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 void G1CollectorPolicy::print_phases() {
1252 phase_times()->print();
1253 }
1254
1255 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1256 double update_rs_processed_buffers,
1257 double goal_ms) {
1258 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1259 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1260
1261 if (G1UseAdaptiveConcRefinement) {
1262 const int k_gy = 3, k_gr = 6;
1263 const double inc_k = 1.1, dec_k = 0.9;
1264
1265 int g = cg1r->green_zone();
1266 if (update_rs_time > goal_ms) {
1267 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1268 } else {
1269 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1270 g = (int)MAX2(g * inc_k, g + 1.0);
1271 }
1272 }
1273 // Change the refinement threads params
1274 cg1r->set_green_zone(g);
1275 cg1r->set_yellow_zone(g * k_gy);
1276 cg1r->set_red_zone(g * k_gr);
1277 cg1r->reinitialize_threads();
1278
1279 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * _predictor.sigma()), 1);
1280 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1281 cg1r->yellow_zone());
1282 // Change the barrier params
1283 dcqs.set_process_completed_threshold(processing_threshold);
1284 dcqs.set_max_completed_queue(cg1r->red_zone());
1285 }
1286
1287 int curr_queue_size = dcqs.completed_buffers_num();
1288 if (curr_queue_size >= cg1r->yellow_zone()) {
1289 dcqs.set_completed_queue_padding(curr_queue_size);
1290 } else {
1291 dcqs.set_completed_queue_padding(0);
1292 }
1293 dcqs.notify_if_necessary();
1294 }
1295
1296 size_t G1CollectorPolicy::predict_rs_length_diff() const {
1297 return get_new_size_prediction(_rs_length_diff_seq);
1298 }
1299
1300 double G1CollectorPolicy::predict_alloc_rate_ms() const {
1301 return get_new_prediction(_alloc_rate_ms_seq);
1302 }
1303
1304 double G1CollectorPolicy::predict_cost_per_card_ms() const {
1305 return get_new_prediction(_cost_per_card_ms_seq);
1306 }
1307
1308 double G1CollectorPolicy::predict_scan_hcc_ms() const {
1309 return get_new_prediction(_cost_scan_hcc_seq);
1310 }
1311
1312 double G1CollectorPolicy::predict_rs_update_time_ms(size_t pending_cards) const {
1313 return pending_cards * predict_cost_per_card_ms() + predict_scan_hcc_ms();
1314 }
1315
1316 double G1CollectorPolicy::predict_young_cards_per_entry_ratio() const {
1317 return get_new_prediction(_young_cards_per_entry_ratio_seq);
1318 }
1319
1320 double G1CollectorPolicy::predict_mixed_cards_per_entry_ratio() const {
1321 if (_mixed_cards_per_entry_ratio_seq->num() < 2) {
1322 return predict_young_cards_per_entry_ratio();
1323 } else {
1324 return get_new_prediction(_mixed_cards_per_entry_ratio_seq);
1325 }
1326 }
1327
1328 size_t G1CollectorPolicy::predict_young_card_num(size_t rs_length) const {
1329 return (size_t) (rs_length * predict_young_cards_per_entry_ratio());
1330 }
1331
1332 size_t G1CollectorPolicy::predict_non_young_card_num(size_t rs_length) const {
1333 return (size_t)(rs_length * predict_mixed_cards_per_entry_ratio());
1334 }
1335
1336 double G1CollectorPolicy::predict_rs_scan_time_ms(size_t card_num) const {
1337 if (collector_state()->gcs_are_young()) {
1338 return card_num * get_new_prediction(_cost_per_entry_ms_seq);
1339 } else {
1340 return predict_mixed_rs_scan_time_ms(card_num);
1341 }
1342 }
1343
1344 double G1CollectorPolicy::predict_mixed_rs_scan_time_ms(size_t card_num) const {
1345 if (_mixed_cost_per_entry_ms_seq->num() < 3) {
1346 return card_num * get_new_prediction(_cost_per_entry_ms_seq);
1347 } else {
1348 return card_num * get_new_prediction(_mixed_cost_per_entry_ms_seq);
1349 }
1350 }
1351
1352 double G1CollectorPolicy::predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) const {
1353 if (_cost_per_byte_ms_during_cm_seq->num() < 3) {
1354 return (1.1 * bytes_to_copy) * get_new_prediction(_cost_per_byte_ms_seq);
1355 } else {
1356 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_during_cm_seq);
1357 }
1358 }
1359
1360 double G1CollectorPolicy::predict_object_copy_time_ms(size_t bytes_to_copy) const {
1361 if (collector_state()->during_concurrent_mark()) {
1362 return predict_object_copy_time_ms_during_cm(bytes_to_copy);
1363 } else {
1364 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_seq);
1365 }
1366 }
1367
1368 double G1CollectorPolicy::predict_constant_other_time_ms() const {
1369 return get_new_prediction(_constant_other_time_ms_seq);
1370 }
1371
1372 double G1CollectorPolicy::predict_young_other_time_ms(size_t young_num) const {
1373 return young_num * get_new_prediction(_young_other_cost_per_region_ms_seq);
1374 }
1375
1376 double G1CollectorPolicy::predict_non_young_other_time_ms(size_t non_young_num) const {
1377 return non_young_num * get_new_prediction(_non_young_other_cost_per_region_ms_seq);
1378 }
1379
1380 double G1CollectorPolicy::predict_remark_time_ms() const {
1381 return get_new_prediction(_concurrent_mark_remark_times_ms);
1382 }
1383
1384 double G1CollectorPolicy::predict_cleanup_time_ms() const {
1385 return get_new_prediction(_concurrent_mark_cleanup_times_ms);
1386 }
1387
1388 double G1CollectorPolicy::predict_yg_surv_rate(int age, SurvRateGroup* surv_rate_group) const {
1389 TruncatedSeq* seq = surv_rate_group->get_seq(age);
1390 guarantee(seq->num() > 0, "There should be some young gen survivor samples available. Tried to access with age %d", age);
1391 double pred = get_new_prediction(seq);
1392 if (pred > 1.0) {
1393 pred = 1.0;
1394 }
1395 return pred;
1396 }
1397
1398 double G1CollectorPolicy::predict_yg_surv_rate(int age) const {
1399 return predict_yg_surv_rate(age, _short_lived_surv_rate_group);
1400 }
1401
1402 double G1CollectorPolicy::accum_yg_surv_rate_pred(int age) const {
1403 return _short_lived_surv_rate_group->accum_surv_rate_pred(age);
1404 }
1405
1406 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
1407 size_t scanned_cards) const {
1408 return
1409 predict_rs_update_time_ms(pending_cards) +
1410 predict_rs_scan_time_ms(scanned_cards) +
1411 predict_constant_other_time_ms();
1412 }
1413
1414 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) const {
1415 size_t rs_length = predict_rs_length_diff();
1416 size_t card_num;
1417 if (collector_state()->gcs_are_young()) {
1418 card_num = predict_young_card_num(rs_length);
1419 } else {
1420 card_num = predict_non_young_card_num(rs_length);
1421 }
1422 return predict_base_elapsed_time_ms(pending_cards, card_num);
1423 }
1424
1425 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) const {
1426 size_t bytes_to_copy;
1427 if (hr->is_marked())
1428 bytes_to_copy = hr->max_live_bytes();
1429 else {
1430 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant");
1431 int age = hr->age_in_surv_rate_group();
1432 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
1433 bytes_to_copy = (size_t) (hr->used() * yg_surv_rate);
1434 }
1435 return bytes_to_copy;
1436 }
1437
1438 double G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
1439 bool for_young_gc) const {
1440 size_t rs_length = hr->rem_set()->occupied();
1441 size_t card_num;
1442
1443 // Predicting the number of cards is based on which type of GC
1444 // we're predicting for.
1445 if (for_young_gc) {
1446 card_num = predict_young_card_num(rs_length);
1447 } else {
1448 card_num = predict_non_young_card_num(rs_length);
1449 }
1450 size_t bytes_to_copy = predict_bytes_to_copy(hr);
1451
1452 double region_elapsed_time_ms =
1453 predict_rs_scan_time_ms(card_num) +
1454 predict_object_copy_time_ms(bytes_to_copy);
1455
1456 // The prediction of the "other" time for this region is based
1457 // upon the region type and NOT the GC type.
1458 if (hr->is_young()) {
1459 region_elapsed_time_ms += predict_young_other_time_ms(1);
1460 } else {
1461 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
1462 }
1463 return region_elapsed_time_ms;
1464 }
1465
1466 void G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length,
1467 uint survivor_cset_region_length) {
1468 _eden_cset_region_length = eden_cset_region_length;
1469 _survivor_cset_region_length = survivor_cset_region_length;
1470 _old_cset_region_length = 0;
1471 }
1472
1473 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) {
1474 _recorded_rs_lengths = rs_lengths;
1475 }
1476
1477 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
1478 double elapsed_ms) {
1479 _recent_gc_times_ms->add(elapsed_ms);
1480 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
1481 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
1482 }
1483
1484 void G1CollectorPolicy::clear_ratio_check_data() {
1485 _ratio_over_threshold_count = 0;
1486 _ratio_over_threshold_sum = 0.0;
1487 _pauses_since_start = 0;
1488 }
1489
1490 size_t G1CollectorPolicy::expansion_amount() {
1491 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0;
1492 double last_gc_overhead = _last_pause_time_ratio * 100.0;
1493 double threshold = _gc_overhead_perc;
1494 size_t expand_bytes = 0;
1495
1496 // If the heap is at less than half its maximum size, scale the threshold down,
1497 // to a limit of 1. Thus the smaller the heap is, the more likely it is to expand,
1498 // though the scaling code will likely keep the increase small.
1499 if (_g1->capacity() <= _g1->max_capacity() / 2) {
1500 threshold *= (double)_g1->capacity() / (double)(_g1->max_capacity() / 2);
1501 threshold = MAX2(threshold, 1.0);
1502 }
1503
1504 // If the last GC time ratio is over the threshold, increment the count of
1505 // times it has been exceeded, and add this ratio to the sum of exceeded
1506 // ratios.
1507 if (last_gc_overhead > threshold) {
1508 _ratio_over_threshold_count++;
1509 _ratio_over_threshold_sum += last_gc_overhead;
1510 }
1511
1512 // Check if we've had enough GC time ratio checks that were over the
1513 // threshold to trigger an expansion. We'll also expand if we've
1514 // reached the end of the history buffer and the average of all entries
1515 // is still over the threshold. This indicates a smaller number of GCs were
1516 // long enough to make the average exceed the threshold.
1517 bool filled_history_buffer = _pauses_since_start == NumPrevPausesForHeuristics;
1518 if ((_ratio_over_threshold_count == MinOverThresholdForGrowth) ||
1519 (filled_history_buffer && (recent_gc_overhead > threshold))) {
1520 size_t min_expand_bytes = HeapRegion::GrainBytes;
1521 size_t reserved_bytes = _g1->max_capacity();
1522 size_t committed_bytes = _g1->capacity();
1523 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
1524 size_t expand_bytes_via_pct =
1525 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
1526 double scale_factor = 1.0;
1527
1528 // If the current size is less than 1/4 of the Initial heap size, expand
1529 // by half of the delta between the current and Initial sizes. IE, grow
1530 // back quickly.
1531 //
1532 // Otherwise, take the current size, or G1ExpandByPercentOfAvailable % of
1533 // the available expansion space, whichever is smaller, as the base
1534 // expansion size. Then possibly scale this size according to how much the
1535 // threshold has (on average) been exceeded by. If the delta is small
1536 // (less than the StartScaleDownAt value), scale the size down linearly, but
1537 // not by less than MinScaleDownFactor. If the delta is large (greater than
1538 // the StartScaleUpAt value), scale up, but adding no more than MaxScaleUpFactor
1539 // times the base size. The scaling will be linear in the range from
1540 // StartScaleUpAt to (StartScaleUpAt + ScaleUpRange). In other words,
1541 // ScaleUpRange sets the rate of scaling up.
1542 if (committed_bytes < InitialHeapSize / 4) {
1543 expand_bytes = (InitialHeapSize - committed_bytes) / 2;
1544 } else {
1545 double const MinScaleDownFactor = 0.2;
1546 double const MaxScaleUpFactor = 2;
1547 double const StartScaleDownAt = _gc_overhead_perc;
1548 double const StartScaleUpAt = _gc_overhead_perc * 1.5;
1549 double const ScaleUpRange = _gc_overhead_perc * 2.0;
1550
1551 double ratio_delta;
1552 if (filled_history_buffer) {
1553 ratio_delta = recent_gc_overhead - threshold;
1554 } else {
1555 ratio_delta = (_ratio_over_threshold_sum/_ratio_over_threshold_count) - threshold;
1556 }
1557
1558 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
1559 if (ratio_delta < StartScaleDownAt) {
1560 scale_factor = ratio_delta / StartScaleDownAt;
1561 scale_factor = MAX2(scale_factor, MinScaleDownFactor);
1562 } else if (ratio_delta > StartScaleUpAt) {
1563 scale_factor = 1 + ((ratio_delta - StartScaleUpAt) / ScaleUpRange);
1564 scale_factor = MIN2(scale_factor, MaxScaleUpFactor);
1565 }
1566 }
1567
1568 log_debug(gc, ergo, heap)("Attempt heap expansion (recent GC overhead higher than threshold after GC) "
1569 "recent GC overhead: %1.2f %% threshold: %1.2f %% uncommitted: " SIZE_FORMAT "B base expansion amount and scale: " SIZE_FORMAT "B (%1.2f%%)",
1570 recent_gc_overhead, threshold, uncommitted_bytes, expand_bytes, scale_factor * 100);
1571
1572 expand_bytes = static_cast<size_t>(expand_bytes * scale_factor);
1573
1574 // Ensure the expansion size is at least the minimum growth amount
1575 // and at most the remaining uncommitted byte size.
1576 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
1577 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
1578
1579 clear_ratio_check_data();
1580 } else {
1581 // An expansion was not triggered. If we've started counting, increment
1582 // the number of checks we've made in the current window. If we've
1583 // reached the end of the window without resizing, clear the counters to
1584 // start again the next time we see a ratio above the threshold.
1585 if (_ratio_over_threshold_count > 0) {
1586 _pauses_since_start++;
1587 if (_pauses_since_start > NumPrevPausesForHeuristics) {
1588 clear_ratio_check_data();
1589 }
1590 }
1591 }
1592
1593 return expand_bytes;
1594 }
1595
1596 void G1CollectorPolicy::print_tracing_info() const {
1597 _trace_young_gen_time_data.print();
1598 _trace_old_gen_time_data.print();
1599 }
1600
1601 void G1CollectorPolicy::print_yg_surv_rate_info() const {
1602 #ifndef PRODUCT
1603 _short_lived_surv_rate_group->print_surv_rate_summary();
1604 // add this call for any other surv rate groups
1605 #endif // PRODUCT
1606 }
1607
1608 bool G1CollectorPolicy::is_young_list_full() const {
1609 uint young_list_length = _g1->young_list()->length();
1610 uint young_list_target_length = _young_list_target_length;
1611 return young_list_length >= young_list_target_length;
1612 }
1613
1614 bool G1CollectorPolicy::can_expand_young_list() const {
1615 uint young_list_length = _g1->young_list()->length();
1616 uint young_list_max_length = _young_list_max_length;
1617 return young_list_length < young_list_max_length;
1618 }
1619
1620 void G1CollectorPolicy::update_max_gc_locker_expansion() {
1621 uint expansion_region_num = 0;
1622 if (GCLockerEdenExpansionPercent > 0) {
1623 double perc = (double) GCLockerEdenExpansionPercent / 100.0;
1624 double expansion_region_num_d = perc * (double) _young_list_target_length;
1625 // We use ceiling so that if expansion_region_num_d is > 0.0 (but
1626 // less than 1.0) we'll get 1.
1627 expansion_region_num = (uint) ceil(expansion_region_num_d);
1628 } else {
1629 assert(expansion_region_num == 0, "sanity");
1630 }
1631 _young_list_max_length = _young_list_target_length + expansion_region_num;
1632 assert(_young_list_target_length <= _young_list_max_length, "post-condition");
1633 }
1634
1635 // Calculates survivor space parameters.
1636 void G1CollectorPolicy::update_survivors_policy() {
1637 double max_survivor_regions_d =
1638 (double) _young_list_target_length / (double) SurvivorRatio;
1639 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but
1640 // smaller than 1.0) we'll get 1.
1641 _max_survivor_regions = (uint) ceil(max_survivor_regions_d);
1642
1643 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
1644 HeapRegion::GrainWords * _max_survivor_regions, counters());
1645 }
1646
1647 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(GCCause::Cause gc_cause) {
1648 // We actually check whether we are marking here and not if we are in a
1649 // reclamation phase. This means that we will schedule a concurrent mark
1650 // even while we are still in the process of reclaiming memory.
1651 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
1652 if (!during_cycle) {
1653 log_debug(gc, ergo)("Request concurrent cycle initiation (requested by GC cause). GC cause: %s", GCCause::to_string(gc_cause));
1654 collector_state()->set_initiate_conc_mark_if_possible(true);
1655 return true;
1656 } else {
1657 log_debug(gc, ergo)("Do not request concurrent cycle initiation (concurrent cycle already in progress). GC cause: %s", GCCause::to_string(gc_cause));
1658 return false;
1659 }
1660 }
1661
1662 void G1CollectorPolicy::initiate_conc_mark() {
1663 collector_state()->set_during_initial_mark_pause(true);
1664 collector_state()->set_initiate_conc_mark_if_possible(false);
1665 }
1666
1667 void G1CollectorPolicy::decide_on_conc_mark_initiation() {
1668 // We are about to decide on whether this pause will be an
1669 // initial-mark pause.
1670
1671 // First, collector_state()->during_initial_mark_pause() should not be already set. We
1672 // will set it here if we have to. However, it should be cleared by
1673 // the end of the pause (it's only set for the duration of an
1674 // initial-mark pause).
1675 assert(!collector_state()->during_initial_mark_pause(), "pre-condition");
1676
1677 if (collector_state()->initiate_conc_mark_if_possible()) {
1678 // We had noticed on a previous pause that the heap occupancy has
1679 // gone over the initiating threshold and we should start a
1680 // concurrent marking cycle. So we might initiate one.
1681
1682 if (!about_to_start_mixed_phase() && collector_state()->gcs_are_young()) {
1683 // Initiate a new initial mark if there is no marking or reclamation going on.
1684 initiate_conc_mark();
1685 log_debug(gc, ergo)("Initiate concurrent cycle (concurrent cycle initiation requested)");
1686 } else if (_g1->is_user_requested_concurrent_full_gc(_g1->gc_cause())) {
1687 // Initiate a user requested initial mark. An initial mark must be young only
1688 // GC, so the collector state must be updated to reflect this.
1689 collector_state()->set_gcs_are_young(true);
1690 collector_state()->set_last_young_gc(false);
1691
1692 abort_time_to_mixed_tracking();
1693 initiate_conc_mark();
1694 log_debug(gc, ergo)("Initiate concurrent cycle (user requested concurrent cycle)");
1695 } else {
1696 // The concurrent marking thread is still finishing up the
1697 // previous cycle. If we start one right now the two cycles
1698 // overlap. In particular, the concurrent marking thread might
1699 // be in the process of clearing the next marking bitmap (which
1700 // we will use for the next cycle if we start one). Starting a
1701 // cycle now will be bad given that parts of the marking
1702 // information might get cleared by the marking thread. And we
1703 // cannot wait for the marking thread to finish the cycle as it
1704 // periodically yields while clearing the next marking bitmap
1705 // and, if it's in a yield point, it's waiting for us to
1706 // finish. So, at this point we will not start a cycle and we'll
1707 // let the concurrent marking thread complete the last one.
1708 log_debug(gc, ergo)("Do not initiate concurrent cycle (concurrent cycle already in progress)");
1709 }
1710 }
1711 }
1712
1713 class ParKnownGarbageHRClosure: public HeapRegionClosure {
1714 G1CollectedHeap* _g1h;
1715 CSetChooserParUpdater _cset_updater;
1716
1717 public:
1718 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
1719 uint chunk_size) :
1720 _g1h(G1CollectedHeap::heap()),
1721 _cset_updater(hrSorted, true /* parallel */, chunk_size) { }
1722
1723 bool doHeapRegion(HeapRegion* r) {
1724 // Do we have any marking information for this region?
1725 if (r->is_marked()) {
1726 // We will skip any region that's currently used as an old GC
1727 // alloc region (we should not consider those for collection
1728 // before we fill them up).
1729 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) {
1730 _cset_updater.add_region(r);
1731 }
1732 }
1733 return false;
1734 }
1735 };
1736
1737 class ParKnownGarbageTask: public AbstractGangTask {
1738 CollectionSetChooser* _hrSorted;
1739 uint _chunk_size;
1740 G1CollectedHeap* _g1;
1741 HeapRegionClaimer _hrclaimer;
1742
1743 public:
1744 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size, uint n_workers) :
1745 AbstractGangTask("ParKnownGarbageTask"),
1746 _hrSorted(hrSorted), _chunk_size(chunk_size),
1747 _g1(G1CollectedHeap::heap()), _hrclaimer(n_workers) {}
1748
1749 void work(uint worker_id) {
1750 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size);
1751 _g1->heap_region_par_iterate(&parKnownGarbageCl, worker_id, &_hrclaimer);
1752 }
1753 };
1754
1755 uint G1CollectorPolicy::calculate_parallel_work_chunk_size(uint n_workers, uint n_regions) const {
1756 assert(n_workers > 0, "Active gc workers should be greater than 0");
1757 const uint overpartition_factor = 4;
1758 const uint min_chunk_size = MAX2(n_regions / n_workers, 1U);
1759 return MAX2(n_regions / (n_workers * overpartition_factor), min_chunk_size);
1760 }
1761
1762 void G1CollectorPolicy::record_concurrent_mark_cleanup_end() {
1763 cset_chooser()->clear();
1764
1765 WorkGang* workers = _g1->workers();
1766 uint n_workers = workers->active_workers();
1767
1768 uint n_regions = _g1->num_regions();
1769 uint chunk_size = calculate_parallel_work_chunk_size(n_workers, n_regions);
1770 cset_chooser()->prepare_for_par_region_addition(n_workers, n_regions, chunk_size);
1771 ParKnownGarbageTask par_known_garbage_task(cset_chooser(), chunk_size, n_workers);
1772 workers->run_task(&par_known_garbage_task);
1773
1774 cset_chooser()->sort_regions();
1775
1776 double end_sec = os::elapsedTime();
1777 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0;
1778 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
1779 _prev_collection_pause_end_ms += elapsed_time_ms;
1780
1781 record_pause(Cleanup, _mark_cleanup_start_sec, end_sec);
1782 }
1783
1784 // Add the heap region at the head of the non-incremental collection set
1785 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) {
1786 assert(_inc_cset_build_state == Active, "Precondition");
1787 assert(hr->is_old(), "the region should be old");
1788
1789 assert(!hr->in_collection_set(), "should not already be in the CSet");
1790 _g1->register_old_region_with_cset(hr);
1791 hr->set_next_in_collection_set(_collection_set);
1792 _collection_set = hr;
1793 _collection_set_bytes_used_before += hr->used();
1794 size_t rs_length = hr->rem_set()->occupied();
1795 _recorded_rs_lengths += rs_length;
1796 _old_cset_region_length += 1;
1797 }
1798
1799 // Initialize the per-collection-set information
1800 void G1CollectorPolicy::start_incremental_cset_building() {
1801 assert(_inc_cset_build_state == Inactive, "Precondition");
1802
1803 _inc_cset_head = NULL;
1804 _inc_cset_tail = NULL;
1805 _inc_cset_bytes_used_before = 0;
1806
1807 _inc_cset_max_finger = 0;
1808 _inc_cset_recorded_rs_lengths = 0;
1809 _inc_cset_recorded_rs_lengths_diffs = 0;
1810 _inc_cset_predicted_elapsed_time_ms = 0.0;
1811 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
1812 _inc_cset_build_state = Active;
1813 }
1814
1815 void G1CollectorPolicy::finalize_incremental_cset_building() {
1816 assert(_inc_cset_build_state == Active, "Precondition");
1817 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
1818
1819 // The two "main" fields, _inc_cset_recorded_rs_lengths and
1820 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread
1821 // that adds a new region to the CSet. Further updates by the
1822 // concurrent refinement thread that samples the young RSet lengths
1823 // are accumulated in the *_diffs fields. Here we add the diffs to
1824 // the "main" fields.
1825
1826 if (_inc_cset_recorded_rs_lengths_diffs >= 0) {
1827 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs;
1828 } else {
1829 // This is defensive. The diff should in theory be always positive
1830 // as RSets can only grow between GCs. However, given that we
1831 // sample their size concurrently with other threads updating them
1832 // it's possible that we might get the wrong size back, which
1833 // could make the calculations somewhat inaccurate.
1834 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs);
1835 if (_inc_cset_recorded_rs_lengths >= diffs) {
1836 _inc_cset_recorded_rs_lengths -= diffs;
1837 } else {
1838 _inc_cset_recorded_rs_lengths = 0;
1839 }
1840 }
1841 _inc_cset_predicted_elapsed_time_ms +=
1842 _inc_cset_predicted_elapsed_time_ms_diffs;
1843
1844 _inc_cset_recorded_rs_lengths_diffs = 0;
1845 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
1846 }
1847
1848 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) {
1849 // This routine is used when:
1850 // * adding survivor regions to the incremental cset at the end of an
1851 // evacuation pause,
1852 // * adding the current allocation region to the incremental cset
1853 // when it is retired, and
1854 // * updating existing policy information for a region in the
1855 // incremental cset via young list RSet sampling.
1856 // Therefore this routine may be called at a safepoint by the
1857 // VM thread, or in-between safepoints by mutator threads (when
1858 // retiring the current allocation region) or a concurrent
1859 // refine thread (RSet sampling).
1860
1861 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young());
1862 size_t used_bytes = hr->used();
1863 _inc_cset_recorded_rs_lengths += rs_length;
1864 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms;
1865 _inc_cset_bytes_used_before += used_bytes;
1866
1867 // Cache the values we have added to the aggregated information
1868 // in the heap region in case we have to remove this region from
1869 // the incremental collection set, or it is updated by the
1870 // rset sampling code
1871 hr->set_recorded_rs_length(rs_length);
1872 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms);
1873 }
1874
1875 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr,
1876 size_t new_rs_length) {
1877 // Update the CSet information that is dependent on the new RS length
1878 assert(hr->is_young(), "Precondition");
1879 assert(!SafepointSynchronize::is_at_safepoint(),
1880 "should not be at a safepoint");
1881
1882 // We could have updated _inc_cset_recorded_rs_lengths and
1883 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do
1884 // that atomically, as this code is executed by a concurrent
1885 // refinement thread, potentially concurrently with a mutator thread
1886 // allocating a new region and also updating the same fields. To
1887 // avoid the atomic operations we accumulate these updates on two
1888 // separate fields (*_diffs) and we'll just add them to the "main"
1889 // fields at the start of a GC.
1890
1891 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length();
1892 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length;
1893 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff;
1894
1895 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms();
1896 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young());
1897 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms;
1898 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff;
1899
1900 hr->set_recorded_rs_length(new_rs_length);
1901 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms);
1902 }
1903
1904 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) {
1905 assert(hr->is_young(), "invariant");
1906 assert(hr->young_index_in_cset() > -1, "should have already been set");
1907 assert(_inc_cset_build_state == Active, "Precondition");
1908
1909 // We need to clear and set the cached recorded/cached collection set
1910 // information in the heap region here (before the region gets added
1911 // to the collection set). An individual heap region's cached values
1912 // are calculated, aggregated with the policy collection set info,
1913 // and cached in the heap region here (initially) and (subsequently)
1914 // by the Young List sampling code.
1915
1916 size_t rs_length = hr->rem_set()->occupied();
1917 add_to_incremental_cset_info(hr, rs_length);
1918
1919 HeapWord* hr_end = hr->end();
1920 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end);
1921
1922 assert(!hr->in_collection_set(), "invariant");
1923 _g1->register_young_region_with_cset(hr);
1924 assert(hr->next_in_collection_set() == NULL, "invariant");
1925 }
1926
1927 // Add the region at the RHS of the incremental cset
1928 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) {
1929 // We should only ever be appending survivors at the end of a pause
1930 assert(hr->is_survivor(), "Logic");
1931
1932 // Do the 'common' stuff
1933 add_region_to_incremental_cset_common(hr);
1934
1935 // Now add the region at the right hand side
1936 if (_inc_cset_tail == NULL) {
1937 assert(_inc_cset_head == NULL, "invariant");
1938 _inc_cset_head = hr;
1939 } else {
1940 _inc_cset_tail->set_next_in_collection_set(hr);
1941 }
1942 _inc_cset_tail = hr;
1943 }
1944
1945 // Add the region to the LHS of the incremental cset
1946 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) {
1947 // Survivors should be added to the RHS at the end of a pause
1948 assert(hr->is_eden(), "Logic");
1949
1950 // Do the 'common' stuff
1951 add_region_to_incremental_cset_common(hr);
1952
1953 // Add the region at the left hand side
1954 hr->set_next_in_collection_set(_inc_cset_head);
1955 if (_inc_cset_head == NULL) {
1956 assert(_inc_cset_tail == NULL, "Invariant");
1957 _inc_cset_tail = hr;
1958 }
1959 _inc_cset_head = hr;
1960 }
1961
1962 #ifndef PRODUCT
1963 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) {
1964 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be");
1965
1966 st->print_cr("\nCollection_set:");
1967 HeapRegion* csr = list_head;
1968 while (csr != NULL) {
1969 HeapRegion* next = csr->next_in_collection_set();
1970 assert(csr->in_collection_set(), "bad CS");
1971 st->print_cr(" " HR_FORMAT ", P: " PTR_FORMAT "N: " PTR_FORMAT ", age: %4d",
1972 HR_FORMAT_PARAMS(csr),
1973 p2i(csr->prev_top_at_mark_start()), p2i(csr->next_top_at_mark_start()),
1974 csr->age_in_surv_rate_group_cond());
1975 csr = next;
1976 }
1977 }
1978 #endif // !PRODUCT
1979
1980 double G1CollectorPolicy::reclaimable_bytes_perc(size_t reclaimable_bytes) const {
1981 // Returns the given amount of reclaimable bytes (that represents
1982 // the amount of reclaimable space still to be collected) as a
1983 // percentage of the current heap capacity.
1984 size_t capacity_bytes = _g1->capacity();
1985 return (double) reclaimable_bytes * 100.0 / (double) capacity_bytes;
1986 }
1987
1988 void G1CollectorPolicy::maybe_start_marking() {
1989 if (need_to_start_conc_mark("end of GC")) {
1990 // Note: this might have already been set, if during the last
1991 // pause we decided to start a cycle but at the beginning of
1992 // this pause we decided to postpone it. That's OK.
1993 collector_state()->set_initiate_conc_mark_if_possible(true);
1994 }
1995 }
1996
1997 G1CollectorPolicy::PauseKind G1CollectorPolicy::young_gc_pause_kind() const {
1998 assert(!collector_state()->full_collection(), "must be");
1999 if (collector_state()->during_initial_mark_pause()) {
2000 assert(collector_state()->last_gc_was_young(), "must be");
2001 assert(!collector_state()->last_young_gc(), "must be");
2002 return InitialMarkGC;
2003 } else if (collector_state()->last_young_gc()) {
2004 assert(!collector_state()->during_initial_mark_pause(), "must be");
2005 assert(collector_state()->last_gc_was_young(), "must be");
2006 return LastYoungGC;
2007 } else if (!collector_state()->last_gc_was_young()) {
2008 assert(!collector_state()->during_initial_mark_pause(), "must be");
2009 assert(!collector_state()->last_young_gc(), "must be");
2010 return MixedGC;
2011 } else {
2012 assert(collector_state()->last_gc_was_young(), "must be");
2013 assert(!collector_state()->during_initial_mark_pause(), "must be");
2014 assert(!collector_state()->last_young_gc(), "must be");
2015 return YoungOnlyGC;
2016 }
2017 }
2018
2019 void G1CollectorPolicy::record_pause(PauseKind kind, double start, double end) {
2020 // Manage the MMU tracker. For some reason it ignores Full GCs.
2021 if (kind != FullGC) {
2022 _mmu_tracker->add_pause(start, end);
2023 }
2024 // Manage the mutator time tracking from initial mark to first mixed gc.
2025 switch (kind) {
2026 case FullGC:
2027 abort_time_to_mixed_tracking();
2028 break;
2029 case Cleanup:
2030 case Remark:
2031 case YoungOnlyGC:
2032 case LastYoungGC:
2033 _initial_mark_to_mixed.add_pause(end - start);
2034 break;
2035 case InitialMarkGC:
2036 _initial_mark_to_mixed.record_initial_mark_end(end);
2037 break;
2038 case MixedGC:
2039 _initial_mark_to_mixed.record_mixed_gc_start(start);
2040 break;
2041 default:
2042 ShouldNotReachHere();
2043 }
2044 }
2045
2046 void G1CollectorPolicy::abort_time_to_mixed_tracking() {
2047 _initial_mark_to_mixed.reset();
2048 }
2049
2050 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str,
2051 const char* false_action_str) const {
2052 if (cset_chooser()->is_empty()) {
2053 log_debug(gc, ergo)("%s (candidate old regions not available)", false_action_str);
2054 return false;
2055 }
2056
2057 // Is the amount of uncollected reclaimable space above G1HeapWastePercent?
2058 size_t reclaimable_bytes = cset_chooser()->remaining_reclaimable_bytes();
2059 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes);
2060 double threshold = (double) G1HeapWastePercent;
2061 if (reclaimable_perc <= threshold) {
2062 log_debug(gc, ergo)("%s (reclaimable percentage not over threshold). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT,
2063 false_action_str, cset_chooser()->remaining_regions(), reclaimable_bytes, reclaimable_perc, G1HeapWastePercent);
2064 return false;
2065 }
2066 log_debug(gc, ergo)("%s (candidate old regions available). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT,
2067 true_action_str, cset_chooser()->remaining_regions(), reclaimable_bytes, reclaimable_perc, G1HeapWastePercent);
2068 return true;
2069 }
2070
2071 uint G1CollectorPolicy::calc_min_old_cset_length() const {
2072 // The min old CSet region bound is based on the maximum desired
2073 // number of mixed GCs after a cycle. I.e., even if some old regions
2074 // look expensive, we should add them to the CSet anyway to make
2075 // sure we go through the available old regions in no more than the
2076 // maximum desired number of mixed GCs.
2077 //
2078 // The calculation is based on the number of marked regions we added
2079 // to the CSet chooser in the first place, not how many remain, so
2080 // that the result is the same during all mixed GCs that follow a cycle.
2081
2082 const size_t region_num = (size_t) cset_chooser()->length();
2083 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1);
2084 size_t result = region_num / gc_num;
2085 // emulate ceiling
2086 if (result * gc_num < region_num) {
2087 result += 1;
2088 }
2089 return (uint) result;
2090 }
2091
2092 uint G1CollectorPolicy::calc_max_old_cset_length() const {
2093 // The max old CSet region bound is based on the threshold expressed
2094 // as a percentage of the heap size. I.e., it should bound the
2095 // number of old regions added to the CSet irrespective of how many
2096 // of them are available.
2097
2098 const G1CollectedHeap* g1h = G1CollectedHeap::heap();
2099 const size_t region_num = g1h->num_regions();
2100 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent;
2101 size_t result = region_num * perc / 100;
2102 // emulate ceiling
2103 if (100 * result < region_num * perc) {
2104 result += 1;
2105 }
2106 return (uint) result;
2107 }
2108
2109
2110 double G1CollectorPolicy::finalize_young_cset_part(double target_pause_time_ms) {
2111 double young_start_time_sec = os::elapsedTime();
2112
2113 YoungList* young_list = _g1->young_list();
2114 finalize_incremental_cset_building();
2115
2116 guarantee(target_pause_time_ms > 0.0,
2117 "target_pause_time_ms = %1.6lf should be positive", target_pause_time_ms);
2118 guarantee(_collection_set == NULL, "Precondition");
2119
2120 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
2121 double time_remaining_ms = MAX2(target_pause_time_ms - base_time_ms, 0.0);
2122
2123 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",
2124 _pending_cards, base_time_ms, time_remaining_ms, target_pause_time_ms);
2125
2126 collector_state()->set_last_gc_was_young(collector_state()->gcs_are_young());
2127
2128 if (collector_state()->last_gc_was_young()) {
2129 _trace_young_gen_time_data.increment_young_collection_count();
2130 } else {
2131 _trace_young_gen_time_data.increment_mixed_collection_count();
2132 }
2133
2134 // The young list is laid with the survivor regions from the previous
2135 // pause are appended to the RHS of the young list, i.e.
2136 // [Newly Young Regions ++ Survivors from last pause].
2137
2138 uint survivor_region_length = young_list->survivor_length();
2139 uint eden_region_length = young_list->eden_length();
2140 init_cset_region_lengths(eden_region_length, survivor_region_length);
2141
2142 HeapRegion* hr = young_list->first_survivor_region();
2143 while (hr != NULL) {
2144 assert(hr->is_survivor(), "badly formed young list");
2145 // There is a convention that all the young regions in the CSet
2146 // are tagged as "eden", so we do this for the survivors here. We
2147 // use the special set_eden_pre_gc() as it doesn't check that the
2148 // region is free (which is not the case here).
2149 hr->set_eden_pre_gc();
2150 hr = hr->get_next_young_region();
2151 }
2152
2153 // Clear the fields that point to the survivor list - they are all young now.
2154 young_list->clear_survivors();
2155
2156 _collection_set = _inc_cset_head;
2157 _collection_set_bytes_used_before = _inc_cset_bytes_used_before;
2158 time_remaining_ms = MAX2(time_remaining_ms - _inc_cset_predicted_elapsed_time_ms, 0.0);
2159
2160 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",
2161 eden_region_length, survivor_region_length, _inc_cset_predicted_elapsed_time_ms, target_pause_time_ms);
2162
2163 // The number of recorded young regions is the incremental
2164 // collection set's current size
2165 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths);
2166
2167 double young_end_time_sec = os::elapsedTime();
2168 phase_times()->record_young_cset_choice_time_ms((young_end_time_sec - young_start_time_sec) * 1000.0);
2169
2170 return time_remaining_ms;
2171 }
2172
2173 void G1CollectorPolicy::finalize_old_cset_part(double time_remaining_ms) {
2174 double non_young_start_time_sec = os::elapsedTime();
2175 double predicted_old_time_ms = 0.0;
2176
2177
2178 if (!collector_state()->gcs_are_young()) {
2179 cset_chooser()->verify();
2180 const uint min_old_cset_length = calc_min_old_cset_length();
2181 const uint max_old_cset_length = calc_max_old_cset_length();
2182
2183 uint expensive_region_num = 0;
2184 bool check_time_remaining = adaptive_young_list_length();
2185
2186 HeapRegion* hr = cset_chooser()->peek();
2187 while (hr != NULL) {
2188 if (old_cset_region_length() >= max_old_cset_length) {
2189 // Added maximum number of old regions to the CSet.
2190 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (old CSet region num reached max). old %u regions, max %u regions",
2191 old_cset_region_length(), max_old_cset_length);
2192 break;
2193 }
2194
2195
2196 // Stop adding regions if the remaining reclaimable space is
2197 // not above G1HeapWastePercent.
2198 size_t reclaimable_bytes = cset_chooser()->remaining_reclaimable_bytes();
2199 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes);
2200 double threshold = (double) G1HeapWastePercent;
2201 if (reclaimable_perc <= threshold) {
2202 // We've added enough old regions that the amount of uncollected
2203 // reclaimable space is at or below the waste threshold. Stop
2204 // adding old regions to the CSet.
2205 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (reclaimable percentage not over threshold). "
2206 "old %u regions, max %u regions, reclaimable: " SIZE_FORMAT "B (%1.2f%%) threshold: " UINTX_FORMAT "%%",
2207 old_cset_region_length(), max_old_cset_length, reclaimable_bytes, reclaimable_perc, G1HeapWastePercent);
2208 break;
2209 }
2210
2211 double predicted_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young());
2212 if (check_time_remaining) {
2213 if (predicted_time_ms > time_remaining_ms) {
2214 // Too expensive for the current CSet.
2215
2216 if (old_cset_region_length() >= min_old_cset_length) {
2217 // We have added the minimum number of old regions to the CSet,
2218 // we are done with this CSet.
2219 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (predicted time is too high). "
2220 "predicted time: %1.2fms, remaining time: %1.2fms old %u regions, min %u regions",
2221 predicted_time_ms, time_remaining_ms, old_cset_region_length(), min_old_cset_length);
2222 break;
2223 }
2224
2225 // We'll add it anyway given that we haven't reached the
2226 // minimum number of old regions.
2227 expensive_region_num += 1;
2228 }
2229 } else {
2230 if (old_cset_region_length() >= min_old_cset_length) {
2231 // In the non-auto-tuning case, we'll finish adding regions
2232 // to the CSet if we reach the minimum.
2233
2234 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (old CSet region num reached min). old %u regions, min %u regions",
2235 old_cset_region_length(), min_old_cset_length);
2236 break;
2237 }
2238 }
2239
2240 // We will add this region to the CSet.
2241 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0);
2242 predicted_old_time_ms += predicted_time_ms;
2243 cset_chooser()->pop(); // already have region via peek()
2244 _g1->old_set_remove(hr);
2245 add_old_region_to_cset(hr);
2246
2247 hr = cset_chooser()->peek();
2248 }
2249 if (hr == NULL) {
2250 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (candidate old regions not available)");
2251 }
2252
2253 if (expensive_region_num > 0) {
2254 // We print the information once here at the end, predicated on
2255 // whether we added any apparently expensive regions or not, to
2256 // avoid generating output per region.
2257 log_debug(gc, ergo, cset)("Added expensive regions to CSet (old CSet region num not reached min)."
2258 "old %u regions, expensive: %u regions, min %u regions, remaining time: %1.2fms",
2259 old_cset_region_length(), expensive_region_num, min_old_cset_length, time_remaining_ms);
2260 }
2261
2262 cset_chooser()->verify();
2263 }
2264
2265 stop_incremental_cset_building();
2266
2267 log_debug(gc, ergo, cset)("Finish choosing CSet. old %u regions, predicted old region time: %1.2fms, time remaining: %1.2f",
2268 old_cset_region_length(), predicted_old_time_ms, time_remaining_ms);
2269
2270 double non_young_end_time_sec = os::elapsedTime();
2271 phase_times()->record_non_young_cset_choice_time_ms((non_young_end_time_sec - non_young_start_time_sec) * 1000.0);
2272 }
2273
2274 void TraceYoungGenTimeData::record_start_collection(double time_to_stop_the_world_ms) {
2275 if(TraceYoungGenTime) {
2276 _all_stop_world_times_ms.add(time_to_stop_the_world_ms);
2277 }
2278 }
2279
2280 void TraceYoungGenTimeData::record_yield_time(double yield_time_ms) {
2281 if(TraceYoungGenTime) {
2282 _all_yield_times_ms.add(yield_time_ms);
2283 }
2284 }
2285
2286 void TraceYoungGenTimeData::record_end_collection(double pause_time_ms, G1GCPhaseTimes* phase_times) {
2287 if(TraceYoungGenTime) {
2288 _total.add(pause_time_ms);
2289 _other.add(pause_time_ms - phase_times->accounted_time_ms());
2290 _root_region_scan_wait.add(phase_times->root_region_scan_wait_time_ms());
2291 _parallel.add(phase_times->cur_collection_par_time_ms());
2292 _ext_root_scan.add(phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan));
2293 _satb_filtering.add(phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering));
2294 _update_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS));
2295 _scan_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::ScanRS));
2296 _obj_copy.add(phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy));
2297 _termination.add(phase_times->average_time_ms(G1GCPhaseTimes::Termination));
2298
2299 double parallel_known_time = phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan) +
2300 phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering) +
2301 phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS) +
2302 phase_times->average_time_ms(G1GCPhaseTimes::ScanRS) +
2303 phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy) +
2304 phase_times->average_time_ms(G1GCPhaseTimes::Termination);
2305
2306 double parallel_other_time = phase_times->cur_collection_par_time_ms() - parallel_known_time;
2307 _parallel_other.add(parallel_other_time);
2308 _clear_ct.add(phase_times->cur_clear_ct_time_ms());
2309 }
2310 }
2311
2312 void TraceYoungGenTimeData::increment_young_collection_count() {
2313 if(TraceYoungGenTime) {
2314 ++_young_pause_num;
2315 }
2316 }
2317
2318 void TraceYoungGenTimeData::increment_mixed_collection_count() {
2319 if(TraceYoungGenTime) {
2320 ++_mixed_pause_num;
2321 }
2322 }
2323
2324 void TraceYoungGenTimeData::print_summary(const char* str,
2325 const NumberSeq* seq) const {
2326 double sum = seq->sum();
2327 tty->print_cr("%-27s = %8.2lf s (avg = %8.2lf ms)",
2328 str, sum / 1000.0, seq->avg());
2329 }
2330
2331 void TraceYoungGenTimeData::print_summary_sd(const char* str,
2332 const NumberSeq* seq) const {
2333 print_summary(str, seq);
2334 tty->print_cr("%45s = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
2335 "(num", seq->num(), seq->sd(), seq->maximum());
2336 }
2337
2338 void TraceYoungGenTimeData::print() const {
2339 if (!TraceYoungGenTime) {
2340 return;
2341 }
2342
2343 tty->print_cr("ALL PAUSES");
2344 print_summary_sd(" Total", &_total);
2345 tty->cr();
2346 tty->cr();
2347 tty->print_cr(" Young GC Pauses: %8d", _young_pause_num);
2348 tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num);
2349 tty->cr();
2350
2351 tty->print_cr("EVACUATION PAUSES");
2352
2353 if (_young_pause_num == 0 && _mixed_pause_num == 0) {
2354 tty->print_cr("none");
2355 } else {
2356 print_summary_sd(" Evacuation Pauses", &_total);
2357 print_summary(" Root Region Scan Wait", &_root_region_scan_wait);
2358 print_summary(" Parallel Time", &_parallel);
2359 print_summary(" Ext Root Scanning", &_ext_root_scan);
2360 print_summary(" SATB Filtering", &_satb_filtering);
2361 print_summary(" Update RS", &_update_rs);
2362 print_summary(" Scan RS", &_scan_rs);
2363 print_summary(" Object Copy", &_obj_copy);
2364 print_summary(" Termination", &_termination);
2365 print_summary(" Parallel Other", &_parallel_other);
2366 print_summary(" Clear CT", &_clear_ct);
2367 print_summary(" Other", &_other);
2368 }
2369 tty->cr();
2370
2371 tty->print_cr("MISC");
2372 print_summary_sd(" Stop World", &_all_stop_world_times_ms);
2373 print_summary_sd(" Yields", &_all_yield_times_ms);
2374 }
2375
2376 void TraceOldGenTimeData::record_full_collection(double full_gc_time_ms) {
2377 if (TraceOldGenTime) {
2378 _all_full_gc_times.add(full_gc_time_ms);
2379 }
2380 }
2381
2382 void TraceOldGenTimeData::print() const {
2383 if (!TraceOldGenTime) {
2384 return;
2385 }
2386
2387 if (_all_full_gc_times.num() > 0) {
2388 tty->print("\n%4d full_gcs: total time = %8.2f s",
2389 _all_full_gc_times.num(),
2390 _all_full_gc_times.sum() / 1000.0);
2391 tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg());
2392 tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]",
2393 _all_full_gc_times.sd(),
2394 _all_full_gc_times.maximum());
2395 }
2396 }