Print this page
--- 1.300 ---
Merged changes between child workspace "/net/spot/workspaces/ysr/cms_bugs" and
parent workspace "/net/jano2/export2/hotspot/ws/main/gc_baseline".
--- 1.297.1.1 ---
6621144 CMS: assertion failure "is_cms_thread == Thread::current()->is_ConcurrentGC_thread()"
Split |
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 2001-2007 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 21 * have any questions.
22 22 *
23 23 */
24 24
25 25 # include "incls/_precompiled.incl"
26 26 # include "incls/_concurrentMarkSweepGeneration.cpp.incl"
27 27
28 28 // statics
29 29 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
30 30 bool CMSCollector::_full_gc_requested = false;
31 31
32 32 //////////////////////////////////////////////////////////////////
33 33 // In support of CMS/VM thread synchronization
34 34 //////////////////////////////////////////////////////////////////
35 35 // We split use of the CGC_lock into 2 "levels".
36 36 // The low-level locking is of the usual CGC_lock monitor. We introduce
37 37 // a higher level "token" (hereafter "CMS token") built on top of the
38 38 // low level monitor (hereafter "CGC lock").
39 39 // The token-passing protocol gives priority to the VM thread. The
40 40 // CMS-lock doesn't provide any fairness guarantees, but clients
41 41 // should ensure that it is only held for very short, bounded
42 42 // durations.
43 43 //
44 44 // When either of the CMS thread or the VM thread is involved in
45 45 // collection operations during which it does not want the other
46 46 // thread to interfere, it obtains the CMS token.
47 47 //
48 48 // If either thread tries to get the token while the other has
49 49 // it, that thread waits. However, if the VM thread and CMS thread
50 50 // both want the token, then the VM thread gets priority while the
51 51 // CMS thread waits. This ensures, for instance, that the "concurrent"
52 52 // phases of the CMS thread's work do not block out the VM thread
53 53 // for long periods of time as the CMS thread continues to hog
54 54 // the token. (See bug 4616232).
55 55 //
56 56 // The baton-passing functions are, however, controlled by the
57 57 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
58 58 // and here the low-level CMS lock, not the high level token,
59 59 // ensures mutual exclusion.
60 60 //
61 61 // Two important conditions that we have to satisfy:
62 62 // 1. if a thread does a low-level wait on the CMS lock, then it
63 63 // relinquishes the CMS token if it were holding that token
64 64 // when it acquired the low-level CMS lock.
65 65 // 2. any low-level notifications on the low-level lock
66 66 // should only be sent when a thread has relinquished the token.
67 67 //
68 68 // In the absence of either property, we'd have potential deadlock.
69 69 //
70 70 // We protect each of the CMS (concurrent and sequential) phases
71 71 // with the CMS _token_, not the CMS _lock_.
72 72 //
73 73 // The only code protected by CMS lock is the token acquisition code
74 74 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
75 75 // baton-passing code.
76 76 //
77 77 // Unfortunately, i couldn't come up with a good abstraction to factor and
78 78 // hide the naked CGC_lock manipulation in the baton-passing code
79 79 // further below. That's something we should try to do. Also, the proof
80 80 // of correctness of this 2-level locking scheme is far from obvious,
81 81 // and potentially quite slippery. We have an uneasy supsicion, for instance,
82 82 // that there may be a theoretical possibility of delay/starvation in the
83 83 // low-level lock/wait/notify scheme used for the baton-passing because of
84 84 // potential intereference with the priority scheme embodied in the
85 85 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
86 86 // invocation further below and marked with "XXX 20011219YSR".
87 87 // Indeed, as we note elsewhere, this may become yet more slippery
88 88 // in the presence of multiple CMS and/or multiple VM threads. XXX
89 89
90 90 class CMSTokenSync: public StackObj {
91 91 private:
92 92 bool _is_cms_thread;
93 93 public:
94 94 CMSTokenSync(bool is_cms_thread):
95 95 _is_cms_thread(is_cms_thread) {
96 96 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
97 97 "Incorrect argument to constructor");
98 98 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
99 99 }
100 100
101 101 ~CMSTokenSync() {
102 102 assert(_is_cms_thread ?
103 103 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
104 104 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
105 105 "Incorrect state");
106 106 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
107 107 }
108 108 };
109 109
110 110 // Convenience class that does a CMSTokenSync, and then acquires
111 111 // upto three locks.
112 112 class CMSTokenSyncWithLocks: public CMSTokenSync {
113 113 private:
114 114 // Note: locks are acquired in textual declaration order
115 115 // and released in the opposite order
116 116 MutexLockerEx _locker1, _locker2, _locker3;
117 117 public:
118 118 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
119 119 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
120 120 CMSTokenSync(is_cms_thread),
121 121 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
122 122 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
123 123 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
124 124 { }
125 125 };
126 126
127 127
128 128 // Wrapper class to temporarily disable icms during a foreground cms collection.
129 129 class ICMSDisabler: public StackObj {
130 130 public:
131 131 // The ctor disables icms and wakes up the thread so it notices the change;
132 132 // the dtor re-enables icms. Note that the CMSCollector methods will check
133 133 // CMSIncrementalMode.
134 134 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); }
135 135 ~ICMSDisabler() { CMSCollector::enable_icms(); }
136 136 };
137 137
138 138 //////////////////////////////////////////////////////////////////
139 139 // Concurrent Mark-Sweep Generation /////////////////////////////
140 140 //////////////////////////////////////////////////////////////////
141 141
142 142 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
143 143
144 144 // This struct contains per-thread things necessary to support parallel
145 145 // young-gen collection.
146 146 class CMSParGCThreadState: public CHeapObj {
147 147 public:
148 148 CFLS_LAB lab;
149 149 PromotionInfo promo;
150 150
151 151 // Constructor.
152 152 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
153 153 promo.setSpace(cfls);
154 154 }
155 155 };
156 156
157 157 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
158 158 ReservedSpace rs, size_t initial_byte_size, int level,
159 159 CardTableRS* ct, bool use_adaptive_freelists,
160 160 FreeBlockDictionary::DictionaryChoice dictionaryChoice) :
161 161 CardGeneration(rs, initial_byte_size, level, ct),
162 162 _dilatation_factor(((double)MinChunkSize)/((double)(oopDesc::header_size()))),
163 163 _debug_collection_type(Concurrent_collection_type)
164 164 {
165 165 HeapWord* bottom = (HeapWord*) _virtual_space.low();
166 166 HeapWord* end = (HeapWord*) _virtual_space.high();
167 167
168 168 _direct_allocated_words = 0;
169 169 NOT_PRODUCT(
170 170 _numObjectsPromoted = 0;
171 171 _numWordsPromoted = 0;
172 172 _numObjectsAllocated = 0;
173 173 _numWordsAllocated = 0;
174 174 )
175 175
176 176 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
177 177 use_adaptive_freelists,
178 178 dictionaryChoice);
179 179 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
180 180 if (_cmsSpace == NULL) {
181 181 vm_exit_during_initialization(
182 182 "CompactibleFreeListSpace allocation failure");
183 183 }
184 184 _cmsSpace->_gen = this;
185 185
186 186 _gc_stats = new CMSGCStats();
187 187
188 188 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
189 189 // offsets match. The ability to tell free chunks from objects
190 190 // depends on this property.
191 191 debug_only(
192 192 FreeChunk* junk = NULL;
193 193 assert(junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
194 194 "Offset of FreeChunk::_prev within FreeChunk must match"
195 195 " that of OopDesc::_klass within OopDesc");
196 196 )
197 197 if (ParallelGCThreads > 0) {
198 198 typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
199 199 _par_gc_thread_states =
200 200 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads);
201 201 if (_par_gc_thread_states == NULL) {
202 202 vm_exit_during_initialization("Could not allocate par gc structs");
203 203 }
204 204 for (uint i = 0; i < ParallelGCThreads; i++) {
205 205 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
206 206 if (_par_gc_thread_states[i] == NULL) {
207 207 vm_exit_during_initialization("Could not allocate par gc structs");
208 208 }
209 209 }
210 210 } else {
211 211 _par_gc_thread_states = NULL;
212 212 }
213 213 _incremental_collection_failed = false;
214 214 // The "dilatation_factor" is the expansion that can occur on
215 215 // account of the fact that the minimum object size in the CMS
216 216 // generation may be larger than that in, say, a contiguous young
217 217 // generation.
↓ open down ↓ |
217 lines elided |
↑ open up ↑ |
218 218 // Ideally, in the calculation below, we'd compute the dilatation
219 219 // factor as: MinChunkSize/(promoting_gen's min object size)
220 220 // Since we do not have such a general query interface for the
221 221 // promoting generation, we'll instead just use the mimimum
222 222 // object size (which today is a header's worth of space);
223 223 // note that all arithmetic is in units of HeapWords.
224 224 assert(MinChunkSize >= oopDesc::header_size(), "just checking");
225 225 assert(_dilatation_factor >= 1.0, "from previous assert");
226 226 }
227 227
228 +
229 +// The field "_initiating_occupancy" represents the occupancy percentage
230 +// at which we trigger a new collection cycle. Unless explicitly specified
231 +// via CMSInitiating[Perm]OccupancyFraction (argument "io" below), it
232 +// is calculated by:
233 +//
234 +// Let "f" be MinHeapFreeRatio in
235 +//
236 +// _intiating_occupancy = 100-f +
237 +// f * (CMSTrigger[Perm]Ratio/100)
238 +// where CMSTrigger[Perm]Ratio is the argument "tr" below.
239 +//
240 +// That is, if we assume the heap is at its desired maximum occupancy at the
241 +// end of a collection, we let CMSTrigger[Perm]Ratio of the (purported) free
242 +// space be allocated before initiating a new collection cycle.
243 +//
244 +void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) {
245 + assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments");
246 + if (io > 0) {
247 + _initiating_occupancy = (double)io / 100.0;
248 + } else {
249 + _initiating_occupancy = ((100 - MinHeapFreeRatio) +
250 + (double)(tr * MinHeapFreeRatio) / 100.0)
251 + / 100.0;
252 + }
253 +}
254 +
255 +
228 256 void ConcurrentMarkSweepGeneration::ref_processor_init() {
229 257 assert(collector() != NULL, "no collector");
230 258 collector()->ref_processor_init();
231 259 }
232 260
233 261 void CMSCollector::ref_processor_init() {
234 262 if (_ref_processor == NULL) {
235 263 // Allocate and initialize a reference processor
236 264 _ref_processor = ReferenceProcessor::create_ref_processor(
237 265 _span, // span
238 266 _cmsGen->refs_discovery_is_atomic(), // atomic_discovery
239 267 _cmsGen->refs_discovery_is_mt(), // mt_discovery
240 268 &_is_alive_closure,
241 269 ParallelGCThreads,
242 270 ParallelRefProcEnabled);
243 271 // Initialize the _ref_processor field of CMSGen
244 272 _cmsGen->set_ref_processor(_ref_processor);
245 273
246 274 // Allocate a dummy ref processor for perm gen.
247 275 ReferenceProcessor* rp2 = new ReferenceProcessor();
248 276 if (rp2 == NULL) {
249 277 vm_exit_during_initialization("Could not allocate ReferenceProcessor object");
250 278 }
251 279 _permGen->set_ref_processor(rp2);
252 280 }
253 281 }
254 282
255 283 CMSAdaptiveSizePolicy* CMSCollector::size_policy() {
256 284 GenCollectedHeap* gch = GenCollectedHeap::heap();
257 285 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
258 286 "Wrong type of heap");
259 287 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
260 288 gch->gen_policy()->size_policy();
261 289 assert(sp->is_gc_cms_adaptive_size_policy(),
262 290 "Wrong type of size policy");
263 291 return sp;
264 292 }
265 293
266 294 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() {
267 295 CMSGCAdaptivePolicyCounters* results =
268 296 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters();
269 297 assert(
270 298 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
271 299 "Wrong gc policy counter kind");
272 300 return results;
273 301 }
274 302
275 303
276 304 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
277 305
278 306 const char* gen_name = "old";
279 307
280 308 // Generation Counters - generation 1, 1 subspace
281 309 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space);
282 310
283 311 _space_counters = new GSpaceCounters(gen_name, 0,
284 312 _virtual_space.reserved_size(),
285 313 this, _gen_counters);
286 314 }
287 315
288 316 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
289 317 _cms_gen(cms_gen)
290 318 {
291 319 assert(alpha <= 100, "bad value");
292 320 _saved_alpha = alpha;
293 321
294 322 // Initialize the alphas to the bootstrap value of 100.
295 323 _gc0_alpha = _cms_alpha = 100;
296 324
297 325 _cms_begin_time.update();
298 326 _cms_end_time.update();
299 327
300 328 _gc0_duration = 0.0;
301 329 _gc0_period = 0.0;
302 330 _gc0_promoted = 0;
303 331
304 332 _cms_duration = 0.0;
305 333 _cms_period = 0.0;
306 334 _cms_allocated = 0;
307 335
308 336 _cms_used_at_gc0_begin = 0;
309 337 _cms_used_at_gc0_end = 0;
310 338 _allow_duty_cycle_reduction = false;
311 339 _valid_bits = 0;
312 340 _icms_duty_cycle = CMSIncrementalDutyCycle;
313 341 }
314 342
315 343 // If promotion failure handling is on use
316 344 // the padded average size of the promotion for each
317 345 // young generation collection.
318 346 double CMSStats::time_until_cms_gen_full() const {
319 347 size_t cms_free = _cms_gen->cmsSpace()->free();
320 348 GenCollectedHeap* gch = GenCollectedHeap::heap();
321 349 size_t expected_promotion = gch->get_gen(0)->capacity();
322 350 if (HandlePromotionFailure) {
323 351 expected_promotion = MIN2(
324 352 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average(),
325 353 expected_promotion);
326 354 }
327 355 if (cms_free > expected_promotion) {
328 356 // Start a cms collection if there isn't enough space to promote
329 357 // for the next minor collection. Use the padded average as
330 358 // a safety factor.
331 359 cms_free -= expected_promotion;
332 360
333 361 // Adjust by the safety factor.
334 362 double cms_free_dbl = (double)cms_free;
335 363 cms_free_dbl = cms_free_dbl * (100.0 - CMSIncrementalSafetyFactor) / 100.0;
336 364
337 365 if (PrintGCDetails && Verbose) {
338 366 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
339 367 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
340 368 cms_free, expected_promotion);
341 369 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
342 370 cms_free_dbl, cms_consumption_rate() + 1.0);
343 371 }
344 372 // Add 1 in case the consumption rate goes to zero.
345 373 return cms_free_dbl / (cms_consumption_rate() + 1.0);
346 374 }
347 375 return 0.0;
348 376 }
349 377
350 378 // Compare the duration of the cms collection to the
351 379 // time remaining before the cms generation is empty.
352 380 // Note that the time from the start of the cms collection
353 381 // to the start of the cms sweep (less than the total
354 382 // duration of the cms collection) can be used. This
355 383 // has been tried and some applications experienced
356 384 // promotion failures early in execution. This was
357 385 // possibly because the averages were not accurate
358 386 // enough at the beginning.
359 387 double CMSStats::time_until_cms_start() const {
360 388 // We add "gc0_period" to the "work" calculation
361 389 // below because this query is done (mostly) at the
362 390 // end of a scavenge, so we need to conservatively
363 391 // account for that much possible delay
364 392 // in the query so as to avoid concurrent mode failures
365 393 // due to starting the collection just a wee bit too
366 394 // late.
367 395 double work = cms_duration() + gc0_period();
368 396 double deadline = time_until_cms_gen_full();
369 397 if (work > deadline) {
370 398 if (Verbose && PrintGCDetails) {
371 399 gclog_or_tty->print(
372 400 " CMSCollector: collect because of anticipated promotion "
373 401 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
374 402 gc0_period(), time_until_cms_gen_full());
375 403 }
376 404 return 0.0;
377 405 }
378 406 return work - deadline;
379 407 }
380 408
381 409 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the
382 410 // amount of change to prevent wild oscillation.
383 411 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle,
384 412 unsigned int new_duty_cycle) {
385 413 assert(old_duty_cycle <= 100, "bad input value");
386 414 assert(new_duty_cycle <= 100, "bad input value");
387 415
388 416 // Note: use subtraction with caution since it may underflow (values are
389 417 // unsigned). Addition is safe since we're in the range 0-100.
390 418 unsigned int damped_duty_cycle = new_duty_cycle;
391 419 if (new_duty_cycle < old_duty_cycle) {
392 420 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U);
393 421 if (new_duty_cycle + largest_delta < old_duty_cycle) {
394 422 damped_duty_cycle = old_duty_cycle - largest_delta;
395 423 }
396 424 } else if (new_duty_cycle > old_duty_cycle) {
397 425 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U);
398 426 if (new_duty_cycle > old_duty_cycle + largest_delta) {
399 427 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U);
400 428 }
401 429 }
402 430 assert(damped_duty_cycle <= 100, "invalid duty cycle computed");
403 431
404 432 if (CMSTraceIncrementalPacing) {
405 433 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ",
406 434 old_duty_cycle, new_duty_cycle, damped_duty_cycle);
407 435 }
408 436 return damped_duty_cycle;
409 437 }
410 438
411 439 unsigned int CMSStats::icms_update_duty_cycle_impl() {
412 440 assert(CMSIncrementalPacing && valid(),
413 441 "should be handled in icms_update_duty_cycle()");
414 442
415 443 double cms_time_so_far = cms_timer().seconds();
416 444 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M;
417 445 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far);
418 446
419 447 // Avoid division by 0.
420 448 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01);
421 449 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full;
422 450
423 451 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U);
424 452 if (new_duty_cycle > _icms_duty_cycle) {
425 453 // Avoid very small duty cycles (1 or 2); 0 is allowed.
426 454 if (new_duty_cycle > 2) {
427 455 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle,
428 456 new_duty_cycle);
429 457 }
430 458 } else if (_allow_duty_cycle_reduction) {
431 459 // The duty cycle is reduced only once per cms cycle (see record_cms_end()).
432 460 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle);
433 461 // Respect the minimum duty cycle.
434 462 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin;
435 463 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle);
436 464 }
437 465
438 466 if (PrintGCDetails || CMSTraceIncrementalPacing) {
439 467 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle);
440 468 }
441 469
442 470 _allow_duty_cycle_reduction = false;
443 471 return _icms_duty_cycle;
444 472 }
445 473
446 474 #ifndef PRODUCT
447 475 void CMSStats::print_on(outputStream *st) const {
448 476 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
449 477 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
450 478 gc0_duration(), gc0_period(), gc0_promoted());
451 479 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
452 480 cms_duration(), cms_duration_per_mb(),
453 481 cms_period(), cms_allocated());
454 482 st->print(",cms_since_beg=%g,cms_since_end=%g",
455 483 cms_time_since_begin(), cms_time_since_end());
456 484 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
457 485 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
458 486 if (CMSIncrementalMode) {
459 487 st->print(",dc=%d", icms_duty_cycle());
460 488 }
461 489
462 490 if (valid()) {
463 491 st->print(",promo_rate=%g,cms_alloc_rate=%g",
464 492 promotion_rate(), cms_allocation_rate());
465 493 st->print(",cms_consumption_rate=%g,time_until_full=%g",
466 494 cms_consumption_rate(), time_until_cms_gen_full());
467 495 }
468 496 st->print(" ");
469 497 }
470 498 #endif // #ifndef PRODUCT
471 499
472 500 CMSCollector::CollectorState CMSCollector::_collectorState =
473 501 CMSCollector::Idling;
474 502 bool CMSCollector::_foregroundGCIsActive = false;
475 503 bool CMSCollector::_foregroundGCShouldWait = false;
476 504
477 505 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
478 506 ConcurrentMarkSweepGeneration* permGen,
479 507 CardTableRS* ct,
480 508 ConcurrentMarkSweepPolicy* cp):
481 509 _cmsGen(cmsGen),
482 510 _permGen(permGen),
483 511 _ct(ct),
484 512 _ref_processor(NULL), // will be set later
485 513 _conc_workers(NULL), // may be set later
486 514 _abort_preclean(false),
487 515 _start_sampling(false),
488 516 _between_prologue_and_epilogue(false),
489 517 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
490 518 _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"),
491 519 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
492 520 -1 /* lock-free */, "No_lock" /* dummy */),
493 521 _modUnionClosure(&_modUnionTable),
494 522 _modUnionClosurePar(&_modUnionTable),
495 523 _is_alive_closure(&_markBitMap),
496 524 _restart_addr(NULL),
497 525 _overflow_list(NULL),
498 526 _preserved_oop_stack(NULL),
499 527 _preserved_mark_stack(NULL),
500 528 _stats(cmsGen),
501 529 _eden_chunk_array(NULL), // may be set in ctor body
502 530 _eden_chunk_capacity(0), // -- ditto --
503 531 _eden_chunk_index(0), // -- ditto --
504 532 _survivor_plab_array(NULL), // -- ditto --
505 533 _survivor_chunk_array(NULL), // -- ditto --
506 534 _survivor_chunk_capacity(0), // -- ditto --
507 535 _survivor_chunk_index(0), // -- ditto --
508 536 _ser_pmc_preclean_ovflw(0),
509 537 _ser_pmc_remark_ovflw(0),
510 538 _par_pmc_remark_ovflw(0),
511 539 _ser_kac_ovflw(0),
512 540 _par_kac_ovflw(0),
↓ open down ↓ |
275 lines elided |
↑ open up ↑ |
513 541 #ifndef PRODUCT
514 542 _num_par_pushes(0),
515 543 #endif
516 544 _collection_count_start(0),
517 545 _verifying(false),
518 546 _icms_start_limit(NULL),
519 547 _icms_stop_limit(NULL),
520 548 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
521 549 _completed_initialization(false),
522 550 _collector_policy(cp),
523 - _unload_classes(false),
524 - _unloaded_classes_last_cycle(false),
551 + _should_unload_classes(false),
552 + _concurrent_cycles_since_last_unload(0),
525 553 _sweep_estimate(CMS_SweepWeight, CMS_SweepPadding)
526 554 {
527 555 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
528 556 ExplicitGCInvokesConcurrent = true;
529 557 }
530 558 // Now expand the span and allocate the collection support structures
531 559 // (MUT, marking bit map etc.) to cover both generations subject to
532 560 // collection.
533 561
534 562 // First check that _permGen is adjacent to _cmsGen and above it.
535 563 assert( _cmsGen->reserved().word_size() > 0
536 564 && _permGen->reserved().word_size() > 0,
537 565 "generations should not be of zero size");
538 566 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(),
539 567 "_cmsGen and _permGen should not overlap");
540 568 assert(_cmsGen->reserved().end() == _permGen->reserved().start(),
541 569 "_cmsGen->end() different from _permGen->start()");
542 570
543 571 // For use by dirty card to oop closures.
544 572 _cmsGen->cmsSpace()->set_collector(this);
545 573 _permGen->cmsSpace()->set_collector(this);
546 574
547 575 // Adjust my span to cover old (cms) gen and perm gen
548 576 _span = _cmsGen->reserved()._union(_permGen->reserved());
549 577 // Initialize the span of is_alive_closure
550 578 _is_alive_closure.set_span(_span);
551 579
552 580 // Allocate MUT and marking bit map
553 581 {
554 582 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
555 583 if (!_markBitMap.allocate(_span)) {
556 584 warning("Failed to allocate CMS Bit Map");
557 585 return;
558 586 }
559 587 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
560 588 }
561 589 {
562 590 _modUnionTable.allocate(_span);
563 591 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
564 592 }
565 593
566 594 if (!_markStack.allocate(CMSMarkStackSize)) {
567 595 warning("Failed to allocate CMS Marking Stack");
568 596 return;
569 597 }
570 598 if (!_revisitStack.allocate(CMSRevisitStackSize)) {
571 599 warning("Failed to allocate CMS Revisit Stack");
572 600 return;
573 601 }
574 602
575 603 // Support for multi-threaded concurrent phases
576 604 if (ParallelGCThreads > 0 && CMSConcurrentMTEnabled) {
577 605 if (FLAG_IS_DEFAULT(ParallelCMSThreads)) {
578 606 // just for now
579 607 FLAG_SET_DEFAULT(ParallelCMSThreads, (ParallelGCThreads + 3)/4);
580 608 }
581 609 if (ParallelCMSThreads > 1) {
582 610 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
583 611 ParallelCMSThreads, true);
584 612 if (_conc_workers == NULL) {
585 613 warning("GC/CMS: _conc_workers allocation failure: "
586 614 "forcing -CMSConcurrentMTEnabled");
587 615 CMSConcurrentMTEnabled = false;
588 616 }
589 617 } else {
590 618 CMSConcurrentMTEnabled = false;
591 619 }
592 620 }
593 621 if (!CMSConcurrentMTEnabled) {
594 622 ParallelCMSThreads = 0;
595 623 } else {
596 624 // Turn off CMSCleanOnEnter optimization temporarily for
597 625 // the MT case where it's not fixed yet; see 6178663.
598 626 CMSCleanOnEnter = false;
599 627 }
600 628 assert((_conc_workers != NULL) == (ParallelCMSThreads > 1),
601 629 "Inconsistency");
602 630
603 631 // Parallel task queues; these are shared for the
604 632 // concurrent and stop-world phases of CMS, but
605 633 // are not shared with parallel scavenge (ParNew).
606 634 {
607 635 uint i;
608 636 uint num_queues = (uint) MAX2(ParallelGCThreads, ParallelCMSThreads);
609 637
610 638 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
611 639 || ParallelRefProcEnabled)
612 640 && num_queues > 0) {
613 641 _task_queues = new OopTaskQueueSet(num_queues);
614 642 if (_task_queues == NULL) {
615 643 warning("task_queues allocation failure.");
616 644 return;
617 645 }
618 646 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues);
619 647 if (_hash_seed == NULL) {
620 648 warning("_hash_seed array allocation failure");
621 649 return;
622 650 }
623 651
624 652 // XXX use a global constant instead of 64!
625 653 typedef struct OopTaskQueuePadded {
626 654 OopTaskQueue work_queue;
627 655 char pad[64 - sizeof(OopTaskQueue)]; // prevent false sharing
628 656 } OopTaskQueuePadded;
629 657
630 658 for (i = 0; i < num_queues; i++) {
631 659 OopTaskQueuePadded *q_padded = new OopTaskQueuePadded();
632 660 if (q_padded == NULL) {
633 661 warning("work_queue allocation failure.");
634 662 return;
↓ open down ↓ |
100 lines elided |
↑ open up ↑ |
635 663 }
636 664 _task_queues->register_queue(i, &q_padded->work_queue);
637 665 }
638 666 for (i = 0; i < num_queues; i++) {
639 667 _task_queues->queue(i)->initialize();
640 668 _hash_seed[i] = 17; // copied from ParNew
641 669 }
642 670 }
643 671 }
644 672
645 - // "initiatingOccupancy" is the occupancy ratio at which we trigger
646 - // a new collection cycle. Unless explicitly specified via
647 - // CMSTriggerRatio, it is calculated by:
648 - // Let "f" be MinHeapFreeRatio in
649 - //
650 - // intiatingOccupancy = 100-f +
651 - // f * (CMSTriggerRatio/100)
652 - // That is, if we assume the heap is at its desired maximum occupancy at the
653 - // end of a collection, we let CMSTriggerRatio of the (purported) free
654 - // space be allocated before initiating a new collection cycle.
655 - if (CMSInitiatingOccupancyFraction > 0) {
656 - _initiatingOccupancy = (double)CMSInitiatingOccupancyFraction / 100.0;
657 - } else {
658 - _initiatingOccupancy = ((100 - MinHeapFreeRatio) +
659 - (double)(CMSTriggerRatio *
660 - MinHeapFreeRatio) / 100.0)
661 - / 100.0;
662 - }
673 + _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
674 + _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio);
675 +
663 676 // Clip CMSBootstrapOccupancy between 0 and 100.
664 - _bootstrap_occupancy = ((double)MIN2((intx)100, MAX2((intx)0, CMSBootstrapOccupancy)))
677 + _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy)))
665 678 /(double)100;
666 679
667 680 _full_gcs_since_conc_gc = 0;
668 681
669 682 // Now tell CMS generations the identity of their collector
670 683 ConcurrentMarkSweepGeneration::set_collector(this);
671 684
672 685 // Create & start a CMS thread for this CMS collector
673 686 _cmsThread = ConcurrentMarkSweepThread::start(this);
674 687 assert(cmsThread() != NULL, "CMS Thread should have been created");
675 688 assert(cmsThread()->collector() == this,
676 689 "CMS Thread should refer to this gen");
677 690 assert(CGC_lock != NULL, "Where's the CGC_lock?");
678 691
679 692 // Support for parallelizing young gen rescan
680 693 GenCollectedHeap* gch = GenCollectedHeap::heap();
681 694 _young_gen = gch->prev_gen(_cmsGen);
682 695 if (gch->supports_inline_contig_alloc()) {
683 696 _top_addr = gch->top_addr();
684 697 _end_addr = gch->end_addr();
685 698 assert(_young_gen != NULL, "no _young_gen");
686 699 _eden_chunk_index = 0;
687 700 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
688 701 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity);
689 702 if (_eden_chunk_array == NULL) {
690 703 _eden_chunk_capacity = 0;
691 704 warning("GC/CMS: _eden_chunk_array allocation failure");
692 705 }
693 706 }
694 707 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
695 708
696 709 // Support for parallelizing survivor space rescan
697 710 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) {
698 711 size_t max_plab_samples = MaxNewSize/((SurvivorRatio+2)*MinTLABSize);
699 712 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads);
700 713 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples);
701 714 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads);
702 715 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
703 716 || _cursor == NULL) {
704 717 warning("Failed to allocate survivor plab/chunk array");
705 718 if (_survivor_plab_array != NULL) {
706 719 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
707 720 _survivor_plab_array = NULL;
708 721 }
709 722 if (_survivor_chunk_array != NULL) {
710 723 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
711 724 _survivor_chunk_array = NULL;
712 725 }
713 726 if (_cursor != NULL) {
714 727 FREE_C_HEAP_ARRAY(size_t, _cursor);
715 728 _cursor = NULL;
716 729 }
717 730 } else {
718 731 _survivor_chunk_capacity = 2*max_plab_samples;
719 732 for (uint i = 0; i < ParallelGCThreads; i++) {
720 733 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples);
721 734 if (vec == NULL) {
722 735 warning("Failed to allocate survivor plab array");
723 736 for (int j = i; j > 0; j--) {
724 737 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array());
725 738 }
726 739 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
727 740 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
728 741 _survivor_plab_array = NULL;
729 742 _survivor_chunk_array = NULL;
730 743 _survivor_chunk_capacity = 0;
731 744 break;
732 745 } else {
733 746 ChunkArray* cur =
734 747 ::new (&_survivor_plab_array[i]) ChunkArray(vec,
735 748 max_plab_samples);
736 749 assert(cur->end() == 0, "Should be 0");
737 750 assert(cur->array() == vec, "Should be vec");
738 751 assert(cur->capacity() == max_plab_samples, "Error");
739 752 }
740 753 }
741 754 }
742 755 }
743 756 assert( ( _survivor_plab_array != NULL
744 757 && _survivor_chunk_array != NULL)
745 758 || ( _survivor_chunk_capacity == 0
746 759 && _survivor_chunk_index == 0),
747 760 "Error");
748 761
749 762 // Choose what strong roots should be scanned depending on verification options
750 763 // and perm gen collection mode.
751 764 if (!CMSClassUnloadingEnabled) {
752 765 // If class unloading is disabled we want to include all classes into the root set.
753 766 add_root_scanning_option(SharedHeap::SO_AllClasses);
754 767 } else {
755 768 add_root_scanning_option(SharedHeap::SO_SystemClasses);
756 769 }
757 770
758 771 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
759 772 _gc_counters = new CollectorCounters("CMS", 1);
760 773 _completed_initialization = true;
761 774 _sweep_timer.start(); // start of time
762 775 }
763 776
764 777 const char* ConcurrentMarkSweepGeneration::name() const {
765 778 return "concurrent mark-sweep generation";
766 779 }
767 780 void ConcurrentMarkSweepGeneration::update_counters() {
768 781 if (UsePerfData) {
769 782 _space_counters->update_all();
770 783 _gen_counters->update_all();
771 784 }
772 785 }
773 786
774 787 // this is an optimized version of update_counters(). it takes the
775 788 // used value as a parameter rather than computing it.
776 789 //
777 790 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
778 791 if (UsePerfData) {
779 792 _space_counters->update_used(used);
780 793 _space_counters->update_capacity();
781 794 _gen_counters->update_all();
782 795 }
783 796 }
784 797
785 798 void ConcurrentMarkSweepGeneration::print() const {
786 799 Generation::print();
787 800 cmsSpace()->print();
788 801 }
789 802
790 803 #ifndef PRODUCT
791 804 void ConcurrentMarkSweepGeneration::print_statistics() {
792 805 cmsSpace()->printFLCensus(0);
793 806 }
794 807 #endif
795 808
796 809 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
797 810 GenCollectedHeap* gch = GenCollectedHeap::heap();
798 811 if (PrintGCDetails) {
799 812 if (Verbose) {
800 813 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
801 814 level(), short_name(), s, used(), capacity());
802 815 } else {
803 816 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
804 817 level(), short_name(), s, used() / K, capacity() / K);
805 818 }
806 819 }
807 820 if (Verbose) {
808 821 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
809 822 gch->used(), gch->capacity());
810 823 } else {
811 824 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
812 825 gch->used() / K, gch->capacity() / K);
813 826 }
814 827 }
815 828
816 829 size_t
817 830 ConcurrentMarkSweepGeneration::contiguous_available() const {
818 831 // dld proposes an improvement in precision here. If the committed
819 832 // part of the space ends in a free block we should add that to
820 833 // uncommitted size in the calculation below. Will make this
821 834 // change later, staying with the approximation below for the
822 835 // time being. -- ysr.
823 836 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
824 837 }
825 838
826 839 size_t
827 840 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
828 841 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
829 842 }
830 843
831 844 size_t ConcurrentMarkSweepGeneration::max_available() const {
832 845 return free() + _virtual_space.uncommitted_size();
833 846 }
834 847
835 848 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(
836 849 size_t max_promotion_in_bytes,
837 850 bool younger_handles_promotion_failure) const {
838 851
839 852 // This is the most conservative test. Full promotion is
840 853 // guaranteed if this is used. The multiplicative factor is to
841 854 // account for the worst case "dilatation".
842 855 double adjusted_max_promo_bytes = _dilatation_factor * max_promotion_in_bytes;
843 856 if (adjusted_max_promo_bytes > (double)max_uintx) { // larger than size_t
844 857 adjusted_max_promo_bytes = (double)max_uintx;
845 858 }
846 859 bool result = (max_contiguous_available() >= (size_t)adjusted_max_promo_bytes);
847 860
848 861 if (younger_handles_promotion_failure && !result) {
849 862 // Full promotion is not guaranteed because fragmentation
850 863 // of the cms generation can prevent the full promotion.
851 864 result = (max_available() >= (size_t)adjusted_max_promo_bytes);
852 865
853 866 if (!result) {
854 867 // With promotion failure handling the test for the ability
855 868 // to support the promotion does not have to be guaranteed.
856 869 // Use an average of the amount promoted.
857 870 result = max_available() >= (size_t)
858 871 gc_stats()->avg_promoted()->padded_average();
859 872 if (PrintGC && Verbose && result) {
860 873 gclog_or_tty->print_cr(
861 874 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
862 875 " max_available: " SIZE_FORMAT
863 876 " avg_promoted: " SIZE_FORMAT,
864 877 max_available(), (size_t)
865 878 gc_stats()->avg_promoted()->padded_average());
866 879 }
867 880 } else {
868 881 if (PrintGC && Verbose) {
869 882 gclog_or_tty->print_cr(
870 883 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
871 884 " max_available: " SIZE_FORMAT
872 885 " adj_max_promo_bytes: " SIZE_FORMAT,
873 886 max_available(), (size_t)adjusted_max_promo_bytes);
874 887 }
875 888 }
876 889 } else {
877 890 if (PrintGC && Verbose) {
878 891 gclog_or_tty->print_cr(
879 892 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
880 893 " contiguous_available: " SIZE_FORMAT
881 894 " adj_max_promo_bytes: " SIZE_FORMAT,
882 895 max_contiguous_available(), (size_t)adjusted_max_promo_bytes);
883 896 }
884 897 }
885 898 return result;
886 899 }
887 900
888 901 CompactibleSpace*
889 902 ConcurrentMarkSweepGeneration::first_compaction_space() const {
890 903 return _cmsSpace;
891 904 }
892 905
893 906 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
894 907 // Clear the promotion information. These pointers can be adjusted
895 908 // along with all the other pointers into the heap but
896 909 // compaction is expected to be a rare event with
897 910 // a heap using cms so don't do it without seeing the need.
898 911 if (ParallelGCThreads > 0) {
899 912 for (uint i = 0; i < ParallelGCThreads; i++) {
900 913 _par_gc_thread_states[i]->promo.reset();
901 914 }
902 915 }
903 916 }
904 917
905 918 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
906 919 blk->do_space(_cmsSpace);
907 920 }
908 921
909 922 void ConcurrentMarkSweepGeneration::compute_new_size() {
910 923 assert_locked_or_safepoint(Heap_lock);
911 924
912 925 // If incremental collection failed, we just want to expand
913 926 // to the limit.
914 927 if (incremental_collection_failed()) {
915 928 clear_incremental_collection_failed();
916 929 grow_to_reserved();
917 930 return;
918 931 }
919 932
920 933 size_t expand_bytes = 0;
921 934 double free_percentage = ((double) free()) / capacity();
922 935 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
923 936 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
924 937
925 938 // compute expansion delta needed for reaching desired free percentage
926 939 if (free_percentage < desired_free_percentage) {
927 940 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
928 941 assert(desired_capacity >= capacity(), "invalid expansion size");
929 942 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
930 943 }
931 944 if (expand_bytes > 0) {
932 945 if (PrintGCDetails && Verbose) {
933 946 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
934 947 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
935 948 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
936 949 gclog_or_tty->print_cr(" Desired free fraction %f",
937 950 desired_free_percentage);
938 951 gclog_or_tty->print_cr(" Maximum free fraction %f",
939 952 maximum_free_percentage);
940 953 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000);
941 954 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
942 955 desired_capacity/1000);
943 956 int prev_level = level() - 1;
944 957 if (prev_level >= 0) {
945 958 size_t prev_size = 0;
946 959 GenCollectedHeap* gch = GenCollectedHeap::heap();
947 960 Generation* prev_gen = gch->_gens[prev_level];
948 961 prev_size = prev_gen->capacity();
949 962 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
950 963 prev_size/1000);
951 964 }
952 965 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
953 966 unsafe_max_alloc_nogc()/1000);
954 967 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
955 968 contiguous_available()/1000);
956 969 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
957 970 expand_bytes);
958 971 }
959 972 // safe if expansion fails
960 973 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
961 974 if (PrintGCDetails && Verbose) {
962 975 gclog_or_tty->print_cr(" Expanded free fraction %f",
963 976 ((double) free()) / capacity());
964 977 }
965 978 }
966 979 }
967 980
968 981 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
969 982 return cmsSpace()->freelistLock();
970 983 }
971 984
972 985 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
973 986 bool tlab) {
974 987 CMSSynchronousYieldRequest yr;
975 988 MutexLockerEx x(freelistLock(),
976 989 Mutex::_no_safepoint_check_flag);
977 990 return have_lock_and_allocate(size, tlab);
978 991 }
979 992
980 993 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
981 994 bool tlab) {
982 995 assert_lock_strong(freelistLock());
983 996 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
984 997 HeapWord* res = cmsSpace()->allocate(adjustedSize);
985 998 // Allocate the object live (grey) if the background collector has
986 999 // started marking. This is necessary because the marker may
987 1000 // have passed this address and consequently this object will
988 1001 // not otherwise be greyed and would be incorrectly swept up.
989 1002 // Note that if this object contains references, the writing
990 1003 // of those references will dirty the card containing this object
991 1004 // allowing the object to be blackened (and its references scanned)
992 1005 // either during a preclean phase or at the final checkpoint.
993 1006 if (res != NULL) {
994 1007 collector()->direct_allocated(res, adjustedSize);
995 1008 _direct_allocated_words += adjustedSize;
996 1009 // allocation counters
997 1010 NOT_PRODUCT(
998 1011 _numObjectsAllocated++;
999 1012 _numWordsAllocated += (int)adjustedSize;
1000 1013 )
1001 1014 }
1002 1015 return res;
1003 1016 }
1004 1017
1005 1018 // In the case of direct allocation by mutators in a generation that
1006 1019 // is being concurrently collected, the object must be allocated
1007 1020 // live (grey) if the background collector has started marking.
1008 1021 // This is necessary because the marker may
1009 1022 // have passed this address and consequently this object will
1010 1023 // not otherwise be greyed and would be incorrectly swept up.
1011 1024 // Note that if this object contains references, the writing
1012 1025 // of those references will dirty the card containing this object
1013 1026 // allowing the object to be blackened (and its references scanned)
1014 1027 // either during a preclean phase or at the final checkpoint.
1015 1028 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
1016 1029 assert(_markBitMap.covers(start, size), "Out of bounds");
1017 1030 if (_collectorState >= Marking) {
1018 1031 MutexLockerEx y(_markBitMap.lock(),
1019 1032 Mutex::_no_safepoint_check_flag);
1020 1033 // [see comments preceding SweepClosure::do_blk() below for details]
1021 1034 // 1. need to mark the object as live so it isn't collected
1022 1035 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
1023 1036 // 3. need to mark the end of the object so sweeper can skip over it
1024 1037 // if it's uninitialized when the sweeper reaches it.
1025 1038 _markBitMap.mark(start); // object is live
1026 1039 _markBitMap.mark(start + 1); // object is potentially uninitialized?
1027 1040 _markBitMap.mark(start + size - 1);
1028 1041 // mark end of object
1029 1042 }
1030 1043 // check that oop looks uninitialized
1031 1044 assert(oop(start)->klass() == NULL, "_klass should be NULL");
1032 1045 }
1033 1046
1034 1047 void CMSCollector::promoted(bool par, HeapWord* start,
1035 1048 bool is_obj_array, size_t obj_size) {
1036 1049 assert(_markBitMap.covers(start), "Out of bounds");
1037 1050 // See comment in direct_allocated() about when objects should
1038 1051 // be allocated live.
1039 1052 if (_collectorState >= Marking) {
1040 1053 // we already hold the marking bit map lock, taken in
1041 1054 // the prologue
1042 1055 if (par) {
1043 1056 _markBitMap.par_mark(start);
1044 1057 } else {
1045 1058 _markBitMap.mark(start);
1046 1059 }
1047 1060 // We don't need to mark the object as uninitialized (as
1048 1061 // in direct_allocated above) because this is being done with the
1049 1062 // world stopped and the object will be initialized by the
1050 1063 // time the sweeper gets to look at it.
1051 1064 assert(SafepointSynchronize::is_at_safepoint(),
1052 1065 "expect promotion only at safepoints");
1053 1066
1054 1067 if (_collectorState < Sweeping) {
1055 1068 // Mark the appropriate cards in the modUnionTable, so that
1056 1069 // this object gets scanned before the sweep. If this is
1057 1070 // not done, CMS generation references in the object might
1058 1071 // not get marked.
1059 1072 // For the case of arrays, which are otherwise precisely
1060 1073 // marked, we need to dirty the entire array, not just its head.
1061 1074 if (is_obj_array) {
1062 1075 // The [par_]mark_range() method expects mr.end() below to
1063 1076 // be aligned to the granularity of a bit's representation
1064 1077 // in the heap. In the case of the MUT below, that's a
1065 1078 // card size.
1066 1079 MemRegion mr(start,
1067 1080 (HeapWord*)round_to((intptr_t)(start + obj_size),
1068 1081 CardTableModRefBS::card_size /* bytes */));
1069 1082 if (par) {
1070 1083 _modUnionTable.par_mark_range(mr);
1071 1084 } else {
1072 1085 _modUnionTable.mark_range(mr);
1073 1086 }
1074 1087 } else { // not an obj array; we can just mark the head
1075 1088 if (par) {
1076 1089 _modUnionTable.par_mark(start);
1077 1090 } else {
1078 1091 _modUnionTable.mark(start);
1079 1092 }
1080 1093 }
1081 1094 }
1082 1095 }
1083 1096 }
1084 1097
1085 1098 static inline size_t percent_of_space(Space* space, HeapWord* addr)
1086 1099 {
1087 1100 size_t delta = pointer_delta(addr, space->bottom());
1088 1101 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
1089 1102 }
1090 1103
1091 1104 void CMSCollector::icms_update_allocation_limits()
1092 1105 {
1093 1106 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
1094 1107 EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
1095 1108
1096 1109 const unsigned int duty_cycle = stats().icms_update_duty_cycle();
1097 1110 if (CMSTraceIncrementalPacing) {
1098 1111 stats().print();
1099 1112 }
1100 1113
1101 1114 assert(duty_cycle <= 100, "invalid duty cycle");
1102 1115 if (duty_cycle != 0) {
1103 1116 // The duty_cycle is a percentage between 0 and 100; convert to words and
1104 1117 // then compute the offset from the endpoints of the space.
1105 1118 size_t free_words = eden->free() / HeapWordSize;
1106 1119 double free_words_dbl = (double)free_words;
1107 1120 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
1108 1121 size_t offset_words = (free_words - duty_cycle_words) / 2;
1109 1122
1110 1123 _icms_start_limit = eden->top() + offset_words;
1111 1124 _icms_stop_limit = eden->end() - offset_words;
1112 1125
1113 1126 // The limits may be adjusted (shifted to the right) by
1114 1127 // CMSIncrementalOffset, to allow the application more mutator time after a
1115 1128 // young gen gc (when all mutators were stopped) and before CMS starts and
1116 1129 // takes away one or more cpus.
1117 1130 if (CMSIncrementalOffset != 0) {
1118 1131 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
1119 1132 size_t adjustment = (size_t)adjustment_dbl;
1120 1133 HeapWord* tmp_stop = _icms_stop_limit + adjustment;
1121 1134 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
1122 1135 _icms_start_limit += adjustment;
1123 1136 _icms_stop_limit = tmp_stop;
1124 1137 }
1125 1138 }
1126 1139 }
1127 1140 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
1128 1141 _icms_start_limit = _icms_stop_limit = eden->end();
1129 1142 }
1130 1143
1131 1144 // Install the new start limit.
1132 1145 eden->set_soft_end(_icms_start_limit);
1133 1146
1134 1147 if (CMSTraceIncrementalMode) {
1135 1148 gclog_or_tty->print(" icms alloc limits: "
1136 1149 PTR_FORMAT "," PTR_FORMAT
1137 1150 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
1138 1151 _icms_start_limit, _icms_stop_limit,
1139 1152 percent_of_space(eden, _icms_start_limit),
1140 1153 percent_of_space(eden, _icms_stop_limit));
1141 1154 if (Verbose) {
1142 1155 gclog_or_tty->print("eden: ");
1143 1156 eden->print_on(gclog_or_tty);
1144 1157 }
1145 1158 }
1146 1159 }
1147 1160
1148 1161 // Any changes here should try to maintain the invariant
1149 1162 // that if this method is called with _icms_start_limit
1150 1163 // and _icms_stop_limit both NULL, then it should return NULL
1151 1164 // and not notify the icms thread.
1152 1165 HeapWord*
1153 1166 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
1154 1167 size_t word_size)
1155 1168 {
1156 1169 // A start_limit equal to end() means the duty cycle is 0, so treat that as a
1157 1170 // nop.
1158 1171 if (CMSIncrementalMode && _icms_start_limit != space->end()) {
1159 1172 if (top <= _icms_start_limit) {
1160 1173 if (CMSTraceIncrementalMode) {
1161 1174 space->print_on(gclog_or_tty);
1162 1175 gclog_or_tty->stamp();
1163 1176 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
1164 1177 ", new limit=" PTR_FORMAT
1165 1178 " (" SIZE_FORMAT "%%)",
1166 1179 top, _icms_stop_limit,
1167 1180 percent_of_space(space, _icms_stop_limit));
1168 1181 }
1169 1182 ConcurrentMarkSweepThread::start_icms();
1170 1183 assert(top < _icms_stop_limit, "Tautology");
1171 1184 if (word_size < pointer_delta(_icms_stop_limit, top)) {
1172 1185 return _icms_stop_limit;
1173 1186 }
1174 1187
1175 1188 // The allocation will cross both the _start and _stop limits, so do the
1176 1189 // stop notification also and return end().
1177 1190 if (CMSTraceIncrementalMode) {
1178 1191 space->print_on(gclog_or_tty);
1179 1192 gclog_or_tty->stamp();
1180 1193 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
1181 1194 ", new limit=" PTR_FORMAT
1182 1195 " (" SIZE_FORMAT "%%)",
1183 1196 top, space->end(),
1184 1197 percent_of_space(space, space->end()));
1185 1198 }
1186 1199 ConcurrentMarkSweepThread::stop_icms();
1187 1200 return space->end();
1188 1201 }
1189 1202
1190 1203 if (top <= _icms_stop_limit) {
1191 1204 if (CMSTraceIncrementalMode) {
1192 1205 space->print_on(gclog_or_tty);
1193 1206 gclog_or_tty->stamp();
1194 1207 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
1195 1208 ", new limit=" PTR_FORMAT
1196 1209 " (" SIZE_FORMAT "%%)",
1197 1210 top, space->end(),
1198 1211 percent_of_space(space, space->end()));
1199 1212 }
1200 1213 ConcurrentMarkSweepThread::stop_icms();
1201 1214 return space->end();
1202 1215 }
1203 1216
1204 1217 if (CMSTraceIncrementalMode) {
1205 1218 space->print_on(gclog_or_tty);
1206 1219 gclog_or_tty->stamp();
1207 1220 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
1208 1221 ", new limit=" PTR_FORMAT,
1209 1222 top, NULL);
1210 1223 }
1211 1224 }
1212 1225
1213 1226 return NULL;
1214 1227 }
1215 1228
1216 1229 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size, oop* ref) {
1217 1230 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1218 1231 // allocate, copy and if necessary update promoinfo --
1219 1232 // delegate to underlying space.
1220 1233 assert_lock_strong(freelistLock());
1221 1234
1222 1235 #ifndef PRODUCT
1223 1236 if (Universe::heap()->promotion_should_fail()) {
1224 1237 return NULL;
1225 1238 }
1226 1239 #endif // #ifndef PRODUCT
1227 1240
1228 1241 oop res = _cmsSpace->promote(obj, obj_size, ref);
1229 1242 if (res == NULL) {
1230 1243 // expand and retry
1231 1244 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
1232 1245 expand(s*HeapWordSize, MinHeapDeltaBytes,
1233 1246 CMSExpansionCause::_satisfy_promotion);
1234 1247 // Since there's currently no next generation, we don't try to promote
1235 1248 // into a more senior generation.
1236 1249 assert(next_gen() == NULL, "assumption, based upon which no attempt "
1237 1250 "is made to pass on a possibly failing "
1238 1251 "promotion to next generation");
1239 1252 res = _cmsSpace->promote(obj, obj_size, ref);
1240 1253 }
1241 1254 if (res != NULL) {
1242 1255 // See comment in allocate() about when objects should
1243 1256 // be allocated live.
1244 1257 assert(obj->is_oop(), "Will dereference klass pointer below");
1245 1258 collector()->promoted(false, // Not parallel
1246 1259 (HeapWord*)res, obj->is_objArray(), obj_size);
1247 1260 // promotion counters
1248 1261 NOT_PRODUCT(
1249 1262 _numObjectsPromoted++;
1250 1263 _numWordsPromoted +=
1251 1264 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1252 1265 )
1253 1266 }
1254 1267 return res;
1255 1268 }
1256 1269
1257 1270
1258 1271 HeapWord*
1259 1272 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
1260 1273 HeapWord* top,
1261 1274 size_t word_sz)
1262 1275 {
1263 1276 return collector()->allocation_limit_reached(space, top, word_sz);
1264 1277 }
1265 1278
1266 1279 // Things to support parallel young-gen collection.
1267 1280 oop
1268 1281 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1269 1282 oop old, markOop m,
1270 1283 size_t word_sz) {
1271 1284 #ifndef PRODUCT
1272 1285 if (Universe::heap()->promotion_should_fail()) {
1273 1286 return NULL;
1274 1287 }
1275 1288 #endif // #ifndef PRODUCT
1276 1289
1277 1290 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1278 1291 PromotionInfo* promoInfo = &ps->promo;
1279 1292 // if we are tracking promotions, then first ensure space for
1280 1293 // promotion (including spooling space for saving header if necessary).
1281 1294 // then allocate and copy, then track promoted info if needed.
1282 1295 // When tracking (see PromotionInfo::track()), the mark word may
1283 1296 // be displaced and in this case restoration of the mark word
1284 1297 // occurs in the (oop_since_save_marks_)iterate phase.
1285 1298 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1286 1299 // Out of space for allocating spooling buffers;
1287 1300 // try expanding and allocating spooling buffers.
1288 1301 if (!expand_and_ensure_spooling_space(promoInfo)) {
1289 1302 return NULL;
1290 1303 }
1291 1304 }
1292 1305 assert(promoInfo->has_spooling_space(), "Control point invariant");
1293 1306 HeapWord* obj_ptr = ps->lab.alloc(word_sz);
1294 1307 if (obj_ptr == NULL) {
1295 1308 obj_ptr = expand_and_par_lab_allocate(ps, word_sz);
1296 1309 if (obj_ptr == NULL) {
1297 1310 return NULL;
1298 1311 }
1299 1312 }
1300 1313 oop obj = oop(obj_ptr);
1301 1314 assert(obj->klass() == NULL, "Object should be uninitialized here.");
1302 1315 // Otherwise, copy the object. Here we must be careful to insert the
1303 1316 // klass pointer last, since this marks the block as an allocated object.
1304 1317 HeapWord* old_ptr = (HeapWord*)old;
1305 1318 if (word_sz > (size_t)oopDesc::header_size()) {
1306 1319 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1307 1320 obj_ptr + oopDesc::header_size(),
1308 1321 word_sz - oopDesc::header_size());
1309 1322 }
1310 1323 // Restore the mark word copied above.
1311 1324 obj->set_mark(m);
1312 1325 // Now we can track the promoted object, if necessary. We take care
1313 1326 // To delay the transition from uninitialized to full object
1314 1327 // (i.e., insertion of klass pointer) until after, so that it
1315 1328 // atomically becomes a promoted object.
1316 1329 if (promoInfo->tracking()) {
1317 1330 promoInfo->track((PromotedObject*)obj, old->klass());
1318 1331 }
1319 1332 // Finally, install the klass pointer.
1320 1333 obj->set_klass(old->klass());
1321 1334
1322 1335 assert(old->is_oop(), "Will dereference klass ptr below");
1323 1336 collector()->promoted(true, // parallel
1324 1337 obj_ptr, old->is_objArray(), word_sz);
1325 1338
1326 1339 NOT_PRODUCT(
1327 1340 Atomic::inc(&_numObjectsPromoted);
1328 1341 Atomic::add((jint)CompactibleFreeListSpace::adjustObjectSize(obj->size()),
1329 1342 &_numWordsPromoted);
1330 1343 )
1331 1344
1332 1345 return obj;
1333 1346 }
1334 1347
1335 1348 void
1336 1349 ConcurrentMarkSweepGeneration::
1337 1350 par_promote_alloc_undo(int thread_num,
1338 1351 HeapWord* obj, size_t word_sz) {
1339 1352 // CMS does not support promotion undo.
1340 1353 ShouldNotReachHere();
1341 1354 }
1342 1355
1343 1356 void
1344 1357 ConcurrentMarkSweepGeneration::
1345 1358 par_promote_alloc_done(int thread_num) {
1346 1359 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1347 1360 ps->lab.retire();
1348 1361 #if CFLS_LAB_REFILL_STATS
1349 1362 if (thread_num == 0) {
1350 1363 _cmsSpace->print_par_alloc_stats();
1351 1364 }
1352 1365 #endif
1353 1366 }
1354 1367
1355 1368 void
1356 1369 ConcurrentMarkSweepGeneration::
1357 1370 par_oop_since_save_marks_iterate_done(int thread_num) {
1358 1371 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1359 1372 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1360 1373 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1361 1374 }
1362 1375
1363 1376 // XXXPERM
1364 1377 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1365 1378 size_t size,
1366 1379 bool tlab)
1367 1380 {
1368 1381 // We allow a STW collection only if a full
1369 1382 // collection was requested.
1370 1383 return full || should_allocate(size, tlab); // FIX ME !!!
1371 1384 // This and promotion failure handling are connected at the
1372 1385 // hip and should be fixed by untying them.
1373 1386 }
1374 1387
1375 1388 bool CMSCollector::shouldConcurrentCollect() {
1376 1389 if (_full_gc_requested) {
1377 1390 assert(ExplicitGCInvokesConcurrent, "Unexpected state");
1378 1391 if (Verbose && PrintGCDetails) {
1379 1392 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1380 1393 " gc request");
1381 1394 }
1382 1395 return true;
1383 1396 }
1384 1397
1385 1398 // For debugging purposes, change the type of collection.
1386 1399 // If the rotation is not on the concurrent collection
1387 1400 // type, don't start a concurrent collection.
1388 1401 NOT_PRODUCT(
1389 1402 if (RotateCMSCollectionTypes &&
1390 1403 (_cmsGen->debug_collection_type() !=
1391 1404 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
1392 1405 assert(_cmsGen->debug_collection_type() !=
1393 1406 ConcurrentMarkSweepGeneration::Unknown_collection_type,
1394 1407 "Bad cms collection type");
1395 1408 return false;
1396 1409 }
1397 1410 )
1398 1411
1399 1412 FreelistLocker x(this);
1400 1413 // ------------------------------------------------------------------
1401 1414 // Print out lots of information which affects the initiation of
1402 1415 // a collection.
1403 1416 if (PrintCMSInitiationStatistics && stats().valid()) {
1404 1417 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1405 1418 gclog_or_tty->stamp();
↓ open down ↓ |
731 lines elided |
↑ open up ↑ |
1406 1419 gclog_or_tty->print_cr("");
1407 1420 stats().print_on(gclog_or_tty);
1408 1421 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1409 1422 stats().time_until_cms_gen_full());
1410 1423 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1411 1424 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1412 1425 _cmsGen->contiguous_available());
1413 1426 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1414 1427 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1415 1428 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1416 - gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", initiatingOccupancy());
1429 + gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1430 + gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy());
1417 1431 }
1418 1432 // ------------------------------------------------------------------
1419 1433
1420 1434 // If the estimated time to complete a cms collection (cms_duration())
1421 1435 // is less than the estimated time remaining until the cms generation
1422 1436 // is full, start a collection.
1423 1437 if (!UseCMSInitiatingOccupancyOnly) {
1424 1438 if (stats().valid()) {
1425 1439 if (stats().time_until_cms_start() == 0.0) {
1426 1440 return true;
1427 1441 }
1428 1442 } else {
1429 1443 // We want to conservatively collect somewhat early in order
1430 1444 // to try and "bootstrap" our CMS/promotion statistics;
1431 1445 // this branch will not fire after the first successful CMS
1432 1446 // collection because the stats should then be valid.
1433 1447 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1434 1448 if (Verbose && PrintGCDetails) {
1435 1449 gclog_or_tty->print_cr(
1436 1450 " CMSCollector: collect for bootstrapping statistics:"
1437 1451 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1438 1452 _bootstrap_occupancy);
↓ open down ↓ |
12 lines elided |
↑ open up ↑ |
1439 1453 }
1440 1454 return true;
1441 1455 }
1442 1456 }
1443 1457 }
1444 1458
1445 1459 // Otherwise, we start a collection cycle if either the perm gen or
1446 1460 // old gen want a collection cycle started. Each may use
1447 1461 // an appropriate criterion for making this decision.
1448 1462 // XXX We need to make sure that the gen expansion
1449 - // criterion dovetails well with this.
1450 - if (_cmsGen->shouldConcurrentCollect(initiatingOccupancy())) {
1463 + // criterion dovetails well with this. XXX NEED TO FIX THIS
1464 + if (_cmsGen->should_concurrent_collect()) {
1451 1465 if (Verbose && PrintGCDetails) {
1452 1466 gclog_or_tty->print_cr("CMS old gen initiated");
1453 1467 }
1454 1468 return true;
1455 1469 }
1456 1470
1457 - if (cms_should_unload_classes() &&
1458 - _permGen->shouldConcurrentCollect(initiatingOccupancy())) {
1459 - if (Verbose && PrintGCDetails) {
1460 - gclog_or_tty->print_cr("CMS perm gen initiated");
1471 + // We start a collection if we believe an incremental collection may fail;
1472 + // this is not likely to be productive in practice because it's probably too
1473 + // late anyway.
1474 + GenCollectedHeap* gch = GenCollectedHeap::heap();
1475 + assert(gch->collector_policy()->is_two_generation_policy(),
1476 + "You may want to check the correctness of the following");
1477 + if (gch->incremental_collection_will_fail()) {
1478 + if (PrintGCDetails && Verbose) {
1479 + gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1461 1480 }
1462 1481 return true;
1463 1482 }
1464 1483
1484 + if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) {
1485 + bool res = update_should_unload_classes();
1486 + if (res) {
1487 + if (Verbose && PrintGCDetails) {
1488 + gclog_or_tty->print_cr("CMS perm gen initiated");
1489 + }
1490 + return true;
1491 + }
1492 + }
1465 1493 return false;
1466 1494 }
1467 1495
1468 1496 // Clear _expansion_cause fields of constituent generations
1469 1497 void CMSCollector::clear_expansion_cause() {
1470 1498 _cmsGen->clear_expansion_cause();
1471 1499 _permGen->clear_expansion_cause();
1472 1500 }
1473 1501
1474 -bool ConcurrentMarkSweepGeneration::shouldConcurrentCollect(
1475 - double initiatingOccupancy) {
1476 - // We should be conservative in starting a collection cycle. To
1477 - // start too eagerly runs the risk of collecting too often in the
1478 - // extreme. To collect too rarely falls back on full collections,
1479 - // which works, even if not optimum in terms of concurrent work.
1480 - // As a work around for too eagerly collecting, use the flag
1481 - // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1482 - // giving the user an easily understandable way of controlling the
1483 - // collections.
1484 - // We want to start a new collection cycle if any of the following
1485 - // conditions hold:
1486 - // . our current occupancy exceeds the initiating occupancy, or
1487 - // . we recently needed to expand and have not since that expansion,
1488 - // collected, or
1489 - // . we are not using adaptive free lists and linear allocation is
1490 - // going to fail, or
1491 - // . (for old gen) incremental collection has already failed or
1492 - // may soon fail in the near future as we may not be able to absorb
1493 - // promotions.
1494 - assert_lock_strong(freelistLock());
1502 +// We should be conservative in starting a collection cycle. To
1503 +// start too eagerly runs the risk of collecting too often in the
1504 +// extreme. To collect too rarely falls back on full collections,
1505 +// which works, even if not optimum in terms of concurrent work.
1506 +// As a work around for too eagerly collecting, use the flag
1507 +// UseCMSInitiatingOccupancyOnly. This also has the advantage of
1508 +// giving the user an easily understandable way of controlling the
1509 +// collections.
1510 +// We want to start a new collection cycle if any of the following
1511 +// conditions hold:
1512 +// . our current occupancy exceeds the configured initiating occupancy
1513 +// for this generation, or
1514 +// . we recently needed to expand this space and have not, since that
1515 +// expansion, done a collection of this generation, or
1516 +// . the underlying space believes that it may be a good idea to initiate
1517 +// a concurrent collection (this may be based on criteria such as the
1518 +// following: the space uses linear allocation and linear allocation is
1519 +// going to fail, or there is believed to be excessive fragmentation in
1520 +// the generation, etc... or ...
1521 +// [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1522 +// the case of the old generation, not the perm generation; see CR 6543076):
1523 +// we may be approaching a point at which allocation requests may fail because
1524 +// we will be out of sufficient free space given allocation rate estimates.]
1525 +bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1495 1526
1496 - if (occupancy() > initiatingOccupancy) {
1527 + assert_lock_strong(freelistLock());
1528 + if (occupancy() > initiating_occupancy()) {
1497 1529 if (PrintGCDetails && Verbose) {
1498 1530 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1499 - short_name(), occupancy(), initiatingOccupancy);
1531 + short_name(), occupancy(), initiating_occupancy());
1500 1532 }
1501 1533 return true;
1502 1534 }
1503 1535 if (UseCMSInitiatingOccupancyOnly) {
1504 1536 return false;
1505 1537 }
1506 1538 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1507 1539 if (PrintGCDetails && Verbose) {
1508 1540 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1509 1541 short_name());
1510 1542 }
1511 1543 return true;
1512 1544 }
1513 - GenCollectedHeap* gch = GenCollectedHeap::heap();
1514 - assert(gch->collector_policy()->is_two_generation_policy(),
1515 - "You may want to check the correctness of the following");
1516 - if (gch->incremental_collection_will_fail()) {
1545 + if (_cmsSpace->should_concurrent_collect()) {
1517 1546 if (PrintGCDetails && Verbose) {
1518 - gclog_or_tty->print(" %s: collect because incremental collection will fail ",
1547 + gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1519 1548 short_name());
1520 1549 }
1521 1550 return true;
1522 1551 }
1523 - if (!_cmsSpace->adaptive_freelists() &&
1524 - _cmsSpace->linearAllocationWouldFail()) {
1525 - if (PrintGCDetails && Verbose) {
1526 - gclog_or_tty->print(" %s: collect because of linAB ",
1527 - short_name());
1528 - }
1529 - return true;
1530 - }
1531 1552 return false;
1532 1553 }
1533 1554
1534 1555 void ConcurrentMarkSweepGeneration::collect(bool full,
1535 1556 bool clear_all_soft_refs,
1536 1557 size_t size,
1537 1558 bool tlab)
1538 1559 {
1539 1560 collector()->collect(full, clear_all_soft_refs, size, tlab);
1540 1561 }
1541 1562
1542 1563 void CMSCollector::collect(bool full,
1543 1564 bool clear_all_soft_refs,
1544 1565 size_t size,
1545 1566 bool tlab)
1546 1567 {
1547 1568 if (!UseCMSCollectionPassing && _collectorState > Idling) {
1548 1569 // For debugging purposes skip the collection if the state
1549 1570 // is not currently idle
1550 1571 if (TraceCMSState) {
1551 1572 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
1552 1573 Thread::current(), full, _collectorState);
1553 1574 }
1554 1575 return;
1555 1576 }
1556 1577
1557 1578 // The following "if" branch is present for defensive reasons.
1558 1579 // In the current uses of this interface, it can be replaced with:
1559 1580 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1560 1581 // But I am not placing that assert here to allow future
1561 1582 // generality in invoking this interface.
1562 1583 if (GC_locker::is_active()) {
1563 1584 // A consistency test for GC_locker
1564 1585 assert(GC_locker::needs_gc(), "Should have been set already");
1565 1586 // Skip this foreground collection, instead
1566 1587 // expanding the heap if necessary.
1567 1588 // Need the free list locks for the call to free() in compute_new_size()
1568 1589 compute_new_size();
1569 1590 return;
1570 1591 }
1571 1592 acquire_control_and_collect(full, clear_all_soft_refs);
1572 1593 _full_gcs_since_conc_gc++;
1573 1594
1574 1595 }
1575 1596
1576 1597 void CMSCollector::request_full_gc(unsigned int full_gc_count) {
1577 1598 GenCollectedHeap* gch = GenCollectedHeap::heap();
1578 1599 unsigned int gc_count = gch->total_full_collections();
1579 1600 if (gc_count == full_gc_count) {
1580 1601 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1581 1602 _full_gc_requested = true;
1582 1603 CGC_lock->notify(); // nudge CMS thread
1583 1604 }
1584 1605 }
1585 1606
1586 1607
1587 1608 // The foreground and background collectors need to coordinate in order
1588 1609 // to make sure that they do not mutually interfere with CMS collections.
1589 1610 // When a background collection is active,
1590 1611 // the foreground collector may need to take over (preempt) and
1591 1612 // synchronously complete an ongoing collection. Depending on the
1592 1613 // frequency of the background collections and the heap usage
1593 1614 // of the application, this preemption can be seldom or frequent.
1594 1615 // There are only certain
1595 1616 // points in the background collection that the "collection-baton"
1596 1617 // can be passed to the foreground collector.
1597 1618 //
1598 1619 // The foreground collector will wait for the baton before
1599 1620 // starting any part of the collection. The foreground collector
1600 1621 // will only wait at one location.
1601 1622 //
1602 1623 // The background collector will yield the baton before starting a new
1603 1624 // phase of the collection (e.g., before initial marking, marking from roots,
1604 1625 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1605 1626 // of the loop which switches the phases. The background collector does some
1606 1627 // of the phases (initial mark, final re-mark) with the world stopped.
1607 1628 // Because of locking involved in stopping the world,
1608 1629 // the foreground collector should not block waiting for the background
1609 1630 // collector when it is doing a stop-the-world phase. The background
1610 1631 // collector will yield the baton at an additional point just before
1611 1632 // it enters a stop-the-world phase. Once the world is stopped, the
1612 1633 // background collector checks the phase of the collection. If the
1613 1634 // phase has not changed, it proceeds with the collection. If the
1614 1635 // phase has changed, it skips that phase of the collection. See
1615 1636 // the comments on the use of the Heap_lock in collect_in_background().
1616 1637 //
1617 1638 // Variable used in baton passing.
1618 1639 // _foregroundGCIsActive - Set to true by the foreground collector when
1619 1640 // it wants the baton. The foreground clears it when it has finished
1620 1641 // the collection.
1621 1642 // _foregroundGCShouldWait - Set to true by the background collector
1622 1643 // when it is running. The foreground collector waits while
1623 1644 // _foregroundGCShouldWait is true.
1624 1645 // CGC_lock - monitor used to protect access to the above variables
1625 1646 // and to notify the foreground and background collectors.
1626 1647 // _collectorState - current state of the CMS collection.
1627 1648 //
1628 1649 // The foreground collector
1629 1650 // acquires the CGC_lock
1630 1651 // sets _foregroundGCIsActive
1631 1652 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1632 1653 // various locks acquired in preparation for the collection
1633 1654 // are released so as not to block the background collector
1634 1655 // that is in the midst of a collection
1635 1656 // proceeds with the collection
1636 1657 // clears _foregroundGCIsActive
1637 1658 // returns
1638 1659 //
1639 1660 // The background collector in a loop iterating on the phases of the
1640 1661 // collection
1641 1662 // acquires the CGC_lock
1642 1663 // sets _foregroundGCShouldWait
1643 1664 // if _foregroundGCIsActive is set
1644 1665 // clears _foregroundGCShouldWait, notifies _CGC_lock
1645 1666 // waits on _CGC_lock for _foregroundGCIsActive to become false
1646 1667 // and exits the loop.
1647 1668 // otherwise
1648 1669 // proceed with that phase of the collection
1649 1670 // if the phase is a stop-the-world phase,
1650 1671 // yield the baton once more just before enqueueing
1651 1672 // the stop-world CMS operation (executed by the VM thread).
1652 1673 // returns after all phases of the collection are done
1653 1674 //
1654 1675
1655 1676 void CMSCollector::acquire_control_and_collect(bool full,
1656 1677 bool clear_all_soft_refs) {
1657 1678 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1658 1679 assert(!Thread::current()->is_ConcurrentGC_thread(),
1659 1680 "shouldn't try to acquire control from self!");
1660 1681
1661 1682 // Start the protocol for acquiring control of the
1662 1683 // collection from the background collector (aka CMS thread).
1663 1684 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1664 1685 "VM thread should have CMS token");
1665 1686 // Remember the possibly interrupted state of an ongoing
1666 1687 // concurrent collection
1667 1688 CollectorState first_state = _collectorState;
1668 1689
1669 1690 // Signal to a possibly ongoing concurrent collection that
1670 1691 // we want to do a foreground collection.
1671 1692 _foregroundGCIsActive = true;
1672 1693
1673 1694 // Disable incremental mode during a foreground collection.
1674 1695 ICMSDisabler icms_disabler;
1675 1696
1676 1697 // release locks and wait for a notify from the background collector
1677 1698 // releasing the locks in only necessary for phases which
1678 1699 // do yields to improve the granularity of the collection.
1679 1700 assert_lock_strong(bitMapLock());
1680 1701 // We need to lock the Free list lock for the space that we are
1681 1702 // currently collecting.
1682 1703 assert(haveFreelistLocks(), "Must be holding free list locks");
1683 1704 bitMapLock()->unlock();
1684 1705 releaseFreelistLocks();
1685 1706 {
1686 1707 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1687 1708 if (_foregroundGCShouldWait) {
1688 1709 // We are going to be waiting for action for the CMS thread;
1689 1710 // it had better not be gone (for instance at shutdown)!
1690 1711 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1691 1712 "CMS thread must be running");
1692 1713 // Wait here until the background collector gives us the go-ahead
1693 1714 ConcurrentMarkSweepThread::clear_CMS_flag(
1694 1715 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1695 1716 // Get a possibly blocked CMS thread going:
1696 1717 // Note that we set _foregroundGCIsActive true above,
1697 1718 // without protection of the CGC_lock.
1698 1719 CGC_lock->notify();
1699 1720 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1700 1721 "Possible deadlock");
1701 1722 while (_foregroundGCShouldWait) {
1702 1723 // wait for notification
1703 1724 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1704 1725 // Possibility of delay/starvation here, since CMS token does
1705 1726 // not know to give priority to VM thread? Actually, i think
1706 1727 // there wouldn't be any delay/starvation, but the proof of
1707 1728 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1708 1729 }
1709 1730 ConcurrentMarkSweepThread::set_CMS_flag(
1710 1731 ConcurrentMarkSweepThread::CMS_vm_has_token);
1711 1732 }
1712 1733 }
1713 1734 // The CMS_token is already held. Get back the other locks.
1714 1735 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1715 1736 "VM thread should have CMS token");
1716 1737 getFreelistLocks();
1717 1738 bitMapLock()->lock_without_safepoint_check();
1718 1739 if (TraceCMSState) {
1719 1740 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1720 1741 INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1721 1742 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1722 1743 }
1723 1744
1724 1745 // Check if we need to do a compaction, or if not, whether
1725 1746 // we need to start the mark-sweep from scratch.
1726 1747 bool should_compact = false;
1727 1748 bool should_start_over = false;
1728 1749 decide_foreground_collection_type(clear_all_soft_refs,
1729 1750 &should_compact, &should_start_over);
1730 1751
1731 1752 NOT_PRODUCT(
1732 1753 if (RotateCMSCollectionTypes) {
1733 1754 if (_cmsGen->debug_collection_type() ==
1734 1755 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1735 1756 should_compact = true;
1736 1757 } else if (_cmsGen->debug_collection_type() ==
1737 1758 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1738 1759 should_compact = false;
1739 1760 }
1740 1761 }
1741 1762 )
1742 1763
1743 1764 if (PrintGCDetails && first_state > Idling) {
1744 1765 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1745 1766 if (GCCause::is_user_requested_gc(cause) ||
1746 1767 GCCause::is_serviceability_requested_gc(cause)) {
1747 1768 gclog_or_tty->print(" (concurrent mode interrupted)");
1748 1769 } else {
1749 1770 gclog_or_tty->print(" (concurrent mode failure)");
1750 1771 }
1751 1772 }
1752 1773
1753 1774 if (should_compact) {
1754 1775 // If the collection is being acquired from the background
1755 1776 // collector, there may be references on the discovered
1756 1777 // references lists that have NULL referents (being those
1757 1778 // that were concurrently cleared by a mutator) or
1758 1779 // that are no longer active (having been enqueued concurrently
1759 1780 // by the mutator).
1760 1781 // Scrub the list of those references because Mark-Sweep-Compact
1761 1782 // code assumes referents are not NULL and that all discovered
1762 1783 // Reference objects are active.
1763 1784 ref_processor()->clean_up_discovered_references();
1764 1785
1765 1786 do_compaction_work(clear_all_soft_refs);
1766 1787
1767 1788 // Has the GC time limit been exceeded?
1768 1789 check_gc_time_limit();
1769 1790
1770 1791 } else {
1771 1792 do_mark_sweep_work(clear_all_soft_refs, first_state,
1772 1793 should_start_over);
1773 1794 }
1774 1795 // Reset the expansion cause, now that we just completed
1775 1796 // a collection cycle.
1776 1797 clear_expansion_cause();
1777 1798 _foregroundGCIsActive = false;
1778 1799 return;
1779 1800 }
1780 1801
1781 1802 void CMSCollector::check_gc_time_limit() {
1782 1803
1783 1804 // Ignore explicit GC's. Exiting here does not set the flag and
1784 1805 // does not reset the count. Updating of the averages for system
1785 1806 // GC's is still controlled by UseAdaptiveSizePolicyWithSystemGC.
1786 1807 GCCause::Cause gc_cause = GenCollectedHeap::heap()->gc_cause();
1787 1808 if (GCCause::is_user_requested_gc(gc_cause) ||
1788 1809 GCCause::is_serviceability_requested_gc(gc_cause)) {
1789 1810 return;
1790 1811 }
1791 1812
1792 1813 // Calculate the fraction of the CMS generation was freed during
1793 1814 // the last collection.
1794 1815 // Only consider the STW compacting cost for now.
1795 1816 //
1796 1817 // Note that the gc time limit test only works for the collections
1797 1818 // of the young gen + tenured gen and not for collections of the
1798 1819 // permanent gen. That is because the calculation of the space
1799 1820 // freed by the collection is the free space in the young gen +
1800 1821 // tenured gen.
1801 1822
1802 1823 double fraction_free =
1803 1824 ((double)_cmsGen->free())/((double)_cmsGen->max_capacity());
1804 1825 if ((100.0 * size_policy()->compacting_gc_cost()) >
1805 1826 ((double) GCTimeLimit) &&
1806 1827 ((fraction_free * 100) < GCHeapFreeLimit)) {
1807 1828 size_policy()->inc_gc_time_limit_count();
1808 1829 if (UseGCOverheadLimit &&
1809 1830 (size_policy()->gc_time_limit_count() >
1810 1831 AdaptiveSizePolicyGCTimeLimitThreshold)) {
1811 1832 size_policy()->set_gc_time_limit_exceeded(true);
1812 1833 // Avoid consecutive OOM due to the gc time limit by resetting
1813 1834 // the counter.
1814 1835 size_policy()->reset_gc_time_limit_count();
1815 1836 if (PrintGCDetails) {
1816 1837 gclog_or_tty->print_cr(" GC is exceeding overhead limit "
1817 1838 "of %d%%", GCTimeLimit);
1818 1839 }
1819 1840 } else {
1820 1841 if (PrintGCDetails) {
1821 1842 gclog_or_tty->print_cr(" GC would exceed overhead limit "
1822 1843 "of %d%%", GCTimeLimit);
1823 1844 }
1824 1845 }
1825 1846 } else {
1826 1847 size_policy()->reset_gc_time_limit_count();
1827 1848 }
1828 1849 }
1829 1850
1830 1851 // Resize the perm generation and the tenured generation
1831 1852 // after obtaining the free list locks for the
1832 1853 // two generations.
1833 1854 void CMSCollector::compute_new_size() {
1834 1855 assert_locked_or_safepoint(Heap_lock);
1835 1856 FreelistLocker z(this);
1836 1857 _permGen->compute_new_size();
1837 1858 _cmsGen->compute_new_size();
1838 1859 }
1839 1860
1840 1861 // A work method used by foreground collection to determine
1841 1862 // what type of collection (compacting or not, continuing or fresh)
1842 1863 // it should do.
1843 1864 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1844 1865 // and CMSCompactWhenClearAllSoftRefs the default in the future
1845 1866 // and do away with the flags after a suitable period.
1846 1867 void CMSCollector::decide_foreground_collection_type(
1847 1868 bool clear_all_soft_refs, bool* should_compact,
1848 1869 bool* should_start_over) {
1849 1870 // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1850 1871 // flag is set, and we have either requested a System.gc() or
1851 1872 // the number of full gc's since the last concurrent cycle
1852 1873 // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1853 1874 // or if an incremental collection has failed
1854 1875 GenCollectedHeap* gch = GenCollectedHeap::heap();
1855 1876 assert(gch->collector_policy()->is_two_generation_policy(),
1856 1877 "You may want to check the correctness of the following");
1857 1878 // Inform cms gen if this was due to partial collection failing.
1858 1879 // The CMS gen may use this fact to determine its expansion policy.
1859 1880 if (gch->incremental_collection_will_fail()) {
1860 1881 assert(!_cmsGen->incremental_collection_failed(),
1861 1882 "Should have been noticed, reacted to and cleared");
1862 1883 _cmsGen->set_incremental_collection_failed();
1863 1884 }
1864 1885 *should_compact =
1865 1886 UseCMSCompactAtFullCollection &&
1866 1887 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1867 1888 GCCause::is_user_requested_gc(gch->gc_cause()) ||
1868 1889 gch->incremental_collection_will_fail());
1869 1890 *should_start_over = false;
1870 1891 if (clear_all_soft_refs && !*should_compact) {
1871 1892 // We are about to do a last ditch collection attempt
1872 1893 // so it would normally make sense to do a compaction
1873 1894 // to reclaim as much space as possible.
1874 1895 if (CMSCompactWhenClearAllSoftRefs) {
1875 1896 // Default: The rationale is that in this case either
1876 1897 // we are past the final marking phase, in which case
1877 1898 // we'd have to start over, or so little has been done
1878 1899 // that there's little point in saving that work. Compaction
1879 1900 // appears to be the sensible choice in either case.
1880 1901 *should_compact = true;
1881 1902 } else {
1882 1903 // We have been asked to clear all soft refs, but not to
1883 1904 // compact. Make sure that we aren't past the final checkpoint
1884 1905 // phase, for that is where we process soft refs. If we are already
1885 1906 // past that phase, we'll need to redo the refs discovery phase and
1886 1907 // if necessary clear soft refs that weren't previously
1887 1908 // cleared. We do so by remembering the phase in which
1888 1909 // we came in, and if we are past the refs processing
1889 1910 // phase, we'll choose to just redo the mark-sweep
1890 1911 // collection from scratch.
1891 1912 if (_collectorState > FinalMarking) {
1892 1913 // We are past the refs processing phase;
1893 1914 // start over and do a fresh synchronous CMS cycle
1894 1915 _collectorState = Resetting; // skip to reset to start new cycle
1895 1916 reset(false /* == !asynch */);
1896 1917 *should_start_over = true;
1897 1918 } // else we can continue a possibly ongoing current cycle
1898 1919 }
1899 1920 }
1900 1921 }
1901 1922
1902 1923 // A work method used by the foreground collector to do
1903 1924 // a mark-sweep-compact.
1904 1925 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1905 1926 GenCollectedHeap* gch = GenCollectedHeap::heap();
1906 1927 TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty);
1907 1928 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
1908 1929 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
1909 1930 "collections passed to foreground collector", _full_gcs_since_conc_gc);
1910 1931 }
1911 1932
1912 1933 // Sample collection interval time and reset for collection pause.
1913 1934 if (UseAdaptiveSizePolicy) {
1914 1935 size_policy()->msc_collection_begin();
1915 1936 }
1916 1937
1917 1938 // Temporarily widen the span of the weak reference processing to
1918 1939 // the entire heap.
1919 1940 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1920 1941 ReferenceProcessorSpanMutator x(ref_processor(), new_span);
1921 1942
1922 1943 // Temporarily, clear the "is_alive_non_header" field of the
1923 1944 // reference processor.
1924 1945 ReferenceProcessorIsAliveMutator y(ref_processor(), NULL);
1925 1946
1926 1947 // Temporarily make reference _processing_ single threaded (non-MT).
1927 1948 ReferenceProcessorMTProcMutator z(ref_processor(), false);
1928 1949
1929 1950 // Temporarily make refs discovery atomic
1930 1951 ReferenceProcessorAtomicMutator w(ref_processor(), true);
1931 1952
1932 1953 ref_processor()->set_enqueuing_is_done(false);
1933 1954 ref_processor()->enable_discovery();
1934 1955 // If an asynchronous collection finishes, the _modUnionTable is
1935 1956 // all clear. If we are assuming the collection from an asynchronous
1936 1957 // collection, clear the _modUnionTable.
1937 1958 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1938 1959 "_modUnionTable should be clear if the baton was not passed");
1939 1960 _modUnionTable.clear_all();
1940 1961
1941 1962 // We must adjust the allocation statistics being maintained
1942 1963 // in the free list space. We do so by reading and clearing
1943 1964 // the sweep timer and updating the block flux rate estimates below.
1944 1965 assert(_sweep_timer.is_active(), "We should never see the timer inactive");
1945 1966 _sweep_timer.stop();
1946 1967 // Note that we do not use this sample to update the _sweep_estimate.
1947 1968 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()),
1948 1969 _sweep_estimate.padded_average());
1949 1970
1950 1971 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
1951 1972 ref_processor(), clear_all_soft_refs);
1952 1973 #ifdef ASSERT
1953 1974 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1954 1975 size_t free_size = cms_space->free();
1955 1976 assert(free_size ==
1956 1977 pointer_delta(cms_space->end(), cms_space->compaction_top())
1957 1978 * HeapWordSize,
1958 1979 "All the free space should be compacted into one chunk at top");
1959 1980 assert(cms_space->dictionary()->totalChunkSize(
1960 1981 debug_only(cms_space->freelistLock())) == 0 ||
1961 1982 cms_space->totalSizeInIndexedFreeLists() == 0,
1962 1983 "All the free space should be in a single chunk");
↓ open down ↓ |
422 lines elided |
↑ open up ↑ |
1963 1984 size_t num = cms_space->totalCount();
1964 1985 assert((free_size == 0 && num == 0) ||
1965 1986 (free_size > 0 && (num == 1 || num == 2)),
1966 1987 "There should be at most 2 free chunks after compaction");
1967 1988 #endif // ASSERT
1968 1989 _collectorState = Resetting;
1969 1990 assert(_restart_addr == NULL,
1970 1991 "Should have been NULL'd before baton was passed");
1971 1992 reset(false /* == !asynch */);
1972 1993 _cmsGen->reset_after_compaction();
1994 + _concurrent_cycles_since_last_unload = 0;
1973 1995
1974 - if (verifying() && !cms_should_unload_classes()) {
1996 + if (verifying() && !should_unload_classes()) {
1975 1997 perm_gen_verify_bit_map()->clear_all();
1976 1998 }
1977 1999
1978 2000 // Clear any data recorded in the PLAB chunk arrays.
1979 2001 if (_survivor_plab_array != NULL) {
1980 2002 reset_survivor_plab_arrays();
1981 2003 }
1982 2004
1983 2005 // Adjust the per-size allocation stats for the next epoch.
1984 2006 _cmsGen->cmsSpace()->endSweepFLCensus(sweepCount() /* fake */);
1985 2007 // Restart the "sweep timer" for next epoch.
1986 2008 _sweep_timer.reset();
1987 2009 _sweep_timer.start();
1988 2010
1989 2011 // Sample collection pause time and reset for collection interval.
1990 2012 if (UseAdaptiveSizePolicy) {
1991 2013 size_policy()->msc_collection_end(gch->gc_cause());
1992 2014 }
1993 2015
1994 2016 // For a mark-sweep-compact, compute_new_size() will be called
1995 2017 // in the heap's do_collection() method.
1996 2018 }
1997 2019
1998 2020 // A work method used by the foreground collector to do
1999 2021 // a mark-sweep, after taking over from a possibly on-going
2000 2022 // concurrent mark-sweep collection.
2001 2023 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2002 2024 CollectorState first_state, bool should_start_over) {
2003 2025 if (PrintGC && Verbose) {
2004 2026 gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2005 2027 "collector with count %d",
2006 2028 _full_gcs_since_conc_gc);
2007 2029 }
2008 2030 switch (_collectorState) {
2009 2031 case Idling:
2010 2032 if (first_state == Idling || should_start_over) {
2011 2033 // The background GC was not active, or should
2012 2034 // restarted from scratch; start the cycle.
2013 2035 _collectorState = InitialMarking;
2014 2036 }
2015 2037 // If first_state was not Idling, then a background GC
2016 2038 // was in progress and has now finished. No need to do it
2017 2039 // again. Leave the state as Idling.
2018 2040 break;
2019 2041 case Precleaning:
2020 2042 // In the foreground case don't do the precleaning since
2021 2043 // it is not done concurrently and there is extra work
2022 2044 // required.
2023 2045 _collectorState = FinalMarking;
2024 2046 }
2025 2047 if (PrintGCDetails &&
2026 2048 (_collectorState > Idling ||
2027 2049 !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) {
2028 2050 gclog_or_tty->print(" (concurrent mode failure)");
2029 2051 }
2030 2052 collect_in_foreground(clear_all_soft_refs);
2031 2053
2032 2054 // For a mark-sweep, compute_new_size() will be called
2033 2055 // in the heap's do_collection() method.
2034 2056 }
2035 2057
2036 2058
2037 2059 void CMSCollector::getFreelistLocks() const {
2038 2060 // Get locks for all free lists in all generations that this
2039 2061 // collector is responsible for
2040 2062 _cmsGen->freelistLock()->lock_without_safepoint_check();
2041 2063 _permGen->freelistLock()->lock_without_safepoint_check();
2042 2064 }
2043 2065
2044 2066 void CMSCollector::releaseFreelistLocks() const {
2045 2067 // Release locks for all free lists in all generations that this
2046 2068 // collector is responsible for
2047 2069 _cmsGen->freelistLock()->unlock();
2048 2070 _permGen->freelistLock()->unlock();
2049 2071 }
2050 2072
2051 2073 bool CMSCollector::haveFreelistLocks() const {
2052 2074 // Check locks for all free lists in all generations that this
2053 2075 // collector is responsible for
2054 2076 assert_lock_strong(_cmsGen->freelistLock());
2055 2077 assert_lock_strong(_permGen->freelistLock());
2056 2078 PRODUCT_ONLY(ShouldNotReachHere());
2057 2079 return true;
2058 2080 }
2059 2081
2060 2082 // A utility class that is used by the CMS collector to
2061 2083 // temporarily "release" the foreground collector from its
2062 2084 // usual obligation to wait for the background collector to
2063 2085 // complete an ongoing phase before proceeding.
2064 2086 class ReleaseForegroundGC: public StackObj {
2065 2087 private:
2066 2088 CMSCollector* _c;
2067 2089 public:
2068 2090 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2069 2091 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2070 2092 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2071 2093 // allow a potentially blocked foreground collector to proceed
2072 2094 _c->_foregroundGCShouldWait = false;
2073 2095 if (_c->_foregroundGCIsActive) {
2074 2096 CGC_lock->notify();
2075 2097 }
2076 2098 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2077 2099 "Possible deadlock");
2078 2100 }
2079 2101
2080 2102 ~ReleaseForegroundGC() {
2081 2103 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2082 2104 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2083 2105 _c->_foregroundGCShouldWait = true;
2084 2106 }
2085 2107 };
2086 2108
2087 2109 // There are separate collect_in_background and collect_in_foreground because of
2088 2110 // the different locking requirements of the background collector and the
2089 2111 // foreground collector. There was originally an attempt to share
2090 2112 // one "collect" method between the background collector and the foreground
↓ open down ↓ |
106 lines elided |
↑ open up ↑ |
2091 2113 // collector but the if-then-else required made it cleaner to have
2092 2114 // separate methods.
2093 2115 void CMSCollector::collect_in_background(bool clear_all_soft_refs) {
2094 2116 assert(Thread::current()->is_ConcurrentGC_thread(),
2095 2117 "A CMS asynchronous collection is only allowed on a CMS thread.");
2096 2118
2097 2119 GenCollectedHeap* gch = GenCollectedHeap::heap();
2098 2120 {
2099 2121 bool safepoint_check = Mutex::_no_safepoint_check_flag;
2100 2122 MutexLockerEx hl(Heap_lock, safepoint_check);
2123 + FreelistLocker fll(this);
2101 2124 MutexLockerEx x(CGC_lock, safepoint_check);
2102 2125 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2103 2126 // The foreground collector is active or we're
2104 2127 // not using asynchronous collections. Skip this
2105 2128 // background collection.
2106 2129 assert(!_foregroundGCShouldWait, "Should be clear");
2107 2130 return;
2108 2131 } else {
2109 2132 assert(_collectorState == Idling, "Should be idling before start.");
2110 2133 _collectorState = InitialMarking;
2111 2134 // Reset the expansion cause, now that we are about to begin
2112 2135 // a new cycle.
2113 2136 clear_expansion_cause();
2114 2137 }
2115 - _unloaded_classes_last_cycle = cms_should_unload_classes(); // ... from last cycle
2116 - // This controls class unloading in response to an explicit gc request.
2117 - // If ExplicitGCInvokesConcurrentAndUnloadsClasses is set, then
2118 - // we will unload classes even if CMSClassUnloadingEnabled is not set.
2119 - // See CR 6541037 and related CRs.
2120 - _unload_classes = _full_gc_requested // ... for this cycle
2121 - && ExplicitGCInvokesConcurrentAndUnloadsClasses;
2138 + // Decide if we want to enable class unloading as part of the
2139 + // ensuing concurrent GC cycle.
2140 + update_should_unload_classes();
2122 2141 _full_gc_requested = false; // acks all outstanding full gc requests
2123 2142 // Signal that we are about to start a collection
2124 2143 gch->increment_total_full_collections(); // ... starting a collection cycle
2125 2144 _collection_count_start = gch->total_full_collections();
2126 2145 }
2127 2146
2128 2147 // Used for PrintGC
2129 2148 size_t prev_used;
2130 2149 if (PrintGC && Verbose) {
2131 2150 prev_used = _cmsGen->used(); // XXXPERM
2132 2151 }
2133 2152
2134 2153 // The change of the collection state is normally done at this level;
2135 2154 // the exceptions are phases that are executed while the world is
2136 2155 // stopped. For those phases the change of state is done while the
2137 2156 // world is stopped. For baton passing purposes this allows the
2138 2157 // background collector to finish the phase and change state atomically.
2139 2158 // The foreground collector cannot wait on a phase that is done
2140 2159 // while the world is stopped because the foreground collector already
2141 2160 // has the world stopped and would deadlock.
2142 2161 while (_collectorState != Idling) {
2143 2162 if (TraceCMSState) {
2144 2163 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2145 2164 Thread::current(), _collectorState);
2146 2165 }
2147 2166 // The foreground collector
2148 2167 // holds the Heap_lock throughout its collection.
2149 2168 // holds the CMS token (but not the lock)
2150 2169 // except while it is waiting for the background collector to yield.
2151 2170 //
2152 2171 // The foreground collector should be blocked (not for long)
2153 2172 // if the background collector is about to start a phase
2154 2173 // executed with world stopped. If the background
2155 2174 // collector has already started such a phase, the
2156 2175 // foreground collector is blocked waiting for the
2157 2176 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
2158 2177 // are executed in the VM thread.
2159 2178 //
2160 2179 // The locking order is
2161 2180 // PendingListLock (PLL) -- if applicable (FinalMarking)
2162 2181 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
2163 2182 // CMS token (claimed in
2164 2183 // stop_world_and_do() -->
2165 2184 // safepoint_synchronize() -->
2166 2185 // CMSThread::synchronize())
2167 2186
2168 2187 {
2169 2188 // Check if the FG collector wants us to yield.
2170 2189 CMSTokenSync x(true); // is cms thread
2171 2190 if (waitForForegroundGC()) {
2172 2191 // We yielded to a foreground GC, nothing more to be
2173 2192 // done this round.
2174 2193 assert(_foregroundGCShouldWait == false, "We set it to false in "
2175 2194 "waitForForegroundGC()");
2176 2195 if (TraceCMSState) {
2177 2196 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2178 2197 " exiting collection CMS state %d",
2179 2198 Thread::current(), _collectorState);
2180 2199 }
2181 2200 return;
2182 2201 } else {
2183 2202 // The background collector can run but check to see if the
2184 2203 // foreground collector has done a collection while the
2185 2204 // background collector was waiting to get the CGC_lock
2186 2205 // above. If yes, break so that _foregroundGCShouldWait
2187 2206 // is cleared before returning.
2188 2207 if (_collectorState == Idling) {
2189 2208 break;
2190 2209 }
2191 2210 }
2192 2211 }
2193 2212
2194 2213 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2195 2214 "should be waiting");
2196 2215
2197 2216 switch (_collectorState) {
2198 2217 case InitialMarking:
2199 2218 {
2200 2219 ReleaseForegroundGC x(this);
2201 2220 stats().record_cms_begin();
2202 2221
2203 2222 VM_CMS_Initial_Mark initial_mark_op(this);
2204 2223 VMThread::execute(&initial_mark_op);
2205 2224 }
2206 2225 // The collector state may be any legal state at this point
2207 2226 // since the background collector may have yielded to the
2208 2227 // foreground collector.
2209 2228 break;
2210 2229 case Marking:
2211 2230 // initial marking in checkpointRootsInitialWork has been completed
2212 2231 if (markFromRoots(true)) { // we were successful
2213 2232 assert(_collectorState == Precleaning, "Collector state should "
2214 2233 "have changed");
2215 2234 } else {
2216 2235 assert(_foregroundGCIsActive, "Internal state inconsistency");
2217 2236 }
2218 2237 break;
2219 2238 case Precleaning:
2220 2239 if (UseAdaptiveSizePolicy) {
2221 2240 size_policy()->concurrent_precleaning_begin();
2222 2241 }
2223 2242 // marking from roots in markFromRoots has been completed
2224 2243 preclean();
2225 2244 if (UseAdaptiveSizePolicy) {
2226 2245 size_policy()->concurrent_precleaning_end();
2227 2246 }
2228 2247 assert(_collectorState == AbortablePreclean ||
2229 2248 _collectorState == FinalMarking,
2230 2249 "Collector state should have changed");
2231 2250 break;
2232 2251 case AbortablePreclean:
2233 2252 if (UseAdaptiveSizePolicy) {
2234 2253 size_policy()->concurrent_phases_resume();
2235 2254 }
2236 2255 abortable_preclean();
2237 2256 if (UseAdaptiveSizePolicy) {
2238 2257 size_policy()->concurrent_precleaning_end();
2239 2258 }
2240 2259 assert(_collectorState == FinalMarking, "Collector state should "
2241 2260 "have changed");
2242 2261 break;
2243 2262 case FinalMarking:
2244 2263 {
2245 2264 ReleaseForegroundGC x(this);
2246 2265
2247 2266 VM_CMS_Final_Remark final_remark_op(this);
2248 2267 VMThread::execute(&final_remark_op);
2249 2268 }
2250 2269 assert(_foregroundGCShouldWait, "block post-condition");
2251 2270 break;
2252 2271 case Sweeping:
2253 2272 if (UseAdaptiveSizePolicy) {
2254 2273 size_policy()->concurrent_sweeping_begin();
2255 2274 }
2256 2275 // final marking in checkpointRootsFinal has been completed
2257 2276 sweep(true);
2258 2277 assert(_collectorState == Resizing, "Collector state change "
2259 2278 "to Resizing must be done under the free_list_lock");
2260 2279 _full_gcs_since_conc_gc = 0;
2261 2280
2262 2281 // Stop the timers for adaptive size policy for the concurrent phases
2263 2282 if (UseAdaptiveSizePolicy) {
2264 2283 size_policy()->concurrent_sweeping_end();
2265 2284 size_policy()->concurrent_phases_end(gch->gc_cause(),
2266 2285 gch->prev_gen(_cmsGen)->capacity(),
2267 2286 _cmsGen->free());
2268 2287 }
2269 2288
2270 2289 case Resizing: {
2271 2290 // Sweeping has been completed...
2272 2291 // At this point the background collection has completed.
2273 2292 // Don't move the call to compute_new_size() down
2274 2293 // into code that might be executed if the background
2275 2294 // collection was preempted.
2276 2295 {
2277 2296 ReleaseForegroundGC x(this); // unblock FG collection
2278 2297 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
2279 2298 CMSTokenSync z(true); // not strictly needed.
2280 2299 if (_collectorState == Resizing) {
2281 2300 compute_new_size();
2282 2301 _collectorState = Resetting;
2283 2302 } else {
2284 2303 assert(_collectorState == Idling, "The state should only change"
2285 2304 " because the foreground collector has finished the collection");
2286 2305 }
2287 2306 }
2288 2307 break;
2289 2308 }
2290 2309 case Resetting:
2291 2310 // CMS heap resizing has been completed
2292 2311 reset(true);
2293 2312 assert(_collectorState == Idling, "Collector state should "
2294 2313 "have changed");
2295 2314 stats().record_cms_end();
2296 2315 // Don't move the concurrent_phases_end() and compute_new_size()
2297 2316 // calls to here because a preempted background collection
2298 2317 // has it's state set to "Resetting".
2299 2318 break;
2300 2319 case Idling:
2301 2320 default:
2302 2321 ShouldNotReachHere();
2303 2322 break;
2304 2323 }
2305 2324 if (TraceCMSState) {
2306 2325 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2307 2326 Thread::current(), _collectorState);
2308 2327 }
2309 2328 assert(_foregroundGCShouldWait, "block post-condition");
2310 2329 }
2311 2330
2312 2331 // Should this be in gc_epilogue?
2313 2332 collector_policy()->counters()->update_counters();
2314 2333
2315 2334 {
2316 2335 // Clear _foregroundGCShouldWait and, in the event that the
2317 2336 // foreground collector is waiting, notify it, before
2318 2337 // returning.
2319 2338 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2320 2339 _foregroundGCShouldWait = false;
2321 2340 if (_foregroundGCIsActive) {
2322 2341 CGC_lock->notify();
2323 2342 }
2324 2343 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2325 2344 "Possible deadlock");
2326 2345 }
2327 2346 if (TraceCMSState) {
2328 2347 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2329 2348 " exiting collection CMS state %d",
2330 2349 Thread::current(), _collectorState);
2331 2350 }
2332 2351 if (PrintGC && Verbose) {
2333 2352 _cmsGen->print_heap_change(prev_used);
2334 2353 }
2335 2354 }
2336 2355
2337 2356 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) {
2338 2357 assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2339 2358 "Foreground collector should be waiting, not executing");
2340 2359 assert(Thread::current()->is_VM_thread(), "A foreground collection"
2341 2360 "may only be done by the VM Thread with the world stopped");
2342 2361 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2343 2362 "VM thread should have CMS token");
2344 2363
2345 2364 NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2346 2365 true, gclog_or_tty);)
2347 2366 if (UseAdaptiveSizePolicy) {
2348 2367 size_policy()->ms_collection_begin();
2349 2368 }
2350 2369 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2351 2370
2352 2371 HandleMark hm; // Discard invalid handles created during verification
2353 2372
2354 2373 if (VerifyBeforeGC &&
2355 2374 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2356 2375 Universe::verify(true);
2357 2376 }
2358 2377
2359 2378 bool init_mark_was_synchronous = false; // until proven otherwise
2360 2379 while (_collectorState != Idling) {
2361 2380 if (TraceCMSState) {
2362 2381 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2363 2382 Thread::current(), _collectorState);
2364 2383 }
2365 2384 switch (_collectorState) {
2366 2385 case InitialMarking:
2367 2386 init_mark_was_synchronous = true; // fact to be exploited in re-mark
2368 2387 checkpointRootsInitial(false);
2369 2388 assert(_collectorState == Marking, "Collector state should have changed"
2370 2389 " within checkpointRootsInitial()");
2371 2390 break;
2372 2391 case Marking:
2373 2392 // initial marking in checkpointRootsInitialWork has been completed
2374 2393 if (VerifyDuringGC &&
2375 2394 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2376 2395 gclog_or_tty->print("Verify before initial mark: ");
2377 2396 Universe::verify(true);
2378 2397 }
2379 2398 {
2380 2399 bool res = markFromRoots(false);
2381 2400 assert(res && _collectorState == FinalMarking, "Collector state should "
2382 2401 "have changed");
2383 2402 break;
2384 2403 }
2385 2404 case FinalMarking:
2386 2405 if (VerifyDuringGC &&
2387 2406 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2388 2407 gclog_or_tty->print("Verify before re-mark: ");
2389 2408 Universe::verify(true);
2390 2409 }
2391 2410 checkpointRootsFinal(false, clear_all_soft_refs,
2392 2411 init_mark_was_synchronous);
2393 2412 assert(_collectorState == Sweeping, "Collector state should not "
2394 2413 "have changed within checkpointRootsFinal()");
2395 2414 break;
2396 2415 case Sweeping:
2397 2416 // final marking in checkpointRootsFinal has been completed
2398 2417 if (VerifyDuringGC &&
2399 2418 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2400 2419 gclog_or_tty->print("Verify before sweep: ");
2401 2420 Universe::verify(true);
2402 2421 }
2403 2422 sweep(false);
2404 2423 assert(_collectorState == Resizing, "Incorrect state");
2405 2424 break;
2406 2425 case Resizing: {
2407 2426 // Sweeping has been completed; the actual resize in this case
2408 2427 // is done separately; nothing to be done in this state.
2409 2428 _collectorState = Resetting;
2410 2429 break;
2411 2430 }
2412 2431 case Resetting:
2413 2432 // The heap has been resized.
2414 2433 if (VerifyDuringGC &&
2415 2434 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2416 2435 gclog_or_tty->print("Verify before reset: ");
2417 2436 Universe::verify(true);
2418 2437 }
2419 2438 reset(false);
2420 2439 assert(_collectorState == Idling, "Collector state should "
2421 2440 "have changed");
2422 2441 break;
2423 2442 case Precleaning:
2424 2443 case AbortablePreclean:
2425 2444 // Elide the preclean phase
2426 2445 _collectorState = FinalMarking;
2427 2446 break;
2428 2447 default:
2429 2448 ShouldNotReachHere();
2430 2449 }
2431 2450 if (TraceCMSState) {
2432 2451 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2433 2452 Thread::current(), _collectorState);
2434 2453 }
2435 2454 }
2436 2455
2437 2456 if (UseAdaptiveSizePolicy) {
2438 2457 GenCollectedHeap* gch = GenCollectedHeap::heap();
2439 2458 size_policy()->ms_collection_end(gch->gc_cause());
2440 2459 }
2441 2460
2442 2461 if (VerifyAfterGC &&
2443 2462 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2444 2463 Universe::verify(true);
2445 2464 }
2446 2465 if (TraceCMSState) {
2447 2466 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2448 2467 " exiting collection CMS state %d",
2449 2468 Thread::current(), _collectorState);
2450 2469 }
2451 2470 }
2452 2471
2453 2472 bool CMSCollector::waitForForegroundGC() {
2454 2473 bool res = false;
2455 2474 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2456 2475 "CMS thread should have CMS token");
2457 2476 // Block the foreground collector until the
2458 2477 // background collectors decides whether to
2459 2478 // yield.
2460 2479 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2461 2480 _foregroundGCShouldWait = true;
2462 2481 if (_foregroundGCIsActive) {
2463 2482 // The background collector yields to the
2464 2483 // foreground collector and returns a value
2465 2484 // indicating that it has yielded. The foreground
2466 2485 // collector can proceed.
2467 2486 res = true;
2468 2487 _foregroundGCShouldWait = false;
2469 2488 ConcurrentMarkSweepThread::clear_CMS_flag(
2470 2489 ConcurrentMarkSweepThread::CMS_cms_has_token);
2471 2490 ConcurrentMarkSweepThread::set_CMS_flag(
2472 2491 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2473 2492 // Get a possibly blocked foreground thread going
2474 2493 CGC_lock->notify();
2475 2494 if (TraceCMSState) {
2476 2495 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2477 2496 Thread::current(), _collectorState);
2478 2497 }
2479 2498 while (_foregroundGCIsActive) {
2480 2499 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2481 2500 }
2482 2501 ConcurrentMarkSweepThread::set_CMS_flag(
2483 2502 ConcurrentMarkSweepThread::CMS_cms_has_token);
2484 2503 ConcurrentMarkSweepThread::clear_CMS_flag(
2485 2504 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2486 2505 }
2487 2506 if (TraceCMSState) {
2488 2507 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2489 2508 Thread::current(), _collectorState);
2490 2509 }
2491 2510 return res;
2492 2511 }
2493 2512
2494 2513 // Because of the need to lock the free lists and other structures in
2495 2514 // the collector, common to all the generations that the collector is
2496 2515 // collecting, we need the gc_prologues of individual CMS generations
2497 2516 // delegate to their collector. It may have been simpler had the
2498 2517 // current infrastructure allowed one to call a prologue on a
2499 2518 // collector. In the absence of that we have the generation's
2500 2519 // prologue delegate to the collector, which delegates back
2501 2520 // some "local" work to a worker method in the individual generations
2502 2521 // that it's responsible for collecting, while itself doing any
2503 2522 // work common to all generations it's responsible for. A similar
2504 2523 // comment applies to the gc_epilogue()'s.
2505 2524 // The role of the varaible _between_prologue_and_epilogue is to
2506 2525 // enforce the invocation protocol.
2507 2526 void CMSCollector::gc_prologue(bool full) {
2508 2527 // Call gc_prologue_work() for each CMSGen and PermGen that
2509 2528 // we are responsible for.
2510 2529
2511 2530 // The following locking discipline assumes that we are only called
2512 2531 // when the world is stopped.
2513 2532 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2514 2533
2515 2534 // The CMSCollector prologue must call the gc_prologues for the
2516 2535 // "generations" (including PermGen if any) that it's responsible
2517 2536 // for.
2518 2537
2519 2538 assert( Thread::current()->is_VM_thread()
2520 2539 || ( CMSScavengeBeforeRemark
2521 2540 && Thread::current()->is_ConcurrentGC_thread()),
2522 2541 "Incorrect thread type for prologue execution");
2523 2542
2524 2543 if (_between_prologue_and_epilogue) {
2525 2544 // We have already been invoked; this is a gc_prologue delegation
2526 2545 // from yet another CMS generation that we are responsible for, just
2527 2546 // ignore it since all relevant work has already been done.
2528 2547 return;
2529 2548 }
2530 2549
2531 2550 // set a bit saying prologue has been called; cleared in epilogue
2532 2551 _between_prologue_and_epilogue = true;
2533 2552 // Claim locks for common data structures, then call gc_prologue_work()
2534 2553 // for each CMSGen and PermGen that we are responsible for.
2535 2554
2536 2555 getFreelistLocks(); // gets free list locks on constituent spaces
2537 2556 bitMapLock()->lock_without_safepoint_check();
2538 2557
2539 2558 // Should call gc_prologue_work() for all cms gens we are responsible for
2540 2559 bool registerClosure = _collectorState >= Marking
2541 2560 && _collectorState < Sweeping;
2542 2561 ModUnionClosure* muc = ParallelGCThreads > 0 ? &_modUnionClosurePar
2543 2562 : &_modUnionClosure;
2544 2563 _cmsGen->gc_prologue_work(full, registerClosure, muc);
2545 2564 _permGen->gc_prologue_work(full, registerClosure, muc);
2546 2565
2547 2566 if (!full) {
2548 2567 stats().record_gc0_begin();
2549 2568 }
2550 2569 }
2551 2570
2552 2571 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2553 2572 // Delegate to CMScollector which knows how to coordinate between
2554 2573 // this and any other CMS generations that it is responsible for
2555 2574 // collecting.
2556 2575 collector()->gc_prologue(full);
2557 2576 }
2558 2577
2559 2578 // This is a "private" interface for use by this generation's CMSCollector.
2560 2579 // Not to be called directly by any other entity (for instance,
2561 2580 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2562 2581 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2563 2582 bool registerClosure, ModUnionClosure* modUnionClosure) {
2564 2583 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2565 2584 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2566 2585 "Should be NULL");
2567 2586 if (registerClosure) {
2568 2587 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2569 2588 }
2570 2589 cmsSpace()->gc_prologue();
2571 2590 // Clear stat counters
2572 2591 NOT_PRODUCT(
2573 2592 assert(_numObjectsPromoted == 0, "check");
2574 2593 assert(_numWordsPromoted == 0, "check");
2575 2594 if (Verbose && PrintGC) {
2576 2595 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2577 2596 SIZE_FORMAT" bytes concurrently",
2578 2597 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2579 2598 }
2580 2599 _numObjectsAllocated = 0;
2581 2600 _numWordsAllocated = 0;
2582 2601 )
2583 2602 }
2584 2603
2585 2604 void CMSCollector::gc_epilogue(bool full) {
2586 2605 // The following locking discipline assumes that we are only called
2587 2606 // when the world is stopped.
2588 2607 assert(SafepointSynchronize::is_at_safepoint(),
2589 2608 "world is stopped assumption");
2590 2609
2591 2610 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2592 2611 // if linear allocation blocks need to be appropriately marked to allow the
2593 2612 // the blocks to be parsable. We also check here whether we need to nudge the
2594 2613 // CMS collector thread to start a new cycle (if it's not already active).
2595 2614 assert( Thread::current()->is_VM_thread()
2596 2615 || ( CMSScavengeBeforeRemark
2597 2616 && Thread::current()->is_ConcurrentGC_thread()),
2598 2617 "Incorrect thread type for epilogue execution");
2599 2618
2600 2619 if (!_between_prologue_and_epilogue) {
2601 2620 // We have already been invoked; this is a gc_epilogue delegation
2602 2621 // from yet another CMS generation that we are responsible for, just
2603 2622 // ignore it since all relevant work has already been done.
2604 2623 return;
2605 2624 }
2606 2625 assert(haveFreelistLocks(), "must have freelist locks");
2607 2626 assert_lock_strong(bitMapLock());
2608 2627
2609 2628 _cmsGen->gc_epilogue_work(full);
2610 2629 _permGen->gc_epilogue_work(full);
2611 2630
2612 2631 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2613 2632 // in case sampling was not already enabled, enable it
2614 2633 _start_sampling = true;
2615 2634 }
2616 2635 // reset _eden_chunk_array so sampling starts afresh
2617 2636 _eden_chunk_index = 0;
2618 2637
2619 2638 size_t cms_used = _cmsGen->cmsSpace()->used();
2620 2639 size_t perm_used = _permGen->cmsSpace()->used();
2621 2640
2622 2641 // update performance counters - this uses a special version of
2623 2642 // update_counters() that allows the utilization to be passed as a
2624 2643 // parameter, avoiding multiple calls to used().
2625 2644 //
2626 2645 _cmsGen->update_counters(cms_used);
2627 2646 _permGen->update_counters(perm_used);
2628 2647
2629 2648 if (CMSIncrementalMode) {
2630 2649 icms_update_allocation_limits();
2631 2650 }
2632 2651
2633 2652 bitMapLock()->unlock();
2634 2653 releaseFreelistLocks();
2635 2654
2636 2655 _between_prologue_and_epilogue = false; // ready for next cycle
2637 2656 }
2638 2657
2639 2658 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2640 2659 collector()->gc_epilogue(full);
2641 2660
2642 2661 // Also reset promotion tracking in par gc thread states.
2643 2662 if (ParallelGCThreads > 0) {
2644 2663 for (uint i = 0; i < ParallelGCThreads; i++) {
2645 2664 _par_gc_thread_states[i]->promo.stopTrackingPromotions();
2646 2665 }
2647 2666 }
2648 2667 }
2649 2668
2650 2669 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2651 2670 assert(!incremental_collection_failed(), "Should have been cleared");
2652 2671 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2653 2672 cmsSpace()->gc_epilogue();
2654 2673 // Print stat counters
2655 2674 NOT_PRODUCT(
2656 2675 assert(_numObjectsAllocated == 0, "check");
2657 2676 assert(_numWordsAllocated == 0, "check");
2658 2677 if (Verbose && PrintGC) {
2659 2678 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2660 2679 SIZE_FORMAT" bytes",
2661 2680 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2662 2681 }
2663 2682 _numObjectsPromoted = 0;
2664 2683 _numWordsPromoted = 0;
2665 2684 )
2666 2685
2667 2686 if (PrintGC && Verbose) {
2668 2687 // Call down the chain in contiguous_available needs the freelistLock
2669 2688 // so print this out before releasing the freeListLock.
2670 2689 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2671 2690 contiguous_available());
2672 2691 }
2673 2692 }
2674 2693
2675 2694 #ifndef PRODUCT
2676 2695 bool CMSCollector::have_cms_token() {
2677 2696 Thread* thr = Thread::current();
2678 2697 if (thr->is_VM_thread()) {
2679 2698 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2680 2699 } else if (thr->is_ConcurrentGC_thread()) {
2681 2700 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2682 2701 } else if (thr->is_GC_task_thread()) {
2683 2702 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2684 2703 ParGCRareEvent_lock->owned_by_self();
2685 2704 }
2686 2705 return false;
2687 2706 }
2688 2707 #endif
2689 2708
2690 2709 // Check reachability of the given heap address in CMS generation,
2691 2710 // treating all other generations as roots.
2692 2711 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2693 2712 // We could "guarantee" below, rather than assert, but i'll
2694 2713 // leave these as "asserts" so that an adventurous debugger
2695 2714 // could try this in the product build provided some subset of
2696 2715 // the conditions were met, provided they were intersted in the
2697 2716 // results and knew that the computation below wouldn't interfere
2698 2717 // with other concurrent computations mutating the structures
2699 2718 // being read or written.
2700 2719 assert(SafepointSynchronize::is_at_safepoint(),
2701 2720 "Else mutations in object graph will make answer suspect");
2702 2721 assert(have_cms_token(), "Should hold cms token");
2703 2722 assert(haveFreelistLocks(), "must hold free list locks");
2704 2723 assert_lock_strong(bitMapLock());
2705 2724
2706 2725 // Clear the marking bit map array before starting, but, just
2707 2726 // for kicks, first report if the given address is already marked
2708 2727 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
2709 2728 _markBitMap.isMarked(addr) ? "" : " not");
2710 2729
2711 2730 if (verify_after_remark()) {
2712 2731 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2713 2732 bool result = verification_mark_bm()->isMarked(addr);
2714 2733 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
2715 2734 result ? "IS" : "is NOT");
2716 2735 return result;
2717 2736 } else {
2718 2737 gclog_or_tty->print_cr("Could not compute result");
2719 2738 return false;
2720 2739 }
2721 2740 }
2722 2741
2723 2742 ////////////////////////////////////////////////////////
2724 2743 // CMS Verification Support
2725 2744 ////////////////////////////////////////////////////////
2726 2745 // Following the remark phase, the following invariant
2727 2746 // should hold -- each object in the CMS heap which is
2728 2747 // marked in markBitMap() should be marked in the verification_mark_bm().
2729 2748
2730 2749 class VerifyMarkedClosure: public BitMapClosure {
2731 2750 CMSBitMap* _marks;
2732 2751 bool _failed;
2733 2752
2734 2753 public:
2735 2754 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2736 2755
2737 2756 void do_bit(size_t offset) {
2738 2757 HeapWord* addr = _marks->offsetToHeapWord(offset);
2739 2758 if (!_marks->isMarked(addr)) {
2740 2759 oop(addr)->print();
2741 2760 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2742 2761 _failed = true;
2743 2762 }
2744 2763 }
2745 2764
2746 2765 bool failed() { return _failed; }
2747 2766 };
2748 2767
2749 2768 bool CMSCollector::verify_after_remark() {
2750 2769 gclog_or_tty->print(" [Verifying CMS Marking... ");
2751 2770 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2752 2771 static bool init = false;
2753 2772
2754 2773 assert(SafepointSynchronize::is_at_safepoint(),
2755 2774 "Else mutations in object graph will make answer suspect");
2756 2775 assert(have_cms_token(),
2757 2776 "Else there may be mutual interference in use of "
2758 2777 " verification data structures");
2759 2778 assert(_collectorState > Marking && _collectorState <= Sweeping,
2760 2779 "Else marking info checked here may be obsolete");
2761 2780 assert(haveFreelistLocks(), "must hold free list locks");
2762 2781 assert_lock_strong(bitMapLock());
2763 2782
2764 2783
2765 2784 // Allocate marking bit map if not already allocated
2766 2785 if (!init) { // first time
2767 2786 if (!verification_mark_bm()->allocate(_span)) {
2768 2787 return false;
2769 2788 }
2770 2789 init = true;
2771 2790 }
2772 2791
2773 2792 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2774 2793
2775 2794 // Turn off refs discovery -- so we will be tracing through refs.
2776 2795 // This is as intended, because by this time
2777 2796 // GC must already have cleared any refs that need to be cleared,
2778 2797 // and traced those that need to be marked; moreover,
2779 2798 // the marking done here is not going to intefere in any
2780 2799 // way with the marking information used by GC.
2781 2800 NoRefDiscovery no_discovery(ref_processor());
2782 2801
2783 2802 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2784 2803
2785 2804 // Clear any marks from a previous round
2786 2805 verification_mark_bm()->clear_all();
2787 2806 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2788 2807 assert(overflow_list_is_empty(), "overflow list should be empty");
2789 2808
2790 2809 GenCollectedHeap* gch = GenCollectedHeap::heap();
2791 2810 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2792 2811 // Update the saved marks which may affect the root scans.
2793 2812 gch->save_marks();
2794 2813
2795 2814 if (CMSRemarkVerifyVariant == 1) {
2796 2815 // In this first variant of verification, we complete
2797 2816 // all marking, then check if the new marks-verctor is
2798 2817 // a subset of the CMS marks-vector.
2799 2818 verify_after_remark_work_1();
2800 2819 } else if (CMSRemarkVerifyVariant == 2) {
2801 2820 // In this second variant of verification, we flag an error
2802 2821 // (i.e. an object reachable in the new marks-vector not reachable
2803 2822 // in the CMS marks-vector) immediately, also indicating the
2804 2823 // identify of an object (A) that references the unmarked object (B) --
2805 2824 // presumably, a mutation to A failed to be picked up by preclean/remark?
2806 2825 verify_after_remark_work_2();
2807 2826 } else {
2808 2827 warning("Unrecognized value %d for CMSRemarkVerifyVariant",
2809 2828 CMSRemarkVerifyVariant);
2810 2829 }
2811 2830 gclog_or_tty->print(" done] ");
2812 2831 return true;
2813 2832 }
2814 2833
2815 2834 void CMSCollector::verify_after_remark_work_1() {
2816 2835 ResourceMark rm;
2817 2836 HandleMark hm;
2818 2837 GenCollectedHeap* gch = GenCollectedHeap::heap();
2819 2838
2820 2839 // Mark from roots one level into CMS
2821 2840 MarkRefsIntoClosure notOlder(_span, verification_mark_bm(), true /* nmethods */);
2822 2841 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2823 2842
2824 2843 gch->gen_process_strong_roots(_cmsGen->level(),
2825 2844 true, // younger gens are roots
2826 2845 true, // collecting perm gen
2827 2846 SharedHeap::ScanningOption(roots_scanning_options()),
2828 2847 NULL, ¬Older);
2829 2848
2830 2849 // Now mark from the roots
2831 2850 assert(_revisitStack.isEmpty(), "Should be empty");
2832 2851 MarkFromRootsClosure markFromRootsClosure(this, _span,
2833 2852 verification_mark_bm(), verification_mark_stack(), &_revisitStack,
2834 2853 false /* don't yield */, true /* verifying */);
2835 2854 assert(_restart_addr == NULL, "Expected pre-condition");
2836 2855 verification_mark_bm()->iterate(&markFromRootsClosure);
2837 2856 while (_restart_addr != NULL) {
2838 2857 // Deal with stack overflow: by restarting at the indicated
2839 2858 // address.
2840 2859 HeapWord* ra = _restart_addr;
2841 2860 markFromRootsClosure.reset(ra);
2842 2861 _restart_addr = NULL;
2843 2862 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2844 2863 }
2845 2864 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2846 2865 verify_work_stacks_empty();
2847 2866 // Should reset the revisit stack above, since no class tree
2848 2867 // surgery is forthcoming.
2849 2868 _revisitStack.reset(); // throwing away all contents
2850 2869
2851 2870 // Marking completed -- now verify that each bit marked in
2852 2871 // verification_mark_bm() is also marked in markBitMap(); flag all
2853 2872 // errors by printing corresponding objects.
2854 2873 VerifyMarkedClosure vcl(markBitMap());
2855 2874 verification_mark_bm()->iterate(&vcl);
2856 2875 if (vcl.failed()) {
2857 2876 gclog_or_tty->print("Verification failed");
2858 2877 Universe::heap()->print();
2859 2878 fatal(" ... aborting");
2860 2879 }
2861 2880 }
2862 2881
2863 2882 void CMSCollector::verify_after_remark_work_2() {
2864 2883 ResourceMark rm;
2865 2884 HandleMark hm;
2866 2885 GenCollectedHeap* gch = GenCollectedHeap::heap();
2867 2886
2868 2887 // Mark from roots one level into CMS
2869 2888 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2870 2889 markBitMap(), true /* nmethods */);
2871 2890 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2872 2891 gch->gen_process_strong_roots(_cmsGen->level(),
2873 2892 true, // younger gens are roots
2874 2893 true, // collecting perm gen
2875 2894 SharedHeap::ScanningOption(roots_scanning_options()),
2876 2895 NULL, ¬Older);
2877 2896
2878 2897 // Now mark from the roots
2879 2898 assert(_revisitStack.isEmpty(), "Should be empty");
2880 2899 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2881 2900 verification_mark_bm(), markBitMap(), verification_mark_stack());
2882 2901 assert(_restart_addr == NULL, "Expected pre-condition");
2883 2902 verification_mark_bm()->iterate(&markFromRootsClosure);
2884 2903 while (_restart_addr != NULL) {
2885 2904 // Deal with stack overflow: by restarting at the indicated
2886 2905 // address.
2887 2906 HeapWord* ra = _restart_addr;
2888 2907 markFromRootsClosure.reset(ra);
2889 2908 _restart_addr = NULL;
2890 2909 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2891 2910 }
2892 2911 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2893 2912 verify_work_stacks_empty();
2894 2913 // Should reset the revisit stack above, since no class tree
2895 2914 // surgery is forthcoming.
2896 2915 _revisitStack.reset(); // throwing away all contents
2897 2916
2898 2917 // Marking completed -- now verify that each bit marked in
2899 2918 // verification_mark_bm() is also marked in markBitMap(); flag all
2900 2919 // errors by printing corresponding objects.
2901 2920 VerifyMarkedClosure vcl(markBitMap());
2902 2921 verification_mark_bm()->iterate(&vcl);
2903 2922 assert(!vcl.failed(), "Else verification above should not have succeeded");
2904 2923 }
2905 2924
2906 2925 void ConcurrentMarkSweepGeneration::save_marks() {
2907 2926 // delegate to CMS space
2908 2927 cmsSpace()->save_marks();
2909 2928 for (uint i = 0; i < ParallelGCThreads; i++) {
2910 2929 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2911 2930 }
2912 2931 }
2913 2932
2914 2933 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2915 2934 return cmsSpace()->no_allocs_since_save_marks();
2916 2935 }
2917 2936
2918 2937 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2919 2938 \
2920 2939 void ConcurrentMarkSweepGeneration:: \
2921 2940 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2922 2941 cl->set_generation(this); \
2923 2942 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2924 2943 cl->reset_generation(); \
2925 2944 save_marks(); \
2926 2945 }
2927 2946
2928 2947 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2929 2948
2930 2949 void
2931 2950 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk)
2932 2951 {
2933 2952 // Not currently implemented; need to do the following. -- ysr.
2934 2953 // dld -- I think that is used for some sort of allocation profiler. So it
2935 2954 // really means the objects allocated by the mutator since the last
2936 2955 // GC. We could potentially implement this cheaply by recording only
2937 2956 // the direct allocations in a side data structure.
2938 2957 //
2939 2958 // I think we probably ought not to be required to support these
2940 2959 // iterations at any arbitrary point; I think there ought to be some
2941 2960 // call to enable/disable allocation profiling in a generation/space,
2942 2961 // and the iterator ought to return the objects allocated in the
2943 2962 // gen/space since the enable call, or the last iterator call (which
2944 2963 // will probably be at a GC.) That way, for gens like CM&S that would
2945 2964 // require some extra data structure to support this, we only pay the
2946 2965 // cost when it's in use...
2947 2966 cmsSpace()->object_iterate_since_last_GC(blk);
2948 2967 }
2949 2968
2950 2969 void
2951 2970 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
2952 2971 cl->set_generation(this);
2953 2972 younger_refs_in_space_iterate(_cmsSpace, cl);
2954 2973 cl->reset_generation();
2955 2974 }
2956 2975
2957 2976 void
2958 2977 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) {
2959 2978 if (freelistLock()->owned_by_self()) {
2960 2979 Generation::oop_iterate(mr, cl);
2961 2980 } else {
2962 2981 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2963 2982 Generation::oop_iterate(mr, cl);
2964 2983 }
2965 2984 }
2966 2985
2967 2986 void
2968 2987 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) {
2969 2988 if (freelistLock()->owned_by_self()) {
2970 2989 Generation::oop_iterate(cl);
2971 2990 } else {
2972 2991 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2973 2992 Generation::oop_iterate(cl);
2974 2993 }
2975 2994 }
2976 2995
2977 2996 void
2978 2997 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2979 2998 if (freelistLock()->owned_by_self()) {
2980 2999 Generation::object_iterate(cl);
2981 3000 } else {
2982 3001 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2983 3002 Generation::object_iterate(cl);
2984 3003 }
2985 3004 }
2986 3005
2987 3006 void
2988 3007 ConcurrentMarkSweepGeneration::pre_adjust_pointers() {
2989 3008 }
2990 3009
2991 3010 void
2992 3011 ConcurrentMarkSweepGeneration::post_compact() {
2993 3012 }
2994 3013
2995 3014 void
2996 3015 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2997 3016 // Fix the linear allocation blocks to look like free blocks.
2998 3017
2999 3018 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3000 3019 // are not called when the heap is verified during universe initialization and
3001 3020 // at vm shutdown.
3002 3021 if (freelistLock()->owned_by_self()) {
3003 3022 cmsSpace()->prepare_for_verify();
3004 3023 } else {
3005 3024 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3006 3025 cmsSpace()->prepare_for_verify();
3007 3026 }
3008 3027 }
3009 3028
3010 3029 void
3011 3030 ConcurrentMarkSweepGeneration::verify(bool allow_dirty /* ignored */) {
3012 3031 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3013 3032 // are not called when the heap is verified during universe initialization and
3014 3033 // at vm shutdown.
3015 3034 if (freelistLock()->owned_by_self()) {
3016 3035 cmsSpace()->verify(false /* ignored */);
3017 3036 } else {
3018 3037 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3019 3038 cmsSpace()->verify(false /* ignored */);
3020 3039 }
3021 3040 }
3022 3041
3023 3042 void CMSCollector::verify(bool allow_dirty /* ignored */) {
3024 3043 _cmsGen->verify(allow_dirty);
3025 3044 _permGen->verify(allow_dirty);
3026 3045 }
3027 3046
3028 3047 #ifndef PRODUCT
3029 3048 bool CMSCollector::overflow_list_is_empty() const {
3030 3049 assert(_num_par_pushes >= 0, "Inconsistency");
3031 3050 if (_overflow_list == NULL) {
3032 3051 assert(_num_par_pushes == 0, "Inconsistency");
3033 3052 }
3034 3053 return _overflow_list == NULL;
3035 3054 }
3036 3055
3037 3056 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3038 3057 // merely consolidate assertion checks that appear to occur together frequently.
3039 3058 void CMSCollector::verify_work_stacks_empty() const {
↓ open down ↓ |
908 lines elided |
↑ open up ↑ |
3040 3059 assert(_markStack.isEmpty(), "Marking stack should be empty");
3041 3060 assert(overflow_list_is_empty(), "Overflow list should be empty");
3042 3061 }
3043 3062
3044 3063 void CMSCollector::verify_overflow_empty() const {
3045 3064 assert(overflow_list_is_empty(), "Overflow list should be empty");
3046 3065 assert(no_preserved_marks(), "No preserved marks");
3047 3066 }
3048 3067 #endif // PRODUCT
3049 3068
3069 +// Decide if we want to enable class unloading as part of the
3070 +// ensuing concurrent GC cycle. We will collect the perm gen and
3071 +// unload classes if it's the case that:
3072 +// (1) an explicit gc request has been made and the flag
3073 +// ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3074 +// (2) (a) class unloading is enabled at the command line, and
3075 +// (b) (i) perm gen threshold has been crossed, or
3076 +// (ii) old gen is getting really full, or
3077 +// (iii) the previous N CMS collections did not collect the
3078 +// perm gen
3079 +// NOTE: Provided there is no change in the state of the heap between
3080 +// calls to this method, it should have idempotent results. Moreover,
3081 +// its results should be monotonically increasing (i.e. going from 0 to 1,
3082 +// but not 1 to 0) between successive calls between which the heap was
3083 +// not collected. For the implementation below, it must thus rely on
3084 +// the property that concurrent_cycles_since_last_unload()
3085 +// will not decrease unless a collection cycle happened and that
3086 +// _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are
3087 +// themselves also monotonic in that sense. See check_monotonicity()
3088 +// below.
3089 +bool CMSCollector::update_should_unload_classes() {
3090 + // Condition 1 above
3091 + if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3092 + _should_unload_classes = true;
3093 + } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3094 + // Disjuncts 2.b.(i,ii,iii) above
3095 + _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3096 + CMSClassUnloadingMaxInterval)
3097 + || _permGen->should_concurrent_collect()
3098 + || _cmsGen->is_too_full();
3099 + }
3100 + return _should_unload_classes;
3101 +}
3102 +
3103 +bool ConcurrentMarkSweepGeneration::is_too_full() const {
3104 + bool res = should_concurrent_collect();
3105 +#define CMSIsTooFullPercentage 98
3106 + res = res && occupancy() > (double)CMSIsTooFullPercentage/100.0;
3107 + return res;
3108 +}
3109 +
3050 3110 void CMSCollector::setup_cms_unloading_and_verification_state() {
3051 3111 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3052 3112 || VerifyBeforeExit;
3053 3113 const int rso = SharedHeap::SO_Symbols | SharedHeap::SO_Strings
3054 3114 | SharedHeap::SO_CodeCache;
3055 3115
3056 - if (cms_should_unload_classes()) { // Should unload classes this cycle
3116 + if (should_unload_classes()) { // Should unload classes this cycle
3057 3117 remove_root_scanning_option(rso); // Shrink the root set appropriately
3058 3118 set_verifying(should_verify); // Set verification state for this cycle
3059 3119 return; // Nothing else needs to be done at this time
3060 3120 }
3061 3121
3062 3122 // Not unloading classes this cycle
3063 - assert(!cms_should_unload_classes(), "Inconsitency!");
3064 - if ((!verifying() || cms_unloaded_classes_last_cycle()) && should_verify) {
3123 + assert(!should_unload_classes(), "Inconsitency!");
3124 + if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3065 3125 // We were not verifying, or we _were_ unloading classes in the last cycle,
3066 3126 // AND some verification options are enabled this cycle; in this case,
3067 3127 // we must make sure that the deadness map is allocated if not already so,
3068 3128 // and cleared (if already allocated previously --
3069 3129 // CMSBitMap::sizeInBits() is used to determine if it's allocated).
3070 3130 if (perm_gen_verify_bit_map()->sizeInBits() == 0) {
3071 3131 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) {
3072 3132 warning("Failed to allocate permanent generation verification CMS Bit Map;\n"
3073 3133 "permanent generation verification disabled");
3074 3134 return; // Note that we leave verification disabled, so we'll retry this
3075 3135 // allocation next cycle. We _could_ remember this failure
3076 3136 // and skip further attempts and permanently disable verification
3077 3137 // attempts if that is considered more desirable.
3078 3138 }
3079 3139 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()),
3080 3140 "_perm_gen_ver_bit_map inconsistency?");
3081 3141 } else {
3082 3142 perm_gen_verify_bit_map()->clear_all();
3083 3143 }
3084 3144 // Include symbols, strings and code cache elements to prevent their resurrection.
3085 3145 add_root_scanning_option(rso);
3086 3146 set_verifying(true);
3087 3147 } else if (verifying() && !should_verify) {
3088 3148 // We were verifying, but some verification flags got disabled.
3089 3149 set_verifying(false);
3090 3150 // Exclude symbols, strings and code cache elements from root scanning to
3091 3151 // reduce IM and RM pauses.
3092 3152 remove_root_scanning_option(rso);
3093 3153 }
3094 3154 }
3095 3155
3096 3156
3097 3157 #ifndef PRODUCT
3098 3158 HeapWord* CMSCollector::block_start(const void* p) const {
3099 3159 const HeapWord* addr = (HeapWord*)p;
3100 3160 if (_span.contains(p)) {
3101 3161 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3102 3162 return _cmsGen->cmsSpace()->block_start(p);
3103 3163 } else {
3104 3164 assert(_permGen->cmsSpace()->is_in_reserved(addr),
3105 3165 "Inconsistent _span?");
3106 3166 return _permGen->cmsSpace()->block_start(p);
3107 3167 }
3108 3168 }
3109 3169 return NULL;
3110 3170 }
3111 3171 #endif
3112 3172
3113 3173 HeapWord*
3114 3174 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3115 3175 bool tlab,
3116 3176 bool parallel) {
3117 3177 assert(!tlab, "Can't deal with TLAB allocation");
3118 3178 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3119 3179 expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3120 3180 CMSExpansionCause::_satisfy_allocation);
3121 3181 if (GCExpandToAllocateDelayMillis > 0) {
3122 3182 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3123 3183 }
3124 3184 size_t adj_word_sz = CompactibleFreeListSpace::adjustObjectSize(word_size);
3125 3185 if (parallel) {
3126 3186 return cmsSpace()->par_allocate(adj_word_sz);
3127 3187 } else {
3128 3188 return cmsSpace()->allocate(adj_word_sz);
3129 3189 }
3130 3190 }
3131 3191
3132 3192 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3133 3193 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3134 3194 // to CardGeneration and share it...
3135 3195 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3136 3196 CMSExpansionCause::Cause cause)
3137 3197 {
3138 3198 assert_locked_or_safepoint(Heap_lock);
3139 3199
3140 3200 size_t aligned_bytes = ReservedSpace::page_align_size_up(bytes);
3141 3201 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes);
3142 3202 bool success = false;
3143 3203 if (aligned_expand_bytes > aligned_bytes) {
3144 3204 success = grow_by(aligned_expand_bytes);
3145 3205 }
3146 3206 if (!success) {
3147 3207 success = grow_by(aligned_bytes);
3148 3208 }
3149 3209 if (!success) {
3150 3210 size_t remaining_bytes = _virtual_space.uncommitted_size();
3151 3211 if (remaining_bytes > 0) {
3152 3212 success = grow_by(remaining_bytes);
3153 3213 }
3154 3214 }
3155 3215 if (GC_locker::is_active()) {
3156 3216 if (PrintGC && Verbose) {
3157 3217 gclog_or_tty->print_cr("Garbage collection disabled, expanded heap instead");
3158 3218 }
3159 3219 }
3160 3220 // remember why we expanded; this information is used
3161 3221 // by shouldConcurrentCollect() when making decisions on whether to start
3162 3222 // a new CMS cycle.
3163 3223 if (success) {
3164 3224 set_expansion_cause(cause);
3165 3225 if (PrintGCDetails && Verbose) {
3166 3226 gclog_or_tty->print_cr("Expanded CMS gen for %s",
3167 3227 CMSExpansionCause::to_string(cause));
3168 3228 }
3169 3229 }
3170 3230 }
3171 3231
3172 3232 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3173 3233 HeapWord* res = NULL;
3174 3234 MutexLocker x(ParGCRareEvent_lock);
3175 3235 while (true) {
3176 3236 // Expansion by some other thread might make alloc OK now:
3177 3237 res = ps->lab.alloc(word_sz);
3178 3238 if (res != NULL) return res;
3179 3239 // If there's not enough expansion space available, give up.
3180 3240 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3181 3241 return NULL;
3182 3242 }
3183 3243 // Otherwise, we try expansion.
3184 3244 expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3185 3245 CMSExpansionCause::_allocate_par_lab);
3186 3246 // Now go around the loop and try alloc again;
3187 3247 // A competing par_promote might beat us to the expansion space,
3188 3248 // so we may go around the loop again if promotion fails agaion.
3189 3249 if (GCExpandToAllocateDelayMillis > 0) {
3190 3250 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3191 3251 }
3192 3252 }
3193 3253 }
3194 3254
3195 3255
3196 3256 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3197 3257 PromotionInfo* promo) {
3198 3258 MutexLocker x(ParGCRareEvent_lock);
3199 3259 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3200 3260 while (true) {
3201 3261 // Expansion by some other thread might make alloc OK now:
3202 3262 if (promo->ensure_spooling_space()) {
3203 3263 assert(promo->has_spooling_space(),
3204 3264 "Post-condition of successful ensure_spooling_space()");
3205 3265 return true;
3206 3266 }
3207 3267 // If there's not enough expansion space available, give up.
3208 3268 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3209 3269 return false;
3210 3270 }
3211 3271 // Otherwise, we try expansion.
3212 3272 expand(refill_size_bytes, MinHeapDeltaBytes,
3213 3273 CMSExpansionCause::_allocate_par_spooling_space);
3214 3274 // Now go around the loop and try alloc again;
3215 3275 // A competing allocation might beat us to the expansion space,
3216 3276 // so we may go around the loop again if allocation fails again.
3217 3277 if (GCExpandToAllocateDelayMillis > 0) {
3218 3278 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3219 3279 }
3220 3280 }
3221 3281 }
3222 3282
3223 3283
3224 3284
3225 3285 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3226 3286 assert_locked_or_safepoint(Heap_lock);
3227 3287 size_t size = ReservedSpace::page_align_size_down(bytes);
3228 3288 if (size > 0) {
3229 3289 shrink_by(size);
3230 3290 }
3231 3291 }
3232 3292
3233 3293 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3234 3294 assert_locked_or_safepoint(Heap_lock);
3235 3295 bool result = _virtual_space.expand_by(bytes);
3236 3296 if (result) {
3237 3297 HeapWord* old_end = _cmsSpace->end();
3238 3298 size_t new_word_size =
3239 3299 heap_word_size(_virtual_space.committed_size());
3240 3300 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3241 3301 _bts->resize(new_word_size); // resize the block offset shared array
3242 3302 Universe::heap()->barrier_set()->resize_covered_region(mr);
3243 3303 // Hmmmm... why doesn't CFLS::set_end verify locking?
3244 3304 // This is quite ugly; FIX ME XXX
3245 3305 _cmsSpace->assert_locked();
3246 3306 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3247 3307
3248 3308 // update the space and generation capacity counters
3249 3309 if (UsePerfData) {
3250 3310 _space_counters->update_capacity();
3251 3311 _gen_counters->update_all();
3252 3312 }
3253 3313
3254 3314 if (Verbose && PrintGC) {
3255 3315 size_t new_mem_size = _virtual_space.committed_size();
3256 3316 size_t old_mem_size = new_mem_size - bytes;
3257 3317 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK",
3258 3318 name(), old_mem_size/K, bytes/K, new_mem_size/K);
3259 3319 }
3260 3320 }
3261 3321 return result;
3262 3322 }
3263 3323
3264 3324 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3265 3325 assert_locked_or_safepoint(Heap_lock);
3266 3326 bool success = true;
3267 3327 const size_t remaining_bytes = _virtual_space.uncommitted_size();
3268 3328 if (remaining_bytes > 0) {
3269 3329 success = grow_by(remaining_bytes);
3270 3330 DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3271 3331 }
3272 3332 return success;
3273 3333 }
3274 3334
3275 3335 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3276 3336 assert_locked_or_safepoint(Heap_lock);
3277 3337 assert_lock_strong(freelistLock());
3278 3338 // XXX Fix when compaction is implemented.
3279 3339 warning("Shrinking of CMS not yet implemented");
3280 3340 return;
3281 3341 }
3282 3342
3283 3343
3284 3344 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3285 3345 // phases.
3286 3346 class CMSPhaseAccounting: public StackObj {
3287 3347 public:
3288 3348 CMSPhaseAccounting(CMSCollector *collector,
3289 3349 const char *phase,
3290 3350 bool print_cr = true);
3291 3351 ~CMSPhaseAccounting();
3292 3352
3293 3353 private:
3294 3354 CMSCollector *_collector;
3295 3355 const char *_phase;
3296 3356 elapsedTimer _wallclock;
3297 3357 bool _print_cr;
3298 3358
3299 3359 public:
3300 3360 // Not MT-safe; so do not pass around these StackObj's
3301 3361 // where they may be accessed by other threads.
3302 3362 jlong wallclock_millis() {
3303 3363 assert(_wallclock.is_active(), "Wall clock should not stop");
3304 3364 _wallclock.stop(); // to record time
3305 3365 jlong ret = _wallclock.milliseconds();
3306 3366 _wallclock.start(); // restart
3307 3367 return ret;
3308 3368 }
3309 3369 };
3310 3370
3311 3371 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3312 3372 const char *phase,
3313 3373 bool print_cr) :
3314 3374 _collector(collector), _phase(phase), _print_cr(print_cr) {
3315 3375
3316 3376 if (PrintCMSStatistics != 0) {
3317 3377 _collector->resetYields();
3318 3378 }
3319 3379 if (PrintGCDetails && PrintGCTimeStamps) {
3320 3380 gclog_or_tty->date_stamp(PrintGCDateStamps);
3321 3381 gclog_or_tty->stamp();
3322 3382 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]",
3323 3383 _collector->cmsGen()->short_name(), _phase);
3324 3384 }
3325 3385 _collector->resetTimer();
3326 3386 _wallclock.start();
3327 3387 _collector->startTimer();
3328 3388 }
3329 3389
3330 3390 CMSPhaseAccounting::~CMSPhaseAccounting() {
3331 3391 assert(_wallclock.is_active(), "Wall clock should not have stopped");
3332 3392 _collector->stopTimer();
3333 3393 _wallclock.stop();
3334 3394 if (PrintGCDetails) {
3335 3395 gclog_or_tty->date_stamp(PrintGCDateStamps);
3336 3396 if (PrintGCTimeStamps) {
3337 3397 gclog_or_tty->stamp();
3338 3398 gclog_or_tty->print(": ");
3339 3399 }
3340 3400 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3341 3401 _collector->cmsGen()->short_name(),
3342 3402 _phase, _collector->timerValue(), _wallclock.seconds());
3343 3403 if (_print_cr) {
3344 3404 gclog_or_tty->print_cr("");
3345 3405 }
3346 3406 if (PrintCMSStatistics != 0) {
3347 3407 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3348 3408 _collector->yields());
3349 3409 }
3350 3410 }
3351 3411 }
3352 3412
3353 3413 // CMS work
3354 3414
3355 3415 // Checkpoint the roots into this generation from outside
3356 3416 // this generation. [Note this initial checkpoint need only
3357 3417 // be approximate -- we'll do a catch up phase subsequently.]
3358 3418 void CMSCollector::checkpointRootsInitial(bool asynch) {
3359 3419 assert(_collectorState == InitialMarking, "Wrong collector state");
3360 3420 check_correct_thread_executing();
3361 3421 ReferenceProcessor* rp = ref_processor();
3362 3422 SpecializationStats::clear();
3363 3423 assert(_restart_addr == NULL, "Control point invariant");
3364 3424 if (asynch) {
3365 3425 // acquire locks for subsequent manipulations
3366 3426 MutexLockerEx x(bitMapLock(),
3367 3427 Mutex::_no_safepoint_check_flag);
3368 3428 checkpointRootsInitialWork(asynch);
3369 3429 rp->verify_no_references_recorded();
3370 3430 rp->enable_discovery(); // enable ("weak") refs discovery
3371 3431 _collectorState = Marking;
3372 3432 } else {
3373 3433 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3374 3434 // which recognizes if we are a CMS generation, and doesn't try to turn on
3375 3435 // discovery; verify that they aren't meddling.
3376 3436 assert(!rp->discovery_is_atomic(),
3377 3437 "incorrect setting of discovery predicate");
3378 3438 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3379 3439 "ref discovery for this generation kind");
3380 3440 // already have locks
3381 3441 checkpointRootsInitialWork(asynch);
3382 3442 rp->enable_discovery(); // now enable ("weak") refs discovery
3383 3443 _collectorState = Marking;
3384 3444 }
3385 3445 SpecializationStats::print();
3386 3446 }
3387 3447
3388 3448 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3389 3449 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3390 3450 assert(_collectorState == InitialMarking, "just checking");
3391 3451
3392 3452 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3393 3453 // precede our marking with a collection of all
3394 3454 // younger generations to keep floating garbage to a minimum.
3395 3455 // XXX: we won't do this for now -- it's an optimization to be done later.
3396 3456
3397 3457 // already have locks
3398 3458 assert_lock_strong(bitMapLock());
3399 3459 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3400 3460
3401 3461 // Setup the verification and class unloading state for this
3402 3462 // CMS collection cycle.
3403 3463 setup_cms_unloading_and_verification_state();
3404 3464
3405 3465 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork",
3406 3466 PrintGCDetails && Verbose, true, gclog_or_tty);)
3407 3467 if (UseAdaptiveSizePolicy) {
3408 3468 size_policy()->checkpoint_roots_initial_begin();
3409 3469 }
3410 3470
3411 3471 // Reset all the PLAB chunk arrays if necessary.
3412 3472 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3413 3473 reset_survivor_plab_arrays();
3414 3474 }
3415 3475
3416 3476 ResourceMark rm;
3417 3477 HandleMark hm;
3418 3478
3419 3479 FalseClosure falseClosure;
3420 3480 // In the case of a synchronous collection, we will elide the
3421 3481 // remark step, so it's important to catch all the nmethod oops
3422 3482 // in this step; hence the last argument to the constrcutor below.
3423 3483 MarkRefsIntoClosure notOlder(_span, &_markBitMap, !asynch /* nmethods */);
3424 3484 GenCollectedHeap* gch = GenCollectedHeap::heap();
3425 3485
3426 3486 verify_work_stacks_empty();
3427 3487 verify_overflow_empty();
3428 3488
3429 3489 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3430 3490 // Update the saved marks which may affect the root scans.
3431 3491 gch->save_marks();
3432 3492
3433 3493 // weak reference processing has not started yet.
3434 3494 ref_processor()->set_enqueuing_is_done(false);
3435 3495
3436 3496 {
3437 3497 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3438 3498 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3439 3499 gch->gen_process_strong_roots(_cmsGen->level(),
3440 3500 true, // younger gens are roots
3441 3501 true, // collecting perm gen
3442 3502 SharedHeap::ScanningOption(roots_scanning_options()),
3443 3503 NULL, ¬Older);
3444 3504 }
3445 3505
3446 3506 // Clear mod-union table; it will be dirtied in the prologue of
3447 3507 // CMS generation per each younger generation collection.
3448 3508
3449 3509 assert(_modUnionTable.isAllClear(),
3450 3510 "Was cleared in most recent final checkpoint phase"
3451 3511 " or no bits are set in the gc_prologue before the start of the next "
3452 3512 "subsequent marking phase.");
3453 3513
3454 3514 // Temporarily disabled, since pre/post-consumption closures don't
3455 3515 // care about precleaned cards
3456 3516 #if 0
3457 3517 {
3458 3518 MemRegion mr = MemRegion((HeapWord*)_virtual_space.low(),
3459 3519 (HeapWord*)_virtual_space.high());
3460 3520 _ct->ct_bs()->preclean_dirty_cards(mr);
3461 3521 }
3462 3522 #endif
3463 3523
3464 3524 // Save the end of the used_region of the constituent generations
3465 3525 // to be used to limit the extent of sweep in each generation.
3466 3526 save_sweep_limits();
3467 3527 if (UseAdaptiveSizePolicy) {
3468 3528 size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3469 3529 }
3470 3530 verify_overflow_empty();
3471 3531 }
3472 3532
3473 3533 bool CMSCollector::markFromRoots(bool asynch) {
3474 3534 // we might be tempted to assert that:
3475 3535 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3476 3536 // "inconsistent argument?");
3477 3537 // However that wouldn't be right, because it's possible that
3478 3538 // a safepoint is indeed in progress as a younger generation
3479 3539 // stop-the-world GC happens even as we mark in this generation.
3480 3540 assert(_collectorState == Marking, "inconsistent state?");
3481 3541 check_correct_thread_executing();
3482 3542 verify_overflow_empty();
3483 3543
3484 3544 bool res;
3485 3545 if (asynch) {
3486 3546
3487 3547 // Start the timers for adaptive size policy for the concurrent phases
3488 3548 // Do it here so that the foreground MS can use the concurrent
3489 3549 // timer since a foreground MS might has the sweep done concurrently
3490 3550 // or STW.
3491 3551 if (UseAdaptiveSizePolicy) {
3492 3552 size_policy()->concurrent_marking_begin();
3493 3553 }
3494 3554
3495 3555 // Weak ref discovery note: We may be discovering weak
3496 3556 // refs in this generation concurrent (but interleaved) with
3497 3557 // weak ref discovery by a younger generation collector.
3498 3558
3499 3559 CMSTokenSyncWithLocks ts(true, bitMapLock());
3500 3560 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3501 3561 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3502 3562 res = markFromRootsWork(asynch);
3503 3563 if (res) {
3504 3564 _collectorState = Precleaning;
3505 3565 } else { // We failed and a foreground collection wants to take over
3506 3566 assert(_foregroundGCIsActive, "internal state inconsistency");
3507 3567 assert(_restart_addr == NULL, "foreground will restart from scratch");
3508 3568 if (PrintGCDetails) {
3509 3569 gclog_or_tty->print_cr("bailing out to foreground collection");
3510 3570 }
3511 3571 }
3512 3572 if (UseAdaptiveSizePolicy) {
3513 3573 size_policy()->concurrent_marking_end();
3514 3574 }
3515 3575 } else {
3516 3576 assert(SafepointSynchronize::is_at_safepoint(),
3517 3577 "inconsistent with asynch == false");
3518 3578 if (UseAdaptiveSizePolicy) {
3519 3579 size_policy()->ms_collection_marking_begin();
3520 3580 }
3521 3581 // already have locks
3522 3582 res = markFromRootsWork(asynch);
3523 3583 _collectorState = FinalMarking;
3524 3584 if (UseAdaptiveSizePolicy) {
3525 3585 GenCollectedHeap* gch = GenCollectedHeap::heap();
3526 3586 size_policy()->ms_collection_marking_end(gch->gc_cause());
3527 3587 }
3528 3588 }
3529 3589 verify_overflow_empty();
3530 3590 return res;
3531 3591 }
3532 3592
3533 3593 bool CMSCollector::markFromRootsWork(bool asynch) {
3534 3594 // iterate over marked bits in bit map, doing a full scan and mark
3535 3595 // from these roots using the following algorithm:
3536 3596 // . if oop is to the right of the current scan pointer,
3537 3597 // mark corresponding bit (we'll process it later)
3538 3598 // . else (oop is to left of current scan pointer)
3539 3599 // push oop on marking stack
3540 3600 // . drain the marking stack
3541 3601
3542 3602 // Note that when we do a marking step we need to hold the
3543 3603 // bit map lock -- recall that direct allocation (by mutators)
3544 3604 // and promotion (by younger generation collectors) is also
3545 3605 // marking the bit map. [the so-called allocate live policy.]
3546 3606 // Because the implementation of bit map marking is not
3547 3607 // robust wrt simultaneous marking of bits in the same word,
3548 3608 // we need to make sure that there is no such interference
3549 3609 // between concurrent such updates.
3550 3610
3551 3611 // already have locks
3552 3612 assert_lock_strong(bitMapLock());
3553 3613
3554 3614 // Clear the revisit stack, just in case there are any
3555 3615 // obsolete contents from a short-circuited previous CMS cycle.
3556 3616 _revisitStack.reset();
3557 3617 verify_work_stacks_empty();
3558 3618 verify_overflow_empty();
3559 3619 assert(_revisitStack.isEmpty(), "tabula rasa");
3560 3620
3561 3621 bool result = false;
3562 3622 if (CMSConcurrentMTEnabled && ParallelCMSThreads > 0) {
3563 3623 result = do_marking_mt(asynch);
3564 3624 } else {
3565 3625 result = do_marking_st(asynch);
3566 3626 }
3567 3627 return result;
3568 3628 }
3569 3629
3570 3630 // Forward decl
3571 3631 class CMSConcMarkingTask;
3572 3632
3573 3633 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3574 3634 CMSCollector* _collector;
3575 3635 CMSConcMarkingTask* _task;
3576 3636 bool _yield;
3577 3637 protected:
3578 3638 virtual void yield();
3579 3639 public:
3580 3640 // "n_threads" is the number of threads to be terminated.
3581 3641 // "queue_set" is a set of work queues of other threads.
3582 3642 // "collector" is the CMS collector associated with this task terminator.
3583 3643 // "yield" indicates whether we need the gang as a whole to yield.
3584 3644 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set,
3585 3645 CMSCollector* collector, bool yield) :
3586 3646 ParallelTaskTerminator(n_threads, queue_set),
3587 3647 _collector(collector),
3588 3648 _yield(yield) { }
3589 3649
3590 3650 void set_task(CMSConcMarkingTask* task) {
3591 3651 _task = task;
3592 3652 }
3593 3653 };
3594 3654
3595 3655 // MT Concurrent Marking Task
3596 3656 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3597 3657 CMSCollector* _collector;
3598 3658 YieldingFlexibleWorkGang* _workers; // the whole gang
3599 3659 int _n_workers; // requested/desired # workers
3600 3660 bool _asynch;
3601 3661 bool _result;
3602 3662 CompactibleFreeListSpace* _cms_space;
3603 3663 CompactibleFreeListSpace* _perm_space;
3604 3664 HeapWord* _global_finger;
3605 3665
3606 3666 // Exposed here for yielding support
3607 3667 Mutex* const _bit_map_lock;
3608 3668
3609 3669 // The per thread work queues, available here for stealing
3610 3670 OopTaskQueueSet* _task_queues;
3611 3671 CMSConcMarkingTerminator _term;
3612 3672
3613 3673 public:
3614 3674 CMSConcMarkingTask(CMSCollector* collector,
3615 3675 CompactibleFreeListSpace* cms_space,
3616 3676 CompactibleFreeListSpace* perm_space,
3617 3677 bool asynch, int n_workers,
3618 3678 YieldingFlexibleWorkGang* workers,
3619 3679 OopTaskQueueSet* task_queues):
3620 3680 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3621 3681 _collector(collector),
3622 3682 _cms_space(cms_space),
3623 3683 _perm_space(perm_space),
3624 3684 _asynch(asynch), _n_workers(n_workers), _result(true),
3625 3685 _workers(workers), _task_queues(task_queues),
3626 3686 _term(n_workers, task_queues, _collector, asynch),
3627 3687 _bit_map_lock(collector->bitMapLock())
3628 3688 {
3629 3689 assert(n_workers <= workers->total_workers(),
3630 3690 "Else termination won't work correctly today"); // XXX FIX ME!
3631 3691 _requested_size = n_workers;
3632 3692 _term.set_task(this);
3633 3693 assert(_cms_space->bottom() < _perm_space->bottom(),
3634 3694 "Finger incorrectly initialized below");
3635 3695 _global_finger = _cms_space->bottom();
3636 3696 }
3637 3697
3638 3698
3639 3699 OopTaskQueueSet* task_queues() { return _task_queues; }
3640 3700
3641 3701 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3642 3702
3643 3703 HeapWord** global_finger_addr() { return &_global_finger; }
3644 3704
3645 3705 CMSConcMarkingTerminator* terminator() { return &_term; }
3646 3706
3647 3707 void work(int i);
3648 3708
3649 3709 virtual void coordinator_yield(); // stuff done by coordinator
3650 3710 bool result() { return _result; }
3651 3711
3652 3712 void reset(HeapWord* ra) {
3653 3713 _term.reset_for_reuse();
3654 3714 }
3655 3715
3656 3716 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3657 3717 OopTaskQueue* work_q);
3658 3718
3659 3719 private:
3660 3720 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3661 3721 void do_work_steal(int i);
3662 3722 void bump_global_finger(HeapWord* f);
3663 3723 };
3664 3724
3665 3725 void CMSConcMarkingTerminator::yield() {
3666 3726 if (ConcurrentMarkSweepThread::should_yield() &&
3667 3727 !_collector->foregroundGCIsActive() &&
3668 3728 _yield) {
3669 3729 _task->yield();
3670 3730 } else {
3671 3731 ParallelTaskTerminator::yield();
3672 3732 }
3673 3733 }
3674 3734
3675 3735 ////////////////////////////////////////////////////////////////
3676 3736 // Concurrent Marking Algorithm Sketch
3677 3737 ////////////////////////////////////////////////////////////////
3678 3738 // Until all tasks exhausted (both spaces):
3679 3739 // -- claim next available chunk
3680 3740 // -- bump global finger via CAS
3681 3741 // -- find first object that starts in this chunk
3682 3742 // and start scanning bitmap from that position
3683 3743 // -- scan marked objects for oops
3684 3744 // -- CAS-mark target, and if successful:
3685 3745 // . if target oop is above global finger (volatile read)
3686 3746 // nothing to do
3687 3747 // . if target oop is in chunk and above local finger
3688 3748 // then nothing to do
3689 3749 // . else push on work-queue
3690 3750 // -- Deal with possible overflow issues:
3691 3751 // . local work-queue overflow causes stuff to be pushed on
3692 3752 // global (common) overflow queue
3693 3753 // . always first empty local work queue
3694 3754 // . then get a batch of oops from global work queue if any
3695 3755 // . then do work stealing
3696 3756 // -- When all tasks claimed (both spaces)
3697 3757 // and local work queue empty,
3698 3758 // then in a loop do:
3699 3759 // . check global overflow stack; steal a batch of oops and trace
3700 3760 // . try to steal from other threads oif GOS is empty
3701 3761 // . if neither is available, offer termination
3702 3762 // -- Terminate and return result
3703 3763 //
3704 3764 void CMSConcMarkingTask::work(int i) {
3705 3765 elapsedTimer _timer;
3706 3766 ResourceMark rm;
3707 3767 HandleMark hm;
3708 3768
3709 3769 DEBUG_ONLY(_collector->verify_overflow_empty();)
3710 3770
3711 3771 // Before we begin work, our work queue should be empty
3712 3772 assert(work_queue(i)->size() == 0, "Expected to be empty");
3713 3773 // Scan the bitmap covering _cms_space, tracing through grey objects.
3714 3774 _timer.start();
3715 3775 do_scan_and_mark(i, _cms_space);
3716 3776 _timer.stop();
3717 3777 if (PrintCMSStatistics != 0) {
3718 3778 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3719 3779 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3720 3780 }
3721 3781
3722 3782 // ... do the same for the _perm_space
3723 3783 _timer.reset();
3724 3784 _timer.start();
3725 3785 do_scan_and_mark(i, _perm_space);
3726 3786 _timer.stop();
3727 3787 if (PrintCMSStatistics != 0) {
3728 3788 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec",
3729 3789 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3730 3790 }
3731 3791
3732 3792 // ... do work stealing
3733 3793 _timer.reset();
3734 3794 _timer.start();
3735 3795 do_work_steal(i);
3736 3796 _timer.stop();
3737 3797 if (PrintCMSStatistics != 0) {
3738 3798 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3739 3799 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3740 3800 }
3741 3801 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3742 3802 assert(work_queue(i)->size() == 0, "Should have been emptied");
3743 3803 // Note that under the current task protocol, the
3744 3804 // following assertion is true even of the spaces
3745 3805 // expanded since the completion of the concurrent
3746 3806 // marking. XXX This will likely change under a strict
3747 3807 // ABORT semantics.
3748 3808 assert(_global_finger > _cms_space->end() &&
3749 3809 _global_finger >= _perm_space->end(),
3750 3810 "All tasks have been completed");
3751 3811 DEBUG_ONLY(_collector->verify_overflow_empty();)
3752 3812 }
3753 3813
3754 3814 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3755 3815 HeapWord* read = _global_finger;
3756 3816 HeapWord* cur = read;
3757 3817 while (f > read) {
3758 3818 cur = read;
3759 3819 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3760 3820 if (cur == read) {
3761 3821 // our cas succeeded
3762 3822 assert(_global_finger >= f, "protocol consistency");
3763 3823 break;
3764 3824 }
3765 3825 }
3766 3826 }
3767 3827
3768 3828 // This is really inefficient, and should be redone by
3769 3829 // using (not yet available) block-read and -write interfaces to the
3770 3830 // stack and the work_queue. XXX FIX ME !!!
3771 3831 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3772 3832 OopTaskQueue* work_q) {
3773 3833 // Fast lock-free check
3774 3834 if (ovflw_stk->length() == 0) {
3775 3835 return false;
3776 3836 }
3777 3837 assert(work_q->size() == 0, "Shouldn't steal");
3778 3838 MutexLockerEx ml(ovflw_stk->par_lock(),
3779 3839 Mutex::_no_safepoint_check_flag);
3780 3840 // Grab up to 1/4 the size of the work queue
3781 3841 size_t num = MIN2((size_t)work_q->max_elems()/4,
3782 3842 (size_t)ParGCDesiredObjsFromOverflowList);
3783 3843 num = MIN2(num, ovflw_stk->length());
3784 3844 for (int i = (int) num; i > 0; i--) {
3785 3845 oop cur = ovflw_stk->pop();
3786 3846 assert(cur != NULL, "Counted wrong?");
3787 3847 work_q->push(cur);
3788 3848 }
3789 3849 return num > 0;
3790 3850 }
3791 3851
3792 3852 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3793 3853 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3794 3854 int n_tasks = pst->n_tasks();
3795 3855 // We allow that there may be no tasks to do here because
3796 3856 // we are restarting after a stack overflow.
3797 3857 assert(pst->valid() || n_tasks == 0, "Uninitializd use?");
3798 3858 int nth_task = 0;
3799 3859
3800 3860 HeapWord* start = sp->bottom();
3801 3861 size_t chunk_size = sp->marking_task_size();
3802 3862 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3803 3863 // Having claimed the nth task in this space,
3804 3864 // compute the chunk that it corresponds to:
3805 3865 MemRegion span = MemRegion(start + nth_task*chunk_size,
3806 3866 start + (nth_task+1)*chunk_size);
3807 3867 // Try and bump the global finger via a CAS;
3808 3868 // note that we need to do the global finger bump
3809 3869 // _before_ taking the intersection below, because
3810 3870 // the task corresponding to that region will be
3811 3871 // deemed done even if the used_region() expands
3812 3872 // because of allocation -- as it almost certainly will
3813 3873 // during start-up while the threads yield in the
3814 3874 // closure below.
3815 3875 HeapWord* finger = span.end();
3816 3876 bump_global_finger(finger); // atomically
3817 3877 // There are null tasks here corresponding to chunks
3818 3878 // beyond the "top" address of the space.
3819 3879 span = span.intersection(sp->used_region());
3820 3880 if (!span.is_empty()) { // Non-null task
3821 3881 // We want to skip the first object because
3822 3882 // the protocol is to scan any object in its entirety
3823 3883 // that _starts_ in this span; a fortiori, any
3824 3884 // object starting in an earlier span is scanned
3825 3885 // as part of an earlier claimed task.
3826 3886 // Below we use the "careful" version of block_start
3827 3887 // so we do not try to navigate uninitialized objects.
3828 3888 HeapWord* prev_obj = sp->block_start_careful(span.start());
3829 3889 // Below we use a variant of block_size that uses the
3830 3890 // Printezis bits to avoid waiting for allocated
3831 3891 // objects to become initialized/parsable.
3832 3892 while (prev_obj < span.start()) {
3833 3893 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3834 3894 if (sz > 0) {
3835 3895 prev_obj += sz;
3836 3896 } else {
3837 3897 // In this case we may end up doing a bit of redundant
3838 3898 // scanning, but that appears unavoidable, short of
3839 3899 // locking the free list locks; see bug 6324141.
3840 3900 break;
3841 3901 }
3842 3902 }
3843 3903 if (prev_obj < span.end()) {
3844 3904 MemRegion my_span = MemRegion(prev_obj, span.end());
3845 3905 // Do the marking work within a non-empty span --
3846 3906 // the last argument to the constructor indicates whether the
3847 3907 // iteration should be incremental with periodic yields.
3848 3908 Par_MarkFromRootsClosure cl(this, _collector, my_span,
3849 3909 &_collector->_markBitMap,
3850 3910 work_queue(i),
3851 3911 &_collector->_markStack,
3852 3912 &_collector->_revisitStack,
3853 3913 _asynch);
3854 3914 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3855 3915 } // else nothing to do for this task
3856 3916 } // else nothing to do for this task
3857 3917 }
3858 3918 // We'd be tempted to assert here that since there are no
3859 3919 // more tasks left to claim in this space, the global_finger
3860 3920 // must exceed space->top() and a fortiori space->end(). However,
3861 3921 // that would not quite be correct because the bumping of
3862 3922 // global_finger occurs strictly after the claiming of a task,
3863 3923 // so by the time we reach here the global finger may not yet
3864 3924 // have been bumped up by the thread that claimed the last
3865 3925 // task.
3866 3926 pst->all_tasks_completed();
3867 3927 }
3868 3928
3869 3929 class Par_ConcMarkingClosure: public OopClosure {
3870 3930 CMSCollector* _collector;
3871 3931 MemRegion _span;
3872 3932 CMSBitMap* _bit_map;
3873 3933 CMSMarkStack* _overflow_stack;
3874 3934 CMSMarkStack* _revisit_stack; // XXXXXX Check proper use
3875 3935 OopTaskQueue* _work_queue;
3876 3936
3877 3937 public:
3878 3938 Par_ConcMarkingClosure(CMSCollector* collector, OopTaskQueue* work_queue,
3879 3939 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3880 3940 _collector(collector),
3881 3941 _span(_collector->_span),
3882 3942 _work_queue(work_queue),
3883 3943 _bit_map(bit_map),
3884 3944 _overflow_stack(overflow_stack) { } // need to initialize revisit stack etc.
3885 3945
3886 3946 void do_oop(oop* p);
3887 3947 void trim_queue(size_t max);
3888 3948 void handle_stack_overflow(HeapWord* lost);
3889 3949 };
3890 3950
3891 3951 // Grey object rescan during work stealing phase --
3892 3952 // the salient assumption here is that stolen oops must
3893 3953 // always be initialized, so we do not need to check for
3894 3954 // uninitialized objects before scanning here.
3895 3955 void Par_ConcMarkingClosure::do_oop(oop* p) {
3896 3956 oop this_oop = *p;
3897 3957 assert(this_oop->is_oop_or_null(),
3898 3958 "expected an oop or NULL");
3899 3959 HeapWord* addr = (HeapWord*)this_oop;
3900 3960 // Check if oop points into the CMS generation
3901 3961 // and is not marked
3902 3962 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3903 3963 // a white object ...
3904 3964 // If we manage to "claim" the object, by being the
3905 3965 // first thread to mark it, then we push it on our
3906 3966 // marking stack
3907 3967 if (_bit_map->par_mark(addr)) { // ... now grey
3908 3968 // push on work queue (grey set)
3909 3969 bool simulate_overflow = false;
3910 3970 NOT_PRODUCT(
3911 3971 if (CMSMarkStackOverflowALot &&
3912 3972 _collector->simulate_overflow()) {
3913 3973 // simulate a stack overflow
3914 3974 simulate_overflow = true;
3915 3975 }
3916 3976 )
3917 3977 if (simulate_overflow ||
3918 3978 !(_work_queue->push(this_oop) || _overflow_stack->par_push(this_oop))) {
3919 3979 // stack overflow
3920 3980 if (PrintCMSStatistics != 0) {
3921 3981 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
3922 3982 SIZE_FORMAT, _overflow_stack->capacity());
3923 3983 }
3924 3984 // We cannot assert that the overflow stack is full because
3925 3985 // it may have been emptied since.
3926 3986 assert(simulate_overflow ||
3927 3987 _work_queue->size() == _work_queue->max_elems(),
3928 3988 "Else push should have succeeded");
3929 3989 handle_stack_overflow(addr);
3930 3990 }
3931 3991 } // Else, some other thread got there first
3932 3992 }
3933 3993 }
3934 3994
3935 3995 void Par_ConcMarkingClosure::trim_queue(size_t max) {
3936 3996 while (_work_queue->size() > max) {
3937 3997 oop new_oop;
3938 3998 if (_work_queue->pop_local(new_oop)) {
3939 3999 assert(new_oop->is_oop(), "Should be an oop");
3940 4000 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3941 4001 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3942 4002 assert(new_oop->is_parsable(), "Should be parsable");
3943 4003 new_oop->oop_iterate(this); // do_oop() above
3944 4004 }
3945 4005 }
3946 4006 }
3947 4007
3948 4008 // Upon stack overflow, we discard (part of) the stack,
3949 4009 // remembering the least address amongst those discarded
3950 4010 // in CMSCollector's _restart_address.
3951 4011 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3952 4012 // We need to do this under a mutex to prevent other
3953 4013 // workers from interfering with the expansion below.
3954 4014 MutexLockerEx ml(_overflow_stack->par_lock(),
3955 4015 Mutex::_no_safepoint_check_flag);
3956 4016 // Remember the least grey address discarded
3957 4017 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3958 4018 _collector->lower_restart_addr(ra);
3959 4019 _overflow_stack->reset(); // discard stack contents
3960 4020 _overflow_stack->expand(); // expand the stack if possible
3961 4021 }
3962 4022
3963 4023
3964 4024 void CMSConcMarkingTask::do_work_steal(int i) {
3965 4025 OopTaskQueue* work_q = work_queue(i);
3966 4026 oop obj_to_scan;
3967 4027 CMSBitMap* bm = &(_collector->_markBitMap);
3968 4028 CMSMarkStack* ovflw = &(_collector->_markStack);
3969 4029 int* seed = _collector->hash_seed(i);
3970 4030 Par_ConcMarkingClosure cl(_collector, work_q, bm, ovflw);
3971 4031 while (true) {
3972 4032 cl.trim_queue(0);
3973 4033 assert(work_q->size() == 0, "Should have been emptied above");
3974 4034 if (get_work_from_overflow_stack(ovflw, work_q)) {
3975 4035 // Can't assert below because the work obtained from the
3976 4036 // overflow stack may already have been stolen from us.
3977 4037 // assert(work_q->size() > 0, "Work from overflow stack");
3978 4038 continue;
3979 4039 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3980 4040 assert(obj_to_scan->is_oop(), "Should be an oop");
3981 4041 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3982 4042 obj_to_scan->oop_iterate(&cl);
3983 4043 } else if (terminator()->offer_termination()) {
3984 4044 assert(work_q->size() == 0, "Impossible!");
3985 4045 break;
3986 4046 }
3987 4047 }
3988 4048 }
3989 4049
3990 4050 // This is run by the CMS (coordinator) thread.
3991 4051 void CMSConcMarkingTask::coordinator_yield() {
3992 4052 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3993 4053 "CMS thread should hold CMS token");
3994 4054
3995 4055 // First give up the locks, then yield, then re-lock
3996 4056 // We should probably use a constructor/destructor idiom to
3997 4057 // do this unlock/lock or modify the MutexUnlocker class to
3998 4058 // serve our purpose. XXX
3999 4059 assert_lock_strong(_bit_map_lock);
4000 4060 _bit_map_lock->unlock();
4001 4061 ConcurrentMarkSweepThread::desynchronize(true);
4002 4062 ConcurrentMarkSweepThread::acknowledge_yield_request();
4003 4063 _collector->stopTimer();
4004 4064 if (PrintCMSStatistics != 0) {
4005 4065 _collector->incrementYields();
4006 4066 }
4007 4067 _collector->icms_wait();
4008 4068
4009 4069 // It is possible for whichever thread initiated the yield request
4010 4070 // not to get a chance to wake up and take the bitmap lock between
4011 4071 // this thread releasing it and reacquiring it. So, while the
4012 4072 // should_yield() flag is on, let's sleep for a bit to give the
4013 4073 // other thread a chance to wake up. The limit imposed on the number
4014 4074 // of iterations is defensive, to avoid any unforseen circumstances
4015 4075 // putting us into an infinite loop. Since it's always been this
4016 4076 // (coordinator_yield()) method that was observed to cause the
4017 4077 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4018 4078 // which is by default non-zero. For the other seven methods that
4019 4079 // also perform the yield operation, as are using a different
4020 4080 // parameter (CMSYieldSleepCount) which is by default zero. This way we
4021 4081 // can enable the sleeping for those methods too, if necessary.
4022 4082 // See 6442774.
4023 4083 //
4024 4084 // We really need to reconsider the synchronization between the GC
4025 4085 // thread and the yield-requesting threads in the future and we
4026 4086 // should really use wait/notify, which is the recommended
4027 4087 // way of doing this type of interaction. Additionally, we should
4028 4088 // consolidate the eight methods that do the yield operation and they
4029 4089 // are almost identical into one for better maintenability and
4030 4090 // readability. See 6445193.
4031 4091 //
4032 4092 // Tony 2006.06.29
4033 4093 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4034 4094 ConcurrentMarkSweepThread::should_yield() &&
4035 4095 !CMSCollector::foregroundGCIsActive(); ++i) {
4036 4096 os::sleep(Thread::current(), 1, false);
4037 4097 ConcurrentMarkSweepThread::acknowledge_yield_request();
4038 4098 }
4039 4099
4040 4100 ConcurrentMarkSweepThread::synchronize(true);
4041 4101 _bit_map_lock->lock_without_safepoint_check();
4042 4102 _collector->startTimer();
4043 4103 }
4044 4104
4045 4105 bool CMSCollector::do_marking_mt(bool asynch) {
4046 4106 assert(ParallelCMSThreads > 0 && conc_workers() != NULL, "precondition");
4047 4107 // In the future this would be determined ergonomically, based
4048 4108 // on #cpu's, # active mutator threads (and load), and mutation rate.
4049 4109 int num_workers = ParallelCMSThreads;
4050 4110
4051 4111 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4052 4112 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
4053 4113
4054 4114 CMSConcMarkingTask tsk(this, cms_space, perm_space,
4055 4115 asynch, num_workers /* number requested XXX */,
4056 4116 conc_workers(), task_queues());
4057 4117
4058 4118 // Since the actual number of workers we get may be different
4059 4119 // from the number we requested above, do we need to do anything different
4060 4120 // below? In particular, may be we need to subclass the SequantialSubTasksDone
4061 4121 // class?? XXX
4062 4122 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4063 4123 perm_space->initialize_sequential_subtasks_for_marking(num_workers);
4064 4124
4065 4125 // Refs discovery is already non-atomic.
4066 4126 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4067 4127 // Mutate the Refs discovery so it is MT during the
4068 4128 // multi-threaded marking phase.
4069 4129 ReferenceProcessorMTMutator mt(ref_processor(), num_workers > 1);
4070 4130
4071 4131 conc_workers()->start_task(&tsk);
4072 4132 while (tsk.yielded()) {
4073 4133 tsk.coordinator_yield();
4074 4134 conc_workers()->continue_task(&tsk);
4075 4135 }
4076 4136 // If the task was aborted, _restart_addr will be non-NULL
4077 4137 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4078 4138 while (_restart_addr != NULL) {
4079 4139 // XXX For now we do not make use of ABORTED state and have not
4080 4140 // yet implemented the right abort semantics (even in the original
4081 4141 // single-threaded CMS case). That needs some more investigation
4082 4142 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4083 4143 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4084 4144 // If _restart_addr is non-NULL, a marking stack overflow
4085 4145 // occured; we need to do a fresh marking iteration from the
4086 4146 // indicated restart address.
4087 4147 if (_foregroundGCIsActive && asynch) {
4088 4148 // We may be running into repeated stack overflows, having
4089 4149 // reached the limit of the stack size, while making very
4090 4150 // slow forward progress. It may be best to bail out and
4091 4151 // let the foreground collector do its job.
4092 4152 // Clear _restart_addr, so that foreground GC
4093 4153 // works from scratch. This avoids the headache of
4094 4154 // a "rescan" which would otherwise be needed because
4095 4155 // of the dirty mod union table & card table.
4096 4156 _restart_addr = NULL;
4097 4157 return false;
4098 4158 }
4099 4159 // Adjust the task to restart from _restart_addr
4100 4160 tsk.reset(_restart_addr);
4101 4161 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4102 4162 _restart_addr);
4103 4163 perm_space->initialize_sequential_subtasks_for_marking(num_workers,
4104 4164 _restart_addr);
4105 4165 _restart_addr = NULL;
4106 4166 // Get the workers going again
4107 4167 conc_workers()->start_task(&tsk);
4108 4168 while (tsk.yielded()) {
4109 4169 tsk.coordinator_yield();
4110 4170 conc_workers()->continue_task(&tsk);
4111 4171 }
4112 4172 }
4113 4173 assert(tsk.completed(), "Inconsistency");
4114 4174 assert(tsk.result() == true, "Inconsistency");
4115 4175 return true;
4116 4176 }
4117 4177
4118 4178 bool CMSCollector::do_marking_st(bool asynch) {
4119 4179 ResourceMark rm;
4120 4180 HandleMark hm;
4121 4181
4122 4182 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4123 4183 &_markStack, &_revisitStack, CMSYield && asynch);
4124 4184 // the last argument to iterate indicates whether the iteration
4125 4185 // should be incremental with periodic yields.
4126 4186 _markBitMap.iterate(&markFromRootsClosure);
4127 4187 // If _restart_addr is non-NULL, a marking stack overflow
4128 4188 // occured; we need to do a fresh iteration from the
4129 4189 // indicated restart address.
4130 4190 while (_restart_addr != NULL) {
4131 4191 if (_foregroundGCIsActive && asynch) {
4132 4192 // We may be running into repeated stack overflows, having
4133 4193 // reached the limit of the stack size, while making very
4134 4194 // slow forward progress. It may be best to bail out and
4135 4195 // let the foreground collector do its job.
4136 4196 // Clear _restart_addr, so that foreground GC
4137 4197 // works from scratch. This avoids the headache of
4138 4198 // a "rescan" which would otherwise be needed because
4139 4199 // of the dirty mod union table & card table.
4140 4200 _restart_addr = NULL;
4141 4201 return false; // indicating failure to complete marking
4142 4202 }
4143 4203 // Deal with stack overflow:
4144 4204 // we restart marking from _restart_addr
4145 4205 HeapWord* ra = _restart_addr;
4146 4206 markFromRootsClosure.reset(ra);
4147 4207 _restart_addr = NULL;
4148 4208 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4149 4209 }
4150 4210 return true;
4151 4211 }
4152 4212
4153 4213 void CMSCollector::preclean() {
4154 4214 check_correct_thread_executing();
4155 4215 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4156 4216 verify_work_stacks_empty();
4157 4217 verify_overflow_empty();
4158 4218 _abort_preclean = false;
4159 4219 if (CMSPrecleaningEnabled) {
4160 4220 _eden_chunk_index = 0;
4161 4221 size_t used = get_eden_used();
4162 4222 size_t capacity = get_eden_capacity();
4163 4223 // Don't start sampling unless we will get sufficiently
4164 4224 // many samples.
4165 4225 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4166 4226 * CMSScheduleRemarkEdenPenetration)) {
4167 4227 _start_sampling = true;
4168 4228 } else {
4169 4229 _start_sampling = false;
4170 4230 }
4171 4231 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4172 4232 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
4173 4233 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4174 4234 }
4175 4235 CMSTokenSync x(true); // is cms thread
4176 4236 if (CMSPrecleaningEnabled) {
4177 4237 sample_eden();
4178 4238 _collectorState = AbortablePreclean;
4179 4239 } else {
4180 4240 _collectorState = FinalMarking;
4181 4241 }
4182 4242 verify_work_stacks_empty();
4183 4243 verify_overflow_empty();
4184 4244 }
4185 4245
4186 4246 // Try and schedule the remark such that young gen
4187 4247 // occupancy is CMSScheduleRemarkEdenPenetration %.
4188 4248 void CMSCollector::abortable_preclean() {
4189 4249 check_correct_thread_executing();
4190 4250 assert(CMSPrecleaningEnabled, "Inconsistent control state");
4191 4251 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4192 4252
4193 4253 // If Eden's current occupancy is below this threshold,
4194 4254 // immediately schedule the remark; else preclean
4195 4255 // past the next scavenge in an effort to
4196 4256 // schedule the pause as described avove. By choosing
4197 4257 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4198 4258 // we will never do an actual abortable preclean cycle.
4199 4259 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4200 4260 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4201 4261 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
4202 4262 // We need more smarts in the abortable preclean
4203 4263 // loop below to deal with cases where allocation
4204 4264 // in young gen is very very slow, and our precleaning
4205 4265 // is running a losing race against a horde of
4206 4266 // mutators intent on flooding us with CMS updates
4207 4267 // (dirty cards).
4208 4268 // One, admittedly dumb, strategy is to give up
4209 4269 // after a certain number of abortable precleaning loops
4210 4270 // or after a certain maximum time. We want to make
4211 4271 // this smarter in the next iteration.
4212 4272 // XXX FIX ME!!! YSR
4213 4273 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4214 4274 while (!(should_abort_preclean() ||
4215 4275 ConcurrentMarkSweepThread::should_terminate())) {
4216 4276 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4217 4277 cumworkdone += workdone;
4218 4278 loops++;
4219 4279 // Voluntarily terminate abortable preclean phase if we have
4220 4280 // been at it for too long.
4221 4281 if ((CMSMaxAbortablePrecleanLoops != 0) &&
4222 4282 loops >= CMSMaxAbortablePrecleanLoops) {
4223 4283 if (PrintGCDetails) {
4224 4284 gclog_or_tty->print(" CMS: abort preclean due to loops ");
4225 4285 }
4226 4286 break;
4227 4287 }
4228 4288 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4229 4289 if (PrintGCDetails) {
4230 4290 gclog_or_tty->print(" CMS: abort preclean due to time ");
4231 4291 }
4232 4292 break;
4233 4293 }
4234 4294 // If we are doing little work each iteration, we should
4235 4295 // take a short break.
4236 4296 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4237 4297 // Sleep for some time, waiting for work to accumulate
4238 4298 stopTimer();
4239 4299 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4240 4300 startTimer();
4241 4301 waited++;
4242 4302 }
4243 4303 }
4244 4304 if (PrintCMSStatistics > 0) {
4245 4305 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4246 4306 loops, waited, cumworkdone);
4247 4307 }
4248 4308 }
4249 4309 CMSTokenSync x(true); // is cms thread
4250 4310 if (_collectorState != Idling) {
4251 4311 assert(_collectorState == AbortablePreclean,
4252 4312 "Spontaneous state transition?");
4253 4313 _collectorState = FinalMarking;
4254 4314 } // Else, a foreground collection completed this CMS cycle.
4255 4315 return;
4256 4316 }
4257 4317
4258 4318 // Respond to an Eden sampling opportunity
4259 4319 void CMSCollector::sample_eden() {
4260 4320 // Make sure a young gc cannot sneak in between our
4261 4321 // reading and recording of a sample.
4262 4322 assert(Thread::current()->is_ConcurrentGC_thread(),
4263 4323 "Only the cms thread may collect Eden samples");
4264 4324 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4265 4325 "Should collect samples while holding CMS token");
4266 4326 if (!_start_sampling) {
4267 4327 return;
4268 4328 }
4269 4329 if (_eden_chunk_array) {
4270 4330 if (_eden_chunk_index < _eden_chunk_capacity) {
4271 4331 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
4272 4332 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4273 4333 "Unexpected state of Eden");
4274 4334 // We'd like to check that what we just sampled is an oop-start address;
4275 4335 // however, we cannot do that here since the object may not yet have been
4276 4336 // initialized. So we'll instead do the check when we _use_ this sample
4277 4337 // later.
4278 4338 if (_eden_chunk_index == 0 ||
4279 4339 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4280 4340 _eden_chunk_array[_eden_chunk_index-1])
4281 4341 >= CMSSamplingGrain)) {
4282 4342 _eden_chunk_index++; // commit sample
4283 4343 }
4284 4344 }
4285 4345 }
4286 4346 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4287 4347 size_t used = get_eden_used();
4288 4348 size_t capacity = get_eden_capacity();
4289 4349 assert(used <= capacity, "Unexpected state of Eden");
4290 4350 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4291 4351 _abort_preclean = true;
4292 4352 }
4293 4353 }
4294 4354 }
4295 4355
4296 4356
4297 4357 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4298 4358 assert(_collectorState == Precleaning ||
4299 4359 _collectorState == AbortablePreclean, "incorrect state");
4300 4360 ResourceMark rm;
4301 4361 HandleMark hm;
4302 4362 // Do one pass of scrubbing the discovered reference lists
4303 4363 // to remove any reference objects with strongly-reachable
4304 4364 // referents.
4305 4365 if (clean_refs) {
4306 4366 ReferenceProcessor* rp = ref_processor();
4307 4367 CMSPrecleanRefsYieldClosure yield_cl(this);
4308 4368 assert(rp->span().equals(_span), "Spans should be equal");
4309 4369 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4310 4370 &_markStack);
4311 4371 CMSDrainMarkingStackClosure complete_trace(this,
4312 4372 _span, &_markBitMap, &_markStack,
4313 4373 &keep_alive);
4314 4374
4315 4375 // We don't want this step to interfere with a young
4316 4376 // collection because we don't want to take CPU
4317 4377 // or memory bandwidth away from the young GC threads
4318 4378 // (which may be as many as there are CPUs).
4319 4379 // Note that we don't need to protect ourselves from
4320 4380 // interference with mutators because they can't
4321 4381 // manipulate the discovered reference lists nor affect
4322 4382 // the computed reachability of the referents, the
4323 4383 // only properties manipulated by the precleaning
4324 4384 // of these reference lists.
4325 4385 stopTimer();
4326 4386 CMSTokenSyncWithLocks x(true /* is cms thread */,
4327 4387 bitMapLock());
4328 4388 startTimer();
4329 4389 sample_eden();
4330 4390 // The following will yield to allow foreground
4331 4391 // collection to proceed promptly. XXX YSR:
4332 4392 // The code in this method may need further
4333 4393 // tweaking for better performance and some restructuring
4334 4394 // for cleaner interfaces.
4335 4395 rp->preclean_discovered_references(
4336 4396 rp->is_alive_non_header(), &keep_alive, &complete_trace,
4337 4397 &yield_cl);
4338 4398 }
4339 4399
4340 4400 if (clean_survivor) { // preclean the active survivor space(s)
4341 4401 assert(_young_gen->kind() == Generation::DefNew ||
4342 4402 _young_gen->kind() == Generation::ParNew ||
4343 4403 _young_gen->kind() == Generation::ASParNew,
4344 4404 "incorrect type for cast");
4345 4405 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4346 4406 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4347 4407 &_markBitMap, &_modUnionTable,
4348 4408 &_markStack, &_revisitStack,
4349 4409 true /* precleaning phase */);
4350 4410 stopTimer();
4351 4411 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4352 4412 bitMapLock());
4353 4413 startTimer();
4354 4414 unsigned int before_count =
4355 4415 GenCollectedHeap::heap()->total_collections();
4356 4416 SurvivorSpacePrecleanClosure
4357 4417 sss_cl(this, _span, &_markBitMap, &_markStack,
4358 4418 &pam_cl, before_count, CMSYield);
4359 4419 dng->from()->object_iterate_careful(&sss_cl);
4360 4420 dng->to()->object_iterate_careful(&sss_cl);
4361 4421 }
4362 4422 MarkRefsIntoAndScanClosure
4363 4423 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4364 4424 &_markStack, &_revisitStack, this, CMSYield,
4365 4425 true /* precleaning phase */);
4366 4426 // CAUTION: The following closure has persistent state that may need to
4367 4427 // be reset upon a decrease in the sequence of addresses it
4368 4428 // processes.
4369 4429 ScanMarkedObjectsAgainCarefullyClosure
4370 4430 smoac_cl(this, _span,
4371 4431 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield);
4372 4432
4373 4433 // Preclean dirty cards in ModUnionTable and CardTable using
4374 4434 // appropriate convergence criterion;
4375 4435 // repeat CMSPrecleanIter times unless we find that
4376 4436 // we are losing.
4377 4437 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4378 4438 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4379 4439 "Bad convergence multiplier");
4380 4440 assert(CMSPrecleanThreshold >= 100,
4381 4441 "Unreasonably low CMSPrecleanThreshold");
4382 4442
4383 4443 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4384 4444 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4385 4445 numIter < CMSPrecleanIter;
4386 4446 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4387 4447 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4388 4448 if (CMSPermGenPrecleaningEnabled) {
4389 4449 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl);
4390 4450 }
4391 4451 if (Verbose && PrintGCDetails) {
4392 4452 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4393 4453 }
4394 4454 // Either there are very few dirty cards, so re-mark
4395 4455 // pause will be small anyway, or our pre-cleaning isn't
4396 4456 // that much faster than the rate at which cards are being
4397 4457 // dirtied, so we might as well stop and re-mark since
4398 4458 // precleaning won't improve our re-mark time by much.
4399 4459 if (curNumCards <= CMSPrecleanThreshold ||
4400 4460 (numIter > 0 &&
4401 4461 (curNumCards * CMSPrecleanDenominator >
4402 4462 lastNumCards * CMSPrecleanNumerator))) {
4403 4463 numIter++;
4404 4464 cumNumCards += curNumCards;
4405 4465 break;
4406 4466 }
4407 4467 }
4408 4468 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4409 4469 if (CMSPermGenPrecleaningEnabled) {
4410 4470 curNumCards += preclean_card_table(_permGen, &smoac_cl);
4411 4471 }
4412 4472 cumNumCards += curNumCards;
4413 4473 if (PrintGCDetails && PrintCMSStatistics != 0) {
4414 4474 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4415 4475 curNumCards, cumNumCards, numIter);
4416 4476 }
4417 4477 return cumNumCards; // as a measure of useful work done
4418 4478 }
4419 4479
4420 4480 // PRECLEANING NOTES:
4421 4481 // Precleaning involves:
4422 4482 // . reading the bits of the modUnionTable and clearing the set bits.
4423 4483 // . For the cards corresponding to the set bits, we scan the
4424 4484 // objects on those cards. This means we need the free_list_lock
4425 4485 // so that we can safely iterate over the CMS space when scanning
4426 4486 // for oops.
4427 4487 // . When we scan the objects, we'll be both reading and setting
4428 4488 // marks in the marking bit map, so we'll need the marking bit map.
4429 4489 // . For protecting _collector_state transitions, we take the CGC_lock.
4430 4490 // Note that any races in the reading of of card table entries by the
4431 4491 // CMS thread on the one hand and the clearing of those entries by the
4432 4492 // VM thread or the setting of those entries by the mutator threads on the
4433 4493 // other are quite benign. However, for efficiency it makes sense to keep
4434 4494 // the VM thread from racing with the CMS thread while the latter is
4435 4495 // dirty card info to the modUnionTable. We therefore also use the
4436 4496 // CGC_lock to protect the reading of the card table and the mod union
4437 4497 // table by the CM thread.
4438 4498 // . We run concurrently with mutator updates, so scanning
4439 4499 // needs to be done carefully -- we should not try to scan
4440 4500 // potentially uninitialized objects.
4441 4501 //
4442 4502 // Locking strategy: While holding the CGC_lock, we scan over and
4443 4503 // reset a maximal dirty range of the mod union / card tables, then lock
4444 4504 // the free_list_lock and bitmap lock to do a full marking, then
4445 4505 // release these locks; and repeat the cycle. This allows for a
4446 4506 // certain amount of fairness in the sharing of these locks between
4447 4507 // the CMS collector on the one hand, and the VM thread and the
4448 4508 // mutators on the other.
4449 4509
4450 4510 // NOTE: preclean_mod_union_table() and preclean_card_table()
4451 4511 // further below are largely identical; if you need to modify
4452 4512 // one of these methods, please check the other method too.
4453 4513
4454 4514 size_t CMSCollector::preclean_mod_union_table(
4455 4515 ConcurrentMarkSweepGeneration* gen,
4456 4516 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4457 4517 verify_work_stacks_empty();
4458 4518 verify_overflow_empty();
4459 4519
4460 4520 // strategy: starting with the first card, accumulate contiguous
4461 4521 // ranges of dirty cards; clear these cards, then scan the region
4462 4522 // covered by these cards.
4463 4523
4464 4524 // Since all of the MUT is committed ahead, we can just use
4465 4525 // that, in case the generations expand while we are precleaning.
4466 4526 // It might also be fine to just use the committed part of the
4467 4527 // generation, but we might potentially miss cards when the
4468 4528 // generation is rapidly expanding while we are in the midst
4469 4529 // of precleaning.
4470 4530 HeapWord* startAddr = gen->reserved().start();
4471 4531 HeapWord* endAddr = gen->reserved().end();
4472 4532
4473 4533 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4474 4534
4475 4535 size_t numDirtyCards, cumNumDirtyCards;
4476 4536 HeapWord *nextAddr, *lastAddr;
4477 4537 for (cumNumDirtyCards = numDirtyCards = 0,
4478 4538 nextAddr = lastAddr = startAddr;
4479 4539 nextAddr < endAddr;
4480 4540 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4481 4541
4482 4542 ResourceMark rm;
4483 4543 HandleMark hm;
4484 4544
4485 4545 MemRegion dirtyRegion;
4486 4546 {
4487 4547 stopTimer();
4488 4548 CMSTokenSync ts(true);
4489 4549 startTimer();
4490 4550 sample_eden();
4491 4551 // Get dirty region starting at nextOffset (inclusive),
4492 4552 // simultaneously clearing it.
4493 4553 dirtyRegion =
4494 4554 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4495 4555 assert(dirtyRegion.start() >= nextAddr,
4496 4556 "returned region inconsistent?");
4497 4557 }
4498 4558 // Remember where the next search should begin.
4499 4559 // The returned region (if non-empty) is a right open interval,
4500 4560 // so lastOffset is obtained from the right end of that
4501 4561 // interval.
4502 4562 lastAddr = dirtyRegion.end();
4503 4563 // Should do something more transparent and less hacky XXX
4504 4564 numDirtyCards =
4505 4565 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4506 4566
4507 4567 // We'll scan the cards in the dirty region (with periodic
4508 4568 // yields for foreground GC as needed).
4509 4569 if (!dirtyRegion.is_empty()) {
4510 4570 assert(numDirtyCards > 0, "consistency check");
4511 4571 HeapWord* stop_point = NULL;
4512 4572 {
4513 4573 stopTimer();
4514 4574 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4515 4575 bitMapLock());
4516 4576 startTimer();
4517 4577 verify_work_stacks_empty();
4518 4578 verify_overflow_empty();
4519 4579 sample_eden();
4520 4580 stop_point =
4521 4581 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4522 4582 }
4523 4583 if (stop_point != NULL) {
4524 4584 // The careful iteration stopped early either because it found an
4525 4585 // uninitialized object, or because we were in the midst of an
4526 4586 // "abortable preclean", which should now be aborted. Redirty
4527 4587 // the bits corresponding to the partially-scanned or unscanned
4528 4588 // cards. We'll either restart at the next block boundary or
4529 4589 // abort the preclean.
4530 4590 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) ||
4531 4591 (_collectorState == AbortablePreclean && should_abort_preclean()),
4532 4592 "Unparsable objects should only be in perm gen.");
4533 4593
4534 4594 stopTimer();
4535 4595 CMSTokenSyncWithLocks ts(true, bitMapLock());
4536 4596 startTimer();
4537 4597 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4538 4598 if (should_abort_preclean()) {
4539 4599 break; // out of preclean loop
4540 4600 } else {
4541 4601 // Compute the next address at which preclean should pick up;
4542 4602 // might need bitMapLock in order to read P-bits.
4543 4603 lastAddr = next_card_start_after_block(stop_point);
4544 4604 }
4545 4605 }
4546 4606 } else {
4547 4607 assert(lastAddr == endAddr, "consistency check");
4548 4608 assert(numDirtyCards == 0, "consistency check");
4549 4609 break;
4550 4610 }
4551 4611 }
4552 4612 verify_work_stacks_empty();
4553 4613 verify_overflow_empty();
4554 4614 return cumNumDirtyCards;
4555 4615 }
4556 4616
4557 4617 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4558 4618 // below are largely identical; if you need to modify
4559 4619 // one of these methods, please check the other method too.
4560 4620
4561 4621 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4562 4622 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4563 4623 // strategy: it's similar to precleamModUnionTable above, in that
4564 4624 // we accumulate contiguous ranges of dirty cards, mark these cards
4565 4625 // precleaned, then scan the region covered by these cards.
4566 4626 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4567 4627 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4568 4628
4569 4629 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4570 4630
4571 4631 size_t numDirtyCards, cumNumDirtyCards;
4572 4632 HeapWord *lastAddr, *nextAddr;
4573 4633
4574 4634 for (cumNumDirtyCards = numDirtyCards = 0,
4575 4635 nextAddr = lastAddr = startAddr;
4576 4636 nextAddr < endAddr;
4577 4637 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4578 4638
4579 4639 ResourceMark rm;
4580 4640 HandleMark hm;
4581 4641
4582 4642 MemRegion dirtyRegion;
4583 4643 {
4584 4644 // See comments in "Precleaning notes" above on why we
4585 4645 // do this locking. XXX Could the locking overheads be
4586 4646 // too high when dirty cards are sparse? [I don't think so.]
4587 4647 stopTimer();
4588 4648 CMSTokenSync x(true); // is cms thread
4589 4649 startTimer();
4590 4650 sample_eden();
4591 4651 // Get and clear dirty region from card table
4592 4652 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_preclean(
4593 4653 MemRegion(nextAddr, endAddr));
4594 4654 assert(dirtyRegion.start() >= nextAddr,
4595 4655 "returned region inconsistent?");
4596 4656 }
4597 4657 lastAddr = dirtyRegion.end();
4598 4658 numDirtyCards =
4599 4659 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4600 4660
4601 4661 if (!dirtyRegion.is_empty()) {
4602 4662 stopTimer();
4603 4663 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4604 4664 startTimer();
4605 4665 sample_eden();
4606 4666 verify_work_stacks_empty();
4607 4667 verify_overflow_empty();
4608 4668 HeapWord* stop_point =
4609 4669 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4610 4670 if (stop_point != NULL) {
4611 4671 // The careful iteration stopped early because it found an
4612 4672 // uninitialized object. Redirty the bits corresponding to the
4613 4673 // partially-scanned or unscanned cards, and start again at the
4614 4674 // next block boundary.
4615 4675 assert(CMSPermGenPrecleaningEnabled ||
4616 4676 (_collectorState == AbortablePreclean && should_abort_preclean()),
4617 4677 "Unparsable objects should only be in perm gen.");
4618 4678 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4619 4679 if (should_abort_preclean()) {
4620 4680 break; // out of preclean loop
4621 4681 } else {
4622 4682 // Compute the next address at which preclean should pick up.
4623 4683 lastAddr = next_card_start_after_block(stop_point);
4624 4684 }
4625 4685 }
4626 4686 } else {
4627 4687 break;
4628 4688 }
4629 4689 }
4630 4690 verify_work_stacks_empty();
4631 4691 verify_overflow_empty();
4632 4692 return cumNumDirtyCards;
4633 4693 }
4634 4694
4635 4695 void CMSCollector::checkpointRootsFinal(bool asynch,
4636 4696 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4637 4697 assert(_collectorState == FinalMarking, "incorrect state transition?");
4638 4698 check_correct_thread_executing();
4639 4699 // world is stopped at this checkpoint
4640 4700 assert(SafepointSynchronize::is_at_safepoint(),
4641 4701 "world should be stopped");
4642 4702 verify_work_stacks_empty();
4643 4703 verify_overflow_empty();
4644 4704
4645 4705 SpecializationStats::clear();
4646 4706 if (PrintGCDetails) {
4647 4707 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4648 4708 _young_gen->used() / K,
4649 4709 _young_gen->capacity() / K);
4650 4710 }
4651 4711 if (asynch) {
4652 4712 if (CMSScavengeBeforeRemark) {
4653 4713 GenCollectedHeap* gch = GenCollectedHeap::heap();
4654 4714 // Temporarily set flag to false, GCH->do_collection will
4655 4715 // expect it to be false and set to true
4656 4716 FlagSetting fl(gch->_is_gc_active, false);
4657 4717 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark",
4658 4718 PrintGCDetails && Verbose, true, gclog_or_tty);)
4659 4719 int level = _cmsGen->level() - 1;
4660 4720 if (level >= 0) {
4661 4721 gch->do_collection(true, // full (i.e. force, see below)
4662 4722 false, // !clear_all_soft_refs
4663 4723 0, // size
4664 4724 false, // is_tlab
4665 4725 level // max_level
4666 4726 );
4667 4727 }
4668 4728 }
4669 4729 FreelistLocker x(this);
4670 4730 MutexLockerEx y(bitMapLock(),
4671 4731 Mutex::_no_safepoint_check_flag);
4672 4732 assert(!init_mark_was_synchronous, "but that's impossible!");
4673 4733 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
4674 4734 } else {
4675 4735 // already have all the locks
4676 4736 checkpointRootsFinalWork(asynch, clear_all_soft_refs,
4677 4737 init_mark_was_synchronous);
4678 4738 }
4679 4739 verify_work_stacks_empty();
4680 4740 verify_overflow_empty();
4681 4741 SpecializationStats::print();
4682 4742 }
4683 4743
4684 4744 void CMSCollector::checkpointRootsFinalWork(bool asynch,
4685 4745 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4686 4746
4687 4747 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);)
4688 4748
4689 4749 assert(haveFreelistLocks(), "must have free list locks");
4690 4750 assert_lock_strong(bitMapLock());
↓ open down ↓ |
1616 lines elided |
↑ open up ↑ |
4691 4751
4692 4752 if (UseAdaptiveSizePolicy) {
4693 4753 size_policy()->checkpoint_roots_final_begin();
4694 4754 }
4695 4755
4696 4756 ResourceMark rm;
4697 4757 HandleMark hm;
4698 4758
4699 4759 GenCollectedHeap* gch = GenCollectedHeap::heap();
4700 4760
4701 - if (cms_should_unload_classes()) {
4761 + if (should_unload_classes()) {
4702 4762 CodeCache::gc_prologue();
4703 4763 }
4704 4764 assert(haveFreelistLocks(), "must have free list locks");
4705 4765 assert_lock_strong(bitMapLock());
4706 4766
4707 4767 if (!init_mark_was_synchronous) {
4708 4768 // We might assume that we need not fill TLAB's when
4709 4769 // CMSScavengeBeforeRemark is set, because we may have just done
4710 4770 // a scavenge which would have filled all TLAB's -- and besides
4711 4771 // Eden would be empty. This however may not always be the case --
4712 4772 // for instance although we asked for a scavenge, it may not have
4713 4773 // happened because of a JNI critical section. We probably need
4714 4774 // a policy for deciding whether we can in that case wait until
4715 4775 // the critical section releases and then do the remark following
4716 4776 // the scavenge, and skip it here. In the absence of that policy,
4717 4777 // or of an indication of whether the scavenge did indeed occur,
4718 4778 // we cannot rely on TLAB's having been filled and must do
4719 4779 // so here just in case a scavenge did not happen.
4720 4780 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4721 4781 // Update the saved marks which may affect the root scans.
4722 4782 gch->save_marks();
4723 4783
4724 4784 {
4725 4785 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4726 4786
4727 4787 // Note on the role of the mod union table:
4728 4788 // Since the marker in "markFromRoots" marks concurrently with
4729 4789 // mutators, it is possible for some reachable objects not to have been
4730 4790 // scanned. For instance, an only reference to an object A was
4731 4791 // placed in object B after the marker scanned B. Unless B is rescanned,
4732 4792 // A would be collected. Such updates to references in marked objects
4733 4793 // are detected via the mod union table which is the set of all cards
4734 4794 // dirtied since the first checkpoint in this GC cycle and prior to
4735 4795 // the most recent young generation GC, minus those cleaned up by the
4736 4796 // concurrent precleaning.
4737 4797 if (CMSParallelRemarkEnabled && ParallelGCThreads > 0) {
4738 4798 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty);
4739 4799 do_remark_parallel();
4740 4800 } else {
4741 4801 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4742 4802 gclog_or_tty);
4743 4803 do_remark_non_parallel();
4744 4804 }
4745 4805 }
4746 4806 } else {
4747 4807 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
4748 4808 // The initial mark was stop-world, so there's no rescanning to
4749 4809 // do; go straight on to the next step below.
4750 4810 }
↓ open down ↓ |
39 lines elided |
↑ open up ↑ |
4751 4811 verify_work_stacks_empty();
4752 4812 verify_overflow_empty();
4753 4813
4754 4814 {
4755 4815 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);)
4756 4816 refProcessingWork(asynch, clear_all_soft_refs);
4757 4817 }
4758 4818 verify_work_stacks_empty();
4759 4819 verify_overflow_empty();
4760 4820
4761 - if (cms_should_unload_classes()) {
4821 + if (should_unload_classes()) {
4762 4822 CodeCache::gc_epilogue();
4763 4823 }
4764 4824
4765 4825 // If we encountered any (marking stack / work queue) overflow
4766 4826 // events during the current CMS cycle, take appropriate
4767 4827 // remedial measures, where possible, so as to try and avoid
4768 4828 // recurrence of that condition.
4769 4829 assert(_markStack.isEmpty(), "No grey objects");
4770 4830 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4771 4831 _ser_kac_ovflw;
4772 4832 if (ser_ovflw > 0) {
4773 4833 if (PrintCMSStatistics != 0) {
4774 4834 gclog_or_tty->print_cr("Marking stack overflow (benign) "
4775 4835 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4776 4836 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4777 4837 _ser_kac_ovflw);
4778 4838 }
4779 4839 _markStack.expand();
4780 4840 _ser_pmc_remark_ovflw = 0;
4781 4841 _ser_pmc_preclean_ovflw = 0;
4782 4842 _ser_kac_ovflw = 0;
4783 4843 }
4784 4844 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4785 4845 if (PrintCMSStatistics != 0) {
4786 4846 gclog_or_tty->print_cr("Work queue overflow (benign) "
4787 4847 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4788 4848 _par_pmc_remark_ovflw, _par_kac_ovflw);
4789 4849 }
4790 4850 _par_pmc_remark_ovflw = 0;
4791 4851 _par_kac_ovflw = 0;
4792 4852 }
4793 4853 if (PrintCMSStatistics != 0) {
4794 4854 if (_markStack._hit_limit > 0) {
4795 4855 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
4796 4856 _markStack._hit_limit);
4797 4857 }
4798 4858 if (_markStack._failed_double > 0) {
4799 4859 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
4800 4860 " current capacity "SIZE_FORMAT,
4801 4861 _markStack._failed_double,
4802 4862 _markStack.capacity());
4803 4863 }
4804 4864 }
4805 4865 _markStack._hit_limit = 0;
4806 4866 _markStack._failed_double = 0;
4807 4867
4808 4868 if ((VerifyAfterGC || VerifyDuringGC) &&
4809 4869 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4810 4870 verify_after_remark();
4811 4871 }
4812 4872
4813 4873 // Change under the freelistLocks.
4814 4874 _collectorState = Sweeping;
4815 4875 // Call isAllClear() under bitMapLock
4816 4876 assert(_modUnionTable.isAllClear(), "Should be clear by end of the"
4817 4877 " final marking");
4818 4878 if (UseAdaptiveSizePolicy) {
4819 4879 size_policy()->checkpoint_roots_final_end(gch->gc_cause());
4820 4880 }
4821 4881 }
4822 4882
4823 4883 // Parallel remark task
4824 4884 class CMSParRemarkTask: public AbstractGangTask {
4825 4885 CMSCollector* _collector;
4826 4886 WorkGang* _workers;
4827 4887 int _n_workers;
4828 4888 CompactibleFreeListSpace* _cms_space;
4829 4889 CompactibleFreeListSpace* _perm_space;
4830 4890
4831 4891 // The per-thread work queues, available here for stealing.
4832 4892 OopTaskQueueSet* _task_queues;
4833 4893 ParallelTaskTerminator _term;
4834 4894
4835 4895 public:
4836 4896 CMSParRemarkTask(CMSCollector* collector,
4837 4897 CompactibleFreeListSpace* cms_space,
4838 4898 CompactibleFreeListSpace* perm_space,
4839 4899 int n_workers, WorkGang* workers,
4840 4900 OopTaskQueueSet* task_queues):
4841 4901 AbstractGangTask("Rescan roots and grey objects in parallel"),
4842 4902 _collector(collector),
4843 4903 _cms_space(cms_space), _perm_space(perm_space),
4844 4904 _n_workers(n_workers),
4845 4905 _workers(workers),
4846 4906 _task_queues(task_queues),
4847 4907 _term(workers->total_workers(), task_queues) { }
4848 4908
4849 4909 OopTaskQueueSet* task_queues() { return _task_queues; }
4850 4910
4851 4911 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4852 4912
4853 4913 ParallelTaskTerminator* terminator() { return &_term; }
4854 4914
4855 4915 void work(int i);
4856 4916
4857 4917 private:
4858 4918 // Work method in support of parallel rescan ... of young gen spaces
4859 4919 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl,
4860 4920 ContiguousSpace* space,
4861 4921 HeapWord** chunk_array, size_t chunk_top);
4862 4922
4863 4923 // ... of dirty cards in old space
4864 4924 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4865 4925 Par_MarkRefsIntoAndScanClosure* cl);
4866 4926
4867 4927 // ... work stealing for the above
4868 4928 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4869 4929 };
4870 4930
4871 4931 void CMSParRemarkTask::work(int i) {
4872 4932 elapsedTimer _timer;
4873 4933 ResourceMark rm;
4874 4934 HandleMark hm;
4875 4935
4876 4936 // ---------- rescan from roots --------------
4877 4937 _timer.start();
4878 4938 GenCollectedHeap* gch = GenCollectedHeap::heap();
4879 4939 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4880 4940 _collector->_span, _collector->ref_processor(),
4881 4941 &(_collector->_markBitMap),
4882 4942 work_queue(i), &(_collector->_revisitStack));
4883 4943
4884 4944 // Rescan young gen roots first since these are likely
4885 4945 // coarsely partitioned and may, on that account, constitute
4886 4946 // the critical path; thus, it's best to start off that
4887 4947 // work first.
4888 4948 // ---------- young gen roots --------------
4889 4949 {
4890 4950 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
4891 4951 EdenSpace* eden_space = dng->eden();
4892 4952 ContiguousSpace* from_space = dng->from();
4893 4953 ContiguousSpace* to_space = dng->to();
4894 4954
4895 4955 HeapWord** eca = _collector->_eden_chunk_array;
4896 4956 size_t ect = _collector->_eden_chunk_index;
4897 4957 HeapWord** sca = _collector->_survivor_chunk_array;
4898 4958 size_t sct = _collector->_survivor_chunk_index;
4899 4959
4900 4960 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4901 4961 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4902 4962
4903 4963 do_young_space_rescan(i, &par_mrias_cl, to_space, NULL, 0);
4904 4964 do_young_space_rescan(i, &par_mrias_cl, from_space, sca, sct);
4905 4965 do_young_space_rescan(i, &par_mrias_cl, eden_space, eca, ect);
4906 4966
4907 4967 _timer.stop();
4908 4968 if (PrintCMSStatistics != 0) {
4909 4969 gclog_or_tty->print_cr(
4910 4970 "Finished young gen rescan work in %dth thread: %3.3f sec",
4911 4971 i, _timer.seconds());
4912 4972 }
4913 4973 }
4914 4974
4915 4975 // ---------- remaining roots --------------
4916 4976 _timer.reset();
4917 4977 _timer.start();
4918 4978 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
4919 4979 false, // yg was scanned above
4920 4980 true, // collecting perm gen
4921 4981 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4922 4982 NULL, &par_mrias_cl);
4923 4983 _timer.stop();
4924 4984 if (PrintCMSStatistics != 0) {
4925 4985 gclog_or_tty->print_cr(
4926 4986 "Finished remaining root rescan work in %dth thread: %3.3f sec",
4927 4987 i, _timer.seconds());
4928 4988 }
4929 4989
4930 4990 // ---------- rescan dirty cards ------------
4931 4991 _timer.reset();
4932 4992 _timer.start();
4933 4993
4934 4994 // Do the rescan tasks for each of the two spaces
4935 4995 // (cms_space and perm_space) in turn.
4936 4996 do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl);
4937 4997 do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl);
4938 4998 _timer.stop();
4939 4999 if (PrintCMSStatistics != 0) {
4940 5000 gclog_or_tty->print_cr(
4941 5001 "Finished dirty card rescan work in %dth thread: %3.3f sec",
4942 5002 i, _timer.seconds());
4943 5003 }
4944 5004
4945 5005 // ---------- steal work from other threads ...
4946 5006 // ---------- ... and drain overflow list.
4947 5007 _timer.reset();
4948 5008 _timer.start();
4949 5009 do_work_steal(i, &par_mrias_cl, _collector->hash_seed(i));
4950 5010 _timer.stop();
4951 5011 if (PrintCMSStatistics != 0) {
4952 5012 gclog_or_tty->print_cr(
4953 5013 "Finished work stealing in %dth thread: %3.3f sec",
4954 5014 i, _timer.seconds());
4955 5015 }
4956 5016 }
4957 5017
4958 5018 void
4959 5019 CMSParRemarkTask::do_young_space_rescan(int i,
4960 5020 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space,
4961 5021 HeapWord** chunk_array, size_t chunk_top) {
4962 5022 // Until all tasks completed:
4963 5023 // . claim an unclaimed task
4964 5024 // . compute region boundaries corresponding to task claimed
4965 5025 // using chunk_array
4966 5026 // . par_oop_iterate(cl) over that region
4967 5027
4968 5028 ResourceMark rm;
4969 5029 HandleMark hm;
4970 5030
4971 5031 SequentialSubTasksDone* pst = space->par_seq_tasks();
4972 5032 assert(pst->valid(), "Uninitialized use?");
4973 5033
4974 5034 int nth_task = 0;
4975 5035 int n_tasks = pst->n_tasks();
4976 5036
4977 5037 HeapWord *start, *end;
4978 5038 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4979 5039 // We claimed task # nth_task; compute its boundaries.
4980 5040 if (chunk_top == 0) { // no samples were taken
4981 5041 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
4982 5042 start = space->bottom();
4983 5043 end = space->top();
4984 5044 } else if (nth_task == 0) {
4985 5045 start = space->bottom();
4986 5046 end = chunk_array[nth_task];
4987 5047 } else if (nth_task < (jint)chunk_top) {
4988 5048 assert(nth_task >= 1, "Control point invariant");
4989 5049 start = chunk_array[nth_task - 1];
4990 5050 end = chunk_array[nth_task];
4991 5051 } else {
4992 5052 assert(nth_task == (jint)chunk_top, "Control point invariant");
4993 5053 start = chunk_array[chunk_top - 1];
4994 5054 end = space->top();
4995 5055 }
4996 5056 MemRegion mr(start, end);
4997 5057 // Verify that mr is in space
4998 5058 assert(mr.is_empty() || space->used_region().contains(mr),
4999 5059 "Should be in space");
5000 5060 // Verify that "start" is an object boundary
5001 5061 assert(mr.is_empty() || oop(mr.start())->is_oop(),
5002 5062 "Should be an oop");
5003 5063 space->par_oop_iterate(mr, cl);
5004 5064 }
5005 5065 pst->all_tasks_completed();
5006 5066 }
5007 5067
5008 5068 void
5009 5069 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5010 5070 CompactibleFreeListSpace* sp, int i,
5011 5071 Par_MarkRefsIntoAndScanClosure* cl) {
5012 5072 // Until all tasks completed:
5013 5073 // . claim an unclaimed task
5014 5074 // . compute region boundaries corresponding to task claimed
5015 5075 // . transfer dirty bits ct->mut for that region
5016 5076 // . apply rescanclosure to dirty mut bits for that region
5017 5077
5018 5078 ResourceMark rm;
5019 5079 HandleMark hm;
5020 5080
5021 5081 OopTaskQueue* work_q = work_queue(i);
5022 5082 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5023 5083 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5024 5084 // CAUTION: This closure has state that persists across calls to
5025 5085 // the work method dirty_range_iterate_clear() in that it has
5026 5086 // imbedded in it a (subtype of) UpwardsObjectClosure. The
5027 5087 // use of that state in the imbedded UpwardsObjectClosure instance
5028 5088 // assumes that the cards are always iterated (even if in parallel
5029 5089 // by several threads) in monotonically increasing order per each
5030 5090 // thread. This is true of the implementation below which picks
5031 5091 // card ranges (chunks) in monotonically increasing order globally
5032 5092 // and, a-fortiori, in monotonically increasing order per thread
5033 5093 // (the latter order being a subsequence of the former).
5034 5094 // If the work code below is ever reorganized into a more chaotic
5035 5095 // work-partitioning form than the current "sequential tasks"
5036 5096 // paradigm, the use of that persistent state will have to be
5037 5097 // revisited and modified appropriately. See also related
5038 5098 // bug 4756801 work on which should examine this code to make
5039 5099 // sure that the changes there do not run counter to the
5040 5100 // assumptions made here and necessary for correctness and
5041 5101 // efficiency. Note also that this code might yield inefficient
5042 5102 // behaviour in the case of very large objects that span one or
5043 5103 // more work chunks. Such objects would potentially be scanned
5044 5104 // several times redundantly. Work on 4756801 should try and
5045 5105 // address that performance anomaly if at all possible. XXX
5046 5106 MemRegion full_span = _collector->_span;
5047 5107 CMSBitMap* bm = &(_collector->_markBitMap); // shared
5048 5108 CMSMarkStack* rs = &(_collector->_revisitStack); // shared
5049 5109 MarkFromDirtyCardsClosure
5050 5110 greyRescanClosure(_collector, full_span, // entire span of interest
5051 5111 sp, bm, work_q, rs, cl);
5052 5112
5053 5113 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5054 5114 assert(pst->valid(), "Uninitialized use?");
5055 5115 int nth_task = 0;
5056 5116 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5057 5117 MemRegion span = sp->used_region();
5058 5118 HeapWord* start_addr = span.start();
5059 5119 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5060 5120 alignment);
5061 5121 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5062 5122 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5063 5123 start_addr, "Check alignment");
5064 5124 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5065 5125 chunk_size, "Check alignment");
5066 5126
5067 5127 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5068 5128 // Having claimed the nth_task, compute corresponding mem-region,
5069 5129 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
5070 5130 // The alignment restriction ensures that we do not need any
5071 5131 // synchronization with other gang-workers while setting or
5072 5132 // clearing bits in thus chunk of the MUT.
5073 5133 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5074 5134 start_addr + (nth_task+1)*chunk_size);
5075 5135 // The last chunk's end might be way beyond end of the
5076 5136 // used region. In that case pull back appropriately.
5077 5137 if (this_span.end() > end_addr) {
5078 5138 this_span.set_end(end_addr);
5079 5139 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5080 5140 }
5081 5141 // Iterate over the dirty cards covering this chunk, marking them
5082 5142 // precleaned, and setting the corresponding bits in the mod union
5083 5143 // table. Since we have been careful to partition at Card and MUT-word
5084 5144 // boundaries no synchronization is needed between parallel threads.
5085 5145 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5086 5146 &modUnionClosure);
5087 5147
5088 5148 // Having transferred these marks into the modUnionTable,
5089 5149 // rescan the marked objects on the dirty cards in the modUnionTable.
5090 5150 // Even if this is at a synchronous collection, the initial marking
5091 5151 // may have been done during an asynchronous collection so there
5092 5152 // may be dirty bits in the mod-union table.
5093 5153 _collector->_modUnionTable.dirty_range_iterate_clear(
5094 5154 this_span, &greyRescanClosure);
5095 5155 _collector->_modUnionTable.verifyNoOneBitsInRange(
5096 5156 this_span.start(),
5097 5157 this_span.end());
5098 5158 }
5099 5159 pst->all_tasks_completed(); // declare that i am done
5100 5160 }
5101 5161
5102 5162 // . see if we can share work_queues with ParNew? XXX
5103 5163 void
5104 5164 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5105 5165 int* seed) {
5106 5166 OopTaskQueue* work_q = work_queue(i);
5107 5167 NOT_PRODUCT(int num_steals = 0;)
5108 5168 oop obj_to_scan;
5109 5169 CMSBitMap* bm = &(_collector->_markBitMap);
5110 5170 size_t num_from_overflow_list =
5111 5171 MIN2((size_t)work_q->max_elems()/4,
5112 5172 (size_t)ParGCDesiredObjsFromOverflowList);
5113 5173
5114 5174 while (true) {
5115 5175 // Completely finish any left over work from (an) earlier round(s)
5116 5176 cl->trim_queue(0);
5117 5177 // Now check if there's any work in the overflow list
5118 5178 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5119 5179 work_q)) {
5120 5180 // found something in global overflow list;
5121 5181 // not yet ready to go stealing work from others.
5122 5182 // We'd like to assert(work_q->size() != 0, ...)
5123 5183 // because we just took work from the overflow list,
5124 5184 // but of course we can't since all of that could have
5125 5185 // been already stolen from us.
5126 5186 // "He giveth and He taketh away."
5127 5187 continue;
5128 5188 }
5129 5189 // Verify that we have no work before we resort to stealing
5130 5190 assert(work_q->size() == 0, "Have work, shouldn't steal");
5131 5191 // Try to steal from other queues that have work
5132 5192 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5133 5193 NOT_PRODUCT(num_steals++;)
5134 5194 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5135 5195 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5136 5196 // Do scanning work
5137 5197 obj_to_scan->oop_iterate(cl);
5138 5198 // Loop around, finish this work, and try to steal some more
5139 5199 } else if (terminator()->offer_termination()) {
5140 5200 break; // nirvana from the infinite cycle
5141 5201 }
5142 5202 }
5143 5203 NOT_PRODUCT(
5144 5204 if (PrintCMSStatistics != 0) {
5145 5205 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5146 5206 }
5147 5207 )
5148 5208 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5149 5209 "Else our work is not yet done");
5150 5210 }
5151 5211
5152 5212 // Return a thread-local PLAB recording array, as appropriate.
5153 5213 void* CMSCollector::get_data_recorder(int thr_num) {
5154 5214 if (_survivor_plab_array != NULL &&
5155 5215 (CMSPLABRecordAlways ||
5156 5216 (_collectorState > Marking && _collectorState < FinalMarking))) {
5157 5217 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5158 5218 ChunkArray* ca = &_survivor_plab_array[thr_num];
5159 5219 ca->reset(); // clear it so that fresh data is recorded
5160 5220 return (void*) ca;
5161 5221 } else {
5162 5222 return NULL;
5163 5223 }
5164 5224 }
5165 5225
5166 5226 // Reset all the thread-local PLAB recording arrays
5167 5227 void CMSCollector::reset_survivor_plab_arrays() {
5168 5228 for (uint i = 0; i < ParallelGCThreads; i++) {
5169 5229 _survivor_plab_array[i].reset();
5170 5230 }
5171 5231 }
5172 5232
5173 5233 // Merge the per-thread plab arrays into the global survivor chunk
5174 5234 // array which will provide the partitioning of the survivor space
5175 5235 // for CMS rescan.
5176 5236 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv) {
5177 5237 assert(_survivor_plab_array != NULL, "Error");
5178 5238 assert(_survivor_chunk_array != NULL, "Error");
5179 5239 assert(_collectorState == FinalMarking, "Error");
5180 5240 for (uint j = 0; j < ParallelGCThreads; j++) {
5181 5241 _cursor[j] = 0;
5182 5242 }
5183 5243 HeapWord* top = surv->top();
5184 5244 size_t i;
5185 5245 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
5186 5246 HeapWord* min_val = top; // Higher than any PLAB address
5187 5247 uint min_tid = 0; // position of min_val this round
5188 5248 for (uint j = 0; j < ParallelGCThreads; j++) {
5189 5249 ChunkArray* cur_sca = &_survivor_plab_array[j];
5190 5250 if (_cursor[j] == cur_sca->end()) {
5191 5251 continue;
5192 5252 }
5193 5253 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5194 5254 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5195 5255 assert(surv->used_region().contains(cur_val), "Out of bounds value");
5196 5256 if (cur_val < min_val) {
5197 5257 min_tid = j;
5198 5258 min_val = cur_val;
5199 5259 } else {
5200 5260 assert(cur_val < top, "All recorded addresses should be less");
5201 5261 }
5202 5262 }
5203 5263 // At this point min_val and min_tid are respectively
5204 5264 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5205 5265 // and the thread (j) that witnesses that address.
5206 5266 // We record this address in the _survivor_chunk_array[i]
5207 5267 // and increment _cursor[min_tid] prior to the next round i.
5208 5268 if (min_val == top) {
5209 5269 break;
5210 5270 }
5211 5271 _survivor_chunk_array[i] = min_val;
5212 5272 _cursor[min_tid]++;
5213 5273 }
5214 5274 // We are all done; record the size of the _survivor_chunk_array
5215 5275 _survivor_chunk_index = i; // exclusive: [0, i)
5216 5276 if (PrintCMSStatistics > 0) {
5217 5277 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5218 5278 }
5219 5279 // Verify that we used up all the recorded entries
5220 5280 #ifdef ASSERT
5221 5281 size_t total = 0;
5222 5282 for (uint j = 0; j < ParallelGCThreads; j++) {
5223 5283 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5224 5284 total += _cursor[j];
5225 5285 }
5226 5286 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5227 5287 // Check that the merged array is in sorted order
5228 5288 if (total > 0) {
5229 5289 for (size_t i = 0; i < total - 1; i++) {
5230 5290 if (PrintCMSStatistics > 0) {
5231 5291 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5232 5292 i, _survivor_chunk_array[i]);
5233 5293 }
5234 5294 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5235 5295 "Not sorted");
5236 5296 }
5237 5297 }
5238 5298 #endif // ASSERT
5239 5299 }
5240 5300
5241 5301 // Set up the space's par_seq_tasks structure for work claiming
5242 5302 // for parallel rescan of young gen.
5243 5303 // See ParRescanTask where this is currently used.
5244 5304 void
5245 5305 CMSCollector::
5246 5306 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5247 5307 assert(n_threads > 0, "Unexpected n_threads argument");
5248 5308 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5249 5309
5250 5310 // Eden space
5251 5311 {
5252 5312 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5253 5313 assert(!pst->valid(), "Clobbering existing data?");
5254 5314 // Each valid entry in [0, _eden_chunk_index) represents a task.
5255 5315 size_t n_tasks = _eden_chunk_index + 1;
5256 5316 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5257 5317 pst->set_par_threads(n_threads);
5258 5318 pst->set_n_tasks((int)n_tasks);
5259 5319 }
5260 5320
5261 5321 // Merge the survivor plab arrays into _survivor_chunk_array
5262 5322 if (_survivor_plab_array != NULL) {
5263 5323 merge_survivor_plab_arrays(dng->from());
5264 5324 } else {
5265 5325 assert(_survivor_chunk_index == 0, "Error");
5266 5326 }
5267 5327
5268 5328 // To space
5269 5329 {
5270 5330 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5271 5331 assert(!pst->valid(), "Clobbering existing data?");
5272 5332 pst->set_par_threads(n_threads);
5273 5333 pst->set_n_tasks(1);
5274 5334 assert(pst->valid(), "Error");
5275 5335 }
5276 5336
5277 5337 // From space
5278 5338 {
5279 5339 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5280 5340 assert(!pst->valid(), "Clobbering existing data?");
5281 5341 size_t n_tasks = _survivor_chunk_index + 1;
5282 5342 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5283 5343 pst->set_par_threads(n_threads);
5284 5344 pst->set_n_tasks((int)n_tasks);
5285 5345 assert(pst->valid(), "Error");
5286 5346 }
5287 5347 }
5288 5348
5289 5349 // Parallel version of remark
5290 5350 void CMSCollector::do_remark_parallel() {
5291 5351 GenCollectedHeap* gch = GenCollectedHeap::heap();
5292 5352 WorkGang* workers = gch->workers();
5293 5353 assert(workers != NULL, "Need parallel worker threads.");
5294 5354 int n_workers = workers->total_workers();
5295 5355 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5296 5356 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
5297 5357
5298 5358 CMSParRemarkTask tsk(this,
5299 5359 cms_space, perm_space,
5300 5360 n_workers, workers, task_queues());
5301 5361
5302 5362 // Set up for parallel process_strong_roots work.
5303 5363 gch->set_par_threads(n_workers);
5304 5364 gch->change_strong_roots_parity();
5305 5365 // We won't be iterating over the cards in the card table updating
5306 5366 // the younger_gen cards, so we shouldn't call the following else
5307 5367 // the verification code as well as subsequent younger_refs_iterate
5308 5368 // code would get confused. XXX
5309 5369 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5310 5370
5311 5371 // The young gen rescan work will not be done as part of
5312 5372 // process_strong_roots (which currently doesn't knw how to
5313 5373 // parallelize such a scan), but rather will be broken up into
5314 5374 // a set of parallel tasks (via the sampling that the [abortable]
5315 5375 // preclean phase did of EdenSpace, plus the [two] tasks of
5316 5376 // scanning the [two] survivor spaces. Further fine-grain
5317 5377 // parallelization of the scanning of the survivor spaces
5318 5378 // themselves, and of precleaning of the younger gen itself
5319 5379 // is deferred to the future.
5320 5380 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5321 5381
5322 5382 // The dirty card rescan work is broken up into a "sequence"
5323 5383 // of parallel tasks (per constituent space) that are dynamically
5324 5384 // claimed by the parallel threads.
5325 5385 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5326 5386 perm_space->initialize_sequential_subtasks_for_rescan(n_workers);
5327 5387
5328 5388 // It turns out that even when we're using 1 thread, doing the work in a
5329 5389 // separate thread causes wide variance in run times. We can't help this
5330 5390 // in the multi-threaded case, but we special-case n=1 here to get
5331 5391 // repeatable measurements of the 1-thread overhead of the parallel code.
5332 5392 if (n_workers > 1) {
5333 5393 // Make refs discovery MT-safe
5334 5394 ReferenceProcessorMTMutator mt(ref_processor(), true);
5335 5395 workers->run_task(&tsk);
5336 5396 } else {
5337 5397 tsk.work(0);
5338 5398 }
5339 5399 gch->set_par_threads(0); // 0 ==> non-parallel.
5340 5400 // restore, single-threaded for now, any preserved marks
5341 5401 // as a result of work_q overflow
5342 5402 restore_preserved_marks_if_any();
5343 5403 }
5344 5404
5345 5405 // Non-parallel version of remark
5346 5406 void CMSCollector::do_remark_non_parallel() {
5347 5407 ResourceMark rm;
5348 5408 HandleMark hm;
5349 5409 GenCollectedHeap* gch = GenCollectedHeap::heap();
5350 5410 MarkRefsIntoAndScanClosure
5351 5411 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
5352 5412 &_markStack, &_revisitStack, this,
5353 5413 false /* should_yield */, false /* not precleaning */);
5354 5414 MarkFromDirtyCardsClosure
5355 5415 markFromDirtyCardsClosure(this, _span,
5356 5416 NULL, // space is set further below
5357 5417 &_markBitMap, &_markStack, &_revisitStack,
5358 5418 &mrias_cl);
5359 5419 {
5360 5420 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty);
5361 5421 // Iterate over the dirty cards, marking them precleaned, and
5362 5422 // setting the corresponding bits in the mod union table.
5363 5423 {
5364 5424 ModUnionClosure modUnionClosure(&_modUnionTable);
5365 5425 _ct->ct_bs()->dirty_card_iterate(
5366 5426 _cmsGen->used_region(),
5367 5427 &modUnionClosure);
5368 5428 _ct->ct_bs()->dirty_card_iterate(
5369 5429 _permGen->used_region(),
5370 5430 &modUnionClosure);
5371 5431 }
5372 5432 // Having transferred these marks into the modUnionTable, we just need
5373 5433 // to rescan the marked objects on the dirty cards in the modUnionTable.
5374 5434 // The initial marking may have been done during an asynchronous
5375 5435 // collection so there may be dirty bits in the mod-union table.
5376 5436 const int alignment =
5377 5437 CardTableModRefBS::card_size * BitsPerWord;
5378 5438 {
5379 5439 // ... First handle dirty cards in CMS gen
5380 5440 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5381 5441 MemRegion ur = _cmsGen->used_region();
5382 5442 HeapWord* lb = ur.start();
5383 5443 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5384 5444 MemRegion cms_span(lb, ub);
5385 5445 _modUnionTable.dirty_range_iterate_clear(cms_span,
5386 5446 &markFromDirtyCardsClosure);
5387 5447 verify_work_stacks_empty();
5388 5448 if (PrintCMSStatistics != 0) {
5389 5449 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5390 5450 markFromDirtyCardsClosure.num_dirty_cards());
5391 5451 }
5392 5452 }
5393 5453 {
5394 5454 // .. and then repeat for dirty cards in perm gen
5395 5455 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace());
5396 5456 MemRegion ur = _permGen->used_region();
5397 5457 HeapWord* lb = ur.start();
5398 5458 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5399 5459 MemRegion perm_span(lb, ub);
5400 5460 _modUnionTable.dirty_range_iterate_clear(perm_span,
5401 5461 &markFromDirtyCardsClosure);
5402 5462 verify_work_stacks_empty();
5403 5463 if (PrintCMSStatistics != 0) {
5404 5464 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ",
5405 5465 markFromDirtyCardsClosure.num_dirty_cards());
5406 5466 }
5407 5467 }
5408 5468 }
5409 5469 if (VerifyDuringGC &&
5410 5470 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5411 5471 HandleMark hm; // Discard invalid handles created during verification
5412 5472 Universe::verify(true);
5413 5473 }
5414 5474 {
5415 5475 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty);
5416 5476
5417 5477 verify_work_stacks_empty();
5418 5478
5419 5479 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5420 5480 gch->gen_process_strong_roots(_cmsGen->level(),
5421 5481 true, // younger gens as roots
5422 5482 true, // collecting perm gen
5423 5483 SharedHeap::ScanningOption(roots_scanning_options()),
5424 5484 NULL, &mrias_cl);
5425 5485 }
5426 5486 verify_work_stacks_empty();
5427 5487 // Restore evacuated mark words, if any, used for overflow list links
5428 5488 if (!CMSOverflowEarlyRestoration) {
5429 5489 restore_preserved_marks_if_any();
5430 5490 }
5431 5491 verify_overflow_empty();
5432 5492 }
5433 5493
5434 5494 ////////////////////////////////////////////////////////
5435 5495 // Parallel Reference Processing Task Proxy Class
5436 5496 ////////////////////////////////////////////////////////
5437 5497 class CMSRefProcTaskProxy: public AbstractGangTask {
5438 5498 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5439 5499 CMSCollector* _collector;
5440 5500 CMSBitMap* _mark_bit_map;
5441 5501 MemRegion _span;
5442 5502 OopTaskQueueSet* _task_queues;
5443 5503 ParallelTaskTerminator _term;
5444 5504 ProcessTask& _task;
5445 5505
5446 5506 public:
5447 5507 CMSRefProcTaskProxy(ProcessTask& task,
5448 5508 CMSCollector* collector,
5449 5509 const MemRegion& span,
5450 5510 CMSBitMap* mark_bit_map,
5451 5511 int total_workers,
5452 5512 OopTaskQueueSet* task_queues):
5453 5513 AbstractGangTask("Process referents by policy in parallel"),
5454 5514 _task(task),
5455 5515 _collector(collector), _span(span), _mark_bit_map(mark_bit_map),
5456 5516 _task_queues(task_queues),
5457 5517 _term(total_workers, task_queues)
5458 5518 { }
5459 5519
5460 5520 OopTaskQueueSet* task_queues() { return _task_queues; }
5461 5521
5462 5522 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5463 5523
5464 5524 ParallelTaskTerminator* terminator() { return &_term; }
5465 5525
5466 5526 void do_work_steal(int i,
5467 5527 CMSParDrainMarkingStackClosure* drain,
5468 5528 CMSParKeepAliveClosure* keep_alive,
5469 5529 int* seed);
5470 5530
5471 5531 virtual void work(int i);
5472 5532 };
5473 5533
5474 5534 void CMSRefProcTaskProxy::work(int i) {
5475 5535 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5476 5536 _mark_bit_map, work_queue(i));
5477 5537 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5478 5538 _mark_bit_map, work_queue(i));
5479 5539 CMSIsAliveClosure is_alive_closure(_mark_bit_map);
5480 5540 _task.work(i, is_alive_closure, par_keep_alive, par_drain_stack);
5481 5541 if (_task.marks_oops_alive()) {
5482 5542 do_work_steal(i, &par_drain_stack, &par_keep_alive,
5483 5543 _collector->hash_seed(i));
5484 5544 }
5485 5545 assert(work_queue(i)->size() == 0, "work_queue should be empty");
5486 5546 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5487 5547 }
5488 5548
5489 5549 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5490 5550 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5491 5551 EnqueueTask& _task;
5492 5552
5493 5553 public:
5494 5554 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5495 5555 : AbstractGangTask("Enqueue reference objects in parallel"),
5496 5556 _task(task)
5497 5557 { }
5498 5558
5499 5559 virtual void work(int i)
5500 5560 {
5501 5561 _task.work(i);
5502 5562 }
5503 5563 };
5504 5564
5505 5565 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5506 5566 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5507 5567 _collector(collector),
5508 5568 _span(span),
5509 5569 _bit_map(bit_map),
5510 5570 _work_queue(work_queue),
5511 5571 _mark_and_push(collector, span, bit_map, work_queue),
5512 5572 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5513 5573 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5514 5574 { }
5515 5575
5516 5576 // . see if we can share work_queues with ParNew? XXX
5517 5577 void CMSRefProcTaskProxy::do_work_steal(int i,
5518 5578 CMSParDrainMarkingStackClosure* drain,
5519 5579 CMSParKeepAliveClosure* keep_alive,
5520 5580 int* seed) {
5521 5581 OopTaskQueue* work_q = work_queue(i);
5522 5582 NOT_PRODUCT(int num_steals = 0;)
5523 5583 oop obj_to_scan;
5524 5584 size_t num_from_overflow_list =
5525 5585 MIN2((size_t)work_q->max_elems()/4,
5526 5586 (size_t)ParGCDesiredObjsFromOverflowList);
5527 5587
5528 5588 while (true) {
5529 5589 // Completely finish any left over work from (an) earlier round(s)
5530 5590 drain->trim_queue(0);
5531 5591 // Now check if there's any work in the overflow list
5532 5592 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5533 5593 work_q)) {
5534 5594 // Found something in global overflow list;
5535 5595 // not yet ready to go stealing work from others.
5536 5596 // We'd like to assert(work_q->size() != 0, ...)
5537 5597 // because we just took work from the overflow list,
5538 5598 // but of course we can't, since all of that might have
5539 5599 // been already stolen from us.
5540 5600 continue;
5541 5601 }
5542 5602 // Verify that we have no work before we resort to stealing
5543 5603 assert(work_q->size() == 0, "Have work, shouldn't steal");
5544 5604 // Try to steal from other queues that have work
5545 5605 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5546 5606 NOT_PRODUCT(num_steals++;)
5547 5607 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5548 5608 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5549 5609 // Do scanning work
5550 5610 obj_to_scan->oop_iterate(keep_alive);
5551 5611 // Loop around, finish this work, and try to steal some more
5552 5612 } else if (terminator()->offer_termination()) {
5553 5613 break; // nirvana from the infinite cycle
5554 5614 }
5555 5615 }
5556 5616 NOT_PRODUCT(
5557 5617 if (PrintCMSStatistics != 0) {
5558 5618 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5559 5619 }
5560 5620 )
5561 5621 }
5562 5622
5563 5623 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5564 5624 {
5565 5625 GenCollectedHeap* gch = GenCollectedHeap::heap();
5566 5626 WorkGang* workers = gch->workers();
5567 5627 assert(workers != NULL, "Need parallel worker threads.");
5568 5628 int n_workers = workers->total_workers();
5569 5629 CMSRefProcTaskProxy rp_task(task, &_collector,
5570 5630 _collector.ref_processor()->span(),
5571 5631 _collector.markBitMap(),
5572 5632 n_workers, _collector.task_queues());
5573 5633 workers->run_task(&rp_task);
5574 5634 }
5575 5635
5576 5636 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5577 5637 {
5578 5638
5579 5639 GenCollectedHeap* gch = GenCollectedHeap::heap();
5580 5640 WorkGang* workers = gch->workers();
5581 5641 assert(workers != NULL, "Need parallel worker threads.");
5582 5642 CMSRefEnqueueTaskProxy enq_task(task);
5583 5643 workers->run_task(&enq_task);
5584 5644 }
5585 5645
5586 5646 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
5587 5647
5588 5648 ResourceMark rm;
5589 5649 HandleMark hm;
5590 5650 ReferencePolicy* soft_ref_policy;
5591 5651
5592 5652 assert(!ref_processor()->enqueuing_is_done(), "Enqueuing should not be complete");
5593 5653 // Process weak references.
5594 5654 if (clear_all_soft_refs) {
5595 5655 soft_ref_policy = new AlwaysClearPolicy();
5596 5656 } else {
5597 5657 #ifdef COMPILER2
5598 5658 soft_ref_policy = new LRUMaxHeapPolicy();
5599 5659 #else
5600 5660 soft_ref_policy = new LRUCurrentHeapPolicy();
5601 5661 #endif // COMPILER2
5602 5662 }
5603 5663 verify_work_stacks_empty();
5604 5664
5605 5665 ReferenceProcessor* rp = ref_processor();
5606 5666 assert(rp->span().equals(_span), "Spans should be equal");
5607 5667 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5608 5668 &_markStack);
5609 5669 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5610 5670 _span, &_markBitMap, &_markStack,
5611 5671 &cmsKeepAliveClosure);
5612 5672 {
5613 5673 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty);
5614 5674 if (rp->processing_is_mt()) {
5615 5675 CMSRefProcTaskExecutor task_executor(*this);
5616 5676 rp->process_discovered_references(soft_ref_policy,
5617 5677 &_is_alive_closure,
5618 5678 &cmsKeepAliveClosure,
5619 5679 &cmsDrainMarkingStackClosure,
5620 5680 &task_executor);
↓ open down ↓ |
849 lines elided |
↑ open up ↑ |
5621 5681 } else {
5622 5682 rp->process_discovered_references(soft_ref_policy,
5623 5683 &_is_alive_closure,
5624 5684 &cmsKeepAliveClosure,
5625 5685 &cmsDrainMarkingStackClosure,
5626 5686 NULL);
5627 5687 }
5628 5688 verify_work_stacks_empty();
5629 5689 }
5630 5690
5631 - if (cms_should_unload_classes()) {
5691 + if (should_unload_classes()) {
5632 5692 {
5633 5693 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty);
5634 5694
5635 5695 // Follow SystemDictionary roots and unload classes
5636 5696 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5637 5697
5638 5698 // Follow CodeCache roots and unload any methods marked for unloading
5639 5699 CodeCache::do_unloading(&_is_alive_closure,
5640 5700 &cmsKeepAliveClosure,
5641 5701 purged_class);
5642 5702
5643 5703 cmsDrainMarkingStackClosure.do_void();
5644 5704 verify_work_stacks_empty();
5645 5705
5646 5706 // Update subklass/sibling/implementor links in KlassKlass descendants
5647 5707 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty");
5648 5708 oop k;
5649 5709 while ((k = _revisitStack.pop()) != NULL) {
5650 5710 ((Klass*)(oopDesc*)k)->follow_weak_klass_links(
5651 5711 &_is_alive_closure,
5652 5712 &cmsKeepAliveClosure);
5653 5713 }
5654 5714 assert(!ClassUnloading ||
5655 5715 (_markStack.isEmpty() && overflow_list_is_empty()),
5656 5716 "Should not have found new reachable objects");
5657 5717 assert(_revisitStack.isEmpty(), "revisit stack should have been drained");
5658 5718 cmsDrainMarkingStackClosure.do_void();
5659 5719 verify_work_stacks_empty();
5660 5720 }
5661 5721
5662 5722 {
5663 5723 TraceTime t("scrub symbol & string tables", PrintGCDetails, false, gclog_or_tty);
5664 5724 // Now clean up stale oops in SymbolTable and StringTable
5665 5725 SymbolTable::unlink(&_is_alive_closure);
5666 5726 StringTable::unlink(&_is_alive_closure);
5667 5727 }
5668 5728 }
5669 5729
5670 5730 verify_work_stacks_empty();
5671 5731 // Restore any preserved marks as a result of mark stack or
5672 5732 // work queue overflow
5673 5733 restore_preserved_marks_if_any(); // done single-threaded for now
5674 5734
5675 5735 rp->set_enqueuing_is_done(true);
5676 5736 if (rp->processing_is_mt()) {
5677 5737 CMSRefProcTaskExecutor task_executor(*this);
5678 5738 rp->enqueue_discovered_references(&task_executor);
5679 5739 } else {
5680 5740 rp->enqueue_discovered_references(NULL);
5681 5741 }
5682 5742 rp->verify_no_references_recorded();
5683 5743 assert(!rp->discovery_enabled(), "should have been disabled");
5684 5744
5685 5745 // JVMTI object tagging is based on JNI weak refs. If any of these
5686 5746 // refs were cleared then JVMTI needs to update its maps and
5687 5747 // maybe post ObjectFrees to agents.
5688 5748 JvmtiExport::cms_ref_processing_epilogue();
5689 5749 }
5690 5750
5691 5751 #ifndef PRODUCT
5692 5752 void CMSCollector::check_correct_thread_executing() {
5693 5753 Thread* t = Thread::current();
5694 5754 // Only the VM thread or the CMS thread should be here.
5695 5755 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5696 5756 "Unexpected thread type");
5697 5757 // If this is the vm thread, the foreground process
5698 5758 // should not be waiting. Note that _foregroundGCIsActive is
5699 5759 // true while the foreground collector is waiting.
5700 5760 if (_foregroundGCShouldWait) {
5701 5761 // We cannot be the VM thread
5702 5762 assert(t->is_ConcurrentGC_thread(),
5703 5763 "Should be CMS thread");
5704 5764 } else {
5705 5765 // We can be the CMS thread only if we are in a stop-world
5706 5766 // phase of CMS collection.
5707 5767 if (t->is_ConcurrentGC_thread()) {
5708 5768 assert(_collectorState == InitialMarking ||
5709 5769 _collectorState == FinalMarking,
5710 5770 "Should be a stop-world phase");
5711 5771 // The CMS thread should be holding the CMS_token.
5712 5772 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5713 5773 "Potential interference with concurrently "
5714 5774 "executing VM thread");
5715 5775 }
5716 5776 }
5717 5777 }
5718 5778 #endif
5719 5779
5720 5780 void CMSCollector::sweep(bool asynch) {
5721 5781 assert(_collectorState == Sweeping, "just checking");
5722 5782 check_correct_thread_executing();
5723 5783 verify_work_stacks_empty();
↓ open down ↓ |
82 lines elided |
↑ open up ↑ |
5724 5784 verify_overflow_empty();
5725 5785 incrementSweepCount();
5726 5786 _sweep_timer.stop();
5727 5787 _sweep_estimate.sample(_sweep_timer.seconds());
5728 5788 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
5729 5789
5730 5790 // PermGen verification support: If perm gen sweeping is disabled in
5731 5791 // this cycle, we preserve the perm gen object "deadness" information
5732 5792 // in the perm_gen_verify_bit_map. In order to do that we traverse
5733 5793 // all blocks in perm gen and mark all dead objects.
5734 - if (verifying() && !cms_should_unload_classes()) {
5735 - CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
5736 - bitMapLock());
5794 + if (verifying() && !should_unload_classes()) {
5737 5795 assert(perm_gen_verify_bit_map()->sizeInBits() != 0,
5738 5796 "Should have already been allocated");
5739 5797 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(),
5740 5798 markBitMap(), perm_gen_verify_bit_map());
5741 - _permGen->cmsSpace()->blk_iterate(&mdo);
5799 + if (asynch) {
5800 + CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
5801 + bitMapLock());
5802 + _permGen->cmsSpace()->blk_iterate(&mdo);
5803 + } else {
5804 + // In the case of synchronous sweep, we already have
5805 + // the requisite locks/tokens.
5806 + _permGen->cmsSpace()->blk_iterate(&mdo);
5807 + }
5742 5808 }
5743 5809
5744 5810 if (asynch) {
5745 5811 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5746 5812 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
5747 5813 // First sweep the old gen then the perm gen
5748 5814 {
5749 5815 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5750 5816 bitMapLock());
5751 5817 sweepWork(_cmsGen, asynch);
5752 5818 }
5753 5819
5754 5820 // Now repeat for perm gen
5755 - if (cms_should_unload_classes()) {
5821 + if (should_unload_classes()) {
5756 5822 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
5757 5823 bitMapLock());
5758 5824 sweepWork(_permGen, asynch);
5759 5825 }
5760 5826
5761 5827 // Update Universe::_heap_*_at_gc figures.
5762 5828 // We need all the free list locks to make the abstract state
5763 5829 // transition from Sweeping to Resetting. See detailed note
5764 5830 // further below.
5765 5831 {
5766 5832 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
↓ open down ↓ |
1 lines elided |
↑ open up ↑ |
5767 5833 _permGen->freelistLock());
5768 5834 // Update heap occupancy information which is used as
5769 5835 // input to soft ref clearing policy at the next gc.
5770 5836 Universe::update_heap_info_at_gc();
5771 5837 _collectorState = Resizing;
5772 5838 }
5773 5839 } else {
5774 5840 // already have needed locks
5775 5841 sweepWork(_cmsGen, asynch);
5776 5842
5777 - if (cms_should_unload_classes()) {
5843 + if (should_unload_classes()) {
5778 5844 sweepWork(_permGen, asynch);
5779 5845 }
5780 5846 // Update heap occupancy information which is used as
5781 5847 // input to soft ref clearing policy at the next gc.
5782 5848 Universe::update_heap_info_at_gc();
5783 5849 _collectorState = Resizing;
5784 5850 }
5785 5851 verify_work_stacks_empty();
5786 5852 verify_overflow_empty();
5787 5853
5788 5854 _sweep_timer.reset();
5789 5855 _sweep_timer.start();
5790 5856
5791 5857 update_time_of_last_gc(os::javaTimeMillis());
5792 5858
5793 5859 // NOTE on abstract state transitions:
5794 5860 // Mutators allocate-live and/or mark the mod-union table dirty
5795 5861 // based on the state of the collection. The former is done in
5796 5862 // the interval [Marking, Sweeping] and the latter in the interval
5797 5863 // [Marking, Sweeping). Thus the transitions into the Marking state
5798 5864 // and out of the Sweeping state must be synchronously visible
5799 5865 // globally to the mutators.
5800 5866 // The transition into the Marking state happens with the world
5801 5867 // stopped so the mutators will globally see it. Sweeping is
5802 5868 // done asynchronously by the background collector so the transition
5803 5869 // from the Sweeping state to the Resizing state must be done
5804 5870 // under the freelistLock (as is the check for whether to
5805 5871 // allocate-live and whether to dirty the mod-union table).
5806 5872 assert(_collectorState == Resizing, "Change of collector state to"
5807 5873 " Resizing must be done under the freelistLocks (plural)");
5808 5874
5809 5875 // Now that sweeping has been completed, if the GCH's
5810 5876 // incremental_collection_will_fail flag is set, clear it,
5811 5877 // thus inviting a younger gen collection to promote into
5812 5878 // this generation. If such a promotion may still fail,
5813 5879 // the flag will be set again when a young collection is
5814 5880 // attempted.
5815 5881 // I think the incremental_collection_will_fail flag's use
5816 5882 // is specific to a 2 generation collection policy, so i'll
5817 5883 // assert that that's the configuration we are operating within.
5818 5884 // The use of the flag can and should be generalized appropriately
5819 5885 // in the future to deal with a general n-generation system.
5820 5886
5821 5887 GenCollectedHeap* gch = GenCollectedHeap::heap();
5822 5888 assert(gch->collector_policy()->is_two_generation_policy(),
5823 5889 "Resetting of incremental_collection_will_fail flag"
5824 5890 " may be incorrect otherwise");
5825 5891 gch->clear_incremental_collection_will_fail();
5826 5892 gch->update_full_collections_completed(_collection_count_start);
5827 5893 }
5828 5894
5829 5895 // FIX ME!!! Looks like this belongs in CFLSpace, with
5830 5896 // CMSGen merely delegating to it.
5831 5897 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5832 5898 double nearLargestPercent = 0.999;
5833 5899 HeapWord* minAddr = _cmsSpace->bottom();
5834 5900 HeapWord* largestAddr =
5835 5901 (HeapWord*) _cmsSpace->dictionary()->findLargestDict();
5836 5902 if (largestAddr == 0) {
5837 5903 // The dictionary appears to be empty. In this case
5838 5904 // try to coalesce at the end of the heap.
5839 5905 largestAddr = _cmsSpace->end();
5840 5906 }
5841 5907 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5842 5908 size_t nearLargestOffset =
5843 5909 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5844 5910 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5845 5911 }
5846 5912
5847 5913 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5848 5914 return addr >= _cmsSpace->nearLargestChunk();
5849 5915 }
5850 5916
5851 5917 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5852 5918 return _cmsSpace->find_chunk_at_end();
5853 5919 }
5854 5920
5855 5921 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
5856 5922 bool full) {
5857 5923 // The next lower level has been collected. Gather any statistics
5858 5924 // that are of interest at this point.
5859 5925 if (!full && (current_level + 1) == level()) {
5860 5926 // Gather statistics on the young generation collection.
5861 5927 collector()->stats().record_gc0_end(used());
5862 5928 }
5863 5929 }
5864 5930
5865 5931 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
5866 5932 GenCollectedHeap* gch = GenCollectedHeap::heap();
5867 5933 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
5868 5934 "Wrong type of heap");
5869 5935 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
5870 5936 gch->gen_policy()->size_policy();
5871 5937 assert(sp->is_gc_cms_adaptive_size_policy(),
5872 5938 "Wrong type of size policy");
5873 5939 return sp;
5874 5940 }
5875 5941
5876 5942 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
5877 5943 if (PrintGCDetails && Verbose) {
5878 5944 gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
5879 5945 }
5880 5946 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
5881 5947 _debug_collection_type =
5882 5948 (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
5883 5949 if (PrintGCDetails && Verbose) {
5884 5950 gclog_or_tty->print_cr("to %d ", _debug_collection_type);
5885 5951 }
5886 5952 }
5887 5953
5888 5954 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
5889 5955 bool asynch) {
5890 5956 // We iterate over the space(s) underlying this generation,
5891 5957 // checking the mark bit map to see if the bits corresponding
5892 5958 // to specific blocks are marked or not. Blocks that are
5893 5959 // marked are live and are not swept up. All remaining blocks
5894 5960 // are swept up, with coalescing on-the-fly as we sweep up
5895 5961 // contiguous free and/or garbage blocks:
5896 5962 // We need to ensure that the sweeper synchronizes with allocators
5897 5963 // and stop-the-world collectors. In particular, the following
5898 5964 // locks are used:
5899 5965 // . CMS token: if this is held, a stop the world collection cannot occur
5900 5966 // . freelistLock: if this is held no allocation can occur from this
5901 5967 // generation by another thread
5902 5968 // . bitMapLock: if this is held, no other thread can access or update
5903 5969 //
5904 5970
5905 5971 // Note that we need to hold the freelistLock if we use
5906 5972 // block iterate below; else the iterator might go awry if
5907 5973 // a mutator (or promotion) causes block contents to change
5908 5974 // (for instance if the allocator divvies up a block).
5909 5975 // If we hold the free list lock, for all practical purposes
5910 5976 // young generation GC's can't occur (they'll usually need to
5911 5977 // promote), so we might as well prevent all young generation
5912 5978 // GC's while we do a sweeping step. For the same reason, we might
5913 5979 // as well take the bit map lock for the entire duration
5914 5980
5915 5981 // check that we hold the requisite locks
5916 5982 assert(have_cms_token(), "Should hold cms token");
5917 5983 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
5918 5984 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
5919 5985 "Should possess CMS token to sweep");
5920 5986 assert_lock_strong(gen->freelistLock());
5921 5987 assert_lock_strong(bitMapLock());
5922 5988
5923 5989 assert(!_sweep_timer.is_active(), "Was switched off in an outer context");
5924 5990 gen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()),
5925 5991 _sweep_estimate.padded_average());
5926 5992 gen->setNearLargestChunk();
5927 5993
5928 5994 {
↓ open down ↓ |
141 lines elided |
↑ open up ↑ |
5929 5995 SweepClosure sweepClosure(this, gen, &_markBitMap,
5930 5996 CMSYield && asynch);
5931 5997 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5932 5998 // We need to free-up/coalesce garbage/blocks from a
5933 5999 // co-terminal free run. This is done in the SweepClosure
5934 6000 // destructor; so, do not remove this scope, else the
5935 6001 // end-of-sweep-census below will be off by a little bit.
5936 6002 }
5937 6003 gen->cmsSpace()->sweep_completed();
5938 6004 gen->cmsSpace()->endSweepFLCensus(sweepCount());
6005 + if (should_unload_classes()) { // unloaded classes this cycle,
6006 + _concurrent_cycles_since_last_unload = 0; // ... reset count
6007 + } else { // did not unload classes,
6008 + _concurrent_cycles_since_last_unload++; // ... increment count
6009 + }
5939 6010 }
5940 6011
5941 6012 // Reset CMS data structures (for now just the marking bit map)
5942 6013 // preparatory for the next cycle.
5943 6014 void CMSCollector::reset(bool asynch) {
5944 6015 GenCollectedHeap* gch = GenCollectedHeap::heap();
5945 6016 CMSAdaptiveSizePolicy* sp = size_policy();
5946 6017 AdaptiveSizePolicyOutput(sp, gch->total_collections());
5947 6018 if (asynch) {
5948 6019 CMSTokenSyncWithLocks ts(true, bitMapLock());
5949 6020
5950 6021 // If the state is not "Resetting", the foreground thread
5951 6022 // has done a collection and the resetting.
5952 6023 if (_collectorState != Resetting) {
5953 6024 assert(_collectorState == Idling, "The state should only change"
5954 6025 " because the foreground collector has finished the collection");
5955 6026 return;
5956 6027 }
5957 6028
5958 6029 // Clear the mark bitmap (no grey objects to start with)
5959 6030 // for the next cycle.
5960 6031 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5961 6032 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
5962 6033
5963 6034 HeapWord* curAddr = _markBitMap.startWord();
5964 6035 while (curAddr < _markBitMap.endWord()) {
5965 6036 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
5966 6037 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5967 6038 _markBitMap.clear_large_range(chunk);
5968 6039 if (ConcurrentMarkSweepThread::should_yield() &&
5969 6040 !foregroundGCIsActive() &&
5970 6041 CMSYield) {
5971 6042 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5972 6043 "CMS thread should hold CMS token");
5973 6044 assert_lock_strong(bitMapLock());
5974 6045 bitMapLock()->unlock();
5975 6046 ConcurrentMarkSweepThread::desynchronize(true);
5976 6047 ConcurrentMarkSweepThread::acknowledge_yield_request();
5977 6048 stopTimer();
5978 6049 if (PrintCMSStatistics != 0) {
5979 6050 incrementYields();
5980 6051 }
5981 6052 icms_wait();
5982 6053
5983 6054 // See the comment in coordinator_yield()
5984 6055 for (unsigned i = 0; i < CMSYieldSleepCount &&
5985 6056 ConcurrentMarkSweepThread::should_yield() &&
5986 6057 !CMSCollector::foregroundGCIsActive(); ++i) {
5987 6058 os::sleep(Thread::current(), 1, false);
5988 6059 ConcurrentMarkSweepThread::acknowledge_yield_request();
5989 6060 }
5990 6061
5991 6062 ConcurrentMarkSweepThread::synchronize(true);
5992 6063 bitMapLock()->lock_without_safepoint_check();
5993 6064 startTimer();
5994 6065 }
5995 6066 curAddr = chunk.end();
5996 6067 }
5997 6068 _collectorState = Idling;
5998 6069 } else {
5999 6070 // already have the lock
6000 6071 assert(_collectorState == Resetting, "just checking");
6001 6072 assert_lock_strong(bitMapLock());
6002 6073 _markBitMap.clear_all();
6003 6074 _collectorState = Idling;
6004 6075 }
6005 6076
6006 6077 // Stop incremental mode after a cycle completes, so that any future cycles
6007 6078 // are triggered by allocation.
6008 6079 stop_icms();
6009 6080
6010 6081 NOT_PRODUCT(
6011 6082 if (RotateCMSCollectionTypes) {
6012 6083 _cmsGen->rotate_debug_collection_type();
6013 6084 }
6014 6085 )
6015 6086 }
6016 6087
6017 6088 void CMSCollector::do_CMS_operation(CMS_op_type op) {
6018 6089 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
6019 6090 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6020 6091 TraceTime t("GC", PrintGC, !PrintGCDetails, gclog_or_tty);
6021 6092 TraceCollectorStats tcs(counters());
6022 6093
6023 6094 switch (op) {
6024 6095 case CMS_op_checkpointRootsInitial: {
6025 6096 checkpointRootsInitial(true); // asynch
6026 6097 if (PrintGC) {
6027 6098 _cmsGen->printOccupancy("initial-mark");
6028 6099 }
6029 6100 break;
6030 6101 }
6031 6102 case CMS_op_checkpointRootsFinal: {
6032 6103 checkpointRootsFinal(true, // asynch
6033 6104 false, // !clear_all_soft_refs
6034 6105 false); // !init_mark_was_synchronous
6035 6106 if (PrintGC) {
6036 6107 _cmsGen->printOccupancy("remark");
6037 6108 }
6038 6109 break;
6039 6110 }
6040 6111 default:
6041 6112 fatal("No such CMS_op");
6042 6113 }
6043 6114 }
6044 6115
6045 6116 #ifndef PRODUCT
6046 6117 size_t const CMSCollector::skip_header_HeapWords() {
6047 6118 return FreeChunk::header_size();
6048 6119 }
6049 6120
6050 6121 // Try and collect here conditions that should hold when
6051 6122 // CMS thread is exiting. The idea is that the foreground GC
6052 6123 // thread should not be blocked if it wants to terminate
6053 6124 // the CMS thread and yet continue to run the VM for a while
6054 6125 // after that.
6055 6126 void CMSCollector::verify_ok_to_terminate() const {
6056 6127 assert(Thread::current()->is_ConcurrentGC_thread(),
6057 6128 "should be called by CMS thread");
6058 6129 assert(!_foregroundGCShouldWait, "should be false");
6059 6130 // We could check here that all the various low-level locks
6060 6131 // are not held by the CMS thread, but that is overkill; see
6061 6132 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6062 6133 // is checked.
6063 6134 }
6064 6135 #endif
6065 6136
6066 6137 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6067 6138 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6068 6139 "missing Printezis mark?");
6069 6140 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6070 6141 size_t size = pointer_delta(nextOneAddr + 1, addr);
6071 6142 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6072 6143 "alignment problem");
6073 6144 assert(size >= 3, "Necessary for Printezis marks to work");
6074 6145 return size;
6075 6146 }
6076 6147
6077 6148 // A variant of the above (block_size_using_printezis_bits()) except
6078 6149 // that we return 0 if the P-bits are not yet set.
6079 6150 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6080 6151 if (_markBitMap.isMarked(addr)) {
6081 6152 assert(_markBitMap.isMarked(addr + 1), "Missing Printezis bit?");
6082 6153 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6083 6154 size_t size = pointer_delta(nextOneAddr + 1, addr);
6084 6155 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6085 6156 "alignment problem");
6086 6157 assert(size >= 3, "Necessary for Printezis marks to work");
6087 6158 return size;
6088 6159 } else {
6089 6160 assert(!_markBitMap.isMarked(addr + 1), "Bit map inconsistency?");
6090 6161 return 0;
6091 6162 }
6092 6163 }
6093 6164
6094 6165 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6095 6166 size_t sz = 0;
6096 6167 oop p = (oop)addr;
6097 6168 if (p->klass() != NULL && p->is_parsable()) {
6098 6169 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6099 6170 } else {
6100 6171 sz = block_size_using_printezis_bits(addr);
6101 6172 }
6102 6173 assert(sz > 0, "size must be nonzero");
6103 6174 HeapWord* next_block = addr + sz;
6104 6175 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
6105 6176 CardTableModRefBS::card_size);
6106 6177 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
6107 6178 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6108 6179 "must be different cards");
6109 6180 return next_card;
6110 6181 }
6111 6182
6112 6183
6113 6184 // CMS Bit Map Wrapper /////////////////////////////////////////
6114 6185
6115 6186 // Construct a CMS bit map infrastructure, but don't create the
6116 6187 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6117 6188 // further below.
6118 6189 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6119 6190 _bm(NULL,0),
6120 6191 _shifter(shifter),
6121 6192 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6122 6193 {
6123 6194 _bmStartWord = 0;
6124 6195 _bmWordSize = 0;
6125 6196 }
6126 6197
6127 6198 bool CMSBitMap::allocate(MemRegion mr) {
6128 6199 _bmStartWord = mr.start();
6129 6200 _bmWordSize = mr.word_size();
6130 6201 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6131 6202 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6132 6203 if (!brs.is_reserved()) {
6133 6204 warning("CMS bit map allocation failure");
6134 6205 return false;
6135 6206 }
6136 6207 // For now we'll just commit all of the bit map up fromt.
6137 6208 // Later on we'll try to be more parsimonious with swap.
6138 6209 if (!_virtual_space.initialize(brs, brs.size())) {
6139 6210 warning("CMS bit map backing store failure");
6140 6211 return false;
6141 6212 }
6142 6213 assert(_virtual_space.committed_size() == brs.size(),
6143 6214 "didn't reserve backing store for all of CMS bit map?");
6144 6215 _bm.set_map((uintptr_t*)_virtual_space.low());
6145 6216 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6146 6217 _bmWordSize, "inconsistency in bit map sizing");
6147 6218 _bm.set_size(_bmWordSize >> _shifter);
6148 6219
6149 6220 // bm.clear(); // can we rely on getting zero'd memory? verify below
6150 6221 assert(isAllClear(),
6151 6222 "Expected zero'd memory from ReservedSpace constructor");
6152 6223 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6153 6224 "consistency check");
6154 6225 return true;
6155 6226 }
6156 6227
6157 6228 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6158 6229 HeapWord *next_addr, *end_addr, *last_addr;
6159 6230 assert_locked();
6160 6231 assert(covers(mr), "out-of-range error");
6161 6232 // XXX assert that start and end are appropriately aligned
6162 6233 for (next_addr = mr.start(), end_addr = mr.end();
6163 6234 next_addr < end_addr; next_addr = last_addr) {
6164 6235 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6165 6236 last_addr = dirty_region.end();
6166 6237 if (!dirty_region.is_empty()) {
6167 6238 cl->do_MemRegion(dirty_region);
6168 6239 } else {
6169 6240 assert(last_addr == end_addr, "program logic");
6170 6241 return;
6171 6242 }
6172 6243 }
6173 6244 }
6174 6245
6175 6246 #ifndef PRODUCT
6176 6247 void CMSBitMap::assert_locked() const {
6177 6248 CMSLockVerifier::assert_locked(lock());
6178 6249 }
6179 6250
6180 6251 bool CMSBitMap::covers(MemRegion mr) const {
6181 6252 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6182 6253 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6183 6254 "size inconsistency");
6184 6255 return (mr.start() >= _bmStartWord) &&
6185 6256 (mr.end() <= endWord());
6186 6257 }
6187 6258
6188 6259 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6189 6260 return (start >= _bmStartWord && (start + size) <= endWord());
6190 6261 }
6191 6262
6192 6263 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6193 6264 // verify that there are no 1 bits in the interval [left, right)
6194 6265 FalseBitMapClosure falseBitMapClosure;
6195 6266 iterate(&falseBitMapClosure, left, right);
6196 6267 }
6197 6268
6198 6269 void CMSBitMap::region_invariant(MemRegion mr)
6199 6270 {
6200 6271 assert_locked();
6201 6272 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6202 6273 assert(!mr.is_empty(), "unexpected empty region");
6203 6274 assert(covers(mr), "mr should be covered by bit map");
6204 6275 // convert address range into offset range
6205 6276 size_t start_ofs = heapWordToOffset(mr.start());
6206 6277 // Make sure that end() is appropriately aligned
6207 6278 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6208 6279 (1 << (_shifter+LogHeapWordSize))),
6209 6280 "Misaligned mr.end()");
6210 6281 size_t end_ofs = heapWordToOffset(mr.end());
6211 6282 assert(end_ofs > start_ofs, "Should mark at least one bit");
6212 6283 }
6213 6284
6214 6285 #endif
6215 6286
6216 6287 bool CMSMarkStack::allocate(size_t size) {
6217 6288 // allocate a stack of the requisite depth
6218 6289 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6219 6290 size * sizeof(oop)));
6220 6291 if (!rs.is_reserved()) {
6221 6292 warning("CMSMarkStack allocation failure");
6222 6293 return false;
6223 6294 }
6224 6295 if (!_virtual_space.initialize(rs, rs.size())) {
6225 6296 warning("CMSMarkStack backing store failure");
6226 6297 return false;
6227 6298 }
6228 6299 assert(_virtual_space.committed_size() == rs.size(),
6229 6300 "didn't reserve backing store for all of CMS stack?");
6230 6301 _base = (oop*)(_virtual_space.low());
6231 6302 _index = 0;
6232 6303 _capacity = size;
6233 6304 NOT_PRODUCT(_max_depth = 0);
6234 6305 return true;
6235 6306 }
6236 6307
6237 6308 // XXX FIX ME !!! In the MT case we come in here holding a
6238 6309 // leaf lock. For printing we need to take a further lock
6239 6310 // which has lower rank. We need to recallibrate the two
6240 6311 // lock-ranks involved in order to be able to rpint the
6241 6312 // messages below. (Or defer the printing to the caller.
6242 6313 // For now we take the expedient path of just disabling the
6243 6314 // messages for the problematic case.)
6244 6315 void CMSMarkStack::expand() {
6245 6316 assert(_capacity <= CMSMarkStackSizeMax, "stack bigger than permitted");
6246 6317 if (_capacity == CMSMarkStackSizeMax) {
6247 6318 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6248 6319 // We print a warning message only once per CMS cycle.
6249 6320 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6250 6321 }
6251 6322 return;
6252 6323 }
6253 6324 // Double capacity if possible
6254 6325 size_t new_capacity = MIN2(_capacity*2, CMSMarkStackSizeMax);
6255 6326 // Do not give up existing stack until we have managed to
6256 6327 // get the double capacity that we desired.
6257 6328 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6258 6329 new_capacity * sizeof(oop)));
6259 6330 if (rs.is_reserved()) {
6260 6331 // Release the backing store associated with old stack
6261 6332 _virtual_space.release();
6262 6333 // Reinitialize virtual space for new stack
6263 6334 if (!_virtual_space.initialize(rs, rs.size())) {
6264 6335 fatal("Not enough swap for expanded marking stack");
6265 6336 }
6266 6337 _base = (oop*)(_virtual_space.low());
6267 6338 _index = 0;
6268 6339 _capacity = new_capacity;
6269 6340 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6270 6341 // Failed to double capacity, continue;
6271 6342 // we print a detail message only once per CMS cycle.
6272 6343 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6273 6344 SIZE_FORMAT"K",
6274 6345 _capacity / K, new_capacity / K);
6275 6346 }
6276 6347 }
6277 6348
6278 6349
6279 6350 // Closures
6280 6351 // XXX: there seems to be a lot of code duplication here;
6281 6352 // should refactor and consolidate common code.
6282 6353
6283 6354 // This closure is used to mark refs into the CMS generation in
6284 6355 // the CMS bit map. Called at the first checkpoint. This closure
6285 6356 // assumes that we do not need to re-mark dirty cards; if the CMS
6286 6357 // generation on which this is used is not an oldest (modulo perm gen)
6287 6358 // generation then this will lose younger_gen cards!
6288 6359
6289 6360 MarkRefsIntoClosure::MarkRefsIntoClosure(
6290 6361 MemRegion span, CMSBitMap* bitMap, bool should_do_nmethods):
6291 6362 _span(span),
6292 6363 _bitMap(bitMap),
6293 6364 _should_do_nmethods(should_do_nmethods)
6294 6365 {
6295 6366 assert(_ref_processor == NULL, "deliberately left NULL");
6296 6367 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6297 6368 }
6298 6369
6299 6370 void MarkRefsIntoClosure::do_oop(oop* p) {
6300 6371 // if p points into _span, then mark corresponding bit in _markBitMap
6301 6372 oop thisOop = *p;
6302 6373 if (thisOop != NULL) {
6303 6374 assert(thisOop->is_oop(), "expected an oop");
6304 6375 HeapWord* addr = (HeapWord*)thisOop;
6305 6376 if (_span.contains(addr)) {
6306 6377 // this should be made more efficient
6307 6378 _bitMap->mark(addr);
6308 6379 }
6309 6380 }
6310 6381 }
6311 6382
6312 6383 // A variant of the above, used for CMS marking verification.
6313 6384 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6314 6385 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6315 6386 bool should_do_nmethods):
6316 6387 _span(span),
6317 6388 _verification_bm(verification_bm),
6318 6389 _cms_bm(cms_bm),
6319 6390 _should_do_nmethods(should_do_nmethods) {
6320 6391 assert(_ref_processor == NULL, "deliberately left NULL");
6321 6392 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6322 6393 }
6323 6394
6324 6395 void MarkRefsIntoVerifyClosure::do_oop(oop* p) {
6325 6396 // if p points into _span, then mark corresponding bit in _markBitMap
6326 6397 oop this_oop = *p;
6327 6398 if (this_oop != NULL) {
6328 6399 assert(this_oop->is_oop(), "expected an oop");
6329 6400 HeapWord* addr = (HeapWord*)this_oop;
6330 6401 if (_span.contains(addr)) {
6331 6402 _verification_bm->mark(addr);
6332 6403 if (!_cms_bm->isMarked(addr)) {
6333 6404 oop(addr)->print();
6334 6405 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
6335 6406 fatal("... aborting");
6336 6407 }
6337 6408 }
6338 6409 }
6339 6410 }
6340 6411
6341 6412 //////////////////////////////////////////////////
6342 6413 // MarkRefsIntoAndScanClosure
6343 6414 //////////////////////////////////////////////////
6344 6415
6345 6416 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6346 6417 ReferenceProcessor* rp,
6347 6418 CMSBitMap* bit_map,
6348 6419 CMSBitMap* mod_union_table,
6349 6420 CMSMarkStack* mark_stack,
6350 6421 CMSMarkStack* revisit_stack,
6351 6422 CMSCollector* collector,
6352 6423 bool should_yield,
6353 6424 bool concurrent_precleaning):
6354 6425 _collector(collector),
6355 6426 _span(span),
6356 6427 _bit_map(bit_map),
6357 6428 _mark_stack(mark_stack),
6358 6429 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6359 6430 mark_stack, revisit_stack, concurrent_precleaning),
6360 6431 _yield(should_yield),
6361 6432 _concurrent_precleaning(concurrent_precleaning),
6362 6433 _freelistLock(NULL)
6363 6434 {
6364 6435 _ref_processor = rp;
6365 6436 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6366 6437 }
6367 6438
6368 6439 // This closure is used to mark refs into the CMS generation at the
6369 6440 // second (final) checkpoint, and to scan and transitively follow
6370 6441 // the unmarked oops. It is also used during the concurrent precleaning
6371 6442 // phase while scanning objects on dirty cards in the CMS generation.
6372 6443 // The marks are made in the marking bit map and the marking stack is
6373 6444 // used for keeping the (newly) grey objects during the scan.
6374 6445 // The parallel version (Par_...) appears further below.
6375 6446 void MarkRefsIntoAndScanClosure::do_oop(oop* p) {
6376 6447 oop this_oop = *p;
6377 6448 if (this_oop != NULL) {
6378 6449 assert(this_oop->is_oop(), "expected an oop");
6379 6450 HeapWord* addr = (HeapWord*)this_oop;
6380 6451 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6381 6452 assert(_collector->overflow_list_is_empty(), "should be empty");
6382 6453 if (_span.contains(addr) &&
6383 6454 !_bit_map->isMarked(addr)) {
6384 6455 // mark bit map (object is now grey)
6385 6456 _bit_map->mark(addr);
6386 6457 // push on marking stack (stack should be empty), and drain the
6387 6458 // stack by applying this closure to the oops in the oops popped
6388 6459 // from the stack (i.e. blacken the grey objects)
6389 6460 bool res = _mark_stack->push(this_oop);
6390 6461 assert(res, "Should have space to push on empty stack");
6391 6462 do {
6392 6463 oop new_oop = _mark_stack->pop();
6393 6464 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6394 6465 assert(new_oop->is_parsable(), "Found unparsable oop");
6395 6466 assert(_bit_map->isMarked((HeapWord*)new_oop),
6396 6467 "only grey objects on this stack");
6397 6468 // iterate over the oops in this oop, marking and pushing
6398 6469 // the ones in CMS heap (i.e. in _span).
6399 6470 new_oop->oop_iterate(&_pushAndMarkClosure);
6400 6471 // check if it's time to yield
6401 6472 do_yield_check();
6402 6473 } while (!_mark_stack->isEmpty() ||
6403 6474 (!_concurrent_precleaning && take_from_overflow_list()));
6404 6475 // if marking stack is empty, and we are not doing this
6405 6476 // during precleaning, then check the overflow list
6406 6477 }
6407 6478 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6408 6479 assert(_collector->overflow_list_is_empty(),
6409 6480 "overflow list was drained above");
6410 6481 // We could restore evacuated mark words, if any, used for
6411 6482 // overflow list links here because the overflow list is
6412 6483 // provably empty here. That would reduce the maximum
6413 6484 // size requirements for preserved_{oop,mark}_stack.
6414 6485 // But we'll just postpone it until we are all done
6415 6486 // so we can just stream through.
6416 6487 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6417 6488 _collector->restore_preserved_marks_if_any();
6418 6489 assert(_collector->no_preserved_marks(), "No preserved marks");
6419 6490 }
6420 6491 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6421 6492 "All preserved marks should have been restored above");
6422 6493 }
6423 6494 }
6424 6495
6425 6496 void MarkRefsIntoAndScanClosure::do_yield_work() {
6426 6497 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6427 6498 "CMS thread should hold CMS token");
6428 6499 assert_lock_strong(_freelistLock);
6429 6500 assert_lock_strong(_bit_map->lock());
6430 6501 // relinquish the free_list_lock and bitMaplock()
6431 6502 _bit_map->lock()->unlock();
6432 6503 _freelistLock->unlock();
6433 6504 ConcurrentMarkSweepThread::desynchronize(true);
6434 6505 ConcurrentMarkSweepThread::acknowledge_yield_request();
6435 6506 _collector->stopTimer();
6436 6507 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6437 6508 if (PrintCMSStatistics != 0) {
6438 6509 _collector->incrementYields();
6439 6510 }
6440 6511 _collector->icms_wait();
6441 6512
6442 6513 // See the comment in coordinator_yield()
6443 6514 for (unsigned i = 0; i < CMSYieldSleepCount &&
6444 6515 ConcurrentMarkSweepThread::should_yield() &&
6445 6516 !CMSCollector::foregroundGCIsActive(); ++i) {
6446 6517 os::sleep(Thread::current(), 1, false);
6447 6518 ConcurrentMarkSweepThread::acknowledge_yield_request();
6448 6519 }
6449 6520
6450 6521 ConcurrentMarkSweepThread::synchronize(true);
6451 6522 _freelistLock->lock_without_safepoint_check();
6452 6523 _bit_map->lock()->lock_without_safepoint_check();
6453 6524 _collector->startTimer();
6454 6525 }
6455 6526
6456 6527 ///////////////////////////////////////////////////////////
6457 6528 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6458 6529 // MarkRefsIntoAndScanClosure
6459 6530 ///////////////////////////////////////////////////////////
6460 6531 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6461 6532 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6462 6533 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack):
6463 6534 _span(span),
6464 6535 _bit_map(bit_map),
6465 6536 _work_queue(work_queue),
6466 6537 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6467 6538 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6468 6539 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue,
6469 6540 revisit_stack)
6470 6541 {
6471 6542 _ref_processor = rp;
6472 6543 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6473 6544 }
6474 6545
6475 6546 // This closure is used to mark refs into the CMS generation at the
6476 6547 // second (final) checkpoint, and to scan and transitively follow
6477 6548 // the unmarked oops. The marks are made in the marking bit map and
6478 6549 // the work_queue is used for keeping the (newly) grey objects during
6479 6550 // the scan phase whence they are also available for stealing by parallel
6480 6551 // threads. Since the marking bit map is shared, updates are
6481 6552 // synchronized (via CAS).
6482 6553 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) {
6483 6554 oop this_oop = *p;
6484 6555 if (this_oop != NULL) {
6485 6556 // Ignore mark word because this could be an already marked oop
6486 6557 // that may be chained at the end of the overflow list.
6487 6558 assert(this_oop->is_oop(true /* ignore mark word */), "expected an oop");
6488 6559 HeapWord* addr = (HeapWord*)this_oop;
6489 6560 if (_span.contains(addr) &&
6490 6561 !_bit_map->isMarked(addr)) {
6491 6562 // mark bit map (object will become grey):
6492 6563 // It is possible for several threads to be
6493 6564 // trying to "claim" this object concurrently;
6494 6565 // the unique thread that succeeds in marking the
6495 6566 // object first will do the subsequent push on
6496 6567 // to the work queue (or overflow list).
6497 6568 if (_bit_map->par_mark(addr)) {
6498 6569 // push on work_queue (which may not be empty), and trim the
6499 6570 // queue to an appropriate length by applying this closure to
6500 6571 // the oops in the oops popped from the stack (i.e. blacken the
6501 6572 // grey objects)
6502 6573 bool res = _work_queue->push(this_oop);
6503 6574 assert(res, "Low water mark should be less than capacity?");
6504 6575 trim_queue(_low_water_mark);
6505 6576 } // Else, another thread claimed the object
6506 6577 }
6507 6578 }
6508 6579 }
6509 6580
6510 6581 // This closure is used to rescan the marked objects on the dirty cards
6511 6582 // in the mod union table and the card table proper.
6512 6583 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6513 6584 oop p, MemRegion mr) {
6514 6585
6515 6586 size_t size = 0;
6516 6587 HeapWord* addr = (HeapWord*)p;
6517 6588 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6518 6589 assert(_span.contains(addr), "we are scanning the CMS generation");
6519 6590 // check if it's time to yield
6520 6591 if (do_yield_check()) {
6521 6592 // We yielded for some foreground stop-world work,
6522 6593 // and we have been asked to abort this ongoing preclean cycle.
6523 6594 return 0;
6524 6595 }
6525 6596 if (_bitMap->isMarked(addr)) {
6526 6597 // it's marked; is it potentially uninitialized?
6527 6598 if (p->klass() != NULL) {
6528 6599 if (CMSPermGenPrecleaningEnabled && !p->is_parsable()) {
6529 6600 // Signal precleaning to redirty the card since
6530 6601 // the klass pointer is already installed.
6531 6602 assert(size == 0, "Initial value");
6532 6603 } else {
6533 6604 assert(p->is_parsable(), "must be parsable.");
6534 6605 // an initialized object; ignore mark word in verification below
6535 6606 // since we are running concurrent with mutators
6536 6607 assert(p->is_oop(true), "should be an oop");
6537 6608 if (p->is_objArray()) {
6538 6609 // objArrays are precisely marked; restrict scanning
6539 6610 // to dirty cards only.
6540 6611 size = p->oop_iterate(_scanningClosure, mr);
6541 6612 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6542 6613 "adjustObjectSize should be the identity for array sizes, "
6543 6614 "which are necessarily larger than minimum object size of "
6544 6615 "two heap words");
6545 6616 } else {
6546 6617 // A non-array may have been imprecisely marked; we need
6547 6618 // to scan object in its entirety.
6548 6619 size = CompactibleFreeListSpace::adjustObjectSize(
6549 6620 p->oop_iterate(_scanningClosure));
6550 6621 }
6551 6622 #ifdef DEBUG
6552 6623 size_t direct_size =
6553 6624 CompactibleFreeListSpace::adjustObjectSize(p->size());
6554 6625 assert(size == direct_size, "Inconsistency in size");
6555 6626 assert(size >= 3, "Necessary for Printezis marks to work");
6556 6627 if (!_bitMap->isMarked(addr+1)) {
6557 6628 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6558 6629 } else {
6559 6630 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6560 6631 assert(_bitMap->isMarked(addr+size-1),
6561 6632 "inconsistent Printezis mark");
6562 6633 }
6563 6634 #endif // DEBUG
6564 6635 }
6565 6636 } else {
6566 6637 // an unitialized object
6567 6638 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6568 6639 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6569 6640 size = pointer_delta(nextOneAddr + 1, addr);
6570 6641 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6571 6642 "alignment problem");
6572 6643 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6573 6644 // will dirty the card when the klass pointer is installed in the
6574 6645 // object (signalling the completion of initialization).
6575 6646 }
6576 6647 } else {
6577 6648 // Either a not yet marked object or an uninitialized object
6578 6649 if (p->klass() == NULL || !p->is_parsable()) {
6579 6650 // An uninitialized object, skip to the next card, since
6580 6651 // we may not be able to read its P-bits yet.
6581 6652 assert(size == 0, "Initial value");
6582 6653 } else {
6583 6654 // An object not (yet) reached by marking: we merely need to
6584 6655 // compute its size so as to go look at the next block.
6585 6656 assert(p->is_oop(true), "should be an oop");
6586 6657 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6587 6658 }
6588 6659 }
6589 6660 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6590 6661 return size;
6591 6662 }
6592 6663
6593 6664 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6594 6665 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6595 6666 "CMS thread should hold CMS token");
6596 6667 assert_lock_strong(_freelistLock);
6597 6668 assert_lock_strong(_bitMap->lock());
6598 6669 // relinquish the free_list_lock and bitMaplock()
6599 6670 _bitMap->lock()->unlock();
6600 6671 _freelistLock->unlock();
6601 6672 ConcurrentMarkSweepThread::desynchronize(true);
6602 6673 ConcurrentMarkSweepThread::acknowledge_yield_request();
6603 6674 _collector->stopTimer();
6604 6675 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6605 6676 if (PrintCMSStatistics != 0) {
6606 6677 _collector->incrementYields();
6607 6678 }
6608 6679 _collector->icms_wait();
6609 6680
6610 6681 // See the comment in coordinator_yield()
6611 6682 for (unsigned i = 0; i < CMSYieldSleepCount &&
6612 6683 ConcurrentMarkSweepThread::should_yield() &&
6613 6684 !CMSCollector::foregroundGCIsActive(); ++i) {
6614 6685 os::sleep(Thread::current(), 1, false);
6615 6686 ConcurrentMarkSweepThread::acknowledge_yield_request();
6616 6687 }
6617 6688
6618 6689 ConcurrentMarkSweepThread::synchronize(true);
6619 6690 _freelistLock->lock_without_safepoint_check();
6620 6691 _bitMap->lock()->lock_without_safepoint_check();
6621 6692 _collector->startTimer();
6622 6693 }
6623 6694
6624 6695
6625 6696 //////////////////////////////////////////////////////////////////
6626 6697 // SurvivorSpacePrecleanClosure
6627 6698 //////////////////////////////////////////////////////////////////
6628 6699 // This (single-threaded) closure is used to preclean the oops in
6629 6700 // the survivor spaces.
6630 6701 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6631 6702
6632 6703 HeapWord* addr = (HeapWord*)p;
6633 6704 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6634 6705 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6635 6706 assert(p->klass() != NULL, "object should be initializd");
6636 6707 assert(p->is_parsable(), "must be parsable.");
6637 6708 // an initialized object; ignore mark word in verification below
6638 6709 // since we are running concurrent with mutators
6639 6710 assert(p->is_oop(true), "should be an oop");
6640 6711 // Note that we do not yield while we iterate over
6641 6712 // the interior oops of p, pushing the relevant ones
6642 6713 // on our marking stack.
6643 6714 size_t size = p->oop_iterate(_scanning_closure);
6644 6715 do_yield_check();
6645 6716 // Observe that below, we do not abandon the preclean
6646 6717 // phase as soon as we should; rather we empty the
6647 6718 // marking stack before returning. This is to satisfy
6648 6719 // some existing assertions. In general, it may be a
6649 6720 // good idea to abort immediately and complete the marking
6650 6721 // from the grey objects at a later time.
6651 6722 while (!_mark_stack->isEmpty()) {
6652 6723 oop new_oop = _mark_stack->pop();
6653 6724 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6654 6725 assert(new_oop->is_parsable(), "Found unparsable oop");
6655 6726 assert(_bit_map->isMarked((HeapWord*)new_oop),
6656 6727 "only grey objects on this stack");
6657 6728 // iterate over the oops in this oop, marking and pushing
6658 6729 // the ones in CMS heap (i.e. in _span).
6659 6730 new_oop->oop_iterate(_scanning_closure);
6660 6731 // check if it's time to yield
6661 6732 do_yield_check();
6662 6733 }
6663 6734 unsigned int after_count =
6664 6735 GenCollectedHeap::heap()->total_collections();
6665 6736 bool abort = (_before_count != after_count) ||
6666 6737 _collector->should_abort_preclean();
6667 6738 return abort ? 0 : size;
6668 6739 }
6669 6740
6670 6741 void SurvivorSpacePrecleanClosure::do_yield_work() {
6671 6742 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6672 6743 "CMS thread should hold CMS token");
6673 6744 assert_lock_strong(_bit_map->lock());
6674 6745 // Relinquish the bit map lock
6675 6746 _bit_map->lock()->unlock();
6676 6747 ConcurrentMarkSweepThread::desynchronize(true);
6677 6748 ConcurrentMarkSweepThread::acknowledge_yield_request();
6678 6749 _collector->stopTimer();
6679 6750 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6680 6751 if (PrintCMSStatistics != 0) {
6681 6752 _collector->incrementYields();
6682 6753 }
6683 6754 _collector->icms_wait();
6684 6755
6685 6756 // See the comment in coordinator_yield()
6686 6757 for (unsigned i = 0; i < CMSYieldSleepCount &&
6687 6758 ConcurrentMarkSweepThread::should_yield() &&
6688 6759 !CMSCollector::foregroundGCIsActive(); ++i) {
6689 6760 os::sleep(Thread::current(), 1, false);
6690 6761 ConcurrentMarkSweepThread::acknowledge_yield_request();
6691 6762 }
6692 6763
6693 6764 ConcurrentMarkSweepThread::synchronize(true);
6694 6765 _bit_map->lock()->lock_without_safepoint_check();
6695 6766 _collector->startTimer();
6696 6767 }
6697 6768
6698 6769 // This closure is used to rescan the marked objects on the dirty cards
6699 6770 // in the mod union table and the card table proper. In the parallel
6700 6771 // case, although the bitMap is shared, we do a single read so the
6701 6772 // isMarked() query is "safe".
6702 6773 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6703 6774 // Ignore mark word because we are running concurrent with mutators
6704 6775 assert(p->is_oop_or_null(true), "expected an oop or null");
6705 6776 HeapWord* addr = (HeapWord*)p;
6706 6777 assert(_span.contains(addr), "we are scanning the CMS generation");
6707 6778 bool is_obj_array = false;
6708 6779 #ifdef DEBUG
6709 6780 if (!_parallel) {
6710 6781 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6711 6782 assert(_collector->overflow_list_is_empty(),
6712 6783 "overflow list should be empty");
6713 6784
6714 6785 }
6715 6786 #endif // DEBUG
6716 6787 if (_bit_map->isMarked(addr)) {
6717 6788 // Obj arrays are precisely marked, non-arrays are not;
6718 6789 // so we scan objArrays precisely and non-arrays in their
6719 6790 // entirety.
6720 6791 if (p->is_objArray()) {
6721 6792 is_obj_array = true;
6722 6793 if (_parallel) {
6723 6794 p->oop_iterate(_par_scan_closure, mr);
6724 6795 } else {
6725 6796 p->oop_iterate(_scan_closure, mr);
6726 6797 }
6727 6798 } else {
6728 6799 if (_parallel) {
6729 6800 p->oop_iterate(_par_scan_closure);
6730 6801 } else {
6731 6802 p->oop_iterate(_scan_closure);
6732 6803 }
6733 6804 }
6734 6805 }
6735 6806 #ifdef DEBUG
6736 6807 if (!_parallel) {
6737 6808 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6738 6809 assert(_collector->overflow_list_is_empty(),
6739 6810 "overflow list should be empty");
6740 6811
6741 6812 }
6742 6813 #endif // DEBUG
6743 6814 return is_obj_array;
6744 6815 }
6745 6816
6746 6817 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6747 6818 MemRegion span,
6748 6819 CMSBitMap* bitMap, CMSMarkStack* markStack,
6749 6820 CMSMarkStack* revisitStack,
6750 6821 bool should_yield, bool verifying):
6751 6822 _collector(collector),
6752 6823 _span(span),
6753 6824 _bitMap(bitMap),
6754 6825 _mut(&collector->_modUnionTable),
6755 6826 _markStack(markStack),
6756 6827 _revisitStack(revisitStack),
6757 6828 _yield(should_yield),
6758 6829 _skipBits(0)
6759 6830 {
6760 6831 assert(_markStack->isEmpty(), "stack should be empty");
6761 6832 _finger = _bitMap->startWord();
6762 6833 _threshold = _finger;
6763 6834 assert(_collector->_restart_addr == NULL, "Sanity check");
6764 6835 assert(_span.contains(_finger), "Out of bounds _finger?");
6765 6836 DEBUG_ONLY(_verifying = verifying;)
6766 6837 }
6767 6838
6768 6839 void MarkFromRootsClosure::reset(HeapWord* addr) {
6769 6840 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6770 6841 assert(_span.contains(addr), "Out of bounds _finger?");
6771 6842 _finger = addr;
6772 6843 _threshold = (HeapWord*)round_to(
6773 6844 (intptr_t)_finger, CardTableModRefBS::card_size);
6774 6845 }
6775 6846
6776 6847 // Should revisit to see if this should be restructured for
6777 6848 // greater efficiency.
6778 6849 void MarkFromRootsClosure::do_bit(size_t offset) {
6779 6850 if (_skipBits > 0) {
6780 6851 _skipBits--;
6781 6852 return;
6782 6853 }
6783 6854 // convert offset into a HeapWord*
6784 6855 HeapWord* addr = _bitMap->startWord() + offset;
6785 6856 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6786 6857 "address out of range");
6787 6858 assert(_bitMap->isMarked(addr), "tautology");
6788 6859 if (_bitMap->isMarked(addr+1)) {
6789 6860 // this is an allocated but not yet initialized object
6790 6861 assert(_skipBits == 0, "tautology");
6791 6862 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6792 6863 oop p = oop(addr);
6793 6864 if (p->klass() == NULL || !p->is_parsable()) {
6794 6865 DEBUG_ONLY(if (!_verifying) {)
6795 6866 // We re-dirty the cards on which this object lies and increase
6796 6867 // the _threshold so that we'll come back to scan this object
6797 6868 // during the preclean or remark phase. (CMSCleanOnEnter)
6798 6869 if (CMSCleanOnEnter) {
6799 6870 size_t sz = _collector->block_size_using_printezis_bits(addr);
6800 6871 HeapWord* start_card_addr = (HeapWord*)round_down(
6801 6872 (intptr_t)addr, CardTableModRefBS::card_size);
6802 6873 HeapWord* end_card_addr = (HeapWord*)round_to(
6803 6874 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6804 6875 MemRegion redirty_range = MemRegion(start_card_addr, end_card_addr);
6805 6876 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6806 6877 // Bump _threshold to end_card_addr; note that
6807 6878 // _threshold cannot possibly exceed end_card_addr, anyhow.
6808 6879 // This prevents future clearing of the card as the scan proceeds
6809 6880 // to the right.
6810 6881 assert(_threshold <= end_card_addr,
6811 6882 "Because we are just scanning into this object");
6812 6883 if (_threshold < end_card_addr) {
6813 6884 _threshold = end_card_addr;
6814 6885 }
6815 6886 if (p->klass() != NULL) {
6816 6887 // Redirty the range of cards...
6817 6888 _mut->mark_range(redirty_range);
6818 6889 } // ...else the setting of klass will dirty the card anyway.
6819 6890 }
6820 6891 DEBUG_ONLY(})
6821 6892 return;
6822 6893 }
6823 6894 }
6824 6895 scanOopsInOop(addr);
6825 6896 }
6826 6897
6827 6898 // We take a break if we've been at this for a while,
6828 6899 // so as to avoid monopolizing the locks involved.
6829 6900 void MarkFromRootsClosure::do_yield_work() {
6830 6901 // First give up the locks, then yield, then re-lock
6831 6902 // We should probably use a constructor/destructor idiom to
6832 6903 // do this unlock/lock or modify the MutexUnlocker class to
6833 6904 // serve our purpose. XXX
6834 6905 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6835 6906 "CMS thread should hold CMS token");
6836 6907 assert_lock_strong(_bitMap->lock());
6837 6908 _bitMap->lock()->unlock();
6838 6909 ConcurrentMarkSweepThread::desynchronize(true);
6839 6910 ConcurrentMarkSweepThread::acknowledge_yield_request();
6840 6911 _collector->stopTimer();
6841 6912 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6842 6913 if (PrintCMSStatistics != 0) {
6843 6914 _collector->incrementYields();
6844 6915 }
6845 6916 _collector->icms_wait();
6846 6917
6847 6918 // See the comment in coordinator_yield()
6848 6919 for (unsigned i = 0; i < CMSYieldSleepCount &&
6849 6920 ConcurrentMarkSweepThread::should_yield() &&
6850 6921 !CMSCollector::foregroundGCIsActive(); ++i) {
6851 6922 os::sleep(Thread::current(), 1, false);
6852 6923 ConcurrentMarkSweepThread::acknowledge_yield_request();
6853 6924 }
6854 6925
6855 6926 ConcurrentMarkSweepThread::synchronize(true);
6856 6927 _bitMap->lock()->lock_without_safepoint_check();
6857 6928 _collector->startTimer();
6858 6929 }
6859 6930
6860 6931 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6861 6932 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6862 6933 assert(_markStack->isEmpty(),
6863 6934 "should drain stack to limit stack usage");
6864 6935 // convert ptr to an oop preparatory to scanning
6865 6936 oop this_oop = oop(ptr);
6866 6937 // Ignore mark word in verification below, since we
6867 6938 // may be running concurrent with mutators.
6868 6939 assert(this_oop->is_oop(true), "should be an oop");
6869 6940 assert(_finger <= ptr, "_finger runneth ahead");
6870 6941 // advance the finger to right end of this object
6871 6942 _finger = ptr + this_oop->size();
6872 6943 assert(_finger > ptr, "we just incremented it above");
6873 6944 // On large heaps, it may take us some time to get through
6874 6945 // the marking phase (especially if running iCMS). During
6875 6946 // this time it's possible that a lot of mutations have
6876 6947 // accumulated in the card table and the mod union table --
6877 6948 // these mutation records are redundant until we have
6878 6949 // actually traced into the corresponding card.
6879 6950 // Here, we check whether advancing the finger would make
6880 6951 // us cross into a new card, and if so clear corresponding
6881 6952 // cards in the MUT (preclean them in the card-table in the
6882 6953 // future).
6883 6954
6884 6955 DEBUG_ONLY(if (!_verifying) {)
6885 6956 // The clean-on-enter optimization is disabled by default,
6886 6957 // until we fix 6178663.
6887 6958 if (CMSCleanOnEnter && (_finger > _threshold)) {
6888 6959 // [_threshold, _finger) represents the interval
6889 6960 // of cards to be cleared in MUT (or precleaned in card table).
6890 6961 // The set of cards to be cleared is all those that overlap
6891 6962 // with the interval [_threshold, _finger); note that
6892 6963 // _threshold is always kept card-aligned but _finger isn't
6893 6964 // always card-aligned.
6894 6965 HeapWord* old_threshold = _threshold;
6895 6966 assert(old_threshold == (HeapWord*)round_to(
6896 6967 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6897 6968 "_threshold should always be card-aligned");
6898 6969 _threshold = (HeapWord*)round_to(
6899 6970 (intptr_t)_finger, CardTableModRefBS::card_size);
6900 6971 MemRegion mr(old_threshold, _threshold);
6901 6972 assert(!mr.is_empty(), "Control point invariant");
6902 6973 assert(_span.contains(mr), "Should clear within span");
6903 6974 // XXX When _finger crosses from old gen into perm gen
6904 6975 // we may be doing unnecessary cleaning; do better in the
6905 6976 // future by detecting that condition and clearing fewer
6906 6977 // MUT/CT entries.
6907 6978 _mut->clear_range(mr);
6908 6979 }
6909 6980 DEBUG_ONLY(})
6910 6981
6911 6982 // Note: the finger doesn't advance while we drain
6912 6983 // the stack below.
6913 6984 PushOrMarkClosure pushOrMarkClosure(_collector,
6914 6985 _span, _bitMap, _markStack,
6915 6986 _revisitStack,
6916 6987 _finger, this);
6917 6988 bool res = _markStack->push(this_oop);
6918 6989 assert(res, "Empty non-zero size stack should have space for single push");
6919 6990 while (!_markStack->isEmpty()) {
6920 6991 oop new_oop = _markStack->pop();
6921 6992 // Skip verifying header mark word below because we are
6922 6993 // running concurrent with mutators.
6923 6994 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6924 6995 // now scan this oop's oops
6925 6996 new_oop->oop_iterate(&pushOrMarkClosure);
6926 6997 do_yield_check();
6927 6998 }
6928 6999 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6929 7000 }
6930 7001
6931 7002 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
6932 7003 CMSCollector* collector, MemRegion span,
6933 7004 CMSBitMap* bit_map,
6934 7005 OopTaskQueue* work_queue,
6935 7006 CMSMarkStack* overflow_stack,
6936 7007 CMSMarkStack* revisit_stack,
6937 7008 bool should_yield):
6938 7009 _collector(collector),
6939 7010 _whole_span(collector->_span),
6940 7011 _span(span),
6941 7012 _bit_map(bit_map),
6942 7013 _mut(&collector->_modUnionTable),
6943 7014 _work_queue(work_queue),
6944 7015 _overflow_stack(overflow_stack),
6945 7016 _revisit_stack(revisit_stack),
6946 7017 _yield(should_yield),
6947 7018 _skip_bits(0),
6948 7019 _task(task)
6949 7020 {
6950 7021 assert(_work_queue->size() == 0, "work_queue should be empty");
6951 7022 _finger = span.start();
6952 7023 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
6953 7024 assert(_span.contains(_finger), "Out of bounds _finger?");
6954 7025 }
6955 7026
6956 7027 // Should revisit to see if this should be restructured for
6957 7028 // greater efficiency.
6958 7029 void Par_MarkFromRootsClosure::do_bit(size_t offset) {
6959 7030 if (_skip_bits > 0) {
6960 7031 _skip_bits--;
6961 7032 return;
6962 7033 }
6963 7034 // convert offset into a HeapWord*
6964 7035 HeapWord* addr = _bit_map->startWord() + offset;
6965 7036 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6966 7037 "address out of range");
6967 7038 assert(_bit_map->isMarked(addr), "tautology");
6968 7039 if (_bit_map->isMarked(addr+1)) {
6969 7040 // this is an allocated object that might not yet be initialized
6970 7041 assert(_skip_bits == 0, "tautology");
6971 7042 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
6972 7043 oop p = oop(addr);
6973 7044 if (p->klass() == NULL || !p->is_parsable()) {
6974 7045 // in the case of Clean-on-Enter optimization, redirty card
6975 7046 // and avoid clearing card by increasing the threshold.
6976 7047 return;
6977 7048 }
6978 7049 }
6979 7050 scan_oops_in_oop(addr);
6980 7051 }
6981 7052
6982 7053 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6983 7054 assert(_bit_map->isMarked(ptr), "expected bit to be set");
6984 7055 // Should we assert that our work queue is empty or
6985 7056 // below some drain limit?
6986 7057 assert(_work_queue->size() == 0,
6987 7058 "should drain stack to limit stack usage");
6988 7059 // convert ptr to an oop preparatory to scanning
6989 7060 oop this_oop = oop(ptr);
6990 7061 // Ignore mark word in verification below, since we
6991 7062 // may be running concurrent with mutators.
6992 7063 assert(this_oop->is_oop(true), "should be an oop");
6993 7064 assert(_finger <= ptr, "_finger runneth ahead");
6994 7065 // advance the finger to right end of this object
6995 7066 _finger = ptr + this_oop->size();
6996 7067 assert(_finger > ptr, "we just incremented it above");
6997 7068 // On large heaps, it may take us some time to get through
6998 7069 // the marking phase (especially if running iCMS). During
6999 7070 // this time it's possible that a lot of mutations have
7000 7071 // accumulated in the card table and the mod union table --
7001 7072 // these mutation records are redundant until we have
7002 7073 // actually traced into the corresponding card.
7003 7074 // Here, we check whether advancing the finger would make
7004 7075 // us cross into a new card, and if so clear corresponding
7005 7076 // cards in the MUT (preclean them in the card-table in the
7006 7077 // future).
7007 7078
7008 7079 // The clean-on-enter optimization is disabled by default,
7009 7080 // until we fix 6178663.
7010 7081 if (CMSCleanOnEnter && (_finger > _threshold)) {
7011 7082 // [_threshold, _finger) represents the interval
7012 7083 // of cards to be cleared in MUT (or precleaned in card table).
7013 7084 // The set of cards to be cleared is all those that overlap
7014 7085 // with the interval [_threshold, _finger); note that
7015 7086 // _threshold is always kept card-aligned but _finger isn't
7016 7087 // always card-aligned.
7017 7088 HeapWord* old_threshold = _threshold;
7018 7089 assert(old_threshold == (HeapWord*)round_to(
7019 7090 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7020 7091 "_threshold should always be card-aligned");
7021 7092 _threshold = (HeapWord*)round_to(
7022 7093 (intptr_t)_finger, CardTableModRefBS::card_size);
7023 7094 MemRegion mr(old_threshold, _threshold);
7024 7095 assert(!mr.is_empty(), "Control point invariant");
7025 7096 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7026 7097 // XXX When _finger crosses from old gen into perm gen
7027 7098 // we may be doing unnecessary cleaning; do better in the
7028 7099 // future by detecting that condition and clearing fewer
7029 7100 // MUT/CT entries.
7030 7101 _mut->clear_range(mr);
7031 7102 }
7032 7103
7033 7104 // Note: the local finger doesn't advance while we drain
7034 7105 // the stack below, but the global finger sure can and will.
7035 7106 HeapWord** gfa = _task->global_finger_addr();
7036 7107 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7037 7108 _span, _bit_map,
7038 7109 _work_queue,
7039 7110 _overflow_stack,
7040 7111 _revisit_stack,
7041 7112 _finger,
7042 7113 gfa, this);
7043 7114 bool res = _work_queue->push(this_oop); // overflow could occur here
7044 7115 assert(res, "Will hold once we use workqueues");
7045 7116 while (true) {
7046 7117 oop new_oop;
7047 7118 if (!_work_queue->pop_local(new_oop)) {
7048 7119 // We emptied our work_queue; check if there's stuff that can
7049 7120 // be gotten from the overflow stack.
7050 7121 if (CMSConcMarkingTask::get_work_from_overflow_stack(
7051 7122 _overflow_stack, _work_queue)) {
7052 7123 do_yield_check();
7053 7124 continue;
7054 7125 } else { // done
7055 7126 break;
7056 7127 }
7057 7128 }
7058 7129 // Skip verifying header mark word below because we are
7059 7130 // running concurrent with mutators.
7060 7131 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7061 7132 // now scan this oop's oops
7062 7133 new_oop->oop_iterate(&pushOrMarkClosure);
7063 7134 do_yield_check();
7064 7135 }
7065 7136 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7066 7137 }
7067 7138
7068 7139 // Yield in response to a request from VM Thread or
7069 7140 // from mutators.
7070 7141 void Par_MarkFromRootsClosure::do_yield_work() {
7071 7142 assert(_task != NULL, "sanity");
7072 7143 _task->yield();
7073 7144 }
7074 7145
7075 7146 // A variant of the above used for verifying CMS marking work.
7076 7147 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7077 7148 MemRegion span,
7078 7149 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7079 7150 CMSMarkStack* mark_stack):
7080 7151 _collector(collector),
7081 7152 _span(span),
7082 7153 _verification_bm(verification_bm),
7083 7154 _cms_bm(cms_bm),
7084 7155 _mark_stack(mark_stack),
7085 7156 _pam_verify_closure(collector, span, verification_bm, cms_bm,
7086 7157 mark_stack)
7087 7158 {
7088 7159 assert(_mark_stack->isEmpty(), "stack should be empty");
7089 7160 _finger = _verification_bm->startWord();
7090 7161 assert(_collector->_restart_addr == NULL, "Sanity check");
7091 7162 assert(_span.contains(_finger), "Out of bounds _finger?");
7092 7163 }
7093 7164
7094 7165 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7095 7166 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7096 7167 assert(_span.contains(addr), "Out of bounds _finger?");
7097 7168 _finger = addr;
7098 7169 }
7099 7170
7100 7171 // Should revisit to see if this should be restructured for
7101 7172 // greater efficiency.
7102 7173 void MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7103 7174 // convert offset into a HeapWord*
7104 7175 HeapWord* addr = _verification_bm->startWord() + offset;
7105 7176 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7106 7177 "address out of range");
7107 7178 assert(_verification_bm->isMarked(addr), "tautology");
7108 7179 assert(_cms_bm->isMarked(addr), "tautology");
7109 7180
7110 7181 assert(_mark_stack->isEmpty(),
7111 7182 "should drain stack to limit stack usage");
7112 7183 // convert addr to an oop preparatory to scanning
7113 7184 oop this_oop = oop(addr);
7114 7185 assert(this_oop->is_oop(), "should be an oop");
7115 7186 assert(_finger <= addr, "_finger runneth ahead");
7116 7187 // advance the finger to right end of this object
7117 7188 _finger = addr + this_oop->size();
7118 7189 assert(_finger > addr, "we just incremented it above");
7119 7190 // Note: the finger doesn't advance while we drain
7120 7191 // the stack below.
7121 7192 bool res = _mark_stack->push(this_oop);
7122 7193 assert(res, "Empty non-zero size stack should have space for single push");
7123 7194 while (!_mark_stack->isEmpty()) {
7124 7195 oop new_oop = _mark_stack->pop();
7125 7196 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7126 7197 // now scan this oop's oops
7127 7198 new_oop->oop_iterate(&_pam_verify_closure);
7128 7199 }
7129 7200 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7130 7201 }
7131 7202
7132 7203 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7133 7204 CMSCollector* collector, MemRegion span,
7134 7205 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7135 7206 CMSMarkStack* mark_stack):
7136 7207 OopClosure(collector->ref_processor()),
7137 7208 _collector(collector),
7138 7209 _span(span),
7139 7210 _verification_bm(verification_bm),
7140 7211 _cms_bm(cms_bm),
7141 7212 _mark_stack(mark_stack)
7142 7213 { }
7143 7214
7144 7215
7145 7216 // Upon stack overflow, we discard (part of) the stack,
7146 7217 // remembering the least address amongst those discarded
7147 7218 // in CMSCollector's _restart_address.
7148 7219 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7149 7220 // Remember the least grey address discarded
7150 7221 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7151 7222 _collector->lower_restart_addr(ra);
7152 7223 _mark_stack->reset(); // discard stack contents
7153 7224 _mark_stack->expand(); // expand the stack if possible
7154 7225 }
7155 7226
7156 7227 void PushAndMarkVerifyClosure::do_oop(oop* p) {
7157 7228 oop this_oop = *p;
7158 7229 assert(this_oop->is_oop_or_null(), "expected an oop or NULL");
7159 7230 HeapWord* addr = (HeapWord*)this_oop;
7160 7231 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7161 7232 // Oop lies in _span and isn't yet grey or black
7162 7233 _verification_bm->mark(addr); // now grey
7163 7234 if (!_cms_bm->isMarked(addr)) {
7164 7235 oop(addr)->print();
7165 7236 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
7166 7237 fatal("... aborting");
7167 7238 }
7168 7239
7169 7240 if (!_mark_stack->push(this_oop)) { // stack overflow
7170 7241 if (PrintCMSStatistics != 0) {
7171 7242 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7172 7243 SIZE_FORMAT, _mark_stack->capacity());
7173 7244 }
7174 7245 assert(_mark_stack->isFull(), "Else push should have succeeded");
7175 7246 handle_stack_overflow(addr);
7176 7247 }
7177 7248 // anything including and to the right of _finger
7178 7249 // will be scanned as we iterate over the remainder of the
7179 7250 // bit map
7180 7251 }
7181 7252 }
7182 7253
7183 7254 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7184 7255 MemRegion span,
7185 7256 CMSBitMap* bitMap, CMSMarkStack* markStack,
↓ open down ↓ |
1237 lines elided |
↑ open up ↑ |
7186 7257 CMSMarkStack* revisitStack,
7187 7258 HeapWord* finger, MarkFromRootsClosure* parent) :
7188 7259 OopClosure(collector->ref_processor()),
7189 7260 _collector(collector),
7190 7261 _span(span),
7191 7262 _bitMap(bitMap),
7192 7263 _markStack(markStack),
7193 7264 _revisitStack(revisitStack),
7194 7265 _finger(finger),
7195 7266 _parent(parent),
7196 - _should_remember_klasses(collector->cms_should_unload_classes())
7267 + _should_remember_klasses(collector->should_unload_classes())
7197 7268 { }
7198 7269
7199 7270 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7200 7271 MemRegion span,
7201 7272 CMSBitMap* bit_map,
7202 7273 OopTaskQueue* work_queue,
7203 7274 CMSMarkStack* overflow_stack,
7204 7275 CMSMarkStack* revisit_stack,
7205 7276 HeapWord* finger,
7206 7277 HeapWord** global_finger_addr,
7207 7278 Par_MarkFromRootsClosure* parent) :
7208 7279 OopClosure(collector->ref_processor()),
↓ open down ↓ |
2 lines elided |
↑ open up ↑ |
7209 7280 _collector(collector),
7210 7281 _whole_span(collector->_span),
7211 7282 _span(span),
7212 7283 _bit_map(bit_map),
7213 7284 _work_queue(work_queue),
7214 7285 _overflow_stack(overflow_stack),
7215 7286 _revisit_stack(revisit_stack),
7216 7287 _finger(finger),
7217 7288 _global_finger_addr(global_finger_addr),
7218 7289 _parent(parent),
7219 - _should_remember_klasses(collector->cms_should_unload_classes())
7290 + _should_remember_klasses(collector->should_unload_classes())
7220 7291 { }
7221 7292
7222 7293
7223 7294 void CMSCollector::lower_restart_addr(HeapWord* low) {
7224 7295 assert(_span.contains(low), "Out of bounds addr");
7225 7296 if (_restart_addr == NULL) {
7226 7297 _restart_addr = low;
7227 7298 } else {
7228 7299 _restart_addr = MIN2(_restart_addr, low);
7229 7300 }
7230 7301 }
7231 7302
7232 7303 // Upon stack overflow, we discard (part of) the stack,
7233 7304 // remembering the least address amongst those discarded
7234 7305 // in CMSCollector's _restart_address.
7235 7306 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7236 7307 // Remember the least grey address discarded
7237 7308 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7238 7309 _collector->lower_restart_addr(ra);
7239 7310 _markStack->reset(); // discard stack contents
7240 7311 _markStack->expand(); // expand the stack if possible
7241 7312 }
7242 7313
7243 7314 // Upon stack overflow, we discard (part of) the stack,
7244 7315 // remembering the least address amongst those discarded
7245 7316 // in CMSCollector's _restart_address.
7246 7317 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7247 7318 // We need to do this under a mutex to prevent other
7248 7319 // workers from interfering with the expansion below.
7249 7320 MutexLockerEx ml(_overflow_stack->par_lock(),
7250 7321 Mutex::_no_safepoint_check_flag);
7251 7322 // Remember the least grey address discarded
7252 7323 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7253 7324 _collector->lower_restart_addr(ra);
7254 7325 _overflow_stack->reset(); // discard stack contents
7255 7326 _overflow_stack->expand(); // expand the stack if possible
7256 7327 }
7257 7328
7258 7329
7259 7330 void PushOrMarkClosure::do_oop(oop* p) {
7260 7331 oop thisOop = *p;
7261 7332 // Ignore mark word because we are running concurrent with mutators.
7262 7333 assert(thisOop->is_oop_or_null(true), "expected an oop or NULL");
7263 7334 HeapWord* addr = (HeapWord*)thisOop;
7264 7335 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7265 7336 // Oop lies in _span and isn't yet grey or black
7266 7337 _bitMap->mark(addr); // now grey
7267 7338 if (addr < _finger) {
7268 7339 // the bit map iteration has already either passed, or
7269 7340 // sampled, this bit in the bit map; we'll need to
7270 7341 // use the marking stack to scan this oop's oops.
7271 7342 bool simulate_overflow = false;
7272 7343 NOT_PRODUCT(
7273 7344 if (CMSMarkStackOverflowALot &&
7274 7345 _collector->simulate_overflow()) {
7275 7346 // simulate a stack overflow
7276 7347 simulate_overflow = true;
7277 7348 }
7278 7349 )
7279 7350 if (simulate_overflow || !_markStack->push(thisOop)) { // stack overflow
7280 7351 if (PrintCMSStatistics != 0) {
7281 7352 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7282 7353 SIZE_FORMAT, _markStack->capacity());
7283 7354 }
7284 7355 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7285 7356 handle_stack_overflow(addr);
7286 7357 }
7287 7358 }
7288 7359 // anything including and to the right of _finger
7289 7360 // will be scanned as we iterate over the remainder of the
7290 7361 // bit map
7291 7362 do_yield_check();
7292 7363 }
7293 7364 }
7294 7365
7295 7366 void Par_PushOrMarkClosure::do_oop(oop* p) {
7296 7367 oop this_oop = *p;
7297 7368 // Ignore mark word because we are running concurrent with mutators.
7298 7369 assert(this_oop->is_oop_or_null(true), "expected an oop or NULL");
7299 7370 HeapWord* addr = (HeapWord*)this_oop;
7300 7371 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7301 7372 // Oop lies in _span and isn't yet grey or black
7302 7373 // We read the global_finger (volatile read) strictly after marking oop
7303 7374 bool res = _bit_map->par_mark(addr); // now grey
7304 7375 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7305 7376 // Should we push this marked oop on our stack?
7306 7377 // -- if someone else marked it, nothing to do
7307 7378 // -- if target oop is above global finger nothing to do
7308 7379 // -- if target oop is in chunk and above local finger
7309 7380 // then nothing to do
7310 7381 // -- else push on work queue
7311 7382 if ( !res // someone else marked it, they will deal with it
7312 7383 || (addr >= *gfa) // will be scanned in a later task
7313 7384 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7314 7385 return;
7315 7386 }
7316 7387 // the bit map iteration has already either passed, or
7317 7388 // sampled, this bit in the bit map; we'll need to
7318 7389 // use the marking stack to scan this oop's oops.
7319 7390 bool simulate_overflow = false;
7320 7391 NOT_PRODUCT(
7321 7392 if (CMSMarkStackOverflowALot &&
7322 7393 _collector->simulate_overflow()) {
7323 7394 // simulate a stack overflow
7324 7395 simulate_overflow = true;
7325 7396 }
7326 7397 )
7327 7398 if (simulate_overflow ||
7328 7399 !(_work_queue->push(this_oop) || _overflow_stack->par_push(this_oop))) {
7329 7400 // stack overflow
7330 7401 if (PrintCMSStatistics != 0) {
7331 7402 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7332 7403 SIZE_FORMAT, _overflow_stack->capacity());
7333 7404 }
7334 7405 // We cannot assert that the overflow stack is full because
7335 7406 // it may have been emptied since.
7336 7407 assert(simulate_overflow ||
7337 7408 _work_queue->size() == _work_queue->max_elems(),
7338 7409 "Else push should have succeeded");
7339 7410 handle_stack_overflow(addr);
7340 7411 }
7341 7412 do_yield_check();
7342 7413 }
7343 7414 }
7344 7415
7345 7416
7346 7417 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7347 7418 MemRegion span,
7348 7419 ReferenceProcessor* rp,
7349 7420 CMSBitMap* bit_map,
7350 7421 CMSBitMap* mod_union_table,
7351 7422 CMSMarkStack* mark_stack,
↓ open down ↓ |
122 lines elided |
↑ open up ↑ |
7352 7423 CMSMarkStack* revisit_stack,
7353 7424 bool concurrent_precleaning):
7354 7425 OopClosure(rp),
7355 7426 _collector(collector),
7356 7427 _span(span),
7357 7428 _bit_map(bit_map),
7358 7429 _mod_union_table(mod_union_table),
7359 7430 _mark_stack(mark_stack),
7360 7431 _revisit_stack(revisit_stack),
7361 7432 _concurrent_precleaning(concurrent_precleaning),
7362 - _should_remember_klasses(collector->cms_should_unload_classes())
7433 + _should_remember_klasses(collector->should_unload_classes())
7363 7434 {
7364 7435 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7365 7436 }
7366 7437
7367 7438 // Grey object rescan during pre-cleaning and second checkpoint phases --
7368 7439 // the non-parallel version (the parallel version appears further below.)
7369 7440 void PushAndMarkClosure::do_oop(oop* p) {
7370 7441 oop this_oop = *p;
7371 7442 // Ignore mark word verification. If during concurrent precleaning
7372 7443 // the object monitor may be locked. If during the checkpoint
7373 7444 // phases, the object may already have been reached by a different
7374 7445 // path and may be at the end of the global overflow list (so
7375 7446 // the mark word may be NULL).
7376 7447 assert(this_oop->is_oop_or_null(true/* ignore mark word */),
7377 7448 "expected an oop or NULL");
7378 7449 HeapWord* addr = (HeapWord*)this_oop;
7379 7450 // Check if oop points into the CMS generation
7380 7451 // and is not marked
7381 7452 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7382 7453 // a white object ...
7383 7454 _bit_map->mark(addr); // ... now grey
7384 7455 // push on the marking stack (grey set)
7385 7456 bool simulate_overflow = false;
7386 7457 NOT_PRODUCT(
7387 7458 if (CMSMarkStackOverflowALot &&
7388 7459 _collector->simulate_overflow()) {
7389 7460 // simulate a stack overflow
7390 7461 simulate_overflow = true;
7391 7462 }
7392 7463 )
7393 7464 if (simulate_overflow || !_mark_stack->push(this_oop)) {
7394 7465 if (_concurrent_precleaning) {
7395 7466 // During precleaning we can just dirty the appropriate card
7396 7467 // in the mod union table, thus ensuring that the object remains
7397 7468 // in the grey set and continue. Note that no one can be intefering
7398 7469 // with us in this action of dirtying the mod union table, so
7399 7470 // no locking is required.
7400 7471 _mod_union_table->mark(addr);
7401 7472 _collector->_ser_pmc_preclean_ovflw++;
7402 7473 } else {
7403 7474 // During the remark phase, we need to remember this oop
7404 7475 // in the overflow list.
7405 7476 _collector->push_on_overflow_list(this_oop);
7406 7477 _collector->_ser_pmc_remark_ovflw++;
7407 7478 }
7408 7479 }
7409 7480 }
7410 7481 }
7411 7482
7412 7483 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7413 7484 MemRegion span,
↓ open down ↓ |
41 lines elided |
↑ open up ↑ |
7414 7485 ReferenceProcessor* rp,
7415 7486 CMSBitMap* bit_map,
7416 7487 OopTaskQueue* work_queue,
7417 7488 CMSMarkStack* revisit_stack):
7418 7489 OopClosure(rp),
7419 7490 _collector(collector),
7420 7491 _span(span),
7421 7492 _bit_map(bit_map),
7422 7493 _work_queue(work_queue),
7423 7494 _revisit_stack(revisit_stack),
7424 - _should_remember_klasses(collector->cms_should_unload_classes())
7495 + _should_remember_klasses(collector->should_unload_classes())
7425 7496 {
7426 7497 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7427 7498 }
7428 7499
7429 7500 // Grey object rescan during second checkpoint phase --
7430 7501 // the parallel version.
7431 7502 void Par_PushAndMarkClosure::do_oop(oop* p) {
7432 7503 oop this_oop = *p;
7433 7504 // In the assert below, we ignore the mark word because
7434 7505 // this oop may point to an already visited object that is
7435 7506 // on the overflow stack (in which case the mark word has
7436 7507 // been hijacked for chaining into the overflow stack --
7437 7508 // if this is the last object in the overflow stack then
7438 7509 // its mark word will be NULL). Because this object may
7439 7510 // have been subsequently popped off the global overflow
7440 7511 // stack, and the mark word possibly restored to the prototypical
7441 7512 // value, by the time we get to examined this failing assert in
7442 7513 // the debugger, is_oop_or_null(false) may subsequently start
7443 7514 // to hold.
7444 7515 assert(this_oop->is_oop_or_null(true),
7445 7516 "expected an oop or NULL");
7446 7517 HeapWord* addr = (HeapWord*)this_oop;
7447 7518 // Check if oop points into the CMS generation
7448 7519 // and is not marked
7449 7520 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7450 7521 // a white object ...
7451 7522 // If we manage to "claim" the object, by being the
7452 7523 // first thread to mark it, then we push it on our
7453 7524 // marking stack
7454 7525 if (_bit_map->par_mark(addr)) { // ... now grey
7455 7526 // push on work queue (grey set)
7456 7527 bool simulate_overflow = false;
7457 7528 NOT_PRODUCT(
7458 7529 if (CMSMarkStackOverflowALot &&
7459 7530 _collector->par_simulate_overflow()) {
7460 7531 // simulate a stack overflow
7461 7532 simulate_overflow = true;
7462 7533 }
7463 7534 )
7464 7535 if (simulate_overflow || !_work_queue->push(this_oop)) {
7465 7536 _collector->par_push_on_overflow_list(this_oop);
7466 7537 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7467 7538 }
7468 7539 } // Else, some other thread got there first
7469 7540 }
7470 7541 }
7471 7542
7472 7543 void PushAndMarkClosure::remember_klass(Klass* k) {
7473 7544 if (!_revisit_stack->push(oop(k))) {
7474 7545 fatal("Revisit stack overflowed in PushAndMarkClosure");
7475 7546 }
7476 7547 }
7477 7548
7478 7549 void Par_PushAndMarkClosure::remember_klass(Klass* k) {
7479 7550 if (!_revisit_stack->par_push(oop(k))) {
7480 7551 fatal("Revist stack overflowed in Par_PushAndMarkClosure");
7481 7552 }
7482 7553 }
7483 7554
7484 7555 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7485 7556 Mutex* bml = _collector->bitMapLock();
7486 7557 assert_lock_strong(bml);
7487 7558 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7488 7559 "CMS thread should hold CMS token");
7489 7560
7490 7561 bml->unlock();
7491 7562 ConcurrentMarkSweepThread::desynchronize(true);
7492 7563
7493 7564 ConcurrentMarkSweepThread::acknowledge_yield_request();
7494 7565
7495 7566 _collector->stopTimer();
7496 7567 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7497 7568 if (PrintCMSStatistics != 0) {
7498 7569 _collector->incrementYields();
7499 7570 }
7500 7571 _collector->icms_wait();
7501 7572
7502 7573 // See the comment in coordinator_yield()
7503 7574 for (unsigned i = 0; i < CMSYieldSleepCount &&
7504 7575 ConcurrentMarkSweepThread::should_yield() &&
7505 7576 !CMSCollector::foregroundGCIsActive(); ++i) {
7506 7577 os::sleep(Thread::current(), 1, false);
7507 7578 ConcurrentMarkSweepThread::acknowledge_yield_request();
7508 7579 }
7509 7580
7510 7581 ConcurrentMarkSweepThread::synchronize(true);
7511 7582 bml->lock();
7512 7583
7513 7584 _collector->startTimer();
7514 7585 }
7515 7586
7516 7587 bool CMSPrecleanRefsYieldClosure::should_return() {
7517 7588 if (ConcurrentMarkSweepThread::should_yield()) {
7518 7589 do_yield_work();
7519 7590 }
7520 7591 return _collector->foregroundGCIsActive();
7521 7592 }
7522 7593
7523 7594 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7524 7595 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7525 7596 "mr should be aligned to start at a card boundary");
7526 7597 // We'd like to assert:
7527 7598 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7528 7599 // "mr should be a range of cards");
7529 7600 // However, that would be too strong in one case -- the last
7530 7601 // partition ends at _unallocated_block which, in general, can be
7531 7602 // an arbitrary boundary, not necessarily card aligned.
7532 7603 if (PrintCMSStatistics != 0) {
7533 7604 _num_dirty_cards +=
7534 7605 mr.word_size()/CardTableModRefBS::card_size_in_words;
7535 7606 }
7536 7607 _space->object_iterate_mem(mr, &_scan_cl);
7537 7608 }
7538 7609
7539 7610 SweepClosure::SweepClosure(CMSCollector* collector,
7540 7611 ConcurrentMarkSweepGeneration* g,
7541 7612 CMSBitMap* bitMap, bool should_yield) :
7542 7613 _collector(collector),
7543 7614 _g(g),
7544 7615 _sp(g->cmsSpace()),
7545 7616 _limit(_sp->sweep_limit()),
7546 7617 _freelistLock(_sp->freelistLock()),
7547 7618 _bitMap(bitMap),
7548 7619 _yield(should_yield),
7549 7620 _inFreeRange(false), // No free range at beginning of sweep
7550 7621 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7551 7622 _lastFreeRangeCoalesced(false),
7552 7623 _freeFinger(g->used_region().start())
7553 7624 {
7554 7625 NOT_PRODUCT(
7555 7626 _numObjectsFreed = 0;
7556 7627 _numWordsFreed = 0;
7557 7628 _numObjectsLive = 0;
7558 7629 _numWordsLive = 0;
7559 7630 _numObjectsAlreadyFree = 0;
7560 7631 _numWordsAlreadyFree = 0;
7561 7632 _last_fc = NULL;
7562 7633
7563 7634 _sp->initializeIndexedFreeListArrayReturnedBytes();
7564 7635 _sp->dictionary()->initializeDictReturnedBytes();
7565 7636 )
7566 7637 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7567 7638 "sweep _limit out of bounds");
7568 7639 if (CMSTraceSweeper) {
7569 7640 gclog_or_tty->print("\n====================\nStarting new sweep\n");
7570 7641 }
7571 7642 }
7572 7643
7573 7644 // We need this destructor to reclaim any space at the end
7574 7645 // of the space, which do_blk below may not have added back to
7575 7646 // the free lists. [basically dealing with the "fringe effect"]
7576 7647 SweepClosure::~SweepClosure() {
7577 7648 assert_lock_strong(_freelistLock);
7578 7649 // this should be treated as the end of a free run if any
7579 7650 // The current free range should be returned to the free lists
7580 7651 // as one coalesced chunk.
7581 7652 if (inFreeRange()) {
7582 7653 flushCurFreeChunk(freeFinger(),
7583 7654 pointer_delta(_limit, freeFinger()));
7584 7655 assert(freeFinger() < _limit, "the finger pointeth off base");
7585 7656 if (CMSTraceSweeper) {
7586 7657 gclog_or_tty->print("destructor:");
7587 7658 gclog_or_tty->print("Sweep:put_free_blk 0x%x ("SIZE_FORMAT") "
7588 7659 "[coalesced:"SIZE_FORMAT"]\n",
7589 7660 freeFinger(), pointer_delta(_limit, freeFinger()),
7590 7661 lastFreeRangeCoalesced());
7591 7662 }
7592 7663 }
7593 7664 NOT_PRODUCT(
7594 7665 if (Verbose && PrintGC) {
7595 7666 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, "
7596 7667 SIZE_FORMAT " bytes",
7597 7668 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7598 7669 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
7599 7670 SIZE_FORMAT" bytes "
7600 7671 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
7601 7672 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7602 7673 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7603 7674 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) *
7604 7675 sizeof(HeapWord);
7605 7676 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
7606 7677
7607 7678 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7608 7679 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7609 7680 size_t dictReturnedBytes = _sp->dictionary()->sumDictReturnedBytes();
7610 7681 size_t returnedBytes = indexListReturnedBytes + dictReturnedBytes;
7611 7682 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returnedBytes);
7612 7683 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
7613 7684 indexListReturnedBytes);
7614 7685 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
7615 7686 dictReturnedBytes);
7616 7687 }
7617 7688 }
7618 7689 )
7619 7690 // Now, in debug mode, just null out the sweep_limit
7620 7691 NOT_PRODUCT(_sp->clear_sweep_limit();)
7621 7692 if (CMSTraceSweeper) {
7622 7693 gclog_or_tty->print("end of sweep\n================\n");
7623 7694 }
7624 7695 }
7625 7696
7626 7697 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7627 7698 bool freeRangeInFreeLists) {
7628 7699 if (CMSTraceSweeper) {
7629 7700 gclog_or_tty->print("---- Start free range 0x%x with free block [%d] (%d)\n",
7630 7701 freeFinger, _sp->block_size(freeFinger),
7631 7702 freeRangeInFreeLists);
7632 7703 }
7633 7704 assert(!inFreeRange(), "Trampling existing free range");
7634 7705 set_inFreeRange(true);
7635 7706 set_lastFreeRangeCoalesced(false);
7636 7707
7637 7708 set_freeFinger(freeFinger);
7638 7709 set_freeRangeInFreeLists(freeRangeInFreeLists);
7639 7710 if (CMSTestInFreeList) {
7640 7711 if (freeRangeInFreeLists) {
7641 7712 FreeChunk* fc = (FreeChunk*) freeFinger;
7642 7713 assert(fc->isFree(), "A chunk on the free list should be free.");
7643 7714 assert(fc->size() > 0, "Free range should have a size");
7644 7715 assert(_sp->verifyChunkInFreeLists(fc), "Chunk is not in free lists");
7645 7716 }
7646 7717 }
7647 7718 }
7648 7719
7649 7720 // Note that the sweeper runs concurrently with mutators. Thus,
7650 7721 // it is possible for direct allocation in this generation to happen
7651 7722 // in the middle of the sweep. Note that the sweeper also coalesces
7652 7723 // contiguous free blocks. Thus, unless the sweeper and the allocator
7653 7724 // synchronize appropriately freshly allocated blocks may get swept up.
7654 7725 // This is accomplished by the sweeper locking the free lists while
7655 7726 // it is sweeping. Thus blocks that are determined to be free are
7656 7727 // indeed free. There is however one additional complication:
7657 7728 // blocks that have been allocated since the final checkpoint and
7658 7729 // mark, will not have been marked and so would be treated as
7659 7730 // unreachable and swept up. To prevent this, the allocator marks
7660 7731 // the bit map when allocating during the sweep phase. This leads,
7661 7732 // however, to a further complication -- objects may have been allocated
7662 7733 // but not yet initialized -- in the sense that the header isn't yet
7663 7734 // installed. The sweeper can not then determine the size of the block
7664 7735 // in order to skip over it. To deal with this case, we use a technique
7665 7736 // (due to Printezis) to encode such uninitialized block sizes in the
7666 7737 // bit map. Since the bit map uses a bit per every HeapWord, but the
7667 7738 // CMS generation has a minimum object size of 3 HeapWords, it follows
7668 7739 // that "normal marks" won't be adjacent in the bit map (there will
7669 7740 // always be at least two 0 bits between successive 1 bits). We make use
7670 7741 // of these "unused" bits to represent uninitialized blocks -- the bit
7671 7742 // corresponding to the start of the uninitialized object and the next
7672 7743 // bit are both set. Finally, a 1 bit marks the end of the object that
7673 7744 // started with the two consecutive 1 bits to indicate its potentially
7674 7745 // uninitialized state.
7675 7746
7676 7747 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7677 7748 FreeChunk* fc = (FreeChunk*)addr;
7678 7749 size_t res;
7679 7750
7680 7751 // check if we are done sweepinrg
7681 7752 if (addr == _limit) { // we have swept up to the limit, do nothing more
7682 7753 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7683 7754 "sweep _limit out of bounds");
7684 7755 // help the closure application finish
7685 7756 return pointer_delta(_sp->end(), _limit);
7686 7757 }
7687 7758 assert(addr <= _limit, "sweep invariant");
7688 7759
7689 7760 // check if we should yield
7690 7761 do_yield_check(addr);
7691 7762 if (fc->isFree()) {
7692 7763 // Chunk that is already free
7693 7764 res = fc->size();
7694 7765 doAlreadyFreeChunk(fc);
7695 7766 debug_only(_sp->verifyFreeLists());
7696 7767 assert(res == fc->size(), "Don't expect the size to change");
7697 7768 NOT_PRODUCT(
7698 7769 _numObjectsAlreadyFree++;
7699 7770 _numWordsAlreadyFree += res;
7700 7771 )
7701 7772 NOT_PRODUCT(_last_fc = fc;)
7702 7773 } else if (!_bitMap->isMarked(addr)) {
7703 7774 // Chunk is fresh garbage
7704 7775 res = doGarbageChunk(fc);
7705 7776 debug_only(_sp->verifyFreeLists());
7706 7777 NOT_PRODUCT(
7707 7778 _numObjectsFreed++;
7708 7779 _numWordsFreed += res;
7709 7780 )
7710 7781 } else {
7711 7782 // Chunk that is alive.
7712 7783 res = doLiveChunk(fc);
7713 7784 debug_only(_sp->verifyFreeLists());
7714 7785 NOT_PRODUCT(
7715 7786 _numObjectsLive++;
7716 7787 _numWordsLive += res;
7717 7788 )
7718 7789 }
7719 7790 return res;
7720 7791 }
7721 7792
7722 7793 // For the smart allocation, record following
7723 7794 // split deaths - a free chunk is removed from its free list because
7724 7795 // it is being split into two or more chunks.
7725 7796 // split birth - a free chunk is being added to its free list because
7726 7797 // a larger free chunk has been split and resulted in this free chunk.
7727 7798 // coal death - a free chunk is being removed from its free list because
7728 7799 // it is being coalesced into a large free chunk.
7729 7800 // coal birth - a free chunk is being added to its free list because
7730 7801 // it was created when two or more free chunks where coalesced into
7731 7802 // this free chunk.
7732 7803 //
7733 7804 // These statistics are used to determine the desired number of free
7734 7805 // chunks of a given size. The desired number is chosen to be relative
7735 7806 // to the end of a CMS sweep. The desired number at the end of a sweep
7736 7807 // is the
7737 7808 // count-at-end-of-previous-sweep (an amount that was enough)
7738 7809 // - count-at-beginning-of-current-sweep (the excess)
7739 7810 // + split-births (gains in this size during interval)
7740 7811 // - split-deaths (demands on this size during interval)
7741 7812 // where the interval is from the end of one sweep to the end of the
7742 7813 // next.
7743 7814 //
7744 7815 // When sweeping the sweeper maintains an accumulated chunk which is
7745 7816 // the chunk that is made up of chunks that have been coalesced. That
7746 7817 // will be termed the left-hand chunk. A new chunk of garbage that
7747 7818 // is being considered for coalescing will be referred to as the
7748 7819 // right-hand chunk.
7749 7820 //
7750 7821 // When making a decision on whether to coalesce a right-hand chunk with
7751 7822 // the current left-hand chunk, the current count vs. the desired count
7752 7823 // of the left-hand chunk is considered. Also if the right-hand chunk
7753 7824 // is near the large chunk at the end of the heap (see
7754 7825 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7755 7826 // left-hand chunk is coalesced.
7756 7827 //
7757 7828 // When making a decision about whether to split a chunk, the desired count
7758 7829 // vs. the current count of the candidate to be split is also considered.
7759 7830 // If the candidate is underpopulated (currently fewer chunks than desired)
7760 7831 // a chunk of an overpopulated (currently more chunks than desired) size may
7761 7832 // be chosen. The "hint" associated with a free list, if non-null, points
7762 7833 // to a free list which may be overpopulated.
7763 7834 //
7764 7835
7765 7836 void SweepClosure::doAlreadyFreeChunk(FreeChunk* fc) {
7766 7837 size_t size = fc->size();
7767 7838 // Chunks that cannot be coalesced are not in the
7768 7839 // free lists.
7769 7840 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7770 7841 assert(_sp->verifyChunkInFreeLists(fc),
7771 7842 "free chunk should be in free lists");
7772 7843 }
7773 7844 // a chunk that is already free, should not have been
7774 7845 // marked in the bit map
7775 7846 HeapWord* addr = (HeapWord*) fc;
7776 7847 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7777 7848 // Verify that the bit map has no bits marked between
7778 7849 // addr and purported end of this block.
7779 7850 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7780 7851
7781 7852 // Some chunks cannot be coalesced in under any circumstances.
7782 7853 // See the definition of cantCoalesce().
7783 7854 if (!fc->cantCoalesce()) {
7784 7855 // This chunk can potentially be coalesced.
7785 7856 if (_sp->adaptive_freelists()) {
7786 7857 // All the work is done in
7787 7858 doPostIsFreeOrGarbageChunk(fc, size);
7788 7859 } else { // Not adaptive free lists
7789 7860 // this is a free chunk that can potentially be coalesced by the sweeper;
7790 7861 if (!inFreeRange()) {
7791 7862 // if the next chunk is a free block that can't be coalesced
7792 7863 // it doesn't make sense to remove this chunk from the free lists
7793 7864 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
7794 7865 assert((HeapWord*)nextChunk <= _limit, "sweep invariant");
7795 7866 if ((HeapWord*)nextChunk < _limit && // there's a next chunk...
7796 7867 nextChunk->isFree() && // which is free...
7797 7868 nextChunk->cantCoalesce()) { // ... but cant be coalesced
7798 7869 // nothing to do
7799 7870 } else {
7800 7871 // Potentially the start of a new free range:
7801 7872 // Don't eagerly remove it from the free lists.
7802 7873 // No need to remove it if it will just be put
7803 7874 // back again. (Also from a pragmatic point of view
7804 7875 // if it is a free block in a region that is beyond
7805 7876 // any allocated blocks, an assertion will fail)
7806 7877 // Remember the start of a free run.
7807 7878 initialize_free_range(addr, true);
7808 7879 // end - can coalesce with next chunk
7809 7880 }
7810 7881 } else {
7811 7882 // the midst of a free range, we are coalescing
7812 7883 debug_only(record_free_block_coalesced(fc);)
7813 7884 if (CMSTraceSweeper) {
7814 7885 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size);
7815 7886 }
7816 7887 // remove it from the free lists
7817 7888 _sp->removeFreeChunkFromFreeLists(fc);
7818 7889 set_lastFreeRangeCoalesced(true);
7819 7890 // If the chunk is being coalesced and the current free range is
7820 7891 // in the free lists, remove the current free range so that it
7821 7892 // will be returned to the free lists in its entirety - all
7822 7893 // the coalesced pieces included.
7823 7894 if (freeRangeInFreeLists()) {
7824 7895 FreeChunk* ffc = (FreeChunk*) freeFinger();
7825 7896 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7826 7897 "Size of free range is inconsistent with chunk size.");
7827 7898 if (CMSTestInFreeList) {
7828 7899 assert(_sp->verifyChunkInFreeLists(ffc),
7829 7900 "free range is not in free lists");
7830 7901 }
7831 7902 _sp->removeFreeChunkFromFreeLists(ffc);
7832 7903 set_freeRangeInFreeLists(false);
7833 7904 }
7834 7905 }
7835 7906 }
7836 7907 } else {
7837 7908 // Code path common to both original and adaptive free lists.
7838 7909
7839 7910 // cant coalesce with previous block; this should be treated
7840 7911 // as the end of a free run if any
7841 7912 if (inFreeRange()) {
7842 7913 // we kicked some butt; time to pick up the garbage
7843 7914 assert(freeFinger() < addr, "the finger pointeth off base");
7844 7915 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger()));
7845 7916 }
7846 7917 // else, nothing to do, just continue
7847 7918 }
7848 7919 }
7849 7920
7850 7921 size_t SweepClosure::doGarbageChunk(FreeChunk* fc) {
7851 7922 // This is a chunk of garbage. It is not in any free list.
7852 7923 // Add it to a free list or let it possibly be coalesced into
7853 7924 // a larger chunk.
7854 7925 HeapWord* addr = (HeapWord*) fc;
7855 7926 size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7856 7927
7857 7928 if (_sp->adaptive_freelists()) {
7858 7929 // Verify that the bit map has no bits marked between
7859 7930 // addr and purported end of just dead object.
7860 7931 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7861 7932
7862 7933 doPostIsFreeOrGarbageChunk(fc, size);
7863 7934 } else {
7864 7935 if (!inFreeRange()) {
7865 7936 // start of a new free range
7866 7937 assert(size > 0, "A free range should have a size");
7867 7938 initialize_free_range(addr, false);
7868 7939
7869 7940 } else {
7870 7941 // this will be swept up when we hit the end of the
7871 7942 // free range
7872 7943 if (CMSTraceSweeper) {
7873 7944 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size);
7874 7945 }
7875 7946 // If the chunk is being coalesced and the current free range is
7876 7947 // in the free lists, remove the current free range so that it
7877 7948 // will be returned to the free lists in its entirety - all
7878 7949 // the coalesced pieces included.
7879 7950 if (freeRangeInFreeLists()) {
7880 7951 FreeChunk* ffc = (FreeChunk*)freeFinger();
7881 7952 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7882 7953 "Size of free range is inconsistent with chunk size.");
7883 7954 if (CMSTestInFreeList) {
7884 7955 assert(_sp->verifyChunkInFreeLists(ffc),
7885 7956 "free range is not in free lists");
7886 7957 }
7887 7958 _sp->removeFreeChunkFromFreeLists(ffc);
7888 7959 set_freeRangeInFreeLists(false);
7889 7960 }
7890 7961 set_lastFreeRangeCoalesced(true);
7891 7962 }
7892 7963 // this will be swept up when we hit the end of the free range
7893 7964
7894 7965 // Verify that the bit map has no bits marked between
7895 7966 // addr and purported end of just dead object.
7896 7967 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7897 7968 }
7898 7969 return size;
7899 7970 }
7900 7971
7901 7972 size_t SweepClosure::doLiveChunk(FreeChunk* fc) {
7902 7973 HeapWord* addr = (HeapWord*) fc;
7903 7974 // The sweeper has just found a live object. Return any accumulated
7904 7975 // left hand chunk to the free lists.
7905 7976 if (inFreeRange()) {
7906 7977 if (_sp->adaptive_freelists()) {
7907 7978 flushCurFreeChunk(freeFinger(),
7908 7979 pointer_delta(addr, freeFinger()));
7909 7980 } else { // not adaptive freelists
7910 7981 set_inFreeRange(false);
7911 7982 // Add the free range back to the free list if it is not already
7912 7983 // there.
7913 7984 if (!freeRangeInFreeLists()) {
7914 7985 assert(freeFinger() < addr, "the finger pointeth off base");
7915 7986 if (CMSTraceSweeper) {
7916 7987 gclog_or_tty->print("Sweep:put_free_blk 0x%x (%d) "
7917 7988 "[coalesced:%d]\n",
7918 7989 freeFinger(), pointer_delta(addr, freeFinger()),
7919 7990 lastFreeRangeCoalesced());
7920 7991 }
7921 7992 _sp->addChunkAndRepairOffsetTable(freeFinger(),
7922 7993 pointer_delta(addr, freeFinger()), lastFreeRangeCoalesced());
7923 7994 }
7924 7995 }
7925 7996 }
7926 7997
7927 7998 // Common code path for original and adaptive free lists.
7928 7999
7929 8000 // this object is live: we'd normally expect this to be
7930 8001 // an oop, and like to assert the following:
7931 8002 // assert(oop(addr)->is_oop(), "live block should be an oop");
7932 8003 // However, as we commented above, this may be an object whose
7933 8004 // header hasn't yet been initialized.
7934 8005 size_t size;
7935 8006 assert(_bitMap->isMarked(addr), "Tautology for this control point");
↓ open down ↓ |
501 lines elided |
↑ open up ↑ |
7936 8007 if (_bitMap->isMarked(addr + 1)) {
7937 8008 // Determine the size from the bit map, rather than trying to
7938 8009 // compute it from the object header.
7939 8010 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7940 8011 size = pointer_delta(nextOneAddr + 1, addr);
7941 8012 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7942 8013 "alignment problem");
7943 8014
7944 8015 #ifdef DEBUG
7945 8016 if (oop(addr)->klass() != NULL &&
7946 - ( !_collector->cms_should_unload_classes()
8017 + ( !_collector->should_unload_classes()
7947 8018 || oop(addr)->is_parsable())) {
7948 8019 // Ignore mark word because we are running concurrent with mutators
7949 8020 assert(oop(addr)->is_oop(true), "live block should be an oop");
7950 8021 assert(size ==
7951 8022 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7952 8023 "P-mark and computed size do not agree");
7953 8024 }
7954 8025 #endif
7955 8026
7956 8027 } else {
7957 8028 // This should be an initialized object that's alive.
7958 8029 assert(oop(addr)->klass() != NULL &&
7959 - (!_collector->cms_should_unload_classes()
8030 + (!_collector->should_unload_classes()
7960 8031 || oop(addr)->is_parsable()),
7961 8032 "Should be an initialized object");
7962 8033 // Ignore mark word because we are running concurrent with mutators
7963 8034 assert(oop(addr)->is_oop(true), "live block should be an oop");
7964 8035 // Verify that the bit map has no bits marked between
7965 8036 // addr and purported end of this block.
7966 8037 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7967 8038 assert(size >= 3, "Necessary for Printezis marks to work");
7968 8039 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7969 8040 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7970 8041 }
7971 8042 return size;
7972 8043 }
7973 8044
7974 8045 void SweepClosure::doPostIsFreeOrGarbageChunk(FreeChunk* fc,
7975 8046 size_t chunkSize) {
7976 8047 // doPostIsFreeOrGarbageChunk() should only be called in the smart allocation
7977 8048 // scheme.
7978 8049 bool fcInFreeLists = fc->isFree();
7979 8050 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
7980 8051 assert((HeapWord*)fc <= _limit, "sweep invariant");
7981 8052 if (CMSTestInFreeList && fcInFreeLists) {
7982 8053 assert(_sp->verifyChunkInFreeLists(fc),
7983 8054 "free chunk is not in free lists");
7984 8055 }
7985 8056
7986 8057
7987 8058 if (CMSTraceSweeper) {
7988 8059 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
7989 8060 }
7990 8061
7991 8062 HeapWord* addr = (HeapWord*) fc;
7992 8063
7993 8064 bool coalesce;
7994 8065 size_t left = pointer_delta(addr, freeFinger());
7995 8066 size_t right = chunkSize;
7996 8067 switch (FLSCoalescePolicy) {
7997 8068 // numeric value forms a coalition aggressiveness metric
7998 8069 case 0: { // never coalesce
7999 8070 coalesce = false;
8000 8071 break;
8001 8072 }
8002 8073 case 1: { // coalesce if left & right chunks on overpopulated lists
8003 8074 coalesce = _sp->coalOverPopulated(left) &&
8004 8075 _sp->coalOverPopulated(right);
8005 8076 break;
8006 8077 }
8007 8078 case 2: { // coalesce if left chunk on overpopulated list (default)
8008 8079 coalesce = _sp->coalOverPopulated(left);
8009 8080 break;
8010 8081 }
8011 8082 case 3: { // coalesce if left OR right chunk on overpopulated list
8012 8083 coalesce = _sp->coalOverPopulated(left) ||
8013 8084 _sp->coalOverPopulated(right);
8014 8085 break;
8015 8086 }
8016 8087 case 4: { // always coalesce
8017 8088 coalesce = true;
8018 8089 break;
8019 8090 }
8020 8091 default:
8021 8092 ShouldNotReachHere();
8022 8093 }
8023 8094
8024 8095 // Should the current free range be coalesced?
8025 8096 // If the chunk is in a free range and either we decided to coalesce above
8026 8097 // or the chunk is near the large block at the end of the heap
8027 8098 // (isNearLargestChunk() returns true), then coalesce this chunk.
8028 8099 bool doCoalesce = inFreeRange() &&
8029 8100 (coalesce || _g->isNearLargestChunk((HeapWord*)fc));
8030 8101 if (doCoalesce) {
8031 8102 // Coalesce the current free range on the left with the new
8032 8103 // chunk on the right. If either is on a free list,
8033 8104 // it must be removed from the list and stashed in the closure.
8034 8105 if (freeRangeInFreeLists()) {
8035 8106 FreeChunk* ffc = (FreeChunk*)freeFinger();
8036 8107 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8037 8108 "Size of free range is inconsistent with chunk size.");
8038 8109 if (CMSTestInFreeList) {
8039 8110 assert(_sp->verifyChunkInFreeLists(ffc),
8040 8111 "Chunk is not in free lists");
8041 8112 }
8042 8113 _sp->coalDeath(ffc->size());
8043 8114 _sp->removeFreeChunkFromFreeLists(ffc);
8044 8115 set_freeRangeInFreeLists(false);
8045 8116 }
8046 8117 if (fcInFreeLists) {
8047 8118 _sp->coalDeath(chunkSize);
8048 8119 assert(fc->size() == chunkSize,
8049 8120 "The chunk has the wrong size or is not in the free lists");
8050 8121 _sp->removeFreeChunkFromFreeLists(fc);
8051 8122 }
8052 8123 set_lastFreeRangeCoalesced(true);
8053 8124 } else { // not in a free range and/or should not coalesce
8054 8125 // Return the current free range and start a new one.
8055 8126 if (inFreeRange()) {
8056 8127 // In a free range but cannot coalesce with the right hand chunk.
8057 8128 // Put the current free range into the free lists.
8058 8129 flushCurFreeChunk(freeFinger(),
8059 8130 pointer_delta(addr, freeFinger()));
8060 8131 }
8061 8132 // Set up for new free range. Pass along whether the right hand
8062 8133 // chunk is in the free lists.
8063 8134 initialize_free_range((HeapWord*)fc, fcInFreeLists);
8064 8135 }
8065 8136 }
8066 8137 void SweepClosure::flushCurFreeChunk(HeapWord* chunk, size_t size) {
8067 8138 assert(inFreeRange(), "Should only be called if currently in a free range.");
8068 8139 assert(size > 0,
8069 8140 "A zero sized chunk cannot be added to the free lists.");
8070 8141 if (!freeRangeInFreeLists()) {
8071 8142 if(CMSTestInFreeList) {
8072 8143 FreeChunk* fc = (FreeChunk*) chunk;
8073 8144 fc->setSize(size);
8074 8145 assert(!_sp->verifyChunkInFreeLists(fc),
8075 8146 "chunk should not be in free lists yet");
8076 8147 }
8077 8148 if (CMSTraceSweeper) {
8078 8149 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8079 8150 chunk, size);
8080 8151 }
8081 8152 // A new free range is going to be starting. The current
8082 8153 // free range has not been added to the free lists yet or
8083 8154 // was removed so add it back.
8084 8155 // If the current free range was coalesced, then the death
8085 8156 // of the free range was recorded. Record a birth now.
8086 8157 if (lastFreeRangeCoalesced()) {
8087 8158 _sp->coalBirth(size);
8088 8159 }
8089 8160 _sp->addChunkAndRepairOffsetTable(chunk, size,
8090 8161 lastFreeRangeCoalesced());
8091 8162 }
8092 8163 set_inFreeRange(false);
8093 8164 set_freeRangeInFreeLists(false);
8094 8165 }
8095 8166
8096 8167 // We take a break if we've been at this for a while,
8097 8168 // so as to avoid monopolizing the locks involved.
8098 8169 void SweepClosure::do_yield_work(HeapWord* addr) {
8099 8170 // Return current free chunk being used for coalescing (if any)
8100 8171 // to the appropriate freelist. After yielding, the next
8101 8172 // free block encountered will start a coalescing range of
8102 8173 // free blocks. If the next free block is adjacent to the
8103 8174 // chunk just flushed, they will need to wait for the next
8104 8175 // sweep to be coalesced.
8105 8176 if (inFreeRange()) {
8106 8177 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger()));
8107 8178 }
8108 8179
8109 8180 // First give up the locks, then yield, then re-lock.
8110 8181 // We should probably use a constructor/destructor idiom to
8111 8182 // do this unlock/lock or modify the MutexUnlocker class to
8112 8183 // serve our purpose. XXX
8113 8184 assert_lock_strong(_bitMap->lock());
8114 8185 assert_lock_strong(_freelistLock);
8115 8186 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8116 8187 "CMS thread should hold CMS token");
8117 8188 _bitMap->lock()->unlock();
8118 8189 _freelistLock->unlock();
8119 8190 ConcurrentMarkSweepThread::desynchronize(true);
8120 8191 ConcurrentMarkSweepThread::acknowledge_yield_request();
8121 8192 _collector->stopTimer();
8122 8193 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8123 8194 if (PrintCMSStatistics != 0) {
8124 8195 _collector->incrementYields();
8125 8196 }
8126 8197 _collector->icms_wait();
8127 8198
8128 8199 // See the comment in coordinator_yield()
8129 8200 for (unsigned i = 0; i < CMSYieldSleepCount &&
8130 8201 ConcurrentMarkSweepThread::should_yield() &&
8131 8202 !CMSCollector::foregroundGCIsActive(); ++i) {
8132 8203 os::sleep(Thread::current(), 1, false);
8133 8204 ConcurrentMarkSweepThread::acknowledge_yield_request();
8134 8205 }
8135 8206
8136 8207 ConcurrentMarkSweepThread::synchronize(true);
8137 8208 _freelistLock->lock();
8138 8209 _bitMap->lock()->lock_without_safepoint_check();
8139 8210 _collector->startTimer();
8140 8211 }
8141 8212
8142 8213 #ifndef PRODUCT
8143 8214 // This is actually very useful in a product build if it can
8144 8215 // be called from the debugger. Compile it into the product
8145 8216 // as needed.
8146 8217 bool debug_verifyChunkInFreeLists(FreeChunk* fc) {
8147 8218 return debug_cms_space->verifyChunkInFreeLists(fc);
8148 8219 }
8149 8220
8150 8221 void SweepClosure::record_free_block_coalesced(FreeChunk* fc) const {
8151 8222 if (CMSTraceSweeper) {
8152 8223 gclog_or_tty->print("Sweep:coal_free_blk 0x%x (%d)\n", fc, fc->size());
8153 8224 }
8154 8225 }
8155 8226 #endif
8156 8227
8157 8228 // CMSIsAliveClosure
8158 8229 bool CMSIsAliveClosure::do_object_b(oop obj) {
8159 8230 HeapWord* addr = (HeapWord*)obj;
8160 8231 return addr != NULL &&
8161 8232 (!_span.contains(addr) || _bit_map->isMarked(addr));
8162 8233 }
8163 8234
8164 8235 // CMSKeepAliveClosure: the serial version
8165 8236 void CMSKeepAliveClosure::do_oop(oop* p) {
8166 8237 oop this_oop = *p;
8167 8238 HeapWord* addr = (HeapWord*)this_oop;
8168 8239 if (_span.contains(addr) &&
8169 8240 !_bit_map->isMarked(addr)) {
8170 8241 _bit_map->mark(addr);
8171 8242 bool simulate_overflow = false;
8172 8243 NOT_PRODUCT(
8173 8244 if (CMSMarkStackOverflowALot &&
8174 8245 _collector->simulate_overflow()) {
8175 8246 // simulate a stack overflow
8176 8247 simulate_overflow = true;
8177 8248 }
8178 8249 )
8179 8250 if (simulate_overflow || !_mark_stack->push(this_oop)) {
8180 8251 _collector->push_on_overflow_list(this_oop);
8181 8252 _collector->_ser_kac_ovflw++;
8182 8253 }
8183 8254 }
8184 8255 }
8185 8256
8186 8257 // CMSParKeepAliveClosure: a parallel version of the above.
8187 8258 // The work queues are private to each closure (thread),
8188 8259 // but (may be) available for stealing by other threads.
8189 8260 void CMSParKeepAliveClosure::do_oop(oop* p) {
8190 8261 oop this_oop = *p;
8191 8262 HeapWord* addr = (HeapWord*)this_oop;
8192 8263 if (_span.contains(addr) &&
8193 8264 !_bit_map->isMarked(addr)) {
8194 8265 // In general, during recursive tracing, several threads
8195 8266 // may be concurrently getting here; the first one to
8196 8267 // "tag" it, claims it.
8197 8268 if (_bit_map->par_mark(addr)) {
8198 8269 bool res = _work_queue->push(this_oop);
8199 8270 assert(res, "Low water mark should be much less than capacity");
8200 8271 // Do a recursive trim in the hope that this will keep
8201 8272 // stack usage lower, but leave some oops for potential stealers
8202 8273 trim_queue(_low_water_mark);
8203 8274 } // Else, another thread got there first
8204 8275 }
8205 8276 }
8206 8277
8207 8278 void CMSParKeepAliveClosure::trim_queue(uint max) {
8208 8279 while (_work_queue->size() > max) {
8209 8280 oop new_oop;
8210 8281 if (_work_queue->pop_local(new_oop)) {
8211 8282 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8212 8283 assert(_bit_map->isMarked((HeapWord*)new_oop),
8213 8284 "no white objects on this stack!");
8214 8285 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8215 8286 // iterate over the oops in this oop, marking and pushing
8216 8287 // the ones in CMS heap (i.e. in _span).
8217 8288 new_oop->oop_iterate(&_mark_and_push);
8218 8289 }
8219 8290 }
8220 8291 }
8221 8292
8222 8293 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) {
8223 8294 oop this_oop = *p;
8224 8295 HeapWord* addr = (HeapWord*)this_oop;
8225 8296 if (_span.contains(addr) &&
8226 8297 !_bit_map->isMarked(addr)) {
8227 8298 if (_bit_map->par_mark(addr)) {
8228 8299 bool simulate_overflow = false;
8229 8300 NOT_PRODUCT(
8230 8301 if (CMSMarkStackOverflowALot &&
8231 8302 _collector->par_simulate_overflow()) {
8232 8303 // simulate a stack overflow
8233 8304 simulate_overflow = true;
8234 8305 }
8235 8306 )
8236 8307 if (simulate_overflow || !_work_queue->push(this_oop)) {
8237 8308 _collector->par_push_on_overflow_list(this_oop);
8238 8309 _collector->_par_kac_ovflw++;
8239 8310 }
8240 8311 } // Else another thread got there already
8241 8312 }
8242 8313 }
8243 8314
8244 8315 //////////////////////////////////////////////////////////////////
8245 8316 // CMSExpansionCause /////////////////////////////
8246 8317 //////////////////////////////////////////////////////////////////
8247 8318 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8248 8319 switch (cause) {
8249 8320 case _no_expansion:
8250 8321 return "No expansion";
8251 8322 case _satisfy_free_ratio:
8252 8323 return "Free ratio";
8253 8324 case _satisfy_promotion:
8254 8325 return "Satisfy promotion";
8255 8326 case _satisfy_allocation:
8256 8327 return "allocation";
8257 8328 case _allocate_par_lab:
8258 8329 return "Par LAB";
8259 8330 case _allocate_par_spooling_space:
8260 8331 return "Par Spooling Space";
8261 8332 case _adaptive_size_policy:
8262 8333 return "Ergonomics";
8263 8334 default:
8264 8335 return "unknown";
8265 8336 }
8266 8337 }
8267 8338
8268 8339 void CMSDrainMarkingStackClosure::do_void() {
8269 8340 // the max number to take from overflow list at a time
8270 8341 const size_t num = _mark_stack->capacity()/4;
8271 8342 while (!_mark_stack->isEmpty() ||
8272 8343 // if stack is empty, check the overflow list
8273 8344 _collector->take_from_overflow_list(num, _mark_stack)) {
8274 8345 oop this_oop = _mark_stack->pop();
8275 8346 HeapWord* addr = (HeapWord*)this_oop;
8276 8347 assert(_span.contains(addr), "Should be within span");
8277 8348 assert(_bit_map->isMarked(addr), "Should be marked");
8278 8349 assert(this_oop->is_oop(), "Should be an oop");
8279 8350 this_oop->oop_iterate(_keep_alive);
8280 8351 }
8281 8352 }
8282 8353
8283 8354 void CMSParDrainMarkingStackClosure::do_void() {
8284 8355 // drain queue
8285 8356 trim_queue(0);
8286 8357 }
8287 8358
8288 8359 // Trim our work_queue so its length is below max at return
8289 8360 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8290 8361 while (_work_queue->size() > max) {
8291 8362 oop new_oop;
8292 8363 if (_work_queue->pop_local(new_oop)) {
8293 8364 assert(new_oop->is_oop(), "Expected an oop");
8294 8365 assert(_bit_map->isMarked((HeapWord*)new_oop),
8295 8366 "no white objects on this stack!");
8296 8367 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8297 8368 // iterate over the oops in this oop, marking and pushing
8298 8369 // the ones in CMS heap (i.e. in _span).
8299 8370 new_oop->oop_iterate(&_mark_and_push);
8300 8371 }
8301 8372 }
8302 8373 }
8303 8374
8304 8375 ////////////////////////////////////////////////////////////////////
8305 8376 // Support for Marking Stack Overflow list handling and related code
8306 8377 ////////////////////////////////////////////////////////////////////
8307 8378 // Much of the following code is similar in shape and spirit to the
8308 8379 // code used in ParNewGC. We should try and share that code
8309 8380 // as much as possible in the future.
8310 8381
8311 8382 #ifndef PRODUCT
8312 8383 // Debugging support for CMSStackOverflowALot
8313 8384
8314 8385 // It's OK to call this multi-threaded; the worst thing
8315 8386 // that can happen is that we'll get a bunch of closely
8316 8387 // spaced simulated oveflows, but that's OK, in fact
8317 8388 // probably good as it would exercise the overflow code
8318 8389 // under contention.
8319 8390 bool CMSCollector::simulate_overflow() {
8320 8391 if (_overflow_counter-- <= 0) { // just being defensive
8321 8392 _overflow_counter = CMSMarkStackOverflowInterval;
8322 8393 return true;
8323 8394 } else {
8324 8395 return false;
8325 8396 }
8326 8397 }
8327 8398
8328 8399 bool CMSCollector::par_simulate_overflow() {
8329 8400 return simulate_overflow();
8330 8401 }
8331 8402 #endif
8332 8403
8333 8404 // Single-threaded
8334 8405 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8335 8406 assert(stack->isEmpty(), "Expected precondition");
8336 8407 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8337 8408 size_t i = num;
8338 8409 oop cur = _overflow_list;
8339 8410 const markOop proto = markOopDesc::prototype();
8340 8411 NOT_PRODUCT(size_t n = 0;)
8341 8412 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8342 8413 next = oop(cur->mark());
8343 8414 cur->set_mark(proto); // until proven otherwise
8344 8415 assert(cur->is_oop(), "Should be an oop");
8345 8416 bool res = stack->push(cur);
8346 8417 assert(res, "Bit off more than can chew?");
8347 8418 NOT_PRODUCT(n++;)
8348 8419 }
8349 8420 _overflow_list = cur;
8350 8421 #ifndef PRODUCT
8351 8422 assert(_num_par_pushes >= n, "Too many pops?");
8352 8423 _num_par_pushes -=n;
8353 8424 #endif
8354 8425 return !stack->isEmpty();
8355 8426 }
8356 8427
8357 8428 // Multi-threaded; use CAS to break off a prefix
8358 8429 bool CMSCollector::par_take_from_overflow_list(size_t num,
8359 8430 OopTaskQueue* work_q) {
8360 8431 assert(work_q->size() == 0, "That's the current policy");
8361 8432 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8362 8433 if (_overflow_list == NULL) {
8363 8434 return false;
8364 8435 }
8365 8436 // Grab the entire list; we'll put back a suffix
8366 8437 oop prefix = (oop)Atomic::xchg_ptr(NULL, &_overflow_list);
8367 8438 if (prefix == NULL) { // someone grabbed it before we did ...
8368 8439 // ... we could spin for a short while, but for now we don't
8369 8440 return false;
8370 8441 }
8371 8442 size_t i = num;
8372 8443 oop cur = prefix;
8373 8444 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8374 8445 if (cur->mark() != NULL) {
8375 8446 oop suffix_head = cur->mark(); // suffix will be put back on global list
8376 8447 cur->set_mark(NULL); // break off suffix
8377 8448 // Find tail of suffix so we can prepend suffix to global list
8378 8449 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8379 8450 oop suffix_tail = cur;
8380 8451 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8381 8452 "Tautology");
8382 8453 oop observed_overflow_list = _overflow_list;
8383 8454 do {
8384 8455 cur = observed_overflow_list;
8385 8456 suffix_tail->set_mark(markOop(cur));
8386 8457 observed_overflow_list =
8387 8458 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur);
8388 8459 } while (cur != observed_overflow_list);
8389 8460 }
8390 8461
8391 8462 // Push the prefix elements on work_q
8392 8463 assert(prefix != NULL, "control point invariant");
8393 8464 const markOop proto = markOopDesc::prototype();
8394 8465 oop next;
8395 8466 NOT_PRODUCT(size_t n = 0;)
8396 8467 for (cur = prefix; cur != NULL; cur = next) {
8397 8468 next = oop(cur->mark());
8398 8469 cur->set_mark(proto); // until proven otherwise
8399 8470 assert(cur->is_oop(), "Should be an oop");
8400 8471 bool res = work_q->push(cur);
8401 8472 assert(res, "Bit off more than we can chew?");
8402 8473 NOT_PRODUCT(n++;)
8403 8474 }
8404 8475 #ifndef PRODUCT
8405 8476 assert(_num_par_pushes >= n, "Too many pops?");
8406 8477 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8407 8478 #endif
8408 8479 return true;
8409 8480 }
8410 8481
8411 8482 // Single-threaded
8412 8483 void CMSCollector::push_on_overflow_list(oop p) {
8413 8484 NOT_PRODUCT(_num_par_pushes++;)
8414 8485 assert(p->is_oop(), "Not an oop");
8415 8486 preserve_mark_if_necessary(p);
8416 8487 p->set_mark((markOop)_overflow_list);
8417 8488 _overflow_list = p;
8418 8489 }
8419 8490
8420 8491 // Multi-threaded; use CAS to prepend to overflow list
8421 8492 void CMSCollector::par_push_on_overflow_list(oop p) {
8422 8493 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8423 8494 assert(p->is_oop(), "Not an oop");
8424 8495 par_preserve_mark_if_necessary(p);
8425 8496 oop observed_overflow_list = _overflow_list;
8426 8497 oop cur_overflow_list;
8427 8498 do {
8428 8499 cur_overflow_list = observed_overflow_list;
8429 8500 p->set_mark(markOop(cur_overflow_list));
8430 8501 observed_overflow_list =
8431 8502 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8432 8503 } while (cur_overflow_list != observed_overflow_list);
8433 8504 }
8434 8505
8435 8506 // Single threaded
8436 8507 // General Note on GrowableArray: pushes may silently fail
8437 8508 // because we are (temporarily) out of C-heap for expanding
8438 8509 // the stack. The problem is quite ubiquitous and affects
8439 8510 // a lot of code in the JVM. The prudent thing for GrowableArray
8440 8511 // to do (for now) is to exit with an error. However, that may
8441 8512 // be too draconian in some cases because the caller may be
8442 8513 // able to recover without much harm. For suych cases, we
8443 8514 // should probably introduce a "soft_push" method which returns
8444 8515 // an indication of success or failure with the assumption that
8445 8516 // the caller may be able to recover from a failure; code in
8446 8517 // the VM can then be changed, incrementally, to deal with such
8447 8518 // failures where possible, thus, incrementally hardening the VM
8448 8519 // in such low resource situations.
8449 8520 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8450 8521 int PreserveMarkStackSize = 128;
8451 8522
8452 8523 if (_preserved_oop_stack == NULL) {
8453 8524 assert(_preserved_mark_stack == NULL,
8454 8525 "bijection with preserved_oop_stack");
8455 8526 // Allocate the stacks
8456 8527 _preserved_oop_stack = new (ResourceObj::C_HEAP)
8457 8528 GrowableArray<oop>(PreserveMarkStackSize, true);
8458 8529 _preserved_mark_stack = new (ResourceObj::C_HEAP)
8459 8530 GrowableArray<markOop>(PreserveMarkStackSize, true);
8460 8531 if (_preserved_oop_stack == NULL || _preserved_mark_stack == NULL) {
8461 8532 vm_exit_out_of_memory(2* PreserveMarkStackSize * sizeof(oop) /* punt */,
8462 8533 "Preserved Mark/Oop Stack for CMS (C-heap)");
8463 8534 }
8464 8535 }
8465 8536 _preserved_oop_stack->push(p);
8466 8537 _preserved_mark_stack->push(m);
8467 8538 assert(m == p->mark(), "Mark word changed");
8468 8539 assert(_preserved_oop_stack->length() == _preserved_mark_stack->length(),
8469 8540 "bijection");
8470 8541 }
8471 8542
8472 8543 // Single threaded
8473 8544 void CMSCollector::preserve_mark_if_necessary(oop p) {
8474 8545 markOop m = p->mark();
8475 8546 if (m->must_be_preserved(p)) {
8476 8547 preserve_mark_work(p, m);
8477 8548 }
8478 8549 }
8479 8550
8480 8551 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8481 8552 markOop m = p->mark();
8482 8553 if (m->must_be_preserved(p)) {
8483 8554 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8484 8555 // Even though we read the mark word without holding
8485 8556 // the lock, we are assured that it will not change
8486 8557 // because we "own" this oop, so no other thread can
8487 8558 // be trying to push it on the overflow list; see
8488 8559 // the assertion in preserve_mark_work() that checks
8489 8560 // that m == p->mark().
8490 8561 preserve_mark_work(p, m);
8491 8562 }
8492 8563 }
8493 8564
8494 8565 // We should be able to do this multi-threaded,
8495 8566 // a chunk of stack being a task (this is
8496 8567 // correct because each oop only ever appears
8497 8568 // once in the overflow list. However, it's
8498 8569 // not very easy to completely overlap this with
8499 8570 // other operations, so will generally not be done
8500 8571 // until all work's been completed. Because we
8501 8572 // expect the preserved oop stack (set) to be small,
8502 8573 // it's probably fine to do this single-threaded.
8503 8574 // We can explore cleverer concurrent/overlapped/parallel
8504 8575 // processing of preserved marks if we feel the
8505 8576 // need for this in the future. Stack overflow should
8506 8577 // be so rare in practice and, when it happens, its
8507 8578 // effect on performance so great that this will
8508 8579 // likely just be in the noise anyway.
8509 8580 void CMSCollector::restore_preserved_marks_if_any() {
8510 8581 if (_preserved_oop_stack == NULL) {
8511 8582 assert(_preserved_mark_stack == NULL,
8512 8583 "bijection with preserved_oop_stack");
8513 8584 return;
8514 8585 }
8515 8586
8516 8587 assert(SafepointSynchronize::is_at_safepoint(),
8517 8588 "world should be stopped");
8518 8589 assert(Thread::current()->is_ConcurrentGC_thread() ||
8519 8590 Thread::current()->is_VM_thread(),
8520 8591 "should be single-threaded");
8521 8592
8522 8593 int length = _preserved_oop_stack->length();
8523 8594 assert(_preserved_mark_stack->length() == length, "bijection");
8524 8595 for (int i = 0; i < length; i++) {
8525 8596 oop p = _preserved_oop_stack->at(i);
8526 8597 assert(p->is_oop(), "Should be an oop");
8527 8598 assert(_span.contains(p), "oop should be in _span");
8528 8599 assert(p->mark() == markOopDesc::prototype(),
8529 8600 "Set when taken from overflow list");
8530 8601 markOop m = _preserved_mark_stack->at(i);
8531 8602 p->set_mark(m);
8532 8603 }
8533 8604 _preserved_mark_stack->clear();
8534 8605 _preserved_oop_stack->clear();
8535 8606 assert(_preserved_mark_stack->is_empty() &&
8536 8607 _preserved_oop_stack->is_empty(),
8537 8608 "stacks were cleared above");
8538 8609 }
8539 8610
8540 8611 #ifndef PRODUCT
8541 8612 bool CMSCollector::no_preserved_marks() const {
8542 8613 return ( ( _preserved_mark_stack == NULL
8543 8614 && _preserved_oop_stack == NULL)
8544 8615 || ( _preserved_mark_stack->is_empty()
8545 8616 && _preserved_oop_stack->is_empty()));
8546 8617 }
8547 8618 #endif
8548 8619
8549 8620 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
8550 8621 {
8551 8622 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
8552 8623 CMSAdaptiveSizePolicy* size_policy =
8553 8624 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
8554 8625 assert(size_policy->is_gc_cms_adaptive_size_policy(),
8555 8626 "Wrong type for size policy");
8556 8627 return size_policy;
8557 8628 }
8558 8629
8559 8630 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
8560 8631 size_t desired_promo_size) {
8561 8632 if (cur_promo_size < desired_promo_size) {
8562 8633 size_t expand_bytes = desired_promo_size - cur_promo_size;
8563 8634 if (PrintAdaptiveSizePolicy && Verbose) {
8564 8635 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
8565 8636 "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
8566 8637 expand_bytes);
8567 8638 }
8568 8639 expand(expand_bytes,
8569 8640 MinHeapDeltaBytes,
8570 8641 CMSExpansionCause::_adaptive_size_policy);
8571 8642 } else if (desired_promo_size < cur_promo_size) {
8572 8643 size_t shrink_bytes = cur_promo_size - desired_promo_size;
8573 8644 if (PrintAdaptiveSizePolicy && Verbose) {
8574 8645 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
8575 8646 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
8576 8647 shrink_bytes);
8577 8648 }
8578 8649 shrink(shrink_bytes);
8579 8650 }
8580 8651 }
8581 8652
8582 8653 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
8583 8654 GenCollectedHeap* gch = GenCollectedHeap::heap();
8584 8655 CMSGCAdaptivePolicyCounters* counters =
8585 8656 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
8586 8657 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
8587 8658 "Wrong kind of counters");
8588 8659 return counters;
8589 8660 }
8590 8661
8591 8662
8592 8663 void ASConcurrentMarkSweepGeneration::update_counters() {
8593 8664 if (UsePerfData) {
8594 8665 _space_counters->update_all();
8595 8666 _gen_counters->update_all();
8596 8667 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8597 8668 GenCollectedHeap* gch = GenCollectedHeap::heap();
8598 8669 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
8599 8670 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
8600 8671 "Wrong gc statistics type");
8601 8672 counters->update_counters(gc_stats_l);
8602 8673 }
8603 8674 }
8604 8675
8605 8676 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
8606 8677 if (UsePerfData) {
8607 8678 _space_counters->update_used(used);
8608 8679 _space_counters->update_capacity();
8609 8680 _gen_counters->update_all();
8610 8681
8611 8682 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8612 8683 GenCollectedHeap* gch = GenCollectedHeap::heap();
8613 8684 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
8614 8685 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
8615 8686 "Wrong gc statistics type");
8616 8687 counters->update_counters(gc_stats_l);
8617 8688 }
8618 8689 }
8619 8690
8620 8691 // The desired expansion delta is computed so that:
8621 8692 // . desired free percentage or greater is used
8622 8693 void ASConcurrentMarkSweepGeneration::compute_new_size() {
8623 8694 assert_locked_or_safepoint(Heap_lock);
8624 8695
8625 8696 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
8626 8697
8627 8698 // If incremental collection failed, we just want to expand
8628 8699 // to the limit.
8629 8700 if (incremental_collection_failed()) {
8630 8701 clear_incremental_collection_failed();
8631 8702 grow_to_reserved();
8632 8703 return;
8633 8704 }
8634 8705
8635 8706 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing");
8636 8707
8637 8708 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
8638 8709 "Wrong type of heap");
8639 8710 int prev_level = level() - 1;
8640 8711 assert(prev_level >= 0, "The cms generation is the lowest generation");
8641 8712 Generation* prev_gen = gch->get_gen(prev_level);
8642 8713 assert(prev_gen->kind() == Generation::ASParNew,
8643 8714 "Wrong type of young generation");
8644 8715 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen;
8645 8716 size_t cur_eden = younger_gen->eden()->capacity();
8646 8717 CMSAdaptiveSizePolicy* size_policy = cms_size_policy();
8647 8718 size_t cur_promo = free();
8648 8719 size_policy->compute_tenured_generation_free_space(cur_promo,
8649 8720 max_available(),
8650 8721 cur_eden);
8651 8722 resize(cur_promo, size_policy->promo_size());
8652 8723
8653 8724 // Record the new size of the space in the cms generation
8654 8725 // that is available for promotions. This is temporary.
8655 8726 // It should be the desired promo size.
8656 8727 size_policy->avg_cms_promo()->sample(free());
8657 8728 size_policy->avg_old_live()->sample(used());
8658 8729
8659 8730 if (UsePerfData) {
8660 8731 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8661 8732 counters->update_cms_capacity_counter(capacity());
8662 8733 }
8663 8734 }
8664 8735
8665 8736 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
8666 8737 assert_locked_or_safepoint(Heap_lock);
8667 8738 assert_lock_strong(freelistLock());
8668 8739 HeapWord* old_end = _cmsSpace->end();
8669 8740 HeapWord* unallocated_start = _cmsSpace->unallocated_block();
8670 8741 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
8671 8742 FreeChunk* chunk_at_end = find_chunk_at_end();
8672 8743 if (chunk_at_end == NULL) {
8673 8744 // No room to shrink
8674 8745 if (PrintGCDetails && Verbose) {
8675 8746 gclog_or_tty->print_cr("No room to shrink: old_end "
8676 8747 PTR_FORMAT " unallocated_start " PTR_FORMAT
8677 8748 " chunk_at_end " PTR_FORMAT,
8678 8749 old_end, unallocated_start, chunk_at_end);
8679 8750 }
8680 8751 return;
8681 8752 } else {
8682 8753
8683 8754 // Find the chunk at the end of the space and determine
8684 8755 // how much it can be shrunk.
8685 8756 size_t shrinkable_size_in_bytes = chunk_at_end->size();
8686 8757 size_t aligned_shrinkable_size_in_bytes =
8687 8758 align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
8688 8759 assert(unallocated_start <= chunk_at_end->end(),
8689 8760 "Inconsistent chunk at end of space");
8690 8761 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
8691 8762 size_t word_size_before = heap_word_size(_virtual_space.committed_size());
8692 8763
8693 8764 // Shrink the underlying space
8694 8765 _virtual_space.shrink_by(bytes);
8695 8766 if (PrintGCDetails && Verbose) {
8696 8767 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
8697 8768 " desired_bytes " SIZE_FORMAT
8698 8769 " shrinkable_size_in_bytes " SIZE_FORMAT
8699 8770 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
8700 8771 " bytes " SIZE_FORMAT,
8701 8772 desired_bytes, shrinkable_size_in_bytes,
8702 8773 aligned_shrinkable_size_in_bytes, bytes);
8703 8774 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT
8704 8775 " unallocated_start " SIZE_FORMAT,
8705 8776 old_end, unallocated_start);
8706 8777 }
8707 8778
8708 8779 // If the space did shrink (shrinking is not guaranteed),
8709 8780 // shrink the chunk at the end by the appropriate amount.
8710 8781 if (((HeapWord*)_virtual_space.high()) < old_end) {
8711 8782 size_t new_word_size =
8712 8783 heap_word_size(_virtual_space.committed_size());
8713 8784
8714 8785 // Have to remove the chunk from the dictionary because it is changing
8715 8786 // size and might be someplace elsewhere in the dictionary.
8716 8787
8717 8788 // Get the chunk at end, shrink it, and put it
8718 8789 // back.
8719 8790 _cmsSpace->removeChunkFromDictionary(chunk_at_end);
8720 8791 size_t word_size_change = word_size_before - new_word_size;
8721 8792 size_t chunk_at_end_old_size = chunk_at_end->size();
8722 8793 assert(chunk_at_end_old_size >= word_size_change,
8723 8794 "Shrink is too large");
8724 8795 chunk_at_end->setSize(chunk_at_end_old_size -
8725 8796 word_size_change);
8726 8797 _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
8727 8798 word_size_change);
8728 8799
8729 8800 _cmsSpace->returnChunkToDictionary(chunk_at_end);
8730 8801
8731 8802 MemRegion mr(_cmsSpace->bottom(), new_word_size);
8732 8803 _bts->resize(new_word_size); // resize the block offset shared array
8733 8804 Universe::heap()->barrier_set()->resize_covered_region(mr);
8734 8805 _cmsSpace->assert_locked();
8735 8806 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
8736 8807
8737 8808 NOT_PRODUCT(_cmsSpace->dictionary()->verify());
8738 8809
8739 8810 // update the space and generation capacity counters
8740 8811 if (UsePerfData) {
8741 8812 _space_counters->update_capacity();
8742 8813 _gen_counters->update_all();
8743 8814 }
8744 8815
8745 8816 if (Verbose && PrintGCDetails) {
8746 8817 size_t new_mem_size = _virtual_space.committed_size();
8747 8818 size_t old_mem_size = new_mem_size + bytes;
8748 8819 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK",
8749 8820 name(), old_mem_size/K, bytes/K, new_mem_size/K);
8750 8821 }
8751 8822 }
8752 8823
8753 8824 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
8754 8825 "Inconsistency at end of space");
8755 8826 assert(chunk_at_end->end() == _cmsSpace->end(),
8756 8827 "Shrinking is inconsistent");
8757 8828 return;
8758 8829 }
8759 8830 }
8760 8831
8761 8832 // Transfer some number of overflown objects to usual marking
8762 8833 // stack. Return true if some objects were transferred.
8763 8834 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8764 8835 size_t num = MIN2((size_t)_mark_stack->capacity()/4,
8765 8836 (size_t)ParGCDesiredObjsFromOverflowList);
8766 8837
8767 8838 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8768 8839 assert(_collector->overflow_list_is_empty() || res,
8769 8840 "If list is not empty, we should have taken something");
8770 8841 assert(!res || !_mark_stack->isEmpty(),
8771 8842 "If we took something, it should now be on our stack");
8772 8843 return res;
8773 8844 }
8774 8845
8775 8846 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8776 8847 size_t res = _sp->block_size_no_stall(addr, _collector);
8777 8848 assert(res != 0, "Should always be able to compute a size");
8778 8849 if (_sp->block_is_obj(addr)) {
8779 8850 if (_live_bit_map->isMarked(addr)) {
8780 8851 // It can't have been dead in a previous cycle
8781 8852 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8782 8853 } else {
8783 8854 _dead_bit_map->mark(addr); // mark the dead object
8784 8855 }
8785 8856 }
8786 8857 return res;
8787 8858 }
↓ open down ↓ |
818 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX