 bool CMSCollector::par_simulate_overflow() {
8774 8775 return simulate_overflow();
8775 8776 }
8776 8777 #endif
8777 8778
8778 8779 // Single-threaded
8779 8780 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8780 8781 assert(stack->isEmpty(), "Expected precondition");
8781 8782 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8782 8783 size_t i = num;
8783 8784 oop cur = _overflow_list;
8784 8785 const markOop proto = markOopDesc::prototype();
8785 8786 NOT_PRODUCT(ssize_t n = 0;)
8786 8787 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8787 8788 next = oop(cur->mark());
8788 8789 cur->set_mark(proto); // until proven otherwise
8789 8790 assert(cur->is_oop(), "Should be an oop");
8790 8791 bool res = stack->push(cur);
8791 8792 assert(res, "Bit off more than can chew?");
8792 8793 NOT_PRODUCT(n++;)
8793 8794 }
8794 8795 _overflow_list = cur;
8795 8796 #ifndef PRODUCT
8796 8797 assert(_num_par_pushes >= n, "Too many pops?");
8797 8798 _num_par_pushes -=n;
8798 8799 #endif
8799 8800 return !stack->isEmpty();
8800 8801 }
8801 8802
8802 8803 #define BUSY (oop(0x1aff1aff))
8803 8804 // (MT-safe) Get a prefix of at most "num" from the list.
8804 8805 // The overflow list is chained through the mark word of
8805 8806 // each object in the list. We fetch the entire list,
8806 8807 // break off a prefix of the right size and return the
8807 8808 // remainder. If other threads try to take objects from
8808 8809 // the overflow list at that time, they will wait for
8809 8810 // some time to see if data becomes available. If (and
8810 8811 // only if) another thread places one or more object(s)
8811 8812 // on the global list before we have returned the suffix
8812 8813 // to the global list, we will walk down our local list
8813 8814 // to find its end and append the global list to
8814 8815 // our suffix before returning it. This suffix walk can
8815 8816 // prove to be expensive (quadratic in the amount of traffic)
8816 8817 // when there are many objects in the overflow list and
8817 8818 // there is much producer-consumer contention on the list.
8818 8819 // *NOTE*: The overflow list manipulation code here and
8819 8820 // in ParNewGeneration:: are very similar in shape,
8820 8821 // except that in the ParNew case we use the old (from/eden)
8821 8822 // copy of the object to thread the list via its klass word.
8822 8823 // Because of the common code, if you make any changes in
8823 8824 // the code below, please check the ParNew version to see if
8824 8825 // similar changes might be needed.
8825 8826 // CR 6797058 has been filed to consolidate the common code.
8826 8827 bool CMSCollector::par_take_from_overflow_list(size_t num,
8827 8828 OopTaskQueue* work_q,
8828 8829 int no_of_gc_threads) {
8829 8830 assert(work_q->size() == 0, "First empty local work queue");
8830 8831 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8831 8832 if (_overflow_list == NULL) {
8832 8833 return false;
8833 8834 }
8834 8835 // Grab the entire list; we'll put back a suffix
8835 8836 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
8836 8837 Thread* tid = Thread::current();
8837 8838 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
8838 8839 // set to ParallelGCThreads.
8839 8840 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
8840 8841 size_t sleep_time_millis = MAX2((size_t)1, num/100);
8841 8842 // If the list is busy, we spin for a short while,
8842 8843 // sleeping between attempts to get the list.
8843 8844 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
8844 8845 os::sleep(tid, sleep_time_millis, false);
8845 8846 if (_overflow_list == NULL) {
8846 8847 // Nothing left to take
8847 8848 return false;
8848 8849 } else if (_overflow_list != BUSY) {
8849 8850 // Try and grab the prefix
8850 8851 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
8851 8852 }
8852 8853 }
8853 8854 // If the list was found to be empty, or we spun long
8854 8855 // enough, we give up and return empty-handed. If we leave
8855 8856 // the list in the BUSY state below, it must be the case that
8856 8857 // some other thread holds the overflow list and will set it
8857 8858 // to a non-BUSY state in the future.
8858 8859 if (prefix == NULL || prefix == BUSY) {
8859 8860 // Nothing to take or waited long enough
8860 8861 if (prefix == NULL) {
8861 8862 // Write back the NULL in case we overwrote it with BUSY above
8862 8863 // and it is still the same value.
8863 8864 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8864 8865 }
8865 8866 return false;
8866 8867 }
8867 8868 assert(prefix != NULL && prefix != BUSY, "Error");
8868 8869 size_t i = num;
8869 8870 oop cur = prefix;
8870 8871 // Walk down the first "num" objects, unless we reach the end.
8871 8872 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8872 8873 if (cur->mark() == NULL) {
8873 8874 // We have "num" or fewer elements in the list, so there
8874 8875 // is nothing to return to the global list.
8875 8876 // Write back the NULL in lieu of the BUSY we wrote
8876 8877 // above, if it is still the same value.
8877 8878 if (_overflow_list == BUSY) {
8878 8879 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8879 8880 }
8880 8881 } else {
8881 8882 // Chop off the suffix and rerturn it to the global list.
8882 8883 assert(cur->mark() != BUSY, "Error");
8883 8884 oop suffix_head = cur->mark(); // suffix will be put back on global list
8884 8885 cur->set_mark(NULL); // break off suffix
8885 8886 // It's possible that the list is still in the empty(busy) state
8886 8887 // we left it in a short while ago; in that case we may be
8887 8888 // able to place back the suffix without incurring the cost
8888 8889 // of a walk down the list.
8889 8890 oop observed_overflow_list = _overflow_list;
8890 8891 oop cur_overflow_list = observed_overflow_list;
8891 8892 bool attached = false;
8892 8893 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
8893 8894 observed_overflow_list =
8894 8895 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8895 8896 if (cur_overflow_list == observed_overflow_list) {
8896 8897 attached = true;
8897 8898 break;
8898 8899 } else cur_overflow_list = observed_overflow_list;
8899 8900 }
8900 8901 if (!attached) {
8901 8902 // Too bad, someone else sneaked in (at least) an element; we'll need
8902 8903 // to do a splice. Find tail of suffix so we can prepend suffix to global
8903 8904 // list.
8904 8905 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8905 8906 oop suffix_tail = cur;
8906 8907 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8907 8908 "Tautology");
8908 8909 observed_overflow_list = _overflow_list;
8909 8910 do {
8910 8911 cur_overflow_list = observed_overflow_list;
8911 8912 if (cur_overflow_list != BUSY) {
8912 8913 // Do the splice ...
8913 8914 suffix_tail->set_mark(markOop(cur_overflow_list));
8914 8915 } else { // cur_overflow_list == BUSY
8915 8916 suffix_tail->set_mark(NULL);
8916 8917 }
8917 8918 // ... and try to place spliced list back on overflow_list ...
8918 8919 observed_overflow_list =
8919 8920 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8920 8921 } while (cur_overflow_list != observed_overflow_list);
8921 8922 // ... until we have succeeded in doing so.
8922 8923 }
8923 8924 }
8924 8925
8925 8926 // Push the prefix elements on work_q
8926 8927 assert(prefix != NULL, "control point invariant");
8927 8928 const markOop proto = markOopDesc::prototype();
8928 8929 oop next;
8929 8930 NOT_PRODUCT(ssize_t n = 0;)
8930 8931 for (cur = prefix; cur != NULL; cur = next) {
8931 8932 next = oop(cur->mark());
8932 8933 cur->set_mark(proto); // until proven otherwise
8933 8934 assert(cur->is_oop(), "Should be an oop");
8934 8935 bool res = work_q->push(cur);
8935 8936 assert(res, "Bit off more than we can chew?");
8936 8937 NOT_PRODUCT(n++;)
8937 8938 }
8938 8939 #ifndef PRODUCT
8939 8940 assert(_num_par_pushes >= n, "Too many pops?");
8940 8941 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8941 8942 #endif
8942 8943 return true;
8943 8944 }
8944 8945
8945 8946 // Single-threaded
8946 8947 void CMSCollector::push_on_overflow_list(oop p) {
8947 8948 NOT_PRODUCT(_num_par_pushes++;)
8948 8949 assert(p->is_oop(), "Not an oop");
8949 8950 preserve_mark_if_necessary(p);
8950 8951 p->set_mark((markOop)_overflow_list);
8951 8952 _overflow_list = p;
8952 8953 }
8953 8954
8954 8955 // Multi-threaded; use CAS to prepend to overflow list
8955 8956 void CMSCollector::par_push_on_overflow_list(oop p) {
8956 8957 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8957 8958 assert(p->is_oop(), "Not an oop");
8958 8959 par_preserve_mark_if_necessary(p);
8959 8960 oop observed_overflow_list = _overflow_list;
8960 8961 oop cur_overflow_list;
8961 8962 do {
8962 8963 cur_overflow_list = observed_overflow_list;
8963 8964 if (cur_overflow_list != BUSY) {
8964 8965 p->set_mark(markOop(cur_overflow_list));
8965 8966 } else {
8966 8967 p->set_mark(NULL);
8967 8968 }
8968 8969 observed_overflow_list =
8969 8970 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8970 8971 } while (cur_overflow_list != observed_overflow_list);
8971 8972 }
8972 8973 #undef BUSY
8973 8974
8974 8975 // Single threaded
8975 8976 // General Note on GrowableArray: pushes may silently fail
8976 8977 // because we are (temporarily) out of C-heap for expanding
8977 8978 // the stack. The problem is quite ubiquitous and affects
8978 8979 // a lot of code in the JVM. The prudent thing for GrowableArray
8979 8980 // to do (for now) is to exit with an error. However, that may
8980 8981 // be too draconian in some cases because the caller may be
8981 8982 // able to recover without much harm. For such cases, we
8982 8983 // should probably introduce a "soft_push" method which returns
8983 8984 // an indication of success or failure with the assumption that
8984 8985 // the caller may be able to recover from a failure; code in
8985 8986 // the VM can then be changed, incrementally, to deal with such
8986 8987 // failures where possible, thus, incrementally hardening the VM
8987 8988 // in such low resource situations.
8988 8989 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8989 8990 _preserved_oop_stack.push(p);
8990 8991 _preserved_mark_stack.push(m);
8991 8992 assert(m == p->mark(), "Mark word changed");
8992 8993 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8993 8994 "bijection");
8994 8995 }
8995 8996
8996 8997 // Single threaded
8997 8998 void CMSCollector::preserve_mark_if_necessary(oop p) {
8998 8999 markOop m = p->mark();
8999 9000 if (m->must_be_preserved(p)) {
9000 9001 preserve_mark_work(p, m);
9001 9002 }
9002 9003 }
9003 9004
9004 9005 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
9005 9006 markOop m = p->mark();
9006 9007 if (m->must_be_preserved(p)) {
9007 9008 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
9008 9009 // Even though we read the mark word without holding
9009 9010 // the lock, we are assured that it will not change
9010 9011 // because we "own" this oop, so no other thread can
9011 9012 // be trying to push it on the overflow list; see
9012 9013 // the assertion in preserve_mark_work() that checks
9013 9014 // that m == p->mark().
9014 9015 preserve_mark_work(p, m);
9015 9016 }
9016 9017 }
9017 9018
9018 9019 // We should be able to do this multi-threaded,
9019 9020 // a chunk of stack being a task (this is
9020 9021 // correct because each oop only ever appears
9021 9022 // once in the overflow list. However, it's
9022 9023 // not very easy to completely overlap this with
9023 9024 // other operations, so will generally not be done
9024 9025 // until all work's been completed. Because we
9025 9026 // expect the preserved oop stack (set) to be small,
9026 9027 // it's probably fine to do this single-threaded.
9027 9028 // We can explore cleverer concurrent/overlapped/parallel
9028 9029 // processing of preserved marks if we feel the
9029 9030 // need for this in the future. Stack overflow should
9030 9031 // be so rare in practice and, when it happens, its
9031 9032 // effect on performance so great that this will
9032 9033 // likely just be in the noise anyway.
9033 9034 void CMSCollector::restore_preserved_marks_if_any() {
9034 9035 assert(SafepointSynchronize::is_at_safepoint(),
9035 9036 "world should be stopped");
9036 9037 assert(Thread::current()->is_ConcurrentGC_thread() ||
9037 9038 Thread::current()->is_VM_thread(),
9038 9039 "should be single-threaded");
9039 9040 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9040 9041 "bijection");
9041 9042
9042 9043 while (!_preserved_oop_stack.is_empty()) {
9043 9044 oop p = _preserved_oop_stack.pop();
9044 9045 assert(p->is_oop(), "Should be an oop");
9045 9046 assert(_span.contains(p), "oop should be in _span");
9046 9047 assert(p->mark() == markOopDesc::prototype(),
9047 9048 "Set when taken from overflow list");
9048 9049 markOop m = _preserved_mark_stack.pop();
9049 9050 p->set_mark(m);
9050 9051 }
9051 9052 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
9052 9053 "stacks were cleared above");
9053 9054 }
9054 9055
9055 9056 #ifndef PRODUCT
9056 9057 bool CMSCollector::no_preserved_marks() const {
9057 9058 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
9058 9059 }
9059 9060 #endif
9060 9061
9061 9062 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
9062 9063 {
9063 9064 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
9064 9065 CMSAdaptiveSizePolicy* size_policy =
9065 9066 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
9066 9067 assert(size_policy->is_gc_cms_adaptive_size_policy(),
9067 9068 "Wrong type for size policy");
9068 9069 return size_policy;
9069 9070 }
9070 9071
9071 9072 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
9072 9073 size_t desired_promo_size) {
9073 9074 if (cur_promo_size < desired_promo_size) {
9074 9075 size_t expand_bytes = desired_promo_size - cur_promo_size;
9075 9076 if (PrintAdaptiveSizePolicy && Verbose) {
9076 9077 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9077 9078 "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
9078 9079 expand_bytes);
9079 9080 }
9080 9081 expand(expand_bytes,
9081 9082 MinHeapDeltaBytes,
9082 9083 CMSExpansionCause::_adaptive_size_policy);
9083 9084 } else if (desired_promo_size < cur_promo_size) {
9084 9085 size_t shrink_bytes = cur_promo_size - desired_promo_size;
9085 9086 if (PrintAdaptiveSizePolicy && Verbose) {
9086 9087 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9087 9088 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
9088 9089 shrink_bytes);
9089 9090 }
9090 9091 shrink(shrink_bytes);
9091 9092 }
9092 9093 }
9093 9094
9094 9095 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
9095 9096 GenCollectedHeap* gch = GenCollectedHeap::heap();
9096 9097 CMSGCAdaptivePolicyCounters* counters =
9097 9098 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
9098 9099 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
9099 9100 "Wrong kind of counters");
9100 9101 return counters;
9101 9102 }
9102 9103
9103 9104
9104 9105 void ASConcurrentMarkSweepGeneration::update_counters() {
9105 9106 if (UsePerfData) {
9106 9107 _space_counters->update_all();
9107 9108 _gen_counters->update_all();
9108 9109 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9109 9110 GenCollectedHeap* gch = GenCollectedHeap::heap();
9110 9111 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9111 9112 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9112 9113 "Wrong gc statistics type");
9113 9114 counters->update_counters(gc_stats_l);
9114 9115 }
9115 9116 }
9116 9117
9117 9118 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
9118 9119 if (UsePerfData) {
9119 9120 _space_counters->update_used(used);
9120 9121 _space_counters->update_capacity();
9121 9122 _gen_counters->update_all();
9122 9123
9123 9124 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9124 9125 GenCollectedHeap* gch = GenCollectedHeap::heap();
9125 9126 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9126 9127 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9127 9128 "Wrong gc statistics type");
9128 9129 counters->update_counters(gc_stats_l);
9129 9130 }
9130 9131 }
9131 9132
9132 9133 // The desired expansion delta is computed so that:
9133 9134 // . desired free percentage or greater is used
9134 9135 void ASConcurrentMarkSweepGeneration::compute_new_size() {
9135 9136 assert_locked_or_safepoint(Heap_lock);
9136 9137
9137 9138 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
9138 9139
9139 9140 // If incremental collection failed, we just want to expand
9140 9141 // to the limit.
9141 9142 if (incremental_collection_failed()) {
9142 9143 clear_incremental_collection_failed();
9143 9144 grow_to_reserved();
9144 9145 return;
9145 9146 }
9146 9147
9147 9148 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing");
9148 9149
9149 9150 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
9150 9151 "Wrong type of heap");
9151 9152 int prev_level = level() - 1;
9152 9153 assert(prev_level >= 0, "The cms generation is the lowest generation");
9153 9154 Generation* prev_gen = gch->get_gen(prev_level);
9154 9155 assert(prev_gen->kind() == Generation::ASParNew,
9155 9156 "Wrong type of young generation");
9156 9157 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen;
9157 9158 size_t cur_eden = younger_gen->eden()->capacity();
9158 9159 CMSAdaptiveSizePolicy* size_policy = cms_size_policy();
9159 9160 size_t cur_promo = free();
9160 9161 size_policy->compute_tenured_generation_free_space(cur_promo,
9161 9162 max_available(),
9162 9163 cur_eden);
9163 9164 resize(cur_promo, size_policy->promo_size());
9164 9165
9165 9166 // Record the new size of the space in the cms generation
9166 9167 // that is available for promotions. This is temporary.
9167 9168 // It should be the desired promo size.
9168 9169 size_policy->avg_cms_promo()->sample(free());
9169 9170 size_policy->avg_old_live()->sample(used());
9170 9171
9171 9172 if (UsePerfData) {
9172 9173 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9173 9174 counters->update_cms_capacity_counter(capacity());
9174 9175 }
9175 9176 }
9176 9177
9177 9178 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
9178 9179 assert_locked_or_safepoint(Heap_lock);
9179 9180 assert_lock_strong(freelistLock());
9180 9181 HeapWord* old_end = _cmsSpace->end();
9181 9182 HeapWord* unallocated_start = _cmsSpace->unallocated_block();
9182 9183 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
9183 9184 FreeChunk* chunk_at_end = find_chunk_at_end();
9184 9185 if (chunk_at_end == NULL) {
9185 9186 // No room to shrink
9186 9187 if (PrintGCDetails && Verbose) {
9187 9188 gclog_or_tty->print_cr("No room to shrink: old_end "
9188 9189 PTR_FORMAT " unallocated_start " PTR_FORMAT
9189 9190 " chunk_at_end " PTR_FORMAT,
9190 9191 old_end, unallocated_start, chunk_at_end);
9191 9192 }
9192 9193 return;
9193 9194 } else {
9194 9195
9195 9196 // Find the chunk at the end of the space and determine
9196 9197 // how much it can be shrunk.
9197 9198 size_t shrinkable_size_in_bytes = chunk_at_end->size();
9198 9199 size_t aligned_shrinkable_size_in_bytes =
9199 9200 align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
9200 9201 assert(unallocated_start <= chunk_at_end->end(),
9201 9202 "Inconsistent chunk at end of space");
9202 9203 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
9203 9204 size_t word_size_before = heap_word_size(_virtual_space.committed_size());
9204 9205
9205 9206 // Shrink the underlying space
9206 9207 _virtual_space.shrink_by(bytes);
9207 9208 if (PrintGCDetails && Verbose) {
9208 9209 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
9209 9210 " desired_bytes " SIZE_FORMAT
9210 9211 " shrinkable_size_in_bytes " SIZE_FORMAT
9211 9212 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
9212 9213 " bytes " SIZE_FORMAT,
9213 9214 desired_bytes, shrinkable_size_in_bytes,
9214 9215 aligned_shrinkable_size_in_bytes, bytes);
9215 9216 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT
9216 9217 " unallocated_start " SIZE_FORMAT,
9217 9218 old_end, unallocated_start);
9218 9219 }
9219 9220
9220 9221 // If the space did shrink (shrinking is not guaranteed),
9221 9222 // shrink the chunk at the end by the appropriate amount.
9222 9223 if (((HeapWord*)_virtual_space.high()) < old_end) {
9223 9224 size_t new_word_size =
9224 9225 heap_word_size(_virtual_space.committed_size());
9225 9226
9226 9227 // Have to remove the chunk from the dictionary because it is changing
9227 9228 // size and might be someplace elsewhere in the dictionary.
9228 9229
9229 9230 // Get the chunk at end, shrink it, and put it
9230 9231 // back.
9231 9232 _cmsSpace->removeChunkFromDictionary(chunk_at_end);
9232 9233 size_t word_size_change = word_size_before - new_word_size;
9233 9234 size_t chunk_at_end_old_size = chunk_at_end->size();
9234 9235 assert(chunk_at_end_old_size >= word_size_change,
9235 9236 "Shrink is too large");
9236 9237 chunk_at_end->setSize(chunk_at_end_old_size -
9237 9238 word_size_change);
9238 9239 _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
9239 9240 word_size_change);
9240 9241
9241 9242 _cmsSpace->returnChunkToDictionary(chunk_at_end);
9242 9243
9243 9244 MemRegion mr(_cmsSpace->bottom(), new_word_size);
9244 9245 _bts->resize(new_word_size); // resize the block offset shared array
9245 9246 Universe::heap()->barrier_set()->resize_covered_region(mr);
9246 9247 _cmsSpace->assert_locked();
9247 9248 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
9248 9249
9249 9250 NOT_PRODUCT(_cmsSpace->dictionary()->verify());
9250 9251
9251 9252 // update the space and generation capacity counters
9252 9253 if (UsePerfData) {
9253 9254 _space_counters->update_capacity();
9254 9255 _gen_counters->update_all();
9255 9256 }
9256 9257
9257 9258 if (Verbose && PrintGCDetails) {
9258 9259 size_t new_mem_size = _virtual_space.committed_size();
9259 9260 size_t old_mem_size = new_mem_size + bytes;
9260 9261 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK",
9261 9262 name(), old_mem_size/K, bytes/K, new_mem_size/K);
9262 9263 }
9263 9264 }
9264 9265
9265 9266 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
9266 9267 "Inconsistency at end of space");
9267 9268 assert(chunk_at_end->end() == _cmsSpace->end(),
9268 9269 "Shrinking is inconsistent");
9269 9270 return;
9270 9271 }
9271 9272 }
9272 9273
9273 9274 // Transfer some number of overflown objects to usual marking
9274 9275 // stack. Return true if some objects were transferred.
9275 9276 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
9276 9277 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
9277 9278 (size_t)ParGCDesiredObjsFromOverflowList);
9278 9279
9279 9280 bool res = _collector->take_from_overflow_list(num, _mark_stack);
9280 9281 assert(_collector->overflow_list_is_empty() || res,
9281 9282 "If list is not empty, we should have taken something");
9282 9283 assert(!res || !_mark_stack->isEmpty(),
9283 9284 "If we took something, it should now be on our stack");
9284 9285 return res;
9285 9286 }
9286 9287
9287 9288 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
9288 9289 size_t res = _sp->block_size_no_stall(addr, _collector);
9289 9290 if (_sp->block_is_obj(addr)) {
9290 9291 if (_live_bit_map->isMarked(addr)) {
9291 9292 // It can't have been dead in a previous cycle
9292 9293 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
9293 9294 } else {
9294 9295 _dead_bit_map->mark(addr); // mark the dead object
9295 9296 }
9296 9297 }
9297 9298 // Could be 0, if the block size could not be computed without stalling.
9298 9299 return res;
9299 9300 }
9300 9301
9301 9302 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
9302 9303
9303 9304 switch (phase) {
9304 9305 case CMSCollector::InitialMarking:
9305 9306 initialize(true /* fullGC */ ,
9306 9307 cause /* cause of the GC */,
9307 9308 true /* recordGCBeginTime */,
9308 9309 true /* recordPreGCUsage */,
9309 9310 false /* recordPeakUsage */,
9310 9311 false /* recordPostGCusage */,
9311 9312 true /* recordAccumulatedGCTime */,
9312 9313 false /* recordGCEndTime */,
9313 9314 false /* countCollection */ );
9314 9315 break;
9315 9316
9316 9317 case CMSCollector::FinalMarking:
9317 9318 initialize(true /* fullGC */ ,
9318 9319 cause /* cause of the GC */,
9319 9320 false /* recordGCBeginTime */,
9320 9321 false /* recordPreGCUsage */,
9321 9322 false /* recordPeakUsage */,
9322 9323 false /* recordPostGCusage */,
9323 9324 true /* recordAccumulatedGCTime */,
9324 9325 false /* recordGCEndTime */,
9325 9326 false /* countCollection */ );
9326 9327 break;
9327 9328
9328 9329 case CMSCollector::Sweeping:
9329 9330 initialize(true /* fullGC */ ,
9330 9331 cause /* cause of the GC */,
9331 9332 false /* recordGCBeginTime */,
9332 9333 false /* recordPreGCUsage */,
9333 9334 true /* recordPeakUsage */,
9334 9335 true /* recordPostGCusage */,
9335 9336 false /* recordAccumulatedGCTime */,
9336 9337 true /* recordGCEndTime */,
9337 9338 true /* countCollection */ );
9338 9339 break;
9339 9340
9340 9341 default:
9341 9342 ShouldNotReachHere();
9342 9343 }
9343 9344 }
9344 9345
|
↓ open down ↓ |
5828 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX