1 /*
2 * Copyright (c) 2001, 2015, 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/concurrentG1Refine.hpp"
27 #include "gc/g1/concurrentMark.hpp"
28 #include "gc/g1/concurrentMarkThread.inline.hpp"
29 #include "gc/g1/g1CollectedHeap.inline.hpp"
30 #include "gc/g1/g1CollectorPolicy.hpp"
31 #include "gc/g1/g1IHOPControl.hpp"
32 #include "gc/g1/g1GCPhaseTimes.hpp"
33 #include "gc/g1/heapRegion.inline.hpp"
34 #include "gc/g1/heapRegionRemSet.hpp"
35 #include "gc/shared/gcPolicyCounters.hpp"
36 #include "runtime/arguments.hpp"
37 #include "runtime/java.hpp"
38 #include "runtime/mutexLocker.hpp"
39 #include "utilities/debug.hpp"
40 #include "utilities/pair.hpp"
41
42 // Different defaults for different number of GC threads
43 // They were chosen by running GCOld and SPECjbb on debris with different
44 // numbers of GC threads and choosing them based on the results
45
46 // all the same
47 static double rs_length_diff_defaults[] = {
48 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
49 };
50
51 static double cost_per_card_ms_defaults[] = {
52 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015
53 };
54
55 // all the same
56 static double young_cards_per_entry_ratio_defaults[] = {
57 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0
58 };
59
60 static double cost_per_entry_ms_defaults[] = {
61 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005
62 };
63
64 static double cost_per_byte_ms_defaults[] = {
65 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009
66 };
67
68 // these should be pretty consistent
69 static double constant_other_time_ms_defaults[] = {
70 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0
71 };
72
73
74 static double young_other_cost_per_region_ms_defaults[] = {
75 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1
76 };
77
78 static double non_young_other_cost_per_region_ms_defaults[] = {
79 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30
80 };
81
82 G1CollectorPolicy::G1CollectorPolicy() :
83 _predictor(G1ConfidencePercent / 100.0),
84 _parallel_gc_threads(ParallelGCThreads),
85
86 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
87 _stop_world_start(0.0),
88
89 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
90 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
91
92 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
93 _prev_collection_pause_end_ms(0.0),
94 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
95 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
96 _cost_scan_hcc_seq(new TruncatedSeq(TruncatedSeqLength)),
97 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
98 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
99 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
100 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
101 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
102 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)),
103 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
104 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
105 _non_young_other_cost_per_region_ms_seq(
106 new TruncatedSeq(TruncatedSeqLength)),
107
108 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)),
109 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)),
110
111 _pause_time_target_ms((double) MaxGCPauseMillis),
112
113 _recent_prev_end_times_for_all_gcs_sec(
114 new TruncatedSeq(NumPrevPausesForHeuristics)),
115
116 _recent_avg_pause_time_ratio(0.0),
117 _rs_lengths_prediction(0),
118 _max_survivor_regions(0),
119
120 _eden_cset_region_length(0),
121 _survivor_cset_region_length(0),
122 _old_cset_region_length(0),
123
124 _collection_set(NULL),
125 _collection_set_bytes_used_before(0),
126
127 // Incremental CSet attributes
128 _inc_cset_build_state(Inactive),
129 _inc_cset_head(NULL),
130 _inc_cset_tail(NULL),
131 _inc_cset_bytes_used_before(0),
132 _inc_cset_max_finger(NULL),
133 _inc_cset_recorded_rs_lengths(0),
134 _inc_cset_recorded_rs_lengths_diffs(0),
135 _inc_cset_predicted_elapsed_time_ms(0.0),
136 _inc_cset_predicted_elapsed_time_ms_diffs(0.0),
137
138 // add here any more surv rate groups
139 _recorded_survivor_regions(0),
140 _recorded_survivor_head(NULL),
141 _recorded_survivor_tail(NULL),
142 _survivors_age_table(true),
143
144 _gc_overhead_perc(0.0),
145
146 _bytes_allocated_in_old_since_last_gc(0),
147 _ihop_control(NULL),
148 _initial_mark_to_mixed() {
149
150 // SurvRateGroups below must be initialized after the predictor because they
151 // indirectly use it through this object passed to their constructor.
152 _short_lived_surv_rate_group =
153 new SurvRateGroup(&_predictor, "Short Lived", G1YoungSurvRateNumRegionsSummary);
154 _survivor_surv_rate_group =
155 new SurvRateGroup(&_predictor, "Survivor", G1YoungSurvRateNumRegionsSummary);
156
157 // Set up the region size and associated fields. Given that the
158 // policy is created before the heap, we have to set this up here,
159 // so it's done as soon as possible.
160
161 // It would have been natural to pass initial_heap_byte_size() and
162 // max_heap_byte_size() to setup_heap_region_size() but those have
163 // not been set up at this point since they should be aligned with
164 // the region size. So, there is a circular dependency here. We base
165 // the region size on the heap size, but the heap size should be
166 // aligned with the region size. To get around this we use the
167 // unaligned values for the heap.
168 HeapRegion::setup_heap_region_size(InitialHeapSize, MaxHeapSize);
169 HeapRegionRemSet::setup_remset_size();
170
171 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime());
172 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0;
173 clear_ratio_check_data();
174
175 _phase_times = new G1GCPhaseTimes(_parallel_gc_threads);
176
177 int index = MIN2(_parallel_gc_threads - 1, 7);
178
179 _rs_length_diff_seq->add(rs_length_diff_defaults[index]);
180 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]);
181 _cost_scan_hcc_seq->add(0.0);
182 _young_cards_per_entry_ratio_seq->add(
183 young_cards_per_entry_ratio_defaults[index]);
184 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]);
185 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]);
186 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]);
187 _young_other_cost_per_region_ms_seq->add(
188 young_other_cost_per_region_ms_defaults[index]);
189 _non_young_other_cost_per_region_ms_seq->add(
190 non_young_other_cost_per_region_ms_defaults[index]);
191
192 // Below, we might need to calculate the pause time target based on
193 // the pause interval. When we do so we are going to give G1 maximum
194 // flexibility and allow it to do pauses when it needs to. So, we'll
195 // arrange that the pause interval to be pause time target + 1 to
196 // ensure that a) the pause time target is maximized with respect to
197 // the pause interval and b) we maintain the invariant that pause
198 // time target < pause interval. If the user does not want this
199 // maximum flexibility, they will have to set the pause interval
200 // explicitly.
201
202 // First make sure that, if either parameter is set, its value is
203 // reasonable.
204 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
205 if (MaxGCPauseMillis < 1) {
206 vm_exit_during_initialization("MaxGCPauseMillis should be "
207 "greater than 0");
208 }
209 }
210 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
211 if (GCPauseIntervalMillis < 1) {
212 vm_exit_during_initialization("GCPauseIntervalMillis should be "
213 "greater than 0");
214 }
215 }
216
217 // Then, if the pause time target parameter was not set, set it to
218 // the default value.
219 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
220 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
221 // The default pause time target in G1 is 200ms
222 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200);
223 } else {
224 // We do not allow the pause interval to be set without the
225 // pause time target
226 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set "
227 "without setting MaxGCPauseMillis");
228 }
229 }
230
231 // Then, if the interval parameter was not set, set it according to
232 // the pause time target (this will also deal with the case when the
233 // pause time target is the default value).
234 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
235 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1);
236 }
237
238 // Finally, make sure that the two parameters are consistent.
239 if (MaxGCPauseMillis >= GCPauseIntervalMillis) {
240 char buffer[256];
241 jio_snprintf(buffer, 256,
242 "MaxGCPauseMillis (%u) should be less than "
243 "GCPauseIntervalMillis (%u)",
244 MaxGCPauseMillis, GCPauseIntervalMillis);
245 vm_exit_during_initialization(buffer);
246 }
247
248 double max_gc_time = (double) MaxGCPauseMillis / 1000.0;
249 double time_slice = (double) GCPauseIntervalMillis / 1000.0;
250 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time);
251
252 // start conservatively (around 50ms is about right)
253 _concurrent_mark_remark_times_ms->add(0.05);
254 _concurrent_mark_cleanup_times_ms->add(0.20);
255 _tenuring_threshold = MaxTenuringThreshold;
256
257 assert(GCTimeRatio > 0,
258 "we should have set it to a default value set_g1_gc_flags() "
259 "if a user set it to 0");
260 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio));
261
262 uintx reserve_perc = G1ReservePercent;
263 // Put an artificial ceiling on this so that it's not set to a silly value.
264 if (reserve_perc > 50) {
265 reserve_perc = 50;
266 warning("G1ReservePercent is set to a value that is too large, "
267 "it's been updated to " UINTX_FORMAT, reserve_perc);
268 }
269 _reserve_factor = (double) reserve_perc / 100.0;
270 // This will be set when the heap is expanded
271 // for the first time during initialization.
272 _reserve_regions = 0;
273
274 _cset_chooser = new CollectionSetChooser();
275 }
276
277 G1CollectorPolicy::~G1CollectorPolicy() {
278 delete _ihop_control;
279 }
280
281 double G1CollectorPolicy::get_new_prediction(TruncatedSeq const* seq) const {
282 return _predictor.get_new_prediction(seq);
283 }
284
285 size_t G1CollectorPolicy::get_new_size_prediction(TruncatedSeq const* seq) const {
286 return (size_t)get_new_prediction(seq);
287 }
288
289 void G1CollectorPolicy::initialize_alignments() {
290 _space_alignment = HeapRegion::GrainBytes;
291 size_t card_table_alignment = CardTableRS::ct_max_alignment_constraint();
292 size_t page_size = UseLargePages ? os::large_page_size() : os::vm_page_size();
293 _heap_alignment = MAX3(card_table_alignment, _space_alignment, page_size);
294 }
295
296 void G1CollectorPolicy::initialize_flags() {
297 if (G1HeapRegionSize != HeapRegion::GrainBytes) {
298 FLAG_SET_ERGO(size_t, G1HeapRegionSize, HeapRegion::GrainBytes);
299 }
300
301 if (SurvivorRatio < 1) {
302 vm_exit_during_initialization("Invalid survivor ratio specified");
303 }
304 CollectorPolicy::initialize_flags();
305 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags
306 }
307
308 void G1CollectorPolicy::post_heap_initialize() {
309 uintx max_regions = G1CollectedHeap::heap()->max_regions();
310 size_t max_young_size = (size_t)_young_gen_sizer->max_young_length(max_regions) * HeapRegion::GrainBytes;
311 if (max_young_size != MaxNewSize) {
312 FLAG_SET_ERGO(size_t, MaxNewSize, max_young_size);
313 }
314
315 _ihop_control = create_ihop_control();
316 }
317
318 G1CollectorState* G1CollectorPolicy::collector_state() const { return _g1->collector_state(); }
319
320 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true),
321 _min_desired_young_length(0), _max_desired_young_length(0) {
322 if (FLAG_IS_CMDLINE(NewRatio)) {
323 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) {
324 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio");
325 } else {
326 _sizer_kind = SizerNewRatio;
327 _adaptive_size = false;
328 return;
329 }
330 }
331
332 if (NewSize > MaxNewSize) {
333 if (FLAG_IS_CMDLINE(MaxNewSize)) {
334 warning("NewSize (" SIZE_FORMAT "k) is greater than the MaxNewSize (" SIZE_FORMAT "k). "
335 "A new max generation size of " SIZE_FORMAT "k will be used.",
336 NewSize/K, MaxNewSize/K, NewSize/K);
337 }
338 MaxNewSize = NewSize;
339 }
340
341 if (FLAG_IS_CMDLINE(NewSize)) {
342 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes),
343 1U);
344 if (FLAG_IS_CMDLINE(MaxNewSize)) {
345 _max_desired_young_length =
346 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes),
347 1U);
348 _sizer_kind = SizerMaxAndNewSize;
349 _adaptive_size = _min_desired_young_length == _max_desired_young_length;
350 } else {
351 _sizer_kind = SizerNewSizeOnly;
352 }
353 } else if (FLAG_IS_CMDLINE(MaxNewSize)) {
354 _max_desired_young_length =
355 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes),
356 1U);
357 _sizer_kind = SizerMaxNewSizeOnly;
358 }
359 }
360
361 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) {
362 uint default_value = (new_number_of_heap_regions * G1NewSizePercent) / 100;
363 return MAX2(1U, default_value);
364 }
365
366 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) {
367 uint default_value = (new_number_of_heap_regions * G1MaxNewSizePercent) / 100;
368 return MAX2(1U, default_value);
369 }
370
371 void G1YoungGenSizer::recalculate_min_max_young_length(uint number_of_heap_regions, uint* min_young_length, uint* max_young_length) {
372 assert(number_of_heap_regions > 0, "Heap must be initialized");
373
374 switch (_sizer_kind) {
375 case SizerDefaults:
376 *min_young_length = calculate_default_min_length(number_of_heap_regions);
377 *max_young_length = calculate_default_max_length(number_of_heap_regions);
378 break;
379 case SizerNewSizeOnly:
380 *max_young_length = calculate_default_max_length(number_of_heap_regions);
381 *max_young_length = MAX2(*min_young_length, *max_young_length);
382 break;
383 case SizerMaxNewSizeOnly:
384 *min_young_length = calculate_default_min_length(number_of_heap_regions);
385 *min_young_length = MIN2(*min_young_length, *max_young_length);
386 break;
387 case SizerMaxAndNewSize:
388 // Do nothing. Values set on the command line, don't update them at runtime.
389 break;
390 case SizerNewRatio:
391 *min_young_length = number_of_heap_regions / (NewRatio + 1);
392 *max_young_length = *min_young_length;
393 break;
394 default:
395 ShouldNotReachHere();
396 }
397
398 assert(*min_young_length <= *max_young_length, "Invalid min/max young gen size values");
399 }
400
401 uint G1YoungGenSizer::max_young_length(uint number_of_heap_regions) {
402 // We need to pass the desired values because recalculation may not update these
403 // values in some cases.
404 uint temp = _min_desired_young_length;
405 uint result = _max_desired_young_length;
406 recalculate_min_max_young_length(number_of_heap_regions, &temp, &result);
407 return result;
408 }
409
410 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) {
411 recalculate_min_max_young_length(new_number_of_heap_regions, &_min_desired_young_length,
412 &_max_desired_young_length);
413 }
414
415 void G1CollectorPolicy::init() {
416 // Set aside an initial future to_space.
417 _g1 = G1CollectedHeap::heap();
418
419 assert(Heap_lock->owned_by_self(), "Locking discipline.");
420
421 initialize_gc_policy_counters();
422
423 if (adaptive_young_list_length()) {
424 _young_list_fixed_length = 0;
425 } else {
426 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length();
427 }
428 _free_regions_at_end_of_collection = _g1->num_free_regions();
429
430 update_young_list_max_and_target_length();
431 // We may immediately start allocating regions and placing them on the
432 // collection set list. Initialize the per-collection set info
433 start_incremental_cset_building();
434 }
435
436 void G1CollectorPolicy::note_gc_start(uint num_active_workers) {
437 phase_times()->note_gc_start(num_active_workers);
438 }
439
440 // Create the jstat counters for the policy.
441 void G1CollectorPolicy::initialize_gc_policy_counters() {
442 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3);
443 }
444
445 bool G1CollectorPolicy::predict_will_fit(uint young_length,
446 double base_time_ms,
447 uint base_free_regions,
448 double target_pause_time_ms) const {
449 if (young_length >= base_free_regions) {
450 // end condition 1: not enough space for the young regions
451 return false;
452 }
453
454 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1);
455 size_t bytes_to_copy =
456 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
457 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
458 double young_other_time_ms = predict_young_other_time_ms(young_length);
459 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms;
460 if (pause_time_ms > target_pause_time_ms) {
461 // end condition 2: prediction is over the target pause time
462 return false;
463 }
464
465 size_t free_bytes = (base_free_regions - young_length) * HeapRegion::GrainBytes;
466
467 // When copying, we will likely need more bytes free than is live in the region.
468 // Add some safety margin to factor in the confidence of our guess, and the
469 // natural expected waste.
470 // (100.0 / G1ConfidencePercent) is a scale factor that expresses the uncertainty
471 // of the calculation: the lower the confidence, the more headroom.
472 // (100 + TargetPLABWastePct) represents the increase in expected bytes during
473 // copying due to anticipated waste in the PLABs.
474 double safety_factor = (100.0 / G1ConfidencePercent) * (100 + TargetPLABWastePct) / 100.0;
475 size_t expected_bytes_to_copy = (size_t)(safety_factor * bytes_to_copy);
476
477 if (expected_bytes_to_copy > free_bytes) {
478 // end condition 3: out-of-space
479 return false;
480 }
481
482 // success!
483 return true;
484 }
485
486 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) {
487 // re-calculate the necessary reserve
488 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor;
489 // We use ceiling so that if reserve_regions_d is > 0.0 (but
490 // smaller than 1.0) we'll get 1.
491 _reserve_regions = (uint) ceil(reserve_regions_d);
492
493 _young_gen_sizer->heap_size_changed(new_number_of_regions);
494 }
495
496 uint G1CollectorPolicy::calculate_young_list_desired_min_length(
497 uint base_min_length) const {
498 uint desired_min_length = 0;
499 if (adaptive_young_list_length()) {
500 if (_alloc_rate_ms_seq->num() > 3) {
501 double now_sec = os::elapsedTime();
502 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
503 double alloc_rate_ms = predict_alloc_rate_ms();
504 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms);
505 } else {
506 // otherwise we don't have enough info to make the prediction
507 }
508 }
509 desired_min_length += base_min_length;
510 // make sure we don't go below any user-defined minimum bound
511 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length);
512 }
513
514 uint G1CollectorPolicy::calculate_young_list_desired_max_length() const {
515 // Here, we might want to also take into account any additional
516 // constraints (i.e., user-defined minimum bound). Currently, we
517 // effectively don't set this bound.
518 return _young_gen_sizer->max_desired_young_length();
519 }
520
521 uint G1CollectorPolicy::update_young_list_max_and_target_length() {
522 return update_young_list_max_and_target_length(get_new_size_prediction(_rs_lengths_seq));
523 }
524
525 uint G1CollectorPolicy::update_young_list_max_and_target_length(size_t rs_lengths) {
526 uint unbounded_target_length = update_young_list_target_length(rs_lengths);
527 update_max_gc_locker_expansion();
528 return unbounded_target_length;
529 }
530
531 uint G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) {
532 YoungTargetLengths young_lengths = young_list_target_lengths(rs_lengths);
533 _young_list_target_length = young_lengths.first;
534 return young_lengths.second;
535 }
536
537 G1CollectorPolicy::YoungTargetLengths G1CollectorPolicy::young_list_target_lengths(size_t rs_lengths) const {
538 YoungTargetLengths result;
539
540 // Calculate the absolute and desired min bounds first.
541
542 // This is how many young regions we already have (currently: the survivors).
543 uint base_min_length = recorded_survivor_regions();
544 uint desired_min_length = calculate_young_list_desired_min_length(base_min_length);
545 // This is the absolute minimum young length. Ensure that we
546 // will at least have one eden region available for allocation.
547 uint absolute_min_length = base_min_length + MAX2(_g1->young_list()->eden_length(), (uint)1);
548 // If we shrank the young list target it should not shrink below the current size.
549 desired_min_length = MAX2(desired_min_length, absolute_min_length);
550 // Calculate the absolute and desired max bounds.
551
552 uint desired_max_length = calculate_young_list_desired_max_length();
553
554 uint young_list_target_length = 0;
555 if (adaptive_young_list_length()) {
556 if (collector_state()->gcs_are_young()) {
557 young_list_target_length =
558 calculate_young_list_target_length(rs_lengths,
559 base_min_length,
560 desired_min_length,
561 desired_max_length);
562 } else {
563 // Don't calculate anything and let the code below bound it to
564 // the desired_min_length, i.e., do the next GC as soon as
565 // possible to maximize how many old regions we can add to it.
566 }
567 } else {
568 // The user asked for a fixed young gen so we'll fix the young gen
569 // whether the next GC is young or mixed.
570 young_list_target_length = _young_list_fixed_length;
571 }
572
573 result.second = young_list_target_length;
574
575 // We will try our best not to "eat" into the reserve.
576 uint absolute_max_length = 0;
577 if (_free_regions_at_end_of_collection > _reserve_regions) {
578 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions;
579 }
580 if (desired_max_length > absolute_max_length) {
581 desired_max_length = absolute_max_length;
582 }
583
584 // Make sure we don't go over the desired max length, nor under the
585 // desired min length. In case they clash, desired_min_length wins
586 // which is why that test is second.
587 if (young_list_target_length > desired_max_length) {
588 young_list_target_length = desired_max_length;
589 }
590 if (young_list_target_length < desired_min_length) {
591 young_list_target_length = desired_min_length;
592 }
593
594 assert(young_list_target_length > recorded_survivor_regions(),
595 "we should be able to allocate at least one eden region");
596 assert(young_list_target_length >= absolute_min_length, "post-condition");
597
598 result.first = young_list_target_length;
599 return result;
600 }
601
602 uint
603 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths,
604 uint base_min_length,
605 uint desired_min_length,
606 uint desired_max_length) const {
607 assert(adaptive_young_list_length(), "pre-condition");
608 assert(collector_state()->gcs_are_young(), "only call this for young GCs");
609
610 // In case some edge-condition makes the desired max length too small...
611 if (desired_max_length <= desired_min_length) {
612 return desired_min_length;
613 }
614
615 // We'll adjust min_young_length and max_young_length not to include
616 // the already allocated young regions (i.e., so they reflect the
617 // min and max eden regions we'll allocate). The base_min_length
618 // will be reflected in the predictions by the
619 // survivor_regions_evac_time prediction.
620 assert(desired_min_length > base_min_length, "invariant");
621 uint min_young_length = desired_min_length - base_min_length;
622 assert(desired_max_length > base_min_length, "invariant");
623 uint max_young_length = desired_max_length - base_min_length;
624
625 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
626 double survivor_regions_evac_time = predict_survivor_regions_evac_time();
627 size_t pending_cards = get_new_size_prediction(_pending_cards_seq);
628 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
629 size_t scanned_cards = predict_young_card_num(adj_rs_lengths);
630 double base_time_ms =
631 predict_base_elapsed_time_ms(pending_cards, scanned_cards) +
632 survivor_regions_evac_time;
633 uint available_free_regions = _free_regions_at_end_of_collection;
634 uint base_free_regions = 0;
635 if (available_free_regions > _reserve_regions) {
636 base_free_regions = available_free_regions - _reserve_regions;
637 }
638
639 // Here, we will make sure that the shortest young length that
640 // makes sense fits within the target pause time.
641
642 if (predict_will_fit(min_young_length, base_time_ms,
643 base_free_regions, target_pause_time_ms)) {
644 // The shortest young length will fit into the target pause time;
645 // we'll now check whether the absolute maximum number of young
646 // regions will fit in the target pause time. If not, we'll do
647 // a binary search between min_young_length and max_young_length.
648 if (predict_will_fit(max_young_length, base_time_ms,
649 base_free_regions, target_pause_time_ms)) {
650 // The maximum young length will fit into the target pause time.
651 // We are done so set min young length to the maximum length (as
652 // the result is assumed to be returned in min_young_length).
653 min_young_length = max_young_length;
654 } else {
655 // The maximum possible number of young regions will not fit within
656 // the target pause time so we'll search for the optimal
657 // length. The loop invariants are:
658 //
659 // min_young_length < max_young_length
660 // min_young_length is known to fit into the target pause time
661 // max_young_length is known not to fit into the target pause time
662 //
663 // Going into the loop we know the above hold as we've just
664 // checked them. Every time around the loop we check whether
665 // the middle value between min_young_length and
666 // max_young_length fits into the target pause time. If it
667 // does, it becomes the new min. If it doesn't, it becomes
668 // the new max. This way we maintain the loop invariants.
669
670 assert(min_young_length < max_young_length, "invariant");
671 uint diff = (max_young_length - min_young_length) / 2;
672 while (diff > 0) {
673 uint young_length = min_young_length + diff;
674 if (predict_will_fit(young_length, base_time_ms,
675 base_free_regions, target_pause_time_ms)) {
676 min_young_length = young_length;
677 } else {
678 max_young_length = young_length;
679 }
680 assert(min_young_length < max_young_length, "invariant");
681 diff = (max_young_length - min_young_length) / 2;
682 }
683 // The results is min_young_length which, according to the
684 // loop invariants, should fit within the target pause time.
685
686 // These are the post-conditions of the binary search above:
687 assert(min_young_length < max_young_length,
688 "otherwise we should have discovered that max_young_length "
689 "fits into the pause target and not done the binary search");
690 assert(predict_will_fit(min_young_length, base_time_ms,
691 base_free_regions, target_pause_time_ms),
692 "min_young_length, the result of the binary search, should "
693 "fit into the pause target");
694 assert(!predict_will_fit(min_young_length + 1, base_time_ms,
695 base_free_regions, target_pause_time_ms),
696 "min_young_length, the result of the binary search, should be "
697 "optimal, so no larger length should fit into the pause target");
698 }
699 } else {
700 // Even the minimum length doesn't fit into the pause time
701 // target, return it as the result nevertheless.
702 }
703 return base_min_length + min_young_length;
704 }
705
706 double G1CollectorPolicy::predict_survivor_regions_evac_time() const {
707 double survivor_regions_evac_time = 0.0;
708 for (HeapRegion * r = _recorded_survivor_head;
709 r != NULL && r != _recorded_survivor_tail->get_next_young_region();
710 r = r->get_next_young_region()) {
711 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, collector_state()->gcs_are_young());
712 }
713 return survivor_regions_evac_time;
714 }
715
716 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() {
717 guarantee( adaptive_young_list_length(), "should not call this otherwise" );
718
719 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths();
720 if (rs_lengths > _rs_lengths_prediction) {
721 // add 10% to avoid having to recalculate often
722 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000;
723 update_rs_lengths_prediction(rs_lengths_prediction);
724
725 update_young_list_max_and_target_length(rs_lengths_prediction);
726 }
727 }
728
729 void G1CollectorPolicy::update_rs_lengths_prediction() {
730 update_rs_lengths_prediction(get_new_size_prediction(_rs_lengths_seq));
731 }
732
733 void G1CollectorPolicy::update_rs_lengths_prediction(size_t prediction) {
734 if (collector_state()->gcs_are_young() && adaptive_young_list_length()) {
735 _rs_lengths_prediction = prediction;
736 }
737 }
738
739 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size,
740 bool is_tlab,
741 bool* gc_overhead_limit_was_exceeded) {
742 guarantee(false, "Not using this policy feature yet.");
743 return NULL;
744 }
745
746 // This method controls how a collector handles one or more
747 // of its generations being fully allocated.
748 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size,
749 bool is_tlab) {
750 guarantee(false, "Not using this policy feature yet.");
751 return NULL;
752 }
753
754
755 #ifndef PRODUCT
756 bool G1CollectorPolicy::verify_young_ages() {
757 HeapRegion* head = _g1->young_list()->first_region();
758 return
759 verify_young_ages(head, _short_lived_surv_rate_group);
760 // also call verify_young_ages on any additional surv rate groups
761 }
762
763 bool
764 G1CollectorPolicy::verify_young_ages(HeapRegion* head,
765 SurvRateGroup *surv_rate_group) {
766 guarantee( surv_rate_group != NULL, "pre-condition" );
767
768 const char* name = surv_rate_group->name();
769 bool ret = true;
770 int prev_age = -1;
771
772 for (HeapRegion* curr = head;
773 curr != NULL;
774 curr = curr->get_next_young_region()) {
775 SurvRateGroup* group = curr->surv_rate_group();
776 if (group == NULL && !curr->is_survivor()) {
777 log_info(gc, verify)("## %s: encountered NULL surv_rate_group", name);
778 ret = false;
779 }
780
781 if (surv_rate_group == group) {
782 int age = curr->age_in_surv_rate_group();
783
784 if (age < 0) {
785 log_info(gc, verify)("## %s: encountered negative age", name);
786 ret = false;
787 }
788
789 if (age <= prev_age) {
790 log_info(gc, verify)("## %s: region ages are not strictly increasing (%d, %d)", name, age, prev_age);
791 ret = false;
792 }
793 prev_age = age;
794 }
795 }
796
797 return ret;
798 }
799 #endif // PRODUCT
800
801 void G1CollectorPolicy::record_full_collection_start() {
802 _full_collection_start_sec = os::elapsedTime();
803 // Release the future to-space so that it is available for compaction into.
804 collector_state()->set_full_collection(true);
805 }
806
807 void G1CollectorPolicy::record_full_collection_end() {
808 // Consider this like a collection pause for the purposes of allocation
809 // since last pause.
810 double end_sec = os::elapsedTime();
811 double full_gc_time_sec = end_sec - _full_collection_start_sec;
812 double full_gc_time_ms = full_gc_time_sec * 1000.0;
813
814 _trace_old_gen_time_data.record_full_collection(full_gc_time_ms);
815
816 update_recent_gc_times(end_sec, full_gc_time_ms);
817
818 collector_state()->set_full_collection(false);
819
820 // "Nuke" the heuristics that control the young/mixed GC
821 // transitions and make sure we start with young GCs after the Full GC.
822 collector_state()->set_gcs_are_young(true);
823 collector_state()->set_last_young_gc(false);
824 collector_state()->set_initiate_conc_mark_if_possible(need_to_start_conc_mark("end of Full GC", 0));
825 collector_state()->set_during_initial_mark_pause(false);
826 collector_state()->set_in_marking_window(false);
827 collector_state()->set_in_marking_window_im(false);
828
829 _short_lived_surv_rate_group->start_adding_regions();
830 // also call this on any additional surv rate groups
831
832 record_survivor_regions(0, NULL, NULL);
833
834 _free_regions_at_end_of_collection = _g1->num_free_regions();
835 // Reset survivors SurvRateGroup.
836 _survivor_surv_rate_group->reset();
837 update_young_list_max_and_target_length();
838 update_rs_lengths_prediction();
839 cset_chooser()->clear();
840
841 _bytes_allocated_in_old_since_last_gc = 0;
842
843 record_pause(FullGC, _full_collection_start_sec, end_sec);
844 }
845
846 void G1CollectorPolicy::record_stop_world_start() {
847 _stop_world_start = os::elapsedTime();
848 }
849
850 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec) {
851 // We only need to do this here as the policy will only be applied
852 // to the GC we're about to start. so, no point is calculating this
853 // every time we calculate / recalculate the target young length.
854 update_survivors_policy();
855
856 assert(_g1->used() == _g1->recalculate_used(),
857 "sanity, used: " SIZE_FORMAT " recalculate_used: " SIZE_FORMAT,
858 _g1->used(), _g1->recalculate_used());
859
860 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0;
861 _trace_young_gen_time_data.record_start_collection(s_w_t_ms);
862 _stop_world_start = 0.0;
863
864 phase_times()->record_cur_collection_start_sec(start_time_sec);
865 _pending_cards = _g1->pending_card_num();
866
867 _collection_set_bytes_used_before = 0;
868 _bytes_copied_during_gc = 0;
869
870 collector_state()->set_last_gc_was_young(false);
871
872 // do that for any other surv rate groups
873 _short_lived_surv_rate_group->stop_adding_regions();
874 _survivors_age_table.clear();
875
876 assert( verify_young_ages(), "region age verification" );
877 }
878
879 void G1CollectorPolicy::record_concurrent_mark_init_end(double
880 mark_init_elapsed_time_ms) {
881 collector_state()->set_during_marking(true);
882 assert(!collector_state()->initiate_conc_mark_if_possible(), "we should have cleared it by now");
883 collector_state()->set_during_initial_mark_pause(false);
884 }
885
886 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
887 _mark_remark_start_sec = os::elapsedTime();
888 collector_state()->set_during_marking(false);
889 }
890
891 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
892 double end_time_sec = os::elapsedTime();
893 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
894 _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
895 _prev_collection_pause_end_ms += elapsed_time_ms;
896
897 record_pause(Remark, _mark_remark_start_sec, end_time_sec);
898 }
899
900 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
901 _mark_cleanup_start_sec = os::elapsedTime();
902 }
903
904 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
905 bool should_continue_with_reclaim = next_gc_should_be_mixed("request last young-only gc",
906 "skip last young-only gc");
907 collector_state()->set_last_young_gc(should_continue_with_reclaim);
908 // We skip the marking phase.
909 if (!should_continue_with_reclaim) {
910 abort_time_to_mixed_tracking();
911 }
912 collector_state()->set_in_marking_window(false);
913 }
914
915 void G1CollectorPolicy::record_concurrent_pause() {
916 if (_stop_world_start > 0.0) {
917 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0;
918 _trace_young_gen_time_data.record_yield_time(yield_ms);
919 }
920 }
921
922 double G1CollectorPolicy::average_time_ms(G1GCPhaseTimes::GCParPhases phase) const {
923 return phase_times()->average_time_ms(phase);
924 }
925
926 double G1CollectorPolicy::young_other_time_ms() const {
927 return phase_times()->young_cset_choice_time_ms() +
928 phase_times()->young_free_cset_time_ms();
929 }
930
931 double G1CollectorPolicy::non_young_other_time_ms() const {
932 return phase_times()->non_young_cset_choice_time_ms() +
933 phase_times()->non_young_free_cset_time_ms();
934
935 }
936
937 double G1CollectorPolicy::other_time_ms(double pause_time_ms) const {
938 return pause_time_ms -
939 average_time_ms(G1GCPhaseTimes::UpdateRS) -
940 average_time_ms(G1GCPhaseTimes::ScanRS) -
941 average_time_ms(G1GCPhaseTimes::ObjCopy) -
942 average_time_ms(G1GCPhaseTimes::Termination);
943 }
944
945 double G1CollectorPolicy::constant_other_time_ms(double pause_time_ms) const {
946 return other_time_ms(pause_time_ms) - young_other_time_ms() - non_young_other_time_ms();
947 }
948
949 bool G1CollectorPolicy::about_to_start_mixed_phase() const {
950 return _g1->concurrent_mark()->cmThread()->during_cycle() || collector_state()->last_young_gc();
951 }
952
953 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) {
954 if (about_to_start_mixed_phase()) {
955 return false;
956 }
957
958 size_t marking_initiating_used_threshold = _ihop_control->get_conc_mark_start_threshold();
959
960 size_t cur_used_bytes = _g1->non_young_capacity_bytes();
961 size_t alloc_byte_size = alloc_word_size * HeapWordSize;
962 size_t marking_request_bytes = cur_used_bytes + alloc_byte_size;
963
964 bool result = false;
965 if (marking_request_bytes > marking_initiating_used_threshold) {
966 result = collector_state()->gcs_are_young() && !collector_state()->last_young_gc();
967 log_debug(gc, ergo, ihop)("%s occupancy: " SIZE_FORMAT "B allocation request: " SIZE_FORMAT "B threshold: " SIZE_FORMAT "B (%1.2f) source: %s",
968 result ? "Request concurrent cycle initiation (occupancy higher than threshold)" : "Do not request concurrent cycle initiation (still doing mixed collections)",
969 cur_used_bytes, alloc_byte_size, marking_initiating_used_threshold, (double) marking_initiating_used_threshold / _g1->capacity() * 100, source);
970 }
971
972 return result;
973 }
974
975 // Anything below that is considered to be zero
976 #define MIN_TIMER_GRANULARITY 0.0000001
977
978 void G1CollectorPolicy::record_collection_pause_end(double pause_time_ms, size_t cards_scanned, size_t heap_used_bytes_before_gc) {
979 double end_time_sec = os::elapsedTime();
980
981 size_t cur_used_bytes = _g1->used();
982 assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
983 bool last_pause_included_initial_mark = false;
984 bool update_stats = !_g1->evacuation_failed();
985
986 NOT_PRODUCT(_short_lived_surv_rate_group->print());
987
988 record_pause(young_gc_pause_kind(), end_time_sec - pause_time_ms / 1000.0, end_time_sec);
989
990 last_pause_included_initial_mark = collector_state()->during_initial_mark_pause();
991 if (last_pause_included_initial_mark) {
992 record_concurrent_mark_init_end(0.0);
993 } else {
994 maybe_start_marking();
995 }
996
997 double app_time_ms = (phase_times()->cur_collection_start_sec() * 1000.0 - _prev_collection_pause_end_ms);
998 if (app_time_ms < MIN_TIMER_GRANULARITY) {
999 // This usually happens due to the timer not having the required
1000 // granularity. Some Linuxes are the usual culprits.
1001 // We'll just set it to something (arbitrarily) small.
1002 app_time_ms = 1.0;
1003 }
1004
1005 if (update_stats) {
1006 _trace_young_gen_time_data.record_end_collection(pause_time_ms, phase_times());
1007 // We maintain the invariant that all objects allocated by mutator
1008 // threads will be allocated out of eden regions. So, we can use
1009 // the eden region number allocated since the previous GC to
1010 // calculate the application's allocate rate. The only exception
1011 // to that is humongous objects that are allocated separately. But
1012 // given that humongous object allocations do not really affect
1013 // either the pause's duration nor when the next pause will take
1014 // place we can safely ignore them here.
1015 uint regions_allocated = eden_cset_region_length();
1016 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
1017 _alloc_rate_ms_seq->add(alloc_rate_ms);
1018
1019 double interval_ms =
1020 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
1021 update_recent_gc_times(end_time_sec, pause_time_ms);
1022 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
1023 if (recent_avg_pause_time_ratio() < 0.0 ||
1024 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) {
1025 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
1026 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
1027 if (_recent_avg_pause_time_ratio < 0.0) {
1028 _recent_avg_pause_time_ratio = 0.0;
1029 } else {
1030 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
1031 _recent_avg_pause_time_ratio = 1.0;
1032 }
1033 }
1034
1035 // Compute the ratio of just this last pause time to the entire time range stored
1036 // in the vectors. Comparing this pause to the entire range, rather than only the
1037 // most recent interval, has the effect of smoothing over a possible transient 'burst'
1038 // of more frequent pauses that don't really reflect a change in heap occupancy.
1039 // This reduces the likelihood of a needless heap expansion being triggered.
1040 _last_pause_time_ratio =
1041 (pause_time_ms * _recent_prev_end_times_for_all_gcs_sec->num()) / interval_ms;
1042 }
1043
1044 bool new_in_marking_window = collector_state()->in_marking_window();
1045 bool new_in_marking_window_im = false;
1046 if (last_pause_included_initial_mark) {
1047 new_in_marking_window = true;
1048 new_in_marking_window_im = true;
1049 }
1050
1051 if (collector_state()->last_young_gc()) {
1052 // This is supposed to to be the "last young GC" before we start
1053 // doing mixed GCs. Here we decide whether to start mixed GCs or not.
1054 assert(!last_pause_included_initial_mark, "The last young GC is not allowed to be an initial mark GC");
1055
1056 if (next_gc_should_be_mixed("start mixed GCs",
1057 "do not start mixed GCs")) {
1058 collector_state()->set_gcs_are_young(false);
1059 } else {
1060 // We aborted the mixed GC phase early.
1061 abort_time_to_mixed_tracking();
1062 }
1063
1064 collector_state()->set_last_young_gc(false);
1065 }
1066
1067 if (!collector_state()->last_gc_was_young()) {
1068 // This is a mixed GC. Here we decide whether to continue doing
1069 // mixed GCs or not.
1070 if (!next_gc_should_be_mixed("continue mixed GCs",
1071 "do not continue mixed GCs")) {
1072 collector_state()->set_gcs_are_young(true);
1073
1074 maybe_start_marking();
1075 }
1076 }
1077
1078 _short_lived_surv_rate_group->start_adding_regions();
1079 // Do that for any other surv rate groups
1080
1081 if (update_stats) {
1082 double cost_per_card_ms = 0.0;
1083 double cost_scan_hcc = average_time_ms(G1GCPhaseTimes::ScanHCC);
1084 if (_pending_cards > 0) {
1085 cost_per_card_ms = (average_time_ms(G1GCPhaseTimes::UpdateRS) - cost_scan_hcc) / (double) _pending_cards;
1086 _cost_per_card_ms_seq->add(cost_per_card_ms);
1087 }
1088 _cost_scan_hcc_seq->add(cost_scan_hcc);
1089
1090 double cost_per_entry_ms = 0.0;
1091 if (cards_scanned > 10) {
1092 cost_per_entry_ms = average_time_ms(G1GCPhaseTimes::ScanRS) / (double) cards_scanned;
1093 if (collector_state()->last_gc_was_young()) {
1094 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
1095 } else {
1096 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms);
1097 }
1098 }
1099
1100 if (_max_rs_lengths > 0) {
1101 double cards_per_entry_ratio =
1102 (double) cards_scanned / (double) _max_rs_lengths;
1103 if (collector_state()->last_gc_was_young()) {
1104 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1105 } else {
1106 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1107 }
1108 }
1109
1110 // This is defensive. For a while _max_rs_lengths could get
1111 // smaller than _recorded_rs_lengths which was causing
1112 // rs_length_diff to get very large and mess up the RSet length
1113 // predictions. The reason was unsafe concurrent updates to the
1114 // _inc_cset_recorded_rs_lengths field which the code below guards
1115 // against (see CR 7118202). This bug has now been fixed (see CR
1116 // 7119027). However, I'm still worried that
1117 // _inc_cset_recorded_rs_lengths might still end up somewhat
1118 // inaccurate. The concurrent refinement thread calculates an
1119 // RSet's length concurrently with other CR threads updating it
1120 // which might cause it to calculate the length incorrectly (if,
1121 // say, it's in mid-coarsening). So I'll leave in the defensive
1122 // conditional below just in case.
1123 size_t rs_length_diff = 0;
1124 if (_max_rs_lengths > _recorded_rs_lengths) {
1125 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths;
1126 }
1127 _rs_length_diff_seq->add((double) rs_length_diff);
1128
1129 size_t freed_bytes = heap_used_bytes_before_gc - cur_used_bytes;
1130 size_t copied_bytes = _collection_set_bytes_used_before - freed_bytes;
1131 double cost_per_byte_ms = 0.0;
1132
1133 if (copied_bytes > 0) {
1134 cost_per_byte_ms = average_time_ms(G1GCPhaseTimes::ObjCopy) / (double) copied_bytes;
1135 if (collector_state()->in_marking_window()) {
1136 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
1137 } else {
1138 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
1139 }
1140 }
1141
1142 if (young_cset_region_length() > 0) {
1143 _young_other_cost_per_region_ms_seq->add(young_other_time_ms() /
1144 young_cset_region_length());
1145 }
1146
1147 if (old_cset_region_length() > 0) {
1148 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms() /
1149 old_cset_region_length());
1150 }
1151
1152 _constant_other_time_ms_seq->add(constant_other_time_ms(pause_time_ms));
1153
1154 _pending_cards_seq->add((double) _pending_cards);
1155 _rs_lengths_seq->add((double) _max_rs_lengths);
1156 }
1157
1158 collector_state()->set_in_marking_window(new_in_marking_window);
1159 collector_state()->set_in_marking_window_im(new_in_marking_window_im);
1160 _free_regions_at_end_of_collection = _g1->num_free_regions();
1161 // IHOP control wants to know the expected young gen length if it were not
1162 // restrained by the heap reserve. Using the actual length would make the
1163 // prediction too small and the limit the young gen every time we get to the
1164 // predicted target occupancy.
1165 size_t last_unrestrained_young_length = update_young_list_max_and_target_length();
1166 update_rs_lengths_prediction();
1167
1168 update_ihop_prediction(app_time_ms / 1000.0,
1169 _bytes_allocated_in_old_since_last_gc,
1170 last_unrestrained_young_length * HeapRegion::GrainBytes);
1171 _bytes_allocated_in_old_since_last_gc = 0;
1172
1173 _ihop_control->send_trace_event(_g1->gc_tracer_stw());
1174
1175 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
1176 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
1177
1178 double scan_hcc_time_ms = average_time_ms(G1GCPhaseTimes::ScanHCC);
1179
1180 if (update_rs_time_goal_ms < scan_hcc_time_ms) {
1181 log_debug(gc, ergo, refine)("Adjust concurrent refinement thresholds (scanning the HCC expected to take longer than Update RS time goal)."
1182 "Update RS time goal: %1.2fms Scan HCC time: %1.2fms",
1183 update_rs_time_goal_ms, scan_hcc_time_ms);
1184
1185 update_rs_time_goal_ms = 0;
1186 } else {
1187 update_rs_time_goal_ms -= scan_hcc_time_ms;
1188 }
1189 adjust_concurrent_refinement(average_time_ms(G1GCPhaseTimes::UpdateRS) - scan_hcc_time_ms,
1190 phase_times()->sum_thread_work_items(G1GCPhaseTimes::UpdateRS),
1191 update_rs_time_goal_ms);
1192
1193 cset_chooser()->verify();
1194 }
1195
1196 G1IHOPControl* G1CollectorPolicy::create_ihop_control() const {
1197 if (G1UseAdaptiveIHOP) {
1198 return new G1AdaptiveIHOPControl(InitiatingHeapOccupancyPercent,
1199 G1CollectedHeap::heap()->max_capacity(),
1200 &_predictor,
1201 G1ReservePercent,
1202 G1HeapWastePercent);
1203 } else {
1204 return new G1StaticIHOPControl(InitiatingHeapOccupancyPercent,
1205 G1CollectedHeap::heap()->max_capacity());
1206 }
1207 }
1208
1209 void G1CollectorPolicy::update_ihop_prediction(double mutator_time_s,
1210 size_t mutator_alloc_bytes,
1211 size_t young_gen_size) {
1212 // Always try to update IHOP prediction. Even evacuation failures give information
1213 // about e.g. whether to start IHOP earlier next time.
1214
1215 // Avoid using really small application times that might create samples with
1216 // very high or very low values. They may be caused by e.g. back-to-back gcs.
1217 double const min_valid_time = 1e-6;
1218
1219 bool report = false;
1220
1221 double marking_to_mixed_time = -1.0;
1222 if (!collector_state()->last_gc_was_young() && _initial_mark_to_mixed.has_result()) {
1223 marking_to_mixed_time = _initial_mark_to_mixed.last_marking_time();
1224 assert(marking_to_mixed_time > 0.0,
1225 "Initial mark to mixed time must be larger than zero but is %.3f",
1226 marking_to_mixed_time);
1227 if (marking_to_mixed_time > min_valid_time) {
1228 _ihop_control->update_marking_length(marking_to_mixed_time);
1229 report = true;
1230 }
1231 }
1232
1233 // As an approximation for the young gc promotion rates during marking we use
1234 // all of them. In many applications there are only a few if any young gcs during
1235 // marking, which makes any prediction useless. This increases the accuracy of the
1236 // prediction.
1237 if (collector_state()->last_gc_was_young() && mutator_time_s > min_valid_time) {
1238 _ihop_control->update_allocation_info(mutator_time_s, mutator_alloc_bytes, young_gen_size);
1239 report = true;
1240 }
1241
1242 if (report) {
1243 report_ihop_statistics();
1244 }
1245 }
1246
1247 void G1CollectorPolicy::report_ihop_statistics() {
1248 _ihop_control->print();
1249 }
1250
1251 #define EXT_SIZE_FORMAT "%.1f%s"
1252 #define EXT_SIZE_PARAMS(bytes) \
1253 byte_size_in_proper_unit((double)(bytes)), \
1254 proper_unit_for_byte_size((bytes))
1255
1256 void G1CollectorPolicy::print_phases() {
1257 phase_times()->print();
1258 }
1259
1260 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1261 double update_rs_processed_buffers,
1262 double goal_ms) {
1263 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1264 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1265
1266 if (G1UseAdaptiveConcRefinement) {
1267 const int k_gy = 3, k_gr = 6;
1268 const double inc_k = 1.1, dec_k = 0.9;
1269
1270 int g = cg1r->green_zone();
1271 if (update_rs_time > goal_ms) {
1272 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1273 } else {
1274 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1275 g = (int)MAX2(g * inc_k, g + 1.0);
1276 }
1277 }
1278 // Change the refinement threads params
1279 cg1r->set_green_zone(g);
1280 cg1r->set_yellow_zone(g * k_gy);
1281 cg1r->set_red_zone(g * k_gr);
1282 cg1r->reinitialize_threads();
1283
1284 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * _predictor.sigma()), 1);
1285 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1286 cg1r->yellow_zone());
1287 // Change the barrier params
1288 dcqs.set_process_completed_threshold(processing_threshold);
1289 dcqs.set_max_completed_queue(cg1r->red_zone());
1290 }
1291
1292 int curr_queue_size = dcqs.completed_buffers_num();
1293 if (curr_queue_size >= cg1r->yellow_zone()) {
1294 dcqs.set_completed_queue_padding(curr_queue_size);
1295 } else {
1296 dcqs.set_completed_queue_padding(0);
1297 }
1298 dcqs.notify_if_necessary();
1299 }
1300
1301 size_t G1CollectorPolicy::predict_rs_length_diff() const {
1302 return get_new_size_prediction(_rs_length_diff_seq);
1303 }
1304
1305 double G1CollectorPolicy::predict_alloc_rate_ms() const {
1306 return get_new_prediction(_alloc_rate_ms_seq);
1307 }
1308
1309 double G1CollectorPolicy::predict_cost_per_card_ms() const {
1310 return get_new_prediction(_cost_per_card_ms_seq);
1311 }
1312
1313 double G1CollectorPolicy::predict_scan_hcc_ms() const {
1314 return get_new_prediction(_cost_scan_hcc_seq);
1315 }
1316
1317 double G1CollectorPolicy::predict_rs_update_time_ms(size_t pending_cards) const {
1318 return pending_cards * predict_cost_per_card_ms() + predict_scan_hcc_ms();
1319 }
1320
1321 double G1CollectorPolicy::predict_young_cards_per_entry_ratio() const {
1322 return get_new_prediction(_young_cards_per_entry_ratio_seq);
1323 }
1324
1325 double G1CollectorPolicy::predict_mixed_cards_per_entry_ratio() const {
1326 if (_mixed_cards_per_entry_ratio_seq->num() < 2) {
1327 return predict_young_cards_per_entry_ratio();
1328 } else {
1329 return get_new_prediction(_mixed_cards_per_entry_ratio_seq);
1330 }
1331 }
1332
1333 size_t G1CollectorPolicy::predict_young_card_num(size_t rs_length) const {
1334 return (size_t) (rs_length * predict_young_cards_per_entry_ratio());
1335 }
1336
1337 size_t G1CollectorPolicy::predict_non_young_card_num(size_t rs_length) const {
1338 return (size_t)(rs_length * predict_mixed_cards_per_entry_ratio());
1339 }
1340
1341 double G1CollectorPolicy::predict_rs_scan_time_ms(size_t card_num) const {
1342 if (collector_state()->gcs_are_young()) {
1343 return card_num * get_new_prediction(_cost_per_entry_ms_seq);
1344 } else {
1345 return predict_mixed_rs_scan_time_ms(card_num);
1346 }
1347 }
1348
1349 double G1CollectorPolicy::predict_mixed_rs_scan_time_ms(size_t card_num) const {
1350 if (_mixed_cost_per_entry_ms_seq->num() < 3) {
1351 return card_num * get_new_prediction(_cost_per_entry_ms_seq);
1352 } else {
1353 return card_num * get_new_prediction(_mixed_cost_per_entry_ms_seq);
1354 }
1355 }
1356
1357 double G1CollectorPolicy::predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) const {
1358 if (_cost_per_byte_ms_during_cm_seq->num() < 3) {
1359 return (1.1 * bytes_to_copy) * get_new_prediction(_cost_per_byte_ms_seq);
1360 } else {
1361 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_during_cm_seq);
1362 }
1363 }
1364
1365 double G1CollectorPolicy::predict_object_copy_time_ms(size_t bytes_to_copy) const {
1366 if (collector_state()->during_concurrent_mark()) {
1367 return predict_object_copy_time_ms_during_cm(bytes_to_copy);
1368 } else {
1369 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_seq);
1370 }
1371 }
1372
1373 double G1CollectorPolicy::predict_constant_other_time_ms() const {
1374 return get_new_prediction(_constant_other_time_ms_seq);
1375 }
1376
1377 double G1CollectorPolicy::predict_young_other_time_ms(size_t young_num) const {
1378 return young_num * get_new_prediction(_young_other_cost_per_region_ms_seq);
1379 }
1380
1381 double G1CollectorPolicy::predict_non_young_other_time_ms(size_t non_young_num) const {
1382 return non_young_num * get_new_prediction(_non_young_other_cost_per_region_ms_seq);
1383 }
1384
1385 double G1CollectorPolicy::predict_remark_time_ms() const {
1386 return get_new_prediction(_concurrent_mark_remark_times_ms);
1387 }
1388
1389 double G1CollectorPolicy::predict_cleanup_time_ms() const {
1390 return get_new_prediction(_concurrent_mark_cleanup_times_ms);
1391 }
1392
1393 double G1CollectorPolicy::predict_yg_surv_rate(int age, SurvRateGroup* surv_rate_group) const {
1394 TruncatedSeq* seq = surv_rate_group->get_seq(age);
1395 guarantee(seq->num() > 0, "There should be some young gen survivor samples available. Tried to access with age %d", age);
1396 double pred = get_new_prediction(seq);
1397 if (pred > 1.0) {
1398 pred = 1.0;
1399 }
1400 return pred;
1401 }
1402
1403 double G1CollectorPolicy::predict_yg_surv_rate(int age) const {
1404 return predict_yg_surv_rate(age, _short_lived_surv_rate_group);
1405 }
1406
1407 double G1CollectorPolicy::accum_yg_surv_rate_pred(int age) const {
1408 return _short_lived_surv_rate_group->accum_surv_rate_pred(age);
1409 }
1410
1411 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
1412 size_t scanned_cards) const {
1413 return
1414 predict_rs_update_time_ms(pending_cards) +
1415 predict_rs_scan_time_ms(scanned_cards) +
1416 predict_constant_other_time_ms();
1417 }
1418
1419 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) const {
1420 size_t rs_length = predict_rs_length_diff();
1421 size_t card_num;
1422 if (collector_state()->gcs_are_young()) {
1423 card_num = predict_young_card_num(rs_length);
1424 } else {
1425 card_num = predict_non_young_card_num(rs_length);
1426 }
1427 return predict_base_elapsed_time_ms(pending_cards, card_num);
1428 }
1429
1430 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) const {
1431 size_t bytes_to_copy;
1432 if (hr->is_marked())
1433 bytes_to_copy = hr->max_live_bytes();
1434 else {
1435 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant");
1436 int age = hr->age_in_surv_rate_group();
1437 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
1438 bytes_to_copy = (size_t) (hr->used() * yg_surv_rate);
1439 }
1440 return bytes_to_copy;
1441 }
1442
1443 double G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
1444 bool for_young_gc) const {
1445 size_t rs_length = hr->rem_set()->occupied();
1446 size_t card_num;
1447
1448 // Predicting the number of cards is based on which type of GC
1449 // we're predicting for.
1450 if (for_young_gc) {
1451 card_num = predict_young_card_num(rs_length);
1452 } else {
1453 card_num = predict_non_young_card_num(rs_length);
1454 }
1455 size_t bytes_to_copy = predict_bytes_to_copy(hr);
1456
1457 double region_elapsed_time_ms =
1458 predict_rs_scan_time_ms(card_num) +
1459 predict_object_copy_time_ms(bytes_to_copy);
1460
1461 // The prediction of the "other" time for this region is based
1462 // upon the region type and NOT the GC type.
1463 if (hr->is_young()) {
1464 region_elapsed_time_ms += predict_young_other_time_ms(1);
1465 } else {
1466 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
1467 }
1468 return region_elapsed_time_ms;
1469 }
1470
1471 void G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length,
1472 uint survivor_cset_region_length) {
1473 _eden_cset_region_length = eden_cset_region_length;
1474 _survivor_cset_region_length = survivor_cset_region_length;
1475 _old_cset_region_length = 0;
1476 }
1477
1478 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) {
1479 _recorded_rs_lengths = rs_lengths;
1480 }
1481
1482 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
1483 double elapsed_ms) {
1484 _recent_gc_times_ms->add(elapsed_ms);
1485 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
1486 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
1487 }
1488
1489 void G1CollectorPolicy::clear_ratio_check_data() {
1490 _ratio_over_threshold_count = 0;
1491 _ratio_over_threshold_sum = 0.0;
1492 _pauses_since_start = 0;
1493 }
1494
1495 size_t G1CollectorPolicy::expansion_amount() {
1496 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0;
1497 double last_gc_overhead = _last_pause_time_ratio * 100.0;
1498 double threshold = _gc_overhead_perc;
1499 size_t expand_bytes = 0;
1500
1501 // If the heap is at less than half its maximum size, scale the threshold down,
1502 // to a limit of 1. Thus the smaller the heap is, the more likely it is to expand,
1503 // though the scaling code will likely keep the increase small.
1504 if (_g1->capacity() <= _g1->max_capacity() / 2) {
1505 threshold *= (double)_g1->capacity() / (double)(_g1->max_capacity() / 2);
1506 threshold = MAX2(threshold, 1.0);
1507 }
1508
1509 // If the last GC time ratio is over the threshold, increment the count of
1510 // times it has been exceeded, and add this ratio to the sum of exceeded
1511 // ratios.
1512 if (last_gc_overhead > threshold) {
1513 _ratio_over_threshold_count++;
1514 _ratio_over_threshold_sum += last_gc_overhead;
1515 }
1516
1517 // Check if we've had enough GC time ratio checks that were over the
1518 // threshold to trigger an expansion. We'll also expand if we've
1519 // reached the end of the history buffer and the average of all entries
1520 // is still over the threshold. This indicates a smaller number of GCs were
1521 // long enough to make the average exceed the threshold.
1522 bool filled_history_buffer = _pauses_since_start == NumPrevPausesForHeuristics;
1523 if ((_ratio_over_threshold_count == MinOverThresholdForGrowth) ||
1524 (filled_history_buffer && (recent_gc_overhead > threshold))) {
1525 size_t min_expand_bytes = HeapRegion::GrainBytes;
1526 size_t reserved_bytes = _g1->max_capacity();
1527 size_t committed_bytes = _g1->capacity();
1528 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
1529 size_t expand_bytes_via_pct =
1530 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
1531 double scale_factor = 1.0;
1532
1533 // If the current size is less than 1/4 of the Initial heap size, expand
1534 // by half of the delta between the current and Initial sizes. IE, grow
1535 // back quickly.
1536 //
1537 // Otherwise, take the current size, or G1ExpandByPercentOfAvailable % of
1538 // the available expansion space, whichever is smaller, as the base
1539 // expansion size. Then possibly scale this size according to how much the
1540 // threshold has (on average) been exceeded by. If the delta is small
1541 // (less than the StartScaleDownAt value), scale the size down linearly, but
1542 // not by less than MinScaleDownFactor. If the delta is large (greater than
1543 // the StartScaleUpAt value), scale up, but adding no more than MaxScaleUpFactor
1544 // times the base size. The scaling will be linear in the range from
1545 // StartScaleUpAt to (StartScaleUpAt + ScaleUpRange). In other words,
1546 // ScaleUpRange sets the rate of scaling up.
1547 if (committed_bytes < InitialHeapSize / 4) {
1548 expand_bytes = (InitialHeapSize - committed_bytes) / 2;
1549 } else {
1550 double const MinScaleDownFactor = 0.2;
1551 double const MaxScaleUpFactor = 2;
1552 double const StartScaleDownAt = _gc_overhead_perc;
1553 double const StartScaleUpAt = _gc_overhead_perc * 1.5;
1554 double const ScaleUpRange = _gc_overhead_perc * 2.0;
1555
1556 double ratio_delta;
1557 if (filled_history_buffer) {
1558 ratio_delta = recent_gc_overhead - threshold;
1559 } else {
1560 ratio_delta = (_ratio_over_threshold_sum/_ratio_over_threshold_count) - threshold;
1561 }
1562
1563 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
1564 if (ratio_delta < StartScaleDownAt) {
1565 scale_factor = ratio_delta / StartScaleDownAt;
1566 scale_factor = MAX2(scale_factor, MinScaleDownFactor);
1567 } else if (ratio_delta > StartScaleUpAt) {
1568 scale_factor = 1 + ((ratio_delta - StartScaleUpAt) / ScaleUpRange);
1569 scale_factor = MIN2(scale_factor, MaxScaleUpFactor);
1570 }
1571 }
1572
1573 log_debug(gc, ergo, heap)("Attempt heap expansion (recent GC overhead higher than threshold after GC) "
1574 "recent GC overhead: %1.2f %% threshold: %1.2f %% uncommitted: " SIZE_FORMAT "B base expansion amount and scale: " SIZE_FORMAT "B (%1.2f%%)",
1575 recent_gc_overhead, threshold, uncommitted_bytes, expand_bytes, scale_factor * 100);
1576
1577 expand_bytes = static_cast<size_t>(expand_bytes * scale_factor);
1578
1579 // Ensure the expansion size is at least the minimum growth amount
1580 // and at most the remaining uncommitted byte size.
1581 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
1582 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
1583
1584 clear_ratio_check_data();
1585 } else {
1586 // An expansion was not triggered. If we've started counting, increment
1587 // the number of checks we've made in the current window. If we've
1588 // reached the end of the window without resizing, clear the counters to
1589 // start again the next time we see a ratio above the threshold.
1590 if (_ratio_over_threshold_count > 0) {
1591 _pauses_since_start++;
1592 if (_pauses_since_start > NumPrevPausesForHeuristics) {
1593 clear_ratio_check_data();
1594 }
1595 }
1596 }
1597
1598 return expand_bytes;
1599 }
1600
1601 void G1CollectorPolicy::print_tracing_info() const {
1602 _trace_young_gen_time_data.print();
1603 _trace_old_gen_time_data.print();
1604 }
1605
1606 void G1CollectorPolicy::print_yg_surv_rate_info() const {
1607 #ifndef PRODUCT
1608 _short_lived_surv_rate_group->print_surv_rate_summary();
1609 // add this call for any other surv rate groups
1610 #endif // PRODUCT
1611 }
1612
1613 bool G1CollectorPolicy::is_young_list_full() const {
1614 uint young_list_length = _g1->young_list()->length();
1615 uint young_list_target_length = _young_list_target_length;
1616 return young_list_length >= young_list_target_length;
1617 }
1618
1619 bool G1CollectorPolicy::can_expand_young_list() const {
1620 uint young_list_length = _g1->young_list()->length();
1621 uint young_list_max_length = _young_list_max_length;
1622 return young_list_length < young_list_max_length;
1623 }
1624
1625 void G1CollectorPolicy::update_max_gc_locker_expansion() {
1626 uint expansion_region_num = 0;
1627 if (GCLockerEdenExpansionPercent > 0) {
1628 double perc = (double) GCLockerEdenExpansionPercent / 100.0;
1629 double expansion_region_num_d = perc * (double) _young_list_target_length;
1630 // We use ceiling so that if expansion_region_num_d is > 0.0 (but
1631 // less than 1.0) we'll get 1.
1632 expansion_region_num = (uint) ceil(expansion_region_num_d);
1633 } else {
1634 assert(expansion_region_num == 0, "sanity");
1635 }
1636 _young_list_max_length = _young_list_target_length + expansion_region_num;
1637 assert(_young_list_target_length <= _young_list_max_length, "post-condition");
1638 }
1639
1640 // Calculates survivor space parameters.
1641 void G1CollectorPolicy::update_survivors_policy() {
1642 double max_survivor_regions_d =
1643 (double) _young_list_target_length / (double) SurvivorRatio;
1644 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but
1645 // smaller than 1.0) we'll get 1.
1646 _max_survivor_regions = (uint) ceil(max_survivor_regions_d);
1647
1648 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
1649 HeapRegion::GrainWords * _max_survivor_regions, counters());
1650 }
1651
1652 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(GCCause::Cause gc_cause) {
1653 // We actually check whether we are marking here and not if we are in a
1654 // reclamation phase. This means that we will schedule a concurrent mark
1655 // even while we are still in the process of reclaiming memory.
1656 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
1657 if (!during_cycle) {
1658 log_debug(gc, ergo)("Request concurrent cycle initiation (requested by GC cause). GC cause: %s", GCCause::to_string(gc_cause));
1659 collector_state()->set_initiate_conc_mark_if_possible(true);
1660 return true;
1661 } else {
1662 log_debug(gc, ergo)("Do not request concurrent cycle initiation (concurrent cycle already in progress). GC cause: %s", GCCause::to_string(gc_cause));
1663 return false;
1664 }
1665 }
1666
1667 void G1CollectorPolicy::initiate_conc_mark() {
1668 collector_state()->set_during_initial_mark_pause(true);
1669 collector_state()->set_initiate_conc_mark_if_possible(false);
1670 }
1671
1672 void G1CollectorPolicy::decide_on_conc_mark_initiation() {
1673 // We are about to decide on whether this pause will be an
1674 // initial-mark pause.
1675
1676 // First, collector_state()->during_initial_mark_pause() should not be already set. We
1677 // will set it here if we have to. However, it should be cleared by
1678 // the end of the pause (it's only set for the duration of an
1679 // initial-mark pause).
1680 assert(!collector_state()->during_initial_mark_pause(), "pre-condition");
1681
1682 if (collector_state()->initiate_conc_mark_if_possible()) {
1683 // We had noticed on a previous pause that the heap occupancy has
1684 // gone over the initiating threshold and we should start a
1685 // concurrent marking cycle. So we might initiate one.
1686
1687 if (!about_to_start_mixed_phase() && collector_state()->gcs_are_young()) {
1688 // Initiate a new initial mark if there is no marking or reclamation going on.
1689 initiate_conc_mark();
1690 log_debug(gc, ergo)("Initiate concurrent cycle (concurrent cycle initiation requested)");
1691 } else if (_g1->is_user_requested_concurrent_full_gc(_g1->gc_cause())) {
1692 // Initiate a user requested initial mark. An initial mark must be young only
1693 // GC, so the collector state must be updated to reflect this.
1694 collector_state()->set_gcs_are_young(true);
1695 collector_state()->set_last_young_gc(false);
1696
1697 abort_time_to_mixed_tracking();
1698 initiate_conc_mark();
1699 log_debug(gc, ergo)("Initiate concurrent cycle (user requested concurrent cycle)");
1700 } else {
1701 // The concurrent marking thread is still finishing up the
1702 // previous cycle. If we start one right now the two cycles
1703 // overlap. In particular, the concurrent marking thread might
1704 // be in the process of clearing the next marking bitmap (which
1705 // we will use for the next cycle if we start one). Starting a
1706 // cycle now will be bad given that parts of the marking
1707 // information might get cleared by the marking thread. And we
1708 // cannot wait for the marking thread to finish the cycle as it
1709 // periodically yields while clearing the next marking bitmap
1710 // and, if it's in a yield point, it's waiting for us to
1711 // finish. So, at this point we will not start a cycle and we'll
1712 // let the concurrent marking thread complete the last one.
1713 log_debug(gc, ergo)("Do not initiate concurrent cycle (concurrent cycle already in progress)");
1714 }
1715 }
1716 }
1717
1718 class ParKnownGarbageHRClosure: public HeapRegionClosure {
1719 G1CollectedHeap* _g1h;
1720 CSetChooserParUpdater _cset_updater;
1721
1722 public:
1723 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
1724 uint chunk_size) :
1725 _g1h(G1CollectedHeap::heap()),
1726 _cset_updater(hrSorted, true /* parallel */, chunk_size) { }
1727
1728 bool doHeapRegion(HeapRegion* r) {
1729 // Do we have any marking information for this region?
1730 if (r->is_marked()) {
1731 // We will skip any region that's currently used as an old GC
1732 // alloc region (we should not consider those for collection
1733 // before we fill them up).
1734 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) {
1735 _cset_updater.add_region(r);
1736 }
1737 }
1738 return false;
1739 }
1740 };
1741
1742 class ParKnownGarbageTask: public AbstractGangTask {
1743 CollectionSetChooser* _hrSorted;
1744 uint _chunk_size;
1745 G1CollectedHeap* _g1;
1746 HeapRegionClaimer _hrclaimer;
1747
1748 public:
1749 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size, uint n_workers) :
1750 AbstractGangTask("ParKnownGarbageTask"),
1751 _hrSorted(hrSorted), _chunk_size(chunk_size),
1752 _g1(G1CollectedHeap::heap()), _hrclaimer(n_workers) {}
1753
1754 void work(uint worker_id) {
1755 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size);
1756 _g1->heap_region_par_iterate(&parKnownGarbageCl, worker_id, &_hrclaimer);
1757 }
1758 };
1759
1760 uint G1CollectorPolicy::calculate_parallel_work_chunk_size(uint n_workers, uint n_regions) const {
1761 assert(n_workers > 0, "Active gc workers should be greater than 0");
1762 const uint overpartition_factor = 4;
1763 const uint min_chunk_size = MAX2(n_regions / n_workers, 1U);
1764 return MAX2(n_regions / (n_workers * overpartition_factor), min_chunk_size);
1765 }
1766
1767 void G1CollectorPolicy::record_concurrent_mark_cleanup_end() {
1768 cset_chooser()->clear();
1769
1770 WorkGang* workers = _g1->workers();
1771 uint n_workers = workers->active_workers();
1772
1773 uint n_regions = _g1->num_regions();
1774 uint chunk_size = calculate_parallel_work_chunk_size(n_workers, n_regions);
1775 cset_chooser()->prepare_for_par_region_addition(n_workers, n_regions, chunk_size);
1776 ParKnownGarbageTask par_known_garbage_task(cset_chooser(), chunk_size, n_workers);
1777 workers->run_task(&par_known_garbage_task);
1778
1779 cset_chooser()->sort_regions();
1780
1781 double end_sec = os::elapsedTime();
1782 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0;
1783 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
1784 _prev_collection_pause_end_ms += elapsed_time_ms;
1785
1786 record_pause(Cleanup, _mark_cleanup_start_sec, end_sec);
1787 }
1788
1789 // Add the heap region at the head of the non-incremental collection set
1790 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) {
1791 assert(_inc_cset_build_state == Active, "Precondition");
1792 assert(hr->is_old(), "the region should be old");
1793
1794 assert(!hr->in_collection_set(), "should not already be in the CSet");
1795 _g1->register_old_region_with_cset(hr);
1796 hr->set_next_in_collection_set(_collection_set);
1797 _collection_set = hr;
1798 _collection_set_bytes_used_before += hr->used();
1799 size_t rs_length = hr->rem_set()->occupied();
1800 _recorded_rs_lengths += rs_length;
1801 _old_cset_region_length += 1;
1802 }
1803
1804 // Initialize the per-collection-set information
1805 void G1CollectorPolicy::start_incremental_cset_building() {
1806 assert(_inc_cset_build_state == Inactive, "Precondition");
1807
1808 _inc_cset_head = NULL;
1809 _inc_cset_tail = NULL;
1810 _inc_cset_bytes_used_before = 0;
1811
1812 _inc_cset_max_finger = 0;
1813 _inc_cset_recorded_rs_lengths = 0;
1814 _inc_cset_recorded_rs_lengths_diffs = 0;
1815 _inc_cset_predicted_elapsed_time_ms = 0.0;
1816 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
1817 _inc_cset_build_state = Active;
1818 }
1819
1820 void G1CollectorPolicy::finalize_incremental_cset_building() {
1821 assert(_inc_cset_build_state == Active, "Precondition");
1822 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
1823
1824 // The two "main" fields, _inc_cset_recorded_rs_lengths and
1825 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread
1826 // that adds a new region to the CSet. Further updates by the
1827 // concurrent refinement thread that samples the young RSet lengths
1828 // are accumulated in the *_diffs fields. Here we add the diffs to
1829 // the "main" fields.
1830
1831 if (_inc_cset_recorded_rs_lengths_diffs >= 0) {
1832 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs;
1833 } else {
1834 // This is defensive. The diff should in theory be always positive
1835 // as RSets can only grow between GCs. However, given that we
1836 // sample their size concurrently with other threads updating them
1837 // it's possible that we might get the wrong size back, which
1838 // could make the calculations somewhat inaccurate.
1839 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs);
1840 if (_inc_cset_recorded_rs_lengths >= diffs) {
1841 _inc_cset_recorded_rs_lengths -= diffs;
1842 } else {
1843 _inc_cset_recorded_rs_lengths = 0;
1844 }
1845 }
1846 _inc_cset_predicted_elapsed_time_ms +=
1847 _inc_cset_predicted_elapsed_time_ms_diffs;
1848
1849 _inc_cset_recorded_rs_lengths_diffs = 0;
1850 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
1851 }
1852
1853 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) {
1854 // This routine is used when:
1855 // * adding survivor regions to the incremental cset at the end of an
1856 // evacuation pause,
1857 // * adding the current allocation region to the incremental cset
1858 // when it is retired, and
1859 // * updating existing policy information for a region in the
1860 // incremental cset via young list RSet sampling.
1861 // Therefore this routine may be called at a safepoint by the
1862 // VM thread, or in-between safepoints by mutator threads (when
1863 // retiring the current allocation region) or a concurrent
1864 // refine thread (RSet sampling).
1865
1866 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young());
1867 size_t used_bytes = hr->used();
1868 _inc_cset_recorded_rs_lengths += rs_length;
1869 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms;
1870 _inc_cset_bytes_used_before += used_bytes;
1871
1872 // Cache the values we have added to the aggregated information
1873 // in the heap region in case we have to remove this region from
1874 // the incremental collection set, or it is updated by the
1875 // rset sampling code
1876 hr->set_recorded_rs_length(rs_length);
1877 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms);
1878 }
1879
1880 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr,
1881 size_t new_rs_length) {
1882 // Update the CSet information that is dependent on the new RS length
1883 assert(hr->is_young(), "Precondition");
1884 assert(!SafepointSynchronize::is_at_safepoint(),
1885 "should not be at a safepoint");
1886
1887 // We could have updated _inc_cset_recorded_rs_lengths and
1888 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do
1889 // that atomically, as this code is executed by a concurrent
1890 // refinement thread, potentially concurrently with a mutator thread
1891 // allocating a new region and also updating the same fields. To
1892 // avoid the atomic operations we accumulate these updates on two
1893 // separate fields (*_diffs) and we'll just add them to the "main"
1894 // fields at the start of a GC.
1895
1896 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length();
1897 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length;
1898 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff;
1899
1900 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms();
1901 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young());
1902 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms;
1903 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff;
1904
1905 hr->set_recorded_rs_length(new_rs_length);
1906 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms);
1907 }
1908
1909 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) {
1910 assert(hr->is_young(), "invariant");
1911 assert(hr->young_index_in_cset() > -1, "should have already been set");
1912 assert(_inc_cset_build_state == Active, "Precondition");
1913
1914 // We need to clear and set the cached recorded/cached collection set
1915 // information in the heap region here (before the region gets added
1916 // to the collection set). An individual heap region's cached values
1917 // are calculated, aggregated with the policy collection set info,
1918 // and cached in the heap region here (initially) and (subsequently)
1919 // by the Young List sampling code.
1920
1921 size_t rs_length = hr->rem_set()->occupied();
1922 add_to_incremental_cset_info(hr, rs_length);
1923
1924 HeapWord* hr_end = hr->end();
1925 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end);
1926
1927 assert(!hr->in_collection_set(), "invariant");
1928 _g1->register_young_region_with_cset(hr);
1929 assert(hr->next_in_collection_set() == NULL, "invariant");
1930 }
1931
1932 // Add the region at the RHS of the incremental cset
1933 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) {
1934 // We should only ever be appending survivors at the end of a pause
1935 assert(hr->is_survivor(), "Logic");
1936
1937 // Do the 'common' stuff
1938 add_region_to_incremental_cset_common(hr);
1939
1940 // Now add the region at the right hand side
1941 if (_inc_cset_tail == NULL) {
1942 assert(_inc_cset_head == NULL, "invariant");
1943 _inc_cset_head = hr;
1944 } else {
1945 _inc_cset_tail->set_next_in_collection_set(hr);
1946 }
1947 _inc_cset_tail = hr;
1948 }
1949
1950 // Add the region to the LHS of the incremental cset
1951 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) {
1952 // Survivors should be added to the RHS at the end of a pause
1953 assert(hr->is_eden(), "Logic");
1954
1955 // Do the 'common' stuff
1956 add_region_to_incremental_cset_common(hr);
1957
1958 // Add the region at the left hand side
1959 hr->set_next_in_collection_set(_inc_cset_head);
1960 if (_inc_cset_head == NULL) {
1961 assert(_inc_cset_tail == NULL, "Invariant");
1962 _inc_cset_tail = hr;
1963 }
1964 _inc_cset_head = hr;
1965 }
1966
1967 #ifndef PRODUCT
1968 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) {
1969 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be");
1970
1971 st->print_cr("\nCollection_set:");
1972 HeapRegion* csr = list_head;
1973 while (csr != NULL) {
1974 HeapRegion* next = csr->next_in_collection_set();
1975 assert(csr->in_collection_set(), "bad CS");
1976 st->print_cr(" " HR_FORMAT ", P: " PTR_FORMAT "N: " PTR_FORMAT ", age: %4d",
1977 HR_FORMAT_PARAMS(csr),
1978 p2i(csr->prev_top_at_mark_start()), p2i(csr->next_top_at_mark_start()),
1979 csr->age_in_surv_rate_group_cond());
1980 csr = next;
1981 }
1982 }
1983 #endif // !PRODUCT
1984
1985 double G1CollectorPolicy::reclaimable_bytes_perc(size_t reclaimable_bytes) const {
1986 // Returns the given amount of reclaimable bytes (that represents
1987 // the amount of reclaimable space still to be collected) as a
1988 // percentage of the current heap capacity.
1989 size_t capacity_bytes = _g1->capacity();
1990 return (double) reclaimable_bytes * 100.0 / (double) capacity_bytes;
1991 }
1992
1993 void G1CollectorPolicy::maybe_start_marking() {
1994 if (need_to_start_conc_mark("end of GC")) {
1995 // Note: this might have already been set, if during the last
1996 // pause we decided to start a cycle but at the beginning of
1997 // this pause we decided to postpone it. That's OK.
1998 collector_state()->set_initiate_conc_mark_if_possible(true);
1999 }
2000 }
2001
2002 G1CollectorPolicy::PauseKind G1CollectorPolicy::young_gc_pause_kind() const {
2003 assert(!collector_state()->full_collection(), "must be");
2004 if (collector_state()->during_initial_mark_pause()) {
2005 assert(collector_state()->last_gc_was_young(), "must be");
2006 assert(!collector_state()->last_young_gc(), "must be");
2007 return InitialMarkGC;
2008 } else if (collector_state()->last_young_gc()) {
2009 assert(!collector_state()->during_initial_mark_pause(), "must be");
2010 assert(collector_state()->last_gc_was_young(), "must be");
2011 return LastYoungGC;
2012 } else if (!collector_state()->last_gc_was_young()) {
2013 assert(!collector_state()->during_initial_mark_pause(), "must be");
2014 assert(!collector_state()->last_young_gc(), "must be");
2015 return MixedGC;
2016 } else {
2017 assert(collector_state()->last_gc_was_young(), "must be");
2018 assert(!collector_state()->during_initial_mark_pause(), "must be");
2019 assert(!collector_state()->last_young_gc(), "must be");
2020 return YoungOnlyGC;
2021 }
2022 }
2023
2024 void G1CollectorPolicy::record_pause(PauseKind kind, double start, double end) {
2025 // Manage the MMU tracker. For some reason it ignores Full GCs.
2026 if (kind != FullGC) {
2027 _mmu_tracker->add_pause(start, end);
2028 }
2029 // Manage the mutator time tracking from initial mark to first mixed gc.
2030 switch (kind) {
2031 case FullGC:
2032 abort_time_to_mixed_tracking();
2033 break;
2034 case Cleanup:
2035 case Remark:
2036 case YoungOnlyGC:
2037 case LastYoungGC:
2038 _initial_mark_to_mixed.add_pause(end - start);
2039 break;
2040 case InitialMarkGC:
2041 _initial_mark_to_mixed.record_initial_mark_end(end);
2042 break;
2043 case MixedGC:
2044 _initial_mark_to_mixed.record_mixed_gc_start(start);
2045 break;
2046 default:
2047 ShouldNotReachHere();
2048 }
2049 }
2050
2051 void G1CollectorPolicy::abort_time_to_mixed_tracking() {
2052 _initial_mark_to_mixed.reset();
2053 }
2054
2055 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str,
2056 const char* false_action_str) const {
2057 if (cset_chooser()->is_empty()) {
2058 log_debug(gc, ergo)("%s (candidate old regions not available)", false_action_str);
2059 return false;
2060 }
2061
2062 // Is the amount of uncollected reclaimable space above G1HeapWastePercent?
2063 size_t reclaimable_bytes = cset_chooser()->remaining_reclaimable_bytes();
2064 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes);
2065 double threshold = (double) G1HeapWastePercent;
2066 if (reclaimable_perc <= threshold) {
2067 log_debug(gc, ergo)("%s (reclaimable percentage not over threshold). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT,
2068 false_action_str, cset_chooser()->remaining_regions(), reclaimable_bytes, reclaimable_perc, G1HeapWastePercent);
2069 return false;
2070 }
2071 log_debug(gc, ergo)("%s (candidate old regions available). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT,
2072 true_action_str, cset_chooser()->remaining_regions(), reclaimable_bytes, reclaimable_perc, G1HeapWastePercent);
2073 return true;
2074 }
2075
2076 uint G1CollectorPolicy::calc_min_old_cset_length() const {
2077 // The min old CSet region bound is based on the maximum desired
2078 // number of mixed GCs after a cycle. I.e., even if some old regions
2079 // look expensive, we should add them to the CSet anyway to make
2080 // sure we go through the available old regions in no more than the
2081 // maximum desired number of mixed GCs.
2082 //
2083 // The calculation is based on the number of marked regions we added
2084 // to the CSet chooser in the first place, not how many remain, so
2085 // that the result is the same during all mixed GCs that follow a cycle.
2086
2087 const size_t region_num = (size_t) cset_chooser()->length();
2088 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1);
2089 size_t result = region_num / gc_num;
2090 // emulate ceiling
2091 if (result * gc_num < region_num) {
2092 result += 1;
2093 }
2094 return (uint) result;
2095 }
2096
2097 uint G1CollectorPolicy::calc_max_old_cset_length() const {
2098 // The max old CSet region bound is based on the threshold expressed
2099 // as a percentage of the heap size. I.e., it should bound the
2100 // number of old regions added to the CSet irrespective of how many
2101 // of them are available.
2102
2103 const G1CollectedHeap* g1h = G1CollectedHeap::heap();
2104 const size_t region_num = g1h->num_regions();
2105 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent;
2106 size_t result = region_num * perc / 100;
2107 // emulate ceiling
2108 if (100 * result < region_num * perc) {
2109 result += 1;
2110 }
2111 return (uint) result;
2112 }
2113
2114
2115 double G1CollectorPolicy::finalize_young_cset_part(double target_pause_time_ms) {
2116 double young_start_time_sec = os::elapsedTime();
2117
2118 YoungList* young_list = _g1->young_list();
2119 finalize_incremental_cset_building();
2120
2121 guarantee(target_pause_time_ms > 0.0,
2122 "target_pause_time_ms = %1.6lf should be positive", target_pause_time_ms);
2123 guarantee(_collection_set == NULL, "Precondition");
2124
2125 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
2126 double time_remaining_ms = MAX2(target_pause_time_ms - base_time_ms, 0.0);
2127
2128 log_trace(gc, ergo, cset)("Start choosing CSet. pending cards: " SIZE_FORMAT " predicted base time: %1.2fms remaining time: %1.2fms target pause time: %1.2fms",
2129 _pending_cards, base_time_ms, time_remaining_ms, target_pause_time_ms);
2130
2131 collector_state()->set_last_gc_was_young(collector_state()->gcs_are_young());
2132
2133 if (collector_state()->last_gc_was_young()) {
2134 _trace_young_gen_time_data.increment_young_collection_count();
2135 } else {
2136 _trace_young_gen_time_data.increment_mixed_collection_count();
2137 }
2138
2139 // The young list is laid with the survivor regions from the previous
2140 // pause are appended to the RHS of the young list, i.e.
2141 // [Newly Young Regions ++ Survivors from last pause].
2142
2143 uint survivor_region_length = young_list->survivor_length();
2144 uint eden_region_length = young_list->eden_length();
2145 init_cset_region_lengths(eden_region_length, survivor_region_length);
2146
2147 HeapRegion* hr = young_list->first_survivor_region();
2148 while (hr != NULL) {
2149 assert(hr->is_survivor(), "badly formed young list");
2150 // There is a convention that all the young regions in the CSet
2151 // are tagged as "eden", so we do this for the survivors here. We
2152 // use the special set_eden_pre_gc() as it doesn't check that the
2153 // region is free (which is not the case here).
2154 hr->set_eden_pre_gc();
2155 hr = hr->get_next_young_region();
2156 }
2157
2158 // Clear the fields that point to the survivor list - they are all young now.
2159 young_list->clear_survivors();
2160
2161 _collection_set = _inc_cset_head;
2162 _collection_set_bytes_used_before = _inc_cset_bytes_used_before;
2163 time_remaining_ms = MAX2(time_remaining_ms - _inc_cset_predicted_elapsed_time_ms, 0.0);
2164
2165 log_trace(gc, ergo, cset)("Add young regions to CSet. eden: %u regions, survivors: %u regions, predicted young region time: %1.2fms, target pause time: %1.2fms",
2166 eden_region_length, survivor_region_length, _inc_cset_predicted_elapsed_time_ms, target_pause_time_ms);
2167
2168 // The number of recorded young regions is the incremental
2169 // collection set's current size
2170 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths);
2171
2172 double young_end_time_sec = os::elapsedTime();
2173 phase_times()->record_young_cset_choice_time_ms((young_end_time_sec - young_start_time_sec) * 1000.0);
2174
2175 return time_remaining_ms;
2176 }
2177
2178 void G1CollectorPolicy::finalize_old_cset_part(double time_remaining_ms) {
2179 double non_young_start_time_sec = os::elapsedTime();
2180 double predicted_old_time_ms = 0.0;
2181
2182
2183 if (!collector_state()->gcs_are_young()) {
2184 cset_chooser()->verify();
2185 const uint min_old_cset_length = calc_min_old_cset_length();
2186 const uint max_old_cset_length = calc_max_old_cset_length();
2187
2188 uint expensive_region_num = 0;
2189 bool check_time_remaining = adaptive_young_list_length();
2190
2191 HeapRegion* hr = cset_chooser()->peek();
2192 while (hr != NULL) {
2193 if (old_cset_region_length() >= max_old_cset_length) {
2194 // Added maximum number of old regions to the CSet.
2195 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (old CSet region num reached max). old %u regions, max %u regions",
2196 old_cset_region_length(), max_old_cset_length);
2197 break;
2198 }
2199
2200
2201 // Stop adding regions if the remaining reclaimable space is
2202 // not above G1HeapWastePercent.
2203 size_t reclaimable_bytes = cset_chooser()->remaining_reclaimable_bytes();
2204 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes);
2205 double threshold = (double) G1HeapWastePercent;
2206 if (reclaimable_perc <= threshold) {
2207 // We've added enough old regions that the amount of uncollected
2208 // reclaimable space is at or below the waste threshold. Stop
2209 // adding old regions to the CSet.
2210 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (reclaimable percentage not over threshold). "
2211 "old %u regions, max %u regions, reclaimable: " SIZE_FORMAT "B (%1.2f%%) threshold: " UINTX_FORMAT "%%",
2212 old_cset_region_length(), max_old_cset_length, reclaimable_bytes, reclaimable_perc, G1HeapWastePercent);
2213 break;
2214 }
2215
2216 double predicted_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young());
2217 if (check_time_remaining) {
2218 if (predicted_time_ms > time_remaining_ms) {
2219 // Too expensive for the current CSet.
2220
2221 if (old_cset_region_length() >= min_old_cset_length) {
2222 // We have added the minimum number of old regions to the CSet,
2223 // we are done with this CSet.
2224 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (predicted time is too high). "
2225 "predicted time: %1.2fms, remaining time: %1.2fms old %u regions, min %u regions",
2226 predicted_time_ms, time_remaining_ms, old_cset_region_length(), min_old_cset_length);
2227 break;
2228 }
2229
2230 // We'll add it anyway given that we haven't reached the
2231 // minimum number of old regions.
2232 expensive_region_num += 1;
2233 }
2234 } else {
2235 if (old_cset_region_length() >= min_old_cset_length) {
2236 // In the non-auto-tuning case, we'll finish adding regions
2237 // to the CSet if we reach the minimum.
2238
2239 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (old CSet region num reached min). old %u regions, min %u regions",
2240 old_cset_region_length(), min_old_cset_length);
2241 break;
2242 }
2243 }
2244
2245 // We will add this region to the CSet.
2246 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0);
2247 predicted_old_time_ms += predicted_time_ms;
2248 cset_chooser()->pop(); // already have region via peek()
2249 _g1->old_set_remove(hr);
2250 add_old_region_to_cset(hr);
2251
2252 hr = cset_chooser()->peek();
2253 }
2254 if (hr == NULL) {
2255 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (candidate old regions not available)");
2256 }
2257
2258 if (expensive_region_num > 0) {
2259 // We print the information once here at the end, predicated on
2260 // whether we added any apparently expensive regions or not, to
2261 // avoid generating output per region.
2262 log_debug(gc, ergo, cset)("Added expensive regions to CSet (old CSet region num not reached min)."
2263 "old %u regions, expensive: %u regions, min %u regions, remaining time: %1.2fms",
2264 old_cset_region_length(), expensive_region_num, min_old_cset_length, time_remaining_ms);
2265 }
2266
2267 cset_chooser()->verify();
2268 }
2269
2270 stop_incremental_cset_building();
2271
2272 log_debug(gc, ergo, cset)("Finish choosing CSet. old %u regions, predicted old region time: %1.2fms, time remaining: %1.2f",
2273 old_cset_region_length(), predicted_old_time_ms, time_remaining_ms);
2274
2275 double non_young_end_time_sec = os::elapsedTime();
2276 phase_times()->record_non_young_cset_choice_time_ms((non_young_end_time_sec - non_young_start_time_sec) * 1000.0);
2277 }
2278
2279 void TraceYoungGenTimeData::record_start_collection(double time_to_stop_the_world_ms) {
2280 if(TraceYoungGenTime) {
2281 _all_stop_world_times_ms.add(time_to_stop_the_world_ms);
2282 }
2283 }
2284
2285 void TraceYoungGenTimeData::record_yield_time(double yield_time_ms) {
2286 if(TraceYoungGenTime) {
2287 _all_yield_times_ms.add(yield_time_ms);
2288 }
2289 }
2290
2291 void TraceYoungGenTimeData::record_end_collection(double pause_time_ms, G1GCPhaseTimes* phase_times) {
2292 if(TraceYoungGenTime) {
2293 _total.add(pause_time_ms);
2294 _other.add(pause_time_ms - phase_times->accounted_time_ms());
2295 _root_region_scan_wait.add(phase_times->root_region_scan_wait_time_ms());
2296 _parallel.add(phase_times->cur_collection_par_time_ms());
2297 _ext_root_scan.add(phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan));
2298 _satb_filtering.add(phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering));
2299 _update_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS));
2300 _scan_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::ScanRS));
2301 _obj_copy.add(phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy));
2302 _termination.add(phase_times->average_time_ms(G1GCPhaseTimes::Termination));
2303
2304 double parallel_known_time = phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan) +
2305 phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering) +
2306 phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS) +
2307 phase_times->average_time_ms(G1GCPhaseTimes::ScanRS) +
2308 phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy) +
2309 phase_times->average_time_ms(G1GCPhaseTimes::Termination);
2310
2311 double parallel_other_time = phase_times->cur_collection_par_time_ms() - parallel_known_time;
2312 _parallel_other.add(parallel_other_time);
2313 _clear_ct.add(phase_times->cur_clear_ct_time_ms());
2314 }
2315 }
2316
2317 void TraceYoungGenTimeData::increment_young_collection_count() {
2318 if(TraceYoungGenTime) {
2319 ++_young_pause_num;
2320 }
2321 }
2322
2323 void TraceYoungGenTimeData::increment_mixed_collection_count() {
2324 if(TraceYoungGenTime) {
2325 ++_mixed_pause_num;
2326 }
2327 }
2328
2329 void TraceYoungGenTimeData::print_summary(const char* str,
2330 const NumberSeq* seq) const {
2331 double sum = seq->sum();
2332 tty->print_cr("%-27s = %8.2lf s (avg = %8.2lf ms)",
2333 str, sum / 1000.0, seq->avg());
2334 }
2335
2336 void TraceYoungGenTimeData::print_summary_sd(const char* str,
2337 const NumberSeq* seq) const {
2338 print_summary(str, seq);
2339 tty->print_cr("%45s = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
2340 "(num", seq->num(), seq->sd(), seq->maximum());
2341 }
2342
2343 void TraceYoungGenTimeData::print() const {
2344 if (!TraceYoungGenTime) {
2345 return;
2346 }
2347
2348 tty->print_cr("ALL PAUSES");
2349 print_summary_sd(" Total", &_total);
2350 tty->cr();
2351 tty->cr();
2352 tty->print_cr(" Young GC Pauses: %8d", _young_pause_num);
2353 tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num);
2354 tty->cr();
2355
2356 tty->print_cr("EVACUATION PAUSES");
2357
2358 if (_young_pause_num == 0 && _mixed_pause_num == 0) {
2359 tty->print_cr("none");
2360 } else {
2361 print_summary_sd(" Evacuation Pauses", &_total);
2362 print_summary(" Root Region Scan Wait", &_root_region_scan_wait);
2363 print_summary(" Parallel Time", &_parallel);
2364 print_summary(" Ext Root Scanning", &_ext_root_scan);
2365 print_summary(" SATB Filtering", &_satb_filtering);
2366 print_summary(" Update RS", &_update_rs);
2367 print_summary(" Scan RS", &_scan_rs);
2368 print_summary(" Object Copy", &_obj_copy);
2369 print_summary(" Termination", &_termination);
2370 print_summary(" Parallel Other", &_parallel_other);
2371 print_summary(" Clear CT", &_clear_ct);
2372 print_summary(" Other", &_other);
2373 }
2374 tty->cr();
2375
2376 tty->print_cr("MISC");
2377 print_summary_sd(" Stop World", &_all_stop_world_times_ms);
2378 print_summary_sd(" Yields", &_all_yield_times_ms);
2379 }
2380
2381 void TraceOldGenTimeData::record_full_collection(double full_gc_time_ms) {
2382 if (TraceOldGenTime) {
2383 _all_full_gc_times.add(full_gc_time_ms);
2384 }
2385 }
2386
2387 void TraceOldGenTimeData::print() const {
2388 if (!TraceOldGenTime) {
2389 return;
2390 }
2391
2392 if (_all_full_gc_times.num() > 0) {
2393 tty->print("\n%4d full_gcs: total time = %8.2f s",
2394 _all_full_gc_times.num(),
2395 _all_full_gc_times.sum() / 1000.0);
2396 tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg());
2397 tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]",
2398 _all_full_gc_times.sd(),
2399 _all_full_gc_times.maximum());
2400 }
2401 }