1 /*
2 * Copyright (c) 2001, 2020, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc_implementation/g1/concurrentG1Refine.hpp"
27 #include "gc_implementation/g1/concurrentMark.hpp"
28 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
29 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
30 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
31 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
32 #include "gc_implementation/g1/g1GCPhaseTimes.hpp"
33 #include "gc_implementation/g1/g1Log.hpp"
34 #include "gc_implementation/g1/heapRegionRemSet.hpp"
35 #include "gc_implementation/shared/gcPolicyCounters.hpp"
36 #include "runtime/arguments.hpp"
37 #include "runtime/java.hpp"
38 #include "runtime/mutexLocker.hpp"
39 #include "utilities/debug.hpp"
40
41 // Different defaults for different number of GC threads
42 // They were chosen by running GCOld and SPECjbb on debris with different
43 // numbers of GC threads and choosing them based on the results
44
45 // all the same
46 static double rs_length_diff_defaults[] = {
47 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
48 };
49
50 static double cost_per_card_ms_defaults[] = {
51 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015
52 };
53
54 // all the same
55 static double young_cards_per_entry_ratio_defaults[] = {
56 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0
57 };
58
59 static double cost_per_entry_ms_defaults[] = {
60 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005
61 };
62
63 static double cost_per_byte_ms_defaults[] = {
64 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009
65 };
66
67 // these should be pretty consistent
68 static double constant_other_time_ms_defaults[] = {
69 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0
70 };
71
72
73 static double young_other_cost_per_region_ms_defaults[] = {
74 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1
75 };
76
77 static double non_young_other_cost_per_region_ms_defaults[] = {
78 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30
79 };
80
G1CollectorPolicy()81 G1CollectorPolicy::G1CollectorPolicy() :
82 _parallel_gc_threads(G1CollectedHeap::use_parallel_gc_threads()
83 ? ParallelGCThreads : 1),
84
85 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
86 _stop_world_start(0.0),
87
88 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
89 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
90
91 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
92 _prev_collection_pause_end_ms(0.0),
93 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
94 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
95 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
96 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
97 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
98 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
99 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
100 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)),
101 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
102 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
103 _non_young_other_cost_per_region_ms_seq(
104 new TruncatedSeq(TruncatedSeqLength)),
105
106 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)),
107 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)),
108
109 _pause_time_target_ms((double) MaxGCPauseMillis),
110
111 _gcs_are_young(true),
112
113 _during_marking(false),
114 _in_marking_window(false),
115 _in_marking_window_im(false),
116
117 _recent_prev_end_times_for_all_gcs_sec(
118 new TruncatedSeq(NumPrevPausesForHeuristics)),
119
120 _recent_avg_pause_time_ratio(0.0),
121
122 _initiate_conc_mark_if_possible(false),
123 _during_initial_mark_pause(false),
124 _last_young_gc(false),
125 _last_gc_was_young(false),
126
127 _eden_used_bytes_before_gc(0),
128 _survivor_used_bytes_before_gc(0),
129 _heap_used_bytes_before_gc(0),
130 _metaspace_used_bytes_before_gc(0),
131 _eden_capacity_bytes_before_gc(0),
132 _heap_capacity_bytes_before_gc(0),
133
134 _eden_cset_region_length(0),
135 _survivor_cset_region_length(0),
136 _old_cset_region_length(0),
137
138 _collection_set(NULL),
139 _collection_set_bytes_used_before(0),
140
141 // Incremental CSet attributes
142 _inc_cset_build_state(Inactive),
143 _inc_cset_head(NULL),
144 _inc_cset_tail(NULL),
145 _inc_cset_bytes_used_before(0),
146 _inc_cset_max_finger(NULL),
147 _inc_cset_recorded_rs_lengths(0),
148 _inc_cset_recorded_rs_lengths_diffs(0),
149 _inc_cset_predicted_elapsed_time_ms(0.0),
150 _inc_cset_predicted_elapsed_time_ms_diffs(0.0),
151
152 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
153 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
154 #endif // _MSC_VER
155
156 _short_lived_surv_rate_group(new SurvRateGroup(this, "Short Lived",
157 G1YoungSurvRateNumRegionsSummary)),
158 _survivor_surv_rate_group(new SurvRateGroup(this, "Survivor",
159 G1YoungSurvRateNumRegionsSummary)),
160 // add here any more surv rate groups
161 _recorded_survivor_regions(0),
162 _recorded_survivor_head(NULL),
163 _recorded_survivor_tail(NULL),
164 _survivors_age_table(true),
165
166 _gc_overhead_perc(0.0) {
167
168 // Set up the region size and associated fields. Given that the
169 // policy is created before the heap, we have to set this up here,
170 // so it's done as soon as possible.
171
172 // It would have been natural to pass initial_heap_byte_size() and
173 // max_heap_byte_size() to setup_heap_region_size() but those have
174 // not been set up at this point since they should be aligned with
175 // the region size. So, there is a circular dependency here. We base
176 // the region size on the heap size, but the heap size should be
177 // aligned with the region size. To get around this we use the
178 // unaligned values for the heap.
179 HeapRegion::setup_heap_region_size(InitialHeapSize, MaxHeapSize);
180 HeapRegionRemSet::setup_remset_size();
181
182 G1ErgoVerbose::initialize();
183 if (PrintAdaptiveSizePolicy) {
184 // Currently, we only use a single switch for all the heuristics.
185 G1ErgoVerbose::set_enabled(true);
186 // Given that we don't currently have a verboseness level
187 // parameter, we'll hardcode this to high. This can be easily
188 // changed in the future.
189 G1ErgoVerbose::set_level(ErgoHigh);
190 } else {
191 G1ErgoVerbose::set_enabled(false);
192 }
193
194 // Verify PLAB sizes
195 const size_t region_size = HeapRegion::GrainWords;
196 if (YoungPLABSize > region_size || OldPLABSize > region_size) {
197 char buffer[128];
198 jio_snprintf(buffer, sizeof(buffer), "%sPLABSize should be at most " SIZE_FORMAT,
199 OldPLABSize > region_size ? "Old" : "Young", region_size);
200 vm_exit_during_initialization(buffer);
201 }
202
203 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime());
204 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0;
205
206 _phase_times = new G1GCPhaseTimes(_parallel_gc_threads);
207
208 int index = MIN2(_parallel_gc_threads - 1, 7);
209
210 _rs_length_diff_seq->add(rs_length_diff_defaults[index]);
211 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]);
212 _young_cards_per_entry_ratio_seq->add(
213 young_cards_per_entry_ratio_defaults[index]);
214 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]);
215 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]);
216 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]);
217 _young_other_cost_per_region_ms_seq->add(
218 young_other_cost_per_region_ms_defaults[index]);
219 _non_young_other_cost_per_region_ms_seq->add(
220 non_young_other_cost_per_region_ms_defaults[index]);
221
222 // Below, we might need to calculate the pause time target based on
223 // the pause interval. When we do so we are going to give G1 maximum
224 // flexibility and allow it to do pauses when it needs to. So, we'll
225 // arrange that the pause interval to be pause time target + 1 to
226 // ensure that a) the pause time target is maximized with respect to
227 // the pause interval and b) we maintain the invariant that pause
228 // time target < pause interval. If the user does not want this
229 // maximum flexibility, they will have to set the pause interval
230 // explicitly.
231
232 // First make sure that, if either parameter is set, its value is
233 // reasonable.
234 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
235 if (MaxGCPauseMillis < 1) {
236 vm_exit_during_initialization("MaxGCPauseMillis should be "
237 "greater than 0");
238 }
239 }
240 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
241 if (GCPauseIntervalMillis < 1) {
242 vm_exit_during_initialization("GCPauseIntervalMillis should be "
243 "greater than 0");
244 }
245 }
246
247 // Then, if the pause time target parameter was not set, set it to
248 // the default value.
249 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
250 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
251 // The default pause time target in G1 is 200ms
252 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200);
253 } else {
254 // We do not allow the pause interval to be set without the
255 // pause time target
256 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set "
257 "without setting MaxGCPauseMillis");
258 }
259 }
260
261 // Then, if the interval parameter was not set, set it according to
262 // the pause time target (this will also deal with the case when the
263 // pause time target is the default value).
264 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
265 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1);
266 }
267
268 // Finally, make sure that the two parameters are consistent.
269 if (MaxGCPauseMillis >= GCPauseIntervalMillis) {
270 char buffer[256];
271 jio_snprintf(buffer, 256,
272 "MaxGCPauseMillis (%u) should be less than "
273 "GCPauseIntervalMillis (%u)",
274 MaxGCPauseMillis, GCPauseIntervalMillis);
275 vm_exit_during_initialization(buffer);
276 }
277
278 double max_gc_time = (double) MaxGCPauseMillis / 1000.0;
279 double time_slice = (double) GCPauseIntervalMillis / 1000.0;
280 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time);
281
282 uintx confidence_perc = G1ConfidencePercent;
283 // Put an artificial ceiling on this so that it's not set to a silly value.
284 if (confidence_perc > 100) {
285 confidence_perc = 100;
286 warning("G1ConfidencePercent is set to a value that is too large, "
287 "it's been updated to " UINTX_FORMAT, confidence_perc);
288 }
289 _sigma = (double) confidence_perc / 100.0;
290
291 // start conservatively (around 50ms is about right)
292 _concurrent_mark_remark_times_ms->add(0.05);
293 _concurrent_mark_cleanup_times_ms->add(0.20);
294 _tenuring_threshold = MaxTenuringThreshold;
295 // _max_survivor_regions will be calculated by
296 // update_young_list_target_length() during initialization.
297 _max_survivor_regions = 0;
298
299 assert(GCTimeRatio > 0,
300 "we should have set it to a default value set_g1_gc_flags() "
301 "if a user set it to 0");
302 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio));
303
304 uintx reserve_perc = G1ReservePercent;
305 // Put an artificial ceiling on this so that it's not set to a silly value.
306 if (reserve_perc > 50) {
307 reserve_perc = 50;
308 warning("G1ReservePercent is set to a value that is too large, "
309 "it's been updated to " UINTX_FORMAT, reserve_perc);
310 }
311 _reserve_factor = (double) reserve_perc / 100.0;
312 // This will be set when the heap is expanded
313 // for the first time during initialization.
314 _reserve_regions = 0;
315
316 _collectionSetChooser = new CollectionSetChooser();
317 }
318
initialize_alignments()319 void G1CollectorPolicy::initialize_alignments() {
320 _space_alignment = HeapRegion::GrainBytes;
321 size_t card_table_alignment = GenRemSet::max_alignment_constraint(GenRemSet::CardTable);
322 size_t page_size = UseLargePages ? os::large_page_size() : os::vm_page_size();
323 _heap_alignment = MAX3(card_table_alignment, _space_alignment, page_size);
324 }
325
initialize_flags()326 void G1CollectorPolicy::initialize_flags() {
327 if (G1HeapRegionSize != HeapRegion::GrainBytes) {
328 FLAG_SET_ERGO(uintx, G1HeapRegionSize, HeapRegion::GrainBytes);
329 }
330
331 if (SurvivorRatio < 1) {
332 vm_exit_during_initialization("Invalid survivor ratio specified");
333 }
334 CollectorPolicy::initialize_flags();
335 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags
336 }
337
post_heap_initialize()338 void G1CollectorPolicy::post_heap_initialize() {
339 uintx max_regions = G1CollectedHeap::heap()->max_regions();
340 size_t max_young_size = (size_t)_young_gen_sizer->max_young_length(max_regions) * HeapRegion::GrainBytes;
341 if (max_young_size != MaxNewSize) {
342 FLAG_SET_ERGO(uintx, MaxNewSize, max_young_size);
343 }
344 }
345
G1YoungGenSizer()346 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true),
347 _min_desired_young_length(0), _max_desired_young_length(0) {
348 if (FLAG_IS_CMDLINE(NewRatio)) {
349 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) {
350 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio");
351 } else {
352 _sizer_kind = SizerNewRatio;
353 _adaptive_size = false;
354 return;
355 }
356 }
357
358 if (NewSize > MaxNewSize) {
359 if (FLAG_IS_CMDLINE(MaxNewSize)) {
360 warning("NewSize (" SIZE_FORMAT "k) is greater than the MaxNewSize (" SIZE_FORMAT "k). "
361 "A new max generation size of " SIZE_FORMAT "k will be used.",
362 NewSize/K, MaxNewSize/K, NewSize/K);
363 }
364 MaxNewSize = NewSize;
365 }
366
367 if (FLAG_IS_CMDLINE(NewSize)) {
368 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes),
369 1U);
370 if (FLAG_IS_CMDLINE(MaxNewSize)) {
371 _max_desired_young_length =
372 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes),
373 1U);
374 _sizer_kind = SizerMaxAndNewSize;
375 _adaptive_size = _min_desired_young_length != _max_desired_young_length;
376 } else {
377 _sizer_kind = SizerNewSizeOnly;
378 }
379 } else if (FLAG_IS_CMDLINE(MaxNewSize)) {
380 _max_desired_young_length =
381 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes),
382 1U);
383 _sizer_kind = SizerMaxNewSizeOnly;
384 }
385 }
386
calculate_default_min_length(uint new_number_of_heap_regions)387 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) {
388 uint default_value = (new_number_of_heap_regions * G1NewSizePercent) / 100;
389 return MAX2(1U, default_value);
390 }
391
calculate_default_max_length(uint new_number_of_heap_regions)392 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) {
393 uint default_value = (new_number_of_heap_regions * G1MaxNewSizePercent) / 100;
394 return MAX2(1U, default_value);
395 }
396
recalculate_min_max_young_length(uint number_of_heap_regions,uint * min_young_length,uint * max_young_length)397 void G1YoungGenSizer::recalculate_min_max_young_length(uint number_of_heap_regions, uint* min_young_length, uint* max_young_length) {
398 assert(number_of_heap_regions > 0, "Heap must be initialized");
399
400 switch (_sizer_kind) {
401 case SizerDefaults:
402 *min_young_length = calculate_default_min_length(number_of_heap_regions);
403 *max_young_length = calculate_default_max_length(number_of_heap_regions);
404 break;
405 case SizerNewSizeOnly:
406 *max_young_length = calculate_default_max_length(number_of_heap_regions);
407 *max_young_length = MAX2(*min_young_length, *max_young_length);
408 break;
409 case SizerMaxNewSizeOnly:
410 *min_young_length = calculate_default_min_length(number_of_heap_regions);
411 *min_young_length = MIN2(*min_young_length, *max_young_length);
412 break;
413 case SizerMaxAndNewSize:
414 // Do nothing. Values set on the command line, don't update them at runtime.
415 break;
416 case SizerNewRatio:
417 *min_young_length = number_of_heap_regions / (NewRatio + 1);
418 *max_young_length = *min_young_length;
419 break;
420 default:
421 ShouldNotReachHere();
422 }
423
424 assert(*min_young_length <= *max_young_length, "Invalid min/max young gen size values");
425 }
426
max_young_length(uint number_of_heap_regions)427 uint G1YoungGenSizer::max_young_length(uint number_of_heap_regions) {
428 // We need to pass the desired values because recalculation may not update these
429 // values in some cases.
430 uint temp = _min_desired_young_length;
431 uint result = _max_desired_young_length;
432 recalculate_min_max_young_length(number_of_heap_regions, &temp, &result);
433 return result;
434 }
435
heap_size_changed(uint new_number_of_heap_regions)436 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) {
437 recalculate_min_max_young_length(new_number_of_heap_regions, &_min_desired_young_length,
438 &_max_desired_young_length);
439 }
440
init()441 void G1CollectorPolicy::init() {
442 // Set aside an initial future to_space.
443 _g1 = G1CollectedHeap::heap();
444
445 assert(Heap_lock->owned_by_self(), "Locking discipline.");
446
447 initialize_gc_policy_counters();
448
449 if (adaptive_young_list_length()) {
450 _young_list_fixed_length = 0;
451 } else {
452 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length();
453 }
454 _free_regions_at_end_of_collection = _g1->num_free_regions();
455 update_young_list_target_length();
456
457 // We may immediately start allocating regions and placing them on the
458 // collection set list. Initialize the per-collection set info
459 start_incremental_cset_building();
460 }
461
462 // Create the jstat counters for the policy.
initialize_gc_policy_counters()463 void G1CollectorPolicy::initialize_gc_policy_counters() {
464 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3);
465 }
466
predict_will_fit(uint young_length,double base_time_ms,uint base_free_regions,double target_pause_time_ms)467 bool G1CollectorPolicy::predict_will_fit(uint young_length,
468 double base_time_ms,
469 uint base_free_regions,
470 double target_pause_time_ms) {
471 if (young_length >= base_free_regions) {
472 // end condition 1: not enough space for the young regions
473 return false;
474 }
475
476 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1);
477 size_t bytes_to_copy =
478 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
479 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
480 double young_other_time_ms = predict_young_other_time_ms(young_length);
481 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms;
482 if (pause_time_ms > target_pause_time_ms) {
483 // end condition 2: prediction is over the target pause time
484 return false;
485 }
486
487 size_t free_bytes =
488 (base_free_regions - young_length) * HeapRegion::GrainBytes;
489 if ((2.0 * sigma()) * (double) bytes_to_copy > (double) free_bytes) {
490 // end condition 3: out-of-space (conservatively!)
491 return false;
492 }
493
494 // success!
495 return true;
496 }
497
record_new_heap_size(uint new_number_of_regions)498 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) {
499 // re-calculate the necessary reserve
500 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor;
501 // We use ceiling so that if reserve_regions_d is > 0.0 (but
502 // smaller than 1.0) we'll get 1.
503 _reserve_regions = (uint) ceil(reserve_regions_d);
504
505 _young_gen_sizer->heap_size_changed(new_number_of_regions);
506 }
507
calculate_young_list_desired_min_length(uint base_min_length)508 uint G1CollectorPolicy::calculate_young_list_desired_min_length(
509 uint base_min_length) {
510 uint desired_min_length = 0;
511 if (adaptive_young_list_length()) {
512 if (_alloc_rate_ms_seq->num() > 3) {
513 double now_sec = os::elapsedTime();
514 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
515 double alloc_rate_ms = predict_alloc_rate_ms();
516 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms);
517 } else {
518 // otherwise we don't have enough info to make the prediction
519 }
520 }
521 desired_min_length += base_min_length;
522 // make sure we don't go below any user-defined minimum bound
523 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length);
524 }
525
calculate_young_list_desired_max_length()526 uint G1CollectorPolicy::calculate_young_list_desired_max_length() {
527 // Here, we might want to also take into account any additional
528 // constraints (i.e., user-defined minimum bound). Currently, we
529 // effectively don't set this bound.
530 return _young_gen_sizer->max_desired_young_length();
531 }
532
update_young_list_target_length(size_t rs_lengths)533 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) {
534 if (rs_lengths == (size_t) -1) {
535 // if it's set to the default value (-1), we should predict it;
536 // otherwise, use the given value.
537 rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
538 }
539
540 // Calculate the absolute and desired min bounds.
541
542 // This is how many young regions we already have (currently: the survivors).
543 uint base_min_length = recorded_survivor_regions();
544 // This is the absolute minimum young length, which ensures that we
545 // can allocate one eden region in the worst-case.
546 uint absolute_min_length = base_min_length + 1;
547 uint desired_min_length =
548 calculate_young_list_desired_min_length(base_min_length);
549 if (desired_min_length < absolute_min_length) {
550 desired_min_length = absolute_min_length;
551 }
552
553 // Calculate the absolute and desired max bounds.
554
555 // We will try our best not to "eat" into the reserve.
556 uint absolute_max_length = 0;
557 if (_free_regions_at_end_of_collection > _reserve_regions) {
558 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions;
559 }
560 uint desired_max_length = calculate_young_list_desired_max_length();
561 if (desired_max_length > absolute_max_length) {
562 desired_max_length = absolute_max_length;
563 }
564
565 uint young_list_target_length = 0;
566 if (adaptive_young_list_length()) {
567 if (gcs_are_young()) {
568 young_list_target_length =
569 calculate_young_list_target_length(rs_lengths,
570 base_min_length,
571 desired_min_length,
572 desired_max_length);
573 _rs_lengths_prediction = rs_lengths;
574 } else {
575 // Don't calculate anything and let the code below bound it to
576 // the desired_min_length, i.e., do the next GC as soon as
577 // possible to maximize how many old regions we can add to it.
578 }
579 } else {
580 // The user asked for a fixed young gen so we'll fix the young gen
581 // whether the next GC is young or mixed.
582 young_list_target_length = _young_list_fixed_length;
583 }
584
585 // Make sure we don't go over the desired max length, nor under the
586 // desired min length. In case they clash, desired_min_length wins
587 // which is why that test is second.
588 if (young_list_target_length > desired_max_length) {
589 young_list_target_length = desired_max_length;
590 }
591 if (young_list_target_length < desired_min_length) {
592 young_list_target_length = desired_min_length;
593 }
594
595 assert(young_list_target_length > recorded_survivor_regions(),
596 "we should be able to allocate at least one eden region");
597 assert(young_list_target_length >= absolute_min_length, "post-condition");
598 _young_list_target_length = young_list_target_length;
599
600 update_max_gc_locker_expansion();
601 }
602
603 uint
calculate_young_list_target_length(size_t rs_lengths,uint base_min_length,uint desired_min_length,uint desired_max_length)604 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths,
605 uint base_min_length,
606 uint desired_min_length,
607 uint desired_max_length) {
608 assert(adaptive_young_list_length(), "pre-condition");
609 assert(gcs_are_young(), "only call this for young GCs");
610
611 // In case some edge-condition makes the desired max length too small...
612 if (desired_max_length <= desired_min_length) {
613 return desired_min_length;
614 }
615
616 // We'll adjust min_young_length and max_young_length not to include
617 // the already allocated young regions (i.e., so they reflect the
618 // min and max eden regions we'll allocate). The base_min_length
619 // will be reflected in the predictions by the
620 // survivor_regions_evac_time prediction.
621 assert(desired_min_length > base_min_length, "invariant");
622 uint min_young_length = desired_min_length - base_min_length;
623 assert(desired_max_length > base_min_length, "invariant");
624 uint max_young_length = desired_max_length - base_min_length;
625
626 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
627 double survivor_regions_evac_time = predict_survivor_regions_evac_time();
628 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
629 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
630 size_t scanned_cards = predict_young_card_num(adj_rs_lengths);
631 double base_time_ms =
632 predict_base_elapsed_time_ms(pending_cards, scanned_cards) +
633 survivor_regions_evac_time;
634 uint available_free_regions = _free_regions_at_end_of_collection;
635 uint base_free_regions = 0;
636 if (available_free_regions > _reserve_regions) {
637 base_free_regions = available_free_regions - _reserve_regions;
638 }
639
640 // Here, we will make sure that the shortest young length that
641 // makes sense fits within the target pause time.
642
643 if (predict_will_fit(min_young_length, base_time_ms,
644 base_free_regions, target_pause_time_ms)) {
645 // The shortest young length will fit into the target pause time;
646 // we'll now check whether the absolute maximum number of young
647 // regions will fit in the target pause time. If not, we'll do
648 // a binary search between min_young_length and max_young_length.
649 if (predict_will_fit(max_young_length, base_time_ms,
650 base_free_regions, target_pause_time_ms)) {
651 // The maximum young length will fit into the target pause time.
652 // We are done so set min young length to the maximum length (as
653 // the result is assumed to be returned in min_young_length).
654 min_young_length = max_young_length;
655 } else {
656 // The maximum possible number of young regions will not fit within
657 // the target pause time so we'll search for the optimal
658 // length. The loop invariants are:
659 //
660 // min_young_length < max_young_length
661 // min_young_length is known to fit into the target pause time
662 // max_young_length is known not to fit into the target pause time
663 //
664 // Going into the loop we know the above hold as we've just
665 // checked them. Every time around the loop we check whether
666 // the middle value between min_young_length and
667 // max_young_length fits into the target pause time. If it
668 // does, it becomes the new min. If it doesn't, it becomes
669 // the new max. This way we maintain the loop invariants.
670
671 assert(min_young_length < max_young_length, "invariant");
672 uint diff = (max_young_length - min_young_length) / 2;
673 while (diff > 0) {
674 uint young_length = min_young_length + diff;
675 if (predict_will_fit(young_length, base_time_ms,
676 base_free_regions, target_pause_time_ms)) {
677 min_young_length = young_length;
678 } else {
679 max_young_length = young_length;
680 }
681 assert(min_young_length < max_young_length, "invariant");
682 diff = (max_young_length - min_young_length) / 2;
683 }
684 // The results is min_young_length which, according to the
685 // loop invariants, should fit within the target pause time.
686
687 // These are the post-conditions of the binary search above:
688 assert(min_young_length < max_young_length,
689 "otherwise we should have discovered that max_young_length "
690 "fits into the pause target and not done the binary search");
691 assert(predict_will_fit(min_young_length, base_time_ms,
692 base_free_regions, target_pause_time_ms),
693 "min_young_length, the result of the binary search, should "
694 "fit into the pause target");
695 assert(!predict_will_fit(min_young_length + 1, base_time_ms,
696 base_free_regions, target_pause_time_ms),
697 "min_young_length, the result of the binary search, should be "
698 "optimal, so no larger length should fit into the pause target");
699 }
700 } else {
701 // Even the minimum length doesn't fit into the pause time
702 // target, return it as the result nevertheless.
703 }
704 return base_min_length + min_young_length;
705 }
706
predict_survivor_regions_evac_time()707 double G1CollectorPolicy::predict_survivor_regions_evac_time() {
708 double survivor_regions_evac_time = 0.0;
709 for (HeapRegion * r = _recorded_survivor_head;
710 r != NULL && r != _recorded_survivor_tail->get_next_young_region();
711 r = r->get_next_young_region()) {
712 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, gcs_are_young());
713 }
714 return survivor_regions_evac_time;
715 }
716
revise_young_list_target_length_if_necessary()717 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() {
718 guarantee( adaptive_young_list_length(), "should not call this otherwise" );
719
720 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths();
721 if (rs_lengths > _rs_lengths_prediction) {
722 // add 10% to avoid having to recalculate often
723 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000;
724 update_young_list_target_length(rs_lengths_prediction);
725 }
726 }
727
728
729
mem_allocate_work(size_t size,bool is_tlab,bool * gc_overhead_limit_was_exceeded)730 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size,
731 bool is_tlab,
732 bool* gc_overhead_limit_was_exceeded) {
733 guarantee(false, "Not using this policy feature yet.");
734 return NULL;
735 }
736
737 // This method controls how a collector handles one or more
738 // of its generations being fully allocated.
satisfy_failed_allocation(size_t size,bool is_tlab)739 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size,
740 bool is_tlab) {
741 guarantee(false, "Not using this policy feature yet.");
742 return NULL;
743 }
744
745
746 #ifndef PRODUCT
verify_young_ages()747 bool G1CollectorPolicy::verify_young_ages() {
748 HeapRegion* head = _g1->young_list()->first_region();
749 return
750 verify_young_ages(head, _short_lived_surv_rate_group);
751 // also call verify_young_ages on any additional surv rate groups
752 }
753
754 bool
verify_young_ages(HeapRegion * head,SurvRateGroup * surv_rate_group)755 G1CollectorPolicy::verify_young_ages(HeapRegion* head,
756 SurvRateGroup *surv_rate_group) {
757 guarantee( surv_rate_group != NULL, "pre-condition" );
758
759 const char* name = surv_rate_group->name();
760 bool ret = true;
761 int prev_age = -1;
762
763 for (HeapRegion* curr = head;
764 curr != NULL;
765 curr = curr->get_next_young_region()) {
766 SurvRateGroup* group = curr->surv_rate_group();
767 if (group == NULL && !curr->is_survivor()) {
768 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name);
769 ret = false;
770 }
771
772 if (surv_rate_group == group) {
773 int age = curr->age_in_surv_rate_group();
774
775 if (age < 0) {
776 gclog_or_tty->print_cr("## %s: encountered negative age", name);
777 ret = false;
778 }
779
780 if (age <= prev_age) {
781 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing "
782 "(%d, %d)", name, age, prev_age);
783 ret = false;
784 }
785 prev_age = age;
786 }
787 }
788
789 return ret;
790 }
791 #endif // PRODUCT
792
record_full_collection_start()793 void G1CollectorPolicy::record_full_collection_start() {
794 _full_collection_start_sec = os::elapsedTime();
795 record_heap_size_info_at_start(true /* full */);
796 // Release the future to-space so that it is available for compaction into.
797 _g1->set_full_collection();
798 }
799
record_full_collection_end()800 void G1CollectorPolicy::record_full_collection_end() {
801 // Consider this like a collection pause for the purposes of allocation
802 // since last pause.
803 double end_sec = os::elapsedTime();
804 double full_gc_time_sec = end_sec - _full_collection_start_sec;
805 double full_gc_time_ms = full_gc_time_sec * 1000.0;
806
807 _trace_gen1_time_data.record_full_collection(full_gc_time_ms);
808
809 update_recent_gc_times(end_sec, full_gc_time_ms);
810
811 _g1->clear_full_collection();
812
813 // "Nuke" the heuristics that control the young/mixed GC
814 // transitions and make sure we start with young GCs after the Full GC.
815 set_gcs_are_young(true);
816 _last_young_gc = false;
817 clear_initiate_conc_mark_if_possible();
818 clear_during_initial_mark_pause();
819 _in_marking_window = false;
820 _in_marking_window_im = false;
821
822 _short_lived_surv_rate_group->start_adding_regions();
823 // also call this on any additional surv rate groups
824
825 record_survivor_regions(0, NULL, NULL);
826
827 _free_regions_at_end_of_collection = _g1->num_free_regions();
828 // Reset survivors SurvRateGroup.
829 _survivor_surv_rate_group->reset();
830 update_young_list_target_length();
831 _collectionSetChooser->clear();
832 }
833
record_stop_world_start()834 void G1CollectorPolicy::record_stop_world_start() {
835 _stop_world_start = os::elapsedTime();
836 }
837
record_collection_pause_start(double start_time_sec,GCTracer & tracer)838 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec, GCTracer &tracer) {
839 // We only need to do this here as the policy will only be applied
840 // to the GC we're about to start. so, no point is calculating this
841 // every time we calculate / recalculate the target young length.
842 update_survivors_policy(tracer);
843
844 assert(_g1->used() == _g1->recalculate_used(),
845 err_msg("sanity, used: " SIZE_FORMAT " recalculate_used: " SIZE_FORMAT,
846 _g1->used(), _g1->recalculate_used()));
847
848 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0;
849 _trace_gen0_time_data.record_start_collection(s_w_t_ms);
850 _stop_world_start = 0.0;
851
852 record_heap_size_info_at_start(false /* full */);
853
854 phase_times()->record_cur_collection_start_sec(start_time_sec);
855 _pending_cards = _g1->pending_card_num();
856
857 _collection_set_bytes_used_before = 0;
858 _bytes_copied_during_gc = 0;
859
860 _last_gc_was_young = false;
861
862 // do that for any other surv rate groups
863 _short_lived_surv_rate_group->stop_adding_regions();
864 _survivors_age_table.clear();
865
866 assert( verify_young_ages(), "region age verification" );
867 }
868
record_concurrent_mark_init_end(double mark_init_elapsed_time_ms)869 void G1CollectorPolicy::record_concurrent_mark_init_end(double
870 mark_init_elapsed_time_ms) {
871 _during_marking = true;
872 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now");
873 clear_during_initial_mark_pause();
874 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms;
875 }
876
record_concurrent_mark_remark_start()877 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
878 _mark_remark_start_sec = os::elapsedTime();
879 _during_marking = false;
880 }
881
record_concurrent_mark_remark_end()882 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
883 double end_time_sec = os::elapsedTime();
884 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
885 _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
886 _cur_mark_stop_world_time_ms += elapsed_time_ms;
887 _prev_collection_pause_end_ms += elapsed_time_ms;
888
889 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true);
890 }
891
record_concurrent_mark_cleanup_start()892 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
893 _mark_cleanup_start_sec = os::elapsedTime();
894 }
895
record_concurrent_mark_cleanup_completed()896 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
897 _last_young_gc = true;
898 _in_marking_window = false;
899 }
900
record_concurrent_pause()901 void G1CollectorPolicy::record_concurrent_pause() {
902 if (_stop_world_start > 0.0) {
903 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0;
904 _trace_gen0_time_data.record_yield_time(yield_ms);
905 }
906 }
907
need_to_start_conc_mark(const char * source,size_t alloc_word_size)908 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) {
909 if (_g1->concurrent_mark()->cmThread()->during_cycle()) {
910 return false;
911 }
912
913 size_t marking_initiating_used_threshold =
914 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent;
915 size_t cur_used_bytes = _g1->non_young_capacity_bytes();
916 size_t alloc_byte_size = alloc_word_size * HeapWordSize;
917
918 if ((cur_used_bytes + alloc_byte_size) > marking_initiating_used_threshold) {
919 if (gcs_are_young() && !_last_young_gc) {
920 ergo_verbose5(ErgoConcCycles,
921 "request concurrent cycle initiation",
922 ergo_format_reason("occupancy higher than threshold")
923 ergo_format_byte("occupancy")
924 ergo_format_byte("allocation request")
925 ergo_format_byte_perc("threshold")
926 ergo_format_str("source"),
927 cur_used_bytes,
928 alloc_byte_size,
929 marking_initiating_used_threshold,
930 (double) InitiatingHeapOccupancyPercent,
931 source);
932 return true;
933 } else {
934 ergo_verbose5(ErgoConcCycles,
935 "do not request concurrent cycle initiation",
936 ergo_format_reason("still doing mixed collections")
937 ergo_format_byte("occupancy")
938 ergo_format_byte("allocation request")
939 ergo_format_byte_perc("threshold")
940 ergo_format_str("source"),
941 cur_used_bytes,
942 alloc_byte_size,
943 marking_initiating_used_threshold,
944 (double) InitiatingHeapOccupancyPercent,
945 source);
946 }
947 }
948
949 return false;
950 }
951
952 // Anything below that is considered to be zero
953 #define MIN_TIMER_GRANULARITY 0.0000001
954
record_collection_pause_end(double pause_time_ms,EvacuationInfo & evacuation_info)955 void G1CollectorPolicy::record_collection_pause_end(double pause_time_ms, EvacuationInfo& evacuation_info) {
956 double end_time_sec = os::elapsedTime();
957 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(),
958 "otherwise, the subtraction below does not make sense");
959 size_t rs_size =
960 _cur_collection_pause_used_regions_at_start - cset_region_length();
961 size_t cur_used_bytes = _g1->used();
962 assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
963 bool last_pause_included_initial_mark = false;
964 bool update_stats = !_g1->evacuation_failed();
965
966 #ifndef PRODUCT
967 if (G1YoungSurvRateVerbose) {
968 gclog_or_tty->cr();
969 _short_lived_surv_rate_group->print();
970 // do that for any other surv rate groups too
971 }
972 #endif // PRODUCT
973
974 last_pause_included_initial_mark = during_initial_mark_pause();
975 if (last_pause_included_initial_mark) {
976 record_concurrent_mark_init_end(0.0);
977 } else if (need_to_start_conc_mark("end of GC")) {
978 // Note: this might have already been set, if during the last
979 // pause we decided to start a cycle but at the beginning of
980 // this pause we decided to postpone it. That's OK.
981 set_initiate_conc_mark_if_possible();
982 }
983
984 _mmu_tracker->add_pause(end_time_sec - pause_time_ms/1000.0,
985 end_time_sec, false);
986
987 evacuation_info.set_collectionset_used_before(_collection_set_bytes_used_before);
988 evacuation_info.set_bytes_copied(_bytes_copied_during_gc);
989
990 if (update_stats) {
991 _trace_gen0_time_data.record_end_collection(pause_time_ms, phase_times());
992 // this is where we update the allocation rate of the application
993 double app_time_ms =
994 (phase_times()->cur_collection_start_sec() * 1000.0 - _prev_collection_pause_end_ms);
995 if (app_time_ms < MIN_TIMER_GRANULARITY) {
996 // This usually happens due to the timer not having the required
997 // granularity. Some Linuxes are the usual culprits.
998 // We'll just set it to something (arbitrarily) small.
999 app_time_ms = 1.0;
1000 }
1001 // We maintain the invariant that all objects allocated by mutator
1002 // threads will be allocated out of eden regions. So, we can use
1003 // the eden region number allocated since the previous GC to
1004 // calculate the application's allocate rate. The only exception
1005 // to that is humongous objects that are allocated separately. But
1006 // given that humongous object allocations do not really affect
1007 // either the pause's duration nor when the next pause will take
1008 // place we can safely ignore them here.
1009 uint regions_allocated = eden_cset_region_length();
1010 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
1011 _alloc_rate_ms_seq->add(alloc_rate_ms);
1012
1013 double interval_ms =
1014 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
1015 update_recent_gc_times(end_time_sec, pause_time_ms);
1016 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
1017 if (recent_avg_pause_time_ratio() < 0.0 ||
1018 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) {
1019 #ifndef PRODUCT
1020 // Dump info to allow post-facto debugging
1021 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds");
1022 gclog_or_tty->print_cr("-------------------------------------------");
1023 gclog_or_tty->print_cr("Recent GC Times (ms):");
1024 _recent_gc_times_ms->dump();
1025 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec);
1026 _recent_prev_end_times_for_all_gcs_sec->dump();
1027 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f",
1028 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio());
1029 // In debug mode, terminate the JVM if the user wants to debug at this point.
1030 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above");
1031 #endif // !PRODUCT
1032 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
1033 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
1034 if (_recent_avg_pause_time_ratio < 0.0) {
1035 _recent_avg_pause_time_ratio = 0.0;
1036 } else {
1037 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
1038 _recent_avg_pause_time_ratio = 1.0;
1039 }
1040 }
1041 }
1042
1043 bool new_in_marking_window = _in_marking_window;
1044 bool new_in_marking_window_im = false;
1045 if (last_pause_included_initial_mark) {
1046 new_in_marking_window = true;
1047 new_in_marking_window_im = true;
1048 }
1049
1050 if (_last_young_gc) {
1051 // This is supposed to to be the "last young GC" before we start
1052 // doing mixed GCs. Here we decide whether to start mixed GCs or not.
1053
1054 if (!last_pause_included_initial_mark) {
1055 if (next_gc_should_be_mixed("start mixed GCs",
1056 "do not start mixed GCs")) {
1057 set_gcs_are_young(false);
1058 }
1059 } else {
1060 ergo_verbose0(ErgoMixedGCs,
1061 "do not start mixed GCs",
1062 ergo_format_reason("concurrent cycle is about to start"));
1063 }
1064 _last_young_gc = false;
1065 }
1066
1067 if (!_last_gc_was_young) {
1068 // This is a mixed GC. Here we decide whether to continue doing
1069 // mixed GCs or not.
1070
1071 if (!next_gc_should_be_mixed("continue mixed GCs",
1072 "do not continue mixed GCs")) {
1073 set_gcs_are_young(true);
1074 }
1075 }
1076
1077 _short_lived_surv_rate_group->start_adding_regions();
1078 // do that for any other surv rate groupsx
1079
1080 if (update_stats) {
1081 double cost_per_card_ms = 0.0;
1082 if (_pending_cards > 0) {
1083 cost_per_card_ms = phase_times()->average_time_ms(G1GCPhaseTimes::UpdateRS) / (double) _pending_cards;
1084 _cost_per_card_ms_seq->add(cost_per_card_ms);
1085 }
1086
1087 size_t cards_scanned = _g1->cards_scanned();
1088
1089 double cost_per_entry_ms = 0.0;
1090 if (cards_scanned > 10) {
1091 cost_per_entry_ms = phase_times()->average_time_ms(G1GCPhaseTimes::ScanRS) / (double) cards_scanned;
1092 if (_last_gc_was_young) {
1093 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
1094 } else {
1095 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms);
1096 }
1097 }
1098
1099 if (_max_rs_lengths > 0) {
1100 double cards_per_entry_ratio =
1101 (double) cards_scanned / (double) _max_rs_lengths;
1102 if (_last_gc_was_young) {
1103 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1104 } else {
1105 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1106 }
1107 }
1108
1109 // This is defensive. For a while _max_rs_lengths could get
1110 // smaller than _recorded_rs_lengths which was causing
1111 // rs_length_diff to get very large and mess up the RSet length
1112 // predictions. The reason was unsafe concurrent updates to the
1113 // _inc_cset_recorded_rs_lengths field which the code below guards
1114 // against (see CR 7118202). This bug has now been fixed (see CR
1115 // 7119027). However, I'm still worried that
1116 // _inc_cset_recorded_rs_lengths might still end up somewhat
1117 // inaccurate. The concurrent refinement thread calculates an
1118 // RSet's length concurrently with other CR threads updating it
1119 // which might cause it to calculate the length incorrectly (if,
1120 // say, it's in mid-coarsening). So I'll leave in the defensive
1121 // conditional below just in case.
1122 size_t rs_length_diff = 0;
1123 if (_max_rs_lengths > _recorded_rs_lengths) {
1124 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths;
1125 }
1126 _rs_length_diff_seq->add((double) rs_length_diff);
1127
1128 size_t freed_bytes = _heap_used_bytes_before_gc - cur_used_bytes;
1129
1130 if (_collection_set_bytes_used_before > freed_bytes) {
1131 size_t copied_bytes = _collection_set_bytes_used_before - freed_bytes;
1132 double average_copy_time = phase_times()->average_time_ms(G1GCPhaseTimes::ObjCopy);
1133 double cost_per_byte_ms = average_copy_time / (double) copied_bytes;
1134 if (_in_marking_window) {
1135 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
1136 } else {
1137 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
1138 }
1139 }
1140
1141 double all_other_time_ms = pause_time_ms -
1142 (phase_times()->average_time_ms(G1GCPhaseTimes::UpdateRS) + phase_times()->average_time_ms(G1GCPhaseTimes::ScanRS) +
1143 phase_times()->average_time_ms(G1GCPhaseTimes::ObjCopy) + phase_times()->average_time_ms(G1GCPhaseTimes::Termination));
1144
1145 double young_other_time_ms = 0.0;
1146 if (young_cset_region_length() > 0) {
1147 young_other_time_ms =
1148 phase_times()->young_cset_choice_time_ms() +
1149 phase_times()->young_free_cset_time_ms();
1150 _young_other_cost_per_region_ms_seq->add(young_other_time_ms /
1151 (double) young_cset_region_length());
1152 }
1153 double non_young_other_time_ms = 0.0;
1154 if (old_cset_region_length() > 0) {
1155 non_young_other_time_ms =
1156 phase_times()->non_young_cset_choice_time_ms() +
1157 phase_times()->non_young_free_cset_time_ms();
1158
1159 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms /
1160 (double) old_cset_region_length());
1161 }
1162
1163 double constant_other_time_ms = all_other_time_ms -
1164 (young_other_time_ms + non_young_other_time_ms);
1165 _constant_other_time_ms_seq->add(constant_other_time_ms);
1166
1167 double survival_ratio = 0.0;
1168 if (_collection_set_bytes_used_before > 0) {
1169 survival_ratio = (double) _bytes_copied_during_gc /
1170 (double) _collection_set_bytes_used_before;
1171 }
1172
1173 _pending_cards_seq->add((double) _pending_cards);
1174 _rs_lengths_seq->add((double) _max_rs_lengths);
1175 }
1176
1177 _in_marking_window = new_in_marking_window;
1178 _in_marking_window_im = new_in_marking_window_im;
1179 _free_regions_at_end_of_collection = _g1->num_free_regions();
1180 update_young_list_target_length();
1181
1182 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
1183 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
1184 adjust_concurrent_refinement(phase_times()->average_time_ms(G1GCPhaseTimes::UpdateRS),
1185 phase_times()->sum_thread_work_items(G1GCPhaseTimes::UpdateRS), update_rs_time_goal_ms);
1186
1187 _collectionSetChooser->verify();
1188 }
1189
1190 #define EXT_SIZE_FORMAT "%.1f%s"
1191 #define EXT_SIZE_PARAMS(bytes) \
1192 byte_size_in_proper_unit((double)(bytes)), \
1193 proper_unit_for_byte_size((bytes))
1194
record_heap_size_info_at_start(bool full)1195 void G1CollectorPolicy::record_heap_size_info_at_start(bool full) {
1196 YoungList* young_list = _g1->young_list();
1197 _eden_used_bytes_before_gc = young_list->eden_used_bytes();
1198 _survivor_used_bytes_before_gc = young_list->survivor_used_bytes();
1199 _heap_capacity_bytes_before_gc = _g1->capacity();
1200 _heap_used_bytes_before_gc = _g1->used();
1201 _cur_collection_pause_used_regions_at_start = _g1->num_used_regions();
1202
1203 _eden_capacity_bytes_before_gc =
1204 (_young_list_target_length * HeapRegion::GrainBytes) - _survivor_used_bytes_before_gc;
1205
1206 if (full) {
1207 _metaspace_used_bytes_before_gc = MetaspaceAux::used_bytes();
1208 }
1209 }
1210
print_heap_transition()1211 void G1CollectorPolicy::print_heap_transition() {
1212 _g1->print_size_transition(gclog_or_tty,
1213 _heap_used_bytes_before_gc,
1214 _g1->used(),
1215 _g1->capacity());
1216 }
1217
print_detailed_heap_transition(bool full)1218 void G1CollectorPolicy::print_detailed_heap_transition(bool full) {
1219 YoungList* young_list = _g1->young_list();
1220
1221 size_t eden_used_bytes_after_gc = young_list->eden_used_bytes();
1222 size_t survivor_used_bytes_after_gc = young_list->survivor_used_bytes();
1223 size_t heap_used_bytes_after_gc = _g1->used();
1224
1225 size_t heap_capacity_bytes_after_gc = _g1->capacity();
1226 size_t eden_capacity_bytes_after_gc =
1227 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_used_bytes_after_gc;
1228
1229 gclog_or_tty->print(
1230 " [Eden: " EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")->" EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ") "
1231 "Survivors: " EXT_SIZE_FORMAT "->" EXT_SIZE_FORMAT " "
1232 "Heap: " EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")->"
1233 EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")]",
1234 EXT_SIZE_PARAMS(_eden_used_bytes_before_gc),
1235 EXT_SIZE_PARAMS(_eden_capacity_bytes_before_gc),
1236 EXT_SIZE_PARAMS(eden_used_bytes_after_gc),
1237 EXT_SIZE_PARAMS(eden_capacity_bytes_after_gc),
1238 EXT_SIZE_PARAMS(_survivor_used_bytes_before_gc),
1239 EXT_SIZE_PARAMS(survivor_used_bytes_after_gc),
1240 EXT_SIZE_PARAMS(_heap_used_bytes_before_gc),
1241 EXT_SIZE_PARAMS(_heap_capacity_bytes_before_gc),
1242 EXT_SIZE_PARAMS(heap_used_bytes_after_gc),
1243 EXT_SIZE_PARAMS(heap_capacity_bytes_after_gc));
1244
1245 if (full) {
1246 MetaspaceAux::print_metaspace_change(_metaspace_used_bytes_before_gc);
1247 }
1248
1249 gclog_or_tty->cr();
1250 }
1251
adjust_concurrent_refinement(double update_rs_time,double update_rs_processed_buffers,double goal_ms)1252 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1253 double update_rs_processed_buffers,
1254 double goal_ms) {
1255 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1256 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1257
1258 if (G1UseAdaptiveConcRefinement) {
1259 const int k_gy = 3, k_gr = 6;
1260 const double inc_k = 1.1, dec_k = 0.9;
1261
1262 int g = cg1r->green_zone();
1263 if (update_rs_time > goal_ms) {
1264 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1265 } else {
1266 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1267 g = (int)MAX2(g * inc_k, g + 1.0);
1268 }
1269 }
1270 // Change the refinement threads params
1271 cg1r->set_green_zone(g);
1272 cg1r->set_yellow_zone(g * k_gy);
1273 cg1r->set_red_zone(g * k_gr);
1274 cg1r->reinitialize_threads();
1275
1276 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1);
1277 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1278 cg1r->yellow_zone());
1279 // Change the barrier params
1280 dcqs.set_process_completed_threshold(processing_threshold);
1281 dcqs.set_max_completed_queue(cg1r->red_zone());
1282 }
1283
1284 int curr_queue_size = dcqs.completed_buffers_num();
1285 if (curr_queue_size >= cg1r->yellow_zone()) {
1286 dcqs.set_completed_queue_padding(curr_queue_size);
1287 } else {
1288 dcqs.set_completed_queue_padding(0);
1289 }
1290 dcqs.notify_if_necessary();
1291 }
1292
1293 double
predict_base_elapsed_time_ms(size_t pending_cards,size_t scanned_cards)1294 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
1295 size_t scanned_cards) {
1296 return
1297 predict_rs_update_time_ms(pending_cards) +
1298 predict_rs_scan_time_ms(scanned_cards) +
1299 predict_constant_other_time_ms();
1300 }
1301
1302 double
predict_base_elapsed_time_ms(size_t pending_cards)1303 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) {
1304 size_t rs_length = predict_rs_length_diff();
1305 size_t card_num;
1306 if (gcs_are_young()) {
1307 card_num = predict_young_card_num(rs_length);
1308 } else {
1309 card_num = predict_non_young_card_num(rs_length);
1310 }
1311 return predict_base_elapsed_time_ms(pending_cards, card_num);
1312 }
1313
predict_bytes_to_copy(HeapRegion * hr)1314 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) {
1315 size_t bytes_to_copy;
1316 if (hr->is_marked())
1317 bytes_to_copy = hr->max_live_bytes();
1318 else {
1319 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant");
1320 int age = hr->age_in_surv_rate_group();
1321 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
1322 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate);
1323 }
1324 return bytes_to_copy;
1325 }
1326
1327 double
predict_region_elapsed_time_ms(HeapRegion * hr,bool for_young_gc)1328 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
1329 bool for_young_gc) {
1330 size_t rs_length = hr->rem_set()->occupied();
1331 size_t card_num;
1332
1333 // Predicting the number of cards is based on which type of GC
1334 // we're predicting for.
1335 if (for_young_gc) {
1336 card_num = predict_young_card_num(rs_length);
1337 } else {
1338 card_num = predict_non_young_card_num(rs_length);
1339 }
1340 size_t bytes_to_copy = predict_bytes_to_copy(hr);
1341
1342 double region_elapsed_time_ms =
1343 predict_rs_scan_time_ms(card_num) +
1344 predict_object_copy_time_ms(bytes_to_copy);
1345
1346 // The prediction of the "other" time for this region is based
1347 // upon the region type and NOT the GC type.
1348 if (hr->is_young()) {
1349 region_elapsed_time_ms += predict_young_other_time_ms(1);
1350 } else {
1351 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
1352 }
1353 return region_elapsed_time_ms;
1354 }
1355
1356 void
init_cset_region_lengths(uint eden_cset_region_length,uint survivor_cset_region_length)1357 G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length,
1358 uint survivor_cset_region_length) {
1359 _eden_cset_region_length = eden_cset_region_length;
1360 _survivor_cset_region_length = survivor_cset_region_length;
1361 _old_cset_region_length = 0;
1362 }
1363
set_recorded_rs_lengths(size_t rs_lengths)1364 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) {
1365 _recorded_rs_lengths = rs_lengths;
1366 }
1367
update_recent_gc_times(double end_time_sec,double elapsed_ms)1368 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
1369 double elapsed_ms) {
1370 _recent_gc_times_ms->add(elapsed_ms);
1371 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
1372 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
1373 }
1374
expansion_amount()1375 size_t G1CollectorPolicy::expansion_amount() {
1376 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0;
1377 double threshold = _gc_overhead_perc;
1378 if (recent_gc_overhead > threshold) {
1379 // We will double the existing space, or take
1380 // G1ExpandByPercentOfAvailable % of the available expansion
1381 // space, whichever is smaller, bounded below by a minimum
1382 // expansion (unless that's all that's left.)
1383 const size_t min_expand_bytes = 1*M;
1384 size_t reserved_bytes = _g1->max_capacity();
1385 size_t committed_bytes = _g1->capacity();
1386 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
1387 size_t expand_bytes;
1388 size_t expand_bytes_via_pct =
1389 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
1390 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
1391 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
1392 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
1393
1394 ergo_verbose5(ErgoHeapSizing,
1395 "attempt heap expansion",
1396 ergo_format_reason("recent GC overhead higher than "
1397 "threshold after GC")
1398 ergo_format_perc("recent GC overhead")
1399 ergo_format_perc("threshold")
1400 ergo_format_byte("uncommitted")
1401 ergo_format_byte_perc("calculated expansion amount"),
1402 recent_gc_overhead, threshold,
1403 uncommitted_bytes,
1404 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable);
1405
1406 return expand_bytes;
1407 } else {
1408 return 0;
1409 }
1410 }
1411
print_tracing_info() const1412 void G1CollectorPolicy::print_tracing_info() const {
1413 _trace_gen0_time_data.print();
1414 _trace_gen1_time_data.print();
1415 }
1416
print_yg_surv_rate_info() const1417 void G1CollectorPolicy::print_yg_surv_rate_info() const {
1418 #ifndef PRODUCT
1419 _short_lived_surv_rate_group->print_surv_rate_summary();
1420 // add this call for any other surv rate groups
1421 #endif // PRODUCT
1422 }
1423
is_young_list_full()1424 bool G1CollectorPolicy::is_young_list_full() {
1425 uint young_list_length = _g1->young_list()->length();
1426 uint young_list_target_length = _young_list_target_length;
1427 return young_list_length >= young_list_target_length;
1428 }
1429
can_expand_young_list()1430 bool G1CollectorPolicy::can_expand_young_list() {
1431 uint young_list_length = _g1->young_list()->length();
1432 uint young_list_max_length = _young_list_max_length;
1433 return young_list_length < young_list_max_length;
1434 }
1435
update_max_gc_locker_expansion()1436 void G1CollectorPolicy::update_max_gc_locker_expansion() {
1437 uint expansion_region_num = 0;
1438 if (GCLockerEdenExpansionPercent > 0) {
1439 double perc = (double) GCLockerEdenExpansionPercent / 100.0;
1440 double expansion_region_num_d = perc * (double) _young_list_target_length;
1441 // We use ceiling so that if expansion_region_num_d is > 0.0 (but
1442 // less than 1.0) we'll get 1.
1443 expansion_region_num = (uint) ceil(expansion_region_num_d);
1444 } else {
1445 assert(expansion_region_num == 0, "sanity");
1446 }
1447 _young_list_max_length = _young_list_target_length + expansion_region_num;
1448 assert(_young_list_target_length <= _young_list_max_length, "post-condition");
1449 }
1450
1451 // Calculates survivor space parameters.
update_survivors_policy(GCTracer & tracer)1452 void G1CollectorPolicy::update_survivors_policy(GCTracer &tracer) {
1453 double max_survivor_regions_d =
1454 (double) _young_list_target_length / (double) SurvivorRatio;
1455 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but
1456 // smaller than 1.0) we'll get 1.
1457 _max_survivor_regions = (uint) ceil(max_survivor_regions_d);
1458
1459 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
1460 HeapRegion::GrainWords * _max_survivor_regions, tracer);
1461 }
1462
force_initial_mark_if_outside_cycle(GCCause::Cause gc_cause)1463 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(
1464 GCCause::Cause gc_cause) {
1465 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
1466 if (!during_cycle) {
1467 ergo_verbose1(ErgoConcCycles,
1468 "request concurrent cycle initiation",
1469 ergo_format_reason("requested by GC cause")
1470 ergo_format_str("GC cause"),
1471 GCCause::to_string(gc_cause));
1472 set_initiate_conc_mark_if_possible();
1473 return true;
1474 } else {
1475 ergo_verbose1(ErgoConcCycles,
1476 "do not request concurrent cycle initiation",
1477 ergo_format_reason("concurrent cycle already in progress")
1478 ergo_format_str("GC cause"),
1479 GCCause::to_string(gc_cause));
1480 return false;
1481 }
1482 }
1483
1484 void
decide_on_conc_mark_initiation()1485 G1CollectorPolicy::decide_on_conc_mark_initiation() {
1486 // We are about to decide on whether this pause will be an
1487 // initial-mark pause.
1488
1489 // First, during_initial_mark_pause() should not be already set. We
1490 // will set it here if we have to. However, it should be cleared by
1491 // the end of the pause (it's only set for the duration of an
1492 // initial-mark pause).
1493 assert(!during_initial_mark_pause(), "pre-condition");
1494
1495 if (initiate_conc_mark_if_possible()) {
1496 // We had noticed on a previous pause that the heap occupancy has
1497 // gone over the initiating threshold and we should start a
1498 // concurrent marking cycle. So we might initiate one.
1499
1500 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
1501 if (!during_cycle) {
1502 // The concurrent marking thread is not "during a cycle", i.e.,
1503 // it has completed the last one. So we can go ahead and
1504 // initiate a new cycle.
1505
1506 set_during_initial_mark_pause();
1507 // We do not allow mixed GCs during marking.
1508 if (!gcs_are_young()) {
1509 set_gcs_are_young(true);
1510 ergo_verbose0(ErgoMixedGCs,
1511 "end mixed GCs",
1512 ergo_format_reason("concurrent cycle is about to start"));
1513 }
1514
1515 // And we can now clear initiate_conc_mark_if_possible() as
1516 // we've already acted on it.
1517 clear_initiate_conc_mark_if_possible();
1518
1519 ergo_verbose0(ErgoConcCycles,
1520 "initiate concurrent cycle",
1521 ergo_format_reason("concurrent cycle initiation requested"));
1522 } else {
1523 // The concurrent marking thread is still finishing up the
1524 // previous cycle. If we start one right now the two cycles
1525 // overlap. In particular, the concurrent marking thread might
1526 // be in the process of clearing the next marking bitmap (which
1527 // we will use for the next cycle if we start one). Starting a
1528 // cycle now will be bad given that parts of the marking
1529 // information might get cleared by the marking thread. And we
1530 // cannot wait for the marking thread to finish the cycle as it
1531 // periodically yields while clearing the next marking bitmap
1532 // and, if it's in a yield point, it's waiting for us to
1533 // finish. So, at this point we will not start a cycle and we'll
1534 // let the concurrent marking thread complete the last one.
1535 ergo_verbose0(ErgoConcCycles,
1536 "do not initiate concurrent cycle",
1537 ergo_format_reason("concurrent cycle already in progress"));
1538 }
1539 }
1540 }
1541
1542 class KnownGarbageClosure: public HeapRegionClosure {
1543 G1CollectedHeap* _g1h;
1544 CollectionSetChooser* _hrSorted;
1545
1546 public:
KnownGarbageClosure(CollectionSetChooser * hrSorted)1547 KnownGarbageClosure(CollectionSetChooser* hrSorted) :
1548 _g1h(G1CollectedHeap::heap()), _hrSorted(hrSorted) { }
1549
doHeapRegion(HeapRegion * r)1550 bool doHeapRegion(HeapRegion* r) {
1551 // We only include humongous regions in collection
1552 // sets when concurrent mark shows that their contained object is
1553 // unreachable.
1554
1555 // Do we have any marking information for this region?
1556 if (r->is_marked()) {
1557 // We will skip any region that's currently used as an old GC
1558 // alloc region (we should not consider those for collection
1559 // before we fill them up).
1560 if (_hrSorted->should_add(r) && !_g1h->is_old_gc_alloc_region(r)) {
1561 _hrSorted->add_region(r);
1562 }
1563 }
1564 return false;
1565 }
1566 };
1567
1568 class ParKnownGarbageHRClosure: public HeapRegionClosure {
1569 G1CollectedHeap* _g1h;
1570 CSetChooserParUpdater _cset_updater;
1571
1572 public:
ParKnownGarbageHRClosure(CollectionSetChooser * hrSorted,uint chunk_size)1573 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
1574 uint chunk_size) :
1575 _g1h(G1CollectedHeap::heap()),
1576 _cset_updater(hrSorted, true /* parallel */, chunk_size) { }
1577
doHeapRegion(HeapRegion * r)1578 bool doHeapRegion(HeapRegion* r) {
1579 // Do we have any marking information for this region?
1580 if (r->is_marked()) {
1581 // We will skip any region that's currently used as an old GC
1582 // alloc region (we should not consider those for collection
1583 // before we fill them up).
1584 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) {
1585 _cset_updater.add_region(r);
1586 }
1587 }
1588 return false;
1589 }
1590 };
1591
1592 class ParKnownGarbageTask: public AbstractGangTask {
1593 CollectionSetChooser* _hrSorted;
1594 uint _chunk_size;
1595 G1CollectedHeap* _g1;
1596 public:
ParKnownGarbageTask(CollectionSetChooser * hrSorted,uint chunk_size)1597 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size) :
1598 AbstractGangTask("ParKnownGarbageTask"),
1599 _hrSorted(hrSorted), _chunk_size(chunk_size),
1600 _g1(G1CollectedHeap::heap()) { }
1601
work(uint worker_id)1602 void work(uint worker_id) {
1603 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size);
1604
1605 // Back to zero for the claim value.
1606 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, worker_id,
1607 _g1->workers()->active_workers(),
1608 HeapRegion::InitialClaimValue);
1609 }
1610 };
1611
1612 void
record_concurrent_mark_cleanup_end(int no_of_gc_threads)1613 G1CollectorPolicy::record_concurrent_mark_cleanup_end(int no_of_gc_threads) {
1614 _collectionSetChooser->clear();
1615
1616 uint region_num = _g1->num_regions();
1617 if (G1CollectedHeap::use_parallel_gc_threads()) {
1618 const uint OverpartitionFactor = 4;
1619 uint WorkUnit;
1620 // The use of MinChunkSize = 8 in the original code
1621 // causes some assertion failures when the total number of
1622 // region is less than 8. The code here tries to fix that.
1623 // Should the original code also be fixed?
1624 if (no_of_gc_threads > 0) {
1625 const uint MinWorkUnit = MAX2(region_num / no_of_gc_threads, 1U);
1626 WorkUnit = MAX2(region_num / (no_of_gc_threads * OverpartitionFactor),
1627 MinWorkUnit);
1628 } else {
1629 assert(no_of_gc_threads > 0,
1630 "The active gc workers should be greater than 0");
1631 // In a product build do something reasonable to avoid a crash.
1632 const uint MinWorkUnit = MAX2(region_num / (uint) ParallelGCThreads, 1U);
1633 WorkUnit =
1634 MAX2(region_num / (uint) (ParallelGCThreads * OverpartitionFactor),
1635 MinWorkUnit);
1636 }
1637 _collectionSetChooser->prepare_for_par_region_addition(_g1->num_regions(),
1638 WorkUnit);
1639 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser,
1640 (int) WorkUnit);
1641 _g1->workers()->run_task(&parKnownGarbageTask);
1642
1643 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
1644 "sanity check");
1645 } else {
1646 KnownGarbageClosure knownGarbagecl(_collectionSetChooser);
1647 _g1->heap_region_iterate(&knownGarbagecl);
1648 }
1649
1650 _collectionSetChooser->sort_regions();
1651
1652 double end_sec = os::elapsedTime();
1653 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0;
1654 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
1655 _cur_mark_stop_world_time_ms += elapsed_time_ms;
1656 _prev_collection_pause_end_ms += elapsed_time_ms;
1657 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec, true);
1658 }
1659
1660 // Add the heap region at the head of the non-incremental collection set
add_old_region_to_cset(HeapRegion * hr)1661 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) {
1662 assert(_inc_cset_build_state == Active, "Precondition");
1663 assert(hr->is_old(), "the region should be old");
1664
1665 assert(!hr->in_collection_set(), "should not already be in the CSet");
1666 hr->set_in_collection_set(true);
1667 hr->set_next_in_collection_set(_collection_set);
1668 _collection_set = hr;
1669 _collection_set_bytes_used_before += hr->used();
1670 _g1->register_old_region_with_in_cset_fast_test(hr);
1671 size_t rs_length = hr->rem_set()->occupied();
1672 _recorded_rs_lengths += rs_length;
1673 _old_cset_region_length += 1;
1674 }
1675
1676 // Initialize the per-collection-set information
start_incremental_cset_building()1677 void G1CollectorPolicy::start_incremental_cset_building() {
1678 assert(_inc_cset_build_state == Inactive, "Precondition");
1679
1680 _inc_cset_head = NULL;
1681 _inc_cset_tail = NULL;
1682 _inc_cset_bytes_used_before = 0;
1683
1684 _inc_cset_max_finger = 0;
1685 _inc_cset_recorded_rs_lengths = 0;
1686 _inc_cset_recorded_rs_lengths_diffs = 0;
1687 _inc_cset_predicted_elapsed_time_ms = 0.0;
1688 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
1689 _inc_cset_build_state = Active;
1690 }
1691
finalize_incremental_cset_building()1692 void G1CollectorPolicy::finalize_incremental_cset_building() {
1693 assert(_inc_cset_build_state == Active, "Precondition");
1694 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
1695
1696 // The two "main" fields, _inc_cset_recorded_rs_lengths and
1697 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread
1698 // that adds a new region to the CSet. Further updates by the
1699 // concurrent refinement thread that samples the young RSet lengths
1700 // are accumulated in the *_diffs fields. Here we add the diffs to
1701 // the "main" fields.
1702
1703 if (_inc_cset_recorded_rs_lengths_diffs >= 0) {
1704 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs;
1705 } else {
1706 // This is defensive. The diff should in theory be always positive
1707 // as RSets can only grow between GCs. However, given that we
1708 // sample their size concurrently with other threads updating them
1709 // it's possible that we might get the wrong size back, which
1710 // could make the calculations somewhat inaccurate.
1711 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs);
1712 if (_inc_cset_recorded_rs_lengths >= diffs) {
1713 _inc_cset_recorded_rs_lengths -= diffs;
1714 } else {
1715 _inc_cset_recorded_rs_lengths = 0;
1716 }
1717 }
1718 _inc_cset_predicted_elapsed_time_ms +=
1719 _inc_cset_predicted_elapsed_time_ms_diffs;
1720
1721 _inc_cset_recorded_rs_lengths_diffs = 0;
1722 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
1723 }
1724
add_to_incremental_cset_info(HeapRegion * hr,size_t rs_length)1725 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) {
1726 // This routine is used when:
1727 // * adding survivor regions to the incremental cset at the end of an
1728 // evacuation pause,
1729 // * adding the current allocation region to the incremental cset
1730 // when it is retired, and
1731 // * updating existing policy information for a region in the
1732 // incremental cset via young list RSet sampling.
1733 // Therefore this routine may be called at a safepoint by the
1734 // VM thread, or in-between safepoints by mutator threads (when
1735 // retiring the current allocation region) or a concurrent
1736 // refine thread (RSet sampling).
1737
1738 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, gcs_are_young());
1739 size_t used_bytes = hr->used();
1740 _inc_cset_recorded_rs_lengths += rs_length;
1741 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms;
1742 _inc_cset_bytes_used_before += used_bytes;
1743
1744 // Cache the values we have added to the aggregated informtion
1745 // in the heap region in case we have to remove this region from
1746 // the incremental collection set, or it is updated by the
1747 // rset sampling code
1748 hr->set_recorded_rs_length(rs_length);
1749 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms);
1750 }
1751
update_incremental_cset_info(HeapRegion * hr,size_t new_rs_length)1752 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr,
1753 size_t new_rs_length) {
1754 // Update the CSet information that is dependent on the new RS length
1755 assert(hr->is_young(), "Precondition");
1756 assert(!SafepointSynchronize::is_at_safepoint(),
1757 "should not be at a safepoint");
1758
1759 // We could have updated _inc_cset_recorded_rs_lengths and
1760 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do
1761 // that atomically, as this code is executed by a concurrent
1762 // refinement thread, potentially concurrently with a mutator thread
1763 // allocating a new region and also updating the same fields. To
1764 // avoid the atomic operations we accumulate these updates on two
1765 // separate fields (*_diffs) and we'll just add them to the "main"
1766 // fields at the start of a GC.
1767
1768 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length();
1769 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length;
1770 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff;
1771
1772 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms();
1773 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, gcs_are_young());
1774 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms;
1775 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff;
1776
1777 hr->set_recorded_rs_length(new_rs_length);
1778 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms);
1779 }
1780
add_region_to_incremental_cset_common(HeapRegion * hr)1781 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) {
1782 assert(hr->is_young(), "invariant");
1783 assert(hr->young_index_in_cset() > -1, "should have already been set");
1784 assert(_inc_cset_build_state == Active, "Precondition");
1785
1786 // We need to clear and set the cached recorded/cached collection set
1787 // information in the heap region here (before the region gets added
1788 // to the collection set). An individual heap region's cached values
1789 // are calculated, aggregated with the policy collection set info,
1790 // and cached in the heap region here (initially) and (subsequently)
1791 // by the Young List sampling code.
1792
1793 size_t rs_length = hr->rem_set()->occupied();
1794 add_to_incremental_cset_info(hr, rs_length);
1795
1796 HeapWord* hr_end = hr->end();
1797 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end);
1798
1799 assert(!hr->in_collection_set(), "invariant");
1800 hr->set_in_collection_set(true);
1801 assert( hr->next_in_collection_set() == NULL, "invariant");
1802
1803 _g1->register_young_region_with_in_cset_fast_test(hr);
1804 }
1805
1806 // Add the region at the RHS of the incremental cset
add_region_to_incremental_cset_rhs(HeapRegion * hr)1807 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) {
1808 // We should only ever be appending survivors at the end of a pause
1809 assert(hr->is_survivor(), "Logic");
1810
1811 // Do the 'common' stuff
1812 add_region_to_incremental_cset_common(hr);
1813
1814 // Now add the region at the right hand side
1815 if (_inc_cset_tail == NULL) {
1816 assert(_inc_cset_head == NULL, "invariant");
1817 _inc_cset_head = hr;
1818 } else {
1819 _inc_cset_tail->set_next_in_collection_set(hr);
1820 }
1821 _inc_cset_tail = hr;
1822 }
1823
1824 // Add the region to the LHS of the incremental cset
add_region_to_incremental_cset_lhs(HeapRegion * hr)1825 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) {
1826 // Survivors should be added to the RHS at the end of a pause
1827 assert(hr->is_eden(), "Logic");
1828
1829 // Do the 'common' stuff
1830 add_region_to_incremental_cset_common(hr);
1831
1832 // Add the region at the left hand side
1833 hr->set_next_in_collection_set(_inc_cset_head);
1834 if (_inc_cset_head == NULL) {
1835 assert(_inc_cset_tail == NULL, "Invariant");
1836 _inc_cset_tail = hr;
1837 }
1838 _inc_cset_head = hr;
1839 }
1840
1841 #ifndef PRODUCT
print_collection_set(HeapRegion * list_head,outputStream * st)1842 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) {
1843 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be");
1844
1845 st->print_cr("\nCollection_set:");
1846 HeapRegion* csr = list_head;
1847 while (csr != NULL) {
1848 HeapRegion* next = csr->next_in_collection_set();
1849 assert(csr->in_collection_set(), "bad CS");
1850 st->print_cr(" " HR_FORMAT ", P: " PTR_FORMAT "N: " PTR_FORMAT ", age: %4d",
1851 HR_FORMAT_PARAMS(csr),
1852 p2i(csr->prev_top_at_mark_start()), p2i(csr->next_top_at_mark_start()),
1853 csr->age_in_surv_rate_group_cond());
1854 csr = next;
1855 }
1856 }
1857 #endif // !PRODUCT
1858
reclaimable_bytes_perc(size_t reclaimable_bytes)1859 double G1CollectorPolicy::reclaimable_bytes_perc(size_t reclaimable_bytes) {
1860 // Returns the given amount of reclaimable bytes (that represents
1861 // the amount of reclaimable space still to be collected) as a
1862 // percentage of the current heap capacity.
1863 size_t capacity_bytes = _g1->capacity();
1864 return (double) reclaimable_bytes * 100.0 / (double) capacity_bytes;
1865 }
1866
next_gc_should_be_mixed(const char * true_action_str,const char * false_action_str)1867 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str,
1868 const char* false_action_str) {
1869 CollectionSetChooser* cset_chooser = _collectionSetChooser;
1870 if (cset_chooser->is_empty()) {
1871 ergo_verbose0(ErgoMixedGCs,
1872 false_action_str,
1873 ergo_format_reason("candidate old regions not available"));
1874 return false;
1875 }
1876
1877 // Is the amount of uncollected reclaimable space above G1HeapWastePercent?
1878 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes();
1879 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes);
1880 double threshold = (double) G1HeapWastePercent;
1881 if (reclaimable_perc <= threshold) {
1882 ergo_verbose4(ErgoMixedGCs,
1883 false_action_str,
1884 ergo_format_reason("reclaimable percentage not over threshold")
1885 ergo_format_region("candidate old regions")
1886 ergo_format_byte_perc("reclaimable")
1887 ergo_format_perc("threshold"),
1888 cset_chooser->remaining_regions(),
1889 reclaimable_bytes,
1890 reclaimable_perc, threshold);
1891 return false;
1892 }
1893
1894 ergo_verbose4(ErgoMixedGCs,
1895 true_action_str,
1896 ergo_format_reason("candidate old regions available")
1897 ergo_format_region("candidate old regions")
1898 ergo_format_byte_perc("reclaimable")
1899 ergo_format_perc("threshold"),
1900 cset_chooser->remaining_regions(),
1901 reclaimable_bytes,
1902 reclaimable_perc, threshold);
1903 return true;
1904 }
1905
calc_min_old_cset_length()1906 uint G1CollectorPolicy::calc_min_old_cset_length() {
1907 // The min old CSet region bound is based on the maximum desired
1908 // number of mixed GCs after a cycle. I.e., even if some old regions
1909 // look expensive, we should add them to the CSet anyway to make
1910 // sure we go through the available old regions in no more than the
1911 // maximum desired number of mixed GCs.
1912 //
1913 // The calculation is based on the number of marked regions we added
1914 // to the CSet chooser in the first place, not how many remain, so
1915 // that the result is the same during all mixed GCs that follow a cycle.
1916
1917 const size_t region_num = (size_t) _collectionSetChooser->length();
1918 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1);
1919 size_t result = region_num / gc_num;
1920 // emulate ceiling
1921 if (result * gc_num < region_num) {
1922 result += 1;
1923 }
1924 return (uint) result;
1925 }
1926
calc_max_old_cset_length()1927 uint G1CollectorPolicy::calc_max_old_cset_length() {
1928 // The max old CSet region bound is based on the threshold expressed
1929 // as a percentage of the heap size. I.e., it should bound the
1930 // number of old regions added to the CSet irrespective of how many
1931 // of them are available.
1932
1933 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1934 const size_t region_num = g1h->num_regions();
1935 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent;
1936 size_t result = region_num * perc / 100;
1937 // emulate ceiling
1938 if (100 * result < region_num * perc) {
1939 result += 1;
1940 }
1941 return (uint) result;
1942 }
1943
1944
finalize_cset(double target_pause_time_ms,EvacuationInfo & evacuation_info)1945 void G1CollectorPolicy::finalize_cset(double target_pause_time_ms, EvacuationInfo& evacuation_info) {
1946 double young_start_time_sec = os::elapsedTime();
1947
1948 YoungList* young_list = _g1->young_list();
1949 finalize_incremental_cset_building();
1950
1951 guarantee(target_pause_time_ms > 0.0,
1952 err_msg("target_pause_time_ms = %1.6lf should be positive",
1953 target_pause_time_ms));
1954 guarantee(_collection_set == NULL, "Precondition");
1955
1956 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
1957 double predicted_pause_time_ms = base_time_ms;
1958 double time_remaining_ms = MAX2(target_pause_time_ms - base_time_ms, 0.0);
1959
1960 ergo_verbose4(ErgoCSetConstruction | ErgoHigh,
1961 "start choosing CSet",
1962 ergo_format_size("_pending_cards")
1963 ergo_format_ms("predicted base time")
1964 ergo_format_ms("remaining time")
1965 ergo_format_ms("target pause time"),
1966 _pending_cards, base_time_ms, time_remaining_ms, target_pause_time_ms);
1967
1968 _last_gc_was_young = gcs_are_young() ? true : false;
1969
1970 if (_last_gc_was_young) {
1971 _trace_gen0_time_data.increment_young_collection_count();
1972 } else {
1973 _trace_gen0_time_data.increment_mixed_collection_count();
1974 }
1975
1976 // The young list is laid with the survivor regions from the previous
1977 // pause are appended to the RHS of the young list, i.e.
1978 // [Newly Young Regions ++ Survivors from last pause].
1979
1980 uint survivor_region_length = young_list->survivor_length();
1981 uint eden_region_length = young_list->length() - survivor_region_length;
1982 init_cset_region_lengths(eden_region_length, survivor_region_length);
1983
1984 HeapRegion* hr = young_list->first_survivor_region();
1985 while (hr != NULL) {
1986 assert(hr->is_survivor(), "badly formed young list");
1987 // There is a convention that all the young regions in the CSet
1988 // are tagged as "eden", so we do this for the survivors here. We
1989 // use the special set_eden_pre_gc() as it doesn't check that the
1990 // region is free (which is not the case here).
1991 hr->set_eden_pre_gc();
1992 hr = hr->get_next_young_region();
1993 }
1994
1995 // Clear the fields that point to the survivor list - they are all young now.
1996 young_list->clear_survivors();
1997
1998 _collection_set = _inc_cset_head;
1999 _collection_set_bytes_used_before = _inc_cset_bytes_used_before;
2000 time_remaining_ms = MAX2(time_remaining_ms - _inc_cset_predicted_elapsed_time_ms, 0.0);
2001 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms;
2002
2003 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2004 "add young regions to CSet",
2005 ergo_format_region("eden")
2006 ergo_format_region("survivors")
2007 ergo_format_ms("predicted young region time"),
2008 eden_region_length, survivor_region_length,
2009 _inc_cset_predicted_elapsed_time_ms);
2010
2011 // The number of recorded young regions is the incremental
2012 // collection set's current size
2013 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths);
2014
2015 double young_end_time_sec = os::elapsedTime();
2016 phase_times()->record_young_cset_choice_time_ms((young_end_time_sec - young_start_time_sec) * 1000.0);
2017
2018 // Set the start of the non-young choice time.
2019 double non_young_start_time_sec = young_end_time_sec;
2020
2021 if (!gcs_are_young()) {
2022 CollectionSetChooser* cset_chooser = _collectionSetChooser;
2023 cset_chooser->verify();
2024 const uint min_old_cset_length = calc_min_old_cset_length();
2025 const uint max_old_cset_length = calc_max_old_cset_length();
2026
2027 uint expensive_region_num = 0;
2028 bool check_time_remaining = adaptive_young_list_length();
2029
2030 HeapRegion* hr = cset_chooser->peek();
2031 while (hr != NULL) {
2032 if (old_cset_region_length() >= max_old_cset_length) {
2033 // Added maximum number of old regions to the CSet.
2034 ergo_verbose2(ErgoCSetConstruction,
2035 "finish adding old regions to CSet",
2036 ergo_format_reason("old CSet region num reached max")
2037 ergo_format_region("old")
2038 ergo_format_region("max"),
2039 old_cset_region_length(), max_old_cset_length);
2040 break;
2041 }
2042
2043
2044 // Stop adding regions if the remaining reclaimable space is
2045 // not above G1HeapWastePercent.
2046 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes();
2047 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes);
2048 double threshold = (double) G1HeapWastePercent;
2049 if (reclaimable_perc <= threshold) {
2050 // We've added enough old regions that the amount of uncollected
2051 // reclaimable space is at or below the waste threshold. Stop
2052 // adding old regions to the CSet.
2053 ergo_verbose5(ErgoCSetConstruction,
2054 "finish adding old regions to CSet",
2055 ergo_format_reason("reclaimable percentage not over threshold")
2056 ergo_format_region("old")
2057 ergo_format_region("max")
2058 ergo_format_byte_perc("reclaimable")
2059 ergo_format_perc("threshold"),
2060 old_cset_region_length(),
2061 max_old_cset_length,
2062 reclaimable_bytes,
2063 reclaimable_perc, threshold);
2064 break;
2065 }
2066
2067 double predicted_time_ms = predict_region_elapsed_time_ms(hr, gcs_are_young());
2068 if (check_time_remaining) {
2069 if (predicted_time_ms > time_remaining_ms) {
2070 // Too expensive for the current CSet.
2071
2072 if (old_cset_region_length() >= min_old_cset_length) {
2073 // We have added the minimum number of old regions to the CSet,
2074 // we are done with this CSet.
2075 ergo_verbose4(ErgoCSetConstruction,
2076 "finish adding old regions to CSet",
2077 ergo_format_reason("predicted time is too high")
2078 ergo_format_ms("predicted time")
2079 ergo_format_ms("remaining time")
2080 ergo_format_region("old")
2081 ergo_format_region("min"),
2082 predicted_time_ms, time_remaining_ms,
2083 old_cset_region_length(), min_old_cset_length);
2084 break;
2085 }
2086
2087 // We'll add it anyway given that we haven't reached the
2088 // minimum number of old regions.
2089 expensive_region_num += 1;
2090 }
2091 } else {
2092 if (old_cset_region_length() >= min_old_cset_length) {
2093 // In the non-auto-tuning case, we'll finish adding regions
2094 // to the CSet if we reach the minimum.
2095 ergo_verbose2(ErgoCSetConstruction,
2096 "finish adding old regions to CSet",
2097 ergo_format_reason("old CSet region num reached min")
2098 ergo_format_region("old")
2099 ergo_format_region("min"),
2100 old_cset_region_length(), min_old_cset_length);
2101 break;
2102 }
2103 }
2104
2105 // We will add this region to the CSet.
2106 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0);
2107 predicted_pause_time_ms += predicted_time_ms;
2108 cset_chooser->remove_and_move_to_next(hr);
2109 _g1->old_set_remove(hr);
2110 add_old_region_to_cset(hr);
2111
2112 hr = cset_chooser->peek();
2113 }
2114 if (hr == NULL) {
2115 ergo_verbose0(ErgoCSetConstruction,
2116 "finish adding old regions to CSet",
2117 ergo_format_reason("candidate old regions not available"));
2118 }
2119
2120 if (expensive_region_num > 0) {
2121 // We print the information once here at the end, predicated on
2122 // whether we added any apparently expensive regions or not, to
2123 // avoid generating output per region.
2124 ergo_verbose4(ErgoCSetConstruction,
2125 "added expensive regions to CSet",
2126 ergo_format_reason("old CSet region num not reached min")
2127 ergo_format_region("old")
2128 ergo_format_region("expensive")
2129 ergo_format_region("min")
2130 ergo_format_ms("remaining time"),
2131 old_cset_region_length(),
2132 expensive_region_num,
2133 min_old_cset_length,
2134 time_remaining_ms);
2135 }
2136
2137 cset_chooser->verify();
2138 }
2139
2140 stop_incremental_cset_building();
2141
2142 ergo_verbose5(ErgoCSetConstruction,
2143 "finish choosing CSet",
2144 ergo_format_region("eden")
2145 ergo_format_region("survivors")
2146 ergo_format_region("old")
2147 ergo_format_ms("predicted pause time")
2148 ergo_format_ms("target pause time"),
2149 eden_region_length, survivor_region_length,
2150 old_cset_region_length(),
2151 predicted_pause_time_ms, target_pause_time_ms);
2152
2153 double non_young_end_time_sec = os::elapsedTime();
2154 phase_times()->record_non_young_cset_choice_time_ms((non_young_end_time_sec - non_young_start_time_sec) * 1000.0);
2155 evacuation_info.set_collectionset_regions(cset_region_length());
2156 }
2157
record_start_collection(double time_to_stop_the_world_ms)2158 void TraceGen0TimeData::record_start_collection(double time_to_stop_the_world_ms) {
2159 if(TraceGen0Time) {
2160 _all_stop_world_times_ms.add(time_to_stop_the_world_ms);
2161 }
2162 }
2163
record_yield_time(double yield_time_ms)2164 void TraceGen0TimeData::record_yield_time(double yield_time_ms) {
2165 if(TraceGen0Time) {
2166 _all_yield_times_ms.add(yield_time_ms);
2167 }
2168 }
2169
record_end_collection(double pause_time_ms,G1GCPhaseTimes * phase_times)2170 void TraceGen0TimeData::record_end_collection(double pause_time_ms, G1GCPhaseTimes* phase_times) {
2171 if(TraceGen0Time) {
2172 _total.add(pause_time_ms);
2173 _other.add(pause_time_ms - phase_times->accounted_time_ms());
2174 _root_region_scan_wait.add(phase_times->root_region_scan_wait_time_ms());
2175 _parallel.add(phase_times->cur_collection_par_time_ms());
2176 _ext_root_scan.add(phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan));
2177 _satb_filtering.add(phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering));
2178 _update_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS));
2179 _scan_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::ScanRS));
2180 _obj_copy.add(phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy));
2181 _termination.add(phase_times->average_time_ms(G1GCPhaseTimes::Termination));
2182
2183 double parallel_known_time = phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan) +
2184 phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering) +
2185 phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS) +
2186 phase_times->average_time_ms(G1GCPhaseTimes::ScanRS) +
2187 phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy) +
2188 phase_times->average_time_ms(G1GCPhaseTimes::Termination);
2189
2190 double parallel_other_time = phase_times->cur_collection_par_time_ms() - parallel_known_time;
2191 _parallel_other.add(parallel_other_time);
2192 _clear_ct.add(phase_times->cur_clear_ct_time_ms());
2193 }
2194 }
2195
increment_young_collection_count()2196 void TraceGen0TimeData::increment_young_collection_count() {
2197 if(TraceGen0Time) {
2198 ++_young_pause_num;
2199 }
2200 }
2201
increment_mixed_collection_count()2202 void TraceGen0TimeData::increment_mixed_collection_count() {
2203 if(TraceGen0Time) {
2204 ++_mixed_pause_num;
2205 }
2206 }
2207
print_summary(const char * str,const NumberSeq * seq) const2208 void TraceGen0TimeData::print_summary(const char* str,
2209 const NumberSeq* seq) const {
2210 double sum = seq->sum();
2211 gclog_or_tty->print_cr("%-27s = %8.2lf s (avg = %8.2lf ms)",
2212 str, sum / 1000.0, seq->avg());
2213 }
2214
print_summary_sd(const char * str,const NumberSeq * seq) const2215 void TraceGen0TimeData::print_summary_sd(const char* str,
2216 const NumberSeq* seq) const {
2217 print_summary(str, seq);
2218 gclog_or_tty->print_cr("%45s = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
2219 "(num", seq->num(), seq->sd(), seq->maximum());
2220 }
2221
print() const2222 void TraceGen0TimeData::print() const {
2223 if (!TraceGen0Time) {
2224 return;
2225 }
2226
2227 gclog_or_tty->print_cr("ALL PAUSES");
2228 print_summary_sd(" Total", &_total);
2229 gclog_or_tty->cr();
2230 gclog_or_tty->cr();
2231 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num);
2232 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num);
2233 gclog_or_tty->cr();
2234
2235 gclog_or_tty->print_cr("EVACUATION PAUSES");
2236
2237 if (_young_pause_num == 0 && _mixed_pause_num == 0) {
2238 gclog_or_tty->print_cr("none");
2239 } else {
2240 print_summary_sd(" Evacuation Pauses", &_total);
2241 print_summary(" Root Region Scan Wait", &_root_region_scan_wait);
2242 print_summary(" Parallel Time", &_parallel);
2243 print_summary(" Ext Root Scanning", &_ext_root_scan);
2244 print_summary(" SATB Filtering", &_satb_filtering);
2245 print_summary(" Update RS", &_update_rs);
2246 print_summary(" Scan RS", &_scan_rs);
2247 print_summary(" Object Copy", &_obj_copy);
2248 print_summary(" Termination", &_termination);
2249 print_summary(" Parallel Other", &_parallel_other);
2250 print_summary(" Clear CT", &_clear_ct);
2251 print_summary(" Other", &_other);
2252 }
2253 gclog_or_tty->cr();
2254
2255 gclog_or_tty->print_cr("MISC");
2256 print_summary_sd(" Stop World", &_all_stop_world_times_ms);
2257 print_summary_sd(" Yields", &_all_yield_times_ms);
2258 }
2259
record_full_collection(double full_gc_time_ms)2260 void TraceGen1TimeData::record_full_collection(double full_gc_time_ms) {
2261 if (TraceGen1Time) {
2262 _all_full_gc_times.add(full_gc_time_ms);
2263 }
2264 }
2265
print() const2266 void TraceGen1TimeData::print() const {
2267 if (!TraceGen1Time) {
2268 return;
2269 }
2270
2271 if (_all_full_gc_times.num() > 0) {
2272 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s",
2273 _all_full_gc_times.num(),
2274 _all_full_gc_times.sum() / 1000.0);
2275 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg());
2276 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]",
2277 _all_full_gc_times.sd(),
2278 _all_full_gc_times.maximum());
2279 }
2280 }
2281