1 /*
2 * Copyright (c) 2001, 2020, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/g1/g1Analytics.hpp"
27 #include "gc/g1/g1Arguments.hpp"
28 #include "gc/g1/g1CollectedHeap.inline.hpp"
29 #include "gc/g1/g1CollectionSet.hpp"
30 #include "gc/g1/g1CollectionSetCandidates.hpp"
31 #include "gc/g1/g1ConcurrentMark.hpp"
32 #include "gc/g1/g1ConcurrentMarkThread.inline.hpp"
33 #include "gc/g1/g1ConcurrentRefine.hpp"
34 #include "gc/g1/g1ConcurrentRefineStats.hpp"
35 #include "gc/g1/g1CollectionSetChooser.hpp"
36 #include "gc/g1/g1HeterogeneousHeapPolicy.hpp"
37 #include "gc/g1/g1HotCardCache.hpp"
38 #include "gc/g1/g1IHOPControl.hpp"
39 #include "gc/g1/g1GCPhaseTimes.hpp"
40 #include "gc/g1/g1Policy.hpp"
41 #include "gc/g1/g1SurvivorRegions.hpp"
42 #include "gc/g1/g1YoungGenSizer.hpp"
43 #include "gc/g1/heapRegion.inline.hpp"
44 #include "gc/g1/heapRegionRemSet.hpp"
45 #include "gc/shared/concurrentGCBreakpoints.hpp"
46 #include "gc/shared/gcPolicyCounters.hpp"
47 #include "logging/log.hpp"
48 #include "runtime/arguments.hpp"
49 #include "runtime/globals.hpp"
50 #include "runtime/java.hpp"
51 #include "runtime/mutexLocker.hpp"
52 #include "utilities/debug.hpp"
53 #include "utilities/growableArray.hpp"
54 #include "utilities/pair.hpp"
55
56 G1Policy::G1Policy(STWGCTimer* gc_timer) :
57 _predictor(G1ConfidencePercent / 100.0),
58 _analytics(new G1Analytics(&_predictor)),
59 _remset_tracker(),
60 _mmu_tracker(new G1MMUTrackerQueue(GCPauseIntervalMillis / 1000.0, MaxGCPauseMillis / 1000.0)),
61 _ihop_control(create_ihop_control(&_predictor)),
62 _policy_counters(new GCPolicyCounters("GarbageFirst", 1, 2)),
63 _full_collection_start_sec(0.0),
64 _collection_pause_end_millis(os::javaTimeNanos() / NANOSECS_PER_MILLISEC),
65 _young_list_desired_length(0),
66 _young_list_target_length(0),
67 _young_list_max_length(0),
68 _eden_surv_rate_group(new G1SurvRateGroup()),
69 _survivor_surv_rate_group(new G1SurvRateGroup()),
70 _reserve_factor((double) G1ReservePercent / 100.0),
71 _reserve_regions(0),
72 _young_gen_sizer(G1YoungGenSizer::create_gen_sizer()),
73 _free_regions_at_end_of_collection(0),
74 _rs_length(0),
75 _rs_length_prediction(0),
76 _pending_cards_at_gc_start(0),
77 _old_gen_alloc_tracker(),
78 _initial_mark_to_mixed(),
79 _collection_set(NULL),
80 _g1h(NULL),
81 _phase_times(new G1GCPhaseTimes(gc_timer, ParallelGCThreads)),
82 _mark_remark_start_sec(0),
83 _mark_cleanup_start_sec(0),
84 _tenuring_threshold(MaxTenuringThreshold),
85 _max_survivor_regions(0),
86 _survivors_age_table(true)
87 {
88 }
89
90 G1Policy::~G1Policy() {
91 delete _ihop_control;
92 delete _young_gen_sizer;
93 }
94
95 G1Policy* G1Policy::create_policy(STWGCTimer* gc_timer_stw) {
96 if (G1Arguments::is_heterogeneous_heap()) {
97 return new G1HeterogeneousHeapPolicy(gc_timer_stw);
98 } else {
99 return new G1Policy(gc_timer_stw);
100 }
101 }
102
103 G1CollectorState* G1Policy::collector_state() const { return _g1h->collector_state(); }
104
105 void G1Policy::init(G1CollectedHeap* g1h, G1CollectionSet* collection_set) {
106 _g1h = g1h;
107 _collection_set = collection_set;
108
109 assert(Heap_lock->owned_by_self(), "Locking discipline.");
110
111 _young_gen_sizer->adjust_max_new_size(_g1h->max_expandable_regions());
112
113 _free_regions_at_end_of_collection = _g1h->num_free_regions();
114
115 update_young_length_bounds();
116 // We may immediately start allocating regions and placing them on the
117 // collection set list. Initialize the per-collection set info
118 _collection_set->start_incremental_building();
119 }
120
121 void G1Policy::note_gc_start() {
122 phase_times()->note_gc_start();
123 }
124
125 class G1YoungLengthPredictor {
126 const double _base_time_ms;
127 const double _base_free_regions;
128 const double _target_pause_time_ms;
129 const G1Policy* const _policy;
130
131 public:
132 G1YoungLengthPredictor(double base_time_ms,
133 double base_free_regions,
134 double target_pause_time_ms,
135 const G1Policy* policy) :
136 _base_time_ms(base_time_ms),
137 _base_free_regions(base_free_regions),
138 _target_pause_time_ms(target_pause_time_ms),
139 _policy(policy) {}
140
141 bool will_fit(uint young_length) const {
142 if (young_length >= _base_free_regions) {
143 // end condition 1: not enough space for the young regions
144 return false;
145 }
146
147 size_t bytes_to_copy = 0;
148 const double copy_time_ms = _policy->predict_eden_copy_time_ms(young_length, &bytes_to_copy);
149 const double young_other_time_ms = _policy->analytics()->predict_young_other_time_ms(young_length);
150 const double pause_time_ms = _base_time_ms + copy_time_ms + young_other_time_ms;
151 if (pause_time_ms > _target_pause_time_ms) {
152 // end condition 2: prediction is over the target pause time
153 return false;
154 }
155
156 const size_t free_bytes = (_base_free_regions - young_length) * HeapRegion::GrainBytes;
157
158 // When copying, we will likely need more bytes free than is live in the region.
159 // Add some safety margin to factor in the confidence of our guess, and the
160 // natural expected waste.
161 // (100.0 / G1ConfidencePercent) is a scale factor that expresses the uncertainty
162 // of the calculation: the lower the confidence, the more headroom.
163 // (100 + TargetPLABWastePct) represents the increase in expected bytes during
164 // copying due to anticipated waste in the PLABs.
165 const double safety_factor = (100.0 / G1ConfidencePercent) * (100 + TargetPLABWastePct) / 100.0;
166 const size_t expected_bytes_to_copy = (size_t)(safety_factor * bytes_to_copy);
167
168 if (expected_bytes_to_copy > free_bytes) {
169 // end condition 3: out-of-space
170 return false;
171 }
172
173 // success!
174 return true;
175 }
176 };
177
178 void G1Policy::record_new_heap_size(uint new_number_of_regions) {
179 // re-calculate the necessary reserve
180 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor;
181 // We use ceiling so that if reserve_regions_d is > 0.0 (but
182 // smaller than 1.0) we'll get 1.
183 _reserve_regions = (uint) ceil(reserve_regions_d);
184
185 _young_gen_sizer->heap_size_changed(new_number_of_regions);
186
187 _ihop_control->update_target_occupancy(new_number_of_regions * HeapRegion::GrainBytes);
188 }
189
190 uint G1Policy::calculate_desired_eden_length_by_mmu() const {
191 // One could argue that any useful eden length to keep any MMU would be 1, but
192 // in theory this is possible. Other constraints enforce a minimum eden of 1
193 // anyway.
194 uint desired_min_length = 0;
195 if (use_adaptive_young_list_length()) {
196 double now_sec = os::elapsedTime();
197 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
198 double alloc_rate_ms = _analytics->predict_alloc_rate_ms();
199 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms);
200 }
201 return desired_min_length;
202 }
203
204 void G1Policy::update_young_length_bounds() {
205 update_young_length_bounds(_analytics->predict_rs_length());
206 }
207
208 void G1Policy::update_young_length_bounds(size_t rs_length) {
209 _young_list_desired_length = calculate_young_desired_length(rs_length);
210 _young_list_target_length = calculate_young_target_length(_young_list_desired_length);
211 _young_list_max_length = calculate_young_max_length(_young_list_target_length);
212
213 log_debug(gc,ergo,heap)("Young list lengths: desired: %u, target: %u, max: %u",
214 _young_list_desired_length,
215 _young_list_target_length,
216 _young_list_max_length);
217 }
218
219 // Calculates desired young gen length. It is calculated from:
220 //
221 // - sizer min/max bounds on young gen
222 // - pause time goal for whole young gen evacuation
223 // - MMU goal influencing eden to make GCs spaced apart.
224 // - a minimum one eden region length.
225 //
226 // We may enter with already allocated eden and survivor regions, that may be
227 // higher than the maximum, or the above goals may result in a desired value
228 // smaller than are already allocated.
229 // The main reason is revising young length, with our without the GCLocker being
230 // active.
231 //
232 uint G1Policy::calculate_young_desired_length(size_t rs_length) const {
233 uint min_young_length_by_sizer = _young_gen_sizer->min_desired_young_length();
234 uint max_young_length_by_sizer = _young_gen_sizer->max_desired_young_length();
235
236 assert(min_young_length_by_sizer >= 1, "invariant");
237 assert(max_young_length_by_sizer >= min_young_length_by_sizer, "invariant");
238
239 // Absolute minimum eden length.
240 // Enforcing a minimum eden length helps at startup when the predictors are not
241 // yet trained on the application to avoid unnecessary (but very short) full gcs
242 // on very small (initial) heaps.
243 uint const MinDesiredEdenLength = 1;
244
245 // Calculate the absolute and desired min bounds first.
246
247 // This is how many survivor regions we already have.
248 const uint survivor_length = _g1h->survivor_regions_count();
249 // Size of the already allocated young gen.
250 const uint allocated_young_length = _g1h->young_regions_count();
251 // This is the absolute minimum young length that we can return. Ensure that we
252 // don't go below any user-defined minimum bound; but we might have already
253 // allocated more than that for reasons. In this case, use that.
254 uint absolute_min_young_length = MAX2(allocated_young_length, min_young_length_by_sizer);
255 // Calculate the absolute max bounds. After evac failure or when revising the
256 // young length we might have exceeded absolute min length or absolute_max_length,
257 // so adjust the result accordingly.
258 uint absolute_max_young_length = MAX2(max_young_length_by_sizer, absolute_min_young_length);
259
260 uint desired_eden_length_by_mmu = 0;
261 uint desired_eden_length_by_pause = 0;
262 uint desired_eden_length_before_mixed = 0;
263
264 uint desired_young_length = 0;
265 if (use_adaptive_young_list_length()) {
266 desired_eden_length_by_mmu = calculate_desired_eden_length_by_mmu();
267
268 const size_t pending_cards = _analytics->predict_pending_cards();
269 double survivor_base_time_ms = predict_base_elapsed_time_ms(pending_cards, rs_length);
270
271 if (!next_gc_should_be_mixed(NULL, NULL)) {
272 desired_eden_length_by_pause =
273 calculate_desired_eden_length_by_pause(survivor_base_time_ms,
274 absolute_min_young_length - survivor_length,
275 absolute_max_young_length - survivor_length);
276 } else {
277 desired_eden_length_before_mixed =
278 calculate_desired_eden_length_before_mixed(survivor_base_time_ms,
279 absolute_min_young_length - survivor_length,
280 absolute_max_young_length - survivor_length);
281 }
282 // Above either sets desired_eden_length_by_pause or desired_eden_length_before_mixed,
283 // the other is zero. Use the one that has been set below.
284 uint desired_eden_length = MAX2(desired_eden_length_by_pause,
285 desired_eden_length_before_mixed);
286
287 // Finally incorporate MMU concerns; assume that it overrides the pause time
288 // goal, as the default value has been chosen to effectively disable it.
289 // Also request at least one eden region, see above for reasons.
290 desired_eden_length = MAX3(desired_eden_length,
291 desired_eden_length_by_mmu,
292 MinDesiredEdenLength);
293
294 desired_young_length = desired_eden_length + survivor_length;
295 } else {
296 // The user asked for a fixed young gen so we'll fix the young gen
297 // whether the next GC is young or mixed.
298 desired_young_length = min_young_length_by_sizer;
299 }
300 // Clamp to absolute min/max after we determined desired lengths.
301 desired_young_length = clamp(desired_young_length, absolute_min_young_length, absolute_max_young_length);
302
303 log_trace(gc, ergo, heap)("Young desired length %u "
304 "survivor length %u "
305 "allocated young length %u "
306 "absolute min young length %u "
307 "absolute max young length %u "
308 "desired eden length by mmu %u "
309 "desired eden length by pause %u "
310 "desired eden length before mixed %u"
311 "desired eden length by default %u",
312 desired_young_length, survivor_length,
313 allocated_young_length, absolute_min_young_length,
314 absolute_max_young_length, desired_eden_length_by_mmu,
315 desired_eden_length_by_pause,
316 desired_eden_length_before_mixed,
317 MinDesiredEdenLength);
318
319 assert(desired_young_length >= allocated_young_length, "must be");
320 return desired_young_length;
321 }
322
323 // Limit the desired (wished) young length by current free regions. If the request
324 // can be satisfied without using up reserve regions, do so, otherwise eat into
325 // the reserve, giving away at most what the heap sizer allows.
326 uint G1Policy::calculate_young_target_length(uint desired_young_length) const {
327 uint allocated_young_length = _g1h->young_regions_count();
328
329 uint receiving_additional_eden;
330 if (allocated_young_length >= desired_young_length) {
331 // Already used up all we actually want (may happen as G1 revises the
332 // young list length concurrently, or caused by gclocker). Do not allow more,
333 // potentially resulting in GC.
334 receiving_additional_eden = 0;
335 log_trace(gc, ergo, heap)("Young target length: Already used up desired young %u allocated %u",
336 desired_young_length,
337 allocated_young_length);
338 } else {
339 // Now look at how many free regions are there currently, and the heap reserve.
340 // We will try our best not to "eat" into the reserve as long as we can. If we
341 // do, we at most eat the sizer's minimum regions into the reserve or half the
342 // reserve rounded up (if possible; this is an arbitrary value).
343
344 uint max_to_eat_into_reserve = MIN2(_young_gen_sizer->min_desired_young_length(),
345 (_reserve_regions + 1) / 2);
346
347 log_trace(gc, ergo, heap)("Young target length: Common "
348 "free regions at end of collection %u "
349 "desired young length %u "
350 "reserve region %u "
351 "max to eat into reserve %u",
352 _free_regions_at_end_of_collection,
353 desired_young_length,
354 _reserve_regions,
355 max_to_eat_into_reserve);
356
357 if (_free_regions_at_end_of_collection <= _reserve_regions) {
358 // Fully eat (or already eating) into the reserve, hand back at most absolute_min_length regions.
359 uint receiving_young = MIN3(_free_regions_at_end_of_collection,
360 desired_young_length,
361 max_to_eat_into_reserve);
362 // We could already have allocated more regions than what we could get
363 // above.
364 receiving_additional_eden = allocated_young_length < receiving_young ?
365 receiving_young - allocated_young_length : 0;
366
367 log_trace(gc, ergo, heap)("Young target length: Fully eat into reserve "
368 "receiving young %u receiving additional eden %u",
369 receiving_young,
370 receiving_additional_eden);
371 } else if (_free_regions_at_end_of_collection < (desired_young_length + _reserve_regions)) {
372 // Partially eat into the reserve, at most max_to_eat_into_reserve regions.
373 uint free_outside_reserve = _free_regions_at_end_of_collection - _reserve_regions;
374 assert(free_outside_reserve < desired_young_length,
375 "must be %u %u",
376 free_outside_reserve, desired_young_length);
377
378 uint receiving_within_reserve = MIN2(desired_young_length - free_outside_reserve,
379 max_to_eat_into_reserve);
380 uint receiving_young = free_outside_reserve + receiving_within_reserve;
381 // Again, we could have already allocated more than we could get.
382 receiving_additional_eden = allocated_young_length < receiving_young ?
383 receiving_young - allocated_young_length : 0;
384
385 log_trace(gc, ergo, heap)("Young target length: Partially eat into reserve "
386 "free outside reserve %u "
387 "receiving within reserve %u "
388 "receiving young %u "
389 "receiving additional eden %u",
390 free_outside_reserve, receiving_within_reserve,
391 receiving_young, receiving_additional_eden);
392 } else {
393 // No need to use the reserve.
394 receiving_additional_eden = desired_young_length - allocated_young_length;
395 log_trace(gc, ergo, heap)("Young target length: No need to use reserve "
396 "receiving additional eden %u",
397 receiving_additional_eden);
398 }
399 }
400
401 uint target_young_length = allocated_young_length + receiving_additional_eden;
402
403 assert(target_young_length >= allocated_young_length, "must be");
404
405 log_trace(gc, ergo, heap)("Young target length: "
406 "young target length %u "
407 "allocated young length %u "
408 "received additional eden %u",
409 target_young_length, allocated_young_length,
410 receiving_additional_eden);
411 return target_young_length;
412 }
413
414 uint G1Policy::calculate_desired_eden_length_by_pause(double base_time_ms,
415 uint min_eden_length,
416 uint max_eden_length) const {
417 assert(use_adaptive_young_list_length(), "pre-condition");
418
419 assert(min_eden_length <= max_eden_length, "must be %u %u", min_eden_length, max_eden_length);
420
421 // Here, we will make sure that the shortest young length that
422 // makes sense fits within the target pause time.
423
424 G1YoungLengthPredictor p(base_time_ms,
425 _free_regions_at_end_of_collection,
426 _mmu_tracker->max_gc_time() * 1000.0,
427 this);
428 if (p.will_fit(min_eden_length)) {
429 // The shortest young length will fit into the target pause time;
430 // we'll now check whether the absolute maximum number of young
431 // regions will fit in the target pause time. If not, we'll do
432 // a binary search between min_young_length and max_young_length.
433 if (p.will_fit(max_eden_length)) {
434 // The maximum young length will fit into the target pause time.
435 // We are done so set min young length to the maximum length (as
436 // the result is assumed to be returned in min_young_length).
437 min_eden_length = max_eden_length;
438 } else {
439 // The maximum possible number of young regions will not fit within
440 // the target pause time so we'll search for the optimal
441 // length. The loop invariants are:
442 //
443 // min_young_length < max_young_length
444 // min_young_length is known to fit into the target pause time
445 // max_young_length is known not to fit into the target pause time
446 //
447 // Going into the loop we know the above hold as we've just
448 // checked them. Every time around the loop we check whether
449 // the middle value between min_young_length and
450 // max_young_length fits into the target pause time. If it
451 // does, it becomes the new min. If it doesn't, it becomes
452 // the new max. This way we maintain the loop invariants.
453
454 assert(min_eden_length < max_eden_length, "invariant");
455 uint diff = (max_eden_length - min_eden_length) / 2;
456 while (diff > 0) {
457 uint eden_length = min_eden_length + diff;
458 if (p.will_fit(eden_length)) {
459 min_eden_length = eden_length;
460 } else {
461 max_eden_length = eden_length;
462 }
463 assert(min_eden_length < max_eden_length, "invariant");
464 diff = (max_eden_length - min_eden_length) / 2;
465 }
466 // The results is min_young_length which, according to the
467 // loop invariants, should fit within the target pause time.
468
469 // These are the post-conditions of the binary search above:
470 assert(min_eden_length < max_eden_length,
471 "otherwise we should have discovered that max_eden_length "
472 "fits into the pause target and not done the binary search");
473 assert(p.will_fit(min_eden_length),
474 "min_eden_length, the result of the binary search, should "
475 "fit into the pause target");
476 assert(!p.will_fit(min_eden_length + 1),
477 "min_eden_length, the result of the binary search, should be "
478 "optimal, so no larger length should fit into the pause target");
479 }
480 } else {
481 // Even the minimum length doesn't fit into the pause time
482 // target, return it as the result nevertheless.
483 }
484 return min_eden_length;
485 }
486
487 uint G1Policy::calculate_desired_eden_length_before_mixed(double survivor_base_time_ms,
488 uint min_eden_length,
489 uint max_eden_length) const {
490 G1CollectionSetCandidates* candidates = _collection_set->candidates();
491
492 uint min_old_regions_end = MIN2(candidates->cur_idx() + calc_min_old_cset_length(), candidates->num_regions());
493 double predicted_region_evac_time_ms = survivor_base_time_ms;
494 for (uint i = candidates->cur_idx(); i < min_old_regions_end; i++) {
495 HeapRegion* r = candidates->at(i);
496 predicted_region_evac_time_ms += predict_region_total_time_ms(r, false);
497 }
498 uint desired_eden_length_by_min_cset_length =
499 calculate_desired_eden_length_by_pause(predicted_region_evac_time_ms,
500 min_eden_length,
501 max_eden_length);
502
503 return desired_eden_length_by_min_cset_length;
504 }
505
506 double G1Policy::predict_survivor_regions_evac_time() const {
507 double survivor_regions_evac_time = 0.0;
508 const GrowableArray<HeapRegion*>* survivor_regions = _g1h->survivor()->regions();
509 for (GrowableArrayIterator<HeapRegion*> it = survivor_regions->begin();
510 it != survivor_regions->end();
511 ++it) {
512 survivor_regions_evac_time += predict_region_total_time_ms(*it, collector_state()->in_young_only_phase());
513 }
514 return survivor_regions_evac_time;
515 }
516
517 void G1Policy::revise_young_list_target_length_if_necessary(size_t rs_length) {
518 guarantee(use_adaptive_young_list_length(), "should not call this otherwise" );
519
520 if (rs_length > _rs_length_prediction) {
521 // add 10% to avoid having to recalculate often
522 size_t rs_length_prediction = rs_length * 1100 / 1000;
523 update_rs_length_prediction(rs_length_prediction);
524 update_young_length_bounds(rs_length_prediction);
525 }
526 }
527
528 void G1Policy::update_rs_length_prediction() {
529 update_rs_length_prediction(_analytics->predict_rs_length());
530 }
531
532 void G1Policy::update_rs_length_prediction(size_t prediction) {
533 if (collector_state()->in_young_only_phase() && use_adaptive_young_list_length()) {
534 _rs_length_prediction = prediction;
535 }
536 }
537
538 void G1Policy::record_full_collection_start() {
539 _full_collection_start_sec = os::elapsedTime();
540 // Release the future to-space so that it is available for compaction into.
541 collector_state()->set_in_young_only_phase(false);
542 collector_state()->set_in_full_gc(true);
543 _collection_set->clear_candidates();
544 _pending_cards_at_gc_start = 0;
545 }
546
547 void G1Policy::record_full_collection_end() {
548 // Consider this like a collection pause for the purposes of allocation
549 // since last pause.
550 double end_sec = os::elapsedTime();
551 double full_gc_time_sec = end_sec - _full_collection_start_sec;
552 double full_gc_time_ms = full_gc_time_sec * 1000.0;
553
554 _analytics->update_recent_gc_times(end_sec, full_gc_time_ms);
555
556 collector_state()->set_in_full_gc(false);
557
558 // "Nuke" the heuristics that control the young/mixed GC
559 // transitions and make sure we start with young GCs after the Full GC.
560 collector_state()->set_in_young_only_phase(true);
561 collector_state()->set_in_young_gc_before_mixed(false);
562 collector_state()->set_initiate_conc_mark_if_possible(need_to_start_conc_mark("end of Full GC", 0));
563 collector_state()->set_in_initial_mark_gc(false);
564 collector_state()->set_mark_or_rebuild_in_progress(false);
565 collector_state()->set_clearing_next_bitmap(false);
566
567 _eden_surv_rate_group->start_adding_regions();
568 // also call this on any additional surv rate groups
569
570 _free_regions_at_end_of_collection = _g1h->num_free_regions();
571 _survivor_surv_rate_group->reset();
572 update_young_length_bounds();
573 update_rs_length_prediction();
574
575 _old_gen_alloc_tracker.reset_after_full_gc();
576
577 record_pause(FullGC, _full_collection_start_sec, end_sec);
578 }
579
580 static void log_refinement_stats(const char* kind, const G1ConcurrentRefineStats& stats) {
581 log_debug(gc, refine, stats)
582 ("%s refinement: %.2fms, refined: " SIZE_FORMAT
583 ", precleaned: " SIZE_FORMAT ", dirtied: " SIZE_FORMAT,
584 kind,
585 stats.refinement_time().seconds() * MILLIUNITS,
586 stats.refined_cards(),
587 stats.precleaned_cards(),
588 stats.dirtied_cards());
589 }
590
591 void G1Policy::record_concurrent_refinement_stats() {
592 G1DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set();
593 _pending_cards_at_gc_start = dcqs.num_cards();
594
595 // Collect per-thread stats, mostly from mutator activity.
596 G1ConcurrentRefineStats mut_stats = dcqs.get_and_reset_refinement_stats();
597
598 // Collect specialized concurrent refinement thread stats.
599 G1ConcurrentRefine* cr = _g1h->concurrent_refine();
600 G1ConcurrentRefineStats cr_stats = cr->get_and_reset_refinement_stats();
601
602 G1ConcurrentRefineStats total_stats = mut_stats + cr_stats;
603
604 log_refinement_stats("Mutator", mut_stats);
605 log_refinement_stats("Concurrent", cr_stats);
606 log_refinement_stats("Total", total_stats);
607
608 // Record the rate at which cards were refined.
609 // Don't update the rate if the current sample is empty or time is zero.
610 Tickspan refinement_time = total_stats.refinement_time();
611 size_t refined_cards = total_stats.refined_cards();
612 if ((refined_cards > 0) && (refinement_time > Tickspan())) {
613 double rate = refined_cards / (refinement_time.seconds() * MILLIUNITS);
614 _analytics->report_concurrent_refine_rate_ms(rate);
615 log_debug(gc, refine, stats)("Concurrent refinement rate: %.2f cards/ms", rate);
616 }
617
618 // Record mutator's card logging rate.
619 double mut_start_time = _analytics->prev_collection_pause_end_ms();
620 double mut_end_time = phase_times()->cur_collection_start_sec() * MILLIUNITS;
621 double mut_time = mut_end_time - mut_start_time;
622 // Unlike above for conc-refine rate, here we should not require a
623 // non-empty sample, since an application could go some time with only
624 // young-gen or filtered out writes. But we'll ignore unusually short
625 // sample periods, as they may just pollute the predictions.
626 if (mut_time > 1.0) { // Require > 1ms sample time.
627 double dirtied_rate = total_stats.dirtied_cards() / mut_time;
628 _analytics->report_dirtied_cards_rate_ms(dirtied_rate);
629 log_debug(gc, refine, stats)("Generate dirty cards rate: %.2f cards/ms", dirtied_rate);
630 }
631 }
632
633 void G1Policy::record_collection_pause_start(double start_time_sec) {
634 // We only need to do this here as the policy will only be applied
635 // to the GC we're about to start. so, no point is calculating this
636 // every time we calculate / recalculate the target young length.
637 update_survivors_policy();
638
639 assert(max_survivor_regions() + _g1h->num_used_regions() <= _g1h->max_regions(),
640 "Maximum survivor regions %u plus used regions %u exceeds max regions %u",
641 max_survivor_regions(), _g1h->num_used_regions(), _g1h->max_regions());
642 assert_used_and_recalculate_used_equal(_g1h);
643
644 phase_times()->record_cur_collection_start_sec(start_time_sec);
645
646 record_concurrent_refinement_stats();
647
648 _collection_set->reset_bytes_used_before();
649
650 // do that for any other surv rate groups
651 _eden_surv_rate_group->stop_adding_regions();
652 _survivors_age_table.clear();
653
654 assert(_g1h->collection_set()->verify_young_ages(), "region age verification failed");
655 }
656
657 void G1Policy::record_concurrent_mark_init_end(double mark_init_elapsed_time_ms) {
658 assert(!collector_state()->initiate_conc_mark_if_possible(), "we should have cleared it by now");
659 collector_state()->set_in_initial_mark_gc(false);
660 }
661
662 void G1Policy::record_concurrent_mark_remark_start() {
663 _mark_remark_start_sec = os::elapsedTime();
664 }
665
666 void G1Policy::record_concurrent_mark_remark_end() {
667 double end_time_sec = os::elapsedTime();
668 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
669 _analytics->report_concurrent_mark_remark_times_ms(elapsed_time_ms);
670 _analytics->append_prev_collection_pause_end_ms(elapsed_time_ms);
671
672 record_pause(Remark, _mark_remark_start_sec, end_time_sec);
673 }
674
675 void G1Policy::record_concurrent_mark_cleanup_start() {
676 _mark_cleanup_start_sec = os::elapsedTime();
677 }
678
679 double G1Policy::average_time_ms(G1GCPhaseTimes::GCParPhases phase) const {
680 return phase_times()->average_time_ms(phase);
681 }
682
683 double G1Policy::young_other_time_ms() const {
684 return phase_times()->young_cset_choice_time_ms() +
685 phase_times()->average_time_ms(G1GCPhaseTimes::YoungFreeCSet);
686 }
687
688 double G1Policy::non_young_other_time_ms() const {
689 return phase_times()->non_young_cset_choice_time_ms() +
690 phase_times()->average_time_ms(G1GCPhaseTimes::NonYoungFreeCSet);
691 }
692
693 double G1Policy::other_time_ms(double pause_time_ms) const {
694 return pause_time_ms - phase_times()->cur_collection_par_time_ms();
695 }
696
697 double G1Policy::constant_other_time_ms(double pause_time_ms) const {
698 return other_time_ms(pause_time_ms) - phase_times()->total_free_cset_time_ms() - phase_times()->total_rebuild_freelist_time_ms();
699 }
700
701 bool G1Policy::about_to_start_mixed_phase() const {
702 return _g1h->concurrent_mark()->cm_thread()->during_cycle() || collector_state()->in_young_gc_before_mixed();
703 }
704
705 bool G1Policy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) {
706 if (about_to_start_mixed_phase()) {
707 return false;
708 }
709
710 size_t marking_initiating_used_threshold = _ihop_control->get_conc_mark_start_threshold();
711
712 size_t cur_used_bytes = _g1h->non_young_capacity_bytes();
713 size_t alloc_byte_size = alloc_word_size * HeapWordSize;
714 size_t marking_request_bytes = cur_used_bytes + alloc_byte_size;
715
716 bool result = false;
717 if (marking_request_bytes > marking_initiating_used_threshold) {
718 result = collector_state()->in_young_only_phase() && !collector_state()->in_young_gc_before_mixed();
719 log_debug(gc, ergo, ihop)("%s occupancy: " SIZE_FORMAT "B allocation request: " SIZE_FORMAT "B threshold: " SIZE_FORMAT "B (%1.2f) source: %s",
720 result ? "Request concurrent cycle initiation (occupancy higher than threshold)" : "Do not request concurrent cycle initiation (still doing mixed collections)",
721 cur_used_bytes, alloc_byte_size, marking_initiating_used_threshold, (double) marking_initiating_used_threshold / _g1h->capacity() * 100, source);
722 }
723
724 return result;
725 }
726
727 double G1Policy::logged_cards_processing_time() const {
728 double all_cards_processing_time = average_time_ms(G1GCPhaseTimes::ScanHR) + average_time_ms(G1GCPhaseTimes::OptScanHR);
729 size_t logged_dirty_cards = phase_times()->sum_thread_work_items(G1GCPhaseTimes::MergeLB, G1GCPhaseTimes::MergeLBDirtyCards);
730 size_t scan_heap_roots_cards = phase_times()->sum_thread_work_items(G1GCPhaseTimes::ScanHR, G1GCPhaseTimes::ScanHRScannedCards) +
731 phase_times()->sum_thread_work_items(G1GCPhaseTimes::OptScanHR, G1GCPhaseTimes::ScanHRScannedCards);
732 // This may happen if there are duplicate cards in different log buffers.
733 if (logged_dirty_cards > scan_heap_roots_cards) {
734 return all_cards_processing_time + average_time_ms(G1GCPhaseTimes::MergeLB);
735 }
736 return (all_cards_processing_time * logged_dirty_cards / scan_heap_roots_cards) + average_time_ms(G1GCPhaseTimes::MergeLB);
737 }
738
739 // Anything below that is considered to be zero
740 #define MIN_TIMER_GRANULARITY 0.0000001
741
742 void G1Policy::record_collection_pause_end(double pause_time_ms) {
743 G1GCPhaseTimes* p = phase_times();
744
745 double end_time_sec = os::elapsedTime();
746
747 bool this_pause_included_initial_mark = false;
748 bool this_pause_was_young_only = collector_state()->in_young_only_phase();
749
750 bool update_stats = !_g1h->evacuation_failed();
751
752 record_pause(young_gc_pause_kind(), end_time_sec - pause_time_ms / 1000.0, end_time_sec);
753
754 _collection_pause_end_millis = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
755
756 this_pause_included_initial_mark = collector_state()->in_initial_mark_gc();
757 if (this_pause_included_initial_mark) {
758 record_concurrent_mark_init_end(0.0);
759 } else {
760 maybe_start_marking();
761 }
762
763 double app_time_ms = (phase_times()->cur_collection_start_sec() * 1000.0 - _analytics->prev_collection_pause_end_ms());
764 if (app_time_ms < MIN_TIMER_GRANULARITY) {
765 // This usually happens due to the timer not having the required
766 // granularity. Some Linuxes are the usual culprits.
767 // We'll just set it to something (arbitrarily) small.
768 app_time_ms = 1.0;
769 }
770
771 if (update_stats) {
772 // We maintain the invariant that all objects allocated by mutator
773 // threads will be allocated out of eden regions. So, we can use
774 // the eden region number allocated since the previous GC to
775 // calculate the application's allocate rate. The only exception
776 // to that is humongous objects that are allocated separately. But
777 // given that humongous object allocations do not really affect
778 // either the pause's duration nor when the next pause will take
779 // place we can safely ignore them here.
780 uint regions_allocated = _collection_set->eden_region_length();
781 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
782 _analytics->report_alloc_rate_ms(alloc_rate_ms);
783
784 _analytics->compute_pause_time_ratios(end_time_sec, pause_time_ms);
785 _analytics->update_recent_gc_times(end_time_sec, pause_time_ms);
786 }
787
788 if (collector_state()->in_young_gc_before_mixed()) {
789 assert(!this_pause_included_initial_mark, "The young GC before mixed is not allowed to be an initial mark GC");
790 // This has been the young GC before we start doing mixed GCs. We already
791 // decided to start mixed GCs much earlier, so there is nothing to do except
792 // advancing the state.
793 collector_state()->set_in_young_only_phase(false);
794 collector_state()->set_in_young_gc_before_mixed(false);
795 } else if (!this_pause_was_young_only) {
796 // This is a mixed GC. Here we decide whether to continue doing more
797 // mixed GCs or not.
798 if (!next_gc_should_be_mixed("continue mixed GCs",
799 "do not continue mixed GCs")) {
800 collector_state()->set_in_young_only_phase(true);
801
802 clear_collection_set_candidates();
803 maybe_start_marking();
804 }
805 }
806
807 _eden_surv_rate_group->start_adding_regions();
808
809 double merge_hcc_time_ms = average_time_ms(G1GCPhaseTimes::MergeHCC);
810 if (update_stats) {
811 size_t const total_log_buffer_cards = p->sum_thread_work_items(G1GCPhaseTimes::MergeHCC, G1GCPhaseTimes::MergeHCCDirtyCards) +
812 p->sum_thread_work_items(G1GCPhaseTimes::MergeLB, G1GCPhaseTimes::MergeLBDirtyCards);
813 // Update prediction for card merge; MergeRSDirtyCards includes the cards from the Eager Reclaim phase.
814 size_t const total_cards_merged = p->sum_thread_work_items(G1GCPhaseTimes::MergeRS, G1GCPhaseTimes::MergeRSDirtyCards) +
815 p->sum_thread_work_items(G1GCPhaseTimes::OptMergeRS, G1GCPhaseTimes::MergeRSDirtyCards) +
816 total_log_buffer_cards;
817
818 // The threshold for the number of cards in a given sampling which we consider
819 // large enough so that the impact from setup and other costs is negligible.
820 size_t const CardsNumSamplingThreshold = 10;
821
822 if (total_cards_merged > CardsNumSamplingThreshold) {
823 double avg_time_merge_cards = average_time_ms(G1GCPhaseTimes::MergeER) +
824 average_time_ms(G1GCPhaseTimes::MergeRS) +
825 average_time_ms(G1GCPhaseTimes::MergeHCC) +
826 average_time_ms(G1GCPhaseTimes::MergeLB) +
827 average_time_ms(G1GCPhaseTimes::OptMergeRS);
828 _analytics->report_cost_per_card_merge_ms(avg_time_merge_cards / total_cards_merged, this_pause_was_young_only);
829 }
830
831 // Update prediction for card scan
832 size_t const total_cards_scanned = p->sum_thread_work_items(G1GCPhaseTimes::ScanHR, G1GCPhaseTimes::ScanHRScannedCards) +
833 p->sum_thread_work_items(G1GCPhaseTimes::OptScanHR, G1GCPhaseTimes::ScanHRScannedCards);
834
835 if (total_cards_scanned > CardsNumSamplingThreshold) {
836 double avg_time_dirty_card_scan = average_time_ms(G1GCPhaseTimes::ScanHR) +
837 average_time_ms(G1GCPhaseTimes::OptScanHR);
838
839 _analytics->report_cost_per_card_scan_ms(avg_time_dirty_card_scan / total_cards_scanned, this_pause_was_young_only);
840 }
841
842 // Update prediction for the ratio between cards from the remembered
843 // sets and actually scanned cards from the remembered sets.
844 // Cards from the remembered sets are all cards not duplicated by cards from
845 // the logs.
846 // Due to duplicates in the log buffers, the number of actually scanned cards
847 // can be smaller than the cards in the log buffers.
848 const size_t from_rs_length_cards = (total_cards_scanned > total_log_buffer_cards) ? total_cards_scanned - total_log_buffer_cards : 0;
849 double merge_to_scan_ratio = 0.0;
850 if (total_cards_scanned > 0) {
851 merge_to_scan_ratio = (double) from_rs_length_cards / total_cards_scanned;
852 }
853 _analytics->report_card_merge_to_scan_ratio(merge_to_scan_ratio, this_pause_was_young_only);
854
855 const size_t recorded_rs_length = _collection_set->recorded_rs_length();
856 const size_t rs_length_diff = _rs_length > recorded_rs_length ? _rs_length - recorded_rs_length : 0;
857 _analytics->report_rs_length_diff(rs_length_diff);
858
859 // Update prediction for copy cost per byte
860 size_t copied_bytes = p->sum_thread_work_items(G1GCPhaseTimes::MergePSS, G1GCPhaseTimes::MergePSSCopiedBytes);
861
862 if (copied_bytes > 0) {
863 double cost_per_byte_ms = (average_time_ms(G1GCPhaseTimes::ObjCopy) + average_time_ms(G1GCPhaseTimes::OptObjCopy)) / copied_bytes;
864 _analytics->report_cost_per_byte_ms(cost_per_byte_ms, collector_state()->mark_or_rebuild_in_progress());
865 }
866
867 if (_collection_set->young_region_length() > 0) {
868 _analytics->report_young_other_cost_per_region_ms(young_other_time_ms() /
869 _collection_set->young_region_length());
870 }
871
872 if (_collection_set->old_region_length() > 0) {
873 _analytics->report_non_young_other_cost_per_region_ms(non_young_other_time_ms() /
874 _collection_set->old_region_length());
875 }
876
877 _analytics->report_constant_other_time_ms(constant_other_time_ms(pause_time_ms));
878
879 // Do not update RS lengths and the number of pending cards with information from mixed gc:
880 // these are is wildly different to during young only gc and mess up young gen sizing right
881 // after the mixed gc phase.
882 // During mixed gc we do not use them for young gen sizing.
883 if (this_pause_was_young_only) {
884 _analytics->report_pending_cards((double) _pending_cards_at_gc_start);
885 _analytics->report_rs_length((double) _rs_length);
886 }
887 }
888
889 assert(!(this_pause_included_initial_mark && collector_state()->mark_or_rebuild_in_progress()),
890 "If the last pause has been an initial mark, we should not have been in the marking window");
891 if (this_pause_included_initial_mark) {
892 collector_state()->set_mark_or_rebuild_in_progress(true);
893 }
894
895 _free_regions_at_end_of_collection = _g1h->num_free_regions();
896
897 update_rs_length_prediction();
898
899 // Do not update dynamic IHOP due to G1 periodic collection as it is highly likely
900 // that in this case we are not running in a "normal" operating mode.
901 if (_g1h->gc_cause() != GCCause::_g1_periodic_collection) {
902 update_young_length_bounds();
903
904 _old_gen_alloc_tracker.reset_after_young_gc(app_time_ms / 1000.0);
905 update_ihop_prediction(_old_gen_alloc_tracker.last_cycle_duration(),
906 _old_gen_alloc_tracker.last_cycle_old_bytes(),
907 this_pause_was_young_only);
908
909 _ihop_control->send_trace_event(_g1h->gc_tracer_stw());
910 } else {
911 // Any garbage collection triggered as periodic collection resets the time-to-mixed
912 // measurement. Periodic collection typically means that the application is "inactive", i.e.
913 // the marking threads may have received an uncharacterisic amount of cpu time
914 // for completing the marking, i.e. are faster than expected.
915 // This skews the predicted marking length towards smaller values which might cause
916 // the mark start being too late.
917 _initial_mark_to_mixed.reset();
918 }
919
920 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
921 double scan_logged_cards_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
922
923 if (scan_logged_cards_time_goal_ms < merge_hcc_time_ms) {
924 log_debug(gc, ergo, refine)("Adjust concurrent refinement thresholds (scanning the HCC expected to take longer than Update RS time goal)."
925 "Logged Cards Scan time goal: %1.2fms Scan HCC time: %1.2fms",
926 scan_logged_cards_time_goal_ms, merge_hcc_time_ms);
927
928 scan_logged_cards_time_goal_ms = 0;
929 } else {
930 scan_logged_cards_time_goal_ms -= merge_hcc_time_ms;
931 }
932
933 double const logged_cards_time = logged_cards_processing_time();
934
935 log_debug(gc, ergo, refine)("Concurrent refinement times: Logged Cards Scan time goal: %1.2fms Logged Cards Scan time: %1.2fms HCC time: %1.2fms",
936 scan_logged_cards_time_goal_ms, logged_cards_time, merge_hcc_time_ms);
937
938 _g1h->concurrent_refine()->adjust(logged_cards_time,
939 phase_times()->sum_thread_work_items(G1GCPhaseTimes::MergeLB, G1GCPhaseTimes::MergeLBDirtyCards),
940 scan_logged_cards_time_goal_ms);
941 }
942
943 G1IHOPControl* G1Policy::create_ihop_control(const G1Predictions* predictor){
944 if (G1UseAdaptiveIHOP) {
945 return new G1AdaptiveIHOPControl(InitiatingHeapOccupancyPercent,
946 predictor,
947 G1ReservePercent,
948 G1HeapWastePercent);
949 } else {
950 return new G1StaticIHOPControl(InitiatingHeapOccupancyPercent);
951 }
952 }
953
954 void G1Policy::update_ihop_prediction(double mutator_time_s,
955 size_t mutator_alloc_bytes,
956 bool this_gc_was_young_only) {
957 // Always try to update IHOP prediction. Even evacuation failures give information
958 // about e.g. whether to start IHOP earlier next time.
959
960 // Avoid using really small application times that might create samples with
961 // very high or very low values. They may be caused by e.g. back-to-back gcs.
962 double const min_valid_time = 1e-6;
963
964 bool report = false;
965
966 double marking_to_mixed_time = -1.0;
967 if (!this_gc_was_young_only && _initial_mark_to_mixed.has_result()) {
968 marking_to_mixed_time = _initial_mark_to_mixed.last_marking_time();
969 assert(marking_to_mixed_time > 0.0,
970 "Initial mark to mixed time must be larger than zero but is %.3f",
971 marking_to_mixed_time);
972 if (marking_to_mixed_time > min_valid_time) {
973 _ihop_control->update_marking_length(marking_to_mixed_time);
974 report = true;
975 }
976 }
977
978 // As an approximation for the young gc promotion rates during marking we use
979 // all of them. In many applications there are only a few if any young gcs during
980 // marking, which makes any prediction useless. This increases the accuracy of the
981 // prediction.
982 if (this_gc_was_young_only && mutator_time_s > min_valid_time) {
983 // IHOP control wants to know the expected young gen length if it were not
984 // restrained by the heap reserve. Using the actual length would make the
985 // prediction too small and the limit the young gen every time we get to the
986 // predicted target occupancy.
987 uint young_gen_size = young_list_desired_length() * HeapRegion::GrainBytes;
988 _ihop_control->update_allocation_info(mutator_time_s, mutator_alloc_bytes, young_gen_size);
989 report = true;
990 }
991
992 if (report) {
993 report_ihop_statistics();
994 }
995 }
996
997 void G1Policy::report_ihop_statistics() {
998 _ihop_control->print();
999 }
1000
1001 void G1Policy::print_phases() {
1002 phase_times()->print();
1003 }
1004
1005 double G1Policy::predict_base_elapsed_time_ms(size_t pending_cards,
1006 size_t rs_length) const {
1007 size_t effective_scanned_cards = _analytics->predict_scan_card_num(rs_length, collector_state()->in_young_only_phase());
1008 return
1009 _analytics->predict_card_merge_time_ms(pending_cards + rs_length, collector_state()->in_young_only_phase()) +
1010 _analytics->predict_card_scan_time_ms(effective_scanned_cards, collector_state()->in_young_only_phase()) +
1011 _analytics->predict_constant_other_time_ms() +
1012 predict_survivor_regions_evac_time();
1013 }
1014
1015 double G1Policy::predict_base_elapsed_time_ms(size_t pending_cards) const {
1016 size_t rs_length = _analytics->predict_rs_length();
1017 return predict_base_elapsed_time_ms(pending_cards, rs_length);
1018 }
1019
1020 size_t G1Policy::predict_bytes_to_copy(HeapRegion* hr) const {
1021 size_t bytes_to_copy;
1022 if (!hr->is_young()) {
1023 bytes_to_copy = hr->max_live_bytes();
1024 } else {
1025 bytes_to_copy = (size_t) (hr->used() * hr->surv_rate_prediction(_predictor));
1026 }
1027 return bytes_to_copy;
1028 }
1029
1030 double G1Policy::predict_eden_copy_time_ms(uint count, size_t* bytes_to_copy) const {
1031 if (count == 0) {
1032 return 0.0;
1033 }
1034 size_t const expected_bytes = _eden_surv_rate_group->accum_surv_rate_pred(count) * HeapRegion::GrainBytes;
1035 if (bytes_to_copy != NULL) {
1036 *bytes_to_copy = expected_bytes;
1037 }
1038 return _analytics->predict_object_copy_time_ms(expected_bytes, collector_state()->mark_or_rebuild_in_progress());
1039 }
1040
1041 double G1Policy::predict_region_copy_time_ms(HeapRegion* hr) const {
1042 size_t const bytes_to_copy = predict_bytes_to_copy(hr);
1043 return _analytics->predict_object_copy_time_ms(bytes_to_copy, collector_state()->mark_or_rebuild_in_progress());
1044 }
1045
1046 double G1Policy::predict_region_non_copy_time_ms(HeapRegion* hr,
1047 bool for_young_gc) const {
1048 size_t rs_length = hr->rem_set()->occupied();
1049 size_t scan_card_num = _analytics->predict_scan_card_num(rs_length, for_young_gc);
1050
1051 double region_elapsed_time_ms =
1052 _analytics->predict_card_merge_time_ms(rs_length, collector_state()->in_young_only_phase()) +
1053 _analytics->predict_card_scan_time_ms(scan_card_num, collector_state()->in_young_only_phase());
1054
1055 // The prediction of the "other" time for this region is based
1056 // upon the region type and NOT the GC type.
1057 if (hr->is_young()) {
1058 region_elapsed_time_ms += _analytics->predict_young_other_time_ms(1);
1059 } else {
1060 region_elapsed_time_ms += _analytics->predict_non_young_other_time_ms(1);
1061 }
1062 return region_elapsed_time_ms;
1063 }
1064
1065 double G1Policy::predict_region_total_time_ms(HeapRegion* hr, bool for_young_gc) const {
1066 return predict_region_non_copy_time_ms(hr, for_young_gc) + predict_region_copy_time_ms(hr);
1067 }
1068
1069 bool G1Policy::should_allocate_mutator_region() const {
1070 uint young_list_length = _g1h->young_regions_count();
1071 uint young_list_target_length = _young_list_target_length;
1072 return young_list_length < young_list_target_length;
1073 }
1074
1075 bool G1Policy::can_expand_young_list() const {
1076 uint young_list_length = _g1h->young_regions_count();
1077 uint young_list_max_length = _young_list_max_length;
1078 return young_list_length < young_list_max_length;
1079 }
1080
1081 bool G1Policy::use_adaptive_young_list_length() const {
1082 return _young_gen_sizer->use_adaptive_young_list_length();
1083 }
1084
1085 size_t G1Policy::desired_survivor_size(uint max_regions) const {
1086 size_t const survivor_capacity = HeapRegion::GrainWords * max_regions;
1087 return (size_t)((((double)survivor_capacity) * TargetSurvivorRatio) / 100);
1088 }
1089
1090 void G1Policy::print_age_table() {
1091 _survivors_age_table.print_age_table(_tenuring_threshold);
1092 }
1093
1094 uint G1Policy::calculate_young_max_length(uint target_young_length) const {
1095 uint expansion_region_num = 0;
1096 if (GCLockerEdenExpansionPercent > 0) {
1097 double perc = (double) GCLockerEdenExpansionPercent / 100.0;
1098 double expansion_region_num_d = perc * (double) _young_list_target_length;
1099 // We use ceiling so that if expansion_region_num_d is > 0.0 (but
1100 // less than 1.0) we'll get 1.
1101 expansion_region_num = (uint) ceil(expansion_region_num_d);
1102 } else {
1103 assert(expansion_region_num == 0, "sanity");
1104 }
1105 uint max_length = target_young_length + expansion_region_num;
1106 assert(target_young_length <= max_length, "post-condition");
1107 return max_length;
1108 }
1109
1110 // Calculates survivor space parameters.
1111 void G1Policy::update_survivors_policy() {
1112 double max_survivor_regions_d =
1113 (double) _young_list_target_length / (double) SurvivorRatio;
1114
1115 // Calculate desired survivor size based on desired max survivor regions (unconstrained
1116 // by remaining heap). Otherwise we may cause undesired promotions as we are
1117 // already getting close to end of the heap, impacting performance even more.
1118 uint const desired_max_survivor_regions = ceil(max_survivor_regions_d);
1119 size_t const survivor_size = desired_survivor_size(desired_max_survivor_regions);
1120
1121 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(survivor_size);
1122 if (UsePerfData) {
1123 _policy_counters->tenuring_threshold()->set_value(_tenuring_threshold);
1124 _policy_counters->desired_survivor_size()->set_value(survivor_size * oopSize);
1125 }
1126 // The real maximum survivor size is bounded by the number of regions that can
1127 // be allocated into.
1128 _max_survivor_regions = MIN2(desired_max_survivor_regions,
1129 _g1h->num_free_or_available_regions());
1130 }
1131
1132 bool G1Policy::force_initial_mark_if_outside_cycle(GCCause::Cause gc_cause) {
1133 // We actually check whether we are marking here and not if we are in a
1134 // reclamation phase. This means that we will schedule a concurrent mark
1135 // even while we are still in the process of reclaiming memory.
1136 bool during_cycle = _g1h->concurrent_mark()->cm_thread()->during_cycle();
1137 if (!during_cycle) {
1138 log_debug(gc, ergo)("Request concurrent cycle initiation (requested by GC cause). GC cause: %s", GCCause::to_string(gc_cause));
1139 collector_state()->set_initiate_conc_mark_if_possible(true);
1140 return true;
1141 } else {
1142 log_debug(gc, ergo)("Do not request concurrent cycle initiation (concurrent cycle already in progress). GC cause: %s", GCCause::to_string(gc_cause));
1143 return false;
1144 }
1145 }
1146
1147 void G1Policy::initiate_conc_mark() {
1148 collector_state()->set_in_initial_mark_gc(true);
1149 collector_state()->set_initiate_conc_mark_if_possible(false);
1150 }
1151
1152 void G1Policy::decide_on_conc_mark_initiation() {
1153 // We are about to decide on whether this pause will be an
1154 // initial-mark pause.
1155
1156 // First, collector_state()->in_initial_mark_gc() should not be already set. We
1157 // will set it here if we have to. However, it should be cleared by
1158 // the end of the pause (it's only set for the duration of an
1159 // initial-mark pause).
1160 assert(!collector_state()->in_initial_mark_gc(), "pre-condition");
1161
1162 if (collector_state()->initiate_conc_mark_if_possible()) {
1163 // We had noticed on a previous pause that the heap occupancy has
1164 // gone over the initiating threshold and we should start a
1165 // concurrent marking cycle. Or we've been explicitly requested
1166 // to start a concurrent marking cycle. Either way, we initiate
1167 // one if not inhibited for some reason.
1168
1169 GCCause::Cause cause = _g1h->gc_cause();
1170 if ((cause != GCCause::_wb_breakpoint) &&
1171 ConcurrentGCBreakpoints::is_controlled()) {
1172 log_debug(gc, ergo)("Do not initiate concurrent cycle (whitebox controlled)");
1173 } else if (!about_to_start_mixed_phase() && collector_state()->in_young_only_phase()) {
1174 // Initiate a new initial mark if there is no marking or reclamation going on.
1175 initiate_conc_mark();
1176 log_debug(gc, ergo)("Initiate concurrent cycle (concurrent cycle initiation requested)");
1177 } else if (_g1h->is_user_requested_concurrent_full_gc(cause) ||
1178 (cause == GCCause::_wb_breakpoint)) {
1179 // Initiate a user requested initial mark or run_to a breakpoint.
1180 // An initial mark must be young only GC, so the collector state
1181 // must be updated to reflect this.
1182 collector_state()->set_in_young_only_phase(true);
1183 collector_state()->set_in_young_gc_before_mixed(false);
1184
1185 // We might have ended up coming here about to start a mixed phase with a collection set
1186 // active. The following remark might change the change the "evacuation efficiency" of
1187 // the regions in this set, leading to failing asserts later.
1188 // Since the concurrent cycle will recreate the collection set anyway, simply drop it here.
1189 clear_collection_set_candidates();
1190 abort_time_to_mixed_tracking();
1191 initiate_conc_mark();
1192 log_debug(gc, ergo)("Initiate concurrent cycle (%s requested concurrent cycle)",
1193 (cause == GCCause::_wb_breakpoint) ? "run_to breakpoint" : "user");
1194 } else {
1195 // The concurrent marking thread is still finishing up the
1196 // previous cycle. If we start one right now the two cycles
1197 // overlap. In particular, the concurrent marking thread might
1198 // be in the process of clearing the next marking bitmap (which
1199 // we will use for the next cycle if we start one). Starting a
1200 // cycle now will be bad given that parts of the marking
1201 // information might get cleared by the marking thread. And we
1202 // cannot wait for the marking thread to finish the cycle as it
1203 // periodically yields while clearing the next marking bitmap
1204 // and, if it's in a yield point, it's waiting for us to
1205 // finish. So, at this point we will not start a cycle and we'll
1206 // let the concurrent marking thread complete the last one.
1207 log_debug(gc, ergo)("Do not initiate concurrent cycle (concurrent cycle already in progress)");
1208 }
1209 }
1210 }
1211
1212 void G1Policy::record_concurrent_mark_cleanup_end() {
1213 G1CollectionSetCandidates* candidates = G1CollectionSetChooser::build(_g1h->workers(), _g1h->num_regions());
1214 _collection_set->set_candidates(candidates);
1215
1216 bool mixed_gc_pending = next_gc_should_be_mixed("request mixed gcs", "request young-only gcs");
1217 if (!mixed_gc_pending) {
1218 clear_collection_set_candidates();
1219 abort_time_to_mixed_tracking();
1220 }
1221 collector_state()->set_in_young_gc_before_mixed(mixed_gc_pending);
1222 collector_state()->set_mark_or_rebuild_in_progress(false);
1223
1224 double end_sec = os::elapsedTime();
1225 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0;
1226 _analytics->report_concurrent_mark_cleanup_times_ms(elapsed_time_ms);
1227 _analytics->append_prev_collection_pause_end_ms(elapsed_time_ms);
1228
1229 record_pause(Cleanup, _mark_cleanup_start_sec, end_sec);
1230 }
1231
1232 double G1Policy::reclaimable_bytes_percent(size_t reclaimable_bytes) const {
1233 return percent_of(reclaimable_bytes, _g1h->capacity());
1234 }
1235
1236 class G1ClearCollectionSetCandidateRemSets : public HeapRegionClosure {
1237 virtual bool do_heap_region(HeapRegion* r) {
1238 r->rem_set()->clear_locked(true /* only_cardset */);
1239 return false;
1240 }
1241 };
1242
1243 void G1Policy::clear_collection_set_candidates() {
1244 // Clear remembered sets of remaining candidate regions and the actual candidate
1245 // set.
1246 G1ClearCollectionSetCandidateRemSets cl;
1247 _collection_set->candidates()->iterate(&cl);
1248 _collection_set->clear_candidates();
1249 }
1250
1251 void G1Policy::maybe_start_marking() {
1252 if (need_to_start_conc_mark("end of GC")) {
1253 // Note: this might have already been set, if during the last
1254 // pause we decided to start a cycle but at the beginning of
1255 // this pause we decided to postpone it. That's OK.
1256 collector_state()->set_initiate_conc_mark_if_possible(true);
1257 }
1258 }
1259
1260 G1Policy::PauseKind G1Policy::young_gc_pause_kind() const {
1261 assert(!collector_state()->in_full_gc(), "must be");
1262 if (collector_state()->in_initial_mark_gc()) {
1263 assert(!collector_state()->in_young_gc_before_mixed(), "must be");
1264 return InitialMarkGC;
1265 } else if (collector_state()->in_young_gc_before_mixed()) {
1266 assert(!collector_state()->in_initial_mark_gc(), "must be");
1267 return LastYoungGC;
1268 } else if (collector_state()->in_mixed_phase()) {
1269 assert(!collector_state()->in_initial_mark_gc(), "must be");
1270 assert(!collector_state()->in_young_gc_before_mixed(), "must be");
1271 return MixedGC;
1272 } else {
1273 assert(!collector_state()->in_initial_mark_gc(), "must be");
1274 assert(!collector_state()->in_young_gc_before_mixed(), "must be");
1275 return YoungOnlyGC;
1276 }
1277 }
1278
1279 void G1Policy::record_pause(PauseKind kind, double start, double end) {
1280 // Manage the MMU tracker. For some reason it ignores Full GCs.
1281 if (kind != FullGC) {
1282 _mmu_tracker->add_pause(start, end);
1283 }
1284 // Manage the mutator time tracking from initial mark to first mixed gc.
1285 switch (kind) {
1286 case FullGC:
1287 abort_time_to_mixed_tracking();
1288 break;
1289 case Cleanup:
1290 case Remark:
1291 case YoungOnlyGC:
1292 case LastYoungGC:
1293 _initial_mark_to_mixed.add_pause(end - start);
1294 break;
1295 case InitialMarkGC:
1296 if (_g1h->gc_cause() != GCCause::_g1_periodic_collection) {
1297 _initial_mark_to_mixed.record_initial_mark_end(end);
1298 }
1299 break;
1300 case MixedGC:
1301 _initial_mark_to_mixed.record_mixed_gc_start(start);
1302 break;
1303 default:
1304 ShouldNotReachHere();
1305 }
1306 }
1307
1308 void G1Policy::abort_time_to_mixed_tracking() {
1309 _initial_mark_to_mixed.reset();
1310 }
1311
1312 bool G1Policy::next_gc_should_be_mixed(const char* true_action_str,
1313 const char* false_action_str) const {
1314 G1CollectionSetCandidates* candidates = _collection_set->candidates();
1315
1316 if (candidates == NULL || candidates->is_empty()) {
1317 if (false_action_str != NULL) {
1318 log_debug(gc, ergo)("%s (candidate old regions not available)", false_action_str);
1319 }
1320 return false;
1321 }
1322
1323 // Is the amount of uncollected reclaimable space above G1HeapWastePercent?
1324 size_t reclaimable_bytes = candidates->remaining_reclaimable_bytes();
1325 double reclaimable_percent = reclaimable_bytes_percent(reclaimable_bytes);
1326 double threshold = (double) G1HeapWastePercent;
1327 if (reclaimable_percent <= threshold) {
1328 if (false_action_str != NULL) {
1329 log_debug(gc, ergo)("%s (reclaimable percentage not over threshold). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT,
1330 false_action_str, candidates->num_remaining(), reclaimable_bytes, reclaimable_percent, G1HeapWastePercent);
1331 }
1332 return false;
1333 }
1334 if (true_action_str != NULL) {
1335 log_debug(gc, ergo)("%s (candidate old regions available). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT,
1336 true_action_str, candidates->num_remaining(), reclaimable_bytes, reclaimable_percent, G1HeapWastePercent);
1337 }
1338 return true;
1339 }
1340
1341 uint G1Policy::calc_min_old_cset_length() const {
1342 // The min old CSet region bound is based on the maximum desired
1343 // number of mixed GCs after a cycle. I.e., even if some old regions
1344 // look expensive, we should add them to the CSet anyway to make
1345 // sure we go through the available old regions in no more than the
1346 // maximum desired number of mixed GCs.
1347 //
1348 // The calculation is based on the number of marked regions we added
1349 // to the CSet candidates in the first place, not how many remain, so
1350 // that the result is the same during all mixed GCs that follow a cycle.
1351
1352 const size_t region_num = _collection_set->candidates()->num_regions();
1353 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1);
1354 size_t result = region_num / gc_num;
1355 // emulate ceiling
1356 if (result * gc_num < region_num) {
1357 result += 1;
1358 }
1359 return (uint) result;
1360 }
1361
1362 uint G1Policy::calc_max_old_cset_length() const {
1363 // The max old CSet region bound is based on the threshold expressed
1364 // as a percentage of the heap size. I.e., it should bound the
1365 // number of old regions added to the CSet irrespective of how many
1366 // of them are available.
1367
1368 const G1CollectedHeap* g1h = G1CollectedHeap::heap();
1369 const size_t region_num = g1h->num_regions();
1370 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent;
1371 size_t result = region_num * perc / 100;
1372 // emulate ceiling
1373 if (100 * result < region_num * perc) {
1374 result += 1;
1375 }
1376 return (uint) result;
1377 }
1378
1379 void G1Policy::calculate_old_collection_set_regions(G1CollectionSetCandidates* candidates,
1380 double time_remaining_ms,
1381 uint& num_initial_regions,
1382 uint& num_optional_regions) {
1383 assert(candidates != NULL, "Must be");
1384
1385 num_initial_regions = 0;
1386 num_optional_regions = 0;
1387 uint num_expensive_regions = 0;
1388
1389 double predicted_old_time_ms = 0.0;
1390 double predicted_initial_time_ms = 0.0;
1391 double predicted_optional_time_ms = 0.0;
1392
1393 double optional_threshold_ms = time_remaining_ms * optional_prediction_fraction();
1394
1395 const uint min_old_cset_length = calc_min_old_cset_length();
1396 const uint max_old_cset_length = MAX2(min_old_cset_length, calc_max_old_cset_length());
1397 const uint max_optional_regions = max_old_cset_length - min_old_cset_length;
1398 bool check_time_remaining = use_adaptive_young_list_length();
1399
1400 uint candidate_idx = candidates->cur_idx();
1401
1402 log_debug(gc, ergo, cset)("Start adding old regions to collection set. Min %u regions, max %u regions, "
1403 "time remaining %1.2fms, optional threshold %1.2fms",
1404 min_old_cset_length, max_old_cset_length, time_remaining_ms, optional_threshold_ms);
1405
1406 HeapRegion* hr = candidates->at(candidate_idx);
1407 while (hr != NULL) {
1408 if (num_initial_regions + num_optional_regions >= max_old_cset_length) {
1409 // Added maximum number of old regions to the CSet.
1410 log_debug(gc, ergo, cset)("Finish adding old regions to collection set (Maximum number of regions). "
1411 "Initial %u regions, optional %u regions",
1412 num_initial_regions, num_optional_regions);
1413 break;
1414 }
1415
1416 // Stop adding regions if the remaining reclaimable space is
1417 // not above G1HeapWastePercent.
1418 size_t reclaimable_bytes = candidates->remaining_reclaimable_bytes();
1419 double reclaimable_percent = reclaimable_bytes_percent(reclaimable_bytes);
1420 double threshold = (double) G1HeapWastePercent;
1421 if (reclaimable_percent <= threshold) {
1422 // We've added enough old regions that the amount of uncollected
1423 // reclaimable space is at or below the waste threshold. Stop
1424 // adding old regions to the CSet.
1425 log_debug(gc, ergo, cset)("Finish adding old regions to collection set (Reclaimable percentage below threshold). "
1426 "Reclaimable: " SIZE_FORMAT "%s (%1.2f%%) threshold: " UINTX_FORMAT "%%",
1427 byte_size_in_proper_unit(reclaimable_bytes), proper_unit_for_byte_size(reclaimable_bytes),
1428 reclaimable_percent, G1HeapWastePercent);
1429 break;
1430 }
1431
1432 double predicted_time_ms = predict_region_total_time_ms(hr, false);
1433 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0);
1434 // Add regions to old set until we reach the minimum amount
1435 if (num_initial_regions < min_old_cset_length) {
1436 predicted_old_time_ms += predicted_time_ms;
1437 num_initial_regions++;
1438 // Record the number of regions added with no time remaining
1439 if (time_remaining_ms == 0.0) {
1440 num_expensive_regions++;
1441 }
1442 } else if (!check_time_remaining) {
1443 // In the non-auto-tuning case, we'll finish adding regions
1444 // to the CSet if we reach the minimum.
1445 log_debug(gc, ergo, cset)("Finish adding old regions to collection set (Region amount reached min).");
1446 break;
1447 } else {
1448 // Keep adding regions to old set until we reach the optional threshold
1449 if (time_remaining_ms > optional_threshold_ms) {
1450 predicted_old_time_ms += predicted_time_ms;
1451 num_initial_regions++;
1452 } else if (time_remaining_ms > 0) {
1453 // Keep adding optional regions until time is up.
1454 assert(num_optional_regions < max_optional_regions, "Should not be possible.");
1455 predicted_optional_time_ms += predicted_time_ms;
1456 num_optional_regions++;
1457 } else {
1458 log_debug(gc, ergo, cset)("Finish adding old regions to collection set (Predicted time too high).");
1459 break;
1460 }
1461 }
1462 hr = candidates->at(++candidate_idx);
1463 }
1464 if (hr == NULL) {
1465 log_debug(gc, ergo, cset)("Old candidate collection set empty.");
1466 }
1467
1468 if (num_expensive_regions > 0) {
1469 log_debug(gc, ergo, cset)("Added %u initial old regions to collection set although the predicted time was too high.",
1470 num_expensive_regions);
1471 }
1472
1473 log_debug(gc, ergo, cset)("Finish choosing collection set old regions. Initial: %u, optional: %u, "
1474 "predicted old time: %1.2fms, predicted optional time: %1.2fms, time remaining: %1.2f",
1475 num_initial_regions, num_optional_regions,
1476 predicted_initial_time_ms, predicted_optional_time_ms, time_remaining_ms);
1477 }
1478
1479 void G1Policy::calculate_optional_collection_set_regions(G1CollectionSetCandidates* candidates,
1480 uint const max_optional_regions,
1481 double time_remaining_ms,
1482 uint& num_optional_regions) {
1483 assert(_g1h->collector_state()->in_mixed_phase(), "Should only be called in mixed phase");
1484
1485 num_optional_regions = 0;
1486 double prediction_ms = 0;
1487 uint candidate_idx = candidates->cur_idx();
1488
1489 HeapRegion* r = candidates->at(candidate_idx);
1490 while (num_optional_regions < max_optional_regions) {
1491 assert(r != NULL, "Region must exist");
1492 prediction_ms += predict_region_total_time_ms(r, false);
1493
1494 if (prediction_ms > time_remaining_ms) {
1495 log_debug(gc, ergo, cset)("Prediction %.3fms for region %u does not fit remaining time: %.3fms.",
1496 prediction_ms, r->hrm_index(), time_remaining_ms);
1497 break;
1498 }
1499 // This region will be included in the next optional evacuation.
1500
1501 time_remaining_ms -= prediction_ms;
1502 num_optional_regions++;
1503 r = candidates->at(++candidate_idx);
1504 }
1505
1506 log_debug(gc, ergo, cset)("Prepared %u regions out of %u for optional evacuation. Predicted time: %.3fms",
1507 num_optional_regions, max_optional_regions, prediction_ms);
1508 }
1509
1510 void G1Policy::transfer_survivors_to_cset(const G1SurvivorRegions* survivors) {
1511 note_start_adding_survivor_regions();
1512
1513 HeapRegion* last = NULL;
1514 for (GrowableArrayIterator<HeapRegion*> it = survivors->regions()->begin();
1515 it != survivors->regions()->end();
1516 ++it) {
1517 HeapRegion* curr = *it;
1518 set_region_survivor(curr);
1519
1520 // The region is a non-empty survivor so let's add it to
1521 // the incremental collection set for the next evacuation
1522 // pause.
1523 _collection_set->add_survivor_regions(curr);
1524
1525 last = curr;
1526 }
1527 note_stop_adding_survivor_regions();
1528
1529 // Don't clear the survivor list handles until the start of
1530 // the next evacuation pause - we need it in order to re-tag
1531 // the survivor regions from this evacuation pause as 'young'
1532 // at the start of the next.
1533 }