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),
553 + _sweep_count(0),
525 554 _sweep_estimate(CMS_SweepWeight, CMS_SweepPadding)
526 555 {
527 556 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
528 557 ExplicitGCInvokesConcurrent = true;
529 558 }
530 559 // Now expand the span and allocate the collection support structures
531 560 // (MUT, marking bit map etc.) to cover both generations subject to
532 561 // collection.
533 562
534 563 // First check that _permGen is adjacent to _cmsGen and above it.
535 564 assert( _cmsGen->reserved().word_size() > 0
536 565 && _permGen->reserved().word_size() > 0,
537 566 "generations should not be of zero size");
538 567 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(),
539 568 "_cmsGen and _permGen should not overlap");
540 569 assert(_cmsGen->reserved().end() == _permGen->reserved().start(),
541 570 "_cmsGen->end() different from _permGen->start()");
542 571
543 572 // For use by dirty card to oop closures.
544 573 _cmsGen->cmsSpace()->set_collector(this);
545 574 _permGen->cmsSpace()->set_collector(this);
546 575
547 576 // Adjust my span to cover old (cms) gen and perm gen
548 577 _span = _cmsGen->reserved()._union(_permGen->reserved());
549 578 // Initialize the span of is_alive_closure
550 579 _is_alive_closure.set_span(_span);
551 580
552 581 // Allocate MUT and marking bit map
553 582 {
554 583 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
555 584 if (!_markBitMap.allocate(_span)) {
556 585 warning("Failed to allocate CMS Bit Map");
557 586 return;
558 587 }
559 588 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
560 589 }
561 590 {
562 591 _modUnionTable.allocate(_span);
563 592 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
564 593 }
565 594
566 595 if (!_markStack.allocate(CMSMarkStackSize)) {
567 596 warning("Failed to allocate CMS Marking Stack");
568 597 return;
569 598 }
570 599 if (!_revisitStack.allocate(CMSRevisitStackSize)) {
571 600 warning("Failed to allocate CMS Revisit Stack");
572 601 return;
573 602 }
574 603
575 604 // Support for multi-threaded concurrent phases
576 605 if (ParallelGCThreads > 0 && CMSConcurrentMTEnabled) {
577 606 if (FLAG_IS_DEFAULT(ParallelCMSThreads)) {
578 607 // just for now
579 608 FLAG_SET_DEFAULT(ParallelCMSThreads, (ParallelGCThreads + 3)/4);
580 609 }
581 610 if (ParallelCMSThreads > 1) {
582 611 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
583 612 ParallelCMSThreads, true);
584 613 if (_conc_workers == NULL) {
585 614 warning("GC/CMS: _conc_workers allocation failure: "
586 615 "forcing -CMSConcurrentMTEnabled");
587 616 CMSConcurrentMTEnabled = false;
588 617 }
589 618 } else {
590 619 CMSConcurrentMTEnabled = false;
591 620 }
592 621 }
593 622 if (!CMSConcurrentMTEnabled) {
594 623 ParallelCMSThreads = 0;
595 624 } else {
596 625 // Turn off CMSCleanOnEnter optimization temporarily for
597 626 // the MT case where it's not fixed yet; see 6178663.
598 627 CMSCleanOnEnter = false;
599 628 }
600 629 assert((_conc_workers != NULL) == (ParallelCMSThreads > 1),
601 630 "Inconsistency");
602 631
603 632 // Parallel task queues; these are shared for the
604 633 // concurrent and stop-world phases of CMS, but
605 634 // are not shared with parallel scavenge (ParNew).
606 635 {
607 636 uint i;
608 637 uint num_queues = (uint) MAX2(ParallelGCThreads, ParallelCMSThreads);
609 638
610 639 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
611 640 || ParallelRefProcEnabled)
612 641 && num_queues > 0) {
613 642 _task_queues = new OopTaskQueueSet(num_queues);
614 643 if (_task_queues == NULL) {
615 644 warning("task_queues allocation failure.");
616 645 return;
617 646 }
618 647 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues);
619 648 if (_hash_seed == NULL) {
620 649 warning("_hash_seed array allocation failure");
621 650 return;
622 651 }
623 652
624 653 // XXX use a global constant instead of 64!
625 654 typedef struct OopTaskQueuePadded {
626 655 OopTaskQueue work_queue;
627 656 char pad[64 - sizeof(OopTaskQueue)]; // prevent false sharing
628 657 } OopTaskQueuePadded;
629 658
630 659 for (i = 0; i < num_queues; i++) {
631 660 OopTaskQueuePadded *q_padded = new OopTaskQueuePadded();
632 661 if (q_padded == NULL) {
633 662 warning("work_queue allocation failure.");
634 663 return;
↓ open down ↓ |
100 lines elided |
↑ open up ↑ |
635 664 }
636 665 _task_queues->register_queue(i, &q_padded->work_queue);
637 666 }
638 667 for (i = 0; i < num_queues; i++) {
639 668 _task_queues->queue(i)->initialize();
640 669 _hash_seed[i] = 17; // copied from ParNew
641 670 }
642 671 }
643 672 }
644 673
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 - }
674 + _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
675 + _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio);
676 +
663 677 // Clip CMSBootstrapOccupancy between 0 and 100.
664 - _bootstrap_occupancy = ((double)MIN2((intx)100, MAX2((intx)0, CMSBootstrapOccupancy)))
678 + _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy)))
665 679 /(double)100;
666 680
667 681 _full_gcs_since_conc_gc = 0;
668 682
669 683 // Now tell CMS generations the identity of their collector
670 684 ConcurrentMarkSweepGeneration::set_collector(this);
671 685
672 686 // Create & start a CMS thread for this CMS collector
673 687 _cmsThread = ConcurrentMarkSweepThread::start(this);
674 688 assert(cmsThread() != NULL, "CMS Thread should have been created");
675 689 assert(cmsThread()->collector() == this,
676 690 "CMS Thread should refer to this gen");
677 691 assert(CGC_lock != NULL, "Where's the CGC_lock?");
678 692
679 693 // Support for parallelizing young gen rescan
680 694 GenCollectedHeap* gch = GenCollectedHeap::heap();
681 695 _young_gen = gch->prev_gen(_cmsGen);
682 696 if (gch->supports_inline_contig_alloc()) {
683 697 _top_addr = gch->top_addr();
684 698 _end_addr = gch->end_addr();
685 699 assert(_young_gen != NULL, "no _young_gen");
686 700 _eden_chunk_index = 0;
687 701 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
688 702 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity);
689 703 if (_eden_chunk_array == NULL) {
690 704 _eden_chunk_capacity = 0;
691 705 warning("GC/CMS: _eden_chunk_array allocation failure");
692 706 }
693 707 }
694 708 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
695 709
696 710 // Support for parallelizing survivor space rescan
697 711 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) {
698 712 size_t max_plab_samples = MaxNewSize/((SurvivorRatio+2)*MinTLABSize);
699 713 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads);
700 714 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples);
701 715 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads);
702 716 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
703 717 || _cursor == NULL) {
704 718 warning("Failed to allocate survivor plab/chunk array");
705 719 if (_survivor_plab_array != NULL) {
706 720 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
707 721 _survivor_plab_array = NULL;
708 722 }
709 723 if (_survivor_chunk_array != NULL) {
710 724 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
711 725 _survivor_chunk_array = NULL;
712 726 }
713 727 if (_cursor != NULL) {
714 728 FREE_C_HEAP_ARRAY(size_t, _cursor);
715 729 _cursor = NULL;
716 730 }
717 731 } else {
718 732 _survivor_chunk_capacity = 2*max_plab_samples;
719 733 for (uint i = 0; i < ParallelGCThreads; i++) {
720 734 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples);
721 735 if (vec == NULL) {
722 736 warning("Failed to allocate survivor plab array");
723 737 for (int j = i; j > 0; j--) {
724 738 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array());
725 739 }
726 740 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
727 741 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
728 742 _survivor_plab_array = NULL;
729 743 _survivor_chunk_array = NULL;
730 744 _survivor_chunk_capacity = 0;
731 745 break;
732 746 } else {
733 747 ChunkArray* cur =
734 748 ::new (&_survivor_plab_array[i]) ChunkArray(vec,
735 749 max_plab_samples);
736 750 assert(cur->end() == 0, "Should be 0");
737 751 assert(cur->array() == vec, "Should be vec");
738 752 assert(cur->capacity() == max_plab_samples, "Error");
739 753 }
740 754 }
741 755 }
742 756 }
743 757 assert( ( _survivor_plab_array != NULL
744 758 && _survivor_chunk_array != NULL)
745 759 || ( _survivor_chunk_capacity == 0
746 760 && _survivor_chunk_index == 0),
747 761 "Error");
748 762
749 763 // Choose what strong roots should be scanned depending on verification options
750 764 // and perm gen collection mode.
751 765 if (!CMSClassUnloadingEnabled) {
752 766 // If class unloading is disabled we want to include all classes into the root set.
753 767 add_root_scanning_option(SharedHeap::SO_AllClasses);
754 768 } else {
755 769 add_root_scanning_option(SharedHeap::SO_SystemClasses);
756 770 }
757 771
758 772 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
759 773 _gc_counters = new CollectorCounters("CMS", 1);
760 774 _completed_initialization = true;
761 775 _sweep_timer.start(); // start of time
762 776 }
763 777
764 778 const char* ConcurrentMarkSweepGeneration::name() const {
765 779 return "concurrent mark-sweep generation";
766 780 }
767 781 void ConcurrentMarkSweepGeneration::update_counters() {
768 782 if (UsePerfData) {
769 783 _space_counters->update_all();
770 784 _gen_counters->update_all();
771 785 }
772 786 }
773 787
774 788 // this is an optimized version of update_counters(). it takes the
775 789 // used value as a parameter rather than computing it.
776 790 //
777 791 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
778 792 if (UsePerfData) {
779 793 _space_counters->update_used(used);
780 794 _space_counters->update_capacity();
781 795 _gen_counters->update_all();
782 796 }
783 797 }
784 798
785 799 void ConcurrentMarkSweepGeneration::print() const {
786 800 Generation::print();
787 801 cmsSpace()->print();
788 802 }
789 803
790 804 #ifndef PRODUCT
791 805 void ConcurrentMarkSweepGeneration::print_statistics() {
792 806 cmsSpace()->printFLCensus(0);
793 807 }
794 808 #endif
795 809
796 810 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
797 811 GenCollectedHeap* gch = GenCollectedHeap::heap();
798 812 if (PrintGCDetails) {
799 813 if (Verbose) {
800 814 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
801 815 level(), short_name(), s, used(), capacity());
802 816 } else {
803 817 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
804 818 level(), short_name(), s, used() / K, capacity() / K);
805 819 }
806 820 }
807 821 if (Verbose) {
808 822 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
809 823 gch->used(), gch->capacity());
810 824 } else {
811 825 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
812 826 gch->used() / K, gch->capacity() / K);
813 827 }
814 828 }
815 829
816 830 size_t
817 831 ConcurrentMarkSweepGeneration::contiguous_available() const {
818 832 // dld proposes an improvement in precision here. If the committed
819 833 // part of the space ends in a free block we should add that to
820 834 // uncommitted size in the calculation below. Will make this
821 835 // change later, staying with the approximation below for the
822 836 // time being. -- ysr.
823 837 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
824 838 }
825 839
826 840 size_t
827 841 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
828 842 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
829 843 }
830 844
831 845 size_t ConcurrentMarkSweepGeneration::max_available() const {
832 846 return free() + _virtual_space.uncommitted_size();
833 847 }
834 848
835 849 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(
836 850 size_t max_promotion_in_bytes,
837 851 bool younger_handles_promotion_failure) const {
838 852
839 853 // This is the most conservative test. Full promotion is
840 854 // guaranteed if this is used. The multiplicative factor is to
841 855 // account for the worst case "dilatation".
842 856 double adjusted_max_promo_bytes = _dilatation_factor * max_promotion_in_bytes;
843 857 if (adjusted_max_promo_bytes > (double)max_uintx) { // larger than size_t
844 858 adjusted_max_promo_bytes = (double)max_uintx;
845 859 }
846 860 bool result = (max_contiguous_available() >= (size_t)adjusted_max_promo_bytes);
847 861
848 862 if (younger_handles_promotion_failure && !result) {
849 863 // Full promotion is not guaranteed because fragmentation
850 864 // of the cms generation can prevent the full promotion.
851 865 result = (max_available() >= (size_t)adjusted_max_promo_bytes);
852 866
853 867 if (!result) {
854 868 // With promotion failure handling the test for the ability
855 869 // to support the promotion does not have to be guaranteed.
856 870 // Use an average of the amount promoted.
857 871 result = max_available() >= (size_t)
858 872 gc_stats()->avg_promoted()->padded_average();
859 873 if (PrintGC && Verbose && result) {
860 874 gclog_or_tty->print_cr(
861 875 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
862 876 " max_available: " SIZE_FORMAT
863 877 " avg_promoted: " SIZE_FORMAT,
864 878 max_available(), (size_t)
865 879 gc_stats()->avg_promoted()->padded_average());
866 880 }
867 881 } else {
868 882 if (PrintGC && Verbose) {
869 883 gclog_or_tty->print_cr(
870 884 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
871 885 " max_available: " SIZE_FORMAT
872 886 " adj_max_promo_bytes: " SIZE_FORMAT,
873 887 max_available(), (size_t)adjusted_max_promo_bytes);
874 888 }
875 889 }
876 890 } else {
877 891 if (PrintGC && Verbose) {
878 892 gclog_or_tty->print_cr(
879 893 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
880 894 " contiguous_available: " SIZE_FORMAT
881 895 " adj_max_promo_bytes: " SIZE_FORMAT,
882 896 max_contiguous_available(), (size_t)adjusted_max_promo_bytes);
883 897 }
884 898 }
885 899 return result;
886 900 }
887 901
888 902 CompactibleSpace*
889 903 ConcurrentMarkSweepGeneration::first_compaction_space() const {
890 904 return _cmsSpace;
891 905 }
892 906
893 907 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
894 908 // Clear the promotion information. These pointers can be adjusted
895 909 // along with all the other pointers into the heap but
896 910 // compaction is expected to be a rare event with
897 911 // a heap using cms so don't do it without seeing the need.
898 912 if (ParallelGCThreads > 0) {
899 913 for (uint i = 0; i < ParallelGCThreads; i++) {
900 914 _par_gc_thread_states[i]->promo.reset();
901 915 }
902 916 }
903 917 }
904 918
905 919 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
906 920 blk->do_space(_cmsSpace);
907 921 }
908 922
909 923 void ConcurrentMarkSweepGeneration::compute_new_size() {
910 924 assert_locked_or_safepoint(Heap_lock);
911 925
912 926 // If incremental collection failed, we just want to expand
913 927 // to the limit.
914 928 if (incremental_collection_failed()) {
915 929 clear_incremental_collection_failed();
916 930 grow_to_reserved();
917 931 return;
918 932 }
919 933
920 934 size_t expand_bytes = 0;
921 935 double free_percentage = ((double) free()) / capacity();
922 936 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
923 937 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
924 938
925 939 // compute expansion delta needed for reaching desired free percentage
926 940 if (free_percentage < desired_free_percentage) {
927 941 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
928 942 assert(desired_capacity >= capacity(), "invalid expansion size");
929 943 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
930 944 }
931 945 if (expand_bytes > 0) {
932 946 if (PrintGCDetails && Verbose) {
933 947 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
934 948 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
935 949 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
936 950 gclog_or_tty->print_cr(" Desired free fraction %f",
937 951 desired_free_percentage);
938 952 gclog_or_tty->print_cr(" Maximum free fraction %f",
939 953 maximum_free_percentage);
940 954 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000);
941 955 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
942 956 desired_capacity/1000);
943 957 int prev_level = level() - 1;
944 958 if (prev_level >= 0) {
945 959 size_t prev_size = 0;
946 960 GenCollectedHeap* gch = GenCollectedHeap::heap();
947 961 Generation* prev_gen = gch->_gens[prev_level];
948 962 prev_size = prev_gen->capacity();
949 963 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
950 964 prev_size/1000);
951 965 }
952 966 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
953 967 unsafe_max_alloc_nogc()/1000);
954 968 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
955 969 contiguous_available()/1000);
956 970 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
957 971 expand_bytes);
958 972 }
959 973 // safe if expansion fails
960 974 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
961 975 if (PrintGCDetails && Verbose) {
962 976 gclog_or_tty->print_cr(" Expanded free fraction %f",
963 977 ((double) free()) / capacity());
964 978 }
965 979 }
966 980 }
967 981
968 982 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
969 983 return cmsSpace()->freelistLock();
970 984 }
971 985
972 986 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
973 987 bool tlab) {
974 988 CMSSynchronousYieldRequest yr;
975 989 MutexLockerEx x(freelistLock(),
976 990 Mutex::_no_safepoint_check_flag);
977 991 return have_lock_and_allocate(size, tlab);
978 992 }
979 993
980 994 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
981 995 bool tlab) {
982 996 assert_lock_strong(freelistLock());
983 997 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
984 998 HeapWord* res = cmsSpace()->allocate(adjustedSize);
985 999 // Allocate the object live (grey) if the background collector has
986 1000 // started marking. This is necessary because the marker may
987 1001 // have passed this address and consequently this object will
988 1002 // not otherwise be greyed and would be incorrectly swept up.
989 1003 // Note that if this object contains references, the writing
990 1004 // of those references will dirty the card containing this object
991 1005 // allowing the object to be blackened (and its references scanned)
992 1006 // either during a preclean phase or at the final checkpoint.
993 1007 if (res != NULL) {
994 1008 collector()->direct_allocated(res, adjustedSize);
995 1009 _direct_allocated_words += adjustedSize;
996 1010 // allocation counters
997 1011 NOT_PRODUCT(
998 1012 _numObjectsAllocated++;
999 1013 _numWordsAllocated += (int)adjustedSize;
1000 1014 )
1001 1015 }
1002 1016 return res;
1003 1017 }
1004 1018
1005 1019 // In the case of direct allocation by mutators in a generation that
1006 1020 // is being concurrently collected, the object must be allocated
1007 1021 // live (grey) if the background collector has started marking.
1008 1022 // This is necessary because the marker may
1009 1023 // have passed this address and consequently this object will
1010 1024 // not otherwise be greyed and would be incorrectly swept up.
1011 1025 // Note that if this object contains references, the writing
1012 1026 // of those references will dirty the card containing this object
1013 1027 // allowing the object to be blackened (and its references scanned)
1014 1028 // either during a preclean phase or at the final checkpoint.
1015 1029 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
1016 1030 assert(_markBitMap.covers(start, size), "Out of bounds");
1017 1031 if (_collectorState >= Marking) {
1018 1032 MutexLockerEx y(_markBitMap.lock(),
1019 1033 Mutex::_no_safepoint_check_flag);
1020 1034 // [see comments preceding SweepClosure::do_blk() below for details]
1021 1035 // 1. need to mark the object as live so it isn't collected
1022 1036 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
1023 1037 // 3. need to mark the end of the object so sweeper can skip over it
1024 1038 // if it's uninitialized when the sweeper reaches it.
1025 1039 _markBitMap.mark(start); // object is live
1026 1040 _markBitMap.mark(start + 1); // object is potentially uninitialized?
1027 1041 _markBitMap.mark(start + size - 1);
1028 1042 // mark end of object
1029 1043 }
1030 1044 // check that oop looks uninitialized
1031 1045 assert(oop(start)->klass() == NULL, "_klass should be NULL");
1032 1046 }
1033 1047
1034 1048 void CMSCollector::promoted(bool par, HeapWord* start,
1035 1049 bool is_obj_array, size_t obj_size) {
1036 1050 assert(_markBitMap.covers(start), "Out of bounds");
1037 1051 // See comment in direct_allocated() about when objects should
1038 1052 // be allocated live.
1039 1053 if (_collectorState >= Marking) {
1040 1054 // we already hold the marking bit map lock, taken in
1041 1055 // the prologue
1042 1056 if (par) {
1043 1057 _markBitMap.par_mark(start);
1044 1058 } else {
1045 1059 _markBitMap.mark(start);
1046 1060 }
1047 1061 // We don't need to mark the object as uninitialized (as
1048 1062 // in direct_allocated above) because this is being done with the
1049 1063 // world stopped and the object will be initialized by the
1050 1064 // time the sweeper gets to look at it.
1051 1065 assert(SafepointSynchronize::is_at_safepoint(),
1052 1066 "expect promotion only at safepoints");
1053 1067
1054 1068 if (_collectorState < Sweeping) {
1055 1069 // Mark the appropriate cards in the modUnionTable, so that
1056 1070 // this object gets scanned before the sweep. If this is
1057 1071 // not done, CMS generation references in the object might
1058 1072 // not get marked.
1059 1073 // For the case of arrays, which are otherwise precisely
1060 1074 // marked, we need to dirty the entire array, not just its head.
1061 1075 if (is_obj_array) {
1062 1076 // The [par_]mark_range() method expects mr.end() below to
1063 1077 // be aligned to the granularity of a bit's representation
1064 1078 // in the heap. In the case of the MUT below, that's a
1065 1079 // card size.
1066 1080 MemRegion mr(start,
1067 1081 (HeapWord*)round_to((intptr_t)(start + obj_size),
1068 1082 CardTableModRefBS::card_size /* bytes */));
1069 1083 if (par) {
1070 1084 _modUnionTable.par_mark_range(mr);
1071 1085 } else {
1072 1086 _modUnionTable.mark_range(mr);
1073 1087 }
1074 1088 } else { // not an obj array; we can just mark the head
1075 1089 if (par) {
1076 1090 _modUnionTable.par_mark(start);
1077 1091 } else {
1078 1092 _modUnionTable.mark(start);
1079 1093 }
1080 1094 }
1081 1095 }
1082 1096 }
1083 1097 }
1084 1098
1085 1099 static inline size_t percent_of_space(Space* space, HeapWord* addr)
1086 1100 {
1087 1101 size_t delta = pointer_delta(addr, space->bottom());
1088 1102 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
1089 1103 }
1090 1104
1091 1105 void CMSCollector::icms_update_allocation_limits()
1092 1106 {
1093 1107 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
1094 1108 EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
1095 1109
1096 1110 const unsigned int duty_cycle = stats().icms_update_duty_cycle();
1097 1111 if (CMSTraceIncrementalPacing) {
1098 1112 stats().print();
1099 1113 }
1100 1114
1101 1115 assert(duty_cycle <= 100, "invalid duty cycle");
1102 1116 if (duty_cycle != 0) {
1103 1117 // The duty_cycle is a percentage between 0 and 100; convert to words and
1104 1118 // then compute the offset from the endpoints of the space.
1105 1119 size_t free_words = eden->free() / HeapWordSize;
1106 1120 double free_words_dbl = (double)free_words;
1107 1121 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
1108 1122 size_t offset_words = (free_words - duty_cycle_words) / 2;
1109 1123
1110 1124 _icms_start_limit = eden->top() + offset_words;
1111 1125 _icms_stop_limit = eden->end() - offset_words;
1112 1126
1113 1127 // The limits may be adjusted (shifted to the right) by
1114 1128 // CMSIncrementalOffset, to allow the application more mutator time after a
1115 1129 // young gen gc (when all mutators were stopped) and before CMS starts and
1116 1130 // takes away one or more cpus.
1117 1131 if (CMSIncrementalOffset != 0) {
1118 1132 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
1119 1133 size_t adjustment = (size_t)adjustment_dbl;
1120 1134 HeapWord* tmp_stop = _icms_stop_limit + adjustment;
1121 1135 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
1122 1136 _icms_start_limit += adjustment;
1123 1137 _icms_stop_limit = tmp_stop;
1124 1138 }
1125 1139 }
1126 1140 }
1127 1141 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
1128 1142 _icms_start_limit = _icms_stop_limit = eden->end();
1129 1143 }
1130 1144
1131 1145 // Install the new start limit.
1132 1146 eden->set_soft_end(_icms_start_limit);
1133 1147
1134 1148 if (CMSTraceIncrementalMode) {
1135 1149 gclog_or_tty->print(" icms alloc limits: "
1136 1150 PTR_FORMAT "," PTR_FORMAT
1137 1151 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
1138 1152 _icms_start_limit, _icms_stop_limit,
1139 1153 percent_of_space(eden, _icms_start_limit),
1140 1154 percent_of_space(eden, _icms_stop_limit));
1141 1155 if (Verbose) {
1142 1156 gclog_or_tty->print("eden: ");
1143 1157 eden->print_on(gclog_or_tty);
1144 1158 }
1145 1159 }
1146 1160 }
1147 1161
1148 1162 // Any changes here should try to maintain the invariant
1149 1163 // that if this method is called with _icms_start_limit
1150 1164 // and _icms_stop_limit both NULL, then it should return NULL
1151 1165 // and not notify the icms thread.
1152 1166 HeapWord*
1153 1167 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
1154 1168 size_t word_size)
1155 1169 {
1156 1170 // A start_limit equal to end() means the duty cycle is 0, so treat that as a
1157 1171 // nop.
1158 1172 if (CMSIncrementalMode && _icms_start_limit != space->end()) {
1159 1173 if (top <= _icms_start_limit) {
1160 1174 if (CMSTraceIncrementalMode) {
1161 1175 space->print_on(gclog_or_tty);
1162 1176 gclog_or_tty->stamp();
1163 1177 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
1164 1178 ", new limit=" PTR_FORMAT
1165 1179 " (" SIZE_FORMAT "%%)",
1166 1180 top, _icms_stop_limit,
1167 1181 percent_of_space(space, _icms_stop_limit));
1168 1182 }
1169 1183 ConcurrentMarkSweepThread::start_icms();
1170 1184 assert(top < _icms_stop_limit, "Tautology");
1171 1185 if (word_size < pointer_delta(_icms_stop_limit, top)) {
1172 1186 return _icms_stop_limit;
1173 1187 }
1174 1188
1175 1189 // The allocation will cross both the _start and _stop limits, so do the
1176 1190 // stop notification also and return end().
1177 1191 if (CMSTraceIncrementalMode) {
1178 1192 space->print_on(gclog_or_tty);
1179 1193 gclog_or_tty->stamp();
1180 1194 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
1181 1195 ", new limit=" PTR_FORMAT
1182 1196 " (" SIZE_FORMAT "%%)",
1183 1197 top, space->end(),
1184 1198 percent_of_space(space, space->end()));
1185 1199 }
1186 1200 ConcurrentMarkSweepThread::stop_icms();
1187 1201 return space->end();
1188 1202 }
1189 1203
1190 1204 if (top <= _icms_stop_limit) {
1191 1205 if (CMSTraceIncrementalMode) {
1192 1206 space->print_on(gclog_or_tty);
1193 1207 gclog_or_tty->stamp();
1194 1208 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
1195 1209 ", new limit=" PTR_FORMAT
1196 1210 " (" SIZE_FORMAT "%%)",
1197 1211 top, space->end(),
1198 1212 percent_of_space(space, space->end()));
1199 1213 }
1200 1214 ConcurrentMarkSweepThread::stop_icms();
1201 1215 return space->end();
1202 1216 }
1203 1217
1204 1218 if (CMSTraceIncrementalMode) {
1205 1219 space->print_on(gclog_or_tty);
1206 1220 gclog_or_tty->stamp();
1207 1221 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
1208 1222 ", new limit=" PTR_FORMAT,
1209 1223 top, NULL);
1210 1224 }
1211 1225 }
1212 1226
1213 1227 return NULL;
1214 1228 }
1215 1229
1216 1230 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size, oop* ref) {
1217 1231 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1218 1232 // allocate, copy and if necessary update promoinfo --
1219 1233 // delegate to underlying space.
1220 1234 assert_lock_strong(freelistLock());
1221 1235
1222 1236 #ifndef PRODUCT
1223 1237 if (Universe::heap()->promotion_should_fail()) {
1224 1238 return NULL;
1225 1239 }
1226 1240 #endif // #ifndef PRODUCT
1227 1241
1228 1242 oop res = _cmsSpace->promote(obj, obj_size, ref);
1229 1243 if (res == NULL) {
1230 1244 // expand and retry
1231 1245 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
1232 1246 expand(s*HeapWordSize, MinHeapDeltaBytes,
1233 1247 CMSExpansionCause::_satisfy_promotion);
1234 1248 // Since there's currently no next generation, we don't try to promote
1235 1249 // into a more senior generation.
1236 1250 assert(next_gen() == NULL, "assumption, based upon which no attempt "
1237 1251 "is made to pass on a possibly failing "
1238 1252 "promotion to next generation");
1239 1253 res = _cmsSpace->promote(obj, obj_size, ref);
1240 1254 }
1241 1255 if (res != NULL) {
1242 1256 // See comment in allocate() about when objects should
1243 1257 // be allocated live.
1244 1258 assert(obj->is_oop(), "Will dereference klass pointer below");
1245 1259 collector()->promoted(false, // Not parallel
1246 1260 (HeapWord*)res, obj->is_objArray(), obj_size);
1247 1261 // promotion counters
1248 1262 NOT_PRODUCT(
1249 1263 _numObjectsPromoted++;
1250 1264 _numWordsPromoted +=
1251 1265 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1252 1266 )
1253 1267 }
1254 1268 return res;
1255 1269 }
1256 1270
1257 1271
1258 1272 HeapWord*
1259 1273 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
1260 1274 HeapWord* top,
1261 1275 size_t word_sz)
1262 1276 {
1263 1277 return collector()->allocation_limit_reached(space, top, word_sz);
1264 1278 }
1265 1279
1266 1280 // Things to support parallel young-gen collection.
1267 1281 oop
1268 1282 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1269 1283 oop old, markOop m,
1270 1284 size_t word_sz) {
1271 1285 #ifndef PRODUCT
1272 1286 if (Universe::heap()->promotion_should_fail()) {
1273 1287 return NULL;
1274 1288 }
1275 1289 #endif // #ifndef PRODUCT
1276 1290
1277 1291 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1278 1292 PromotionInfo* promoInfo = &ps->promo;
1279 1293 // if we are tracking promotions, then first ensure space for
1280 1294 // promotion (including spooling space for saving header if necessary).
1281 1295 // then allocate and copy, then track promoted info if needed.
1282 1296 // When tracking (see PromotionInfo::track()), the mark word may
1283 1297 // be displaced and in this case restoration of the mark word
1284 1298 // occurs in the (oop_since_save_marks_)iterate phase.
1285 1299 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1286 1300 // Out of space for allocating spooling buffers;
1287 1301 // try expanding and allocating spooling buffers.
1288 1302 if (!expand_and_ensure_spooling_space(promoInfo)) {
1289 1303 return NULL;
1290 1304 }
1291 1305 }
1292 1306 assert(promoInfo->has_spooling_space(), "Control point invariant");
1293 1307 HeapWord* obj_ptr = ps->lab.alloc(word_sz);
1294 1308 if (obj_ptr == NULL) {
1295 1309 obj_ptr = expand_and_par_lab_allocate(ps, word_sz);
1296 1310 if (obj_ptr == NULL) {
1297 1311 return NULL;
1298 1312 }
1299 1313 }
1300 1314 oop obj = oop(obj_ptr);
1301 1315 assert(obj->klass() == NULL, "Object should be uninitialized here.");
1302 1316 // Otherwise, copy the object. Here we must be careful to insert the
1303 1317 // klass pointer last, since this marks the block as an allocated object.
1304 1318 HeapWord* old_ptr = (HeapWord*)old;
1305 1319 if (word_sz > (size_t)oopDesc::header_size()) {
1306 1320 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1307 1321 obj_ptr + oopDesc::header_size(),
1308 1322 word_sz - oopDesc::header_size());
1309 1323 }
1310 1324 // Restore the mark word copied above.
1311 1325 obj->set_mark(m);
1312 1326 // Now we can track the promoted object, if necessary. We take care
1313 1327 // To delay the transition from uninitialized to full object
1314 1328 // (i.e., insertion of klass pointer) until after, so that it
1315 1329 // atomically becomes a promoted object.
1316 1330 if (promoInfo->tracking()) {
1317 1331 promoInfo->track((PromotedObject*)obj, old->klass());
1318 1332 }
1319 1333 // Finally, install the klass pointer.
1320 1334 obj->set_klass(old->klass());
1321 1335
1322 1336 assert(old->is_oop(), "Will dereference klass ptr below");
1323 1337 collector()->promoted(true, // parallel
1324 1338 obj_ptr, old->is_objArray(), word_sz);
1325 1339
1326 1340 NOT_PRODUCT(
1327 1341 Atomic::inc(&_numObjectsPromoted);
1328 1342 Atomic::add((jint)CompactibleFreeListSpace::adjustObjectSize(obj->size()),
1329 1343 &_numWordsPromoted);
1330 1344 )
1331 1345
1332 1346 return obj;
1333 1347 }
1334 1348
1335 1349 void
1336 1350 ConcurrentMarkSweepGeneration::
1337 1351 par_promote_alloc_undo(int thread_num,
1338 1352 HeapWord* obj, size_t word_sz) {
1339 1353 // CMS does not support promotion undo.
1340 1354 ShouldNotReachHere();
1341 1355 }
1342 1356
1343 1357 void
1344 1358 ConcurrentMarkSweepGeneration::
1345 1359 par_promote_alloc_done(int thread_num) {
1346 1360 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1347 1361 ps->lab.retire();
1348 1362 #if CFLS_LAB_REFILL_STATS
1349 1363 if (thread_num == 0) {
1350 1364 _cmsSpace->print_par_alloc_stats();
1351 1365 }
1352 1366 #endif
1353 1367 }
1354 1368
1355 1369 void
1356 1370 ConcurrentMarkSweepGeneration::
1357 1371 par_oop_since_save_marks_iterate_done(int thread_num) {
1358 1372 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1359 1373 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1360 1374 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1361 1375 }
1362 1376
1363 1377 // XXXPERM
1364 1378 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1365 1379 size_t size,
1366 1380 bool tlab)
1367 1381 {
1368 1382 // We allow a STW collection only if a full
1369 1383 // collection was requested.
1370 1384 return full || should_allocate(size, tlab); // FIX ME !!!
1371 1385 // This and promotion failure handling are connected at the
1372 1386 // hip and should be fixed by untying them.
1373 1387 }
1374 1388
1375 1389 bool CMSCollector::shouldConcurrentCollect() {
1376 1390 if (_full_gc_requested) {
1377 1391 assert(ExplicitGCInvokesConcurrent, "Unexpected state");
1378 1392 if (Verbose && PrintGCDetails) {
1379 1393 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1380 1394 " gc request");
1381 1395 }
1382 1396 return true;
1383 1397 }
1384 1398
1385 1399 // For debugging purposes, change the type of collection.
1386 1400 // If the rotation is not on the concurrent collection
1387 1401 // type, don't start a concurrent collection.
1388 1402 NOT_PRODUCT(
1389 1403 if (RotateCMSCollectionTypes &&
1390 1404 (_cmsGen->debug_collection_type() !=
1391 1405 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
1392 1406 assert(_cmsGen->debug_collection_type() !=
1393 1407 ConcurrentMarkSweepGeneration::Unknown_collection_type,
1394 1408 "Bad cms collection type");
1395 1409 return false;
1396 1410 }
1397 1411 )
1398 1412
1399 1413 FreelistLocker x(this);
1400 1414 // ------------------------------------------------------------------
1401 1415 // Print out lots of information which affects the initiation of
1402 1416 // a collection.
1403 1417 if (PrintCMSInitiationStatistics && stats().valid()) {
1404 1418 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1405 1419 gclog_or_tty->stamp();
↓ open down ↓ |
731 lines elided |
↑ open up ↑ |
1406 1420 gclog_or_tty->print_cr("");
1407 1421 stats().print_on(gclog_or_tty);
1408 1422 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1409 1423 stats().time_until_cms_gen_full());
1410 1424 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1411 1425 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1412 1426 _cmsGen->contiguous_available());
1413 1427 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1414 1428 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1415 1429 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1416 - gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", initiatingOccupancy());
1430 + gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1431 + gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy());
1417 1432 }
1418 1433 // ------------------------------------------------------------------
1419 1434
1420 1435 // If the estimated time to complete a cms collection (cms_duration())
1421 1436 // is less than the estimated time remaining until the cms generation
1422 1437 // is full, start a collection.
1423 1438 if (!UseCMSInitiatingOccupancyOnly) {
1424 1439 if (stats().valid()) {
1425 1440 if (stats().time_until_cms_start() == 0.0) {
1426 1441 return true;
1427 1442 }
1428 1443 } else {
1429 1444 // We want to conservatively collect somewhat early in order
1430 1445 // to try and "bootstrap" our CMS/promotion statistics;
1431 1446 // this branch will not fire after the first successful CMS
1432 1447 // collection because the stats should then be valid.
1433 1448 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1434 1449 if (Verbose && PrintGCDetails) {
1435 1450 gclog_or_tty->print_cr(
1436 1451 " CMSCollector: collect for bootstrapping statistics:"
1437 1452 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1438 1453 _bootstrap_occupancy);
↓ open down ↓ |
12 lines elided |
↑ open up ↑ |
1439 1454 }
1440 1455 return true;
1441 1456 }
1442 1457 }
1443 1458 }
1444 1459
1445 1460 // Otherwise, we start a collection cycle if either the perm gen or
1446 1461 // old gen want a collection cycle started. Each may use
1447 1462 // an appropriate criterion for making this decision.
1448 1463 // XXX We need to make sure that the gen expansion
1449 - // criterion dovetails well with this.
1450 - if (_cmsGen->shouldConcurrentCollect(initiatingOccupancy())) {
1464 + // criterion dovetails well with this. XXX NEED TO FIX THIS
1465 + if (_cmsGen->should_concurrent_collect()) {
1451 1466 if (Verbose && PrintGCDetails) {
1452 1467 gclog_or_tty->print_cr("CMS old gen initiated");
1453 1468 }
1454 1469 return true;
1455 1470 }
1456 1471
1457 - if (cms_should_unload_classes() &&
1458 - _permGen->shouldConcurrentCollect(initiatingOccupancy())) {
1459 - if (Verbose && PrintGCDetails) {
1460 - gclog_or_tty->print_cr("CMS perm gen initiated");
1472 + // We start a collection if we believe an incremental collection may fail;
1473 + // this is not likely to be productive in practice because it's probably too
1474 + // late anyway.
1475 + GenCollectedHeap* gch = GenCollectedHeap::heap();
1476 + assert(gch->collector_policy()->is_two_generation_policy(),
1477 + "You may want to check the correctness of the following");
1478 + if (gch->incremental_collection_will_fail()) {
1479 + if (PrintGCDetails && Verbose) {
1480 + gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1461 1481 }
1462 1482 return true;
1463 1483 }
1464 1484
1485 + if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) {
1486 + bool res = update_should_unload_classes();
1487 + if (res) {
1488 + if (Verbose && PrintGCDetails) {
1489 + gclog_or_tty->print_cr("CMS perm gen initiated");
1490 + }
1491 + return true;
1492 + }
1493 + }
1465 1494 return false;
1466 1495 }
1467 1496
1468 1497 // Clear _expansion_cause fields of constituent generations
1469 1498 void CMSCollector::clear_expansion_cause() {
1470 1499 _cmsGen->clear_expansion_cause();
1471 1500 _permGen->clear_expansion_cause();
1472 1501 }
1473 1502
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());
1503 +// We should be conservative in starting a collection cycle. To
1504 +// start too eagerly runs the risk of collecting too often in the
1505 +// extreme. To collect too rarely falls back on full collections,
1506 +// which works, even if not optimum in terms of concurrent work.
1507 +// As a work around for too eagerly collecting, use the flag
1508 +// UseCMSInitiatingOccupancyOnly. This also has the advantage of
1509 +// giving the user an easily understandable way of controlling the
1510 +// collections.
1511 +// We want to start a new collection cycle if any of the following
1512 +// conditions hold:
1513 +// . our current occupancy exceeds the configured initiating occupancy
1514 +// for this generation, or
1515 +// . we recently needed to expand this space and have not, since that
1516 +// expansion, done a collection of this generation, or
1517 +// . the underlying space believes that it may be a good idea to initiate
1518 +// a concurrent collection (this may be based on criteria such as the
1519 +// following: the space uses linear allocation and linear allocation is
1520 +// going to fail, or there is believed to be excessive fragmentation in
1521 +// the generation, etc... or ...
1522 +// [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1523 +// the case of the old generation, not the perm generation; see CR 6543076):
1524 +// we may be approaching a point at which allocation requests may fail because
1525 +// we will be out of sufficient free space given allocation rate estimates.]
1526 +bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1495 1527
1496 - if (occupancy() > initiatingOccupancy) {
1528 + assert_lock_strong(freelistLock());
1529 + if (occupancy() > initiating_occupancy()) {
1497 1530 if (PrintGCDetails && Verbose) {
1498 1531 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1499 - short_name(), occupancy(), initiatingOccupancy);
1532 + short_name(), occupancy(), initiating_occupancy());
1500 1533 }
1501 1534 return true;
1502 1535 }
1503 1536 if (UseCMSInitiatingOccupancyOnly) {
1504 1537 return false;
1505 1538 }
1506 1539 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1507 1540 if (PrintGCDetails && Verbose) {
1508 1541 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1509 1542 short_name());
1510 1543 }
1511 1544 return true;
1512 1545 }
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()) {
1546 + if (_cmsSpace->should_concurrent_collect()) {
1517 1547 if (PrintGCDetails && Verbose) {
1518 - gclog_or_tty->print(" %s: collect because incremental collection will fail ",
1548 + gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1519 1549 short_name());
1520 1550 }
1521 1551 return true;
1522 1552 }
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 1553 return false;
1532 1554 }
1533 1555
1534 1556 void ConcurrentMarkSweepGeneration::collect(bool full,
1535 1557 bool clear_all_soft_refs,
1536 1558 size_t size,
1537 1559 bool tlab)
1538 1560 {
1539 1561 collector()->collect(full, clear_all_soft_refs, size, tlab);
1540 1562 }
1541 1563
1542 1564 void CMSCollector::collect(bool full,
1543 1565 bool clear_all_soft_refs,
1544 1566 size_t size,
1545 1567 bool tlab)
1546 1568 {
1547 1569 if (!UseCMSCollectionPassing && _collectorState > Idling) {
1548 1570 // For debugging purposes skip the collection if the state
1549 1571 // is not currently idle
1550 1572 if (TraceCMSState) {
1551 1573 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
1552 1574 Thread::current(), full, _collectorState);
1553 1575 }
1554 1576 return;
1555 1577 }
1556 1578
1557 1579 // The following "if" branch is present for defensive reasons.
1558 1580 // In the current uses of this interface, it can be replaced with:
1559 1581 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1560 1582 // But I am not placing that assert here to allow future
1561 1583 // generality in invoking this interface.
1562 1584 if (GC_locker::is_active()) {
1563 1585 // A consistency test for GC_locker
1564 1586 assert(GC_locker::needs_gc(), "Should have been set already");
1565 1587 // Skip this foreground collection, instead
1566 1588 // expanding the heap if necessary.
1567 1589 // Need the free list locks for the call to free() in compute_new_size()
1568 1590 compute_new_size();
1569 1591 return;
1570 1592 }
1571 1593 acquire_control_and_collect(full, clear_all_soft_refs);
1572 1594 _full_gcs_since_conc_gc++;
1573 1595
1574 1596 }
1575 1597
1576 1598 void CMSCollector::request_full_gc(unsigned int full_gc_count) {
1577 1599 GenCollectedHeap* gch = GenCollectedHeap::heap();
1578 1600 unsigned int gc_count = gch->total_full_collections();
1579 1601 if (gc_count == full_gc_count) {
1580 1602 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1581 1603 _full_gc_requested = true;
1582 1604 CGC_lock->notify(); // nudge CMS thread
1583 1605 }
1584 1606 }
1585 1607
1586 1608
1587 1609 // The foreground and background collectors need to coordinate in order
1588 1610 // to make sure that they do not mutually interfere with CMS collections.
1589 1611 // When a background collection is active,
1590 1612 // the foreground collector may need to take over (preempt) and
1591 1613 // synchronously complete an ongoing collection. Depending on the
1592 1614 // frequency of the background collections and the heap usage
1593 1615 // of the application, this preemption can be seldom or frequent.
1594 1616 // There are only certain
1595 1617 // points in the background collection that the "collection-baton"
1596 1618 // can be passed to the foreground collector.
1597 1619 //
1598 1620 // The foreground collector will wait for the baton before
1599 1621 // starting any part of the collection. The foreground collector
1600 1622 // will only wait at one location.
1601 1623 //
1602 1624 // The background collector will yield the baton before starting a new
1603 1625 // phase of the collection (e.g., before initial marking, marking from roots,
1604 1626 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1605 1627 // of the loop which switches the phases. The background collector does some
1606 1628 // of the phases (initial mark, final re-mark) with the world stopped.
1607 1629 // Because of locking involved in stopping the world,
1608 1630 // the foreground collector should not block waiting for the background
1609 1631 // collector when it is doing a stop-the-world phase. The background
1610 1632 // collector will yield the baton at an additional point just before
1611 1633 // it enters a stop-the-world phase. Once the world is stopped, the
1612 1634 // background collector checks the phase of the collection. If the
1613 1635 // phase has not changed, it proceeds with the collection. If the
1614 1636 // phase has changed, it skips that phase of the collection. See
1615 1637 // the comments on the use of the Heap_lock in collect_in_background().
1616 1638 //
1617 1639 // Variable used in baton passing.
1618 1640 // _foregroundGCIsActive - Set to true by the foreground collector when
1619 1641 // it wants the baton. The foreground clears it when it has finished
1620 1642 // the collection.
1621 1643 // _foregroundGCShouldWait - Set to true by the background collector
1622 1644 // when it is running. The foreground collector waits while
1623 1645 // _foregroundGCShouldWait is true.
1624 1646 // CGC_lock - monitor used to protect access to the above variables
1625 1647 // and to notify the foreground and background collectors.
1626 1648 // _collectorState - current state of the CMS collection.
1627 1649 //
1628 1650 // The foreground collector
1629 1651 // acquires the CGC_lock
1630 1652 // sets _foregroundGCIsActive
1631 1653 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1632 1654 // various locks acquired in preparation for the collection
1633 1655 // are released so as not to block the background collector
1634 1656 // that is in the midst of a collection
1635 1657 // proceeds with the collection
1636 1658 // clears _foregroundGCIsActive
1637 1659 // returns
1638 1660 //
1639 1661 // The background collector in a loop iterating on the phases of the
1640 1662 // collection
1641 1663 // acquires the CGC_lock
1642 1664 // sets _foregroundGCShouldWait
1643 1665 // if _foregroundGCIsActive is set
1644 1666 // clears _foregroundGCShouldWait, notifies _CGC_lock
1645 1667 // waits on _CGC_lock for _foregroundGCIsActive to become false
1646 1668 // and exits the loop.
1647 1669 // otherwise
1648 1670 // proceed with that phase of the collection
1649 1671 // if the phase is a stop-the-world phase,
1650 1672 // yield the baton once more just before enqueueing
1651 1673 // the stop-world CMS operation (executed by the VM thread).
1652 1674 // returns after all phases of the collection are done
1653 1675 //
1654 1676
1655 1677 void CMSCollector::acquire_control_and_collect(bool full,
1656 1678 bool clear_all_soft_refs) {
1657 1679 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1658 1680 assert(!Thread::current()->is_ConcurrentGC_thread(),
1659 1681 "shouldn't try to acquire control from self!");
1660 1682
1661 1683 // Start the protocol for acquiring control of the
1662 1684 // collection from the background collector (aka CMS thread).
1663 1685 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1664 1686 "VM thread should have CMS token");
1665 1687 // Remember the possibly interrupted state of an ongoing
1666 1688 // concurrent collection
1667 1689 CollectorState first_state = _collectorState;
1668 1690
1669 1691 // Signal to a possibly ongoing concurrent collection that
1670 1692 // we want to do a foreground collection.
1671 1693 _foregroundGCIsActive = true;
1672 1694
1673 1695 // Disable incremental mode during a foreground collection.
1674 1696 ICMSDisabler icms_disabler;
1675 1697
1676 1698 // release locks and wait for a notify from the background collector
1677 1699 // releasing the locks in only necessary for phases which
1678 1700 // do yields to improve the granularity of the collection.
1679 1701 assert_lock_strong(bitMapLock());
1680 1702 // We need to lock the Free list lock for the space that we are
1681 1703 // currently collecting.
1682 1704 assert(haveFreelistLocks(), "Must be holding free list locks");
1683 1705 bitMapLock()->unlock();
1684 1706 releaseFreelistLocks();
1685 1707 {
1686 1708 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1687 1709 if (_foregroundGCShouldWait) {
1688 1710 // We are going to be waiting for action for the CMS thread;
1689 1711 // it had better not be gone (for instance at shutdown)!
1690 1712 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1691 1713 "CMS thread must be running");
1692 1714 // Wait here until the background collector gives us the go-ahead
1693 1715 ConcurrentMarkSweepThread::clear_CMS_flag(
1694 1716 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1695 1717 // Get a possibly blocked CMS thread going:
1696 1718 // Note that we set _foregroundGCIsActive true above,
1697 1719 // without protection of the CGC_lock.
1698 1720 CGC_lock->notify();
1699 1721 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1700 1722 "Possible deadlock");
1701 1723 while (_foregroundGCShouldWait) {
1702 1724 // wait for notification
1703 1725 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1704 1726 // Possibility of delay/starvation here, since CMS token does
1705 1727 // not know to give priority to VM thread? Actually, i think
1706 1728 // there wouldn't be any delay/starvation, but the proof of
1707 1729 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1708 1730 }
1709 1731 ConcurrentMarkSweepThread::set_CMS_flag(
1710 1732 ConcurrentMarkSweepThread::CMS_vm_has_token);
1711 1733 }
1712 1734 }
1713 1735 // The CMS_token is already held. Get back the other locks.
1714 1736 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1715 1737 "VM thread should have CMS token");
1716 1738 getFreelistLocks();
1717 1739 bitMapLock()->lock_without_safepoint_check();
1718 1740 if (TraceCMSState) {
1719 1741 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1720 1742 INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1721 1743 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1722 1744 }
1723 1745
1724 1746 // Check if we need to do a compaction, or if not, whether
1725 1747 // we need to start the mark-sweep from scratch.
1726 1748 bool should_compact = false;
1727 1749 bool should_start_over = false;
1728 1750 decide_foreground_collection_type(clear_all_soft_refs,
1729 1751 &should_compact, &should_start_over);
1730 1752
1731 1753 NOT_PRODUCT(
1732 1754 if (RotateCMSCollectionTypes) {
1733 1755 if (_cmsGen->debug_collection_type() ==
1734 1756 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1735 1757 should_compact = true;
1736 1758 } else if (_cmsGen->debug_collection_type() ==
1737 1759 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1738 1760 should_compact = false;
1739 1761 }
1740 1762 }
1741 1763 )
1742 1764
1743 1765 if (PrintGCDetails && first_state > Idling) {
1744 1766 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1745 1767 if (GCCause::is_user_requested_gc(cause) ||
1746 1768 GCCause::is_serviceability_requested_gc(cause)) {
1747 1769 gclog_or_tty->print(" (concurrent mode interrupted)");
1748 1770 } else {
1749 1771 gclog_or_tty->print(" (concurrent mode failure)");
1750 1772 }
1751 1773 }
1752 1774
1753 1775 if (should_compact) {
1754 1776 // If the collection is being acquired from the background
1755 1777 // collector, there may be references on the discovered
1756 1778 // references lists that have NULL referents (being those
1757 1779 // that were concurrently cleared by a mutator) or
1758 1780 // that are no longer active (having been enqueued concurrently
1759 1781 // by the mutator).
1760 1782 // Scrub the list of those references because Mark-Sweep-Compact
1761 1783 // code assumes referents are not NULL and that all discovered
1762 1784 // Reference objects are active.
1763 1785 ref_processor()->clean_up_discovered_references();
1764 1786
1765 1787 do_compaction_work(clear_all_soft_refs);
1766 1788
1767 1789 // Has the GC time limit been exceeded?
1768 1790 check_gc_time_limit();
1769 1791
1770 1792 } else {
1771 1793 do_mark_sweep_work(clear_all_soft_refs, first_state,
1772 1794 should_start_over);
1773 1795 }
1774 1796 // Reset the expansion cause, now that we just completed
1775 1797 // a collection cycle.
1776 1798 clear_expansion_cause();
1777 1799 _foregroundGCIsActive = false;
1778 1800 return;
1779 1801 }
1780 1802
1781 1803 void CMSCollector::check_gc_time_limit() {
1782 1804
1783 1805 // Ignore explicit GC's. Exiting here does not set the flag and
1784 1806 // does not reset the count. Updating of the averages for system
1785 1807 // GC's is still controlled by UseAdaptiveSizePolicyWithSystemGC.
1786 1808 GCCause::Cause gc_cause = GenCollectedHeap::heap()->gc_cause();
1787 1809 if (GCCause::is_user_requested_gc(gc_cause) ||
1788 1810 GCCause::is_serviceability_requested_gc(gc_cause)) {
1789 1811 return;
1790 1812 }
1791 1813
1792 1814 // Calculate the fraction of the CMS generation was freed during
1793 1815 // the last collection.
1794 1816 // Only consider the STW compacting cost for now.
1795 1817 //
1796 1818 // Note that the gc time limit test only works for the collections
1797 1819 // of the young gen + tenured gen and not for collections of the
1798 1820 // permanent gen. That is because the calculation of the space
1799 1821 // freed by the collection is the free space in the young gen +
1800 1822 // tenured gen.
1801 1823
1802 1824 double fraction_free =
1803 1825 ((double)_cmsGen->free())/((double)_cmsGen->max_capacity());
1804 1826 if ((100.0 * size_policy()->compacting_gc_cost()) >
1805 1827 ((double) GCTimeLimit) &&
1806 1828 ((fraction_free * 100) < GCHeapFreeLimit)) {
1807 1829 size_policy()->inc_gc_time_limit_count();
1808 1830 if (UseGCOverheadLimit &&
1809 1831 (size_policy()->gc_time_limit_count() >
1810 1832 AdaptiveSizePolicyGCTimeLimitThreshold)) {
1811 1833 size_policy()->set_gc_time_limit_exceeded(true);
1812 1834 // Avoid consecutive OOM due to the gc time limit by resetting
1813 1835 // the counter.
1814 1836 size_policy()->reset_gc_time_limit_count();
1815 1837 if (PrintGCDetails) {
1816 1838 gclog_or_tty->print_cr(" GC is exceeding overhead limit "
1817 1839 "of %d%%", GCTimeLimit);
1818 1840 }
1819 1841 } else {
1820 1842 if (PrintGCDetails) {
1821 1843 gclog_or_tty->print_cr(" GC would exceed overhead limit "
1822 1844 "of %d%%", GCTimeLimit);
1823 1845 }
1824 1846 }
1825 1847 } else {
1826 1848 size_policy()->reset_gc_time_limit_count();
1827 1849 }
1828 1850 }
1829 1851
1830 1852 // Resize the perm generation and the tenured generation
1831 1853 // after obtaining the free list locks for the
1832 1854 // two generations.
1833 1855 void CMSCollector::compute_new_size() {
1834 1856 assert_locked_or_safepoint(Heap_lock);
1835 1857 FreelistLocker z(this);
1836 1858 _permGen->compute_new_size();
1837 1859 _cmsGen->compute_new_size();
1838 1860 }
1839 1861
1840 1862 // A work method used by foreground collection to determine
1841 1863 // what type of collection (compacting or not, continuing or fresh)
1842 1864 // it should do.
1843 1865 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1844 1866 // and CMSCompactWhenClearAllSoftRefs the default in the future
1845 1867 // and do away with the flags after a suitable period.
1846 1868 void CMSCollector::decide_foreground_collection_type(
1847 1869 bool clear_all_soft_refs, bool* should_compact,
1848 1870 bool* should_start_over) {
1849 1871 // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1850 1872 // flag is set, and we have either requested a System.gc() or
1851 1873 // the number of full gc's since the last concurrent cycle
1852 1874 // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1853 1875 // or if an incremental collection has failed
1854 1876 GenCollectedHeap* gch = GenCollectedHeap::heap();
1855 1877 assert(gch->collector_policy()->is_two_generation_policy(),
1856 1878 "You may want to check the correctness of the following");
1857 1879 // Inform cms gen if this was due to partial collection failing.
1858 1880 // The CMS gen may use this fact to determine its expansion policy.
1859 1881 if (gch->incremental_collection_will_fail()) {
1860 1882 assert(!_cmsGen->incremental_collection_failed(),
1861 1883 "Should have been noticed, reacted to and cleared");
1862 1884 _cmsGen->set_incremental_collection_failed();
1863 1885 }
1864 1886 *should_compact =
1865 1887 UseCMSCompactAtFullCollection &&
1866 1888 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1867 1889 GCCause::is_user_requested_gc(gch->gc_cause()) ||
1868 1890 gch->incremental_collection_will_fail());
1869 1891 *should_start_over = false;
1870 1892 if (clear_all_soft_refs && !*should_compact) {
1871 1893 // We are about to do a last ditch collection attempt
1872 1894 // so it would normally make sense to do a compaction
1873 1895 // to reclaim as much space as possible.
1874 1896 if (CMSCompactWhenClearAllSoftRefs) {
1875 1897 // Default: The rationale is that in this case either
1876 1898 // we are past the final marking phase, in which case
1877 1899 // we'd have to start over, or so little has been done
1878 1900 // that there's little point in saving that work. Compaction
1879 1901 // appears to be the sensible choice in either case.
1880 1902 *should_compact = true;
1881 1903 } else {
1882 1904 // We have been asked to clear all soft refs, but not to
1883 1905 // compact. Make sure that we aren't past the final checkpoint
1884 1906 // phase, for that is where we process soft refs. If we are already
1885 1907 // past that phase, we'll need to redo the refs discovery phase and
1886 1908 // if necessary clear soft refs that weren't previously
1887 1909 // cleared. We do so by remembering the phase in which
1888 1910 // we came in, and if we are past the refs processing
1889 1911 // phase, we'll choose to just redo the mark-sweep
1890 1912 // collection from scratch.
1891 1913 if (_collectorState > FinalMarking) {
1892 1914 // We are past the refs processing phase;
1893 1915 // start over and do a fresh synchronous CMS cycle
1894 1916 _collectorState = Resetting; // skip to reset to start new cycle
1895 1917 reset(false /* == !asynch */);
1896 1918 *should_start_over = true;
1897 1919 } // else we can continue a possibly ongoing current cycle
1898 1920 }
1899 1921 }
1900 1922 }
1901 1923
1902 1924 // A work method used by the foreground collector to do
1903 1925 // a mark-sweep-compact.
1904 1926 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1905 1927 GenCollectedHeap* gch = GenCollectedHeap::heap();
1906 1928 TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty);
1907 1929 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
1908 1930 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
1909 1931 "collections passed to foreground collector", _full_gcs_since_conc_gc);
1910 1932 }
1911 1933
1912 1934 // Sample collection interval time and reset for collection pause.
1913 1935 if (UseAdaptiveSizePolicy) {
1914 1936 size_policy()->msc_collection_begin();
1915 1937 }
1916 1938
1917 1939 // Temporarily widen the span of the weak reference processing to
1918 1940 // the entire heap.
1919 1941 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1920 1942 ReferenceProcessorSpanMutator x(ref_processor(), new_span);
1921 1943
1922 1944 // Temporarily, clear the "is_alive_non_header" field of the
1923 1945 // reference processor.
1924 1946 ReferenceProcessorIsAliveMutator y(ref_processor(), NULL);
1925 1947
1926 1948 // Temporarily make reference _processing_ single threaded (non-MT).
1927 1949 ReferenceProcessorMTProcMutator z(ref_processor(), false);
1928 1950
1929 1951 // Temporarily make refs discovery atomic
1930 1952 ReferenceProcessorAtomicMutator w(ref_processor(), true);
1931 1953
1932 1954 ref_processor()->set_enqueuing_is_done(false);
1933 1955 ref_processor()->enable_discovery();
1934 1956 // If an asynchronous collection finishes, the _modUnionTable is
1935 1957 // all clear. If we are assuming the collection from an asynchronous
1936 1958 // collection, clear the _modUnionTable.
1937 1959 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1938 1960 "_modUnionTable should be clear if the baton was not passed");
1939 1961 _modUnionTable.clear_all();
1940 1962
1941 1963 // We must adjust the allocation statistics being maintained
1942 1964 // in the free list space. We do so by reading and clearing
1943 1965 // the sweep timer and updating the block flux rate estimates below.
1944 1966 assert(_sweep_timer.is_active(), "We should never see the timer inactive");
1945 1967 _sweep_timer.stop();
1946 1968 // Note that we do not use this sample to update the _sweep_estimate.
1947 1969 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()),
1948 1970 _sweep_estimate.padded_average());
1949 1971
1950 1972 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
1951 1973 ref_processor(), clear_all_soft_refs);
1952 1974 #ifdef ASSERT
1953 1975 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1954 1976 size_t free_size = cms_space->free();
1955 1977 assert(free_size ==
1956 1978 pointer_delta(cms_space->end(), cms_space->compaction_top())
1957 1979 * HeapWordSize,
1958 1980 "All the free space should be compacted into one chunk at top");
1959 1981 assert(cms_space->dictionary()->totalChunkSize(
1960 1982 debug_only(cms_space->freelistLock())) == 0 ||
1961 1983 cms_space->totalSizeInIndexedFreeLists() == 0,
1962 1984 "All the free space should be in a single chunk");
↓ open down ↓ |
422 lines elided |
↑ open up ↑ |
1963 1985 size_t num = cms_space->totalCount();
1964 1986 assert((free_size == 0 && num == 0) ||
1965 1987 (free_size > 0 && (num == 1 || num == 2)),
1966 1988 "There should be at most 2 free chunks after compaction");
1967 1989 #endif // ASSERT
1968 1990 _collectorState = Resetting;
1969 1991 assert(_restart_addr == NULL,
1970 1992 "Should have been NULL'd before baton was passed");
1971 1993 reset(false /* == !asynch */);
1972 1994 _cmsGen->reset_after_compaction();
1995 + _concurrent_cycles_since_last_unload = 0;
1973 1996
1974 - if (verifying() && !cms_should_unload_classes()) {
1997 + if (verifying() && !should_unload_classes()) {
1975 1998 perm_gen_verify_bit_map()->clear_all();
1976 1999 }
1977 2000
1978 2001 // Clear any data recorded in the PLAB chunk arrays.
1979 2002 if (_survivor_plab_array != NULL) {
1980 2003 reset_survivor_plab_arrays();
1981 2004 }
1982 2005
1983 2006 // Adjust the per-size allocation stats for the next epoch.
1984 - _cmsGen->cmsSpace()->endSweepFLCensus(sweepCount() /* fake */);
2007 + _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1985 2008 // Restart the "sweep timer" for next epoch.
1986 2009 _sweep_timer.reset();
1987 2010 _sweep_timer.start();
1988 2011
1989 2012 // Sample collection pause time and reset for collection interval.
1990 2013 if (UseAdaptiveSizePolicy) {
1991 2014 size_policy()->msc_collection_end(gch->gc_cause());
1992 2015 }
1993 2016
1994 2017 // For a mark-sweep-compact, compute_new_size() will be called
1995 2018 // in the heap's do_collection() method.
1996 2019 }
1997 2020
1998 2021 // A work method used by the foreground collector to do
1999 2022 // a mark-sweep, after taking over from a possibly on-going
2000 2023 // concurrent mark-sweep collection.
2001 2024 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2002 2025 CollectorState first_state, bool should_start_over) {
2003 2026 if (PrintGC && Verbose) {
2004 2027 gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2005 2028 "collector with count %d",
2006 2029 _full_gcs_since_conc_gc);
2007 2030 }
2008 2031 switch (_collectorState) {
2009 2032 case Idling:
2010 2033 if (first_state == Idling || should_start_over) {
2011 2034 // The background GC was not active, or should
2012 2035 // restarted from scratch; start the cycle.
2013 2036 _collectorState = InitialMarking;
2014 2037 }
2015 2038 // If first_state was not Idling, then a background GC
2016 2039 // was in progress and has now finished. No need to do it
2017 2040 // again. Leave the state as Idling.
2018 2041 break;
2019 2042 case Precleaning:
2020 2043 // In the foreground case don't do the precleaning since
2021 2044 // it is not done concurrently and there is extra work
2022 2045 // required.
2023 2046 _collectorState = FinalMarking;
2024 2047 }
2025 2048 if (PrintGCDetails &&
2026 2049 (_collectorState > Idling ||
2027 2050 !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) {
2028 2051 gclog_or_tty->print(" (concurrent mode failure)");
2029 2052 }
2030 2053 collect_in_foreground(clear_all_soft_refs);
2031 2054
2032 2055 // For a mark-sweep, compute_new_size() will be called
2033 2056 // in the heap's do_collection() method.
2034 2057 }
2035 2058
2036 2059
2037 2060 void CMSCollector::getFreelistLocks() const {
2038 2061 // Get locks for all free lists in all generations that this
2039 2062 // collector is responsible for
2040 2063 _cmsGen->freelistLock()->lock_without_safepoint_check();
2041 2064 _permGen->freelistLock()->lock_without_safepoint_check();
2042 2065 }
2043 2066
2044 2067 void CMSCollector::releaseFreelistLocks() const {
2045 2068 // Release locks for all free lists in all generations that this
2046 2069 // collector is responsible for
2047 2070 _cmsGen->freelistLock()->unlock();
2048 2071 _permGen->freelistLock()->unlock();
2049 2072 }
2050 2073
2051 2074 bool CMSCollector::haveFreelistLocks() const {
2052 2075 // Check locks for all free lists in all generations that this
2053 2076 // collector is responsible for
2054 2077 assert_lock_strong(_cmsGen->freelistLock());
2055 2078 assert_lock_strong(_permGen->freelistLock());
2056 2079 PRODUCT_ONLY(ShouldNotReachHere());
2057 2080 return true;
2058 2081 }
2059 2082
2060 2083 // A utility class that is used by the CMS collector to
2061 2084 // temporarily "release" the foreground collector from its
2062 2085 // usual obligation to wait for the background collector to
2063 2086 // complete an ongoing phase before proceeding.
2064 2087 class ReleaseForegroundGC: public StackObj {
2065 2088 private:
2066 2089 CMSCollector* _c;
2067 2090 public:
2068 2091 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2069 2092 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2070 2093 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2071 2094 // allow a potentially blocked foreground collector to proceed
2072 2095 _c->_foregroundGCShouldWait = false;
2073 2096 if (_c->_foregroundGCIsActive) {
2074 2097 CGC_lock->notify();
2075 2098 }
2076 2099 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2077 2100 "Possible deadlock");
2078 2101 }
2079 2102
2080 2103 ~ReleaseForegroundGC() {
2081 2104 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2082 2105 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2083 2106 _c->_foregroundGCShouldWait = true;
2084 2107 }
2085 2108 };
2086 2109
2087 2110 // There are separate collect_in_background and collect_in_foreground because of
2088 2111 // the different locking requirements of the background collector and the
2089 2112 // foreground collector. There was originally an attempt to share
2090 2113 // one "collect" method between the background collector and the foreground
↓ open down ↓ |
96 lines elided |
↑ open up ↑ |
2091 2114 // collector but the if-then-else required made it cleaner to have
2092 2115 // separate methods.
2093 2116 void CMSCollector::collect_in_background(bool clear_all_soft_refs) {
2094 2117 assert(Thread::current()->is_ConcurrentGC_thread(),
2095 2118 "A CMS asynchronous collection is only allowed on a CMS thread.");
2096 2119
2097 2120 GenCollectedHeap* gch = GenCollectedHeap::heap();
2098 2121 {
2099 2122 bool safepoint_check = Mutex::_no_safepoint_check_flag;
2100 2123 MutexLockerEx hl(Heap_lock, safepoint_check);
2124 + FreelistLocker fll(this);
2101 2125 MutexLockerEx x(CGC_lock, safepoint_check);
2102 2126 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2103 2127 // The foreground collector is active or we're
2104 2128 // not using asynchronous collections. Skip this
2105 2129 // background collection.
2106 2130 assert(!_foregroundGCShouldWait, "Should be clear");
2107 2131 return;
2108 2132 } else {
2109 2133 assert(_collectorState == Idling, "Should be idling before start.");
2110 2134 _collectorState = InitialMarking;
2111 2135 // Reset the expansion cause, now that we are about to begin
2112 2136 // a new cycle.
2113 2137 clear_expansion_cause();
2114 2138 }
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;
2139 + // Decide if we want to enable class unloading as part of the
2140 + // ensuing concurrent GC cycle.
2141 + update_should_unload_classes();
2122 2142 _full_gc_requested = false; // acks all outstanding full gc requests
2123 2143 // Signal that we are about to start a collection
2124 2144 gch->increment_total_full_collections(); // ... starting a collection cycle
2125 2145 _collection_count_start = gch->total_full_collections();
2126 2146 }
2127 2147
2128 2148 // Used for PrintGC
2129 2149 size_t prev_used;
2130 2150 if (PrintGC && Verbose) {
2131 2151 prev_used = _cmsGen->used(); // XXXPERM
2132 2152 }
2133 2153
2134 2154 // The change of the collection state is normally done at this level;
2135 2155 // the exceptions are phases that are executed while the world is
2136 2156 // stopped. For those phases the change of state is done while the
2137 2157 // world is stopped. For baton passing purposes this allows the
2138 2158 // background collector to finish the phase and change state atomically.
2139 2159 // The foreground collector cannot wait on a phase that is done
2140 2160 // while the world is stopped because the foreground collector already
2141 2161 // has the world stopped and would deadlock.
2142 2162 while (_collectorState != Idling) {
2143 2163 if (TraceCMSState) {
2144 2164 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2145 2165 Thread::current(), _collectorState);
2146 2166 }
2147 2167 // The foreground collector
2148 2168 // holds the Heap_lock throughout its collection.
2149 2169 // holds the CMS token (but not the lock)
2150 2170 // except while it is waiting for the background collector to yield.
2151 2171 //
2152 2172 // The foreground collector should be blocked (not for long)
2153 2173 // if the background collector is about to start a phase
2154 2174 // executed with world stopped. If the background
2155 2175 // collector has already started such a phase, the
2156 2176 // foreground collector is blocked waiting for the
2157 2177 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
2158 2178 // are executed in the VM thread.
2159 2179 //
2160 2180 // The locking order is
2161 2181 // PendingListLock (PLL) -- if applicable (FinalMarking)
2162 2182 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
2163 2183 // CMS token (claimed in
2164 2184 // stop_world_and_do() -->
2165 2185 // safepoint_synchronize() -->
2166 2186 // CMSThread::synchronize())
2167 2187
2168 2188 {
2169 2189 // Check if the FG collector wants us to yield.
2170 2190 CMSTokenSync x(true); // is cms thread
2171 2191 if (waitForForegroundGC()) {
2172 2192 // We yielded to a foreground GC, nothing more to be
2173 2193 // done this round.
2174 2194 assert(_foregroundGCShouldWait == false, "We set it to false in "
2175 2195 "waitForForegroundGC()");
2176 2196 if (TraceCMSState) {
2177 2197 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2178 2198 " exiting collection CMS state %d",
2179 2199 Thread::current(), _collectorState);
2180 2200 }
2181 2201 return;
2182 2202 } else {
2183 2203 // The background collector can run but check to see if the
2184 2204 // foreground collector has done a collection while the
2185 2205 // background collector was waiting to get the CGC_lock
2186 2206 // above. If yes, break so that _foregroundGCShouldWait
2187 2207 // is cleared before returning.
2188 2208 if (_collectorState == Idling) {
2189 2209 break;
2190 2210 }
2191 2211 }
2192 2212 }
2193 2213
2194 2214 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2195 2215 "should be waiting");
2196 2216
2197 2217 switch (_collectorState) {
2198 2218 case InitialMarking:
2199 2219 {
2200 2220 ReleaseForegroundGC x(this);
2201 2221 stats().record_cms_begin();
2202 2222
2203 2223 VM_CMS_Initial_Mark initial_mark_op(this);
2204 2224 VMThread::execute(&initial_mark_op);
2205 2225 }
2206 2226 // The collector state may be any legal state at this point
2207 2227 // since the background collector may have yielded to the
2208 2228 // foreground collector.
2209 2229 break;
2210 2230 case Marking:
2211 2231 // initial marking in checkpointRootsInitialWork has been completed
2212 2232 if (markFromRoots(true)) { // we were successful
2213 2233 assert(_collectorState == Precleaning, "Collector state should "
2214 2234 "have changed");
2215 2235 } else {
2216 2236 assert(_foregroundGCIsActive, "Internal state inconsistency");
2217 2237 }
2218 2238 break;
2219 2239 case Precleaning:
2220 2240 if (UseAdaptiveSizePolicy) {
2221 2241 size_policy()->concurrent_precleaning_begin();
2222 2242 }
2223 2243 // marking from roots in markFromRoots has been completed
2224 2244 preclean();
2225 2245 if (UseAdaptiveSizePolicy) {
2226 2246 size_policy()->concurrent_precleaning_end();
2227 2247 }
2228 2248 assert(_collectorState == AbortablePreclean ||
2229 2249 _collectorState == FinalMarking,
2230 2250 "Collector state should have changed");
2231 2251 break;
2232 2252 case AbortablePreclean:
2233 2253 if (UseAdaptiveSizePolicy) {
2234 2254 size_policy()->concurrent_phases_resume();
2235 2255 }
2236 2256 abortable_preclean();
2237 2257 if (UseAdaptiveSizePolicy) {
2238 2258 size_policy()->concurrent_precleaning_end();
2239 2259 }
2240 2260 assert(_collectorState == FinalMarking, "Collector state should "
2241 2261 "have changed");
2242 2262 break;
2243 2263 case FinalMarking:
2244 2264 {
2245 2265 ReleaseForegroundGC x(this);
2246 2266
2247 2267 VM_CMS_Final_Remark final_remark_op(this);
2248 2268 VMThread::execute(&final_remark_op);
2249 2269 }
2250 2270 assert(_foregroundGCShouldWait, "block post-condition");
2251 2271 break;
2252 2272 case Sweeping:
2253 2273 if (UseAdaptiveSizePolicy) {
2254 2274 size_policy()->concurrent_sweeping_begin();
2255 2275 }
2256 2276 // final marking in checkpointRootsFinal has been completed
2257 2277 sweep(true);
2258 2278 assert(_collectorState == Resizing, "Collector state change "
2259 2279 "to Resizing must be done under the free_list_lock");
2260 2280 _full_gcs_since_conc_gc = 0;
2261 2281
2262 2282 // Stop the timers for adaptive size policy for the concurrent phases
2263 2283 if (UseAdaptiveSizePolicy) {
2264 2284 size_policy()->concurrent_sweeping_end();
2265 2285 size_policy()->concurrent_phases_end(gch->gc_cause(),
2266 2286 gch->prev_gen(_cmsGen)->capacity(),
2267 2287 _cmsGen->free());
2268 2288 }
2269 2289
2270 2290 case Resizing: {
2271 2291 // Sweeping has been completed...
2272 2292 // At this point the background collection has completed.
2273 2293 // Don't move the call to compute_new_size() down
2274 2294 // into code that might be executed if the background
2275 2295 // collection was preempted.
2276 2296 {
2277 2297 ReleaseForegroundGC x(this); // unblock FG collection
2278 2298 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
2279 2299 CMSTokenSync z(true); // not strictly needed.
2280 2300 if (_collectorState == Resizing) {
2281 2301 compute_new_size();
2282 2302 _collectorState = Resetting;
2283 2303 } else {
2284 2304 assert(_collectorState == Idling, "The state should only change"
2285 2305 " because the foreground collector has finished the collection");
2286 2306 }
2287 2307 }
2288 2308 break;
2289 2309 }
2290 2310 case Resetting:
2291 2311 // CMS heap resizing has been completed
2292 2312 reset(true);
2293 2313 assert(_collectorState == Idling, "Collector state should "
2294 2314 "have changed");
2295 2315 stats().record_cms_end();
2296 2316 // Don't move the concurrent_phases_end() and compute_new_size()
2297 2317 // calls to here because a preempted background collection
2298 2318 // has it's state set to "Resetting".
2299 2319 break;
2300 2320 case Idling:
2301 2321 default:
2302 2322 ShouldNotReachHere();
2303 2323 break;
2304 2324 }
2305 2325 if (TraceCMSState) {
2306 2326 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2307 2327 Thread::current(), _collectorState);
2308 2328 }
2309 2329 assert(_foregroundGCShouldWait, "block post-condition");
2310 2330 }
2311 2331
2312 2332 // Should this be in gc_epilogue?
2313 2333 collector_policy()->counters()->update_counters();
2314 2334
2315 2335 {
2316 2336 // Clear _foregroundGCShouldWait and, in the event that the
2317 2337 // foreground collector is waiting, notify it, before
2318 2338 // returning.
2319 2339 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2320 2340 _foregroundGCShouldWait = false;
2321 2341 if (_foregroundGCIsActive) {
2322 2342 CGC_lock->notify();
2323 2343 }
2324 2344 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2325 2345 "Possible deadlock");
2326 2346 }
2327 2347 if (TraceCMSState) {
2328 2348 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2329 2349 " exiting collection CMS state %d",
2330 2350 Thread::current(), _collectorState);
2331 2351 }
2332 2352 if (PrintGC && Verbose) {
2333 2353 _cmsGen->print_heap_change(prev_used);
2334 2354 }
2335 2355 }
2336 2356
2337 2357 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) {
2338 2358 assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2339 2359 "Foreground collector should be waiting, not executing");
2340 2360 assert(Thread::current()->is_VM_thread(), "A foreground collection"
2341 2361 "may only be done by the VM Thread with the world stopped");
2342 2362 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2343 2363 "VM thread should have CMS token");
2344 2364
2345 2365 NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2346 2366 true, gclog_or_tty);)
2347 2367 if (UseAdaptiveSizePolicy) {
2348 2368 size_policy()->ms_collection_begin();
2349 2369 }
2350 2370 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2351 2371
2352 2372 HandleMark hm; // Discard invalid handles created during verification
2353 2373
2354 2374 if (VerifyBeforeGC &&
2355 2375 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2356 2376 Universe::verify(true);
2357 2377 }
2358 2378
2359 2379 bool init_mark_was_synchronous = false; // until proven otherwise
2360 2380 while (_collectorState != Idling) {
2361 2381 if (TraceCMSState) {
2362 2382 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2363 2383 Thread::current(), _collectorState);
2364 2384 }
2365 2385 switch (_collectorState) {
2366 2386 case InitialMarking:
2367 2387 init_mark_was_synchronous = true; // fact to be exploited in re-mark
2368 2388 checkpointRootsInitial(false);
2369 2389 assert(_collectorState == Marking, "Collector state should have changed"
2370 2390 " within checkpointRootsInitial()");
2371 2391 break;
2372 2392 case Marking:
2373 2393 // initial marking in checkpointRootsInitialWork has been completed
2374 2394 if (VerifyDuringGC &&
2375 2395 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2376 2396 gclog_or_tty->print("Verify before initial mark: ");
2377 2397 Universe::verify(true);
2378 2398 }
2379 2399 {
2380 2400 bool res = markFromRoots(false);
2381 2401 assert(res && _collectorState == FinalMarking, "Collector state should "
2382 2402 "have changed");
2383 2403 break;
2384 2404 }
2385 2405 case FinalMarking:
2386 2406 if (VerifyDuringGC &&
2387 2407 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2388 2408 gclog_or_tty->print("Verify before re-mark: ");
2389 2409 Universe::verify(true);
2390 2410 }
2391 2411 checkpointRootsFinal(false, clear_all_soft_refs,
2392 2412 init_mark_was_synchronous);
2393 2413 assert(_collectorState == Sweeping, "Collector state should not "
2394 2414 "have changed within checkpointRootsFinal()");
2395 2415 break;
2396 2416 case Sweeping:
2397 2417 // final marking in checkpointRootsFinal has been completed
2398 2418 if (VerifyDuringGC &&
2399 2419 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2400 2420 gclog_or_tty->print("Verify before sweep: ");
2401 2421 Universe::verify(true);
2402 2422 }
2403 2423 sweep(false);
2404 2424 assert(_collectorState == Resizing, "Incorrect state");
2405 2425 break;
2406 2426 case Resizing: {
2407 2427 // Sweeping has been completed; the actual resize in this case
2408 2428 // is done separately; nothing to be done in this state.
2409 2429 _collectorState = Resetting;
2410 2430 break;
2411 2431 }
2412 2432 case Resetting:
2413 2433 // The heap has been resized.
2414 2434 if (VerifyDuringGC &&
2415 2435 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2416 2436 gclog_or_tty->print("Verify before reset: ");
2417 2437 Universe::verify(true);
2418 2438 }
2419 2439 reset(false);
2420 2440 assert(_collectorState == Idling, "Collector state should "
2421 2441 "have changed");
2422 2442 break;
2423 2443 case Precleaning:
2424 2444 case AbortablePreclean:
2425 2445 // Elide the preclean phase
2426 2446 _collectorState = FinalMarking;
2427 2447 break;
2428 2448 default:
2429 2449 ShouldNotReachHere();
2430 2450 }
2431 2451 if (TraceCMSState) {
2432 2452 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2433 2453 Thread::current(), _collectorState);
2434 2454 }
2435 2455 }
2436 2456
2437 2457 if (UseAdaptiveSizePolicy) {
2438 2458 GenCollectedHeap* gch = GenCollectedHeap::heap();
2439 2459 size_policy()->ms_collection_end(gch->gc_cause());
2440 2460 }
2441 2461
2442 2462 if (VerifyAfterGC &&
2443 2463 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2444 2464 Universe::verify(true);
2445 2465 }
2446 2466 if (TraceCMSState) {
2447 2467 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2448 2468 " exiting collection CMS state %d",
2449 2469 Thread::current(), _collectorState);
2450 2470 }
2451 2471 }
2452 2472
2453 2473 bool CMSCollector::waitForForegroundGC() {
2454 2474 bool res = false;
2455 2475 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2456 2476 "CMS thread should have CMS token");
2457 2477 // Block the foreground collector until the
2458 2478 // background collectors decides whether to
2459 2479 // yield.
2460 2480 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2461 2481 _foregroundGCShouldWait = true;
2462 2482 if (_foregroundGCIsActive) {
2463 2483 // The background collector yields to the
2464 2484 // foreground collector and returns a value
2465 2485 // indicating that it has yielded. The foreground
2466 2486 // collector can proceed.
2467 2487 res = true;
2468 2488 _foregroundGCShouldWait = false;
2469 2489 ConcurrentMarkSweepThread::clear_CMS_flag(
2470 2490 ConcurrentMarkSweepThread::CMS_cms_has_token);
2471 2491 ConcurrentMarkSweepThread::set_CMS_flag(
2472 2492 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2473 2493 // Get a possibly blocked foreground thread going
2474 2494 CGC_lock->notify();
2475 2495 if (TraceCMSState) {
2476 2496 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2477 2497 Thread::current(), _collectorState);
2478 2498 }
2479 2499 while (_foregroundGCIsActive) {
2480 2500 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2481 2501 }
2482 2502 ConcurrentMarkSweepThread::set_CMS_flag(
2483 2503 ConcurrentMarkSweepThread::CMS_cms_has_token);
2484 2504 ConcurrentMarkSweepThread::clear_CMS_flag(
2485 2505 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2486 2506 }
2487 2507 if (TraceCMSState) {
2488 2508 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2489 2509 Thread::current(), _collectorState);
2490 2510 }
2491 2511 return res;
2492 2512 }
2493 2513
2494 2514 // Because of the need to lock the free lists and other structures in
2495 2515 // the collector, common to all the generations that the collector is
2496 2516 // collecting, we need the gc_prologues of individual CMS generations
2497 2517 // delegate to their collector. It may have been simpler had the
2498 2518 // current infrastructure allowed one to call a prologue on a
2499 2519 // collector. In the absence of that we have the generation's
2500 2520 // prologue delegate to the collector, which delegates back
2501 2521 // some "local" work to a worker method in the individual generations
2502 2522 // that it's responsible for collecting, while itself doing any
2503 2523 // work common to all generations it's responsible for. A similar
2504 2524 // comment applies to the gc_epilogue()'s.
2505 2525 // The role of the varaible _between_prologue_and_epilogue is to
2506 2526 // enforce the invocation protocol.
2507 2527 void CMSCollector::gc_prologue(bool full) {
2508 2528 // Call gc_prologue_work() for each CMSGen and PermGen that
2509 2529 // we are responsible for.
2510 2530
2511 2531 // The following locking discipline assumes that we are only called
2512 2532 // when the world is stopped.
2513 2533 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2514 2534
2515 2535 // The CMSCollector prologue must call the gc_prologues for the
2516 2536 // "generations" (including PermGen if any) that it's responsible
2517 2537 // for.
2518 2538
2519 2539 assert( Thread::current()->is_VM_thread()
2520 2540 || ( CMSScavengeBeforeRemark
2521 2541 && Thread::current()->is_ConcurrentGC_thread()),
2522 2542 "Incorrect thread type for prologue execution");
2523 2543
2524 2544 if (_between_prologue_and_epilogue) {
2525 2545 // We have already been invoked; this is a gc_prologue delegation
2526 2546 // from yet another CMS generation that we are responsible for, just
2527 2547 // ignore it since all relevant work has already been done.
2528 2548 return;
2529 2549 }
2530 2550
2531 2551 // set a bit saying prologue has been called; cleared in epilogue
2532 2552 _between_prologue_and_epilogue = true;
2533 2553 // Claim locks for common data structures, then call gc_prologue_work()
2534 2554 // for each CMSGen and PermGen that we are responsible for.
2535 2555
2536 2556 getFreelistLocks(); // gets free list locks on constituent spaces
2537 2557 bitMapLock()->lock_without_safepoint_check();
2538 2558
2539 2559 // Should call gc_prologue_work() for all cms gens we are responsible for
2540 2560 bool registerClosure = _collectorState >= Marking
2541 2561 && _collectorState < Sweeping;
2542 2562 ModUnionClosure* muc = ParallelGCThreads > 0 ? &_modUnionClosurePar
2543 2563 : &_modUnionClosure;
2544 2564 _cmsGen->gc_prologue_work(full, registerClosure, muc);
2545 2565 _permGen->gc_prologue_work(full, registerClosure, muc);
2546 2566
2547 2567 if (!full) {
2548 2568 stats().record_gc0_begin();
2549 2569 }
2550 2570 }
2551 2571
2552 2572 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2553 2573 // Delegate to CMScollector which knows how to coordinate between
2554 2574 // this and any other CMS generations that it is responsible for
2555 2575 // collecting.
2556 2576 collector()->gc_prologue(full);
2557 2577 }
2558 2578
2559 2579 // This is a "private" interface for use by this generation's CMSCollector.
2560 2580 // Not to be called directly by any other entity (for instance,
2561 2581 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2562 2582 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2563 2583 bool registerClosure, ModUnionClosure* modUnionClosure) {
2564 2584 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2565 2585 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2566 2586 "Should be NULL");
2567 2587 if (registerClosure) {
2568 2588 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2569 2589 }
2570 2590 cmsSpace()->gc_prologue();
2571 2591 // Clear stat counters
2572 2592 NOT_PRODUCT(
2573 2593 assert(_numObjectsPromoted == 0, "check");
2574 2594 assert(_numWordsPromoted == 0, "check");
2575 2595 if (Verbose && PrintGC) {
2576 2596 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2577 2597 SIZE_FORMAT" bytes concurrently",
2578 2598 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2579 2599 }
2580 2600 _numObjectsAllocated = 0;
2581 2601 _numWordsAllocated = 0;
2582 2602 )
2583 2603 }
2584 2604
2585 2605 void CMSCollector::gc_epilogue(bool full) {
2586 2606 // The following locking discipline assumes that we are only called
2587 2607 // when the world is stopped.
2588 2608 assert(SafepointSynchronize::is_at_safepoint(),
2589 2609 "world is stopped assumption");
2590 2610
2591 2611 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2592 2612 // if linear allocation blocks need to be appropriately marked to allow the
2593 2613 // the blocks to be parsable. We also check here whether we need to nudge the
2594 2614 // CMS collector thread to start a new cycle (if it's not already active).
2595 2615 assert( Thread::current()->is_VM_thread()
2596 2616 || ( CMSScavengeBeforeRemark
2597 2617 && Thread::current()->is_ConcurrentGC_thread()),
2598 2618 "Incorrect thread type for epilogue execution");
2599 2619
2600 2620 if (!_between_prologue_and_epilogue) {
2601 2621 // We have already been invoked; this is a gc_epilogue delegation
2602 2622 // from yet another CMS generation that we are responsible for, just
2603 2623 // ignore it since all relevant work has already been done.
2604 2624 return;
2605 2625 }
2606 2626 assert(haveFreelistLocks(), "must have freelist locks");
2607 2627 assert_lock_strong(bitMapLock());
2608 2628
2609 2629 _cmsGen->gc_epilogue_work(full);
2610 2630 _permGen->gc_epilogue_work(full);
2611 2631
2612 2632 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2613 2633 // in case sampling was not already enabled, enable it
2614 2634 _start_sampling = true;
2615 2635 }
2616 2636 // reset _eden_chunk_array so sampling starts afresh
2617 2637 _eden_chunk_index = 0;
2618 2638
2619 2639 size_t cms_used = _cmsGen->cmsSpace()->used();
2620 2640 size_t perm_used = _permGen->cmsSpace()->used();
2621 2641
2622 2642 // update performance counters - this uses a special version of
2623 2643 // update_counters() that allows the utilization to be passed as a
2624 2644 // parameter, avoiding multiple calls to used().
2625 2645 //
2626 2646 _cmsGen->update_counters(cms_used);
2627 2647 _permGen->update_counters(perm_used);
2628 2648
2629 2649 if (CMSIncrementalMode) {
2630 2650 icms_update_allocation_limits();
2631 2651 }
2632 2652
2633 2653 bitMapLock()->unlock();
2634 2654 releaseFreelistLocks();
2635 2655
2636 2656 _between_prologue_and_epilogue = false; // ready for next cycle
2637 2657 }
2638 2658
2639 2659 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2640 2660 collector()->gc_epilogue(full);
2641 2661
2642 2662 // Also reset promotion tracking in par gc thread states.
2643 2663 if (ParallelGCThreads > 0) {
2644 2664 for (uint i = 0; i < ParallelGCThreads; i++) {
2645 2665 _par_gc_thread_states[i]->promo.stopTrackingPromotions();
2646 2666 }
2647 2667 }
2648 2668 }
2649 2669
2650 2670 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2651 2671 assert(!incremental_collection_failed(), "Should have been cleared");
2652 2672 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2653 2673 cmsSpace()->gc_epilogue();
2654 2674 // Print stat counters
2655 2675 NOT_PRODUCT(
2656 2676 assert(_numObjectsAllocated == 0, "check");
2657 2677 assert(_numWordsAllocated == 0, "check");
2658 2678 if (Verbose && PrintGC) {
2659 2679 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2660 2680 SIZE_FORMAT" bytes",
2661 2681 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2662 2682 }
2663 2683 _numObjectsPromoted = 0;
2664 2684 _numWordsPromoted = 0;
2665 2685 )
2666 2686
2667 2687 if (PrintGC && Verbose) {
2668 2688 // Call down the chain in contiguous_available needs the freelistLock
2669 2689 // so print this out before releasing the freeListLock.
2670 2690 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2671 2691 contiguous_available());
2672 2692 }
2673 2693 }
2674 2694
2675 2695 #ifndef PRODUCT
2676 2696 bool CMSCollector::have_cms_token() {
2677 2697 Thread* thr = Thread::current();
2678 2698 if (thr->is_VM_thread()) {
2679 2699 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2680 2700 } else if (thr->is_ConcurrentGC_thread()) {
2681 2701 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2682 2702 } else if (thr->is_GC_task_thread()) {
2683 2703 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2684 2704 ParGCRareEvent_lock->owned_by_self();
2685 2705 }
2686 2706 return false;
2687 2707 }
2688 2708 #endif
2689 2709
2690 2710 // Check reachability of the given heap address in CMS generation,
2691 2711 // treating all other generations as roots.
2692 2712 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2693 2713 // We could "guarantee" below, rather than assert, but i'll
2694 2714 // leave these as "asserts" so that an adventurous debugger
2695 2715 // could try this in the product build provided some subset of
2696 2716 // the conditions were met, provided they were intersted in the
2697 2717 // results and knew that the computation below wouldn't interfere
2698 2718 // with other concurrent computations mutating the structures
2699 2719 // being read or written.
2700 2720 assert(SafepointSynchronize::is_at_safepoint(),
2701 2721 "Else mutations in object graph will make answer suspect");
2702 2722 assert(have_cms_token(), "Should hold cms token");
2703 2723 assert(haveFreelistLocks(), "must hold free list locks");
2704 2724 assert_lock_strong(bitMapLock());
2705 2725
2706 2726 // Clear the marking bit map array before starting, but, just
2707 2727 // for kicks, first report if the given address is already marked
2708 2728 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
2709 2729 _markBitMap.isMarked(addr) ? "" : " not");
2710 2730
2711 2731 if (verify_after_remark()) {
2712 2732 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2713 2733 bool result = verification_mark_bm()->isMarked(addr);
2714 2734 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
2715 2735 result ? "IS" : "is NOT");
2716 2736 return result;
2717 2737 } else {
2718 2738 gclog_or_tty->print_cr("Could not compute result");
2719 2739 return false;
2720 2740 }
2721 2741 }
2722 2742
2723 2743 ////////////////////////////////////////////////////////
2724 2744 // CMS Verification Support
2725 2745 ////////////////////////////////////////////////////////
2726 2746 // Following the remark phase, the following invariant
2727 2747 // should hold -- each object in the CMS heap which is
2728 2748 // marked in markBitMap() should be marked in the verification_mark_bm().
2729 2749
2730 2750 class VerifyMarkedClosure: public BitMapClosure {
2731 2751 CMSBitMap* _marks;
2732 2752 bool _failed;
2733 2753
2734 2754 public:
2735 2755 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2736 2756
2737 2757 void do_bit(size_t offset) {
2738 2758 HeapWord* addr = _marks->offsetToHeapWord(offset);
2739 2759 if (!_marks->isMarked(addr)) {
2740 2760 oop(addr)->print();
2741 2761 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2742 2762 _failed = true;
2743 2763 }
2744 2764 }
2745 2765
2746 2766 bool failed() { return _failed; }
2747 2767 };
2748 2768
2749 2769 bool CMSCollector::verify_after_remark() {
2750 2770 gclog_or_tty->print(" [Verifying CMS Marking... ");
2751 2771 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2752 2772 static bool init = false;
2753 2773
2754 2774 assert(SafepointSynchronize::is_at_safepoint(),
2755 2775 "Else mutations in object graph will make answer suspect");
2756 2776 assert(have_cms_token(),
2757 2777 "Else there may be mutual interference in use of "
2758 2778 " verification data structures");
2759 2779 assert(_collectorState > Marking && _collectorState <= Sweeping,
2760 2780 "Else marking info checked here may be obsolete");
2761 2781 assert(haveFreelistLocks(), "must hold free list locks");
2762 2782 assert_lock_strong(bitMapLock());
2763 2783
2764 2784
2765 2785 // Allocate marking bit map if not already allocated
2766 2786 if (!init) { // first time
2767 2787 if (!verification_mark_bm()->allocate(_span)) {
2768 2788 return false;
2769 2789 }
2770 2790 init = true;
2771 2791 }
2772 2792
2773 2793 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2774 2794
2775 2795 // Turn off refs discovery -- so we will be tracing through refs.
2776 2796 // This is as intended, because by this time
2777 2797 // GC must already have cleared any refs that need to be cleared,
2778 2798 // and traced those that need to be marked; moreover,
2779 2799 // the marking done here is not going to intefere in any
2780 2800 // way with the marking information used by GC.
2781 2801 NoRefDiscovery no_discovery(ref_processor());
2782 2802
2783 2803 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2784 2804
2785 2805 // Clear any marks from a previous round
2786 2806 verification_mark_bm()->clear_all();
2787 2807 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2788 2808 assert(overflow_list_is_empty(), "overflow list should be empty");
2789 2809
2790 2810 GenCollectedHeap* gch = GenCollectedHeap::heap();
2791 2811 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2792 2812 // Update the saved marks which may affect the root scans.
2793 2813 gch->save_marks();
2794 2814
2795 2815 if (CMSRemarkVerifyVariant == 1) {
2796 2816 // In this first variant of verification, we complete
2797 2817 // all marking, then check if the new marks-verctor is
2798 2818 // a subset of the CMS marks-vector.
2799 2819 verify_after_remark_work_1();
2800 2820 } else if (CMSRemarkVerifyVariant == 2) {
2801 2821 // In this second variant of verification, we flag an error
2802 2822 // (i.e. an object reachable in the new marks-vector not reachable
2803 2823 // in the CMS marks-vector) immediately, also indicating the
2804 2824 // identify of an object (A) that references the unmarked object (B) --
2805 2825 // presumably, a mutation to A failed to be picked up by preclean/remark?
2806 2826 verify_after_remark_work_2();
2807 2827 } else {
2808 2828 warning("Unrecognized value %d for CMSRemarkVerifyVariant",
2809 2829 CMSRemarkVerifyVariant);
2810 2830 }
2811 2831 gclog_or_tty->print(" done] ");
2812 2832 return true;
2813 2833 }
2814 2834
2815 2835 void CMSCollector::verify_after_remark_work_1() {
2816 2836 ResourceMark rm;
2817 2837 HandleMark hm;
2818 2838 GenCollectedHeap* gch = GenCollectedHeap::heap();
2819 2839
2820 2840 // Mark from roots one level into CMS
2821 2841 MarkRefsIntoClosure notOlder(_span, verification_mark_bm(), true /* nmethods */);
2822 2842 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2823 2843
2824 2844 gch->gen_process_strong_roots(_cmsGen->level(),
2825 2845 true, // younger gens are roots
2826 2846 true, // collecting perm gen
2827 2847 SharedHeap::ScanningOption(roots_scanning_options()),
2828 2848 NULL, ¬Older);
2829 2849
2830 2850 // Now mark from the roots
2831 2851 assert(_revisitStack.isEmpty(), "Should be empty");
2832 2852 MarkFromRootsClosure markFromRootsClosure(this, _span,
2833 2853 verification_mark_bm(), verification_mark_stack(), &_revisitStack,
2834 2854 false /* don't yield */, true /* verifying */);
2835 2855 assert(_restart_addr == NULL, "Expected pre-condition");
2836 2856 verification_mark_bm()->iterate(&markFromRootsClosure);
2837 2857 while (_restart_addr != NULL) {
2838 2858 // Deal with stack overflow: by restarting at the indicated
2839 2859 // address.
2840 2860 HeapWord* ra = _restart_addr;
2841 2861 markFromRootsClosure.reset(ra);
2842 2862 _restart_addr = NULL;
2843 2863 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2844 2864 }
2845 2865 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2846 2866 verify_work_stacks_empty();
2847 2867 // Should reset the revisit stack above, since no class tree
2848 2868 // surgery is forthcoming.
2849 2869 _revisitStack.reset(); // throwing away all contents
2850 2870
2851 2871 // Marking completed -- now verify that each bit marked in
2852 2872 // verification_mark_bm() is also marked in markBitMap(); flag all
2853 2873 // errors by printing corresponding objects.
2854 2874 VerifyMarkedClosure vcl(markBitMap());
2855 2875 verification_mark_bm()->iterate(&vcl);
2856 2876 if (vcl.failed()) {
2857 2877 gclog_or_tty->print("Verification failed");
2858 2878 Universe::heap()->print();
2859 2879 fatal(" ... aborting");
2860 2880 }
2861 2881 }
2862 2882
2863 2883 void CMSCollector::verify_after_remark_work_2() {
2864 2884 ResourceMark rm;
2865 2885 HandleMark hm;
2866 2886 GenCollectedHeap* gch = GenCollectedHeap::heap();
2867 2887
2868 2888 // Mark from roots one level into CMS
2869 2889 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2870 2890 markBitMap(), true /* nmethods */);
2871 2891 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2872 2892 gch->gen_process_strong_roots(_cmsGen->level(),
2873 2893 true, // younger gens are roots
2874 2894 true, // collecting perm gen
2875 2895 SharedHeap::ScanningOption(roots_scanning_options()),
2876 2896 NULL, ¬Older);
2877 2897
2878 2898 // Now mark from the roots
2879 2899 assert(_revisitStack.isEmpty(), "Should be empty");
2880 2900 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2881 2901 verification_mark_bm(), markBitMap(), verification_mark_stack());
2882 2902 assert(_restart_addr == NULL, "Expected pre-condition");
2883 2903 verification_mark_bm()->iterate(&markFromRootsClosure);
2884 2904 while (_restart_addr != NULL) {
2885 2905 // Deal with stack overflow: by restarting at the indicated
2886 2906 // address.
2887 2907 HeapWord* ra = _restart_addr;
2888 2908 markFromRootsClosure.reset(ra);
2889 2909 _restart_addr = NULL;
2890 2910 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2891 2911 }
2892 2912 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2893 2913 verify_work_stacks_empty();
2894 2914 // Should reset the revisit stack above, since no class tree
2895 2915 // surgery is forthcoming.
2896 2916 _revisitStack.reset(); // throwing away all contents
2897 2917
2898 2918 // Marking completed -- now verify that each bit marked in
2899 2919 // verification_mark_bm() is also marked in markBitMap(); flag all
2900 2920 // errors by printing corresponding objects.
2901 2921 VerifyMarkedClosure vcl(markBitMap());
2902 2922 verification_mark_bm()->iterate(&vcl);
2903 2923 assert(!vcl.failed(), "Else verification above should not have succeeded");
2904 2924 }
2905 2925
2906 2926 void ConcurrentMarkSweepGeneration::save_marks() {
2907 2927 // delegate to CMS space
2908 2928 cmsSpace()->save_marks();
2909 2929 for (uint i = 0; i < ParallelGCThreads; i++) {
2910 2930 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2911 2931 }
2912 2932 }
2913 2933
2914 2934 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2915 2935 return cmsSpace()->no_allocs_since_save_marks();
2916 2936 }
2917 2937
2918 2938 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2919 2939 \
2920 2940 void ConcurrentMarkSweepGeneration:: \
2921 2941 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2922 2942 cl->set_generation(this); \
2923 2943 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2924 2944 cl->reset_generation(); \
2925 2945 save_marks(); \
2926 2946 }
2927 2947
2928 2948 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2929 2949
2930 2950 void
2931 2951 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk)
2932 2952 {
2933 2953 // Not currently implemented; need to do the following. -- ysr.
2934 2954 // dld -- I think that is used for some sort of allocation profiler. So it
2935 2955 // really means the objects allocated by the mutator since the last
2936 2956 // GC. We could potentially implement this cheaply by recording only
2937 2957 // the direct allocations in a side data structure.
2938 2958 //
2939 2959 // I think we probably ought not to be required to support these
2940 2960 // iterations at any arbitrary point; I think there ought to be some
2941 2961 // call to enable/disable allocation profiling in a generation/space,
2942 2962 // and the iterator ought to return the objects allocated in the
2943 2963 // gen/space since the enable call, or the last iterator call (which
2944 2964 // will probably be at a GC.) That way, for gens like CM&S that would
2945 2965 // require some extra data structure to support this, we only pay the
2946 2966 // cost when it's in use...
2947 2967 cmsSpace()->object_iterate_since_last_GC(blk);
2948 2968 }
2949 2969
2950 2970 void
2951 2971 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
2952 2972 cl->set_generation(this);
2953 2973 younger_refs_in_space_iterate(_cmsSpace, cl);
2954 2974 cl->reset_generation();
2955 2975 }
2956 2976
2957 2977 void
2958 2978 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) {
2959 2979 if (freelistLock()->owned_by_self()) {
2960 2980 Generation::oop_iterate(mr, cl);
2961 2981 } else {
2962 2982 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2963 2983 Generation::oop_iterate(mr, cl);
2964 2984 }
2965 2985 }
2966 2986
2967 2987 void
2968 2988 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) {
2969 2989 if (freelistLock()->owned_by_self()) {
2970 2990 Generation::oop_iterate(cl);
2971 2991 } else {
2972 2992 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2973 2993 Generation::oop_iterate(cl);
2974 2994 }
2975 2995 }
2976 2996
2977 2997 void
2978 2998 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2979 2999 if (freelistLock()->owned_by_self()) {
2980 3000 Generation::object_iterate(cl);
2981 3001 } else {
2982 3002 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2983 3003 Generation::object_iterate(cl);
2984 3004 }
2985 3005 }
2986 3006
2987 3007 void
2988 3008 ConcurrentMarkSweepGeneration::pre_adjust_pointers() {
2989 3009 }
2990 3010
2991 3011 void
2992 3012 ConcurrentMarkSweepGeneration::post_compact() {
2993 3013 }
2994 3014
2995 3015 void
2996 3016 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2997 3017 // Fix the linear allocation blocks to look like free blocks.
2998 3018
2999 3019 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3000 3020 // are not called when the heap is verified during universe initialization and
3001 3021 // at vm shutdown.
3002 3022 if (freelistLock()->owned_by_self()) {
3003 3023 cmsSpace()->prepare_for_verify();
3004 3024 } else {
3005 3025 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3006 3026 cmsSpace()->prepare_for_verify();
3007 3027 }
3008 3028 }
3009 3029
3010 3030 void
3011 3031 ConcurrentMarkSweepGeneration::verify(bool allow_dirty /* ignored */) {
3012 3032 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3013 3033 // are not called when the heap is verified during universe initialization and
3014 3034 // at vm shutdown.
3015 3035 if (freelistLock()->owned_by_self()) {
3016 3036 cmsSpace()->verify(false /* ignored */);
3017 3037 } else {
3018 3038 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3019 3039 cmsSpace()->verify(false /* ignored */);
3020 3040 }
3021 3041 }
3022 3042
3023 3043 void CMSCollector::verify(bool allow_dirty /* ignored */) {
3024 3044 _cmsGen->verify(allow_dirty);
3025 3045 _permGen->verify(allow_dirty);
3026 3046 }
3027 3047
3028 3048 #ifndef PRODUCT
3029 3049 bool CMSCollector::overflow_list_is_empty() const {
3030 3050 assert(_num_par_pushes >= 0, "Inconsistency");
3031 3051 if (_overflow_list == NULL) {
3032 3052 assert(_num_par_pushes == 0, "Inconsistency");
3033 3053 }
3034 3054 return _overflow_list == NULL;
3035 3055 }
3036 3056
3037 3057 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3038 3058 // merely consolidate assertion checks that appear to occur together frequently.
3039 3059 void CMSCollector::verify_work_stacks_empty() const {
↓ open down ↓ |
908 lines elided |
↑ open up ↑ |
3040 3060 assert(_markStack.isEmpty(), "Marking stack should be empty");
3041 3061 assert(overflow_list_is_empty(), "Overflow list should be empty");
3042 3062 }
3043 3063
3044 3064 void CMSCollector::verify_overflow_empty() const {
3045 3065 assert(overflow_list_is_empty(), "Overflow list should be empty");
3046 3066 assert(no_preserved_marks(), "No preserved marks");
3047 3067 }
3048 3068 #endif // PRODUCT
3049 3069
3070 +// Decide if we want to enable class unloading as part of the
3071 +// ensuing concurrent GC cycle. We will collect the perm gen and
3072 +// unload classes if it's the case that:
3073 +// (1) an explicit gc request has been made and the flag
3074 +// ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3075 +// (2) (a) class unloading is enabled at the command line, and
3076 +// (b) (i) perm gen threshold has been crossed, or
3077 +// (ii) old gen is getting really full, or
3078 +// (iii) the previous N CMS collections did not collect the
3079 +// perm gen
3080 +// NOTE: Provided there is no change in the state of the heap between
3081 +// calls to this method, it should have idempotent results. Moreover,
3082 +// its results should be monotonically increasing (i.e. going from 0 to 1,
3083 +// but not 1 to 0) between successive calls between which the heap was
3084 +// not collected. For the implementation below, it must thus rely on
3085 +// the property that concurrent_cycles_since_last_unload()
3086 +// will not decrease unless a collection cycle happened and that
3087 +// _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are
3088 +// themselves also monotonic in that sense. See check_monotonicity()
3089 +// below.
3090 +bool CMSCollector::update_should_unload_classes() {
3091 + // Condition 1 above
3092 + if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3093 + _should_unload_classes = true;
3094 + } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3095 + // Disjuncts 2.b.(i,ii,iii) above
3096 + _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3097 + CMSClassUnloadingMaxInterval)
3098 + || _permGen->should_concurrent_collect()
3099 + || _cmsGen->is_too_full();
3100 + }
3101 + return _should_unload_classes;
3102 +}
3103 +
3104 +bool ConcurrentMarkSweepGeneration::is_too_full() const {
3105 + bool res = should_concurrent_collect();
3106 +#define CMSIsTooFullPercentage 98
3107 + res = res && occupancy() > (double)CMSIsTooFullPercentage/100.0;
3108 + return res;
3109 +}
3110 +
3050 3111 void CMSCollector::setup_cms_unloading_and_verification_state() {
3051 3112 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3052 3113 || VerifyBeforeExit;
3053 3114 const int rso = SharedHeap::SO_Symbols | SharedHeap::SO_Strings
3054 3115 | SharedHeap::SO_CodeCache;
3055 3116
3056 - if (cms_should_unload_classes()) { // Should unload classes this cycle
3117 + if (should_unload_classes()) { // Should unload classes this cycle
3057 3118 remove_root_scanning_option(rso); // Shrink the root set appropriately
3058 3119 set_verifying(should_verify); // Set verification state for this cycle
3059 3120 return; // Nothing else needs to be done at this time
3060 3121 }
3061 3122
3062 3123 // Not unloading classes this cycle
3063 - assert(!cms_should_unload_classes(), "Inconsitency!");
3064 - if ((!verifying() || cms_unloaded_classes_last_cycle()) && should_verify) {
3124 + assert(!should_unload_classes(), "Inconsitency!");
3125 + if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3065 3126 // We were not verifying, or we _were_ unloading classes in the last cycle,
3066 3127 // AND some verification options are enabled this cycle; in this case,
3067 3128 // we must make sure that the deadness map is allocated if not already so,
3068 3129 // and cleared (if already allocated previously --
3069 3130 // CMSBitMap::sizeInBits() is used to determine if it's allocated).
3070 3131 if (perm_gen_verify_bit_map()->sizeInBits() == 0) {
3071 3132 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) {
3072 3133 warning("Failed to allocate permanent generation verification CMS Bit Map;\n"
3073 3134 "permanent generation verification disabled");
3074 3135 return; // Note that we leave verification disabled, so we'll retry this
3075 3136 // allocation next cycle. We _could_ remember this failure
3076 3137 // and skip further attempts and permanently disable verification
3077 3138 // attempts if that is considered more desirable.
3078 3139 }
3079 3140 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()),
3080 3141 "_perm_gen_ver_bit_map inconsistency?");
3081 3142 } else {
3082 3143 perm_gen_verify_bit_map()->clear_all();
3083 3144 }
3084 3145 // Include symbols, strings and code cache elements to prevent their resurrection.
3085 3146 add_root_scanning_option(rso);
3086 3147 set_verifying(true);
3087 3148 } else if (verifying() && !should_verify) {
3088 3149 // We were verifying, but some verification flags got disabled.
3089 3150 set_verifying(false);
3090 3151 // Exclude symbols, strings and code cache elements from root scanning to
3091 3152 // reduce IM and RM pauses.
3092 3153 remove_root_scanning_option(rso);
3093 3154 }
3094 3155 }
3095 3156
3096 3157
3097 3158 #ifndef PRODUCT
3098 3159 HeapWord* CMSCollector::block_start(const void* p) const {
3099 3160 const HeapWord* addr = (HeapWord*)p;
3100 3161 if (_span.contains(p)) {
3101 3162 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3102 3163 return _cmsGen->cmsSpace()->block_start(p);
3103 3164 } else {
3104 3165 assert(_permGen->cmsSpace()->is_in_reserved(addr),
3105 3166 "Inconsistent _span?");
3106 3167 return _permGen->cmsSpace()->block_start(p);
3107 3168 }
3108 3169 }
3109 3170 return NULL;
3110 3171 }
3111 3172 #endif
3112 3173
3113 3174 HeapWord*
3114 3175 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3115 3176 bool tlab,
3116 3177 bool parallel) {
3117 3178 assert(!tlab, "Can't deal with TLAB allocation");
3118 3179 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3119 3180 expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3120 3181 CMSExpansionCause::_satisfy_allocation);
3121 3182 if (GCExpandToAllocateDelayMillis > 0) {
3122 3183 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3123 3184 }
3124 3185 size_t adj_word_sz = CompactibleFreeListSpace::adjustObjectSize(word_size);
3125 3186 if (parallel) {
3126 3187 return cmsSpace()->par_allocate(adj_word_sz);
3127 3188 } else {
3128 3189 return cmsSpace()->allocate(adj_word_sz);
3129 3190 }
3130 3191 }
3131 3192
3132 3193 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3133 3194 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3134 3195 // to CardGeneration and share it...
3135 3196 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3136 3197 CMSExpansionCause::Cause cause)
3137 3198 {
3138 3199 assert_locked_or_safepoint(Heap_lock);
3139 3200
3140 3201 size_t aligned_bytes = ReservedSpace::page_align_size_up(bytes);
3141 3202 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes);
3142 3203 bool success = false;
3143 3204 if (aligned_expand_bytes > aligned_bytes) {
3144 3205 success = grow_by(aligned_expand_bytes);
3145 3206 }
3146 3207 if (!success) {
3147 3208 success = grow_by(aligned_bytes);
3148 3209 }
3149 3210 if (!success) {
3150 3211 size_t remaining_bytes = _virtual_space.uncommitted_size();
3151 3212 if (remaining_bytes > 0) {
3152 3213 success = grow_by(remaining_bytes);
3153 3214 }
3154 3215 }
3155 3216 if (GC_locker::is_active()) {
3156 3217 if (PrintGC && Verbose) {
3157 3218 gclog_or_tty->print_cr("Garbage collection disabled, expanded heap instead");
3158 3219 }
3159 3220 }
3160 3221 // remember why we expanded; this information is used
3161 3222 // by shouldConcurrentCollect() when making decisions on whether to start
3162 3223 // a new CMS cycle.
3163 3224 if (success) {
3164 3225 set_expansion_cause(cause);
3165 3226 if (PrintGCDetails && Verbose) {
3166 3227 gclog_or_tty->print_cr("Expanded CMS gen for %s",
3167 3228 CMSExpansionCause::to_string(cause));
3168 3229 }
3169 3230 }
3170 3231 }
3171 3232
3172 3233 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3173 3234 HeapWord* res = NULL;
3174 3235 MutexLocker x(ParGCRareEvent_lock);
3175 3236 while (true) {
3176 3237 // Expansion by some other thread might make alloc OK now:
3177 3238 res = ps->lab.alloc(word_sz);
3178 3239 if (res != NULL) return res;
3179 3240 // If there's not enough expansion space available, give up.
3180 3241 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3181 3242 return NULL;
3182 3243 }
3183 3244 // Otherwise, we try expansion.
3184 3245 expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3185 3246 CMSExpansionCause::_allocate_par_lab);
3186 3247 // Now go around the loop and try alloc again;
3187 3248 // A competing par_promote might beat us to the expansion space,
3188 3249 // so we may go around the loop again if promotion fails agaion.
3189 3250 if (GCExpandToAllocateDelayMillis > 0) {
3190 3251 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3191 3252 }
3192 3253 }
3193 3254 }
3194 3255
3195 3256
3196 3257 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3197 3258 PromotionInfo* promo) {
3198 3259 MutexLocker x(ParGCRareEvent_lock);
3199 3260 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3200 3261 while (true) {
3201 3262 // Expansion by some other thread might make alloc OK now:
3202 3263 if (promo->ensure_spooling_space()) {
3203 3264 assert(promo->has_spooling_space(),
3204 3265 "Post-condition of successful ensure_spooling_space()");
3205 3266 return true;
3206 3267 }
3207 3268 // If there's not enough expansion space available, give up.
3208 3269 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3209 3270 return false;
3210 3271 }
3211 3272 // Otherwise, we try expansion.
3212 3273 expand(refill_size_bytes, MinHeapDeltaBytes,
3213 3274 CMSExpansionCause::_allocate_par_spooling_space);
3214 3275 // Now go around the loop and try alloc again;
3215 3276 // A competing allocation might beat us to the expansion space,
3216 3277 // so we may go around the loop again if allocation fails again.
3217 3278 if (GCExpandToAllocateDelayMillis > 0) {
3218 3279 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3219 3280 }
3220 3281 }
3221 3282 }
3222 3283
3223 3284
3224 3285
3225 3286 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3226 3287 assert_locked_or_safepoint(Heap_lock);
3227 3288 size_t size = ReservedSpace::page_align_size_down(bytes);
3228 3289 if (size > 0) {
3229 3290 shrink_by(size);
3230 3291 }
3231 3292 }
3232 3293
3233 3294 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3234 3295 assert_locked_or_safepoint(Heap_lock);
3235 3296 bool result = _virtual_space.expand_by(bytes);
3236 3297 if (result) {
3237 3298 HeapWord* old_end = _cmsSpace->end();
3238 3299 size_t new_word_size =
3239 3300 heap_word_size(_virtual_space.committed_size());
3240 3301 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3241 3302 _bts->resize(new_word_size); // resize the block offset shared array
3242 3303 Universe::heap()->barrier_set()->resize_covered_region(mr);
3243 3304 // Hmmmm... why doesn't CFLS::set_end verify locking?
3244 3305 // This is quite ugly; FIX ME XXX
3245 3306 _cmsSpace->assert_locked();
3246 3307 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3247 3308
3248 3309 // update the space and generation capacity counters
3249 3310 if (UsePerfData) {
3250 3311 _space_counters->update_capacity();
3251 3312 _gen_counters->update_all();
3252 3313 }
3253 3314
3254 3315 if (Verbose && PrintGC) {
3255 3316 size_t new_mem_size = _virtual_space.committed_size();
3256 3317 size_t old_mem_size = new_mem_size - bytes;
3257 3318 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK",
3258 3319 name(), old_mem_size/K, bytes/K, new_mem_size/K);
3259 3320 }
3260 3321 }
3261 3322 return result;
3262 3323 }
3263 3324
3264 3325 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3265 3326 assert_locked_or_safepoint(Heap_lock);
3266 3327 bool success = true;
3267 3328 const size_t remaining_bytes = _virtual_space.uncommitted_size();
3268 3329 if (remaining_bytes > 0) {
3269 3330 success = grow_by(remaining_bytes);
3270 3331 DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3271 3332 }
3272 3333 return success;
3273 3334 }
3274 3335
3275 3336 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3276 3337 assert_locked_or_safepoint(Heap_lock);
3277 3338 assert_lock_strong(freelistLock());
3278 3339 // XXX Fix when compaction is implemented.
3279 3340 warning("Shrinking of CMS not yet implemented");
3280 3341 return;
3281 3342 }
3282 3343
3283 3344
3284 3345 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3285 3346 // phases.
3286 3347 class CMSPhaseAccounting: public StackObj {
3287 3348 public:
3288 3349 CMSPhaseAccounting(CMSCollector *collector,
3289 3350 const char *phase,
3290 3351 bool print_cr = true);
3291 3352 ~CMSPhaseAccounting();
3292 3353
3293 3354 private:
3294 3355 CMSCollector *_collector;
3295 3356 const char *_phase;
3296 3357 elapsedTimer _wallclock;
3297 3358 bool _print_cr;
3298 3359
3299 3360 public:
3300 3361 // Not MT-safe; so do not pass around these StackObj's
3301 3362 // where they may be accessed by other threads.
3302 3363 jlong wallclock_millis() {
3303 3364 assert(_wallclock.is_active(), "Wall clock should not stop");
3304 3365 _wallclock.stop(); // to record time
3305 3366 jlong ret = _wallclock.milliseconds();
3306 3367 _wallclock.start(); // restart
3307 3368 return ret;
3308 3369 }
3309 3370 };
3310 3371
3311 3372 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3312 3373 const char *phase,
3313 3374 bool print_cr) :
3314 3375 _collector(collector), _phase(phase), _print_cr(print_cr) {
3315 3376
3316 3377 if (PrintCMSStatistics != 0) {
3317 3378 _collector->resetYields();
3318 3379 }
3319 3380 if (PrintGCDetails && PrintGCTimeStamps) {
3320 3381 gclog_or_tty->date_stamp(PrintGCDateStamps);
3321 3382 gclog_or_tty->stamp();
3322 3383 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]",
3323 3384 _collector->cmsGen()->short_name(), _phase);
3324 3385 }
3325 3386 _collector->resetTimer();
3326 3387 _wallclock.start();
3327 3388 _collector->startTimer();
3328 3389 }
3329 3390
3330 3391 CMSPhaseAccounting::~CMSPhaseAccounting() {
3331 3392 assert(_wallclock.is_active(), "Wall clock should not have stopped");
3332 3393 _collector->stopTimer();
3333 3394 _wallclock.stop();
3334 3395 if (PrintGCDetails) {
3335 3396 gclog_or_tty->date_stamp(PrintGCDateStamps);
3336 3397 if (PrintGCTimeStamps) {
3337 3398 gclog_or_tty->stamp();
3338 3399 gclog_or_tty->print(": ");
3339 3400 }
3340 3401 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3341 3402 _collector->cmsGen()->short_name(),
3342 3403 _phase, _collector->timerValue(), _wallclock.seconds());
3343 3404 if (_print_cr) {
3344 3405 gclog_or_tty->print_cr("");
3345 3406 }
3346 3407 if (PrintCMSStatistics != 0) {
3347 3408 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3348 3409 _collector->yields());
3349 3410 }
3350 3411 }
3351 3412 }
3352 3413
3353 3414 // CMS work
3354 3415
3355 3416 // Checkpoint the roots into this generation from outside
3356 3417 // this generation. [Note this initial checkpoint need only
3357 3418 // be approximate -- we'll do a catch up phase subsequently.]
3358 3419 void CMSCollector::checkpointRootsInitial(bool asynch) {
3359 3420 assert(_collectorState == InitialMarking, "Wrong collector state");
3360 3421 check_correct_thread_executing();
3361 3422 ReferenceProcessor* rp = ref_processor();
3362 3423 SpecializationStats::clear();
3363 3424 assert(_restart_addr == NULL, "Control point invariant");
3364 3425 if (asynch) {
3365 3426 // acquire locks for subsequent manipulations
3366 3427 MutexLockerEx x(bitMapLock(),
3367 3428 Mutex::_no_safepoint_check_flag);
3368 3429 checkpointRootsInitialWork(asynch);
3369 3430 rp->verify_no_references_recorded();
3370 3431 rp->enable_discovery(); // enable ("weak") refs discovery
3371 3432 _collectorState = Marking;
3372 3433 } else {
3373 3434 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3374 3435 // which recognizes if we are a CMS generation, and doesn't try to turn on
3375 3436 // discovery; verify that they aren't meddling.
3376 3437 assert(!rp->discovery_is_atomic(),
3377 3438 "incorrect setting of discovery predicate");
3378 3439 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3379 3440 "ref discovery for this generation kind");
3380 3441 // already have locks
3381 3442 checkpointRootsInitialWork(asynch);
3382 3443 rp->enable_discovery(); // now enable ("weak") refs discovery
3383 3444 _collectorState = Marking;
3384 3445 }
3385 3446 SpecializationStats::print();
3386 3447 }
3387 3448
3388 3449 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3389 3450 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3390 3451 assert(_collectorState == InitialMarking, "just checking");
3391 3452
3392 3453 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3393 3454 // precede our marking with a collection of all
3394 3455 // younger generations to keep floating garbage to a minimum.
3395 3456 // XXX: we won't do this for now -- it's an optimization to be done later.
3396 3457
3397 3458 // already have locks
3398 3459 assert_lock_strong(bitMapLock());
3399 3460 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3400 3461
3401 3462 // Setup the verification and class unloading state for this
3402 3463 // CMS collection cycle.
3403 3464 setup_cms_unloading_and_verification_state();
3404 3465
3405 3466 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork",
3406 3467 PrintGCDetails && Verbose, true, gclog_or_tty);)
3407 3468 if (UseAdaptiveSizePolicy) {
3408 3469 size_policy()->checkpoint_roots_initial_begin();
3409 3470 }
3410 3471
3411 3472 // Reset all the PLAB chunk arrays if necessary.
3412 3473 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3413 3474 reset_survivor_plab_arrays();
3414 3475 }
3415 3476
3416 3477 ResourceMark rm;
3417 3478 HandleMark hm;
3418 3479
3419 3480 FalseClosure falseClosure;
3420 3481 // In the case of a synchronous collection, we will elide the
3421 3482 // remark step, so it's important to catch all the nmethod oops
3422 3483 // in this step; hence the last argument to the constrcutor below.
3423 3484 MarkRefsIntoClosure notOlder(_span, &_markBitMap, !asynch /* nmethods */);
3424 3485 GenCollectedHeap* gch = GenCollectedHeap::heap();
3425 3486
3426 3487 verify_work_stacks_empty();
3427 3488 verify_overflow_empty();
3428 3489
3429 3490 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3430 3491 // Update the saved marks which may affect the root scans.
3431 3492 gch->save_marks();
3432 3493
3433 3494 // weak reference processing has not started yet.
3434 3495 ref_processor()->set_enqueuing_is_done(false);
3435 3496
3436 3497 {
3437 3498 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3438 3499 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3439 3500 gch->gen_process_strong_roots(_cmsGen->level(),
3440 3501 true, // younger gens are roots
3441 3502 true, // collecting perm gen
3442 3503 SharedHeap::ScanningOption(roots_scanning_options()),
3443 3504 NULL, ¬Older);
3444 3505 }
3445 3506
3446 3507 // Clear mod-union table; it will be dirtied in the prologue of
3447 3508 // CMS generation per each younger generation collection.
3448 3509
3449 3510 assert(_modUnionTable.isAllClear(),
3450 3511 "Was cleared in most recent final checkpoint phase"
3451 3512 " or no bits are set in the gc_prologue before the start of the next "
3452 3513 "subsequent marking phase.");
3453 3514
3454 3515 // Temporarily disabled, since pre/post-consumption closures don't
3455 3516 // care about precleaned cards
3456 3517 #if 0
3457 3518 {
3458 3519 MemRegion mr = MemRegion((HeapWord*)_virtual_space.low(),
3459 3520 (HeapWord*)_virtual_space.high());
3460 3521 _ct->ct_bs()->preclean_dirty_cards(mr);
3461 3522 }
3462 3523 #endif
3463 3524
3464 3525 // Save the end of the used_region of the constituent generations
3465 3526 // to be used to limit the extent of sweep in each generation.
3466 3527 save_sweep_limits();
3467 3528 if (UseAdaptiveSizePolicy) {
3468 3529 size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3469 3530 }
3470 3531 verify_overflow_empty();
3471 3532 }
3472 3533
3473 3534 bool CMSCollector::markFromRoots(bool asynch) {
3474 3535 // we might be tempted to assert that:
3475 3536 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3476 3537 // "inconsistent argument?");
3477 3538 // However that wouldn't be right, because it's possible that
3478 3539 // a safepoint is indeed in progress as a younger generation
3479 3540 // stop-the-world GC happens even as we mark in this generation.
3480 3541 assert(_collectorState == Marking, "inconsistent state?");
3481 3542 check_correct_thread_executing();
3482 3543 verify_overflow_empty();
3483 3544
3484 3545 bool res;
3485 3546 if (asynch) {
3486 3547
3487 3548 // Start the timers for adaptive size policy for the concurrent phases
3488 3549 // Do it here so that the foreground MS can use the concurrent
3489 3550 // timer since a foreground MS might has the sweep done concurrently
3490 3551 // or STW.
3491 3552 if (UseAdaptiveSizePolicy) {
3492 3553 size_policy()->concurrent_marking_begin();
3493 3554 }
3494 3555
3495 3556 // Weak ref discovery note: We may be discovering weak
3496 3557 // refs in this generation concurrent (but interleaved) with
3497 3558 // weak ref discovery by a younger generation collector.
3498 3559
3499 3560 CMSTokenSyncWithLocks ts(true, bitMapLock());
3500 3561 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3501 3562 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3502 3563 res = markFromRootsWork(asynch);
3503 3564 if (res) {
3504 3565 _collectorState = Precleaning;
3505 3566 } else { // We failed and a foreground collection wants to take over
3506 3567 assert(_foregroundGCIsActive, "internal state inconsistency");
3507 3568 assert(_restart_addr == NULL, "foreground will restart from scratch");
3508 3569 if (PrintGCDetails) {
3509 3570 gclog_or_tty->print_cr("bailing out to foreground collection");
3510 3571 }
3511 3572 }
3512 3573 if (UseAdaptiveSizePolicy) {
3513 3574 size_policy()->concurrent_marking_end();
3514 3575 }
3515 3576 } else {
3516 3577 assert(SafepointSynchronize::is_at_safepoint(),
3517 3578 "inconsistent with asynch == false");
3518 3579 if (UseAdaptiveSizePolicy) {
3519 3580 size_policy()->ms_collection_marking_begin();
3520 3581 }
3521 3582 // already have locks
3522 3583 res = markFromRootsWork(asynch);
3523 3584 _collectorState = FinalMarking;
3524 3585 if (UseAdaptiveSizePolicy) {
3525 3586 GenCollectedHeap* gch = GenCollectedHeap::heap();
3526 3587 size_policy()->ms_collection_marking_end(gch->gc_cause());
3527 3588 }
3528 3589 }
3529 3590 verify_overflow_empty();
3530 3591 return res;
3531 3592 }
3532 3593
3533 3594 bool CMSCollector::markFromRootsWork(bool asynch) {
3534 3595 // iterate over marked bits in bit map, doing a full scan and mark
3535 3596 // from these roots using the following algorithm:
3536 3597 // . if oop is to the right of the current scan pointer,
3537 3598 // mark corresponding bit (we'll process it later)
3538 3599 // . else (oop is to left of current scan pointer)
3539 3600 // push oop on marking stack
3540 3601 // . drain the marking stack
3541 3602
3542 3603 // Note that when we do a marking step we need to hold the
3543 3604 // bit map lock -- recall that direct allocation (by mutators)
3544 3605 // and promotion (by younger generation collectors) is also
3545 3606 // marking the bit map. [the so-called allocate live policy.]
3546 3607 // Because the implementation of bit map marking is not
3547 3608 // robust wrt simultaneous marking of bits in the same word,
3548 3609 // we need to make sure that there is no such interference
3549 3610 // between concurrent such updates.
3550 3611
3551 3612 // already have locks
3552 3613 assert_lock_strong(bitMapLock());
3553 3614
3554 3615 // Clear the revisit stack, just in case there are any
3555 3616 // obsolete contents from a short-circuited previous CMS cycle.
3556 3617 _revisitStack.reset();
3557 3618 verify_work_stacks_empty();
3558 3619 verify_overflow_empty();
3559 3620 assert(_revisitStack.isEmpty(), "tabula rasa");
3560 3621
3561 3622 bool result = false;
3562 3623 if (CMSConcurrentMTEnabled && ParallelCMSThreads > 0) {
3563 3624 result = do_marking_mt(asynch);
3564 3625 } else {
3565 3626 result = do_marking_st(asynch);
3566 3627 }
3567 3628 return result;
3568 3629 }
3569 3630
3570 3631 // Forward decl
3571 3632 class CMSConcMarkingTask;
3572 3633
3573 3634 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3574 3635 CMSCollector* _collector;
3575 3636 CMSConcMarkingTask* _task;
3576 3637 bool _yield;
3577 3638 protected:
3578 3639 virtual void yield();
3579 3640 public:
3580 3641 // "n_threads" is the number of threads to be terminated.
3581 3642 // "queue_set" is a set of work queues of other threads.
3582 3643 // "collector" is the CMS collector associated with this task terminator.
3583 3644 // "yield" indicates whether we need the gang as a whole to yield.
3584 3645 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set,
3585 3646 CMSCollector* collector, bool yield) :
3586 3647 ParallelTaskTerminator(n_threads, queue_set),
3587 3648 _collector(collector),
3588 3649 _yield(yield) { }
3589 3650
3590 3651 void set_task(CMSConcMarkingTask* task) {
3591 3652 _task = task;
3592 3653 }
3593 3654 };
3594 3655
3595 3656 // MT Concurrent Marking Task
3596 3657 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3597 3658 CMSCollector* _collector;
3598 3659 YieldingFlexibleWorkGang* _workers; // the whole gang
3599 3660 int _n_workers; // requested/desired # workers
3600 3661 bool _asynch;
3601 3662 bool _result;
3602 3663 CompactibleFreeListSpace* _cms_space;
3603 3664 CompactibleFreeListSpace* _perm_space;
3604 3665 HeapWord* _global_finger;
3605 3666
3606 3667 // Exposed here for yielding support
3607 3668 Mutex* const _bit_map_lock;
3608 3669
3609 3670 // The per thread work queues, available here for stealing
3610 3671 OopTaskQueueSet* _task_queues;
3611 3672 CMSConcMarkingTerminator _term;
3612 3673
3613 3674 public:
3614 3675 CMSConcMarkingTask(CMSCollector* collector,
3615 3676 CompactibleFreeListSpace* cms_space,
3616 3677 CompactibleFreeListSpace* perm_space,
3617 3678 bool asynch, int n_workers,
3618 3679 YieldingFlexibleWorkGang* workers,
3619 3680 OopTaskQueueSet* task_queues):
3620 3681 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3621 3682 _collector(collector),
3622 3683 _cms_space(cms_space),
3623 3684 _perm_space(perm_space),
3624 3685 _asynch(asynch), _n_workers(n_workers), _result(true),
3625 3686 _workers(workers), _task_queues(task_queues),
3626 3687 _term(n_workers, task_queues, _collector, asynch),
3627 3688 _bit_map_lock(collector->bitMapLock())
3628 3689 {
3629 3690 assert(n_workers <= workers->total_workers(),
3630 3691 "Else termination won't work correctly today"); // XXX FIX ME!
3631 3692 _requested_size = n_workers;
3632 3693 _term.set_task(this);
3633 3694 assert(_cms_space->bottom() < _perm_space->bottom(),
3634 3695 "Finger incorrectly initialized below");
3635 3696 _global_finger = _cms_space->bottom();
3636 3697 }
3637 3698
3638 3699
3639 3700 OopTaskQueueSet* task_queues() { return _task_queues; }
3640 3701
3641 3702 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3642 3703
3643 3704 HeapWord** global_finger_addr() { return &_global_finger; }
3644 3705
3645 3706 CMSConcMarkingTerminator* terminator() { return &_term; }
3646 3707
3647 3708 void work(int i);
3648 3709
3649 3710 virtual void coordinator_yield(); // stuff done by coordinator
3650 3711 bool result() { return _result; }
3651 3712
3652 3713 void reset(HeapWord* ra) {
3653 3714 _term.reset_for_reuse();
3654 3715 }
3655 3716
3656 3717 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3657 3718 OopTaskQueue* work_q);
3658 3719
3659 3720 private:
3660 3721 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3661 3722 void do_work_steal(int i);
3662 3723 void bump_global_finger(HeapWord* f);
3663 3724 };
3664 3725
3665 3726 void CMSConcMarkingTerminator::yield() {
3666 3727 if (ConcurrentMarkSweepThread::should_yield() &&
3667 3728 !_collector->foregroundGCIsActive() &&
3668 3729 _yield) {
3669 3730 _task->yield();
3670 3731 } else {
3671 3732 ParallelTaskTerminator::yield();
3672 3733 }
3673 3734 }
3674 3735
3675 3736 ////////////////////////////////////////////////////////////////
3676 3737 // Concurrent Marking Algorithm Sketch
3677 3738 ////////////////////////////////////////////////////////////////
3678 3739 // Until all tasks exhausted (both spaces):
3679 3740 // -- claim next available chunk
3680 3741 // -- bump global finger via CAS
3681 3742 // -- find first object that starts in this chunk
3682 3743 // and start scanning bitmap from that position
3683 3744 // -- scan marked objects for oops
3684 3745 // -- CAS-mark target, and if successful:
3685 3746 // . if target oop is above global finger (volatile read)
3686 3747 // nothing to do
3687 3748 // . if target oop is in chunk and above local finger
3688 3749 // then nothing to do
3689 3750 // . else push on work-queue
3690 3751 // -- Deal with possible overflow issues:
3691 3752 // . local work-queue overflow causes stuff to be pushed on
3692 3753 // global (common) overflow queue
3693 3754 // . always first empty local work queue
3694 3755 // . then get a batch of oops from global work queue if any
3695 3756 // . then do work stealing
3696 3757 // -- When all tasks claimed (both spaces)
3697 3758 // and local work queue empty,
3698 3759 // then in a loop do:
3699 3760 // . check global overflow stack; steal a batch of oops and trace
3700 3761 // . try to steal from other threads oif GOS is empty
3701 3762 // . if neither is available, offer termination
3702 3763 // -- Terminate and return result
3703 3764 //
3704 3765 void CMSConcMarkingTask::work(int i) {
3705 3766 elapsedTimer _timer;
3706 3767 ResourceMark rm;
3707 3768 HandleMark hm;
3708 3769
3709 3770 DEBUG_ONLY(_collector->verify_overflow_empty();)
3710 3771
3711 3772 // Before we begin work, our work queue should be empty
3712 3773 assert(work_queue(i)->size() == 0, "Expected to be empty");
3713 3774 // Scan the bitmap covering _cms_space, tracing through grey objects.
3714 3775 _timer.start();
3715 3776 do_scan_and_mark(i, _cms_space);
3716 3777 _timer.stop();
3717 3778 if (PrintCMSStatistics != 0) {
3718 3779 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3719 3780 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3720 3781 }
3721 3782
3722 3783 // ... do the same for the _perm_space
3723 3784 _timer.reset();
3724 3785 _timer.start();
3725 3786 do_scan_and_mark(i, _perm_space);
3726 3787 _timer.stop();
3727 3788 if (PrintCMSStatistics != 0) {
3728 3789 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec",
3729 3790 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3730 3791 }
3731 3792
3732 3793 // ... do work stealing
3733 3794 _timer.reset();
3734 3795 _timer.start();
3735 3796 do_work_steal(i);
3736 3797 _timer.stop();
3737 3798 if (PrintCMSStatistics != 0) {
3738 3799 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3739 3800 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3740 3801 }
3741 3802 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3742 3803 assert(work_queue(i)->size() == 0, "Should have been emptied");
3743 3804 // Note that under the current task protocol, the
3744 3805 // following assertion is true even of the spaces
3745 3806 // expanded since the completion of the concurrent
3746 3807 // marking. XXX This will likely change under a strict
3747 3808 // ABORT semantics.
3748 3809 assert(_global_finger > _cms_space->end() &&
3749 3810 _global_finger >= _perm_space->end(),
3750 3811 "All tasks have been completed");
3751 3812 DEBUG_ONLY(_collector->verify_overflow_empty();)
3752 3813 }
3753 3814
3754 3815 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3755 3816 HeapWord* read = _global_finger;
3756 3817 HeapWord* cur = read;
3757 3818 while (f > read) {
3758 3819 cur = read;
3759 3820 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3760 3821 if (cur == read) {
3761 3822 // our cas succeeded
3762 3823 assert(_global_finger >= f, "protocol consistency");
3763 3824 break;
3764 3825 }
3765 3826 }
3766 3827 }
3767 3828
3768 3829 // This is really inefficient, and should be redone by
3769 3830 // using (not yet available) block-read and -write interfaces to the
3770 3831 // stack and the work_queue. XXX FIX ME !!!
3771 3832 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3772 3833 OopTaskQueue* work_q) {
3773 3834 // Fast lock-free check
3774 3835 if (ovflw_stk->length() == 0) {
3775 3836 return false;
3776 3837 }
3777 3838 assert(work_q->size() == 0, "Shouldn't steal");
3778 3839 MutexLockerEx ml(ovflw_stk->par_lock(),
3779 3840 Mutex::_no_safepoint_check_flag);
3780 3841 // Grab up to 1/4 the size of the work queue
3781 3842 size_t num = MIN2((size_t)work_q->max_elems()/4,
3782 3843 (size_t)ParGCDesiredObjsFromOverflowList);
3783 3844 num = MIN2(num, ovflw_stk->length());
3784 3845 for (int i = (int) num; i > 0; i--) {
3785 3846 oop cur = ovflw_stk->pop();
3786 3847 assert(cur != NULL, "Counted wrong?");
3787 3848 work_q->push(cur);
3788 3849 }
3789 3850 return num > 0;
3790 3851 }
3791 3852
3792 3853 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3793 3854 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3794 3855 int n_tasks = pst->n_tasks();
3795 3856 // We allow that there may be no tasks to do here because
3796 3857 // we are restarting after a stack overflow.
3797 3858 assert(pst->valid() || n_tasks == 0, "Uninitializd use?");
3798 3859 int nth_task = 0;
3799 3860
3800 3861 HeapWord* start = sp->bottom();
3801 3862 size_t chunk_size = sp->marking_task_size();
3802 3863 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3803 3864 // Having claimed the nth task in this space,
3804 3865 // compute the chunk that it corresponds to:
3805 3866 MemRegion span = MemRegion(start + nth_task*chunk_size,
3806 3867 start + (nth_task+1)*chunk_size);
3807 3868 // Try and bump the global finger via a CAS;
3808 3869 // note that we need to do the global finger bump
3809 3870 // _before_ taking the intersection below, because
3810 3871 // the task corresponding to that region will be
3811 3872 // deemed done even if the used_region() expands
3812 3873 // because of allocation -- as it almost certainly will
3813 3874 // during start-up while the threads yield in the
3814 3875 // closure below.
3815 3876 HeapWord* finger = span.end();
3816 3877 bump_global_finger(finger); // atomically
3817 3878 // There are null tasks here corresponding to chunks
3818 3879 // beyond the "top" address of the space.
3819 3880 span = span.intersection(sp->used_region());
3820 3881 if (!span.is_empty()) { // Non-null task
3821 3882 // We want to skip the first object because
3822 3883 // the protocol is to scan any object in its entirety
3823 3884 // that _starts_ in this span; a fortiori, any
3824 3885 // object starting in an earlier span is scanned
3825 3886 // as part of an earlier claimed task.
3826 3887 // Below we use the "careful" version of block_start
3827 3888 // so we do not try to navigate uninitialized objects.
3828 3889 HeapWord* prev_obj = sp->block_start_careful(span.start());
3829 3890 // Below we use a variant of block_size that uses the
3830 3891 // Printezis bits to avoid waiting for allocated
3831 3892 // objects to become initialized/parsable.
3832 3893 while (prev_obj < span.start()) {
3833 3894 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3834 3895 if (sz > 0) {
3835 3896 prev_obj += sz;
3836 3897 } else {
3837 3898 // In this case we may end up doing a bit of redundant
3838 3899 // scanning, but that appears unavoidable, short of
3839 3900 // locking the free list locks; see bug 6324141.
3840 3901 break;
3841 3902 }
3842 3903 }
3843 3904 if (prev_obj < span.end()) {
3844 3905 MemRegion my_span = MemRegion(prev_obj, span.end());
3845 3906 // Do the marking work within a non-empty span --
3846 3907 // the last argument to the constructor indicates whether the
3847 3908 // iteration should be incremental with periodic yields.
3848 3909 Par_MarkFromRootsClosure cl(this, _collector, my_span,
3849 3910 &_collector->_markBitMap,
3850 3911 work_queue(i),
3851 3912 &_collector->_markStack,
3852 3913 &_collector->_revisitStack,
3853 3914 _asynch);
3854 3915 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3855 3916 } // else nothing to do for this task
3856 3917 } // else nothing to do for this task
3857 3918 }
3858 3919 // We'd be tempted to assert here that since there are no
3859 3920 // more tasks left to claim in this space, the global_finger
3860 3921 // must exceed space->top() and a fortiori space->end(). However,
3861 3922 // that would not quite be correct because the bumping of
3862 3923 // global_finger occurs strictly after the claiming of a task,
3863 3924 // so by the time we reach here the global finger may not yet
3864 3925 // have been bumped up by the thread that claimed the last
3865 3926 // task.
3866 3927 pst->all_tasks_completed();
3867 3928 }
3868 3929
3869 3930 class Par_ConcMarkingClosure: public OopClosure {
3870 3931 CMSCollector* _collector;
3871 3932 MemRegion _span;
3872 3933 CMSBitMap* _bit_map;
3873 3934 CMSMarkStack* _overflow_stack;
3874 3935 CMSMarkStack* _revisit_stack; // XXXXXX Check proper use
3875 3936 OopTaskQueue* _work_queue;
3876 3937
3877 3938 public:
3878 3939 Par_ConcMarkingClosure(CMSCollector* collector, OopTaskQueue* work_queue,
3879 3940 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3880 3941 _collector(collector),
3881 3942 _span(_collector->_span),
3882 3943 _work_queue(work_queue),
3883 3944 _bit_map(bit_map),
3884 3945 _overflow_stack(overflow_stack) { } // need to initialize revisit stack etc.
3885 3946
3886 3947 void do_oop(oop* p);
3887 3948 void trim_queue(size_t max);
3888 3949 void handle_stack_overflow(HeapWord* lost);
3889 3950 };
3890 3951
3891 3952 // Grey object rescan during work stealing phase --
3892 3953 // the salient assumption here is that stolen oops must
3893 3954 // always be initialized, so we do not need to check for
3894 3955 // uninitialized objects before scanning here.
3895 3956 void Par_ConcMarkingClosure::do_oop(oop* p) {
3896 3957 oop this_oop = *p;
3897 3958 assert(this_oop->is_oop_or_null(),
3898 3959 "expected an oop or NULL");
3899 3960 HeapWord* addr = (HeapWord*)this_oop;
3900 3961 // Check if oop points into the CMS generation
3901 3962 // and is not marked
3902 3963 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3903 3964 // a white object ...
3904 3965 // If we manage to "claim" the object, by being the
3905 3966 // first thread to mark it, then we push it on our
3906 3967 // marking stack
3907 3968 if (_bit_map->par_mark(addr)) { // ... now grey
3908 3969 // push on work queue (grey set)
3909 3970 bool simulate_overflow = false;
3910 3971 NOT_PRODUCT(
3911 3972 if (CMSMarkStackOverflowALot &&
3912 3973 _collector->simulate_overflow()) {
3913 3974 // simulate a stack overflow
3914 3975 simulate_overflow = true;
3915 3976 }
3916 3977 )
3917 3978 if (simulate_overflow ||
3918 3979 !(_work_queue->push(this_oop) || _overflow_stack->par_push(this_oop))) {
3919 3980 // stack overflow
3920 3981 if (PrintCMSStatistics != 0) {
3921 3982 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
3922 3983 SIZE_FORMAT, _overflow_stack->capacity());
3923 3984 }
3924 3985 // We cannot assert that the overflow stack is full because
3925 3986 // it may have been emptied since.
3926 3987 assert(simulate_overflow ||
3927 3988 _work_queue->size() == _work_queue->max_elems(),
3928 3989 "Else push should have succeeded");
3929 3990 handle_stack_overflow(addr);
3930 3991 }
3931 3992 } // Else, some other thread got there first
3932 3993 }
3933 3994 }
3934 3995
3935 3996 void Par_ConcMarkingClosure::trim_queue(size_t max) {
3936 3997 while (_work_queue->size() > max) {
3937 3998 oop new_oop;
3938 3999 if (_work_queue->pop_local(new_oop)) {
3939 4000 assert(new_oop->is_oop(), "Should be an oop");
3940 4001 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3941 4002 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3942 4003 assert(new_oop->is_parsable(), "Should be parsable");
3943 4004 new_oop->oop_iterate(this); // do_oop() above
3944 4005 }
3945 4006 }
3946 4007 }
3947 4008
3948 4009 // Upon stack overflow, we discard (part of) the stack,
3949 4010 // remembering the least address amongst those discarded
3950 4011 // in CMSCollector's _restart_address.
3951 4012 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3952 4013 // We need to do this under a mutex to prevent other
3953 4014 // workers from interfering with the expansion below.
3954 4015 MutexLockerEx ml(_overflow_stack->par_lock(),
3955 4016 Mutex::_no_safepoint_check_flag);
3956 4017 // Remember the least grey address discarded
3957 4018 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3958 4019 _collector->lower_restart_addr(ra);
3959 4020 _overflow_stack->reset(); // discard stack contents
3960 4021 _overflow_stack->expand(); // expand the stack if possible
3961 4022 }
3962 4023
3963 4024
3964 4025 void CMSConcMarkingTask::do_work_steal(int i) {
3965 4026 OopTaskQueue* work_q = work_queue(i);
3966 4027 oop obj_to_scan;
3967 4028 CMSBitMap* bm = &(_collector->_markBitMap);
3968 4029 CMSMarkStack* ovflw = &(_collector->_markStack);
3969 4030 int* seed = _collector->hash_seed(i);
3970 4031 Par_ConcMarkingClosure cl(_collector, work_q, bm, ovflw);
3971 4032 while (true) {
3972 4033 cl.trim_queue(0);
3973 4034 assert(work_q->size() == 0, "Should have been emptied above");
3974 4035 if (get_work_from_overflow_stack(ovflw, work_q)) {
3975 4036 // Can't assert below because the work obtained from the
3976 4037 // overflow stack may already have been stolen from us.
3977 4038 // assert(work_q->size() > 0, "Work from overflow stack");
3978 4039 continue;
3979 4040 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3980 4041 assert(obj_to_scan->is_oop(), "Should be an oop");
3981 4042 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3982 4043 obj_to_scan->oop_iterate(&cl);
3983 4044 } else if (terminator()->offer_termination()) {
3984 4045 assert(work_q->size() == 0, "Impossible!");
3985 4046 break;
3986 4047 }
3987 4048 }
3988 4049 }
3989 4050
3990 4051 // This is run by the CMS (coordinator) thread.
3991 4052 void CMSConcMarkingTask::coordinator_yield() {
3992 4053 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3993 4054 "CMS thread should hold CMS token");
3994 4055
3995 4056 // First give up the locks, then yield, then re-lock
3996 4057 // We should probably use a constructor/destructor idiom to
3997 4058 // do this unlock/lock or modify the MutexUnlocker class to
3998 4059 // serve our purpose. XXX
3999 4060 assert_lock_strong(_bit_map_lock);
4000 4061 _bit_map_lock->unlock();
4001 4062 ConcurrentMarkSweepThread::desynchronize(true);
4002 4063 ConcurrentMarkSweepThread::acknowledge_yield_request();
4003 4064 _collector->stopTimer();
4004 4065 if (PrintCMSStatistics != 0) {
4005 4066 _collector->incrementYields();
4006 4067 }
4007 4068 _collector->icms_wait();
4008 4069
4009 4070 // It is possible for whichever thread initiated the yield request
4010 4071 // not to get a chance to wake up and take the bitmap lock between
4011 4072 // this thread releasing it and reacquiring it. So, while the
4012 4073 // should_yield() flag is on, let's sleep for a bit to give the
4013 4074 // other thread a chance to wake up. The limit imposed on the number
4014 4075 // of iterations is defensive, to avoid any unforseen circumstances
4015 4076 // putting us into an infinite loop. Since it's always been this
4016 4077 // (coordinator_yield()) method that was observed to cause the
4017 4078 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4018 4079 // which is by default non-zero. For the other seven methods that
4019 4080 // also perform the yield operation, as are using a different
4020 4081 // parameter (CMSYieldSleepCount) which is by default zero. This way we
4021 4082 // can enable the sleeping for those methods too, if necessary.
4022 4083 // See 6442774.
4023 4084 //
4024 4085 // We really need to reconsider the synchronization between the GC
4025 4086 // thread and the yield-requesting threads in the future and we
4026 4087 // should really use wait/notify, which is the recommended
4027 4088 // way of doing this type of interaction. Additionally, we should
4028 4089 // consolidate the eight methods that do the yield operation and they
4029 4090 // are almost identical into one for better maintenability and
4030 4091 // readability. See 6445193.
4031 4092 //
4032 4093 // Tony 2006.06.29
4033 4094 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4034 4095 ConcurrentMarkSweepThread::should_yield() &&
4035 4096 !CMSCollector::foregroundGCIsActive(); ++i) {
4036 4097 os::sleep(Thread::current(), 1, false);
4037 4098 ConcurrentMarkSweepThread::acknowledge_yield_request();
4038 4099 }
4039 4100
4040 4101 ConcurrentMarkSweepThread::synchronize(true);
4041 4102 _bit_map_lock->lock_without_safepoint_check();
4042 4103 _collector->startTimer();
4043 4104 }
4044 4105
4045 4106 bool CMSCollector::do_marking_mt(bool asynch) {
4046 4107 assert(ParallelCMSThreads > 0 && conc_workers() != NULL, "precondition");
4047 4108 // In the future this would be determined ergonomically, based
4048 4109 // on #cpu's, # active mutator threads (and load), and mutation rate.
4049 4110 int num_workers = ParallelCMSThreads;
4050 4111
4051 4112 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4052 4113 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
4053 4114
4054 4115 CMSConcMarkingTask tsk(this, cms_space, perm_space,
4055 4116 asynch, num_workers /* number requested XXX */,
4056 4117 conc_workers(), task_queues());
4057 4118
4058 4119 // Since the actual number of workers we get may be different
4059 4120 // from the number we requested above, do we need to do anything different
4060 4121 // below? In particular, may be we need to subclass the SequantialSubTasksDone
4061 4122 // class?? XXX
4062 4123 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4063 4124 perm_space->initialize_sequential_subtasks_for_marking(num_workers);
4064 4125
4065 4126 // Refs discovery is already non-atomic.
4066 4127 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4067 4128 // Mutate the Refs discovery so it is MT during the
4068 4129 // multi-threaded marking phase.
4069 4130 ReferenceProcessorMTMutator mt(ref_processor(), num_workers > 1);
4070 4131
4071 4132 conc_workers()->start_task(&tsk);
4072 4133 while (tsk.yielded()) {
4073 4134 tsk.coordinator_yield();
4074 4135 conc_workers()->continue_task(&tsk);
4075 4136 }
4076 4137 // If the task was aborted, _restart_addr will be non-NULL
4077 4138 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4078 4139 while (_restart_addr != NULL) {
4079 4140 // XXX For now we do not make use of ABORTED state and have not
4080 4141 // yet implemented the right abort semantics (even in the original
4081 4142 // single-threaded CMS case). That needs some more investigation
4082 4143 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4083 4144 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4084 4145 // If _restart_addr is non-NULL, a marking stack overflow
4085 4146 // occured; we need to do a fresh marking iteration from the
4086 4147 // indicated restart address.
4087 4148 if (_foregroundGCIsActive && asynch) {
4088 4149 // We may be running into repeated stack overflows, having
4089 4150 // reached the limit of the stack size, while making very
4090 4151 // slow forward progress. It may be best to bail out and
4091 4152 // let the foreground collector do its job.
4092 4153 // Clear _restart_addr, so that foreground GC
4093 4154 // works from scratch. This avoids the headache of
4094 4155 // a "rescan" which would otherwise be needed because
4095 4156 // of the dirty mod union table & card table.
4096 4157 _restart_addr = NULL;
4097 4158 return false;
4098 4159 }
4099 4160 // Adjust the task to restart from _restart_addr
4100 4161 tsk.reset(_restart_addr);
4101 4162 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4102 4163 _restart_addr);
4103 4164 perm_space->initialize_sequential_subtasks_for_marking(num_workers,
4104 4165 _restart_addr);
4105 4166 _restart_addr = NULL;
4106 4167 // Get the workers going again
4107 4168 conc_workers()->start_task(&tsk);
4108 4169 while (tsk.yielded()) {
4109 4170 tsk.coordinator_yield();
4110 4171 conc_workers()->continue_task(&tsk);
4111 4172 }
4112 4173 }
4113 4174 assert(tsk.completed(), "Inconsistency");
4114 4175 assert(tsk.result() == true, "Inconsistency");
4115 4176 return true;
4116 4177 }
4117 4178
4118 4179 bool CMSCollector::do_marking_st(bool asynch) {
4119 4180 ResourceMark rm;
4120 4181 HandleMark hm;
4121 4182
4122 4183 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4123 4184 &_markStack, &_revisitStack, CMSYield && asynch);
4124 4185 // the last argument to iterate indicates whether the iteration
4125 4186 // should be incremental with periodic yields.
4126 4187 _markBitMap.iterate(&markFromRootsClosure);
4127 4188 // If _restart_addr is non-NULL, a marking stack overflow
4128 4189 // occured; we need to do a fresh iteration from the
4129 4190 // indicated restart address.
4130 4191 while (_restart_addr != NULL) {
4131 4192 if (_foregroundGCIsActive && asynch) {
4132 4193 // We may be running into repeated stack overflows, having
4133 4194 // reached the limit of the stack size, while making very
4134 4195 // slow forward progress. It may be best to bail out and
4135 4196 // let the foreground collector do its job.
4136 4197 // Clear _restart_addr, so that foreground GC
4137 4198 // works from scratch. This avoids the headache of
4138 4199 // a "rescan" which would otherwise be needed because
4139 4200 // of the dirty mod union table & card table.
4140 4201 _restart_addr = NULL;
4141 4202 return false; // indicating failure to complete marking
4142 4203 }
4143 4204 // Deal with stack overflow:
4144 4205 // we restart marking from _restart_addr
4145 4206 HeapWord* ra = _restart_addr;
4146 4207 markFromRootsClosure.reset(ra);
4147 4208 _restart_addr = NULL;
4148 4209 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4149 4210 }
4150 4211 return true;
4151 4212 }
4152 4213
4153 4214 void CMSCollector::preclean() {
4154 4215 check_correct_thread_executing();
4155 4216 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4156 4217 verify_work_stacks_empty();
4157 4218 verify_overflow_empty();
4158 4219 _abort_preclean = false;
4159 4220 if (CMSPrecleaningEnabled) {
4160 4221 _eden_chunk_index = 0;
4161 4222 size_t used = get_eden_used();
4162 4223 size_t capacity = get_eden_capacity();
4163 4224 // Don't start sampling unless we will get sufficiently
4164 4225 // many samples.
4165 4226 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4166 4227 * CMSScheduleRemarkEdenPenetration)) {
4167 4228 _start_sampling = true;
4168 4229 } else {
4169 4230 _start_sampling = false;
4170 4231 }
4171 4232 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4172 4233 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
4173 4234 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4174 4235 }
4175 4236 CMSTokenSync x(true); // is cms thread
4176 4237 if (CMSPrecleaningEnabled) {
4177 4238 sample_eden();
4178 4239 _collectorState = AbortablePreclean;
4179 4240 } else {
4180 4241 _collectorState = FinalMarking;
4181 4242 }
4182 4243 verify_work_stacks_empty();
4183 4244 verify_overflow_empty();
4184 4245 }
4185 4246
4186 4247 // Try and schedule the remark such that young gen
4187 4248 // occupancy is CMSScheduleRemarkEdenPenetration %.
4188 4249 void CMSCollector::abortable_preclean() {
4189 4250 check_correct_thread_executing();
4190 4251 assert(CMSPrecleaningEnabled, "Inconsistent control state");
4191 4252 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4192 4253
4193 4254 // If Eden's current occupancy is below this threshold,
4194 4255 // immediately schedule the remark; else preclean
4195 4256 // past the next scavenge in an effort to
4196 4257 // schedule the pause as described avove. By choosing
4197 4258 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4198 4259 // we will never do an actual abortable preclean cycle.
4199 4260 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4200 4261 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4201 4262 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
4202 4263 // We need more smarts in the abortable preclean
4203 4264 // loop below to deal with cases where allocation
4204 4265 // in young gen is very very slow, and our precleaning
4205 4266 // is running a losing race against a horde of
4206 4267 // mutators intent on flooding us with CMS updates
4207 4268 // (dirty cards).
4208 4269 // One, admittedly dumb, strategy is to give up
4209 4270 // after a certain number of abortable precleaning loops
4210 4271 // or after a certain maximum time. We want to make
4211 4272 // this smarter in the next iteration.
4212 4273 // XXX FIX ME!!! YSR
4213 4274 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4214 4275 while (!(should_abort_preclean() ||
4215 4276 ConcurrentMarkSweepThread::should_terminate())) {
4216 4277 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4217 4278 cumworkdone += workdone;
4218 4279 loops++;
4219 4280 // Voluntarily terminate abortable preclean phase if we have
4220 4281 // been at it for too long.
4221 4282 if ((CMSMaxAbortablePrecleanLoops != 0) &&
4222 4283 loops >= CMSMaxAbortablePrecleanLoops) {
4223 4284 if (PrintGCDetails) {
4224 4285 gclog_or_tty->print(" CMS: abort preclean due to loops ");
4225 4286 }
4226 4287 break;
4227 4288 }
4228 4289 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4229 4290 if (PrintGCDetails) {
4230 4291 gclog_or_tty->print(" CMS: abort preclean due to time ");
4231 4292 }
4232 4293 break;
4233 4294 }
4234 4295 // If we are doing little work each iteration, we should
4235 4296 // take a short break.
4236 4297 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4237 4298 // Sleep for some time, waiting for work to accumulate
4238 4299 stopTimer();
4239 4300 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4240 4301 startTimer();
4241 4302 waited++;
4242 4303 }
4243 4304 }
4244 4305 if (PrintCMSStatistics > 0) {
4245 4306 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4246 4307 loops, waited, cumworkdone);
4247 4308 }
4248 4309 }
4249 4310 CMSTokenSync x(true); // is cms thread
4250 4311 if (_collectorState != Idling) {
4251 4312 assert(_collectorState == AbortablePreclean,
4252 4313 "Spontaneous state transition?");
4253 4314 _collectorState = FinalMarking;
4254 4315 } // Else, a foreground collection completed this CMS cycle.
4255 4316 return;
4256 4317 }
4257 4318
4258 4319 // Respond to an Eden sampling opportunity
4259 4320 void CMSCollector::sample_eden() {
4260 4321 // Make sure a young gc cannot sneak in between our
4261 4322 // reading and recording of a sample.
4262 4323 assert(Thread::current()->is_ConcurrentGC_thread(),
4263 4324 "Only the cms thread may collect Eden samples");
4264 4325 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4265 4326 "Should collect samples while holding CMS token");
4266 4327 if (!_start_sampling) {
4267 4328 return;
4268 4329 }
4269 4330 if (_eden_chunk_array) {
4270 4331 if (_eden_chunk_index < _eden_chunk_capacity) {
4271 4332 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
4272 4333 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4273 4334 "Unexpected state of Eden");
4274 4335 // We'd like to check that what we just sampled is an oop-start address;
4275 4336 // however, we cannot do that here since the object may not yet have been
4276 4337 // initialized. So we'll instead do the check when we _use_ this sample
4277 4338 // later.
4278 4339 if (_eden_chunk_index == 0 ||
4279 4340 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4280 4341 _eden_chunk_array[_eden_chunk_index-1])
4281 4342 >= CMSSamplingGrain)) {
4282 4343 _eden_chunk_index++; // commit sample
4283 4344 }
4284 4345 }
4285 4346 }
4286 4347 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4287 4348 size_t used = get_eden_used();
4288 4349 size_t capacity = get_eden_capacity();
4289 4350 assert(used <= capacity, "Unexpected state of Eden");
4290 4351 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4291 4352 _abort_preclean = true;
4292 4353 }
4293 4354 }
4294 4355 }
4295 4356
4296 4357
4297 4358 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4298 4359 assert(_collectorState == Precleaning ||
4299 4360 _collectorState == AbortablePreclean, "incorrect state");
4300 4361 ResourceMark rm;
4301 4362 HandleMark hm;
4302 4363 // Do one pass of scrubbing the discovered reference lists
4303 4364 // to remove any reference objects with strongly-reachable
4304 4365 // referents.
4305 4366 if (clean_refs) {
4306 4367 ReferenceProcessor* rp = ref_processor();
4307 4368 CMSPrecleanRefsYieldClosure yield_cl(this);
4308 4369 assert(rp->span().equals(_span), "Spans should be equal");
4309 4370 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4310 4371 &_markStack);
4311 4372 CMSDrainMarkingStackClosure complete_trace(this,
4312 4373 _span, &_markBitMap, &_markStack,
4313 4374 &keep_alive);
4314 4375
4315 4376 // We don't want this step to interfere with a young
4316 4377 // collection because we don't want to take CPU
4317 4378 // or memory bandwidth away from the young GC threads
4318 4379 // (which may be as many as there are CPUs).
4319 4380 // Note that we don't need to protect ourselves from
4320 4381 // interference with mutators because they can't
4321 4382 // manipulate the discovered reference lists nor affect
4322 4383 // the computed reachability of the referents, the
4323 4384 // only properties manipulated by the precleaning
4324 4385 // of these reference lists.
4325 4386 stopTimer();
4326 4387 CMSTokenSyncWithLocks x(true /* is cms thread */,
4327 4388 bitMapLock());
4328 4389 startTimer();
4329 4390 sample_eden();
4330 4391 // The following will yield to allow foreground
4331 4392 // collection to proceed promptly. XXX YSR:
4332 4393 // The code in this method may need further
4333 4394 // tweaking for better performance and some restructuring
4334 4395 // for cleaner interfaces.
4335 4396 rp->preclean_discovered_references(
4336 4397 rp->is_alive_non_header(), &keep_alive, &complete_trace,
4337 4398 &yield_cl);
4338 4399 }
4339 4400
4340 4401 if (clean_survivor) { // preclean the active survivor space(s)
4341 4402 assert(_young_gen->kind() == Generation::DefNew ||
4342 4403 _young_gen->kind() == Generation::ParNew ||
4343 4404 _young_gen->kind() == Generation::ASParNew,
4344 4405 "incorrect type for cast");
4345 4406 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4346 4407 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4347 4408 &_markBitMap, &_modUnionTable,
4348 4409 &_markStack, &_revisitStack,
4349 4410 true /* precleaning phase */);
4350 4411 stopTimer();
4351 4412 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4352 4413 bitMapLock());
4353 4414 startTimer();
4354 4415 unsigned int before_count =
4355 4416 GenCollectedHeap::heap()->total_collections();
4356 4417 SurvivorSpacePrecleanClosure
4357 4418 sss_cl(this, _span, &_markBitMap, &_markStack,
4358 4419 &pam_cl, before_count, CMSYield);
4359 4420 dng->from()->object_iterate_careful(&sss_cl);
4360 4421 dng->to()->object_iterate_careful(&sss_cl);
4361 4422 }
4362 4423 MarkRefsIntoAndScanClosure
4363 4424 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4364 4425 &_markStack, &_revisitStack, this, CMSYield,
4365 4426 true /* precleaning phase */);
4366 4427 // CAUTION: The following closure has persistent state that may need to
4367 4428 // be reset upon a decrease in the sequence of addresses it
4368 4429 // processes.
4369 4430 ScanMarkedObjectsAgainCarefullyClosure
4370 4431 smoac_cl(this, _span,
4371 4432 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield);
4372 4433
4373 4434 // Preclean dirty cards in ModUnionTable and CardTable using
4374 4435 // appropriate convergence criterion;
4375 4436 // repeat CMSPrecleanIter times unless we find that
4376 4437 // we are losing.
4377 4438 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4378 4439 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4379 4440 "Bad convergence multiplier");
4380 4441 assert(CMSPrecleanThreshold >= 100,
4381 4442 "Unreasonably low CMSPrecleanThreshold");
4382 4443
4383 4444 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4384 4445 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4385 4446 numIter < CMSPrecleanIter;
4386 4447 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4387 4448 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4388 4449 if (CMSPermGenPrecleaningEnabled) {
4389 4450 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl);
4390 4451 }
4391 4452 if (Verbose && PrintGCDetails) {
4392 4453 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4393 4454 }
4394 4455 // Either there are very few dirty cards, so re-mark
4395 4456 // pause will be small anyway, or our pre-cleaning isn't
4396 4457 // that much faster than the rate at which cards are being
4397 4458 // dirtied, so we might as well stop and re-mark since
4398 4459 // precleaning won't improve our re-mark time by much.
4399 4460 if (curNumCards <= CMSPrecleanThreshold ||
4400 4461 (numIter > 0 &&
4401 4462 (curNumCards * CMSPrecleanDenominator >
4402 4463 lastNumCards * CMSPrecleanNumerator))) {
4403 4464 numIter++;
4404 4465 cumNumCards += curNumCards;
4405 4466 break;
4406 4467 }
4407 4468 }
4408 4469 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4409 4470 if (CMSPermGenPrecleaningEnabled) {
4410 4471 curNumCards += preclean_card_table(_permGen, &smoac_cl);
4411 4472 }
4412 4473 cumNumCards += curNumCards;
4413 4474 if (PrintGCDetails && PrintCMSStatistics != 0) {
4414 4475 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4415 4476 curNumCards, cumNumCards, numIter);
4416 4477 }
4417 4478 return cumNumCards; // as a measure of useful work done
4418 4479 }
4419 4480
4420 4481 // PRECLEANING NOTES:
4421 4482 // Precleaning involves:
4422 4483 // . reading the bits of the modUnionTable and clearing the set bits.
4423 4484 // . For the cards corresponding to the set bits, we scan the
4424 4485 // objects on those cards. This means we need the free_list_lock
4425 4486 // so that we can safely iterate over the CMS space when scanning
4426 4487 // for oops.
4427 4488 // . When we scan the objects, we'll be both reading and setting
4428 4489 // marks in the marking bit map, so we'll need the marking bit map.
4429 4490 // . For protecting _collector_state transitions, we take the CGC_lock.
4430 4491 // Note that any races in the reading of of card table entries by the
4431 4492 // CMS thread on the one hand and the clearing of those entries by the
4432 4493 // VM thread or the setting of those entries by the mutator threads on the
4433 4494 // other are quite benign. However, for efficiency it makes sense to keep
4434 4495 // the VM thread from racing with the CMS thread while the latter is
4435 4496 // dirty card info to the modUnionTable. We therefore also use the
4436 4497 // CGC_lock to protect the reading of the card table and the mod union
4437 4498 // table by the CM thread.
4438 4499 // . We run concurrently with mutator updates, so scanning
4439 4500 // needs to be done carefully -- we should not try to scan
4440 4501 // potentially uninitialized objects.
4441 4502 //
4442 4503 // Locking strategy: While holding the CGC_lock, we scan over and
4443 4504 // reset a maximal dirty range of the mod union / card tables, then lock
4444 4505 // the free_list_lock and bitmap lock to do a full marking, then
4445 4506 // release these locks; and repeat the cycle. This allows for a
4446 4507 // certain amount of fairness in the sharing of these locks between
4447 4508 // the CMS collector on the one hand, and the VM thread and the
4448 4509 // mutators on the other.
4449 4510
4450 4511 // NOTE: preclean_mod_union_table() and preclean_card_table()
4451 4512 // further below are largely identical; if you need to modify
4452 4513 // one of these methods, please check the other method too.
4453 4514
4454 4515 size_t CMSCollector::preclean_mod_union_table(
4455 4516 ConcurrentMarkSweepGeneration* gen,
4456 4517 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4457 4518 verify_work_stacks_empty();
4458 4519 verify_overflow_empty();
4459 4520
4460 4521 // strategy: starting with the first card, accumulate contiguous
4461 4522 // ranges of dirty cards; clear these cards, then scan the region
4462 4523 // covered by these cards.
4463 4524
4464 4525 // Since all of the MUT is committed ahead, we can just use
4465 4526 // that, in case the generations expand while we are precleaning.
4466 4527 // It might also be fine to just use the committed part of the
4467 4528 // generation, but we might potentially miss cards when the
4468 4529 // generation is rapidly expanding while we are in the midst
4469 4530 // of precleaning.
4470 4531 HeapWord* startAddr = gen->reserved().start();
4471 4532 HeapWord* endAddr = gen->reserved().end();
4472 4533
4473 4534 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4474 4535
4475 4536 size_t numDirtyCards, cumNumDirtyCards;
4476 4537 HeapWord *nextAddr, *lastAddr;
4477 4538 for (cumNumDirtyCards = numDirtyCards = 0,
4478 4539 nextAddr = lastAddr = startAddr;
4479 4540 nextAddr < endAddr;
4480 4541 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4481 4542
4482 4543 ResourceMark rm;
4483 4544 HandleMark hm;
4484 4545
4485 4546 MemRegion dirtyRegion;
4486 4547 {
4487 4548 stopTimer();
4488 4549 CMSTokenSync ts(true);
4489 4550 startTimer();
4490 4551 sample_eden();
4491 4552 // Get dirty region starting at nextOffset (inclusive),
4492 4553 // simultaneously clearing it.
4493 4554 dirtyRegion =
4494 4555 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4495 4556 assert(dirtyRegion.start() >= nextAddr,
4496 4557 "returned region inconsistent?");
4497 4558 }
4498 4559 // Remember where the next search should begin.
4499 4560 // The returned region (if non-empty) is a right open interval,
4500 4561 // so lastOffset is obtained from the right end of that
4501 4562 // interval.
4502 4563 lastAddr = dirtyRegion.end();
4503 4564 // Should do something more transparent and less hacky XXX
4504 4565 numDirtyCards =
4505 4566 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4506 4567
4507 4568 // We'll scan the cards in the dirty region (with periodic
4508 4569 // yields for foreground GC as needed).
4509 4570 if (!dirtyRegion.is_empty()) {
4510 4571 assert(numDirtyCards > 0, "consistency check");
4511 4572 HeapWord* stop_point = NULL;
4512 4573 {
4513 4574 stopTimer();
4514 4575 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4515 4576 bitMapLock());
4516 4577 startTimer();
4517 4578 verify_work_stacks_empty();
4518 4579 verify_overflow_empty();
4519 4580 sample_eden();
4520 4581 stop_point =
4521 4582 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4522 4583 }
4523 4584 if (stop_point != NULL) {
4524 4585 // The careful iteration stopped early either because it found an
4525 4586 // uninitialized object, or because we were in the midst of an
4526 4587 // "abortable preclean", which should now be aborted. Redirty
4527 4588 // the bits corresponding to the partially-scanned or unscanned
4528 4589 // cards. We'll either restart at the next block boundary or
4529 4590 // abort the preclean.
4530 4591 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) ||
4531 4592 (_collectorState == AbortablePreclean && should_abort_preclean()),
4532 4593 "Unparsable objects should only be in perm gen.");
4533 4594
4534 4595 stopTimer();
4535 4596 CMSTokenSyncWithLocks ts(true, bitMapLock());
4536 4597 startTimer();
4537 4598 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4538 4599 if (should_abort_preclean()) {
4539 4600 break; // out of preclean loop
4540 4601 } else {
4541 4602 // Compute the next address at which preclean should pick up;
4542 4603 // might need bitMapLock in order to read P-bits.
4543 4604 lastAddr = next_card_start_after_block(stop_point);
4544 4605 }
4545 4606 }
4546 4607 } else {
4547 4608 assert(lastAddr == endAddr, "consistency check");
4548 4609 assert(numDirtyCards == 0, "consistency check");
4549 4610 break;
4550 4611 }
4551 4612 }
4552 4613 verify_work_stacks_empty();
4553 4614 verify_overflow_empty();
4554 4615 return cumNumDirtyCards;
4555 4616 }
4556 4617
4557 4618 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4558 4619 // below are largely identical; if you need to modify
4559 4620 // one of these methods, please check the other method too.
4560 4621
4561 4622 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4562 4623 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4563 4624 // strategy: it's similar to precleamModUnionTable above, in that
4564 4625 // we accumulate contiguous ranges of dirty cards, mark these cards
4565 4626 // precleaned, then scan the region covered by these cards.
4566 4627 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4567 4628 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4568 4629
4569 4630 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4570 4631
4571 4632 size_t numDirtyCards, cumNumDirtyCards;
4572 4633 HeapWord *lastAddr, *nextAddr;
4573 4634
4574 4635 for (cumNumDirtyCards = numDirtyCards = 0,
4575 4636 nextAddr = lastAddr = startAddr;
4576 4637 nextAddr < endAddr;
4577 4638 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4578 4639
4579 4640 ResourceMark rm;
4580 4641 HandleMark hm;
4581 4642
4582 4643 MemRegion dirtyRegion;
4583 4644 {
4584 4645 // See comments in "Precleaning notes" above on why we
4585 4646 // do this locking. XXX Could the locking overheads be
4586 4647 // too high when dirty cards are sparse? [I don't think so.]
4587 4648 stopTimer();
4588 4649 CMSTokenSync x(true); // is cms thread
4589 4650 startTimer();
4590 4651 sample_eden();
4591 4652 // Get and clear dirty region from card table
4592 4653 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_preclean(
4593 4654 MemRegion(nextAddr, endAddr));
4594 4655 assert(dirtyRegion.start() >= nextAddr,
4595 4656 "returned region inconsistent?");
4596 4657 }
4597 4658 lastAddr = dirtyRegion.end();
4598 4659 numDirtyCards =
4599 4660 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4600 4661
4601 4662 if (!dirtyRegion.is_empty()) {
4602 4663 stopTimer();
4603 4664 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4604 4665 startTimer();
4605 4666 sample_eden();
4606 4667 verify_work_stacks_empty();
4607 4668 verify_overflow_empty();
4608 4669 HeapWord* stop_point =
4609 4670 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4610 4671 if (stop_point != NULL) {
4611 4672 // The careful iteration stopped early because it found an
4612 4673 // uninitialized object. Redirty the bits corresponding to the
4613 4674 // partially-scanned or unscanned cards, and start again at the
4614 4675 // next block boundary.
4615 4676 assert(CMSPermGenPrecleaningEnabled ||
4616 4677 (_collectorState == AbortablePreclean && should_abort_preclean()),
4617 4678 "Unparsable objects should only be in perm gen.");
4618 4679 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4619 4680 if (should_abort_preclean()) {
4620 4681 break; // out of preclean loop
4621 4682 } else {
4622 4683 // Compute the next address at which preclean should pick up.
4623 4684 lastAddr = next_card_start_after_block(stop_point);
4624 4685 }
4625 4686 }
4626 4687 } else {
4627 4688 break;
4628 4689 }
4629 4690 }
4630 4691 verify_work_stacks_empty();
4631 4692 verify_overflow_empty();
4632 4693 return cumNumDirtyCards;
4633 4694 }
4634 4695
4635 4696 void CMSCollector::checkpointRootsFinal(bool asynch,
4636 4697 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4637 4698 assert(_collectorState == FinalMarking, "incorrect state transition?");
4638 4699 check_correct_thread_executing();
4639 4700 // world is stopped at this checkpoint
4640 4701 assert(SafepointSynchronize::is_at_safepoint(),
4641 4702 "world should be stopped");
4642 4703 verify_work_stacks_empty();
4643 4704 verify_overflow_empty();
4644 4705
4645 4706 SpecializationStats::clear();
4646 4707 if (PrintGCDetails) {
4647 4708 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4648 4709 _young_gen->used() / K,
4649 4710 _young_gen->capacity() / K);
4650 4711 }
4651 4712 if (asynch) {
4652 4713 if (CMSScavengeBeforeRemark) {
4653 4714 GenCollectedHeap* gch = GenCollectedHeap::heap();
4654 4715 // Temporarily set flag to false, GCH->do_collection will
4655 4716 // expect it to be false and set to true
4656 4717 FlagSetting fl(gch->_is_gc_active, false);
4657 4718 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark",
4658 4719 PrintGCDetails && Verbose, true, gclog_or_tty);)
4659 4720 int level = _cmsGen->level() - 1;
4660 4721 if (level >= 0) {
4661 4722 gch->do_collection(true, // full (i.e. force, see below)
4662 4723 false, // !clear_all_soft_refs
4663 4724 0, // size
4664 4725 false, // is_tlab
4665 4726 level // max_level
4666 4727 );
4667 4728 }
4668 4729 }
4669 4730 FreelistLocker x(this);
4670 4731 MutexLockerEx y(bitMapLock(),
4671 4732 Mutex::_no_safepoint_check_flag);
4672 4733 assert(!init_mark_was_synchronous, "but that's impossible!");
4673 4734 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
4674 4735 } else {
4675 4736 // already have all the locks
4676 4737 checkpointRootsFinalWork(asynch, clear_all_soft_refs,
4677 4738 init_mark_was_synchronous);
4678 4739 }
4679 4740 verify_work_stacks_empty();
4680 4741 verify_overflow_empty();
4681 4742 SpecializationStats::print();
4682 4743 }
4683 4744
4684 4745 void CMSCollector::checkpointRootsFinalWork(bool asynch,
4685 4746 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4686 4747
4687 4748 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);)
4688 4749
4689 4750 assert(haveFreelistLocks(), "must have free list locks");
4690 4751 assert_lock_strong(bitMapLock());
↓ open down ↓ |
1616 lines elided |
↑ open up ↑ |
4691 4752
4692 4753 if (UseAdaptiveSizePolicy) {
4693 4754 size_policy()->checkpoint_roots_final_begin();
4694 4755 }
4695 4756
4696 4757 ResourceMark rm;
4697 4758 HandleMark hm;
4698 4759
4699 4760 GenCollectedHeap* gch = GenCollectedHeap::heap();
4700 4761
4701 - if (cms_should_unload_classes()) {
4762 + if (should_unload_classes()) {
4702 4763 CodeCache::gc_prologue();
4703 4764 }
4704 4765 assert(haveFreelistLocks(), "must have free list locks");
4705 4766 assert_lock_strong(bitMapLock());
4706 4767
4707 4768 if (!init_mark_was_synchronous) {
4708 4769 // We might assume that we need not fill TLAB's when
4709 4770 // CMSScavengeBeforeRemark is set, because we may have just done
4710 4771 // a scavenge which would have filled all TLAB's -- and besides
4711 4772 // Eden would be empty. This however may not always be the case --
4712 4773 // for instance although we asked for a scavenge, it may not have
4713 4774 // happened because of a JNI critical section. We probably need
4714 4775 // a policy for deciding whether we can in that case wait until
4715 4776 // the critical section releases and then do the remark following
4716 4777 // the scavenge, and skip it here. In the absence of that policy,
4717 4778 // or of an indication of whether the scavenge did indeed occur,
4718 4779 // we cannot rely on TLAB's having been filled and must do
4719 4780 // so here just in case a scavenge did not happen.
4720 4781 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4721 4782 // Update the saved marks which may affect the root scans.
4722 4783 gch->save_marks();
4723 4784
4724 4785 {
4725 4786 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4726 4787
4727 4788 // Note on the role of the mod union table:
4728 4789 // Since the marker in "markFromRoots" marks concurrently with
4729 4790 // mutators, it is possible for some reachable objects not to have been
4730 4791 // scanned. For instance, an only reference to an object A was
4731 4792 // placed in object B after the marker scanned B. Unless B is rescanned,
4732 4793 // A would be collected. Such updates to references in marked objects
4733 4794 // are detected via the mod union table which is the set of all cards
4734 4795 // dirtied since the first checkpoint in this GC cycle and prior to
4735 4796 // the most recent young generation GC, minus those cleaned up by the
4736 4797 // concurrent precleaning.
4737 4798 if (CMSParallelRemarkEnabled && ParallelGCThreads > 0) {
4738 4799 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty);
4739 4800 do_remark_parallel();
4740 4801 } else {
4741 4802 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4742 4803 gclog_or_tty);
4743 4804 do_remark_non_parallel();
4744 4805 }
4745 4806 }
4746 4807 } else {
4747 4808 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
4748 4809 // The initial mark was stop-world, so there's no rescanning to
4749 4810 // do; go straight on to the next step below.
4750 4811 }
↓ open down ↓ |
39 lines elided |
↑ open up ↑ |
4751 4812 verify_work_stacks_empty();
4752 4813 verify_overflow_empty();
4753 4814
4754 4815 {
4755 4816 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);)
4756 4817 refProcessingWork(asynch, clear_all_soft_refs);
4757 4818 }
4758 4819 verify_work_stacks_empty();
4759 4820 verify_overflow_empty();
4760 4821
4761 - if (cms_should_unload_classes()) {
4822 + if (should_unload_classes()) {
4762 4823 CodeCache::gc_epilogue();
4763 4824 }
4764 4825
4765 4826 // If we encountered any (marking stack / work queue) overflow
4766 4827 // events during the current CMS cycle, take appropriate
4767 4828 // remedial measures, where possible, so as to try and avoid
4768 4829 // recurrence of that condition.
4769 4830 assert(_markStack.isEmpty(), "No grey objects");
4770 4831 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4771 4832 _ser_kac_ovflw;
4772 4833 if (ser_ovflw > 0) {
4773 4834 if (PrintCMSStatistics != 0) {
4774 4835 gclog_or_tty->print_cr("Marking stack overflow (benign) "
4775 4836 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4776 4837 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4777 4838 _ser_kac_ovflw);
4778 4839 }
4779 4840 _markStack.expand();
4780 4841 _ser_pmc_remark_ovflw = 0;
4781 4842 _ser_pmc_preclean_ovflw = 0;
4782 4843 _ser_kac_ovflw = 0;
4783 4844 }
4784 4845 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4785 4846 if (PrintCMSStatistics != 0) {
4786 4847 gclog_or_tty->print_cr("Work queue overflow (benign) "
4787 4848 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4788 4849 _par_pmc_remark_ovflw, _par_kac_ovflw);
4789 4850 }
4790 4851 _par_pmc_remark_ovflw = 0;
4791 4852 _par_kac_ovflw = 0;
4792 4853 }
4793 4854 if (PrintCMSStatistics != 0) {
4794 4855 if (_markStack._hit_limit > 0) {
4795 4856 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
4796 4857 _markStack._hit_limit);
4797 4858 }
4798 4859 if (_markStack._failed_double > 0) {
4799 4860 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
4800 4861 " current capacity "SIZE_FORMAT,
4801 4862 _markStack._failed_double,
4802 4863 _markStack.capacity());
4803 4864 }
4804 4865 }
4805 4866 _markStack._hit_limit = 0;
4806 4867 _markStack._failed_double = 0;
4807 4868
4808 4869 if ((VerifyAfterGC || VerifyDuringGC) &&
4809 4870 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4810 4871 verify_after_remark();
4811 4872 }
4812 4873
4813 4874 // Change under the freelistLocks.
4814 4875 _collectorState = Sweeping;
4815 4876 // Call isAllClear() under bitMapLock
4816 4877 assert(_modUnionTable.isAllClear(), "Should be clear by end of the"
4817 4878 " final marking");
4818 4879 if (UseAdaptiveSizePolicy) {
4819 4880 size_policy()->checkpoint_roots_final_end(gch->gc_cause());
4820 4881 }
4821 4882 }
4822 4883
4823 4884 // Parallel remark task
4824 4885 class CMSParRemarkTask: public AbstractGangTask {
4825 4886 CMSCollector* _collector;
4826 4887 WorkGang* _workers;
4827 4888 int _n_workers;
4828 4889 CompactibleFreeListSpace* _cms_space;
4829 4890 CompactibleFreeListSpace* _perm_space;
4830 4891
4831 4892 // The per-thread work queues, available here for stealing.
4832 4893 OopTaskQueueSet* _task_queues;
4833 4894 ParallelTaskTerminator _term;
4834 4895
4835 4896 public:
4836 4897 CMSParRemarkTask(CMSCollector* collector,
4837 4898 CompactibleFreeListSpace* cms_space,
4838 4899 CompactibleFreeListSpace* perm_space,
4839 4900 int n_workers, WorkGang* workers,
4840 4901 OopTaskQueueSet* task_queues):
4841 4902 AbstractGangTask("Rescan roots and grey objects in parallel"),
4842 4903 _collector(collector),
4843 4904 _cms_space(cms_space), _perm_space(perm_space),
4844 4905 _n_workers(n_workers),
4845 4906 _workers(workers),
4846 4907 _task_queues(task_queues),
4847 4908 _term(workers->total_workers(), task_queues) { }
4848 4909
4849 4910 OopTaskQueueSet* task_queues() { return _task_queues; }
4850 4911
4851 4912 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4852 4913
4853 4914 ParallelTaskTerminator* terminator() { return &_term; }
4854 4915
4855 4916 void work(int i);
4856 4917
4857 4918 private:
4858 4919 // Work method in support of parallel rescan ... of young gen spaces
4859 4920 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl,
4860 4921 ContiguousSpace* space,
4861 4922 HeapWord** chunk_array, size_t chunk_top);
4862 4923
4863 4924 // ... of dirty cards in old space
4864 4925 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4865 4926 Par_MarkRefsIntoAndScanClosure* cl);
4866 4927
4867 4928 // ... work stealing for the above
4868 4929 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4869 4930 };
4870 4931
4871 4932 void CMSParRemarkTask::work(int i) {
4872 4933 elapsedTimer _timer;
4873 4934 ResourceMark rm;
4874 4935 HandleMark hm;
4875 4936
4876 4937 // ---------- rescan from roots --------------
4877 4938 _timer.start();
4878 4939 GenCollectedHeap* gch = GenCollectedHeap::heap();
4879 4940 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4880 4941 _collector->_span, _collector->ref_processor(),
4881 4942 &(_collector->_markBitMap),
4882 4943 work_queue(i), &(_collector->_revisitStack));
4883 4944
4884 4945 // Rescan young gen roots first since these are likely
4885 4946 // coarsely partitioned and may, on that account, constitute
4886 4947 // the critical path; thus, it's best to start off that
4887 4948 // work first.
4888 4949 // ---------- young gen roots --------------
4889 4950 {
4890 4951 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
4891 4952 EdenSpace* eden_space = dng->eden();
4892 4953 ContiguousSpace* from_space = dng->from();
4893 4954 ContiguousSpace* to_space = dng->to();
4894 4955
4895 4956 HeapWord** eca = _collector->_eden_chunk_array;
4896 4957 size_t ect = _collector->_eden_chunk_index;
4897 4958 HeapWord** sca = _collector->_survivor_chunk_array;
4898 4959 size_t sct = _collector->_survivor_chunk_index;
4899 4960
4900 4961 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4901 4962 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4902 4963
4903 4964 do_young_space_rescan(i, &par_mrias_cl, to_space, NULL, 0);
4904 4965 do_young_space_rescan(i, &par_mrias_cl, from_space, sca, sct);
4905 4966 do_young_space_rescan(i, &par_mrias_cl, eden_space, eca, ect);
4906 4967
4907 4968 _timer.stop();
4908 4969 if (PrintCMSStatistics != 0) {
4909 4970 gclog_or_tty->print_cr(
4910 4971 "Finished young gen rescan work in %dth thread: %3.3f sec",
4911 4972 i, _timer.seconds());
4912 4973 }
4913 4974 }
4914 4975
4915 4976 // ---------- remaining roots --------------
4916 4977 _timer.reset();
4917 4978 _timer.start();
4918 4979 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
4919 4980 false, // yg was scanned above
4920 4981 true, // collecting perm gen
4921 4982 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4922 4983 NULL, &par_mrias_cl);
4923 4984 _timer.stop();
4924 4985 if (PrintCMSStatistics != 0) {
4925 4986 gclog_or_tty->print_cr(
4926 4987 "Finished remaining root rescan work in %dth thread: %3.3f sec",
4927 4988 i, _timer.seconds());
4928 4989 }
4929 4990
4930 4991 // ---------- rescan dirty cards ------------
4931 4992 _timer.reset();
4932 4993 _timer.start();
4933 4994
4934 4995 // Do the rescan tasks for each of the two spaces
4935 4996 // (cms_space and perm_space) in turn.
4936 4997 do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl);
4937 4998 do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl);
4938 4999 _timer.stop();
4939 5000 if (PrintCMSStatistics != 0) {
4940 5001 gclog_or_tty->print_cr(
4941 5002 "Finished dirty card rescan work in %dth thread: %3.3f sec",
4942 5003 i, _timer.seconds());
4943 5004 }
4944 5005
4945 5006 // ---------- steal work from other threads ...
4946 5007 // ---------- ... and drain overflow list.
4947 5008 _timer.reset();
4948 5009 _timer.start();
4949 5010 do_work_steal(i, &par_mrias_cl, _collector->hash_seed(i));
4950 5011 _timer.stop();
4951 5012 if (PrintCMSStatistics != 0) {
4952 5013 gclog_or_tty->print_cr(
4953 5014 "Finished work stealing in %dth thread: %3.3f sec",
4954 5015 i, _timer.seconds());
4955 5016 }
4956 5017 }
4957 5018
4958 5019 void
4959 5020 CMSParRemarkTask::do_young_space_rescan(int i,
4960 5021 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space,
4961 5022 HeapWord** chunk_array, size_t chunk_top) {
4962 5023 // Until all tasks completed:
4963 5024 // . claim an unclaimed task
4964 5025 // . compute region boundaries corresponding to task claimed
4965 5026 // using chunk_array
4966 5027 // . par_oop_iterate(cl) over that region
4967 5028
4968 5029 ResourceMark rm;
4969 5030 HandleMark hm;
4970 5031
4971 5032 SequentialSubTasksDone* pst = space->par_seq_tasks();
4972 5033 assert(pst->valid(), "Uninitialized use?");
4973 5034
4974 5035 int nth_task = 0;
4975 5036 int n_tasks = pst->n_tasks();
4976 5037
4977 5038 HeapWord *start, *end;
4978 5039 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4979 5040 // We claimed task # nth_task; compute its boundaries.
4980 5041 if (chunk_top == 0) { // no samples were taken
4981 5042 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
4982 5043 start = space->bottom();
4983 5044 end = space->top();
4984 5045 } else if (nth_task == 0) {
4985 5046 start = space->bottom();
4986 5047 end = chunk_array[nth_task];
4987 5048 } else if (nth_task < (jint)chunk_top) {
4988 5049 assert(nth_task >= 1, "Control point invariant");
4989 5050 start = chunk_array[nth_task - 1];
4990 5051 end = chunk_array[nth_task];
4991 5052 } else {
4992 5053 assert(nth_task == (jint)chunk_top, "Control point invariant");
4993 5054 start = chunk_array[chunk_top - 1];
4994 5055 end = space->top();
4995 5056 }
4996 5057 MemRegion mr(start, end);
4997 5058 // Verify that mr is in space
4998 5059 assert(mr.is_empty() || space->used_region().contains(mr),
4999 5060 "Should be in space");
5000 5061 // Verify that "start" is an object boundary
5001 5062 assert(mr.is_empty() || oop(mr.start())->is_oop(),
5002 5063 "Should be an oop");
5003 5064 space->par_oop_iterate(mr, cl);
5004 5065 }
5005 5066 pst->all_tasks_completed();
5006 5067 }
5007 5068
5008 5069 void
5009 5070 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5010 5071 CompactibleFreeListSpace* sp, int i,
5011 5072 Par_MarkRefsIntoAndScanClosure* cl) {
5012 5073 // Until all tasks completed:
5013 5074 // . claim an unclaimed task
5014 5075 // . compute region boundaries corresponding to task claimed
5015 5076 // . transfer dirty bits ct->mut for that region
5016 5077 // . apply rescanclosure to dirty mut bits for that region
5017 5078
5018 5079 ResourceMark rm;
5019 5080 HandleMark hm;
5020 5081
5021 5082 OopTaskQueue* work_q = work_queue(i);
5022 5083 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5023 5084 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5024 5085 // CAUTION: This closure has state that persists across calls to
5025 5086 // the work method dirty_range_iterate_clear() in that it has
5026 5087 // imbedded in it a (subtype of) UpwardsObjectClosure. The
5027 5088 // use of that state in the imbedded UpwardsObjectClosure instance
5028 5089 // assumes that the cards are always iterated (even if in parallel
5029 5090 // by several threads) in monotonically increasing order per each
5030 5091 // thread. This is true of the implementation below which picks
5031 5092 // card ranges (chunks) in monotonically increasing order globally
5032 5093 // and, a-fortiori, in monotonically increasing order per thread
5033 5094 // (the latter order being a subsequence of the former).
5034 5095 // If the work code below is ever reorganized into a more chaotic
5035 5096 // work-partitioning form than the current "sequential tasks"
5036 5097 // paradigm, the use of that persistent state will have to be
5037 5098 // revisited and modified appropriately. See also related
5038 5099 // bug 4756801 work on which should examine this code to make
5039 5100 // sure that the changes there do not run counter to the
5040 5101 // assumptions made here and necessary for correctness and
5041 5102 // efficiency. Note also that this code might yield inefficient
5042 5103 // behaviour in the case of very large objects that span one or
5043 5104 // more work chunks. Such objects would potentially be scanned
5044 5105 // several times redundantly. Work on 4756801 should try and
5045 5106 // address that performance anomaly if at all possible. XXX
5046 5107 MemRegion full_span = _collector->_span;
5047 5108 CMSBitMap* bm = &(_collector->_markBitMap); // shared
5048 5109 CMSMarkStack* rs = &(_collector->_revisitStack); // shared
5049 5110 MarkFromDirtyCardsClosure
5050 5111 greyRescanClosure(_collector, full_span, // entire span of interest
5051 5112 sp, bm, work_q, rs, cl);
5052 5113
5053 5114 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5054 5115 assert(pst->valid(), "Uninitialized use?");
5055 5116 int nth_task = 0;
5056 5117 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5057 5118 MemRegion span = sp->used_region();
5058 5119 HeapWord* start_addr = span.start();
5059 5120 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5060 5121 alignment);
5061 5122 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5062 5123 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5063 5124 start_addr, "Check alignment");
5064 5125 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5065 5126 chunk_size, "Check alignment");
5066 5127
5067 5128 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5068 5129 // Having claimed the nth_task, compute corresponding mem-region,
5069 5130 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
5070 5131 // The alignment restriction ensures that we do not need any
5071 5132 // synchronization with other gang-workers while setting or
5072 5133 // clearing bits in thus chunk of the MUT.
5073 5134 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5074 5135 start_addr + (nth_task+1)*chunk_size);
5075 5136 // The last chunk's end might be way beyond end of the
5076 5137 // used region. In that case pull back appropriately.
5077 5138 if (this_span.end() > end_addr) {
5078 5139 this_span.set_end(end_addr);
5079 5140 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5080 5141 }
5081 5142 // Iterate over the dirty cards covering this chunk, marking them
5082 5143 // precleaned, and setting the corresponding bits in the mod union
5083 5144 // table. Since we have been careful to partition at Card and MUT-word
5084 5145 // boundaries no synchronization is needed between parallel threads.
5085 5146 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5086 5147 &modUnionClosure);
5087 5148
5088 5149 // Having transferred these marks into the modUnionTable,
5089 5150 // rescan the marked objects on the dirty cards in the modUnionTable.
5090 5151 // Even if this is at a synchronous collection, the initial marking
5091 5152 // may have been done during an asynchronous collection so there
5092 5153 // may be dirty bits in the mod-union table.
5093 5154 _collector->_modUnionTable.dirty_range_iterate_clear(
5094 5155 this_span, &greyRescanClosure);
5095 5156 _collector->_modUnionTable.verifyNoOneBitsInRange(
5096 5157 this_span.start(),
5097 5158 this_span.end());
5098 5159 }
5099 5160 pst->all_tasks_completed(); // declare that i am done
5100 5161 }
5101 5162
5102 5163 // . see if we can share work_queues with ParNew? XXX
5103 5164 void
5104 5165 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5105 5166 int* seed) {
5106 5167 OopTaskQueue* work_q = work_queue(i);
5107 5168 NOT_PRODUCT(int num_steals = 0;)
5108 5169 oop obj_to_scan;
5109 5170 CMSBitMap* bm = &(_collector->_markBitMap);
5110 5171 size_t num_from_overflow_list =
5111 5172 MIN2((size_t)work_q->max_elems()/4,
5112 5173 (size_t)ParGCDesiredObjsFromOverflowList);
5113 5174
5114 5175 while (true) {
5115 5176 // Completely finish any left over work from (an) earlier round(s)
5116 5177 cl->trim_queue(0);
5117 5178 // Now check if there's any work in the overflow list
5118 5179 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5119 5180 work_q)) {
5120 5181 // found something in global overflow list;
5121 5182 // not yet ready to go stealing work from others.
5122 5183 // We'd like to assert(work_q->size() != 0, ...)
5123 5184 // because we just took work from the overflow list,
5124 5185 // but of course we can't since all of that could have
5125 5186 // been already stolen from us.
5126 5187 // "He giveth and He taketh away."
5127 5188 continue;
5128 5189 }
5129 5190 // Verify that we have no work before we resort to stealing
5130 5191 assert(work_q->size() == 0, "Have work, shouldn't steal");
5131 5192 // Try to steal from other queues that have work
5132 5193 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5133 5194 NOT_PRODUCT(num_steals++;)
5134 5195 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5135 5196 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5136 5197 // Do scanning work
5137 5198 obj_to_scan->oop_iterate(cl);
5138 5199 // Loop around, finish this work, and try to steal some more
5139 5200 } else if (terminator()->offer_termination()) {
5140 5201 break; // nirvana from the infinite cycle
5141 5202 }
5142 5203 }
5143 5204 NOT_PRODUCT(
5144 5205 if (PrintCMSStatistics != 0) {
5145 5206 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5146 5207 }
5147 5208 )
5148 5209 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5149 5210 "Else our work is not yet done");
5150 5211 }
5151 5212
5152 5213 // Return a thread-local PLAB recording array, as appropriate.
5153 5214 void* CMSCollector::get_data_recorder(int thr_num) {
5154 5215 if (_survivor_plab_array != NULL &&
5155 5216 (CMSPLABRecordAlways ||
5156 5217 (_collectorState > Marking && _collectorState < FinalMarking))) {
5157 5218 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5158 5219 ChunkArray* ca = &_survivor_plab_array[thr_num];
5159 5220 ca->reset(); // clear it so that fresh data is recorded
5160 5221 return (void*) ca;
5161 5222 } else {
5162 5223 return NULL;
5163 5224 }
5164 5225 }
5165 5226
5166 5227 // Reset all the thread-local PLAB recording arrays
5167 5228 void CMSCollector::reset_survivor_plab_arrays() {
5168 5229 for (uint i = 0; i < ParallelGCThreads; i++) {
5169 5230 _survivor_plab_array[i].reset();
5170 5231 }
5171 5232 }
5172 5233
5173 5234 // Merge the per-thread plab arrays into the global survivor chunk
5174 5235 // array which will provide the partitioning of the survivor space
5175 5236 // for CMS rescan.
5176 5237 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv) {
5177 5238 assert(_survivor_plab_array != NULL, "Error");
5178 5239 assert(_survivor_chunk_array != NULL, "Error");
5179 5240 assert(_collectorState == FinalMarking, "Error");
5180 5241 for (uint j = 0; j < ParallelGCThreads; j++) {
5181 5242 _cursor[j] = 0;
5182 5243 }
5183 5244 HeapWord* top = surv->top();
5184 5245 size_t i;
5185 5246 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
5186 5247 HeapWord* min_val = top; // Higher than any PLAB address
5187 5248 uint min_tid = 0; // position of min_val this round
5188 5249 for (uint j = 0; j < ParallelGCThreads; j++) {
5189 5250 ChunkArray* cur_sca = &_survivor_plab_array[j];
5190 5251 if (_cursor[j] == cur_sca->end()) {
5191 5252 continue;
5192 5253 }
5193 5254 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5194 5255 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5195 5256 assert(surv->used_region().contains(cur_val), "Out of bounds value");
5196 5257 if (cur_val < min_val) {
5197 5258 min_tid = j;
5198 5259 min_val = cur_val;
5199 5260 } else {
5200 5261 assert(cur_val < top, "All recorded addresses should be less");
5201 5262 }
5202 5263 }
5203 5264 // At this point min_val and min_tid are respectively
5204 5265 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5205 5266 // and the thread (j) that witnesses that address.
5206 5267 // We record this address in the _survivor_chunk_array[i]
5207 5268 // and increment _cursor[min_tid] prior to the next round i.
5208 5269 if (min_val == top) {
5209 5270 break;
5210 5271 }
5211 5272 _survivor_chunk_array[i] = min_val;
5212 5273 _cursor[min_tid]++;
5213 5274 }
5214 5275 // We are all done; record the size of the _survivor_chunk_array
5215 5276 _survivor_chunk_index = i; // exclusive: [0, i)
5216 5277 if (PrintCMSStatistics > 0) {
5217 5278 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5218 5279 }
5219 5280 // Verify that we used up all the recorded entries
5220 5281 #ifdef ASSERT
5221 5282 size_t total = 0;
5222 5283 for (uint j = 0; j < ParallelGCThreads; j++) {
5223 5284 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5224 5285 total += _cursor[j];
5225 5286 }
5226 5287 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5227 5288 // Check that the merged array is in sorted order
5228 5289 if (total > 0) {
5229 5290 for (size_t i = 0; i < total - 1; i++) {
5230 5291 if (PrintCMSStatistics > 0) {
5231 5292 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5232 5293 i, _survivor_chunk_array[i]);
5233 5294 }
5234 5295 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5235 5296 "Not sorted");
5236 5297 }
5237 5298 }
5238 5299 #endif // ASSERT
5239 5300 }
5240 5301
5241 5302 // Set up the space's par_seq_tasks structure for work claiming
5242 5303 // for parallel rescan of young gen.
5243 5304 // See ParRescanTask where this is currently used.
5244 5305 void
5245 5306 CMSCollector::
5246 5307 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5247 5308 assert(n_threads > 0, "Unexpected n_threads argument");
5248 5309 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5249 5310
5250 5311 // Eden space
5251 5312 {
5252 5313 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5253 5314 assert(!pst->valid(), "Clobbering existing data?");
5254 5315 // Each valid entry in [0, _eden_chunk_index) represents a task.
5255 5316 size_t n_tasks = _eden_chunk_index + 1;
5256 5317 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5257 5318 pst->set_par_threads(n_threads);
5258 5319 pst->set_n_tasks((int)n_tasks);
5259 5320 }
5260 5321
5261 5322 // Merge the survivor plab arrays into _survivor_chunk_array
5262 5323 if (_survivor_plab_array != NULL) {
5263 5324 merge_survivor_plab_arrays(dng->from());
5264 5325 } else {
5265 5326 assert(_survivor_chunk_index == 0, "Error");
5266 5327 }
5267 5328
5268 5329 // To space
5269 5330 {
5270 5331 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5271 5332 assert(!pst->valid(), "Clobbering existing data?");
5272 5333 pst->set_par_threads(n_threads);
5273 5334 pst->set_n_tasks(1);
5274 5335 assert(pst->valid(), "Error");
5275 5336 }
5276 5337
5277 5338 // From space
5278 5339 {
5279 5340 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5280 5341 assert(!pst->valid(), "Clobbering existing data?");
5281 5342 size_t n_tasks = _survivor_chunk_index + 1;
5282 5343 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5283 5344 pst->set_par_threads(n_threads);
5284 5345 pst->set_n_tasks((int)n_tasks);
5285 5346 assert(pst->valid(), "Error");
5286 5347 }
5287 5348 }
5288 5349
5289 5350 // Parallel version of remark
5290 5351 void CMSCollector::do_remark_parallel() {
5291 5352 GenCollectedHeap* gch = GenCollectedHeap::heap();
5292 5353 WorkGang* workers = gch->workers();
5293 5354 assert(workers != NULL, "Need parallel worker threads.");
5294 5355 int n_workers = workers->total_workers();
5295 5356 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5296 5357 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
5297 5358
5298 5359 CMSParRemarkTask tsk(this,
5299 5360 cms_space, perm_space,
5300 5361 n_workers, workers, task_queues());
5301 5362
5302 5363 // Set up for parallel process_strong_roots work.
5303 5364 gch->set_par_threads(n_workers);
5304 5365 gch->change_strong_roots_parity();
5305 5366 // We won't be iterating over the cards in the card table updating
5306 5367 // the younger_gen cards, so we shouldn't call the following else
5307 5368 // the verification code as well as subsequent younger_refs_iterate
5308 5369 // code would get confused. XXX
5309 5370 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5310 5371
5311 5372 // The young gen rescan work will not be done as part of
5312 5373 // process_strong_roots (which currently doesn't knw how to
5313 5374 // parallelize such a scan), but rather will be broken up into
5314 5375 // a set of parallel tasks (via the sampling that the [abortable]
5315 5376 // preclean phase did of EdenSpace, plus the [two] tasks of
5316 5377 // scanning the [two] survivor spaces. Further fine-grain
5317 5378 // parallelization of the scanning of the survivor spaces
5318 5379 // themselves, and of precleaning of the younger gen itself
5319 5380 // is deferred to the future.
5320 5381 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5321 5382
5322 5383 // The dirty card rescan work is broken up into a "sequence"
5323 5384 // of parallel tasks (per constituent space) that are dynamically
5324 5385 // claimed by the parallel threads.
5325 5386 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5326 5387 perm_space->initialize_sequential_subtasks_for_rescan(n_workers);
5327 5388
5328 5389 // It turns out that even when we're using 1 thread, doing the work in a
5329 5390 // separate thread causes wide variance in run times. We can't help this
5330 5391 // in the multi-threaded case, but we special-case n=1 here to get
5331 5392 // repeatable measurements of the 1-thread overhead of the parallel code.
5332 5393 if (n_workers > 1) {
5333 5394 // Make refs discovery MT-safe
5334 5395 ReferenceProcessorMTMutator mt(ref_processor(), true);
5335 5396 workers->run_task(&tsk);
5336 5397 } else {
5337 5398 tsk.work(0);
5338 5399 }
5339 5400 gch->set_par_threads(0); // 0 ==> non-parallel.
5340 5401 // restore, single-threaded for now, any preserved marks
5341 5402 // as a result of work_q overflow
5342 5403 restore_preserved_marks_if_any();
5343 5404 }
5344 5405
5345 5406 // Non-parallel version of remark
5346 5407 void CMSCollector::do_remark_non_parallel() {
5347 5408 ResourceMark rm;
5348 5409 HandleMark hm;
5349 5410 GenCollectedHeap* gch = GenCollectedHeap::heap();
5350 5411 MarkRefsIntoAndScanClosure
5351 5412 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
5352 5413 &_markStack, &_revisitStack, this,
5353 5414 false /* should_yield */, false /* not precleaning */);
5354 5415 MarkFromDirtyCardsClosure
5355 5416 markFromDirtyCardsClosure(this, _span,
5356 5417 NULL, // space is set further below
5357 5418 &_markBitMap, &_markStack, &_revisitStack,
5358 5419 &mrias_cl);
5359 5420 {
5360 5421 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty);
5361 5422 // Iterate over the dirty cards, marking them precleaned, and
5362 5423 // setting the corresponding bits in the mod union table.
5363 5424 {
5364 5425 ModUnionClosure modUnionClosure(&_modUnionTable);
5365 5426 _ct->ct_bs()->dirty_card_iterate(
5366 5427 _cmsGen->used_region(),
5367 5428 &modUnionClosure);
5368 5429 _ct->ct_bs()->dirty_card_iterate(
5369 5430 _permGen->used_region(),
5370 5431 &modUnionClosure);
5371 5432 }
5372 5433 // Having transferred these marks into the modUnionTable, we just need
5373 5434 // to rescan the marked objects on the dirty cards in the modUnionTable.
5374 5435 // The initial marking may have been done during an asynchronous
5375 5436 // collection so there may be dirty bits in the mod-union table.
5376 5437 const int alignment =
5377 5438 CardTableModRefBS::card_size * BitsPerWord;
5378 5439 {
5379 5440 // ... First handle dirty cards in CMS gen
5380 5441 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5381 5442 MemRegion ur = _cmsGen->used_region();
5382 5443 HeapWord* lb = ur.start();
5383 5444 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5384 5445 MemRegion cms_span(lb, ub);
5385 5446 _modUnionTable.dirty_range_iterate_clear(cms_span,
5386 5447 &markFromDirtyCardsClosure);
5387 5448 verify_work_stacks_empty();
5388 5449 if (PrintCMSStatistics != 0) {
5389 5450 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5390 5451 markFromDirtyCardsClosure.num_dirty_cards());
5391 5452 }
5392 5453 }
5393 5454 {
5394 5455 // .. and then repeat for dirty cards in perm gen
5395 5456 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace());
5396 5457 MemRegion ur = _permGen->used_region();
5397 5458 HeapWord* lb = ur.start();
5398 5459 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5399 5460 MemRegion perm_span(lb, ub);
5400 5461 _modUnionTable.dirty_range_iterate_clear(perm_span,
5401 5462 &markFromDirtyCardsClosure);
5402 5463 verify_work_stacks_empty();
5403 5464 if (PrintCMSStatistics != 0) {
5404 5465 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ",
5405 5466 markFromDirtyCardsClosure.num_dirty_cards());
5406 5467 }
5407 5468 }
5408 5469 }
5409 5470 if (VerifyDuringGC &&
5410 5471 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5411 5472 HandleMark hm; // Discard invalid handles created during verification
5412 5473 Universe::verify(true);
5413 5474 }
5414 5475 {
5415 5476 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty);
5416 5477
5417 5478 verify_work_stacks_empty();
5418 5479
5419 5480 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5420 5481 gch->gen_process_strong_roots(_cmsGen->level(),
5421 5482 true, // younger gens as roots
5422 5483 true, // collecting perm gen
5423 5484 SharedHeap::ScanningOption(roots_scanning_options()),
5424 5485 NULL, &mrias_cl);
5425 5486 }
5426 5487 verify_work_stacks_empty();
5427 5488 // Restore evacuated mark words, if any, used for overflow list links
5428 5489 if (!CMSOverflowEarlyRestoration) {
5429 5490 restore_preserved_marks_if_any();
5430 5491 }
5431 5492 verify_overflow_empty();
5432 5493 }
5433 5494
5434 5495 ////////////////////////////////////////////////////////
5435 5496 // Parallel Reference Processing Task Proxy Class
5436 5497 ////////////////////////////////////////////////////////
5437 5498 class CMSRefProcTaskProxy: public AbstractGangTask {
5438 5499 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5439 5500 CMSCollector* _collector;
5440 5501 CMSBitMap* _mark_bit_map;
5441 5502 MemRegion _span;
5442 5503 OopTaskQueueSet* _task_queues;
5443 5504 ParallelTaskTerminator _term;
5444 5505 ProcessTask& _task;
5445 5506
5446 5507 public:
5447 5508 CMSRefProcTaskProxy(ProcessTask& task,
5448 5509 CMSCollector* collector,
5449 5510 const MemRegion& span,
5450 5511 CMSBitMap* mark_bit_map,
5451 5512 int total_workers,
5452 5513 OopTaskQueueSet* task_queues):
5453 5514 AbstractGangTask("Process referents by policy in parallel"),
5454 5515 _task(task),
5455 5516 _collector(collector), _span(span), _mark_bit_map(mark_bit_map),
5456 5517 _task_queues(task_queues),
5457 5518 _term(total_workers, task_queues)
5458 5519 { }
5459 5520
5460 5521 OopTaskQueueSet* task_queues() { return _task_queues; }
5461 5522
5462 5523 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5463 5524
5464 5525 ParallelTaskTerminator* terminator() { return &_term; }
5465 5526
5466 5527 void do_work_steal(int i,
5467 5528 CMSParDrainMarkingStackClosure* drain,
5468 5529 CMSParKeepAliveClosure* keep_alive,
5469 5530 int* seed);
5470 5531
5471 5532 virtual void work(int i);
5472 5533 };
5473 5534
5474 5535 void CMSRefProcTaskProxy::work(int i) {
5475 5536 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5476 5537 _mark_bit_map, work_queue(i));
5477 5538 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5478 5539 _mark_bit_map, work_queue(i));
5479 5540 CMSIsAliveClosure is_alive_closure(_mark_bit_map);
5480 5541 _task.work(i, is_alive_closure, par_keep_alive, par_drain_stack);
5481 5542 if (_task.marks_oops_alive()) {
5482 5543 do_work_steal(i, &par_drain_stack, &par_keep_alive,
5483 5544 _collector->hash_seed(i));
5484 5545 }
5485 5546 assert(work_queue(i)->size() == 0, "work_queue should be empty");
5486 5547 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5487 5548 }
5488 5549
5489 5550 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5490 5551 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5491 5552 EnqueueTask& _task;
5492 5553
5493 5554 public:
5494 5555 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5495 5556 : AbstractGangTask("Enqueue reference objects in parallel"),
5496 5557 _task(task)
5497 5558 { }
5498 5559
5499 5560 virtual void work(int i)
5500 5561 {
5501 5562 _task.work(i);
5502 5563 }
5503 5564 };
5504 5565
5505 5566 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5506 5567 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5507 5568 _collector(collector),
5508 5569 _span(span),
5509 5570 _bit_map(bit_map),
5510 5571 _work_queue(work_queue),
5511 5572 _mark_and_push(collector, span, bit_map, work_queue),
5512 5573 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5513 5574 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5514 5575 { }
5515 5576
5516 5577 // . see if we can share work_queues with ParNew? XXX
5517 5578 void CMSRefProcTaskProxy::do_work_steal(int i,
5518 5579 CMSParDrainMarkingStackClosure* drain,
5519 5580 CMSParKeepAliveClosure* keep_alive,
5520 5581 int* seed) {
5521 5582 OopTaskQueue* work_q = work_queue(i);
5522 5583 NOT_PRODUCT(int num_steals = 0;)
5523 5584 oop obj_to_scan;
5524 5585 size_t num_from_overflow_list =
5525 5586 MIN2((size_t)work_q->max_elems()/4,
5526 5587 (size_t)ParGCDesiredObjsFromOverflowList);
5527 5588
5528 5589 while (true) {
5529 5590 // Completely finish any left over work from (an) earlier round(s)
5530 5591 drain->trim_queue(0);
5531 5592 // Now check if there's any work in the overflow list
5532 5593 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5533 5594 work_q)) {
5534 5595 // Found something in global overflow list;
5535 5596 // not yet ready to go stealing work from others.
5536 5597 // We'd like to assert(work_q->size() != 0, ...)
5537 5598 // because we just took work from the overflow list,
5538 5599 // but of course we can't, since all of that might have
5539 5600 // been already stolen from us.
5540 5601 continue;
5541 5602 }
5542 5603 // Verify that we have no work before we resort to stealing
5543 5604 assert(work_q->size() == 0, "Have work, shouldn't steal");
5544 5605 // Try to steal from other queues that have work
5545 5606 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5546 5607 NOT_PRODUCT(num_steals++;)
5547 5608 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5548 5609 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5549 5610 // Do scanning work
5550 5611 obj_to_scan->oop_iterate(keep_alive);
5551 5612 // Loop around, finish this work, and try to steal some more
5552 5613 } else if (terminator()->offer_termination()) {
5553 5614 break; // nirvana from the infinite cycle
5554 5615 }
5555 5616 }
5556 5617 NOT_PRODUCT(
5557 5618 if (PrintCMSStatistics != 0) {
5558 5619 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5559 5620 }
5560 5621 )
5561 5622 }
5562 5623
5563 5624 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5564 5625 {
5565 5626 GenCollectedHeap* gch = GenCollectedHeap::heap();
5566 5627 WorkGang* workers = gch->workers();
5567 5628 assert(workers != NULL, "Need parallel worker threads.");
5568 5629 int n_workers = workers->total_workers();
5569 5630 CMSRefProcTaskProxy rp_task(task, &_collector,
5570 5631 _collector.ref_processor()->span(),
5571 5632 _collector.markBitMap(),
5572 5633 n_workers, _collector.task_queues());
5573 5634 workers->run_task(&rp_task);
5574 5635 }
5575 5636
5576 5637 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5577 5638 {
5578 5639
5579 5640 GenCollectedHeap* gch = GenCollectedHeap::heap();
5580 5641 WorkGang* workers = gch->workers();
5581 5642 assert(workers != NULL, "Need parallel worker threads.");
5582 5643 CMSRefEnqueueTaskProxy enq_task(task);
5583 5644 workers->run_task(&enq_task);
5584 5645 }
5585 5646
5586 5647 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
5587 5648
5588 5649 ResourceMark rm;
5589 5650 HandleMark hm;
5590 5651 ReferencePolicy* soft_ref_policy;
5591 5652
5592 5653 assert(!ref_processor()->enqueuing_is_done(), "Enqueuing should not be complete");
5593 5654 // Process weak references.
5594 5655 if (clear_all_soft_refs) {
5595 5656 soft_ref_policy = new AlwaysClearPolicy();
5596 5657 } else {
5597 5658 #ifdef COMPILER2
5598 5659 soft_ref_policy = new LRUMaxHeapPolicy();
5599 5660 #else
5600 5661 soft_ref_policy = new LRUCurrentHeapPolicy();
5601 5662 #endif // COMPILER2
5602 5663 }
5603 5664 verify_work_stacks_empty();
5604 5665
5605 5666 ReferenceProcessor* rp = ref_processor();
5606 5667 assert(rp->span().equals(_span), "Spans should be equal");
5607 5668 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5608 5669 &_markStack);
5609 5670 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5610 5671 _span, &_markBitMap, &_markStack,
5611 5672 &cmsKeepAliveClosure);
5612 5673 {
5613 5674 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty);
5614 5675 if (rp->processing_is_mt()) {
5615 5676 CMSRefProcTaskExecutor task_executor(*this);
5616 5677 rp->process_discovered_references(soft_ref_policy,
5617 5678 &_is_alive_closure,
5618 5679 &cmsKeepAliveClosure,
5619 5680 &cmsDrainMarkingStackClosure,
5620 5681 &task_executor);
↓ open down ↓ |
849 lines elided |
↑ open up ↑ |
5621 5682 } else {
5622 5683 rp->process_discovered_references(soft_ref_policy,
5623 5684 &_is_alive_closure,
5624 5685 &cmsKeepAliveClosure,
5625 5686 &cmsDrainMarkingStackClosure,
5626 5687 NULL);
5627 5688 }
5628 5689 verify_work_stacks_empty();
5629 5690 }
5630 5691
5631 - if (cms_should_unload_classes()) {
5692 + if (should_unload_classes()) {
5632 5693 {
5633 5694 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty);
5634 5695
5635 5696 // Follow SystemDictionary roots and unload classes
5636 5697 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5637 5698
5638 5699 // Follow CodeCache roots and unload any methods marked for unloading
5639 5700 CodeCache::do_unloading(&_is_alive_closure,
5640 5701 &cmsKeepAliveClosure,
5641 5702 purged_class);
5642 5703
5643 5704 cmsDrainMarkingStackClosure.do_void();
5644 5705 verify_work_stacks_empty();
5645 5706
5646 5707 // Update subklass/sibling/implementor links in KlassKlass descendants
5647 5708 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty");
5648 5709 oop k;
5649 5710 while ((k = _revisitStack.pop()) != NULL) {
5650 5711 ((Klass*)(oopDesc*)k)->follow_weak_klass_links(
5651 5712 &_is_alive_closure,
5652 5713 &cmsKeepAliveClosure);
5653 5714 }
5654 5715 assert(!ClassUnloading ||
5655 5716 (_markStack.isEmpty() && overflow_list_is_empty()),
5656 5717 "Should not have found new reachable objects");
5657 5718 assert(_revisitStack.isEmpty(), "revisit stack should have been drained");
5658 5719 cmsDrainMarkingStackClosure.do_void();
5659 5720 verify_work_stacks_empty();
5660 5721 }
5661 5722
5662 5723 {
5663 5724 TraceTime t("scrub symbol & string tables", PrintGCDetails, false, gclog_or_tty);
5664 5725 // Now clean up stale oops in SymbolTable and StringTable
5665 5726 SymbolTable::unlink(&_is_alive_closure);
5666 5727 StringTable::unlink(&_is_alive_closure);
5667 5728 }
5668 5729 }
5669 5730
5670 5731 verify_work_stacks_empty();
5671 5732 // Restore any preserved marks as a result of mark stack or
5672 5733 // work queue overflow
5673 5734 restore_preserved_marks_if_any(); // done single-threaded for now
5674 5735
5675 5736 rp->set_enqueuing_is_done(true);
5676 5737 if (rp->processing_is_mt()) {
5677 5738 CMSRefProcTaskExecutor task_executor(*this);
5678 5739 rp->enqueue_discovered_references(&task_executor);
5679 5740 } else {
5680 5741 rp->enqueue_discovered_references(NULL);
5681 5742 }
5682 5743 rp->verify_no_references_recorded();
5683 5744 assert(!rp->discovery_enabled(), "should have been disabled");
5684 5745
5685 5746 // JVMTI object tagging is based on JNI weak refs. If any of these
5686 5747 // refs were cleared then JVMTI needs to update its maps and
5687 5748 // maybe post ObjectFrees to agents.
5688 5749 JvmtiExport::cms_ref_processing_epilogue();
5689 5750 }
5690 5751
5691 5752 #ifndef PRODUCT
5692 5753 void CMSCollector::check_correct_thread_executing() {
5693 5754 Thread* t = Thread::current();
5694 5755 // Only the VM thread or the CMS thread should be here.
5695 5756 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5696 5757 "Unexpected thread type");
5697 5758 // If this is the vm thread, the foreground process
5698 5759 // should not be waiting. Note that _foregroundGCIsActive is
5699 5760 // true while the foreground collector is waiting.
5700 5761 if (_foregroundGCShouldWait) {
5701 5762 // We cannot be the VM thread
5702 5763 assert(t->is_ConcurrentGC_thread(),
5703 5764 "Should be CMS thread");
5704 5765 } else {
5705 5766 // We can be the CMS thread only if we are in a stop-world
5706 5767 // phase of CMS collection.
5707 5768 if (t->is_ConcurrentGC_thread()) {
5708 5769 assert(_collectorState == InitialMarking ||
5709 5770 _collectorState == FinalMarking,
5710 5771 "Should be a stop-world phase");
5711 5772 // The CMS thread should be holding the CMS_token.
5712 5773 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5713 5774 "Potential interference with concurrently "
5714 5775 "executing VM thread");
↓ open down ↓ |
73 lines elided |
↑ open up ↑ |
5715 5776 }
5716 5777 }
5717 5778 }
5718 5779 #endif
5719 5780
5720 5781 void CMSCollector::sweep(bool asynch) {
5721 5782 assert(_collectorState == Sweeping, "just checking");
5722 5783 check_correct_thread_executing();
5723 5784 verify_work_stacks_empty();
5724 5785 verify_overflow_empty();
5725 - incrementSweepCount();
5786 + increment_sweep_count();
5726 5787 _sweep_timer.stop();
5727 5788 _sweep_estimate.sample(_sweep_timer.seconds());
5728 5789 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
5729 5790
5730 5791 // PermGen verification support: If perm gen sweeping is disabled in
5731 5792 // this cycle, we preserve the perm gen object "deadness" information
5732 5793 // in the perm_gen_verify_bit_map. In order to do that we traverse
5733 5794 // 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());
5795 + if (verifying() && !should_unload_classes()) {
5737 5796 assert(perm_gen_verify_bit_map()->sizeInBits() != 0,
5738 5797 "Should have already been allocated");
5739 5798 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(),
5740 5799 markBitMap(), perm_gen_verify_bit_map());
5741 - _permGen->cmsSpace()->blk_iterate(&mdo);
5800 + if (asynch) {
5801 + CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
5802 + bitMapLock());
5803 + _permGen->cmsSpace()->blk_iterate(&mdo);
5804 + } else {
5805 + // In the case of synchronous sweep, we already have
5806 + // the requisite locks/tokens.
5807 + _permGen->cmsSpace()->blk_iterate(&mdo);
5808 + }
5742 5809 }
5743 5810
5744 5811 if (asynch) {
5745 5812 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5746 5813 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
5747 5814 // First sweep the old gen then the perm gen
5748 5815 {
5749 5816 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5750 5817 bitMapLock());
5751 5818 sweepWork(_cmsGen, asynch);
5752 5819 }
5753 5820
5754 5821 // Now repeat for perm gen
5755 - if (cms_should_unload_classes()) {
5822 + if (should_unload_classes()) {
5756 5823 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
5757 5824 bitMapLock());
5758 5825 sweepWork(_permGen, asynch);
5759 5826 }
5760 5827
5761 5828 // Update Universe::_heap_*_at_gc figures.
5762 5829 // We need all the free list locks to make the abstract state
5763 5830 // transition from Sweeping to Resetting. See detailed note
5764 5831 // further below.
5765 5832 {
5766 5833 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
↓ open down ↓ |
1 lines elided |
↑ open up ↑ |
5767 5834 _permGen->freelistLock());
5768 5835 // Update heap occupancy information which is used as
5769 5836 // input to soft ref clearing policy at the next gc.
5770 5837 Universe::update_heap_info_at_gc();
5771 5838 _collectorState = Resizing;
5772 5839 }
5773 5840 } else {
5774 5841 // already have needed locks
5775 5842 sweepWork(_cmsGen, asynch);
5776 5843
5777 - if (cms_should_unload_classes()) {
5844 + if (should_unload_classes()) {
5778 5845 sweepWork(_permGen, asynch);
5779 5846 }
5780 5847 // Update heap occupancy information which is used as
5781 5848 // input to soft ref clearing policy at the next gc.
5782 5849 Universe::update_heap_info_at_gc();
5783 5850 _collectorState = Resizing;
5784 5851 }
5785 5852 verify_work_stacks_empty();
5786 5853 verify_overflow_empty();
5787 5854
5788 5855 _sweep_timer.reset();
5789 5856 _sweep_timer.start();
5790 5857
5791 5858 update_time_of_last_gc(os::javaTimeMillis());
5792 5859
5793 5860 // NOTE on abstract state transitions:
5794 5861 // Mutators allocate-live and/or mark the mod-union table dirty
5795 5862 // based on the state of the collection. The former is done in
5796 5863 // the interval [Marking, Sweeping] and the latter in the interval
5797 5864 // [Marking, Sweeping). Thus the transitions into the Marking state
5798 5865 // and out of the Sweeping state must be synchronously visible
5799 5866 // globally to the mutators.
5800 5867 // The transition into the Marking state happens with the world
5801 5868 // stopped so the mutators will globally see it. Sweeping is
5802 5869 // done asynchronously by the background collector so the transition
5803 5870 // from the Sweeping state to the Resizing state must be done
5804 5871 // under the freelistLock (as is the check for whether to
5805 5872 // allocate-live and whether to dirty the mod-union table).
5806 5873 assert(_collectorState == Resizing, "Change of collector state to"
5807 5874 " Resizing must be done under the freelistLocks (plural)");
5808 5875
5809 5876 // Now that sweeping has been completed, if the GCH's
5810 5877 // incremental_collection_will_fail flag is set, clear it,
5811 5878 // thus inviting a younger gen collection to promote into
5812 5879 // this generation. If such a promotion may still fail,
5813 5880 // the flag will be set again when a young collection is
5814 5881 // attempted.
5815 5882 // I think the incremental_collection_will_fail flag's use
5816 5883 // is specific to a 2 generation collection policy, so i'll
5817 5884 // assert that that's the configuration we are operating within.
5818 5885 // The use of the flag can and should be generalized appropriately
5819 5886 // in the future to deal with a general n-generation system.
5820 5887
5821 5888 GenCollectedHeap* gch = GenCollectedHeap::heap();
5822 5889 assert(gch->collector_policy()->is_two_generation_policy(),
5823 5890 "Resetting of incremental_collection_will_fail flag"
5824 5891 " may be incorrect otherwise");
5825 5892 gch->clear_incremental_collection_will_fail();
5826 5893 gch->update_full_collections_completed(_collection_count_start);
5827 5894 }
5828 5895
5829 5896 // FIX ME!!! Looks like this belongs in CFLSpace, with
5830 5897 // CMSGen merely delegating to it.
5831 5898 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5832 5899 double nearLargestPercent = 0.999;
5833 5900 HeapWord* minAddr = _cmsSpace->bottom();
5834 5901 HeapWord* largestAddr =
5835 5902 (HeapWord*) _cmsSpace->dictionary()->findLargestDict();
5836 5903 if (largestAddr == 0) {
5837 5904 // The dictionary appears to be empty. In this case
5838 5905 // try to coalesce at the end of the heap.
5839 5906 largestAddr = _cmsSpace->end();
5840 5907 }
5841 5908 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5842 5909 size_t nearLargestOffset =
5843 5910 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5844 5911 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5845 5912 }
5846 5913
5847 5914 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5848 5915 return addr >= _cmsSpace->nearLargestChunk();
5849 5916 }
5850 5917
5851 5918 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5852 5919 return _cmsSpace->find_chunk_at_end();
5853 5920 }
5854 5921
5855 5922 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
5856 5923 bool full) {
5857 5924 // The next lower level has been collected. Gather any statistics
5858 5925 // that are of interest at this point.
5859 5926 if (!full && (current_level + 1) == level()) {
5860 5927 // Gather statistics on the young generation collection.
5861 5928 collector()->stats().record_gc0_end(used());
5862 5929 }
5863 5930 }
5864 5931
5865 5932 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
5866 5933 GenCollectedHeap* gch = GenCollectedHeap::heap();
5867 5934 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
5868 5935 "Wrong type of heap");
5869 5936 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
5870 5937 gch->gen_policy()->size_policy();
5871 5938 assert(sp->is_gc_cms_adaptive_size_policy(),
5872 5939 "Wrong type of size policy");
5873 5940 return sp;
5874 5941 }
5875 5942
5876 5943 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
5877 5944 if (PrintGCDetails && Verbose) {
5878 5945 gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
5879 5946 }
5880 5947 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
5881 5948 _debug_collection_type =
5882 5949 (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
5883 5950 if (PrintGCDetails && Verbose) {
5884 5951 gclog_or_tty->print_cr("to %d ", _debug_collection_type);
5885 5952 }
5886 5953 }
5887 5954
5888 5955 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
5889 5956 bool asynch) {
5890 5957 // We iterate over the space(s) underlying this generation,
5891 5958 // checking the mark bit map to see if the bits corresponding
5892 5959 // to specific blocks are marked or not. Blocks that are
5893 5960 // marked are live and are not swept up. All remaining blocks
5894 5961 // are swept up, with coalescing on-the-fly as we sweep up
5895 5962 // contiguous free and/or garbage blocks:
5896 5963 // We need to ensure that the sweeper synchronizes with allocators
5897 5964 // and stop-the-world collectors. In particular, the following
5898 5965 // locks are used:
5899 5966 // . CMS token: if this is held, a stop the world collection cannot occur
5900 5967 // . freelistLock: if this is held no allocation can occur from this
5901 5968 // generation by another thread
5902 5969 // . bitMapLock: if this is held, no other thread can access or update
5903 5970 //
5904 5971
5905 5972 // Note that we need to hold the freelistLock if we use
5906 5973 // block iterate below; else the iterator might go awry if
5907 5974 // a mutator (or promotion) causes block contents to change
5908 5975 // (for instance if the allocator divvies up a block).
5909 5976 // If we hold the free list lock, for all practical purposes
5910 5977 // young generation GC's can't occur (they'll usually need to
5911 5978 // promote), so we might as well prevent all young generation
5912 5979 // GC's while we do a sweeping step. For the same reason, we might
5913 5980 // as well take the bit map lock for the entire duration
5914 5981
5915 5982 // check that we hold the requisite locks
5916 5983 assert(have_cms_token(), "Should hold cms token");
5917 5984 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
5918 5985 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
5919 5986 "Should possess CMS token to sweep");
5920 5987 assert_lock_strong(gen->freelistLock());
5921 5988 assert_lock_strong(bitMapLock());
5922 5989
5923 5990 assert(!_sweep_timer.is_active(), "Was switched off in an outer context");
5924 5991 gen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()),
5925 5992 _sweep_estimate.padded_average());
5926 5993 gen->setNearLargestChunk();
5927 5994
↓ open down ↓ |
140 lines elided |
↑ open up ↑ |
5928 5995 {
5929 5996 SweepClosure sweepClosure(this, gen, &_markBitMap,
5930 5997 CMSYield && asynch);
5931 5998 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5932 5999 // We need to free-up/coalesce garbage/blocks from a
5933 6000 // co-terminal free run. This is done in the SweepClosure
5934 6001 // destructor; so, do not remove this scope, else the
5935 6002 // end-of-sweep-census below will be off by a little bit.
5936 6003 }
5937 6004 gen->cmsSpace()->sweep_completed();
5938 - gen->cmsSpace()->endSweepFLCensus(sweepCount());
6005 + gen->cmsSpace()->endSweepFLCensus(sweep_count());
6006 + if (should_unload_classes()) { // unloaded classes this cycle,
6007 + _concurrent_cycles_since_last_unload = 0; // ... reset count
6008 + } else { // did not unload classes,
6009 + _concurrent_cycles_since_last_unload++; // ... increment count
6010 + }
5939 6011 }
5940 6012
5941 6013 // Reset CMS data structures (for now just the marking bit map)
5942 6014 // preparatory for the next cycle.
5943 6015 void CMSCollector::reset(bool asynch) {
5944 6016 GenCollectedHeap* gch = GenCollectedHeap::heap();
5945 6017 CMSAdaptiveSizePolicy* sp = size_policy();
5946 6018 AdaptiveSizePolicyOutput(sp, gch->total_collections());
5947 6019 if (asynch) {
5948 6020 CMSTokenSyncWithLocks ts(true, bitMapLock());
5949 6021
5950 6022 // If the state is not "Resetting", the foreground thread
5951 6023 // has done a collection and the resetting.
5952 6024 if (_collectorState != Resetting) {
5953 6025 assert(_collectorState == Idling, "The state should only change"
5954 6026 " because the foreground collector has finished the collection");
5955 6027 return;
5956 6028 }
5957 6029
5958 6030 // Clear the mark bitmap (no grey objects to start with)
5959 6031 // for the next cycle.
5960 6032 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5961 6033 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
5962 6034
5963 6035 HeapWord* curAddr = _markBitMap.startWord();
5964 6036 while (curAddr < _markBitMap.endWord()) {
5965 6037 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
5966 6038 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5967 6039 _markBitMap.clear_large_range(chunk);
5968 6040 if (ConcurrentMarkSweepThread::should_yield() &&
5969 6041 !foregroundGCIsActive() &&
5970 6042 CMSYield) {
5971 6043 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5972 6044 "CMS thread should hold CMS token");
5973 6045 assert_lock_strong(bitMapLock());
5974 6046 bitMapLock()->unlock();
5975 6047 ConcurrentMarkSweepThread::desynchronize(true);
5976 6048 ConcurrentMarkSweepThread::acknowledge_yield_request();
5977 6049 stopTimer();
5978 6050 if (PrintCMSStatistics != 0) {
5979 6051 incrementYields();
5980 6052 }
5981 6053 icms_wait();
5982 6054
5983 6055 // See the comment in coordinator_yield()
5984 6056 for (unsigned i = 0; i < CMSYieldSleepCount &&
5985 6057 ConcurrentMarkSweepThread::should_yield() &&
5986 6058 !CMSCollector::foregroundGCIsActive(); ++i) {
5987 6059 os::sleep(Thread::current(), 1, false);
5988 6060 ConcurrentMarkSweepThread::acknowledge_yield_request();
5989 6061 }
5990 6062
5991 6063 ConcurrentMarkSweepThread::synchronize(true);
5992 6064 bitMapLock()->lock_without_safepoint_check();
5993 6065 startTimer();
5994 6066 }
5995 6067 curAddr = chunk.end();
5996 6068 }
5997 6069 _collectorState = Idling;
5998 6070 } else {
5999 6071 // already have the lock
6000 6072 assert(_collectorState == Resetting, "just checking");
6001 6073 assert_lock_strong(bitMapLock());
6002 6074 _markBitMap.clear_all();
6003 6075 _collectorState = Idling;
6004 6076 }
6005 6077
6006 6078 // Stop incremental mode after a cycle completes, so that any future cycles
6007 6079 // are triggered by allocation.
6008 6080 stop_icms();
6009 6081
6010 6082 NOT_PRODUCT(
6011 6083 if (RotateCMSCollectionTypes) {
6012 6084 _cmsGen->rotate_debug_collection_type();
6013 6085 }
6014 6086 )
6015 6087 }
6016 6088
6017 6089 void CMSCollector::do_CMS_operation(CMS_op_type op) {
6018 6090 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
6019 6091 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6020 6092 TraceTime t("GC", PrintGC, !PrintGCDetails, gclog_or_tty);
6021 6093 TraceCollectorStats tcs(counters());
6022 6094
6023 6095 switch (op) {
6024 6096 case CMS_op_checkpointRootsInitial: {
6025 6097 checkpointRootsInitial(true); // asynch
6026 6098 if (PrintGC) {
6027 6099 _cmsGen->printOccupancy("initial-mark");
6028 6100 }
6029 6101 break;
6030 6102 }
6031 6103 case CMS_op_checkpointRootsFinal: {
6032 6104 checkpointRootsFinal(true, // asynch
6033 6105 false, // !clear_all_soft_refs
6034 6106 false); // !init_mark_was_synchronous
6035 6107 if (PrintGC) {
6036 6108 _cmsGen->printOccupancy("remark");
6037 6109 }
6038 6110 break;
6039 6111 }
6040 6112 default:
6041 6113 fatal("No such CMS_op");
6042 6114 }
6043 6115 }
6044 6116
6045 6117 #ifndef PRODUCT
6046 6118 size_t const CMSCollector::skip_header_HeapWords() {
6047 6119 return FreeChunk::header_size();
6048 6120 }
6049 6121
6050 6122 // Try and collect here conditions that should hold when
6051 6123 // CMS thread is exiting. The idea is that the foreground GC
6052 6124 // thread should not be blocked if it wants to terminate
6053 6125 // the CMS thread and yet continue to run the VM for a while
6054 6126 // after that.
6055 6127 void CMSCollector::verify_ok_to_terminate() const {
6056 6128 assert(Thread::current()->is_ConcurrentGC_thread(),
6057 6129 "should be called by CMS thread");
6058 6130 assert(!_foregroundGCShouldWait, "should be false");
6059 6131 // We could check here that all the various low-level locks
6060 6132 // are not held by the CMS thread, but that is overkill; see
6061 6133 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6062 6134 // is checked.
6063 6135 }
6064 6136 #endif
6065 6137
6066 6138 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6067 6139 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6068 6140 "missing Printezis mark?");
6069 6141 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6070 6142 size_t size = pointer_delta(nextOneAddr + 1, addr);
6071 6143 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6072 6144 "alignment problem");
6073 6145 assert(size >= 3, "Necessary for Printezis marks to work");
6074 6146 return size;
6075 6147 }
6076 6148
6077 6149 // A variant of the above (block_size_using_printezis_bits()) except
6078 6150 // that we return 0 if the P-bits are not yet set.
6079 6151 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6080 6152 if (_markBitMap.isMarked(addr)) {
6081 6153 assert(_markBitMap.isMarked(addr + 1), "Missing Printezis bit?");
6082 6154 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6083 6155 size_t size = pointer_delta(nextOneAddr + 1, addr);
6084 6156 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6085 6157 "alignment problem");
6086 6158 assert(size >= 3, "Necessary for Printezis marks to work");
6087 6159 return size;
6088 6160 } else {
6089 6161 assert(!_markBitMap.isMarked(addr + 1), "Bit map inconsistency?");
6090 6162 return 0;
6091 6163 }
6092 6164 }
6093 6165
6094 6166 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6095 6167 size_t sz = 0;
6096 6168 oop p = (oop)addr;
6097 6169 if (p->klass() != NULL && p->is_parsable()) {
6098 6170 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6099 6171 } else {
6100 6172 sz = block_size_using_printezis_bits(addr);
6101 6173 }
6102 6174 assert(sz > 0, "size must be nonzero");
6103 6175 HeapWord* next_block = addr + sz;
6104 6176 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
6105 6177 CardTableModRefBS::card_size);
6106 6178 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
6107 6179 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6108 6180 "must be different cards");
6109 6181 return next_card;
6110 6182 }
6111 6183
6112 6184
6113 6185 // CMS Bit Map Wrapper /////////////////////////////////////////
6114 6186
6115 6187 // Construct a CMS bit map infrastructure, but don't create the
6116 6188 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6117 6189 // further below.
6118 6190 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6119 6191 _bm(NULL,0),
6120 6192 _shifter(shifter),
6121 6193 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6122 6194 {
6123 6195 _bmStartWord = 0;
6124 6196 _bmWordSize = 0;
6125 6197 }
6126 6198
6127 6199 bool CMSBitMap::allocate(MemRegion mr) {
6128 6200 _bmStartWord = mr.start();
6129 6201 _bmWordSize = mr.word_size();
6130 6202 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6131 6203 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6132 6204 if (!brs.is_reserved()) {
6133 6205 warning("CMS bit map allocation failure");
6134 6206 return false;
6135 6207 }
6136 6208 // For now we'll just commit all of the bit map up fromt.
6137 6209 // Later on we'll try to be more parsimonious with swap.
6138 6210 if (!_virtual_space.initialize(brs, brs.size())) {
6139 6211 warning("CMS bit map backing store failure");
6140 6212 return false;
6141 6213 }
6142 6214 assert(_virtual_space.committed_size() == brs.size(),
6143 6215 "didn't reserve backing store for all of CMS bit map?");
6144 6216 _bm.set_map((uintptr_t*)_virtual_space.low());
6145 6217 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6146 6218 _bmWordSize, "inconsistency in bit map sizing");
6147 6219 _bm.set_size(_bmWordSize >> _shifter);
6148 6220
6149 6221 // bm.clear(); // can we rely on getting zero'd memory? verify below
6150 6222 assert(isAllClear(),
6151 6223 "Expected zero'd memory from ReservedSpace constructor");
6152 6224 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6153 6225 "consistency check");
6154 6226 return true;
6155 6227 }
6156 6228
6157 6229 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6158 6230 HeapWord *next_addr, *end_addr, *last_addr;
6159 6231 assert_locked();
6160 6232 assert(covers(mr), "out-of-range error");
6161 6233 // XXX assert that start and end are appropriately aligned
6162 6234 for (next_addr = mr.start(), end_addr = mr.end();
6163 6235 next_addr < end_addr; next_addr = last_addr) {
6164 6236 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6165 6237 last_addr = dirty_region.end();
6166 6238 if (!dirty_region.is_empty()) {
6167 6239 cl->do_MemRegion(dirty_region);
6168 6240 } else {
6169 6241 assert(last_addr == end_addr, "program logic");
6170 6242 return;
6171 6243 }
6172 6244 }
6173 6245 }
6174 6246
6175 6247 #ifndef PRODUCT
6176 6248 void CMSBitMap::assert_locked() const {
6177 6249 CMSLockVerifier::assert_locked(lock());
6178 6250 }
6179 6251
6180 6252 bool CMSBitMap::covers(MemRegion mr) const {
6181 6253 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6182 6254 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6183 6255 "size inconsistency");
6184 6256 return (mr.start() >= _bmStartWord) &&
6185 6257 (mr.end() <= endWord());
6186 6258 }
6187 6259
6188 6260 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6189 6261 return (start >= _bmStartWord && (start + size) <= endWord());
6190 6262 }
6191 6263
6192 6264 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6193 6265 // verify that there are no 1 bits in the interval [left, right)
6194 6266 FalseBitMapClosure falseBitMapClosure;
6195 6267 iterate(&falseBitMapClosure, left, right);
6196 6268 }
6197 6269
6198 6270 void CMSBitMap::region_invariant(MemRegion mr)
6199 6271 {
6200 6272 assert_locked();
6201 6273 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6202 6274 assert(!mr.is_empty(), "unexpected empty region");
6203 6275 assert(covers(mr), "mr should be covered by bit map");
6204 6276 // convert address range into offset range
6205 6277 size_t start_ofs = heapWordToOffset(mr.start());
6206 6278 // Make sure that end() is appropriately aligned
6207 6279 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6208 6280 (1 << (_shifter+LogHeapWordSize))),
6209 6281 "Misaligned mr.end()");
6210 6282 size_t end_ofs = heapWordToOffset(mr.end());
6211 6283 assert(end_ofs > start_ofs, "Should mark at least one bit");
6212 6284 }
6213 6285
6214 6286 #endif
6215 6287
6216 6288 bool CMSMarkStack::allocate(size_t size) {
6217 6289 // allocate a stack of the requisite depth
6218 6290 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6219 6291 size * sizeof(oop)));
6220 6292 if (!rs.is_reserved()) {
6221 6293 warning("CMSMarkStack allocation failure");
6222 6294 return false;
6223 6295 }
6224 6296 if (!_virtual_space.initialize(rs, rs.size())) {
6225 6297 warning("CMSMarkStack backing store failure");
6226 6298 return false;
6227 6299 }
6228 6300 assert(_virtual_space.committed_size() == rs.size(),
6229 6301 "didn't reserve backing store for all of CMS stack?");
6230 6302 _base = (oop*)(_virtual_space.low());
6231 6303 _index = 0;
6232 6304 _capacity = size;
6233 6305 NOT_PRODUCT(_max_depth = 0);
6234 6306 return true;
6235 6307 }
6236 6308
6237 6309 // XXX FIX ME !!! In the MT case we come in here holding a
6238 6310 // leaf lock. For printing we need to take a further lock
6239 6311 // which has lower rank. We need to recallibrate the two
6240 6312 // lock-ranks involved in order to be able to rpint the
6241 6313 // messages below. (Or defer the printing to the caller.
6242 6314 // For now we take the expedient path of just disabling the
6243 6315 // messages for the problematic case.)
6244 6316 void CMSMarkStack::expand() {
6245 6317 assert(_capacity <= CMSMarkStackSizeMax, "stack bigger than permitted");
6246 6318 if (_capacity == CMSMarkStackSizeMax) {
6247 6319 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6248 6320 // We print a warning message only once per CMS cycle.
6249 6321 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6250 6322 }
6251 6323 return;
6252 6324 }
6253 6325 // Double capacity if possible
6254 6326 size_t new_capacity = MIN2(_capacity*2, CMSMarkStackSizeMax);
6255 6327 // Do not give up existing stack until we have managed to
6256 6328 // get the double capacity that we desired.
6257 6329 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6258 6330 new_capacity * sizeof(oop)));
6259 6331 if (rs.is_reserved()) {
6260 6332 // Release the backing store associated with old stack
6261 6333 _virtual_space.release();
6262 6334 // Reinitialize virtual space for new stack
6263 6335 if (!_virtual_space.initialize(rs, rs.size())) {
6264 6336 fatal("Not enough swap for expanded marking stack");
6265 6337 }
6266 6338 _base = (oop*)(_virtual_space.low());
6267 6339 _index = 0;
6268 6340 _capacity = new_capacity;
6269 6341 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6270 6342 // Failed to double capacity, continue;
6271 6343 // we print a detail message only once per CMS cycle.
6272 6344 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6273 6345 SIZE_FORMAT"K",
6274 6346 _capacity / K, new_capacity / K);
6275 6347 }
6276 6348 }
6277 6349
6278 6350
6279 6351 // Closures
6280 6352 // XXX: there seems to be a lot of code duplication here;
6281 6353 // should refactor and consolidate common code.
6282 6354
6283 6355 // This closure is used to mark refs into the CMS generation in
6284 6356 // the CMS bit map. Called at the first checkpoint. This closure
6285 6357 // assumes that we do not need to re-mark dirty cards; if the CMS
6286 6358 // generation on which this is used is not an oldest (modulo perm gen)
6287 6359 // generation then this will lose younger_gen cards!
6288 6360
6289 6361 MarkRefsIntoClosure::MarkRefsIntoClosure(
6290 6362 MemRegion span, CMSBitMap* bitMap, bool should_do_nmethods):
6291 6363 _span(span),
6292 6364 _bitMap(bitMap),
6293 6365 _should_do_nmethods(should_do_nmethods)
6294 6366 {
6295 6367 assert(_ref_processor == NULL, "deliberately left NULL");
6296 6368 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6297 6369 }
6298 6370
6299 6371 void MarkRefsIntoClosure::do_oop(oop* p) {
6300 6372 // if p points into _span, then mark corresponding bit in _markBitMap
6301 6373 oop thisOop = *p;
6302 6374 if (thisOop != NULL) {
6303 6375 assert(thisOop->is_oop(), "expected an oop");
6304 6376 HeapWord* addr = (HeapWord*)thisOop;
6305 6377 if (_span.contains(addr)) {
6306 6378 // this should be made more efficient
6307 6379 _bitMap->mark(addr);
6308 6380 }
6309 6381 }
6310 6382 }
6311 6383
6312 6384 // A variant of the above, used for CMS marking verification.
6313 6385 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6314 6386 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6315 6387 bool should_do_nmethods):
6316 6388 _span(span),
6317 6389 _verification_bm(verification_bm),
6318 6390 _cms_bm(cms_bm),
6319 6391 _should_do_nmethods(should_do_nmethods) {
6320 6392 assert(_ref_processor == NULL, "deliberately left NULL");
6321 6393 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6322 6394 }
6323 6395
6324 6396 void MarkRefsIntoVerifyClosure::do_oop(oop* p) {
6325 6397 // if p points into _span, then mark corresponding bit in _markBitMap
6326 6398 oop this_oop = *p;
6327 6399 if (this_oop != NULL) {
6328 6400 assert(this_oop->is_oop(), "expected an oop");
6329 6401 HeapWord* addr = (HeapWord*)this_oop;
6330 6402 if (_span.contains(addr)) {
6331 6403 _verification_bm->mark(addr);
6332 6404 if (!_cms_bm->isMarked(addr)) {
6333 6405 oop(addr)->print();
6334 6406 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
6335 6407 fatal("... aborting");
6336 6408 }
6337 6409 }
6338 6410 }
6339 6411 }
6340 6412
6341 6413 //////////////////////////////////////////////////
6342 6414 // MarkRefsIntoAndScanClosure
6343 6415 //////////////////////////////////////////////////
6344 6416
6345 6417 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6346 6418 ReferenceProcessor* rp,
6347 6419 CMSBitMap* bit_map,
6348 6420 CMSBitMap* mod_union_table,
6349 6421 CMSMarkStack* mark_stack,
6350 6422 CMSMarkStack* revisit_stack,
6351 6423 CMSCollector* collector,
6352 6424 bool should_yield,
6353 6425 bool concurrent_precleaning):
6354 6426 _collector(collector),
6355 6427 _span(span),
6356 6428 _bit_map(bit_map),
6357 6429 _mark_stack(mark_stack),
6358 6430 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6359 6431 mark_stack, revisit_stack, concurrent_precleaning),
6360 6432 _yield(should_yield),
6361 6433 _concurrent_precleaning(concurrent_precleaning),
6362 6434 _freelistLock(NULL)
6363 6435 {
6364 6436 _ref_processor = rp;
6365 6437 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6366 6438 }
6367 6439
6368 6440 // This closure is used to mark refs into the CMS generation at the
6369 6441 // second (final) checkpoint, and to scan and transitively follow
6370 6442 // the unmarked oops. It is also used during the concurrent precleaning
6371 6443 // phase while scanning objects on dirty cards in the CMS generation.
6372 6444 // The marks are made in the marking bit map and the marking stack is
6373 6445 // used for keeping the (newly) grey objects during the scan.
6374 6446 // The parallel version (Par_...) appears further below.
6375 6447 void MarkRefsIntoAndScanClosure::do_oop(oop* p) {
6376 6448 oop this_oop = *p;
6377 6449 if (this_oop != NULL) {
6378 6450 assert(this_oop->is_oop(), "expected an oop");
6379 6451 HeapWord* addr = (HeapWord*)this_oop;
6380 6452 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6381 6453 assert(_collector->overflow_list_is_empty(), "should be empty");
6382 6454 if (_span.contains(addr) &&
6383 6455 !_bit_map->isMarked(addr)) {
6384 6456 // mark bit map (object is now grey)
6385 6457 _bit_map->mark(addr);
6386 6458 // push on marking stack (stack should be empty), and drain the
6387 6459 // stack by applying this closure to the oops in the oops popped
6388 6460 // from the stack (i.e. blacken the grey objects)
6389 6461 bool res = _mark_stack->push(this_oop);
6390 6462 assert(res, "Should have space to push on empty stack");
6391 6463 do {
6392 6464 oop new_oop = _mark_stack->pop();
6393 6465 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6394 6466 assert(new_oop->is_parsable(), "Found unparsable oop");
6395 6467 assert(_bit_map->isMarked((HeapWord*)new_oop),
6396 6468 "only grey objects on this stack");
6397 6469 // iterate over the oops in this oop, marking and pushing
6398 6470 // the ones in CMS heap (i.e. in _span).
6399 6471 new_oop->oop_iterate(&_pushAndMarkClosure);
6400 6472 // check if it's time to yield
6401 6473 do_yield_check();
6402 6474 } while (!_mark_stack->isEmpty() ||
6403 6475 (!_concurrent_precleaning && take_from_overflow_list()));
6404 6476 // if marking stack is empty, and we are not doing this
6405 6477 // during precleaning, then check the overflow list
6406 6478 }
6407 6479 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6408 6480 assert(_collector->overflow_list_is_empty(),
6409 6481 "overflow list was drained above");
6410 6482 // We could restore evacuated mark words, if any, used for
6411 6483 // overflow list links here because the overflow list is
6412 6484 // provably empty here. That would reduce the maximum
6413 6485 // size requirements for preserved_{oop,mark}_stack.
6414 6486 // But we'll just postpone it until we are all done
6415 6487 // so we can just stream through.
6416 6488 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6417 6489 _collector->restore_preserved_marks_if_any();
6418 6490 assert(_collector->no_preserved_marks(), "No preserved marks");
6419 6491 }
6420 6492 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6421 6493 "All preserved marks should have been restored above");
6422 6494 }
6423 6495 }
6424 6496
6425 6497 void MarkRefsIntoAndScanClosure::do_yield_work() {
6426 6498 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6427 6499 "CMS thread should hold CMS token");
6428 6500 assert_lock_strong(_freelistLock);
6429 6501 assert_lock_strong(_bit_map->lock());
6430 6502 // relinquish the free_list_lock and bitMaplock()
6431 6503 _bit_map->lock()->unlock();
6432 6504 _freelistLock->unlock();
6433 6505 ConcurrentMarkSweepThread::desynchronize(true);
6434 6506 ConcurrentMarkSweepThread::acknowledge_yield_request();
6435 6507 _collector->stopTimer();
6436 6508 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6437 6509 if (PrintCMSStatistics != 0) {
6438 6510 _collector->incrementYields();
6439 6511 }
6440 6512 _collector->icms_wait();
6441 6513
6442 6514 // See the comment in coordinator_yield()
6443 6515 for (unsigned i = 0; i < CMSYieldSleepCount &&
6444 6516 ConcurrentMarkSweepThread::should_yield() &&
6445 6517 !CMSCollector::foregroundGCIsActive(); ++i) {
6446 6518 os::sleep(Thread::current(), 1, false);
6447 6519 ConcurrentMarkSweepThread::acknowledge_yield_request();
6448 6520 }
6449 6521
6450 6522 ConcurrentMarkSweepThread::synchronize(true);
6451 6523 _freelistLock->lock_without_safepoint_check();
6452 6524 _bit_map->lock()->lock_without_safepoint_check();
6453 6525 _collector->startTimer();
6454 6526 }
6455 6527
6456 6528 ///////////////////////////////////////////////////////////
6457 6529 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6458 6530 // MarkRefsIntoAndScanClosure
6459 6531 ///////////////////////////////////////////////////////////
6460 6532 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6461 6533 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6462 6534 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack):
6463 6535 _span(span),
6464 6536 _bit_map(bit_map),
6465 6537 _work_queue(work_queue),
6466 6538 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6467 6539 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6468 6540 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue,
6469 6541 revisit_stack)
6470 6542 {
6471 6543 _ref_processor = rp;
6472 6544 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6473 6545 }
6474 6546
6475 6547 // This closure is used to mark refs into the CMS generation at the
6476 6548 // second (final) checkpoint, and to scan and transitively follow
6477 6549 // the unmarked oops. The marks are made in the marking bit map and
6478 6550 // the work_queue is used for keeping the (newly) grey objects during
6479 6551 // the scan phase whence they are also available for stealing by parallel
6480 6552 // threads. Since the marking bit map is shared, updates are
6481 6553 // synchronized (via CAS).
6482 6554 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) {
6483 6555 oop this_oop = *p;
6484 6556 if (this_oop != NULL) {
6485 6557 // Ignore mark word because this could be an already marked oop
6486 6558 // that may be chained at the end of the overflow list.
6487 6559 assert(this_oop->is_oop(true /* ignore mark word */), "expected an oop");
6488 6560 HeapWord* addr = (HeapWord*)this_oop;
6489 6561 if (_span.contains(addr) &&
6490 6562 !_bit_map->isMarked(addr)) {
6491 6563 // mark bit map (object will become grey):
6492 6564 // It is possible for several threads to be
6493 6565 // trying to "claim" this object concurrently;
6494 6566 // the unique thread that succeeds in marking the
6495 6567 // object first will do the subsequent push on
6496 6568 // to the work queue (or overflow list).
6497 6569 if (_bit_map->par_mark(addr)) {
6498 6570 // push on work_queue (which may not be empty), and trim the
6499 6571 // queue to an appropriate length by applying this closure to
6500 6572 // the oops in the oops popped from the stack (i.e. blacken the
6501 6573 // grey objects)
6502 6574 bool res = _work_queue->push(this_oop);
6503 6575 assert(res, "Low water mark should be less than capacity?");
6504 6576 trim_queue(_low_water_mark);
6505 6577 } // Else, another thread claimed the object
6506 6578 }
6507 6579 }
6508 6580 }
6509 6581
6510 6582 // This closure is used to rescan the marked objects on the dirty cards
6511 6583 // in the mod union table and the card table proper.
6512 6584 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6513 6585 oop p, MemRegion mr) {
6514 6586
6515 6587 size_t size = 0;
6516 6588 HeapWord* addr = (HeapWord*)p;
6517 6589 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6518 6590 assert(_span.contains(addr), "we are scanning the CMS generation");
6519 6591 // check if it's time to yield
6520 6592 if (do_yield_check()) {
6521 6593 // We yielded for some foreground stop-world work,
6522 6594 // and we have been asked to abort this ongoing preclean cycle.
6523 6595 return 0;
6524 6596 }
6525 6597 if (_bitMap->isMarked(addr)) {
6526 6598 // it's marked; is it potentially uninitialized?
6527 6599 if (p->klass() != NULL) {
6528 6600 if (CMSPermGenPrecleaningEnabled && !p->is_parsable()) {
6529 6601 // Signal precleaning to redirty the card since
6530 6602 // the klass pointer is already installed.
6531 6603 assert(size == 0, "Initial value");
6532 6604 } else {
6533 6605 assert(p->is_parsable(), "must be parsable.");
6534 6606 // an initialized object; ignore mark word in verification below
6535 6607 // since we are running concurrent with mutators
6536 6608 assert(p->is_oop(true), "should be an oop");
6537 6609 if (p->is_objArray()) {
6538 6610 // objArrays are precisely marked; restrict scanning
6539 6611 // to dirty cards only.
6540 6612 size = p->oop_iterate(_scanningClosure, mr);
6541 6613 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6542 6614 "adjustObjectSize should be the identity for array sizes, "
6543 6615 "which are necessarily larger than minimum object size of "
6544 6616 "two heap words");
6545 6617 } else {
6546 6618 // A non-array may have been imprecisely marked; we need
6547 6619 // to scan object in its entirety.
6548 6620 size = CompactibleFreeListSpace::adjustObjectSize(
6549 6621 p->oop_iterate(_scanningClosure));
6550 6622 }
6551 6623 #ifdef DEBUG
6552 6624 size_t direct_size =
6553 6625 CompactibleFreeListSpace::adjustObjectSize(p->size());
6554 6626 assert(size == direct_size, "Inconsistency in size");
6555 6627 assert(size >= 3, "Necessary for Printezis marks to work");
6556 6628 if (!_bitMap->isMarked(addr+1)) {
6557 6629 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6558 6630 } else {
6559 6631 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6560 6632 assert(_bitMap->isMarked(addr+size-1),
6561 6633 "inconsistent Printezis mark");
6562 6634 }
6563 6635 #endif // DEBUG
6564 6636 }
6565 6637 } else {
6566 6638 // an unitialized object
6567 6639 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6568 6640 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6569 6641 size = pointer_delta(nextOneAddr + 1, addr);
6570 6642 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6571 6643 "alignment problem");
6572 6644 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6573 6645 // will dirty the card when the klass pointer is installed in the
6574 6646 // object (signalling the completion of initialization).
6575 6647 }
6576 6648 } else {
6577 6649 // Either a not yet marked object or an uninitialized object
6578 6650 if (p->klass() == NULL || !p->is_parsable()) {
6579 6651 // An uninitialized object, skip to the next card, since
6580 6652 // we may not be able to read its P-bits yet.
6581 6653 assert(size == 0, "Initial value");
6582 6654 } else {
6583 6655 // An object not (yet) reached by marking: we merely need to
6584 6656 // compute its size so as to go look at the next block.
6585 6657 assert(p->is_oop(true), "should be an oop");
6586 6658 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6587 6659 }
6588 6660 }
6589 6661 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6590 6662 return size;
6591 6663 }
6592 6664
6593 6665 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6594 6666 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6595 6667 "CMS thread should hold CMS token");
6596 6668 assert_lock_strong(_freelistLock);
6597 6669 assert_lock_strong(_bitMap->lock());
6598 6670 // relinquish the free_list_lock and bitMaplock()
6599 6671 _bitMap->lock()->unlock();
6600 6672 _freelistLock->unlock();
6601 6673 ConcurrentMarkSweepThread::desynchronize(true);
6602 6674 ConcurrentMarkSweepThread::acknowledge_yield_request();
6603 6675 _collector->stopTimer();
6604 6676 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6605 6677 if (PrintCMSStatistics != 0) {
6606 6678 _collector->incrementYields();
6607 6679 }
6608 6680 _collector->icms_wait();
6609 6681
6610 6682 // See the comment in coordinator_yield()
6611 6683 for (unsigned i = 0; i < CMSYieldSleepCount &&
6612 6684 ConcurrentMarkSweepThread::should_yield() &&
6613 6685 !CMSCollector::foregroundGCIsActive(); ++i) {
6614 6686 os::sleep(Thread::current(), 1, false);
6615 6687 ConcurrentMarkSweepThread::acknowledge_yield_request();
6616 6688 }
6617 6689
6618 6690 ConcurrentMarkSweepThread::synchronize(true);
6619 6691 _freelistLock->lock_without_safepoint_check();
6620 6692 _bitMap->lock()->lock_without_safepoint_check();
6621 6693 _collector->startTimer();
6622 6694 }
6623 6695
6624 6696
6625 6697 //////////////////////////////////////////////////////////////////
6626 6698 // SurvivorSpacePrecleanClosure
6627 6699 //////////////////////////////////////////////////////////////////
6628 6700 // This (single-threaded) closure is used to preclean the oops in
6629 6701 // the survivor spaces.
6630 6702 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6631 6703
6632 6704 HeapWord* addr = (HeapWord*)p;
6633 6705 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6634 6706 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6635 6707 assert(p->klass() != NULL, "object should be initializd");
6636 6708 assert(p->is_parsable(), "must be parsable.");
6637 6709 // an initialized object; ignore mark word in verification below
6638 6710 // since we are running concurrent with mutators
6639 6711 assert(p->is_oop(true), "should be an oop");
6640 6712 // Note that we do not yield while we iterate over
6641 6713 // the interior oops of p, pushing the relevant ones
6642 6714 // on our marking stack.
6643 6715 size_t size = p->oop_iterate(_scanning_closure);
6644 6716 do_yield_check();
6645 6717 // Observe that below, we do not abandon the preclean
6646 6718 // phase as soon as we should; rather we empty the
6647 6719 // marking stack before returning. This is to satisfy
6648 6720 // some existing assertions. In general, it may be a
6649 6721 // good idea to abort immediately and complete the marking
6650 6722 // from the grey objects at a later time.
6651 6723 while (!_mark_stack->isEmpty()) {
6652 6724 oop new_oop = _mark_stack->pop();
6653 6725 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6654 6726 assert(new_oop->is_parsable(), "Found unparsable oop");
6655 6727 assert(_bit_map->isMarked((HeapWord*)new_oop),
6656 6728 "only grey objects on this stack");
6657 6729 // iterate over the oops in this oop, marking and pushing
6658 6730 // the ones in CMS heap (i.e. in _span).
6659 6731 new_oop->oop_iterate(_scanning_closure);
6660 6732 // check if it's time to yield
6661 6733 do_yield_check();
6662 6734 }
6663 6735 unsigned int after_count =
6664 6736 GenCollectedHeap::heap()->total_collections();
6665 6737 bool abort = (_before_count != after_count) ||
6666 6738 _collector->should_abort_preclean();
6667 6739 return abort ? 0 : size;
6668 6740 }
6669 6741
6670 6742 void SurvivorSpacePrecleanClosure::do_yield_work() {
6671 6743 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6672 6744 "CMS thread should hold CMS token");
6673 6745 assert_lock_strong(_bit_map->lock());
6674 6746 // Relinquish the bit map lock
6675 6747 _bit_map->lock()->unlock();
6676 6748 ConcurrentMarkSweepThread::desynchronize(true);
6677 6749 ConcurrentMarkSweepThread::acknowledge_yield_request();
6678 6750 _collector->stopTimer();
6679 6751 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6680 6752 if (PrintCMSStatistics != 0) {
6681 6753 _collector->incrementYields();
6682 6754 }
6683 6755 _collector->icms_wait();
6684 6756
6685 6757 // See the comment in coordinator_yield()
6686 6758 for (unsigned i = 0; i < CMSYieldSleepCount &&
6687 6759 ConcurrentMarkSweepThread::should_yield() &&
6688 6760 !CMSCollector::foregroundGCIsActive(); ++i) {
6689 6761 os::sleep(Thread::current(), 1, false);
6690 6762 ConcurrentMarkSweepThread::acknowledge_yield_request();
6691 6763 }
6692 6764
6693 6765 ConcurrentMarkSweepThread::synchronize(true);
6694 6766 _bit_map->lock()->lock_without_safepoint_check();
6695 6767 _collector->startTimer();
6696 6768 }
6697 6769
6698 6770 // This closure is used to rescan the marked objects on the dirty cards
6699 6771 // in the mod union table and the card table proper. In the parallel
6700 6772 // case, although the bitMap is shared, we do a single read so the
6701 6773 // isMarked() query is "safe".
6702 6774 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6703 6775 // Ignore mark word because we are running concurrent with mutators
6704 6776 assert(p->is_oop_or_null(true), "expected an oop or null");
6705 6777 HeapWord* addr = (HeapWord*)p;
6706 6778 assert(_span.contains(addr), "we are scanning the CMS generation");
6707 6779 bool is_obj_array = false;
6708 6780 #ifdef DEBUG
6709 6781 if (!_parallel) {
6710 6782 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6711 6783 assert(_collector->overflow_list_is_empty(),
6712 6784 "overflow list should be empty");
6713 6785
6714 6786 }
6715 6787 #endif // DEBUG
6716 6788 if (_bit_map->isMarked(addr)) {
6717 6789 // Obj arrays are precisely marked, non-arrays are not;
6718 6790 // so we scan objArrays precisely and non-arrays in their
6719 6791 // entirety.
6720 6792 if (p->is_objArray()) {
6721 6793 is_obj_array = true;
6722 6794 if (_parallel) {
6723 6795 p->oop_iterate(_par_scan_closure, mr);
6724 6796 } else {
6725 6797 p->oop_iterate(_scan_closure, mr);
6726 6798 }
6727 6799 } else {
6728 6800 if (_parallel) {
6729 6801 p->oop_iterate(_par_scan_closure);
6730 6802 } else {
6731 6803 p->oop_iterate(_scan_closure);
6732 6804 }
6733 6805 }
6734 6806 }
6735 6807 #ifdef DEBUG
6736 6808 if (!_parallel) {
6737 6809 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6738 6810 assert(_collector->overflow_list_is_empty(),
6739 6811 "overflow list should be empty");
6740 6812
6741 6813 }
6742 6814 #endif // DEBUG
6743 6815 return is_obj_array;
6744 6816 }
6745 6817
6746 6818 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6747 6819 MemRegion span,
6748 6820 CMSBitMap* bitMap, CMSMarkStack* markStack,
6749 6821 CMSMarkStack* revisitStack,
6750 6822 bool should_yield, bool verifying):
6751 6823 _collector(collector),
6752 6824 _span(span),
6753 6825 _bitMap(bitMap),
6754 6826 _mut(&collector->_modUnionTable),
6755 6827 _markStack(markStack),
6756 6828 _revisitStack(revisitStack),
6757 6829 _yield(should_yield),
6758 6830 _skipBits(0)
6759 6831 {
6760 6832 assert(_markStack->isEmpty(), "stack should be empty");
6761 6833 _finger = _bitMap->startWord();
6762 6834 _threshold = _finger;
6763 6835 assert(_collector->_restart_addr == NULL, "Sanity check");
6764 6836 assert(_span.contains(_finger), "Out of bounds _finger?");
6765 6837 DEBUG_ONLY(_verifying = verifying;)
6766 6838 }
6767 6839
6768 6840 void MarkFromRootsClosure::reset(HeapWord* addr) {
6769 6841 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6770 6842 assert(_span.contains(addr), "Out of bounds _finger?");
6771 6843 _finger = addr;
6772 6844 _threshold = (HeapWord*)round_to(
6773 6845 (intptr_t)_finger, CardTableModRefBS::card_size);
6774 6846 }
6775 6847
6776 6848 // Should revisit to see if this should be restructured for
6777 6849 // greater efficiency.
6778 6850 void MarkFromRootsClosure::do_bit(size_t offset) {
6779 6851 if (_skipBits > 0) {
6780 6852 _skipBits--;
6781 6853 return;
6782 6854 }
6783 6855 // convert offset into a HeapWord*
6784 6856 HeapWord* addr = _bitMap->startWord() + offset;
6785 6857 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6786 6858 "address out of range");
6787 6859 assert(_bitMap->isMarked(addr), "tautology");
6788 6860 if (_bitMap->isMarked(addr+1)) {
6789 6861 // this is an allocated but not yet initialized object
6790 6862 assert(_skipBits == 0, "tautology");
6791 6863 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6792 6864 oop p = oop(addr);
6793 6865 if (p->klass() == NULL || !p->is_parsable()) {
6794 6866 DEBUG_ONLY(if (!_verifying) {)
6795 6867 // We re-dirty the cards on which this object lies and increase
6796 6868 // the _threshold so that we'll come back to scan this object
6797 6869 // during the preclean or remark phase. (CMSCleanOnEnter)
6798 6870 if (CMSCleanOnEnter) {
6799 6871 size_t sz = _collector->block_size_using_printezis_bits(addr);
6800 6872 HeapWord* start_card_addr = (HeapWord*)round_down(
6801 6873 (intptr_t)addr, CardTableModRefBS::card_size);
6802 6874 HeapWord* end_card_addr = (HeapWord*)round_to(
6803 6875 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6804 6876 MemRegion redirty_range = MemRegion(start_card_addr, end_card_addr);
6805 6877 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6806 6878 // Bump _threshold to end_card_addr; note that
6807 6879 // _threshold cannot possibly exceed end_card_addr, anyhow.
6808 6880 // This prevents future clearing of the card as the scan proceeds
6809 6881 // to the right.
6810 6882 assert(_threshold <= end_card_addr,
6811 6883 "Because we are just scanning into this object");
6812 6884 if (_threshold < end_card_addr) {
6813 6885 _threshold = end_card_addr;
6814 6886 }
6815 6887 if (p->klass() != NULL) {
6816 6888 // Redirty the range of cards...
6817 6889 _mut->mark_range(redirty_range);
6818 6890 } // ...else the setting of klass will dirty the card anyway.
6819 6891 }
6820 6892 DEBUG_ONLY(})
6821 6893 return;
6822 6894 }
6823 6895 }
6824 6896 scanOopsInOop(addr);
6825 6897 }
6826 6898
6827 6899 // We take a break if we've been at this for a while,
6828 6900 // so as to avoid monopolizing the locks involved.
6829 6901 void MarkFromRootsClosure::do_yield_work() {
6830 6902 // First give up the locks, then yield, then re-lock
6831 6903 // We should probably use a constructor/destructor idiom to
6832 6904 // do this unlock/lock or modify the MutexUnlocker class to
6833 6905 // serve our purpose. XXX
6834 6906 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6835 6907 "CMS thread should hold CMS token");
6836 6908 assert_lock_strong(_bitMap->lock());
6837 6909 _bitMap->lock()->unlock();
6838 6910 ConcurrentMarkSweepThread::desynchronize(true);
6839 6911 ConcurrentMarkSweepThread::acknowledge_yield_request();
6840 6912 _collector->stopTimer();
6841 6913 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6842 6914 if (PrintCMSStatistics != 0) {
6843 6915 _collector->incrementYields();
6844 6916 }
6845 6917 _collector->icms_wait();
6846 6918
6847 6919 // See the comment in coordinator_yield()
6848 6920 for (unsigned i = 0; i < CMSYieldSleepCount &&
6849 6921 ConcurrentMarkSweepThread::should_yield() &&
6850 6922 !CMSCollector::foregroundGCIsActive(); ++i) {
6851 6923 os::sleep(Thread::current(), 1, false);
6852 6924 ConcurrentMarkSweepThread::acknowledge_yield_request();
6853 6925 }
6854 6926
6855 6927 ConcurrentMarkSweepThread::synchronize(true);
6856 6928 _bitMap->lock()->lock_without_safepoint_check();
6857 6929 _collector->startTimer();
6858 6930 }
6859 6931
6860 6932 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6861 6933 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6862 6934 assert(_markStack->isEmpty(),
6863 6935 "should drain stack to limit stack usage");
6864 6936 // convert ptr to an oop preparatory to scanning
6865 6937 oop this_oop = oop(ptr);
6866 6938 // Ignore mark word in verification below, since we
6867 6939 // may be running concurrent with mutators.
6868 6940 assert(this_oop->is_oop(true), "should be an oop");
6869 6941 assert(_finger <= ptr, "_finger runneth ahead");
6870 6942 // advance the finger to right end of this object
6871 6943 _finger = ptr + this_oop->size();
6872 6944 assert(_finger > ptr, "we just incremented it above");
6873 6945 // On large heaps, it may take us some time to get through
6874 6946 // the marking phase (especially if running iCMS). During
6875 6947 // this time it's possible that a lot of mutations have
6876 6948 // accumulated in the card table and the mod union table --
6877 6949 // these mutation records are redundant until we have
6878 6950 // actually traced into the corresponding card.
6879 6951 // Here, we check whether advancing the finger would make
6880 6952 // us cross into a new card, and if so clear corresponding
6881 6953 // cards in the MUT (preclean them in the card-table in the
6882 6954 // future).
6883 6955
6884 6956 DEBUG_ONLY(if (!_verifying) {)
6885 6957 // The clean-on-enter optimization is disabled by default,
6886 6958 // until we fix 6178663.
6887 6959 if (CMSCleanOnEnter && (_finger > _threshold)) {
6888 6960 // [_threshold, _finger) represents the interval
6889 6961 // of cards to be cleared in MUT (or precleaned in card table).
6890 6962 // The set of cards to be cleared is all those that overlap
6891 6963 // with the interval [_threshold, _finger); note that
6892 6964 // _threshold is always kept card-aligned but _finger isn't
6893 6965 // always card-aligned.
6894 6966 HeapWord* old_threshold = _threshold;
6895 6967 assert(old_threshold == (HeapWord*)round_to(
6896 6968 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6897 6969 "_threshold should always be card-aligned");
6898 6970 _threshold = (HeapWord*)round_to(
6899 6971 (intptr_t)_finger, CardTableModRefBS::card_size);
6900 6972 MemRegion mr(old_threshold, _threshold);
6901 6973 assert(!mr.is_empty(), "Control point invariant");
6902 6974 assert(_span.contains(mr), "Should clear within span");
6903 6975 // XXX When _finger crosses from old gen into perm gen
6904 6976 // we may be doing unnecessary cleaning; do better in the
6905 6977 // future by detecting that condition and clearing fewer
6906 6978 // MUT/CT entries.
6907 6979 _mut->clear_range(mr);
6908 6980 }
6909 6981 DEBUG_ONLY(})
6910 6982
6911 6983 // Note: the finger doesn't advance while we drain
6912 6984 // the stack below.
6913 6985 PushOrMarkClosure pushOrMarkClosure(_collector,
6914 6986 _span, _bitMap, _markStack,
6915 6987 _revisitStack,
6916 6988 _finger, this);
6917 6989 bool res = _markStack->push(this_oop);
6918 6990 assert(res, "Empty non-zero size stack should have space for single push");
6919 6991 while (!_markStack->isEmpty()) {
6920 6992 oop new_oop = _markStack->pop();
6921 6993 // Skip verifying header mark word below because we are
6922 6994 // running concurrent with mutators.
6923 6995 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6924 6996 // now scan this oop's oops
6925 6997 new_oop->oop_iterate(&pushOrMarkClosure);
6926 6998 do_yield_check();
6927 6999 }
6928 7000 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6929 7001 }
6930 7002
6931 7003 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
6932 7004 CMSCollector* collector, MemRegion span,
6933 7005 CMSBitMap* bit_map,
6934 7006 OopTaskQueue* work_queue,
6935 7007 CMSMarkStack* overflow_stack,
6936 7008 CMSMarkStack* revisit_stack,
6937 7009 bool should_yield):
6938 7010 _collector(collector),
6939 7011 _whole_span(collector->_span),
6940 7012 _span(span),
6941 7013 _bit_map(bit_map),
6942 7014 _mut(&collector->_modUnionTable),
6943 7015 _work_queue(work_queue),
6944 7016 _overflow_stack(overflow_stack),
6945 7017 _revisit_stack(revisit_stack),
6946 7018 _yield(should_yield),
6947 7019 _skip_bits(0),
6948 7020 _task(task)
6949 7021 {
6950 7022 assert(_work_queue->size() == 0, "work_queue should be empty");
6951 7023 _finger = span.start();
6952 7024 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
6953 7025 assert(_span.contains(_finger), "Out of bounds _finger?");
6954 7026 }
6955 7027
6956 7028 // Should revisit to see if this should be restructured for
6957 7029 // greater efficiency.
6958 7030 void Par_MarkFromRootsClosure::do_bit(size_t offset) {
6959 7031 if (_skip_bits > 0) {
6960 7032 _skip_bits--;
6961 7033 return;
6962 7034 }
6963 7035 // convert offset into a HeapWord*
6964 7036 HeapWord* addr = _bit_map->startWord() + offset;
6965 7037 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6966 7038 "address out of range");
6967 7039 assert(_bit_map->isMarked(addr), "tautology");
6968 7040 if (_bit_map->isMarked(addr+1)) {
6969 7041 // this is an allocated object that might not yet be initialized
6970 7042 assert(_skip_bits == 0, "tautology");
6971 7043 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
6972 7044 oop p = oop(addr);
6973 7045 if (p->klass() == NULL || !p->is_parsable()) {
6974 7046 // in the case of Clean-on-Enter optimization, redirty card
6975 7047 // and avoid clearing card by increasing the threshold.
6976 7048 return;
6977 7049 }
6978 7050 }
6979 7051 scan_oops_in_oop(addr);
6980 7052 }
6981 7053
6982 7054 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6983 7055 assert(_bit_map->isMarked(ptr), "expected bit to be set");
6984 7056 // Should we assert that our work queue is empty or
6985 7057 // below some drain limit?
6986 7058 assert(_work_queue->size() == 0,
6987 7059 "should drain stack to limit stack usage");
6988 7060 // convert ptr to an oop preparatory to scanning
6989 7061 oop this_oop = oop(ptr);
6990 7062 // Ignore mark word in verification below, since we
6991 7063 // may be running concurrent with mutators.
6992 7064 assert(this_oop->is_oop(true), "should be an oop");
6993 7065 assert(_finger <= ptr, "_finger runneth ahead");
6994 7066 // advance the finger to right end of this object
6995 7067 _finger = ptr + this_oop->size();
6996 7068 assert(_finger > ptr, "we just incremented it above");
6997 7069 // On large heaps, it may take us some time to get through
6998 7070 // the marking phase (especially if running iCMS). During
6999 7071 // this time it's possible that a lot of mutations have
7000 7072 // accumulated in the card table and the mod union table --
7001 7073 // these mutation records are redundant until we have
7002 7074 // actually traced into the corresponding card.
7003 7075 // Here, we check whether advancing the finger would make
7004 7076 // us cross into a new card, and if so clear corresponding
7005 7077 // cards in the MUT (preclean them in the card-table in the
7006 7078 // future).
7007 7079
7008 7080 // The clean-on-enter optimization is disabled by default,
7009 7081 // until we fix 6178663.
7010 7082 if (CMSCleanOnEnter && (_finger > _threshold)) {
7011 7083 // [_threshold, _finger) represents the interval
7012 7084 // of cards to be cleared in MUT (or precleaned in card table).
7013 7085 // The set of cards to be cleared is all those that overlap
7014 7086 // with the interval [_threshold, _finger); note that
7015 7087 // _threshold is always kept card-aligned but _finger isn't
7016 7088 // always card-aligned.
7017 7089 HeapWord* old_threshold = _threshold;
7018 7090 assert(old_threshold == (HeapWord*)round_to(
7019 7091 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7020 7092 "_threshold should always be card-aligned");
7021 7093 _threshold = (HeapWord*)round_to(
7022 7094 (intptr_t)_finger, CardTableModRefBS::card_size);
7023 7095 MemRegion mr(old_threshold, _threshold);
7024 7096 assert(!mr.is_empty(), "Control point invariant");
7025 7097 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7026 7098 // XXX When _finger crosses from old gen into perm gen
7027 7099 // we may be doing unnecessary cleaning; do better in the
7028 7100 // future by detecting that condition and clearing fewer
7029 7101 // MUT/CT entries.
7030 7102 _mut->clear_range(mr);
7031 7103 }
7032 7104
7033 7105 // Note: the local finger doesn't advance while we drain
7034 7106 // the stack below, but the global finger sure can and will.
7035 7107 HeapWord** gfa = _task->global_finger_addr();
7036 7108 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7037 7109 _span, _bit_map,
7038 7110 _work_queue,
7039 7111 _overflow_stack,
7040 7112 _revisit_stack,
7041 7113 _finger,
7042 7114 gfa, this);
7043 7115 bool res = _work_queue->push(this_oop); // overflow could occur here
7044 7116 assert(res, "Will hold once we use workqueues");
7045 7117 while (true) {
7046 7118 oop new_oop;
7047 7119 if (!_work_queue->pop_local(new_oop)) {
7048 7120 // We emptied our work_queue; check if there's stuff that can
7049 7121 // be gotten from the overflow stack.
7050 7122 if (CMSConcMarkingTask::get_work_from_overflow_stack(
7051 7123 _overflow_stack, _work_queue)) {
7052 7124 do_yield_check();
7053 7125 continue;
7054 7126 } else { // done
7055 7127 break;
7056 7128 }
7057 7129 }
7058 7130 // Skip verifying header mark word below because we are
7059 7131 // running concurrent with mutators.
7060 7132 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7061 7133 // now scan this oop's oops
7062 7134 new_oop->oop_iterate(&pushOrMarkClosure);
7063 7135 do_yield_check();
7064 7136 }
7065 7137 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7066 7138 }
7067 7139
7068 7140 // Yield in response to a request from VM Thread or
7069 7141 // from mutators.
7070 7142 void Par_MarkFromRootsClosure::do_yield_work() {
7071 7143 assert(_task != NULL, "sanity");
7072 7144 _task->yield();
7073 7145 }
7074 7146
7075 7147 // A variant of the above used for verifying CMS marking work.
7076 7148 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7077 7149 MemRegion span,
7078 7150 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7079 7151 CMSMarkStack* mark_stack):
7080 7152 _collector(collector),
7081 7153 _span(span),
7082 7154 _verification_bm(verification_bm),
7083 7155 _cms_bm(cms_bm),
7084 7156 _mark_stack(mark_stack),
7085 7157 _pam_verify_closure(collector, span, verification_bm, cms_bm,
7086 7158 mark_stack)
7087 7159 {
7088 7160 assert(_mark_stack->isEmpty(), "stack should be empty");
7089 7161 _finger = _verification_bm->startWord();
7090 7162 assert(_collector->_restart_addr == NULL, "Sanity check");
7091 7163 assert(_span.contains(_finger), "Out of bounds _finger?");
7092 7164 }
7093 7165
7094 7166 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7095 7167 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7096 7168 assert(_span.contains(addr), "Out of bounds _finger?");
7097 7169 _finger = addr;
7098 7170 }
7099 7171
7100 7172 // Should revisit to see if this should be restructured for
7101 7173 // greater efficiency.
7102 7174 void MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7103 7175 // convert offset into a HeapWord*
7104 7176 HeapWord* addr = _verification_bm->startWord() + offset;
7105 7177 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7106 7178 "address out of range");
7107 7179 assert(_verification_bm->isMarked(addr), "tautology");
7108 7180 assert(_cms_bm->isMarked(addr), "tautology");
7109 7181
7110 7182 assert(_mark_stack->isEmpty(),
7111 7183 "should drain stack to limit stack usage");
7112 7184 // convert addr to an oop preparatory to scanning
7113 7185 oop this_oop = oop(addr);
7114 7186 assert(this_oop->is_oop(), "should be an oop");
7115 7187 assert(_finger <= addr, "_finger runneth ahead");
7116 7188 // advance the finger to right end of this object
7117 7189 _finger = addr + this_oop->size();
7118 7190 assert(_finger > addr, "we just incremented it above");
7119 7191 // Note: the finger doesn't advance while we drain
7120 7192 // the stack below.
7121 7193 bool res = _mark_stack->push(this_oop);
7122 7194 assert(res, "Empty non-zero size stack should have space for single push");
7123 7195 while (!_mark_stack->isEmpty()) {
7124 7196 oop new_oop = _mark_stack->pop();
7125 7197 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7126 7198 // now scan this oop's oops
7127 7199 new_oop->oop_iterate(&_pam_verify_closure);
7128 7200 }
7129 7201 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7130 7202 }
7131 7203
7132 7204 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7133 7205 CMSCollector* collector, MemRegion span,
7134 7206 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7135 7207 CMSMarkStack* mark_stack):
7136 7208 OopClosure(collector->ref_processor()),
7137 7209 _collector(collector),
7138 7210 _span(span),
7139 7211 _verification_bm(verification_bm),
7140 7212 _cms_bm(cms_bm),
7141 7213 _mark_stack(mark_stack)
7142 7214 { }
7143 7215
7144 7216
7145 7217 // Upon stack overflow, we discard (part of) the stack,
7146 7218 // remembering the least address amongst those discarded
7147 7219 // in CMSCollector's _restart_address.
7148 7220 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7149 7221 // Remember the least grey address discarded
7150 7222 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7151 7223 _collector->lower_restart_addr(ra);
7152 7224 _mark_stack->reset(); // discard stack contents
7153 7225 _mark_stack->expand(); // expand the stack if possible
7154 7226 }
7155 7227
7156 7228 void PushAndMarkVerifyClosure::do_oop(oop* p) {
7157 7229 oop this_oop = *p;
7158 7230 assert(this_oop->is_oop_or_null(), "expected an oop or NULL");
7159 7231 HeapWord* addr = (HeapWord*)this_oop;
7160 7232 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7161 7233 // Oop lies in _span and isn't yet grey or black
7162 7234 _verification_bm->mark(addr); // now grey
7163 7235 if (!_cms_bm->isMarked(addr)) {
7164 7236 oop(addr)->print();
7165 7237 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
7166 7238 fatal("... aborting");
7167 7239 }
7168 7240
7169 7241 if (!_mark_stack->push(this_oop)) { // stack overflow
7170 7242 if (PrintCMSStatistics != 0) {
7171 7243 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7172 7244 SIZE_FORMAT, _mark_stack->capacity());
7173 7245 }
7174 7246 assert(_mark_stack->isFull(), "Else push should have succeeded");
7175 7247 handle_stack_overflow(addr);
7176 7248 }
7177 7249 // anything including and to the right of _finger
7178 7250 // will be scanned as we iterate over the remainder of the
7179 7251 // bit map
7180 7252 }
7181 7253 }
7182 7254
7183 7255 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7184 7256 MemRegion span,
7185 7257 CMSBitMap* bitMap, CMSMarkStack* markStack,
↓ open down ↓ |
1237 lines elided |
↑ open up ↑ |
7186 7258 CMSMarkStack* revisitStack,
7187 7259 HeapWord* finger, MarkFromRootsClosure* parent) :
7188 7260 OopClosure(collector->ref_processor()),
7189 7261 _collector(collector),
7190 7262 _span(span),
7191 7263 _bitMap(bitMap),
7192 7264 _markStack(markStack),
7193 7265 _revisitStack(revisitStack),
7194 7266 _finger(finger),
7195 7267 _parent(parent),
7196 - _should_remember_klasses(collector->cms_should_unload_classes())
7268 + _should_remember_klasses(collector->should_unload_classes())
7197 7269 { }
7198 7270
7199 7271 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7200 7272 MemRegion span,
7201 7273 CMSBitMap* bit_map,
7202 7274 OopTaskQueue* work_queue,
7203 7275 CMSMarkStack* overflow_stack,
7204 7276 CMSMarkStack* revisit_stack,
7205 7277 HeapWord* finger,
7206 7278 HeapWord** global_finger_addr,
7207 7279 Par_MarkFromRootsClosure* parent) :
7208 7280 OopClosure(collector->ref_processor()),
↓ open down ↓ |
2 lines elided |
↑ open up ↑ |
7209 7281 _collector(collector),
7210 7282 _whole_span(collector->_span),
7211 7283 _span(span),
7212 7284 _bit_map(bit_map),
7213 7285 _work_queue(work_queue),
7214 7286 _overflow_stack(overflow_stack),
7215 7287 _revisit_stack(revisit_stack),
7216 7288 _finger(finger),
7217 7289 _global_finger_addr(global_finger_addr),
7218 7290 _parent(parent),
7219 - _should_remember_klasses(collector->cms_should_unload_classes())
7291 + _should_remember_klasses(collector->should_unload_classes())
7220 7292 { }
7221 7293
7222 7294
7223 7295 void CMSCollector::lower_restart_addr(HeapWord* low) {
7224 7296 assert(_span.contains(low), "Out of bounds addr");
7225 7297 if (_restart_addr == NULL) {
7226 7298 _restart_addr = low;
7227 7299 } else {
7228 7300 _restart_addr = MIN2(_restart_addr, low);
7229 7301 }
7230 7302 }
7231 7303
7232 7304 // Upon stack overflow, we discard (part of) the stack,
7233 7305 // remembering the least address amongst those discarded
7234 7306 // in CMSCollector's _restart_address.
7235 7307 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7236 7308 // Remember the least grey address discarded
7237 7309 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7238 7310 _collector->lower_restart_addr(ra);
7239 7311 _markStack->reset(); // discard stack contents
7240 7312 _markStack->expand(); // expand the stack if possible
7241 7313 }
7242 7314
7243 7315 // Upon stack overflow, we discard (part of) the stack,
7244 7316 // remembering the least address amongst those discarded
7245 7317 // in CMSCollector's _restart_address.
7246 7318 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7247 7319 // We need to do this under a mutex to prevent other
7248 7320 // workers from interfering with the expansion below.
7249 7321 MutexLockerEx ml(_overflow_stack->par_lock(),
7250 7322 Mutex::_no_safepoint_check_flag);
7251 7323 // Remember the least grey address discarded
7252 7324 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7253 7325 _collector->lower_restart_addr(ra);
7254 7326 _overflow_stack->reset(); // discard stack contents
7255 7327 _overflow_stack->expand(); // expand the stack if possible
7256 7328 }
7257 7329
7258 7330
7259 7331 void PushOrMarkClosure::do_oop(oop* p) {
7260 7332 oop thisOop = *p;
7261 7333 // Ignore mark word because we are running concurrent with mutators.
7262 7334 assert(thisOop->is_oop_or_null(true), "expected an oop or NULL");
7263 7335 HeapWord* addr = (HeapWord*)thisOop;
7264 7336 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7265 7337 // Oop lies in _span and isn't yet grey or black
7266 7338 _bitMap->mark(addr); // now grey
7267 7339 if (addr < _finger) {
7268 7340 // the bit map iteration has already either passed, or
7269 7341 // sampled, this bit in the bit map; we'll need to
7270 7342 // use the marking stack to scan this oop's oops.
7271 7343 bool simulate_overflow = false;
7272 7344 NOT_PRODUCT(
7273 7345 if (CMSMarkStackOverflowALot &&
7274 7346 _collector->simulate_overflow()) {
7275 7347 // simulate a stack overflow
7276 7348 simulate_overflow = true;
7277 7349 }
7278 7350 )
7279 7351 if (simulate_overflow || !_markStack->push(thisOop)) { // stack overflow
7280 7352 if (PrintCMSStatistics != 0) {
7281 7353 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7282 7354 SIZE_FORMAT, _markStack->capacity());
7283 7355 }
7284 7356 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7285 7357 handle_stack_overflow(addr);
7286 7358 }
7287 7359 }
7288 7360 // anything including and to the right of _finger
7289 7361 // will be scanned as we iterate over the remainder of the
7290 7362 // bit map
7291 7363 do_yield_check();
7292 7364 }
7293 7365 }
7294 7366
7295 7367 void Par_PushOrMarkClosure::do_oop(oop* p) {
7296 7368 oop this_oop = *p;
7297 7369 // Ignore mark word because we are running concurrent with mutators.
7298 7370 assert(this_oop->is_oop_or_null(true), "expected an oop or NULL");
7299 7371 HeapWord* addr = (HeapWord*)this_oop;
7300 7372 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7301 7373 // Oop lies in _span and isn't yet grey or black
7302 7374 // We read the global_finger (volatile read) strictly after marking oop
7303 7375 bool res = _bit_map->par_mark(addr); // now grey
7304 7376 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7305 7377 // Should we push this marked oop on our stack?
7306 7378 // -- if someone else marked it, nothing to do
7307 7379 // -- if target oop is above global finger nothing to do
7308 7380 // -- if target oop is in chunk and above local finger
7309 7381 // then nothing to do
7310 7382 // -- else push on work queue
7311 7383 if ( !res // someone else marked it, they will deal with it
7312 7384 || (addr >= *gfa) // will be scanned in a later task
7313 7385 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7314 7386 return;
7315 7387 }
7316 7388 // the bit map iteration has already either passed, or
7317 7389 // sampled, this bit in the bit map; we'll need to
7318 7390 // use the marking stack to scan this oop's oops.
7319 7391 bool simulate_overflow = false;
7320 7392 NOT_PRODUCT(
7321 7393 if (CMSMarkStackOverflowALot &&
7322 7394 _collector->simulate_overflow()) {
7323 7395 // simulate a stack overflow
7324 7396 simulate_overflow = true;
7325 7397 }
7326 7398 )
7327 7399 if (simulate_overflow ||
7328 7400 !(_work_queue->push(this_oop) || _overflow_stack->par_push(this_oop))) {
7329 7401 // stack overflow
7330 7402 if (PrintCMSStatistics != 0) {
7331 7403 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7332 7404 SIZE_FORMAT, _overflow_stack->capacity());
7333 7405 }
7334 7406 // We cannot assert that the overflow stack is full because
7335 7407 // it may have been emptied since.
7336 7408 assert(simulate_overflow ||
7337 7409 _work_queue->size() == _work_queue->max_elems(),
7338 7410 "Else push should have succeeded");
7339 7411 handle_stack_overflow(addr);
7340 7412 }
7341 7413 do_yield_check();
7342 7414 }
7343 7415 }
7344 7416
7345 7417
7346 7418 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7347 7419 MemRegion span,
7348 7420 ReferenceProcessor* rp,
7349 7421 CMSBitMap* bit_map,
7350 7422 CMSBitMap* mod_union_table,
7351 7423 CMSMarkStack* mark_stack,
↓ open down ↓ |
122 lines elided |
↑ open up ↑ |
7352 7424 CMSMarkStack* revisit_stack,
7353 7425 bool concurrent_precleaning):
7354 7426 OopClosure(rp),
7355 7427 _collector(collector),
7356 7428 _span(span),
7357 7429 _bit_map(bit_map),
7358 7430 _mod_union_table(mod_union_table),
7359 7431 _mark_stack(mark_stack),
7360 7432 _revisit_stack(revisit_stack),
7361 7433 _concurrent_precleaning(concurrent_precleaning),
7362 - _should_remember_klasses(collector->cms_should_unload_classes())
7434 + _should_remember_klasses(collector->should_unload_classes())
7363 7435 {
7364 7436 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7365 7437 }
7366 7438
7367 7439 // Grey object rescan during pre-cleaning and second checkpoint phases --
7368 7440 // the non-parallel version (the parallel version appears further below.)
7369 7441 void PushAndMarkClosure::do_oop(oop* p) {
7370 7442 oop this_oop = *p;
7371 7443 // Ignore mark word verification. If during concurrent precleaning
7372 7444 // the object monitor may be locked. If during the checkpoint
7373 7445 // phases, the object may already have been reached by a different
7374 7446 // path and may be at the end of the global overflow list (so
7375 7447 // the mark word may be NULL).
7376 7448 assert(this_oop->is_oop_or_null(true/* ignore mark word */),
7377 7449 "expected an oop or NULL");
7378 7450 HeapWord* addr = (HeapWord*)this_oop;
7379 7451 // Check if oop points into the CMS generation
7380 7452 // and is not marked
7381 7453 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7382 7454 // a white object ...
7383 7455 _bit_map->mark(addr); // ... now grey
7384 7456 // push on the marking stack (grey set)
7385 7457 bool simulate_overflow = false;
7386 7458 NOT_PRODUCT(
7387 7459 if (CMSMarkStackOverflowALot &&
7388 7460 _collector->simulate_overflow()) {
7389 7461 // simulate a stack overflow
7390 7462 simulate_overflow = true;
7391 7463 }
7392 7464 )
7393 7465 if (simulate_overflow || !_mark_stack->push(this_oop)) {
7394 7466 if (_concurrent_precleaning) {
7395 7467 // During precleaning we can just dirty the appropriate card
7396 7468 // in the mod union table, thus ensuring that the object remains
7397 7469 // in the grey set and continue. Note that no one can be intefering
7398 7470 // with us in this action of dirtying the mod union table, so
7399 7471 // no locking is required.
7400 7472 _mod_union_table->mark(addr);
7401 7473 _collector->_ser_pmc_preclean_ovflw++;
7402 7474 } else {
7403 7475 // During the remark phase, we need to remember this oop
7404 7476 // in the overflow list.
7405 7477 _collector->push_on_overflow_list(this_oop);
7406 7478 _collector->_ser_pmc_remark_ovflw++;
7407 7479 }
7408 7480 }
7409 7481 }
7410 7482 }
7411 7483
7412 7484 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7413 7485 MemRegion span,
↓ open down ↓ |
41 lines elided |
↑ open up ↑ |
7414 7486 ReferenceProcessor* rp,
7415 7487 CMSBitMap* bit_map,
7416 7488 OopTaskQueue* work_queue,
7417 7489 CMSMarkStack* revisit_stack):
7418 7490 OopClosure(rp),
7419 7491 _collector(collector),
7420 7492 _span(span),
7421 7493 _bit_map(bit_map),
7422 7494 _work_queue(work_queue),
7423 7495 _revisit_stack(revisit_stack),
7424 - _should_remember_klasses(collector->cms_should_unload_classes())
7496 + _should_remember_klasses(collector->should_unload_classes())
7425 7497 {
7426 7498 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7427 7499 }
7428 7500
7429 7501 // Grey object rescan during second checkpoint phase --
7430 7502 // the parallel version.
7431 7503 void Par_PushAndMarkClosure::do_oop(oop* p) {
7432 7504 oop this_oop = *p;
7433 7505 // In the assert below, we ignore the mark word because
7434 7506 // this oop may point to an already visited object that is
7435 7507 // on the overflow stack (in which case the mark word has
7436 7508 // been hijacked for chaining into the overflow stack --
7437 7509 // if this is the last object in the overflow stack then
7438 7510 // its mark word will be NULL). Because this object may
7439 7511 // have been subsequently popped off the global overflow
7440 7512 // stack, and the mark word possibly restored to the prototypical
7441 7513 // value, by the time we get to examined this failing assert in
7442 7514 // the debugger, is_oop_or_null(false) may subsequently start
7443 7515 // to hold.
7444 7516 assert(this_oop->is_oop_or_null(true),
7445 7517 "expected an oop or NULL");
7446 7518 HeapWord* addr = (HeapWord*)this_oop;
7447 7519 // Check if oop points into the CMS generation
7448 7520 // and is not marked
7449 7521 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7450 7522 // a white object ...
7451 7523 // If we manage to "claim" the object, by being the
7452 7524 // first thread to mark it, then we push it on our
7453 7525 // marking stack
7454 7526 if (_bit_map->par_mark(addr)) { // ... now grey
7455 7527 // push on work queue (grey set)
7456 7528 bool simulate_overflow = false;
7457 7529 NOT_PRODUCT(
7458 7530 if (CMSMarkStackOverflowALot &&
7459 7531 _collector->par_simulate_overflow()) {
7460 7532 // simulate a stack overflow
7461 7533 simulate_overflow = true;
7462 7534 }
7463 7535 )
7464 7536 if (simulate_overflow || !_work_queue->push(this_oop)) {
7465 7537 _collector->par_push_on_overflow_list(this_oop);
7466 7538 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7467 7539 }
7468 7540 } // Else, some other thread got there first
7469 7541 }
7470 7542 }
7471 7543
7472 7544 void PushAndMarkClosure::remember_klass(Klass* k) {
7473 7545 if (!_revisit_stack->push(oop(k))) {
7474 7546 fatal("Revisit stack overflowed in PushAndMarkClosure");
7475 7547 }
7476 7548 }
7477 7549
7478 7550 void Par_PushAndMarkClosure::remember_klass(Klass* k) {
7479 7551 if (!_revisit_stack->par_push(oop(k))) {
7480 7552 fatal("Revist stack overflowed in Par_PushAndMarkClosure");
7481 7553 }
7482 7554 }
7483 7555
7484 7556 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7485 7557 Mutex* bml = _collector->bitMapLock();
7486 7558 assert_lock_strong(bml);
7487 7559 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7488 7560 "CMS thread should hold CMS token");
7489 7561
7490 7562 bml->unlock();
7491 7563 ConcurrentMarkSweepThread::desynchronize(true);
7492 7564
7493 7565 ConcurrentMarkSweepThread::acknowledge_yield_request();
7494 7566
7495 7567 _collector->stopTimer();
7496 7568 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7497 7569 if (PrintCMSStatistics != 0) {
7498 7570 _collector->incrementYields();
7499 7571 }
7500 7572 _collector->icms_wait();
7501 7573
7502 7574 // See the comment in coordinator_yield()
7503 7575 for (unsigned i = 0; i < CMSYieldSleepCount &&
7504 7576 ConcurrentMarkSweepThread::should_yield() &&
7505 7577 !CMSCollector::foregroundGCIsActive(); ++i) {
7506 7578 os::sleep(Thread::current(), 1, false);
7507 7579 ConcurrentMarkSweepThread::acknowledge_yield_request();
7508 7580 }
7509 7581
7510 7582 ConcurrentMarkSweepThread::synchronize(true);
7511 7583 bml->lock();
7512 7584
7513 7585 _collector->startTimer();
7514 7586 }
7515 7587
7516 7588 bool CMSPrecleanRefsYieldClosure::should_return() {
7517 7589 if (ConcurrentMarkSweepThread::should_yield()) {
7518 7590 do_yield_work();
7519 7591 }
7520 7592 return _collector->foregroundGCIsActive();
7521 7593 }
7522 7594
7523 7595 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7524 7596 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7525 7597 "mr should be aligned to start at a card boundary");
7526 7598 // We'd like to assert:
7527 7599 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7528 7600 // "mr should be a range of cards");
7529 7601 // However, that would be too strong in one case -- the last
7530 7602 // partition ends at _unallocated_block which, in general, can be
7531 7603 // an arbitrary boundary, not necessarily card aligned.
7532 7604 if (PrintCMSStatistics != 0) {
7533 7605 _num_dirty_cards +=
7534 7606 mr.word_size()/CardTableModRefBS::card_size_in_words;
7535 7607 }
7536 7608 _space->object_iterate_mem(mr, &_scan_cl);
7537 7609 }
7538 7610
7539 7611 SweepClosure::SweepClosure(CMSCollector* collector,
7540 7612 ConcurrentMarkSweepGeneration* g,
7541 7613 CMSBitMap* bitMap, bool should_yield) :
7542 7614 _collector(collector),
7543 7615 _g(g),
7544 7616 _sp(g->cmsSpace()),
7545 7617 _limit(_sp->sweep_limit()),
7546 7618 _freelistLock(_sp->freelistLock()),
7547 7619 _bitMap(bitMap),
7548 7620 _yield(should_yield),
7549 7621 _inFreeRange(false), // No free range at beginning of sweep
7550 7622 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7551 7623 _lastFreeRangeCoalesced(false),
7552 7624 _freeFinger(g->used_region().start())
7553 7625 {
7554 7626 NOT_PRODUCT(
7555 7627 _numObjectsFreed = 0;
7556 7628 _numWordsFreed = 0;
7557 7629 _numObjectsLive = 0;
7558 7630 _numWordsLive = 0;
7559 7631 _numObjectsAlreadyFree = 0;
7560 7632 _numWordsAlreadyFree = 0;
7561 7633 _last_fc = NULL;
7562 7634
7563 7635 _sp->initializeIndexedFreeListArrayReturnedBytes();
7564 7636 _sp->dictionary()->initializeDictReturnedBytes();
7565 7637 )
7566 7638 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7567 7639 "sweep _limit out of bounds");
7568 7640 if (CMSTraceSweeper) {
7569 7641 gclog_or_tty->print("\n====================\nStarting new sweep\n");
7570 7642 }
7571 7643 }
7572 7644
7573 7645 // We need this destructor to reclaim any space at the end
7574 7646 // of the space, which do_blk below may not have added back to
7575 7647 // the free lists. [basically dealing with the "fringe effect"]
7576 7648 SweepClosure::~SweepClosure() {
7577 7649 assert_lock_strong(_freelistLock);
7578 7650 // this should be treated as the end of a free run if any
7579 7651 // The current free range should be returned to the free lists
7580 7652 // as one coalesced chunk.
7581 7653 if (inFreeRange()) {
7582 7654 flushCurFreeChunk(freeFinger(),
7583 7655 pointer_delta(_limit, freeFinger()));
7584 7656 assert(freeFinger() < _limit, "the finger pointeth off base");
7585 7657 if (CMSTraceSweeper) {
7586 7658 gclog_or_tty->print("destructor:");
7587 7659 gclog_or_tty->print("Sweep:put_free_blk 0x%x ("SIZE_FORMAT") "
7588 7660 "[coalesced:"SIZE_FORMAT"]\n",
7589 7661 freeFinger(), pointer_delta(_limit, freeFinger()),
7590 7662 lastFreeRangeCoalesced());
7591 7663 }
7592 7664 }
7593 7665 NOT_PRODUCT(
7594 7666 if (Verbose && PrintGC) {
7595 7667 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, "
7596 7668 SIZE_FORMAT " bytes",
7597 7669 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7598 7670 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
7599 7671 SIZE_FORMAT" bytes "
7600 7672 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
7601 7673 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7602 7674 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7603 7675 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) *
7604 7676 sizeof(HeapWord);
7605 7677 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
7606 7678
7607 7679 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7608 7680 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7609 7681 size_t dictReturnedBytes = _sp->dictionary()->sumDictReturnedBytes();
7610 7682 size_t returnedBytes = indexListReturnedBytes + dictReturnedBytes;
7611 7683 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returnedBytes);
7612 7684 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
7613 7685 indexListReturnedBytes);
7614 7686 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
7615 7687 dictReturnedBytes);
7616 7688 }
7617 7689 }
7618 7690 )
7619 7691 // Now, in debug mode, just null out the sweep_limit
7620 7692 NOT_PRODUCT(_sp->clear_sweep_limit();)
7621 7693 if (CMSTraceSweeper) {
7622 7694 gclog_or_tty->print("end of sweep\n================\n");
7623 7695 }
7624 7696 }
7625 7697
7626 7698 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7627 7699 bool freeRangeInFreeLists) {
7628 7700 if (CMSTraceSweeper) {
7629 7701 gclog_or_tty->print("---- Start free range 0x%x with free block [%d] (%d)\n",
7630 7702 freeFinger, _sp->block_size(freeFinger),
7631 7703 freeRangeInFreeLists);
7632 7704 }
7633 7705 assert(!inFreeRange(), "Trampling existing free range");
7634 7706 set_inFreeRange(true);
7635 7707 set_lastFreeRangeCoalesced(false);
7636 7708
7637 7709 set_freeFinger(freeFinger);
7638 7710 set_freeRangeInFreeLists(freeRangeInFreeLists);
7639 7711 if (CMSTestInFreeList) {
7640 7712 if (freeRangeInFreeLists) {
7641 7713 FreeChunk* fc = (FreeChunk*) freeFinger;
7642 7714 assert(fc->isFree(), "A chunk on the free list should be free.");
7643 7715 assert(fc->size() > 0, "Free range should have a size");
7644 7716 assert(_sp->verifyChunkInFreeLists(fc), "Chunk is not in free lists");
7645 7717 }
7646 7718 }
7647 7719 }
7648 7720
7649 7721 // Note that the sweeper runs concurrently with mutators. Thus,
7650 7722 // it is possible for direct allocation in this generation to happen
7651 7723 // in the middle of the sweep. Note that the sweeper also coalesces
7652 7724 // contiguous free blocks. Thus, unless the sweeper and the allocator
7653 7725 // synchronize appropriately freshly allocated blocks may get swept up.
7654 7726 // This is accomplished by the sweeper locking the free lists while
7655 7727 // it is sweeping. Thus blocks that are determined to be free are
7656 7728 // indeed free. There is however one additional complication:
7657 7729 // blocks that have been allocated since the final checkpoint and
7658 7730 // mark, will not have been marked and so would be treated as
7659 7731 // unreachable and swept up. To prevent this, the allocator marks
7660 7732 // the bit map when allocating during the sweep phase. This leads,
7661 7733 // however, to a further complication -- objects may have been allocated
7662 7734 // but not yet initialized -- in the sense that the header isn't yet
7663 7735 // installed. The sweeper can not then determine the size of the block
7664 7736 // in order to skip over it. To deal with this case, we use a technique
7665 7737 // (due to Printezis) to encode such uninitialized block sizes in the
7666 7738 // bit map. Since the bit map uses a bit per every HeapWord, but the
7667 7739 // CMS generation has a minimum object size of 3 HeapWords, it follows
7668 7740 // that "normal marks" won't be adjacent in the bit map (there will
7669 7741 // always be at least two 0 bits between successive 1 bits). We make use
7670 7742 // of these "unused" bits to represent uninitialized blocks -- the bit
7671 7743 // corresponding to the start of the uninitialized object and the next
7672 7744 // bit are both set. Finally, a 1 bit marks the end of the object that
7673 7745 // started with the two consecutive 1 bits to indicate its potentially
7674 7746 // uninitialized state.
7675 7747
7676 7748 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7677 7749 FreeChunk* fc = (FreeChunk*)addr;
7678 7750 size_t res;
7679 7751
7680 7752 // check if we are done sweepinrg
7681 7753 if (addr == _limit) { // we have swept up to the limit, do nothing more
7682 7754 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7683 7755 "sweep _limit out of bounds");
7684 7756 // help the closure application finish
7685 7757 return pointer_delta(_sp->end(), _limit);
7686 7758 }
7687 7759 assert(addr <= _limit, "sweep invariant");
7688 7760
7689 7761 // check if we should yield
7690 7762 do_yield_check(addr);
7691 7763 if (fc->isFree()) {
7692 7764 // Chunk that is already free
7693 7765 res = fc->size();
7694 7766 doAlreadyFreeChunk(fc);
7695 7767 debug_only(_sp->verifyFreeLists());
7696 7768 assert(res == fc->size(), "Don't expect the size to change");
7697 7769 NOT_PRODUCT(
7698 7770 _numObjectsAlreadyFree++;
7699 7771 _numWordsAlreadyFree += res;
7700 7772 )
7701 7773 NOT_PRODUCT(_last_fc = fc;)
7702 7774 } else if (!_bitMap->isMarked(addr)) {
7703 7775 // Chunk is fresh garbage
7704 7776 res = doGarbageChunk(fc);
7705 7777 debug_only(_sp->verifyFreeLists());
7706 7778 NOT_PRODUCT(
7707 7779 _numObjectsFreed++;
7708 7780 _numWordsFreed += res;
7709 7781 )
7710 7782 } else {
7711 7783 // Chunk that is alive.
7712 7784 res = doLiveChunk(fc);
7713 7785 debug_only(_sp->verifyFreeLists());
7714 7786 NOT_PRODUCT(
7715 7787 _numObjectsLive++;
7716 7788 _numWordsLive += res;
7717 7789 )
7718 7790 }
7719 7791 return res;
7720 7792 }
7721 7793
7722 7794 // For the smart allocation, record following
7723 7795 // split deaths - a free chunk is removed from its free list because
7724 7796 // it is being split into two or more chunks.
7725 7797 // split birth - a free chunk is being added to its free list because
7726 7798 // a larger free chunk has been split and resulted in this free chunk.
7727 7799 // coal death - a free chunk is being removed from its free list because
7728 7800 // it is being coalesced into a large free chunk.
7729 7801 // coal birth - a free chunk is being added to its free list because
7730 7802 // it was created when two or more free chunks where coalesced into
7731 7803 // this free chunk.
7732 7804 //
7733 7805 // These statistics are used to determine the desired number of free
7734 7806 // chunks of a given size. The desired number is chosen to be relative
7735 7807 // to the end of a CMS sweep. The desired number at the end of a sweep
7736 7808 // is the
7737 7809 // count-at-end-of-previous-sweep (an amount that was enough)
7738 7810 // - count-at-beginning-of-current-sweep (the excess)
7739 7811 // + split-births (gains in this size during interval)
7740 7812 // - split-deaths (demands on this size during interval)
7741 7813 // where the interval is from the end of one sweep to the end of the
7742 7814 // next.
7743 7815 //
7744 7816 // When sweeping the sweeper maintains an accumulated chunk which is
7745 7817 // the chunk that is made up of chunks that have been coalesced. That
7746 7818 // will be termed the left-hand chunk. A new chunk of garbage that
7747 7819 // is being considered for coalescing will be referred to as the
7748 7820 // right-hand chunk.
7749 7821 //
7750 7822 // When making a decision on whether to coalesce a right-hand chunk with
7751 7823 // the current left-hand chunk, the current count vs. the desired count
7752 7824 // of the left-hand chunk is considered. Also if the right-hand chunk
7753 7825 // is near the large chunk at the end of the heap (see
7754 7826 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7755 7827 // left-hand chunk is coalesced.
7756 7828 //
7757 7829 // When making a decision about whether to split a chunk, the desired count
7758 7830 // vs. the current count of the candidate to be split is also considered.
7759 7831 // If the candidate is underpopulated (currently fewer chunks than desired)
7760 7832 // a chunk of an overpopulated (currently more chunks than desired) size may
7761 7833 // be chosen. The "hint" associated with a free list, if non-null, points
7762 7834 // to a free list which may be overpopulated.
7763 7835 //
7764 7836
7765 7837 void SweepClosure::doAlreadyFreeChunk(FreeChunk* fc) {
7766 7838 size_t size = fc->size();
7767 7839 // Chunks that cannot be coalesced are not in the
7768 7840 // free lists.
7769 7841 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7770 7842 assert(_sp->verifyChunkInFreeLists(fc),
7771 7843 "free chunk should be in free lists");
7772 7844 }
7773 7845 // a chunk that is already free, should not have been
7774 7846 // marked in the bit map
7775 7847 HeapWord* addr = (HeapWord*) fc;
7776 7848 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7777 7849 // Verify that the bit map has no bits marked between
7778 7850 // addr and purported end of this block.
7779 7851 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7780 7852
7781 7853 // Some chunks cannot be coalesced in under any circumstances.
7782 7854 // See the definition of cantCoalesce().
7783 7855 if (!fc->cantCoalesce()) {
7784 7856 // This chunk can potentially be coalesced.
7785 7857 if (_sp->adaptive_freelists()) {
7786 7858 // All the work is done in
7787 7859 doPostIsFreeOrGarbageChunk(fc, size);
7788 7860 } else { // Not adaptive free lists
7789 7861 // this is a free chunk that can potentially be coalesced by the sweeper;
7790 7862 if (!inFreeRange()) {
7791 7863 // if the next chunk is a free block that can't be coalesced
7792 7864 // it doesn't make sense to remove this chunk from the free lists
7793 7865 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
7794 7866 assert((HeapWord*)nextChunk <= _limit, "sweep invariant");
7795 7867 if ((HeapWord*)nextChunk < _limit && // there's a next chunk...
7796 7868 nextChunk->isFree() && // which is free...
7797 7869 nextChunk->cantCoalesce()) { // ... but cant be coalesced
7798 7870 // nothing to do
7799 7871 } else {
7800 7872 // Potentially the start of a new free range:
7801 7873 // Don't eagerly remove it from the free lists.
7802 7874 // No need to remove it if it will just be put
7803 7875 // back again. (Also from a pragmatic point of view
7804 7876 // if it is a free block in a region that is beyond
7805 7877 // any allocated blocks, an assertion will fail)
7806 7878 // Remember the start of a free run.
7807 7879 initialize_free_range(addr, true);
7808 7880 // end - can coalesce with next chunk
7809 7881 }
7810 7882 } else {
7811 7883 // the midst of a free range, we are coalescing
7812 7884 debug_only(record_free_block_coalesced(fc);)
7813 7885 if (CMSTraceSweeper) {
7814 7886 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size);
7815 7887 }
7816 7888 // remove it from the free lists
7817 7889 _sp->removeFreeChunkFromFreeLists(fc);
7818 7890 set_lastFreeRangeCoalesced(true);
7819 7891 // If the chunk is being coalesced and the current free range is
7820 7892 // in the free lists, remove the current free range so that it
7821 7893 // will be returned to the free lists in its entirety - all
7822 7894 // the coalesced pieces included.
7823 7895 if (freeRangeInFreeLists()) {
7824 7896 FreeChunk* ffc = (FreeChunk*) freeFinger();
7825 7897 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7826 7898 "Size of free range is inconsistent with chunk size.");
7827 7899 if (CMSTestInFreeList) {
7828 7900 assert(_sp->verifyChunkInFreeLists(ffc),
7829 7901 "free range is not in free lists");
7830 7902 }
7831 7903 _sp->removeFreeChunkFromFreeLists(ffc);
7832 7904 set_freeRangeInFreeLists(false);
7833 7905 }
7834 7906 }
7835 7907 }
7836 7908 } else {
7837 7909 // Code path common to both original and adaptive free lists.
7838 7910
7839 7911 // cant coalesce with previous block; this should be treated
7840 7912 // as the end of a free run if any
7841 7913 if (inFreeRange()) {
7842 7914 // we kicked some butt; time to pick up the garbage
7843 7915 assert(freeFinger() < addr, "the finger pointeth off base");
7844 7916 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger()));
7845 7917 }
7846 7918 // else, nothing to do, just continue
7847 7919 }
7848 7920 }
7849 7921
7850 7922 size_t SweepClosure::doGarbageChunk(FreeChunk* fc) {
7851 7923 // This is a chunk of garbage. It is not in any free list.
7852 7924 // Add it to a free list or let it possibly be coalesced into
7853 7925 // a larger chunk.
7854 7926 HeapWord* addr = (HeapWord*) fc;
7855 7927 size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7856 7928
7857 7929 if (_sp->adaptive_freelists()) {
7858 7930 // Verify that the bit map has no bits marked between
7859 7931 // addr and purported end of just dead object.
7860 7932 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7861 7933
7862 7934 doPostIsFreeOrGarbageChunk(fc, size);
7863 7935 } else {
7864 7936 if (!inFreeRange()) {
7865 7937 // start of a new free range
7866 7938 assert(size > 0, "A free range should have a size");
7867 7939 initialize_free_range(addr, false);
7868 7940
7869 7941 } else {
7870 7942 // this will be swept up when we hit the end of the
7871 7943 // free range
7872 7944 if (CMSTraceSweeper) {
7873 7945 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size);
7874 7946 }
7875 7947 // If the chunk is being coalesced and the current free range is
7876 7948 // in the free lists, remove the current free range so that it
7877 7949 // will be returned to the free lists in its entirety - all
7878 7950 // the coalesced pieces included.
7879 7951 if (freeRangeInFreeLists()) {
7880 7952 FreeChunk* ffc = (FreeChunk*)freeFinger();
7881 7953 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7882 7954 "Size of free range is inconsistent with chunk size.");
7883 7955 if (CMSTestInFreeList) {
7884 7956 assert(_sp->verifyChunkInFreeLists(ffc),
7885 7957 "free range is not in free lists");
7886 7958 }
7887 7959 _sp->removeFreeChunkFromFreeLists(ffc);
7888 7960 set_freeRangeInFreeLists(false);
7889 7961 }
7890 7962 set_lastFreeRangeCoalesced(true);
7891 7963 }
7892 7964 // this will be swept up when we hit the end of the free range
7893 7965
7894 7966 // Verify that the bit map has no bits marked between
7895 7967 // addr and purported end of just dead object.
7896 7968 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7897 7969 }
7898 7970 return size;
7899 7971 }
7900 7972
7901 7973 size_t SweepClosure::doLiveChunk(FreeChunk* fc) {
7902 7974 HeapWord* addr = (HeapWord*) fc;
7903 7975 // The sweeper has just found a live object. Return any accumulated
7904 7976 // left hand chunk to the free lists.
7905 7977 if (inFreeRange()) {
7906 7978 if (_sp->adaptive_freelists()) {
7907 7979 flushCurFreeChunk(freeFinger(),
7908 7980 pointer_delta(addr, freeFinger()));
7909 7981 } else { // not adaptive freelists
7910 7982 set_inFreeRange(false);
7911 7983 // Add the free range back to the free list if it is not already
7912 7984 // there.
7913 7985 if (!freeRangeInFreeLists()) {
7914 7986 assert(freeFinger() < addr, "the finger pointeth off base");
7915 7987 if (CMSTraceSweeper) {
7916 7988 gclog_or_tty->print("Sweep:put_free_blk 0x%x (%d) "
7917 7989 "[coalesced:%d]\n",
7918 7990 freeFinger(), pointer_delta(addr, freeFinger()),
7919 7991 lastFreeRangeCoalesced());
7920 7992 }
7921 7993 _sp->addChunkAndRepairOffsetTable(freeFinger(),
7922 7994 pointer_delta(addr, freeFinger()), lastFreeRangeCoalesced());
7923 7995 }
7924 7996 }
7925 7997 }
7926 7998
7927 7999 // Common code path for original and adaptive free lists.
7928 8000
7929 8001 // this object is live: we'd normally expect this to be
7930 8002 // an oop, and like to assert the following:
7931 8003 // assert(oop(addr)->is_oop(), "live block should be an oop");
7932 8004 // However, as we commented above, this may be an object whose
7933 8005 // header hasn't yet been initialized.
7934 8006 size_t size;
7935 8007 assert(_bitMap->isMarked(addr), "Tautology for this control point");
↓ open down ↓ |
501 lines elided |
↑ open up ↑ |
7936 8008 if (_bitMap->isMarked(addr + 1)) {
7937 8009 // Determine the size from the bit map, rather than trying to
7938 8010 // compute it from the object header.
7939 8011 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7940 8012 size = pointer_delta(nextOneAddr + 1, addr);
7941 8013 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7942 8014 "alignment problem");
7943 8015
7944 8016 #ifdef DEBUG
7945 8017 if (oop(addr)->klass() != NULL &&
7946 - ( !_collector->cms_should_unload_classes()
8018 + ( !_collector->should_unload_classes()
7947 8019 || oop(addr)->is_parsable())) {
7948 8020 // Ignore mark word because we are running concurrent with mutators
7949 8021 assert(oop(addr)->is_oop(true), "live block should be an oop");
7950 8022 assert(size ==
7951 8023 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7952 8024 "P-mark and computed size do not agree");
7953 8025 }
7954 8026 #endif
7955 8027
7956 8028 } else {
7957 8029 // This should be an initialized object that's alive.
7958 8030 assert(oop(addr)->klass() != NULL &&
7959 - (!_collector->cms_should_unload_classes()
8031 + (!_collector->should_unload_classes()
7960 8032 || oop(addr)->is_parsable()),
7961 8033 "Should be an initialized object");
7962 8034 // Ignore mark word because we are running concurrent with mutators
7963 8035 assert(oop(addr)->is_oop(true), "live block should be an oop");
7964 8036 // Verify that the bit map has no bits marked between
7965 8037 // addr and purported end of this block.
7966 8038 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7967 8039 assert(size >= 3, "Necessary for Printezis marks to work");
7968 8040 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7969 8041 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7970 8042 }
7971 8043 return size;
7972 8044 }
7973 8045
7974 8046 void SweepClosure::doPostIsFreeOrGarbageChunk(FreeChunk* fc,
7975 8047 size_t chunkSize) {
7976 8048 // doPostIsFreeOrGarbageChunk() should only be called in the smart allocation
7977 8049 // scheme.
7978 8050 bool fcInFreeLists = fc->isFree();
7979 8051 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
7980 8052 assert((HeapWord*)fc <= _limit, "sweep invariant");
7981 8053 if (CMSTestInFreeList && fcInFreeLists) {
7982 8054 assert(_sp->verifyChunkInFreeLists(fc),
7983 8055 "free chunk is not in free lists");
7984 8056 }
7985 8057
7986 8058
7987 8059 if (CMSTraceSweeper) {
7988 8060 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
7989 8061 }
7990 8062
7991 8063 HeapWord* addr = (HeapWord*) fc;
7992 8064
7993 8065 bool coalesce;
7994 8066 size_t left = pointer_delta(addr, freeFinger());
7995 8067 size_t right = chunkSize;
7996 8068 switch (FLSCoalescePolicy) {
7997 8069 // numeric value forms a coalition aggressiveness metric
7998 8070 case 0: { // never coalesce
7999 8071 coalesce = false;
8000 8072 break;
8001 8073 }
8002 8074 case 1: { // coalesce if left & right chunks on overpopulated lists
8003 8075 coalesce = _sp->coalOverPopulated(left) &&
8004 8076 _sp->coalOverPopulated(right);
8005 8077 break;
8006 8078 }
8007 8079 case 2: { // coalesce if left chunk on overpopulated list (default)
8008 8080 coalesce = _sp->coalOverPopulated(left);
8009 8081 break;
8010 8082 }
8011 8083 case 3: { // coalesce if left OR right chunk on overpopulated list
8012 8084 coalesce = _sp->coalOverPopulated(left) ||
8013 8085 _sp->coalOverPopulated(right);
8014 8086 break;
8015 8087 }
8016 8088 case 4: { // always coalesce
8017 8089 coalesce = true;
8018 8090 break;
8019 8091 }
8020 8092 default:
8021 8093 ShouldNotReachHere();
8022 8094 }
8023 8095
8024 8096 // Should the current free range be coalesced?
8025 8097 // If the chunk is in a free range and either we decided to coalesce above
8026 8098 // or the chunk is near the large block at the end of the heap
8027 8099 // (isNearLargestChunk() returns true), then coalesce this chunk.
8028 8100 bool doCoalesce = inFreeRange() &&
8029 8101 (coalesce || _g->isNearLargestChunk((HeapWord*)fc));
8030 8102 if (doCoalesce) {
8031 8103 // Coalesce the current free range on the left with the new
8032 8104 // chunk on the right. If either is on a free list,
8033 8105 // it must be removed from the list and stashed in the closure.
8034 8106 if (freeRangeInFreeLists()) {
8035 8107 FreeChunk* ffc = (FreeChunk*)freeFinger();
8036 8108 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8037 8109 "Size of free range is inconsistent with chunk size.");
8038 8110 if (CMSTestInFreeList) {
8039 8111 assert(_sp->verifyChunkInFreeLists(ffc),
8040 8112 "Chunk is not in free lists");
8041 8113 }
8042 8114 _sp->coalDeath(ffc->size());
8043 8115 _sp->removeFreeChunkFromFreeLists(ffc);
8044 8116 set_freeRangeInFreeLists(false);
8045 8117 }
8046 8118 if (fcInFreeLists) {
8047 8119 _sp->coalDeath(chunkSize);
8048 8120 assert(fc->size() == chunkSize,
8049 8121 "The chunk has the wrong size or is not in the free lists");
8050 8122 _sp->removeFreeChunkFromFreeLists(fc);
8051 8123 }
8052 8124 set_lastFreeRangeCoalesced(true);
8053 8125 } else { // not in a free range and/or should not coalesce
8054 8126 // Return the current free range and start a new one.
8055 8127 if (inFreeRange()) {
8056 8128 // In a free range but cannot coalesce with the right hand chunk.
8057 8129 // Put the current free range into the free lists.
8058 8130 flushCurFreeChunk(freeFinger(),
8059 8131 pointer_delta(addr, freeFinger()));
8060 8132 }
8061 8133 // Set up for new free range. Pass along whether the right hand
8062 8134 // chunk is in the free lists.
8063 8135 initialize_free_range((HeapWord*)fc, fcInFreeLists);
8064 8136 }
8065 8137 }
8066 8138 void SweepClosure::flushCurFreeChunk(HeapWord* chunk, size_t size) {
8067 8139 assert(inFreeRange(), "Should only be called if currently in a free range.");
8068 8140 assert(size > 0,
8069 8141 "A zero sized chunk cannot be added to the free lists.");
8070 8142 if (!freeRangeInFreeLists()) {
8071 8143 if(CMSTestInFreeList) {
8072 8144 FreeChunk* fc = (FreeChunk*) chunk;
8073 8145 fc->setSize(size);
8074 8146 assert(!_sp->verifyChunkInFreeLists(fc),
8075 8147 "chunk should not be in free lists yet");
8076 8148 }
8077 8149 if (CMSTraceSweeper) {
8078 8150 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8079 8151 chunk, size);
8080 8152 }
8081 8153 // A new free range is going to be starting. The current
8082 8154 // free range has not been added to the free lists yet or
8083 8155 // was removed so add it back.
8084 8156 // If the current free range was coalesced, then the death
8085 8157 // of the free range was recorded. Record a birth now.
8086 8158 if (lastFreeRangeCoalesced()) {
8087 8159 _sp->coalBirth(size);
8088 8160 }
8089 8161 _sp->addChunkAndRepairOffsetTable(chunk, size,
8090 8162 lastFreeRangeCoalesced());
8091 8163 }
8092 8164 set_inFreeRange(false);
8093 8165 set_freeRangeInFreeLists(false);
8094 8166 }
8095 8167
8096 8168 // We take a break if we've been at this for a while,
8097 8169 // so as to avoid monopolizing the locks involved.
8098 8170 void SweepClosure::do_yield_work(HeapWord* addr) {
8099 8171 // Return current free chunk being used for coalescing (if any)
8100 8172 // to the appropriate freelist. After yielding, the next
8101 8173 // free block encountered will start a coalescing range of
8102 8174 // free blocks. If the next free block is adjacent to the
8103 8175 // chunk just flushed, they will need to wait for the next
8104 8176 // sweep to be coalesced.
8105 8177 if (inFreeRange()) {
8106 8178 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger()));
8107 8179 }
8108 8180
8109 8181 // First give up the locks, then yield, then re-lock.
8110 8182 // We should probably use a constructor/destructor idiom to
8111 8183 // do this unlock/lock or modify the MutexUnlocker class to
8112 8184 // serve our purpose. XXX
8113 8185 assert_lock_strong(_bitMap->lock());
8114 8186 assert_lock_strong(_freelistLock);
8115 8187 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8116 8188 "CMS thread should hold CMS token");
8117 8189 _bitMap->lock()->unlock();
8118 8190 _freelistLock->unlock();
8119 8191 ConcurrentMarkSweepThread::desynchronize(true);
8120 8192 ConcurrentMarkSweepThread::acknowledge_yield_request();
8121 8193 _collector->stopTimer();
8122 8194 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8123 8195 if (PrintCMSStatistics != 0) {
8124 8196 _collector->incrementYields();
8125 8197 }
8126 8198 _collector->icms_wait();
8127 8199
8128 8200 // See the comment in coordinator_yield()
8129 8201 for (unsigned i = 0; i < CMSYieldSleepCount &&
8130 8202 ConcurrentMarkSweepThread::should_yield() &&
8131 8203 !CMSCollector::foregroundGCIsActive(); ++i) {
8132 8204 os::sleep(Thread::current(), 1, false);
8133 8205 ConcurrentMarkSweepThread::acknowledge_yield_request();
8134 8206 }
8135 8207
8136 8208 ConcurrentMarkSweepThread::synchronize(true);
8137 8209 _freelistLock->lock();
8138 8210 _bitMap->lock()->lock_without_safepoint_check();
8139 8211 _collector->startTimer();
8140 8212 }
8141 8213
8142 8214 #ifndef PRODUCT
8143 8215 // This is actually very useful in a product build if it can
8144 8216 // be called from the debugger. Compile it into the product
8145 8217 // as needed.
8146 8218 bool debug_verifyChunkInFreeLists(FreeChunk* fc) {
8147 8219 return debug_cms_space->verifyChunkInFreeLists(fc);
8148 8220 }
8149 8221
8150 8222 void SweepClosure::record_free_block_coalesced(FreeChunk* fc) const {
8151 8223 if (CMSTraceSweeper) {
8152 8224 gclog_or_tty->print("Sweep:coal_free_blk 0x%x (%d)\n", fc, fc->size());
8153 8225 }
8154 8226 }
8155 8227 #endif
8156 8228
8157 8229 // CMSIsAliveClosure
8158 8230 bool CMSIsAliveClosure::do_object_b(oop obj) {
8159 8231 HeapWord* addr = (HeapWord*)obj;
8160 8232 return addr != NULL &&
8161 8233 (!_span.contains(addr) || _bit_map->isMarked(addr));
8162 8234 }
8163 8235
8164 8236 // CMSKeepAliveClosure: the serial version
8165 8237 void CMSKeepAliveClosure::do_oop(oop* p) {
8166 8238 oop this_oop = *p;
8167 8239 HeapWord* addr = (HeapWord*)this_oop;
8168 8240 if (_span.contains(addr) &&
8169 8241 !_bit_map->isMarked(addr)) {
8170 8242 _bit_map->mark(addr);
8171 8243 bool simulate_overflow = false;
8172 8244 NOT_PRODUCT(
8173 8245 if (CMSMarkStackOverflowALot &&
8174 8246 _collector->simulate_overflow()) {
8175 8247 // simulate a stack overflow
8176 8248 simulate_overflow = true;
8177 8249 }
8178 8250 )
8179 8251 if (simulate_overflow || !_mark_stack->push(this_oop)) {
8180 8252 _collector->push_on_overflow_list(this_oop);
8181 8253 _collector->_ser_kac_ovflw++;
8182 8254 }
8183 8255 }
8184 8256 }
8185 8257
8186 8258 // CMSParKeepAliveClosure: a parallel version of the above.
8187 8259 // The work queues are private to each closure (thread),
8188 8260 // but (may be) available for stealing by other threads.
8189 8261 void CMSParKeepAliveClosure::do_oop(oop* p) {
8190 8262 oop this_oop = *p;
8191 8263 HeapWord* addr = (HeapWord*)this_oop;
8192 8264 if (_span.contains(addr) &&
8193 8265 !_bit_map->isMarked(addr)) {
8194 8266 // In general, during recursive tracing, several threads
8195 8267 // may be concurrently getting here; the first one to
8196 8268 // "tag" it, claims it.
8197 8269 if (_bit_map->par_mark(addr)) {
8198 8270 bool res = _work_queue->push(this_oop);
8199 8271 assert(res, "Low water mark should be much less than capacity");
8200 8272 // Do a recursive trim in the hope that this will keep
8201 8273 // stack usage lower, but leave some oops for potential stealers
8202 8274 trim_queue(_low_water_mark);
8203 8275 } // Else, another thread got there first
8204 8276 }
8205 8277 }
8206 8278
8207 8279 void CMSParKeepAliveClosure::trim_queue(uint max) {
8208 8280 while (_work_queue->size() > max) {
8209 8281 oop new_oop;
8210 8282 if (_work_queue->pop_local(new_oop)) {
8211 8283 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8212 8284 assert(_bit_map->isMarked((HeapWord*)new_oop),
8213 8285 "no white objects on this stack!");
8214 8286 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8215 8287 // iterate over the oops in this oop, marking and pushing
8216 8288 // the ones in CMS heap (i.e. in _span).
8217 8289 new_oop->oop_iterate(&_mark_and_push);
8218 8290 }
8219 8291 }
8220 8292 }
8221 8293
8222 8294 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) {
8223 8295 oop this_oop = *p;
8224 8296 HeapWord* addr = (HeapWord*)this_oop;
8225 8297 if (_span.contains(addr) &&
8226 8298 !_bit_map->isMarked(addr)) {
8227 8299 if (_bit_map->par_mark(addr)) {
8228 8300 bool simulate_overflow = false;
8229 8301 NOT_PRODUCT(
8230 8302 if (CMSMarkStackOverflowALot &&
8231 8303 _collector->par_simulate_overflow()) {
8232 8304 // simulate a stack overflow
8233 8305 simulate_overflow = true;
8234 8306 }
8235 8307 )
8236 8308 if (simulate_overflow || !_work_queue->push(this_oop)) {
8237 8309 _collector->par_push_on_overflow_list(this_oop);
8238 8310 _collector->_par_kac_ovflw++;
8239 8311 }
8240 8312 } // Else another thread got there already
8241 8313 }
8242 8314 }
8243 8315
8244 8316 //////////////////////////////////////////////////////////////////
8245 8317 // CMSExpansionCause /////////////////////////////
8246 8318 //////////////////////////////////////////////////////////////////
8247 8319 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8248 8320 switch (cause) {
8249 8321 case _no_expansion:
8250 8322 return "No expansion";
8251 8323 case _satisfy_free_ratio:
8252 8324 return "Free ratio";
8253 8325 case _satisfy_promotion:
8254 8326 return "Satisfy promotion";
8255 8327 case _satisfy_allocation:
8256 8328 return "allocation";
8257 8329 case _allocate_par_lab:
8258 8330 return "Par LAB";
8259 8331 case _allocate_par_spooling_space:
8260 8332 return "Par Spooling Space";
8261 8333 case _adaptive_size_policy:
8262 8334 return "Ergonomics";
8263 8335 default:
8264 8336 return "unknown";
8265 8337 }
8266 8338 }
8267 8339
8268 8340 void CMSDrainMarkingStackClosure::do_void() {
8269 8341 // the max number to take from overflow list at a time
8270 8342 const size_t num = _mark_stack->capacity()/4;
8271 8343 while (!_mark_stack->isEmpty() ||
8272 8344 // if stack is empty, check the overflow list
8273 8345 _collector->take_from_overflow_list(num, _mark_stack)) {
8274 8346 oop this_oop = _mark_stack->pop();
8275 8347 HeapWord* addr = (HeapWord*)this_oop;
8276 8348 assert(_span.contains(addr), "Should be within span");
8277 8349 assert(_bit_map->isMarked(addr), "Should be marked");
8278 8350 assert(this_oop->is_oop(), "Should be an oop");
8279 8351 this_oop->oop_iterate(_keep_alive);
8280 8352 }
8281 8353 }
8282 8354
8283 8355 void CMSParDrainMarkingStackClosure::do_void() {
8284 8356 // drain queue
8285 8357 trim_queue(0);
8286 8358 }
8287 8359
8288 8360 // Trim our work_queue so its length is below max at return
8289 8361 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8290 8362 while (_work_queue->size() > max) {
8291 8363 oop new_oop;
8292 8364 if (_work_queue->pop_local(new_oop)) {
8293 8365 assert(new_oop->is_oop(), "Expected an oop");
8294 8366 assert(_bit_map->isMarked((HeapWord*)new_oop),
8295 8367 "no white objects on this stack!");
8296 8368 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8297 8369 // iterate over the oops in this oop, marking and pushing
8298 8370 // the ones in CMS heap (i.e. in _span).
8299 8371 new_oop->oop_iterate(&_mark_and_push);
8300 8372 }
8301 8373 }
8302 8374 }
8303 8375
8304 8376 ////////////////////////////////////////////////////////////////////
8305 8377 // Support for Marking Stack Overflow list handling and related code
8306 8378 ////////////////////////////////////////////////////////////////////
8307 8379 // Much of the following code is similar in shape and spirit to the
8308 8380 // code used in ParNewGC. We should try and share that code
8309 8381 // as much as possible in the future.
8310 8382
8311 8383 #ifndef PRODUCT
8312 8384 // Debugging support for CMSStackOverflowALot
8313 8385
8314 8386 // It's OK to call this multi-threaded; the worst thing
8315 8387 // that can happen is that we'll get a bunch of closely
8316 8388 // spaced simulated oveflows, but that's OK, in fact
8317 8389 // probably good as it would exercise the overflow code
8318 8390 // under contention.
8319 8391 bool CMSCollector::simulate_overflow() {
8320 8392 if (_overflow_counter-- <= 0) { // just being defensive
8321 8393 _overflow_counter = CMSMarkStackOverflowInterval;
8322 8394 return true;
8323 8395 } else {
8324 8396 return false;
8325 8397 }
8326 8398 }
8327 8399
8328 8400 bool CMSCollector::par_simulate_overflow() {
8329 8401 return simulate_overflow();
8330 8402 }
8331 8403 #endif
8332 8404
8333 8405 // Single-threaded
8334 8406 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8335 8407 assert(stack->isEmpty(), "Expected precondition");
8336 8408 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8337 8409 size_t i = num;
8338 8410 oop cur = _overflow_list;
8339 8411 const markOop proto = markOopDesc::prototype();
8340 8412 NOT_PRODUCT(size_t n = 0;)
8341 8413 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8342 8414 next = oop(cur->mark());
8343 8415 cur->set_mark(proto); // until proven otherwise
8344 8416 assert(cur->is_oop(), "Should be an oop");
8345 8417 bool res = stack->push(cur);
8346 8418 assert(res, "Bit off more than can chew?");
8347 8419 NOT_PRODUCT(n++;)
8348 8420 }
8349 8421 _overflow_list = cur;
8350 8422 #ifndef PRODUCT
8351 8423 assert(_num_par_pushes >= n, "Too many pops?");
8352 8424 _num_par_pushes -=n;
8353 8425 #endif
8354 8426 return !stack->isEmpty();
8355 8427 }
8356 8428
8357 8429 // Multi-threaded; use CAS to break off a prefix
8358 8430 bool CMSCollector::par_take_from_overflow_list(size_t num,
8359 8431 OopTaskQueue* work_q) {
8360 8432 assert(work_q->size() == 0, "That's the current policy");
8361 8433 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8362 8434 if (_overflow_list == NULL) {
8363 8435 return false;
8364 8436 }
8365 8437 // Grab the entire list; we'll put back a suffix
8366 8438 oop prefix = (oop)Atomic::xchg_ptr(NULL, &_overflow_list);
8367 8439 if (prefix == NULL) { // someone grabbed it before we did ...
8368 8440 // ... we could spin for a short while, but for now we don't
8369 8441 return false;
8370 8442 }
8371 8443 size_t i = num;
8372 8444 oop cur = prefix;
8373 8445 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8374 8446 if (cur->mark() != NULL) {
8375 8447 oop suffix_head = cur->mark(); // suffix will be put back on global list
8376 8448 cur->set_mark(NULL); // break off suffix
8377 8449 // Find tail of suffix so we can prepend suffix to global list
8378 8450 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8379 8451 oop suffix_tail = cur;
8380 8452 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8381 8453 "Tautology");
8382 8454 oop observed_overflow_list = _overflow_list;
8383 8455 do {
8384 8456 cur = observed_overflow_list;
8385 8457 suffix_tail->set_mark(markOop(cur));
8386 8458 observed_overflow_list =
8387 8459 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur);
8388 8460 } while (cur != observed_overflow_list);
8389 8461 }
8390 8462
8391 8463 // Push the prefix elements on work_q
8392 8464 assert(prefix != NULL, "control point invariant");
8393 8465 const markOop proto = markOopDesc::prototype();
8394 8466 oop next;
8395 8467 NOT_PRODUCT(size_t n = 0;)
8396 8468 for (cur = prefix; cur != NULL; cur = next) {
8397 8469 next = oop(cur->mark());
8398 8470 cur->set_mark(proto); // until proven otherwise
8399 8471 assert(cur->is_oop(), "Should be an oop");
8400 8472 bool res = work_q->push(cur);
8401 8473 assert(res, "Bit off more than we can chew?");
8402 8474 NOT_PRODUCT(n++;)
8403 8475 }
8404 8476 #ifndef PRODUCT
8405 8477 assert(_num_par_pushes >= n, "Too many pops?");
8406 8478 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8407 8479 #endif
8408 8480 return true;
8409 8481 }
8410 8482
8411 8483 // Single-threaded
8412 8484 void CMSCollector::push_on_overflow_list(oop p) {
8413 8485 NOT_PRODUCT(_num_par_pushes++;)
8414 8486 assert(p->is_oop(), "Not an oop");
8415 8487 preserve_mark_if_necessary(p);
8416 8488 p->set_mark((markOop)_overflow_list);
8417 8489 _overflow_list = p;
8418 8490 }
8419 8491
8420 8492 // Multi-threaded; use CAS to prepend to overflow list
8421 8493 void CMSCollector::par_push_on_overflow_list(oop p) {
8422 8494 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8423 8495 assert(p->is_oop(), "Not an oop");
8424 8496 par_preserve_mark_if_necessary(p);
8425 8497 oop observed_overflow_list = _overflow_list;
8426 8498 oop cur_overflow_list;
8427 8499 do {
8428 8500 cur_overflow_list = observed_overflow_list;
8429 8501 p->set_mark(markOop(cur_overflow_list));
8430 8502 observed_overflow_list =
8431 8503 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8432 8504 } while (cur_overflow_list != observed_overflow_list);
8433 8505 }
8434 8506
8435 8507 // Single threaded
8436 8508 // General Note on GrowableArray: pushes may silently fail
8437 8509 // because we are (temporarily) out of C-heap for expanding
8438 8510 // the stack. The problem is quite ubiquitous and affects
8439 8511 // a lot of code in the JVM. The prudent thing for GrowableArray
8440 8512 // to do (for now) is to exit with an error. However, that may
8441 8513 // be too draconian in some cases because the caller may be
8442 8514 // able to recover without much harm. For suych cases, we
8443 8515 // should probably introduce a "soft_push" method which returns
8444 8516 // an indication of success or failure with the assumption that
8445 8517 // the caller may be able to recover from a failure; code in
8446 8518 // the VM can then be changed, incrementally, to deal with such
8447 8519 // failures where possible, thus, incrementally hardening the VM
8448 8520 // in such low resource situations.
8449 8521 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8450 8522 int PreserveMarkStackSize = 128;
8451 8523
8452 8524 if (_preserved_oop_stack == NULL) {
8453 8525 assert(_preserved_mark_stack == NULL,
8454 8526 "bijection with preserved_oop_stack");
8455 8527 // Allocate the stacks
8456 8528 _preserved_oop_stack = new (ResourceObj::C_HEAP)
8457 8529 GrowableArray<oop>(PreserveMarkStackSize, true);
8458 8530 _preserved_mark_stack = new (ResourceObj::C_HEAP)
8459 8531 GrowableArray<markOop>(PreserveMarkStackSize, true);
8460 8532 if (_preserved_oop_stack == NULL || _preserved_mark_stack == NULL) {
8461 8533 vm_exit_out_of_memory(2* PreserveMarkStackSize * sizeof(oop) /* punt */,
8462 8534 "Preserved Mark/Oop Stack for CMS (C-heap)");
8463 8535 }
8464 8536 }
8465 8537 _preserved_oop_stack->push(p);
8466 8538 _preserved_mark_stack->push(m);
8467 8539 assert(m == p->mark(), "Mark word changed");
8468 8540 assert(_preserved_oop_stack->length() == _preserved_mark_stack->length(),
8469 8541 "bijection");
8470 8542 }
8471 8543
8472 8544 // Single threaded
8473 8545 void CMSCollector::preserve_mark_if_necessary(oop p) {
8474 8546 markOop m = p->mark();
8475 8547 if (m->must_be_preserved(p)) {
8476 8548 preserve_mark_work(p, m);
8477 8549 }
8478 8550 }
8479 8551
8480 8552 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8481 8553 markOop m = p->mark();
8482 8554 if (m->must_be_preserved(p)) {
8483 8555 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8484 8556 // Even though we read the mark word without holding
8485 8557 // the lock, we are assured that it will not change
8486 8558 // because we "own" this oop, so no other thread can
8487 8559 // be trying to push it on the overflow list; see
8488 8560 // the assertion in preserve_mark_work() that checks
8489 8561 // that m == p->mark().
8490 8562 preserve_mark_work(p, m);
8491 8563 }
8492 8564 }
8493 8565
8494 8566 // We should be able to do this multi-threaded,
8495 8567 // a chunk of stack being a task (this is
8496 8568 // correct because each oop only ever appears
8497 8569 // once in the overflow list. However, it's
8498 8570 // not very easy to completely overlap this with
8499 8571 // other operations, so will generally not be done
8500 8572 // until all work's been completed. Because we
8501 8573 // expect the preserved oop stack (set) to be small,
8502 8574 // it's probably fine to do this single-threaded.
8503 8575 // We can explore cleverer concurrent/overlapped/parallel
8504 8576 // processing of preserved marks if we feel the
8505 8577 // need for this in the future. Stack overflow should
8506 8578 // be so rare in practice and, when it happens, its
8507 8579 // effect on performance so great that this will
8508 8580 // likely just be in the noise anyway.
8509 8581 void CMSCollector::restore_preserved_marks_if_any() {
8510 8582 if (_preserved_oop_stack == NULL) {
8511 8583 assert(_preserved_mark_stack == NULL,
8512 8584 "bijection with preserved_oop_stack");
8513 8585 return;
8514 8586 }
8515 8587
8516 8588 assert(SafepointSynchronize::is_at_safepoint(),
8517 8589 "world should be stopped");
8518 8590 assert(Thread::current()->is_ConcurrentGC_thread() ||
8519 8591 Thread::current()->is_VM_thread(),
8520 8592 "should be single-threaded");
8521 8593
8522 8594 int length = _preserved_oop_stack->length();
8523 8595 assert(_preserved_mark_stack->length() == length, "bijection");
8524 8596 for (int i = 0; i < length; i++) {
8525 8597 oop p = _preserved_oop_stack->at(i);
8526 8598 assert(p->is_oop(), "Should be an oop");
8527 8599 assert(_span.contains(p), "oop should be in _span");
8528 8600 assert(p->mark() == markOopDesc::prototype(),
8529 8601 "Set when taken from overflow list");
8530 8602 markOop m = _preserved_mark_stack->at(i);
8531 8603 p->set_mark(m);
8532 8604 }
8533 8605 _preserved_mark_stack->clear();
8534 8606 _preserved_oop_stack->clear();
8535 8607 assert(_preserved_mark_stack->is_empty() &&
8536 8608 _preserved_oop_stack->is_empty(),
8537 8609 "stacks were cleared above");
8538 8610 }
8539 8611
8540 8612 #ifndef PRODUCT
8541 8613 bool CMSCollector::no_preserved_marks() const {
8542 8614 return ( ( _preserved_mark_stack == NULL
8543 8615 && _preserved_oop_stack == NULL)
8544 8616 || ( _preserved_mark_stack->is_empty()
8545 8617 && _preserved_oop_stack->is_empty()));
8546 8618 }
8547 8619 #endif
8548 8620
8549 8621 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
8550 8622 {
8551 8623 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
8552 8624 CMSAdaptiveSizePolicy* size_policy =
8553 8625 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
8554 8626 assert(size_policy->is_gc_cms_adaptive_size_policy(),
8555 8627 "Wrong type for size policy");
8556 8628 return size_policy;
8557 8629 }
8558 8630
8559 8631 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
8560 8632 size_t desired_promo_size) {
8561 8633 if (cur_promo_size < desired_promo_size) {
8562 8634 size_t expand_bytes = desired_promo_size - cur_promo_size;
8563 8635 if (PrintAdaptiveSizePolicy && Verbose) {
8564 8636 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
8565 8637 "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
8566 8638 expand_bytes);
8567 8639 }
8568 8640 expand(expand_bytes,
8569 8641 MinHeapDeltaBytes,
8570 8642 CMSExpansionCause::_adaptive_size_policy);
8571 8643 } else if (desired_promo_size < cur_promo_size) {
8572 8644 size_t shrink_bytes = cur_promo_size - desired_promo_size;
8573 8645 if (PrintAdaptiveSizePolicy && Verbose) {
8574 8646 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
8575 8647 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
8576 8648 shrink_bytes);
8577 8649 }
8578 8650 shrink(shrink_bytes);
8579 8651 }
8580 8652 }
8581 8653
8582 8654 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
8583 8655 GenCollectedHeap* gch = GenCollectedHeap::heap();
8584 8656 CMSGCAdaptivePolicyCounters* counters =
8585 8657 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
8586 8658 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
8587 8659 "Wrong kind of counters");
8588 8660 return counters;
8589 8661 }
8590 8662
8591 8663
8592 8664 void ASConcurrentMarkSweepGeneration::update_counters() {
8593 8665 if (UsePerfData) {
8594 8666 _space_counters->update_all();
8595 8667 _gen_counters->update_all();
8596 8668 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8597 8669 GenCollectedHeap* gch = GenCollectedHeap::heap();
8598 8670 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
8599 8671 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
8600 8672 "Wrong gc statistics type");
8601 8673 counters->update_counters(gc_stats_l);
8602 8674 }
8603 8675 }
8604 8676
8605 8677 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
8606 8678 if (UsePerfData) {
8607 8679 _space_counters->update_used(used);
8608 8680 _space_counters->update_capacity();
8609 8681 _gen_counters->update_all();
8610 8682
8611 8683 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8612 8684 GenCollectedHeap* gch = GenCollectedHeap::heap();
8613 8685 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
8614 8686 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
8615 8687 "Wrong gc statistics type");
8616 8688 counters->update_counters(gc_stats_l);
8617 8689 }
8618 8690 }
8619 8691
8620 8692 // The desired expansion delta is computed so that:
8621 8693 // . desired free percentage or greater is used
8622 8694 void ASConcurrentMarkSweepGeneration::compute_new_size() {
8623 8695 assert_locked_or_safepoint(Heap_lock);
8624 8696
8625 8697 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
8626 8698
8627 8699 // If incremental collection failed, we just want to expand
8628 8700 // to the limit.
8629 8701 if (incremental_collection_failed()) {
8630 8702 clear_incremental_collection_failed();
8631 8703 grow_to_reserved();
8632 8704 return;
8633 8705 }
8634 8706
8635 8707 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing");
8636 8708
8637 8709 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
8638 8710 "Wrong type of heap");
8639 8711 int prev_level = level() - 1;
8640 8712 assert(prev_level >= 0, "The cms generation is the lowest generation");
8641 8713 Generation* prev_gen = gch->get_gen(prev_level);
8642 8714 assert(prev_gen->kind() == Generation::ASParNew,
8643 8715 "Wrong type of young generation");
8644 8716 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen;
8645 8717 size_t cur_eden = younger_gen->eden()->capacity();
8646 8718 CMSAdaptiveSizePolicy* size_policy = cms_size_policy();
8647 8719 size_t cur_promo = free();
8648 8720 size_policy->compute_tenured_generation_free_space(cur_promo,
8649 8721 max_available(),
8650 8722 cur_eden);
8651 8723 resize(cur_promo, size_policy->promo_size());
8652 8724
8653 8725 // Record the new size of the space in the cms generation
8654 8726 // that is available for promotions. This is temporary.
8655 8727 // It should be the desired promo size.
8656 8728 size_policy->avg_cms_promo()->sample(free());
8657 8729 size_policy->avg_old_live()->sample(used());
8658 8730
8659 8731 if (UsePerfData) {
8660 8732 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8661 8733 counters->update_cms_capacity_counter(capacity());
8662 8734 }
8663 8735 }
8664 8736
8665 8737 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
8666 8738 assert_locked_or_safepoint(Heap_lock);
8667 8739 assert_lock_strong(freelistLock());
8668 8740 HeapWord* old_end = _cmsSpace->end();
8669 8741 HeapWord* unallocated_start = _cmsSpace->unallocated_block();
8670 8742 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
8671 8743 FreeChunk* chunk_at_end = find_chunk_at_end();
8672 8744 if (chunk_at_end == NULL) {
8673 8745 // No room to shrink
8674 8746 if (PrintGCDetails && Verbose) {
8675 8747 gclog_or_tty->print_cr("No room to shrink: old_end "
8676 8748 PTR_FORMAT " unallocated_start " PTR_FORMAT
8677 8749 " chunk_at_end " PTR_FORMAT,
8678 8750 old_end, unallocated_start, chunk_at_end);
8679 8751 }
8680 8752 return;
8681 8753 } else {
8682 8754
8683 8755 // Find the chunk at the end of the space and determine
8684 8756 // how much it can be shrunk.
8685 8757 size_t shrinkable_size_in_bytes = chunk_at_end->size();
8686 8758 size_t aligned_shrinkable_size_in_bytes =
8687 8759 align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
8688 8760 assert(unallocated_start <= chunk_at_end->end(),
8689 8761 "Inconsistent chunk at end of space");
8690 8762 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
8691 8763 size_t word_size_before = heap_word_size(_virtual_space.committed_size());
8692 8764
8693 8765 // Shrink the underlying space
8694 8766 _virtual_space.shrink_by(bytes);
8695 8767 if (PrintGCDetails && Verbose) {
8696 8768 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
8697 8769 " desired_bytes " SIZE_FORMAT
8698 8770 " shrinkable_size_in_bytes " SIZE_FORMAT
8699 8771 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
8700 8772 " bytes " SIZE_FORMAT,
8701 8773 desired_bytes, shrinkable_size_in_bytes,
8702 8774 aligned_shrinkable_size_in_bytes, bytes);
8703 8775 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT
8704 8776 " unallocated_start " SIZE_FORMAT,
8705 8777 old_end, unallocated_start);
8706 8778 }
8707 8779
8708 8780 // If the space did shrink (shrinking is not guaranteed),
8709 8781 // shrink the chunk at the end by the appropriate amount.
8710 8782 if (((HeapWord*)_virtual_space.high()) < old_end) {
8711 8783 size_t new_word_size =
8712 8784 heap_word_size(_virtual_space.committed_size());
8713 8785
8714 8786 // Have to remove the chunk from the dictionary because it is changing
8715 8787 // size and might be someplace elsewhere in the dictionary.
8716 8788
8717 8789 // Get the chunk at end, shrink it, and put it
8718 8790 // back.
8719 8791 _cmsSpace->removeChunkFromDictionary(chunk_at_end);
8720 8792 size_t word_size_change = word_size_before - new_word_size;
8721 8793 size_t chunk_at_end_old_size = chunk_at_end->size();
8722 8794 assert(chunk_at_end_old_size >= word_size_change,
8723 8795 "Shrink is too large");
8724 8796 chunk_at_end->setSize(chunk_at_end_old_size -
8725 8797 word_size_change);
8726 8798 _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
8727 8799 word_size_change);
8728 8800
8729 8801 _cmsSpace->returnChunkToDictionary(chunk_at_end);
8730 8802
8731 8803 MemRegion mr(_cmsSpace->bottom(), new_word_size);
8732 8804 _bts->resize(new_word_size); // resize the block offset shared array
8733 8805 Universe::heap()->barrier_set()->resize_covered_region(mr);
8734 8806 _cmsSpace->assert_locked();
8735 8807 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
8736 8808
8737 8809 NOT_PRODUCT(_cmsSpace->dictionary()->verify());
8738 8810
8739 8811 // update the space and generation capacity counters
8740 8812 if (UsePerfData) {
8741 8813 _space_counters->update_capacity();
8742 8814 _gen_counters->update_all();
8743 8815 }
8744 8816
8745 8817 if (Verbose && PrintGCDetails) {
8746 8818 size_t new_mem_size = _virtual_space.committed_size();
8747 8819 size_t old_mem_size = new_mem_size + bytes;
8748 8820 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK",
8749 8821 name(), old_mem_size/K, bytes/K, new_mem_size/K);
8750 8822 }
8751 8823 }
8752 8824
8753 8825 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
8754 8826 "Inconsistency at end of space");
8755 8827 assert(chunk_at_end->end() == _cmsSpace->end(),
8756 8828 "Shrinking is inconsistent");
8757 8829 return;
8758 8830 }
8759 8831 }
8760 8832
8761 8833 // Transfer some number of overflown objects to usual marking
8762 8834 // stack. Return true if some objects were transferred.
8763 8835 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8764 8836 size_t num = MIN2((size_t)_mark_stack->capacity()/4,
8765 8837 (size_t)ParGCDesiredObjsFromOverflowList);
8766 8838
8767 8839 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8768 8840 assert(_collector->overflow_list_is_empty() || res,
8769 8841 "If list is not empty, we should have taken something");
8770 8842 assert(!res || !_mark_stack->isEmpty(),
8771 8843 "If we took something, it should now be on our stack");
8772 8844 return res;
8773 8845 }
8774 8846
8775 8847 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8776 8848 size_t res = _sp->block_size_no_stall(addr, _collector);
8777 8849 assert(res != 0, "Should always be able to compute a size");
8778 8850 if (_sp->block_is_obj(addr)) {
8779 8851 if (_live_bit_map->isMarked(addr)) {
8780 8852 // It can't have been dead in a previous cycle
8781 8853 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8782 8854 } else {
8783 8855 _dead_bit_map->mark(addr); // mark the dead object
8784 8856 }
8785 8857 }
8786 8858 return res;
8787 8859 }
↓ open down ↓ |
818 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX