rev 3484 : 7178363: G1: Remove the serial code for PrintGCDetails and make it a special case of the parallel code
Summary: Introduced the WorkerDataArray class.
Reviewed-by: mgerdin
1 /*
2 * Copyright (c) 2001, 2012, 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
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 _pending_card_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
94 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
95 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
96 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
97 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
98 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
99 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
100 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
101 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)),
102 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
103 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
104 _non_young_other_cost_per_region_ms_seq(
105 new TruncatedSeq(TruncatedSeqLength)),
106
107 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)),
108 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)),
109
110 _pause_time_target_ms((double) MaxGCPauseMillis),
111
112 _gcs_are_young(true),
113
114 _during_marking(false),
115 _in_marking_window(false),
116 _in_marking_window_im(false),
117
118 _recent_prev_end_times_for_all_gcs_sec(
119 new TruncatedSeq(NumPrevPausesForHeuristics)),
120
121 _recent_avg_pause_time_ratio(0.0),
122
123 _initiate_conc_mark_if_possible(false),
124 _during_initial_mark_pause(false),
125 _last_young_gc(false),
126 _last_gc_was_young(false),
127
128 _eden_bytes_before_gc(0),
129 _survivor_bytes_before_gc(0),
130 _capacity_before_gc(0),
131
132 _eden_cset_region_length(0),
133 _survivor_cset_region_length(0),
134 _old_cset_region_length(0),
135
136 _collection_set(NULL),
137 _collection_set_bytes_used_before(0),
138
139 // Incremental CSet attributes
140 _inc_cset_build_state(Inactive),
141 _inc_cset_head(NULL),
142 _inc_cset_tail(NULL),
143 _inc_cset_bytes_used_before(0),
144 _inc_cset_max_finger(NULL),
145 _inc_cset_recorded_rs_lengths(0),
146 _inc_cset_recorded_rs_lengths_diffs(0),
147 _inc_cset_predicted_elapsed_time_ms(0.0),
148 _inc_cset_predicted_elapsed_time_ms_diffs(0.0),
149
150 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
151 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
152 #endif // _MSC_VER
153
154 _short_lived_surv_rate_group(new SurvRateGroup(this, "Short Lived",
155 G1YoungSurvRateNumRegionsSummary)),
156 _survivor_surv_rate_group(new SurvRateGroup(this, "Survivor",
157 G1YoungSurvRateNumRegionsSummary)),
158 // add here any more surv rate groups
159 _recorded_survivor_regions(0),
160 _recorded_survivor_head(NULL),
161 _recorded_survivor_tail(NULL),
162 _survivors_age_table(true),
163
164 _gc_overhead_perc(0.0) {
165
166 // Set up the region size and associated fields. Given that the
167 // policy is created before the heap, we have to set this up here,
168 // so it's done as soon as possible.
169 HeapRegion::setup_heap_region_size(Arguments::min_heap_size());
170 HeapRegionRemSet::setup_remset_size();
171
172 G1ErgoVerbose::initialize();
173 if (PrintAdaptiveSizePolicy) {
174 // Currently, we only use a single switch for all the heuristics.
175 G1ErgoVerbose::set_enabled(true);
176 // Given that we don't currently have a verboseness level
177 // parameter, we'll hardcode this to high. This can be easily
178 // changed in the future.
179 G1ErgoVerbose::set_level(ErgoHigh);
180 } else {
181 G1ErgoVerbose::set_enabled(false);
182 }
183
184 // Verify PLAB sizes
185 const size_t region_size = HeapRegion::GrainWords;
186 if (YoungPLABSize > region_size || OldPLABSize > region_size) {
187 char buffer[128];
188 jio_snprintf(buffer, sizeof(buffer), "%sPLABSize should be at most "SIZE_FORMAT,
189 OldPLABSize > region_size ? "Old" : "Young", region_size);
190 vm_exit_during_initialization(buffer);
191 }
192
193 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime());
194 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0;
195
196 _phase_times = new G1GCPhaseTimes(_parallel_gc_threads);
197
198 int index = MIN2(_parallel_gc_threads - 1, 7);
199
200 _pending_card_diff_seq->add(0.0);
201 _rs_length_diff_seq->add(rs_length_diff_defaults[index]);
202 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]);
203 _young_cards_per_entry_ratio_seq->add(
204 young_cards_per_entry_ratio_defaults[index]);
205 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]);
206 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]);
207 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]);
208 _young_other_cost_per_region_ms_seq->add(
209 young_other_cost_per_region_ms_defaults[index]);
210 _non_young_other_cost_per_region_ms_seq->add(
211 non_young_other_cost_per_region_ms_defaults[index]);
212
213 // Below, we might need to calculate the pause time target based on
214 // the pause interval. When we do so we are going to give G1 maximum
215 // flexibility and allow it to do pauses when it needs to. So, we'll
216 // arrange that the pause interval to be pause time target + 1 to
217 // ensure that a) the pause time target is maximized with respect to
218 // the pause interval and b) we maintain the invariant that pause
219 // time target < pause interval. If the user does not want this
220 // maximum flexibility, they will have to set the pause interval
221 // explicitly.
222
223 // First make sure that, if either parameter is set, its value is
224 // reasonable.
225 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
226 if (MaxGCPauseMillis < 1) {
227 vm_exit_during_initialization("MaxGCPauseMillis should be "
228 "greater than 0");
229 }
230 }
231 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
232 if (GCPauseIntervalMillis < 1) {
233 vm_exit_during_initialization("GCPauseIntervalMillis should be "
234 "greater than 0");
235 }
236 }
237
238 // Then, if the pause time target parameter was not set, set it to
239 // the default value.
240 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
241 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
242 // The default pause time target in G1 is 200ms
243 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200);
244 } else {
245 // We do not allow the pause interval to be set without the
246 // pause time target
247 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set "
248 "without setting MaxGCPauseMillis");
249 }
250 }
251
252 // Then, if the interval parameter was not set, set it according to
253 // the pause time target (this will also deal with the case when the
254 // pause time target is the default value).
255 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
256 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1);
257 }
258
259 // Finally, make sure that the two parameters are consistent.
260 if (MaxGCPauseMillis >= GCPauseIntervalMillis) {
261 char buffer[256];
262 jio_snprintf(buffer, 256,
263 "MaxGCPauseMillis (%u) should be less than "
264 "GCPauseIntervalMillis (%u)",
265 MaxGCPauseMillis, GCPauseIntervalMillis);
266 vm_exit_during_initialization(buffer);
267 }
268
269 double max_gc_time = (double) MaxGCPauseMillis / 1000.0;
270 double time_slice = (double) GCPauseIntervalMillis / 1000.0;
271 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time);
272 _sigma = (double) G1ConfidencePercent / 100.0;
273
274 // start conservatively (around 50ms is about right)
275 _concurrent_mark_remark_times_ms->add(0.05);
276 _concurrent_mark_cleanup_times_ms->add(0.20);
277 _tenuring_threshold = MaxTenuringThreshold;
278 // _max_survivor_regions will be calculated by
279 // update_young_list_target_length() during initialization.
280 _max_survivor_regions = 0;
281
282 assert(GCTimeRatio > 0,
283 "we should have set it to a default value set_g1_gc_flags() "
284 "if a user set it to 0");
285 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio));
286
287 uintx reserve_perc = G1ReservePercent;
288 // Put an artificial ceiling on this so that it's not set to a silly value.
289 if (reserve_perc > 50) {
290 reserve_perc = 50;
291 warning("G1ReservePercent is set to a value that is too large, "
292 "it's been updated to %u", reserve_perc);
293 }
294 _reserve_factor = (double) reserve_perc / 100.0;
295 // This will be set when the heap is expanded
296 // for the first time during initialization.
297 _reserve_regions = 0;
298
299 initialize_all();
300 _collectionSetChooser = new CollectionSetChooser();
301 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags
302 }
303
304 void G1CollectorPolicy::initialize_flags() {
305 set_min_alignment(HeapRegion::GrainBytes);
306 set_max_alignment(GenRemSet::max_alignment_constraint(rem_set_name()));
307 if (SurvivorRatio < 1) {
308 vm_exit_during_initialization("Invalid survivor ratio specified");
309 }
310 CollectorPolicy::initialize_flags();
311 }
312
313 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true) {
314 assert(G1DefaultMinNewGenPercent <= G1DefaultMaxNewGenPercent, "Min larger than max");
315 assert(G1DefaultMinNewGenPercent > 0 && G1DefaultMinNewGenPercent < 100, "Min out of bounds");
316 assert(G1DefaultMaxNewGenPercent > 0 && G1DefaultMaxNewGenPercent < 100, "Max out of bounds");
317
318 if (FLAG_IS_CMDLINE(NewRatio)) {
319 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) {
320 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio");
321 } else {
322 _sizer_kind = SizerNewRatio;
323 _adaptive_size = false;
324 return;
325 }
326 }
327
328 if (FLAG_IS_CMDLINE(NewSize)) {
329 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes),
330 1U);
331 if (FLAG_IS_CMDLINE(MaxNewSize)) {
332 _max_desired_young_length =
333 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes),
334 1U);
335 _sizer_kind = SizerMaxAndNewSize;
336 _adaptive_size = _min_desired_young_length == _max_desired_young_length;
337 } else {
338 _sizer_kind = SizerNewSizeOnly;
339 }
340 } else if (FLAG_IS_CMDLINE(MaxNewSize)) {
341 _max_desired_young_length =
342 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes),
343 1U);
344 _sizer_kind = SizerMaxNewSizeOnly;
345 }
346 }
347
348 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) {
349 uint default_value = (new_number_of_heap_regions * G1DefaultMinNewGenPercent) / 100;
350 return MAX2(1U, default_value);
351 }
352
353 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) {
354 uint default_value = (new_number_of_heap_regions * G1DefaultMaxNewGenPercent) / 100;
355 return MAX2(1U, default_value);
356 }
357
358 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) {
359 assert(new_number_of_heap_regions > 0, "Heap must be initialized");
360
361 switch (_sizer_kind) {
362 case SizerDefaults:
363 _min_desired_young_length = calculate_default_min_length(new_number_of_heap_regions);
364 _max_desired_young_length = calculate_default_max_length(new_number_of_heap_regions);
365 break;
366 case SizerNewSizeOnly:
367 _max_desired_young_length = calculate_default_max_length(new_number_of_heap_regions);
368 _max_desired_young_length = MAX2(_min_desired_young_length, _max_desired_young_length);
369 break;
370 case SizerMaxNewSizeOnly:
371 _min_desired_young_length = calculate_default_min_length(new_number_of_heap_regions);
372 _min_desired_young_length = MIN2(_min_desired_young_length, _max_desired_young_length);
373 break;
374 case SizerMaxAndNewSize:
375 // Do nothing. Values set on the command line, don't update them at runtime.
376 break;
377 case SizerNewRatio:
378 _min_desired_young_length = new_number_of_heap_regions / (NewRatio + 1);
379 _max_desired_young_length = _min_desired_young_length;
380 break;
381 default:
382 ShouldNotReachHere();
383 }
384
385 assert(_min_desired_young_length <= _max_desired_young_length, "Invalid min/max young gen size values");
386 }
387
388 void G1CollectorPolicy::init() {
389 // Set aside an initial future to_space.
390 _g1 = G1CollectedHeap::heap();
391
392 assert(Heap_lock->owned_by_self(), "Locking discipline.");
393
394 initialize_gc_policy_counters();
395
396 if (adaptive_young_list_length()) {
397 _young_list_fixed_length = 0;
398 } else {
399 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length();
400 }
401 _free_regions_at_end_of_collection = _g1->free_regions();
402 update_young_list_target_length();
403 _prev_eden_capacity = _young_list_target_length * HeapRegion::GrainBytes;
404
405 // We may immediately start allocating regions and placing them on the
406 // collection set list. Initialize the per-collection set info
407 start_incremental_cset_building();
408 }
409
410 // Create the jstat counters for the policy.
411 void G1CollectorPolicy::initialize_gc_policy_counters() {
412 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3);
413 }
414
415 bool G1CollectorPolicy::predict_will_fit(uint young_length,
416 double base_time_ms,
417 uint base_free_regions,
418 double target_pause_time_ms) {
419 if (young_length >= base_free_regions) {
420 // end condition 1: not enough space for the young regions
421 return false;
422 }
423
424 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1);
425 size_t bytes_to_copy =
426 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
427 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
428 double young_other_time_ms = predict_young_other_time_ms(young_length);
429 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms;
430 if (pause_time_ms > target_pause_time_ms) {
431 // end condition 2: prediction is over the target pause time
432 return false;
433 }
434
435 size_t free_bytes =
436 (base_free_regions - young_length) * HeapRegion::GrainBytes;
437 if ((2.0 * sigma()) * (double) bytes_to_copy > (double) free_bytes) {
438 // end condition 3: out-of-space (conservatively!)
439 return false;
440 }
441
442 // success!
443 return true;
444 }
445
446 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) {
447 // re-calculate the necessary reserve
448 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor;
449 // We use ceiling so that if reserve_regions_d is > 0.0 (but
450 // smaller than 1.0) we'll get 1.
451 _reserve_regions = (uint) ceil(reserve_regions_d);
452
453 _young_gen_sizer->heap_size_changed(new_number_of_regions);
454 }
455
456 uint G1CollectorPolicy::calculate_young_list_desired_min_length(
457 uint base_min_length) {
458 uint desired_min_length = 0;
459 if (adaptive_young_list_length()) {
460 if (_alloc_rate_ms_seq->num() > 3) {
461 double now_sec = os::elapsedTime();
462 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
463 double alloc_rate_ms = predict_alloc_rate_ms();
464 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms);
465 } else {
466 // otherwise we don't have enough info to make the prediction
467 }
468 }
469 desired_min_length += base_min_length;
470 // make sure we don't go below any user-defined minimum bound
471 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length);
472 }
473
474 uint G1CollectorPolicy::calculate_young_list_desired_max_length() {
475 // Here, we might want to also take into account any additional
476 // constraints (i.e., user-defined minimum bound). Currently, we
477 // effectively don't set this bound.
478 return _young_gen_sizer->max_desired_young_length();
479 }
480
481 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) {
482 if (rs_lengths == (size_t) -1) {
483 // if it's set to the default value (-1), we should predict it;
484 // otherwise, use the given value.
485 rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
486 }
487
488 // Calculate the absolute and desired min bounds.
489
490 // This is how many young regions we already have (currently: the survivors).
491 uint base_min_length = recorded_survivor_regions();
492 // This is the absolute minimum young length, which ensures that we
493 // can allocate one eden region in the worst-case.
494 uint absolute_min_length = base_min_length + 1;
495 uint desired_min_length =
496 calculate_young_list_desired_min_length(base_min_length);
497 if (desired_min_length < absolute_min_length) {
498 desired_min_length = absolute_min_length;
499 }
500
501 // Calculate the absolute and desired max bounds.
502
503 // We will try our best not to "eat" into the reserve.
504 uint absolute_max_length = 0;
505 if (_free_regions_at_end_of_collection > _reserve_regions) {
506 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions;
507 }
508 uint desired_max_length = calculate_young_list_desired_max_length();
509 if (desired_max_length > absolute_max_length) {
510 desired_max_length = absolute_max_length;
511 }
512
513 uint young_list_target_length = 0;
514 if (adaptive_young_list_length()) {
515 if (gcs_are_young()) {
516 young_list_target_length =
517 calculate_young_list_target_length(rs_lengths,
518 base_min_length,
519 desired_min_length,
520 desired_max_length);
521 _rs_lengths_prediction = rs_lengths;
522 } else {
523 // Don't calculate anything and let the code below bound it to
524 // the desired_min_length, i.e., do the next GC as soon as
525 // possible to maximize how many old regions we can add to it.
526 }
527 } else {
528 // The user asked for a fixed young gen so we'll fix the young gen
529 // whether the next GC is young or mixed.
530 young_list_target_length = _young_list_fixed_length;
531 }
532
533 // Make sure we don't go over the desired max length, nor under the
534 // desired min length. In case they clash, desired_min_length wins
535 // which is why that test is second.
536 if (young_list_target_length > desired_max_length) {
537 young_list_target_length = desired_max_length;
538 }
539 if (young_list_target_length < desired_min_length) {
540 young_list_target_length = desired_min_length;
541 }
542
543 assert(young_list_target_length > recorded_survivor_regions(),
544 "we should be able to allocate at least one eden region");
545 assert(young_list_target_length >= absolute_min_length, "post-condition");
546 _young_list_target_length = young_list_target_length;
547
548 update_max_gc_locker_expansion();
549 }
550
551 uint
552 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths,
553 uint base_min_length,
554 uint desired_min_length,
555 uint desired_max_length) {
556 assert(adaptive_young_list_length(), "pre-condition");
557 assert(gcs_are_young(), "only call this for young GCs");
558
559 // In case some edge-condition makes the desired max length too small...
560 if (desired_max_length <= desired_min_length) {
561 return desired_min_length;
562 }
563
564 // We'll adjust min_young_length and max_young_length not to include
565 // the already allocated young regions (i.e., so they reflect the
566 // min and max eden regions we'll allocate). The base_min_length
567 // will be reflected in the predictions by the
568 // survivor_regions_evac_time prediction.
569 assert(desired_min_length > base_min_length, "invariant");
570 uint min_young_length = desired_min_length - base_min_length;
571 assert(desired_max_length > base_min_length, "invariant");
572 uint max_young_length = desired_max_length - base_min_length;
573
574 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
575 double survivor_regions_evac_time = predict_survivor_regions_evac_time();
576 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
577 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
578 size_t scanned_cards = predict_young_card_num(adj_rs_lengths);
579 double base_time_ms =
580 predict_base_elapsed_time_ms(pending_cards, scanned_cards) +
581 survivor_regions_evac_time;
582 uint available_free_regions = _free_regions_at_end_of_collection;
583 uint base_free_regions = 0;
584 if (available_free_regions > _reserve_regions) {
585 base_free_regions = available_free_regions - _reserve_regions;
586 }
587
588 // Here, we will make sure that the shortest young length that
589 // makes sense fits within the target pause time.
590
591 if (predict_will_fit(min_young_length, base_time_ms,
592 base_free_regions, target_pause_time_ms)) {
593 // The shortest young length will fit into the target pause time;
594 // we'll now check whether the absolute maximum number of young
595 // regions will fit in the target pause time. If not, we'll do
596 // a binary search between min_young_length and max_young_length.
597 if (predict_will_fit(max_young_length, base_time_ms,
598 base_free_regions, target_pause_time_ms)) {
599 // The maximum young length will fit into the target pause time.
600 // We are done so set min young length to the maximum length (as
601 // the result is assumed to be returned in min_young_length).
602 min_young_length = max_young_length;
603 } else {
604 // The maximum possible number of young regions will not fit within
605 // the target pause time so we'll search for the optimal
606 // length. The loop invariants are:
607 //
608 // min_young_length < max_young_length
609 // min_young_length is known to fit into the target pause time
610 // max_young_length is known not to fit into the target pause time
611 //
612 // Going into the loop we know the above hold as we've just
613 // checked them. Every time around the loop we check whether
614 // the middle value between min_young_length and
615 // max_young_length fits into the target pause time. If it
616 // does, it becomes the new min. If it doesn't, it becomes
617 // the new max. This way we maintain the loop invariants.
618
619 assert(min_young_length < max_young_length, "invariant");
620 uint diff = (max_young_length - min_young_length) / 2;
621 while (diff > 0) {
622 uint young_length = min_young_length + diff;
623 if (predict_will_fit(young_length, base_time_ms,
624 base_free_regions, target_pause_time_ms)) {
625 min_young_length = young_length;
626 } else {
627 max_young_length = young_length;
628 }
629 assert(min_young_length < max_young_length, "invariant");
630 diff = (max_young_length - min_young_length) / 2;
631 }
632 // The results is min_young_length which, according to the
633 // loop invariants, should fit within the target pause time.
634
635 // These are the post-conditions of the binary search above:
636 assert(min_young_length < max_young_length,
637 "otherwise we should have discovered that max_young_length "
638 "fits into the pause target and not done the binary search");
639 assert(predict_will_fit(min_young_length, base_time_ms,
640 base_free_regions, target_pause_time_ms),
641 "min_young_length, the result of the binary search, should "
642 "fit into the pause target");
643 assert(!predict_will_fit(min_young_length + 1, base_time_ms,
644 base_free_regions, target_pause_time_ms),
645 "min_young_length, the result of the binary search, should be "
646 "optimal, so no larger length should fit into the pause target");
647 }
648 } else {
649 // Even the minimum length doesn't fit into the pause time
650 // target, return it as the result nevertheless.
651 }
652 return base_min_length + min_young_length;
653 }
654
655 double G1CollectorPolicy::predict_survivor_regions_evac_time() {
656 double survivor_regions_evac_time = 0.0;
657 for (HeapRegion * r = _recorded_survivor_head;
658 r != NULL && r != _recorded_survivor_tail->get_next_young_region();
659 r = r->get_next_young_region()) {
660 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, true);
661 }
662 return survivor_regions_evac_time;
663 }
664
665 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() {
666 guarantee( adaptive_young_list_length(), "should not call this otherwise" );
667
668 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths();
669 if (rs_lengths > _rs_lengths_prediction) {
670 // add 10% to avoid having to recalculate often
671 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000;
672 update_young_list_target_length(rs_lengths_prediction);
673 }
674 }
675
676
677
678 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size,
679 bool is_tlab,
680 bool* gc_overhead_limit_was_exceeded) {
681 guarantee(false, "Not using this policy feature yet.");
682 return NULL;
683 }
684
685 // This method controls how a collector handles one or more
686 // of its generations being fully allocated.
687 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size,
688 bool is_tlab) {
689 guarantee(false, "Not using this policy feature yet.");
690 return NULL;
691 }
692
693
694 #ifndef PRODUCT
695 bool G1CollectorPolicy::verify_young_ages() {
696 HeapRegion* head = _g1->young_list()->first_region();
697 return
698 verify_young_ages(head, _short_lived_surv_rate_group);
699 // also call verify_young_ages on any additional surv rate groups
700 }
701
702 bool
703 G1CollectorPolicy::verify_young_ages(HeapRegion* head,
704 SurvRateGroup *surv_rate_group) {
705 guarantee( surv_rate_group != NULL, "pre-condition" );
706
707 const char* name = surv_rate_group->name();
708 bool ret = true;
709 int prev_age = -1;
710
711 for (HeapRegion* curr = head;
712 curr != NULL;
713 curr = curr->get_next_young_region()) {
714 SurvRateGroup* group = curr->surv_rate_group();
715 if (group == NULL && !curr->is_survivor()) {
716 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name);
717 ret = false;
718 }
719
720 if (surv_rate_group == group) {
721 int age = curr->age_in_surv_rate_group();
722
723 if (age < 0) {
724 gclog_or_tty->print_cr("## %s: encountered negative age", name);
725 ret = false;
726 }
727
728 if (age <= prev_age) {
729 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing "
730 "(%d, %d)", name, age, prev_age);
731 ret = false;
732 }
733 prev_age = age;
734 }
735 }
736
737 return ret;
738 }
739 #endif // PRODUCT
740
741 void G1CollectorPolicy::record_full_collection_start() {
742 _full_collection_start_sec = os::elapsedTime();
743 // Release the future to-space so that it is available for compaction into.
744 _g1->set_full_collection();
745 }
746
747 void G1CollectorPolicy::record_full_collection_end() {
748 // Consider this like a collection pause for the purposes of allocation
749 // since last pause.
750 double end_sec = os::elapsedTime();
751 double full_gc_time_sec = end_sec - _full_collection_start_sec;
752 double full_gc_time_ms = full_gc_time_sec * 1000.0;
753
754 _trace_gen1_time_data.record_full_collection(full_gc_time_ms);
755
756 update_recent_gc_times(end_sec, full_gc_time_ms);
757
758 _g1->clear_full_collection();
759
760 // "Nuke" the heuristics that control the young/mixed GC
761 // transitions and make sure we start with young GCs after the Full GC.
762 set_gcs_are_young(true);
763 _last_young_gc = false;
764 clear_initiate_conc_mark_if_possible();
765 clear_during_initial_mark_pause();
766 _in_marking_window = false;
767 _in_marking_window_im = false;
768
769 _short_lived_surv_rate_group->start_adding_regions();
770 // also call this on any additional surv rate groups
771
772 record_survivor_regions(0, NULL, NULL);
773
774 _free_regions_at_end_of_collection = _g1->free_regions();
775 // Reset survivors SurvRateGroup.
776 _survivor_surv_rate_group->reset();
777 update_young_list_target_length();
778 _collectionSetChooser->clear();
779 }
780
781 void G1CollectorPolicy::record_stop_world_start() {
782 _stop_world_start = os::elapsedTime();
783 }
784
785 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec,
786 size_t start_used) {
787 // We only need to do this here as the policy will only be applied
788 // to the GC we're about to start. so, no point is calculating this
789 // every time we calculate / recalculate the target young length.
790 update_survivors_policy();
791
792 assert(_g1->used() == _g1->recalculate_used(),
793 err_msg("sanity, used: "SIZE_FORMAT" recalculate_used: "SIZE_FORMAT,
794 _g1->used(), _g1->recalculate_used()));
795
796 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0;
797 _trace_gen0_time_data.record_start_collection(s_w_t_ms);
798 _stop_world_start = 0.0;
799
800 phase_times()->_cur_collection_start_sec = start_time_sec;
801 _cur_collection_pause_used_at_start_bytes = start_used;
802 _cur_collection_pause_used_regions_at_start = _g1->used_regions();
803 _pending_cards = _g1->pending_card_num();
804 _max_pending_cards = _g1->max_pending_card_num();
805
806 _bytes_in_collection_set_before_gc = 0;
807 _bytes_copied_during_gc = 0;
808
809 YoungList* young_list = _g1->young_list();
810 _eden_bytes_before_gc = young_list->eden_used_bytes();
811 _survivor_bytes_before_gc = young_list->survivor_used_bytes();
812 _capacity_before_gc = _g1->capacity();
813
814 _last_gc_was_young = false;
815
816 // do that for any other surv rate groups
817 _short_lived_surv_rate_group->stop_adding_regions();
818 _survivors_age_table.clear();
819
820 assert( verify_young_ages(), "region age verification" );
821 }
822
823 void G1CollectorPolicy::record_concurrent_mark_init_end(double
824 mark_init_elapsed_time_ms) {
825 _during_marking = true;
826 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now");
827 clear_during_initial_mark_pause();
828 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms;
829 }
830
831 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
832 _mark_remark_start_sec = os::elapsedTime();
833 _during_marking = false;
834 }
835
836 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
837 double end_time_sec = os::elapsedTime();
838 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
839 _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
840 _cur_mark_stop_world_time_ms += elapsed_time_ms;
841 _prev_collection_pause_end_ms += elapsed_time_ms;
842
843 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true);
844 }
845
846 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
847 _mark_cleanup_start_sec = os::elapsedTime();
848 }
849
850 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
851 _last_young_gc = true;
852 _in_marking_window = false;
853 }
854
855 void G1CollectorPolicy::record_concurrent_pause() {
856 if (_stop_world_start > 0.0) {
857 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0;
858 _trace_gen0_time_data.record_yield_time(yield_ms);
859 }
860 }
861
862 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) {
863 if (_g1->concurrent_mark()->cmThread()->during_cycle()) {
864 return false;
865 }
866
867 size_t marking_initiating_used_threshold =
868 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent;
869 size_t cur_used_bytes = _g1->non_young_capacity_bytes();
870 size_t alloc_byte_size = alloc_word_size * HeapWordSize;
871
872 if ((cur_used_bytes + alloc_byte_size) > marking_initiating_used_threshold) {
873 if (gcs_are_young()) {
874 ergo_verbose5(ErgoConcCycles,
875 "request concurrent cycle initiation",
876 ergo_format_reason("occupancy higher than threshold")
877 ergo_format_byte("occupancy")
878 ergo_format_byte("allocation request")
879 ergo_format_byte_perc("threshold")
880 ergo_format_str("source"),
881 cur_used_bytes,
882 alloc_byte_size,
883 marking_initiating_used_threshold,
884 (double) InitiatingHeapOccupancyPercent,
885 source);
886 return true;
887 } else {
888 ergo_verbose5(ErgoConcCycles,
889 "do not request concurrent cycle initiation",
890 ergo_format_reason("still doing mixed collections")
891 ergo_format_byte("occupancy")
892 ergo_format_byte("allocation request")
893 ergo_format_byte_perc("threshold")
894 ergo_format_str("source"),
895 cur_used_bytes,
896 alloc_byte_size,
897 marking_initiating_used_threshold,
898 (double) InitiatingHeapOccupancyPercent,
899 source);
900 }
901 }
902
903 return false;
904 }
905
906 // Anything below that is considered to be zero
907 #define MIN_TIMER_GRANULARITY 0.0000001
908
909 void G1CollectorPolicy::record_collection_pause_end(double pause_time_ms) {
910 double end_time_sec = os::elapsedTime();
911 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(),
912 "otherwise, the subtraction below does not make sense");
913 size_t rs_size =
914 _cur_collection_pause_used_regions_at_start - cset_region_length();
915 size_t cur_used_bytes = _g1->used();
916 assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
917 bool last_pause_included_initial_mark = false;
918 bool update_stats = !_g1->evacuation_failed();
919
920 #ifndef PRODUCT
921 if (G1YoungSurvRateVerbose) {
922 gclog_or_tty->print_cr("");
923 _short_lived_surv_rate_group->print();
924 // do that for any other surv rate groups too
925 }
926 #endif // PRODUCT
927
928 last_pause_included_initial_mark = during_initial_mark_pause();
929 if (last_pause_included_initial_mark) {
930 record_concurrent_mark_init_end(0.0);
931 } else if (!_last_young_gc && need_to_start_conc_mark("end of GC")) {
932 // Note: this might have already been set, if during the last
933 // pause we decided to start a cycle but at the beginning of
934 // this pause we decided to postpone it. That's OK.
935 set_initiate_conc_mark_if_possible();
936 }
937
938 _mmu_tracker->add_pause(end_time_sec - pause_time_ms/1000.0,
939 end_time_sec, false);
940
941 size_t freed_bytes =
942 _cur_collection_pause_used_at_start_bytes - cur_used_bytes;
943 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes;
944
945 double survival_fraction =
946 (double)surviving_bytes/
947 (double)_collection_set_bytes_used_before;
948
949 if (update_stats) {
950 _trace_gen0_time_data.record_end_collection(pause_time_ms, phase_times());
951 // this is where we update the allocation rate of the application
952 double app_time_ms =
953 (phase_times()->_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms);
954 if (app_time_ms < MIN_TIMER_GRANULARITY) {
955 // This usually happens due to the timer not having the required
956 // granularity. Some Linuxes are the usual culprits.
957 // We'll just set it to something (arbitrarily) small.
958 app_time_ms = 1.0;
959 }
960 // We maintain the invariant that all objects allocated by mutator
961 // threads will be allocated out of eden regions. So, we can use
962 // the eden region number allocated since the previous GC to
963 // calculate the application's allocate rate. The only exception
964 // to that is humongous objects that are allocated separately. But
965 // given that humongous object allocations do not really affect
966 // either the pause's duration nor when the next pause will take
967 // place we can safely ignore them here.
968 uint regions_allocated = eden_cset_region_length();
969 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
970 _alloc_rate_ms_seq->add(alloc_rate_ms);
971
972 double interval_ms =
973 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
974 update_recent_gc_times(end_time_sec, pause_time_ms);
975 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
976 if (recent_avg_pause_time_ratio() < 0.0 ||
977 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) {
978 #ifndef PRODUCT
979 // Dump info to allow post-facto debugging
980 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds");
981 gclog_or_tty->print_cr("-------------------------------------------");
982 gclog_or_tty->print_cr("Recent GC Times (ms):");
983 _recent_gc_times_ms->dump();
984 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec);
985 _recent_prev_end_times_for_all_gcs_sec->dump();
986 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f",
987 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio());
988 // In debug mode, terminate the JVM if the user wants to debug at this point.
989 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above");
990 #endif // !PRODUCT
991 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
992 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
993 if (_recent_avg_pause_time_ratio < 0.0) {
994 _recent_avg_pause_time_ratio = 0.0;
995 } else {
996 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
997 _recent_avg_pause_time_ratio = 1.0;
998 }
999 }
1000 }
1001 bool new_in_marking_window = _in_marking_window;
1002 bool new_in_marking_window_im = false;
1003 if (during_initial_mark_pause()) {
1004 new_in_marking_window = true;
1005 new_in_marking_window_im = true;
1006 }
1007
1008 if (_last_young_gc) {
1009 // This is supposed to to be the "last young GC" before we start
1010 // doing mixed GCs. Here we decide whether to start mixed GCs or not.
1011
1012 if (!last_pause_included_initial_mark) {
1013 if (next_gc_should_be_mixed("start mixed GCs",
1014 "do not start mixed GCs")) {
1015 set_gcs_are_young(false);
1016 }
1017 } else {
1018 ergo_verbose0(ErgoMixedGCs,
1019 "do not start mixed GCs",
1020 ergo_format_reason("concurrent cycle is about to start"));
1021 }
1022 _last_young_gc = false;
1023 }
1024
1025 if (!_last_gc_was_young) {
1026 // This is a mixed GC. Here we decide whether to continue doing
1027 // mixed GCs or not.
1028
1029 if (!next_gc_should_be_mixed("continue mixed GCs",
1030 "do not continue mixed GCs")) {
1031 set_gcs_are_young(true);
1032 }
1033 }
1034
1035 _short_lived_surv_rate_group->start_adding_regions();
1036 // do that for any other surv rate groupsx
1037
1038 if (update_stats) {
1039 size_t diff = 0;
1040 if (_max_pending_cards >= _pending_cards) {
1041 diff = _max_pending_cards - _pending_cards;
1042 }
1043 _pending_card_diff_seq->add((double) diff);
1044
1045 double cost_per_card_ms = 0.0;
1046 if (_pending_cards > 0) {
1047 cost_per_card_ms = phase_times()->_update_rs_time / (double) _pending_cards;
1048 _cost_per_card_ms_seq->add(cost_per_card_ms);
1049 }
1050
1051 size_t cards_scanned = _g1->cards_scanned();
1052
1053 double cost_per_entry_ms = 0.0;
1054 if (cards_scanned > 10) {
1055 cost_per_entry_ms = phase_times()->_scan_rs_time / (double) cards_scanned;
1056 if (_last_gc_was_young) {
1057 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
1058 } else {
1059 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms);
1060 }
1061 }
1062
1063 if (_max_rs_lengths > 0) {
1064 double cards_per_entry_ratio =
1065 (double) cards_scanned / (double) _max_rs_lengths;
1066 if (_last_gc_was_young) {
1067 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1068 } else {
1069 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1070 }
1071 }
1072
1073 // This is defensive. For a while _max_rs_lengths could get
1074 // smaller than _recorded_rs_lengths which was causing
1075 // rs_length_diff to get very large and mess up the RSet length
1076 // predictions. The reason was unsafe concurrent updates to the
1077 // _inc_cset_recorded_rs_lengths field which the code below guards
1078 // against (see CR 7118202). This bug has now been fixed (see CR
1079 // 7119027). However, I'm still worried that
1080 // _inc_cset_recorded_rs_lengths might still end up somewhat
1081 // inaccurate. The concurrent refinement thread calculates an
1082 // RSet's length concurrently with other CR threads updating it
1083 // which might cause it to calculate the length incorrectly (if,
1084 // say, it's in mid-coarsening). So I'll leave in the defensive
1085 // conditional below just in case.
1086 size_t rs_length_diff = 0;
1087 if (_max_rs_lengths > _recorded_rs_lengths) {
1088 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths;
1089 }
1090 _rs_length_diff_seq->add((double) rs_length_diff);
1091
1092 size_t copied_bytes = surviving_bytes;
1093 double cost_per_byte_ms = 0.0;
1094 if (copied_bytes > 0) {
1095 cost_per_byte_ms = phase_times()->_obj_copy_time / (double) copied_bytes;
1096 if (_in_marking_window) {
1097 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
1098 } else {
1099 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
1100 }
1101 }
1102
1103 double all_other_time_ms = pause_time_ms -
1104 (phase_times()->_update_rs_time + phase_times()->_scan_rs_time + phase_times()->_obj_copy_time + phase_times()->_termination_time);
1105
1106 double young_other_time_ms = 0.0;
1107 if (young_cset_region_length() > 0) {
1108 young_other_time_ms =
1109 phase_times()->_recorded_young_cset_choice_time_ms +
1110 phase_times()->_recorded_young_free_cset_time_ms;
1111 _young_other_cost_per_region_ms_seq->add(young_other_time_ms /
1112 (double) young_cset_region_length());
1113 }
1114 double non_young_other_time_ms = 0.0;
1115 if (old_cset_region_length() > 0) {
1116 non_young_other_time_ms =
1117 phase_times()->_recorded_non_young_cset_choice_time_ms +
1118 phase_times()->_recorded_non_young_free_cset_time_ms;
1119
1120 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms /
1121 (double) old_cset_region_length());
1122 }
1123
1124 double constant_other_time_ms = all_other_time_ms -
1125 (young_other_time_ms + non_young_other_time_ms);
1126 _constant_other_time_ms_seq->add(constant_other_time_ms);
1127
1128 double survival_ratio = 0.0;
1129 if (_bytes_in_collection_set_before_gc > 0) {
1130 survival_ratio = (double) _bytes_copied_during_gc /
1131 (double) _bytes_in_collection_set_before_gc;
1132 }
1133
1134 _pending_cards_seq->add((double) _pending_cards);
1135 _rs_lengths_seq->add((double) _max_rs_lengths);
1136 }
1137
1138 _in_marking_window = new_in_marking_window;
1139 _in_marking_window_im = new_in_marking_window_im;
1140 _free_regions_at_end_of_collection = _g1->free_regions();
1141 update_young_list_target_length();
1142
1143 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
1144 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
1145 adjust_concurrent_refinement(phase_times()->_update_rs_time, phase_times()->_update_rs_processed_buffers, update_rs_time_goal_ms);
1146
1147 _collectionSetChooser->verify();
1148 }
1149
1150 #define EXT_SIZE_FORMAT "%.1f%s"
1151 #define EXT_SIZE_PARAMS(bytes) \
1152 byte_size_in_proper_unit((double)(bytes)), \
1153 proper_unit_for_byte_size((bytes))
1154
1155 void G1CollectorPolicy::print_heap_transition() {
1156 if (G1Log::finer()) {
1157 YoungList* young_list = _g1->young_list();
1158 size_t eden_bytes = young_list->eden_used_bytes();
1159 size_t survivor_bytes = young_list->survivor_used_bytes();
1160 size_t used_before_gc = _cur_collection_pause_used_at_start_bytes;
1161 size_t used = _g1->used();
1162 size_t capacity = _g1->capacity();
1163 size_t eden_capacity =
1164 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_bytes;
1165
1166 gclog_or_tty->print_cr(
1167 " [Eden: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT") "
1168 "Survivors: "EXT_SIZE_FORMAT"->"EXT_SIZE_FORMAT" "
1169 "Heap: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"
1170 EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")]",
1171 EXT_SIZE_PARAMS(_eden_bytes_before_gc),
1172 EXT_SIZE_PARAMS(_prev_eden_capacity),
1173 EXT_SIZE_PARAMS(eden_bytes),
1174 EXT_SIZE_PARAMS(eden_capacity),
1175 EXT_SIZE_PARAMS(_survivor_bytes_before_gc),
1176 EXT_SIZE_PARAMS(survivor_bytes),
1177 EXT_SIZE_PARAMS(used_before_gc),
1178 EXT_SIZE_PARAMS(_capacity_before_gc),
1179 EXT_SIZE_PARAMS(used),
1180 EXT_SIZE_PARAMS(capacity));
1181
1182 _prev_eden_capacity = eden_capacity;
1183 } else if (G1Log::fine()) {
1184 _g1->print_size_transition(gclog_or_tty,
1185 _cur_collection_pause_used_at_start_bytes,
1186 _g1->used(), _g1->capacity());
1187 }
1188 }
1189
1190 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1191 double update_rs_processed_buffers,
1192 double goal_ms) {
1193 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1194 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1195
1196 if (G1UseAdaptiveConcRefinement) {
1197 const int k_gy = 3, k_gr = 6;
1198 const double inc_k = 1.1, dec_k = 0.9;
1199
1200 int g = cg1r->green_zone();
1201 if (update_rs_time > goal_ms) {
1202 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1203 } else {
1204 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1205 g = (int)MAX2(g * inc_k, g + 1.0);
1206 }
1207 }
1208 // Change the refinement threads params
1209 cg1r->set_green_zone(g);
1210 cg1r->set_yellow_zone(g * k_gy);
1211 cg1r->set_red_zone(g * k_gr);
1212 cg1r->reinitialize_threads();
1213
1214 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1);
1215 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1216 cg1r->yellow_zone());
1217 // Change the barrier params
1218 dcqs.set_process_completed_threshold(processing_threshold);
1219 dcqs.set_max_completed_queue(cg1r->red_zone());
1220 }
1221
1222 int curr_queue_size = dcqs.completed_buffers_num();
1223 if (curr_queue_size >= cg1r->yellow_zone()) {
1224 dcqs.set_completed_queue_padding(curr_queue_size);
1225 } else {
1226 dcqs.set_completed_queue_padding(0);
1227 }
1228 dcqs.notify_if_necessary();
1229 }
1230
1231 double
1232 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) {
1233 size_t rs_length = predict_rs_length_diff();
1234 size_t card_num;
1235 if (gcs_are_young()) {
1236 card_num = predict_young_card_num(rs_length);
1237 } else {
1238 card_num = predict_non_young_card_num(rs_length);
1239 }
1240 return predict_base_elapsed_time_ms(pending_cards, card_num);
1241 }
1242
1243 double
1244 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
1245 size_t scanned_cards) {
1246 return
1247 predict_rs_update_time_ms(pending_cards) +
1248 predict_rs_scan_time_ms(scanned_cards) +
1249 predict_constant_other_time_ms();
1250 }
1251
1252 double
1253 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
1254 bool young) {
1255 size_t rs_length = hr->rem_set()->occupied();
1256 size_t card_num;
1257 if (gcs_are_young()) {
1258 card_num = predict_young_card_num(rs_length);
1259 } else {
1260 card_num = predict_non_young_card_num(rs_length);
1261 }
1262 size_t bytes_to_copy = predict_bytes_to_copy(hr);
1263
1264 double region_elapsed_time_ms =
1265 predict_rs_scan_time_ms(card_num) +
1266 predict_object_copy_time_ms(bytes_to_copy);
1267
1268 if (young)
1269 region_elapsed_time_ms += predict_young_other_time_ms(1);
1270 else
1271 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
1272
1273 return region_elapsed_time_ms;
1274 }
1275
1276 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) {
1277 size_t bytes_to_copy;
1278 if (hr->is_marked())
1279 bytes_to_copy = hr->max_live_bytes();
1280 else {
1281 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant");
1282 int age = hr->age_in_surv_rate_group();
1283 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
1284 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate);
1285 }
1286 return bytes_to_copy;
1287 }
1288
1289 void
1290 G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length,
1291 uint survivor_cset_region_length) {
1292 _eden_cset_region_length = eden_cset_region_length;
1293 _survivor_cset_region_length = survivor_cset_region_length;
1294 _old_cset_region_length = 0;
1295 }
1296
1297 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) {
1298 _recorded_rs_lengths = rs_lengths;
1299 }
1300
1301 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
1302 double elapsed_ms) {
1303 _recent_gc_times_ms->add(elapsed_ms);
1304 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
1305 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
1306 }
1307
1308 size_t G1CollectorPolicy::expansion_amount() {
1309 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0;
1310 double threshold = _gc_overhead_perc;
1311 if (recent_gc_overhead > threshold) {
1312 // We will double the existing space, or take
1313 // G1ExpandByPercentOfAvailable % of the available expansion
1314 // space, whichever is smaller, bounded below by a minimum
1315 // expansion (unless that's all that's left.)
1316 const size_t min_expand_bytes = 1*M;
1317 size_t reserved_bytes = _g1->max_capacity();
1318 size_t committed_bytes = _g1->capacity();
1319 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
1320 size_t expand_bytes;
1321 size_t expand_bytes_via_pct =
1322 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
1323 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
1324 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
1325 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
1326
1327 ergo_verbose5(ErgoHeapSizing,
1328 "attempt heap expansion",
1329 ergo_format_reason("recent GC overhead higher than "
1330 "threshold after GC")
1331 ergo_format_perc("recent GC overhead")
1332 ergo_format_perc("threshold")
1333 ergo_format_byte("uncommitted")
1334 ergo_format_byte_perc("calculated expansion amount"),
1335 recent_gc_overhead, threshold,
1336 uncommitted_bytes,
1337 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable);
1338
1339 return expand_bytes;
1340 } else {
1341 return 0;
1342 }
1343 }
1344
1345 class CountCSClosure: public HeapRegionClosure {
1346 G1CollectorPolicy* _g1_policy;
1347 public:
1348 CountCSClosure(G1CollectorPolicy* g1_policy) :
1349 _g1_policy(g1_policy) {}
1350 bool doHeapRegion(HeapRegion* r) {
1351 _g1_policy->_bytes_in_collection_set_before_gc += r->used();
1352 return false;
1353 }
1354 };
1355
1356 void G1CollectorPolicy::count_CS_bytes_used() {
1357 CountCSClosure cs_closure(this);
1358 _g1->collection_set_iterate(&cs_closure);
1359 }
1360
1361 void G1CollectorPolicy::print_tracing_info() const {
1362 _trace_gen0_time_data.print();
1363 _trace_gen1_time_data.print();
1364 }
1365
1366 void G1CollectorPolicy::print_yg_surv_rate_info() const {
1367 #ifndef PRODUCT
1368 _short_lived_surv_rate_group->print_surv_rate_summary();
1369 // add this call for any other surv rate groups
1370 #endif // PRODUCT
1371 }
1372
1373 #ifndef PRODUCT
1374 // for debugging, bit of a hack...
1375 static char*
1376 region_num_to_mbs(int length) {
1377 static char buffer[64];
1378 double bytes = (double) (length * HeapRegion::GrainBytes);
1379 double mbs = bytes / (double) (1024 * 1024);
1380 sprintf(buffer, "%7.2lfMB", mbs);
1381 return buffer;
1382 }
1383 #endif // PRODUCT
1384
1385 uint G1CollectorPolicy::max_regions(int purpose) {
1386 switch (purpose) {
1387 case GCAllocForSurvived:
1388 return _max_survivor_regions;
1389 case GCAllocForTenured:
1390 return REGIONS_UNLIMITED;
1391 default:
1392 ShouldNotReachHere();
1393 return REGIONS_UNLIMITED;
1394 };
1395 }
1396
1397 void G1CollectorPolicy::update_max_gc_locker_expansion() {
1398 uint expansion_region_num = 0;
1399 if (GCLockerEdenExpansionPercent > 0) {
1400 double perc = (double) GCLockerEdenExpansionPercent / 100.0;
1401 double expansion_region_num_d = perc * (double) _young_list_target_length;
1402 // We use ceiling so that if expansion_region_num_d is > 0.0 (but
1403 // less than 1.0) we'll get 1.
1404 expansion_region_num = (uint) ceil(expansion_region_num_d);
1405 } else {
1406 assert(expansion_region_num == 0, "sanity");
1407 }
1408 _young_list_max_length = _young_list_target_length + expansion_region_num;
1409 assert(_young_list_target_length <= _young_list_max_length, "post-condition");
1410 }
1411
1412 // Calculates survivor space parameters.
1413 void G1CollectorPolicy::update_survivors_policy() {
1414 double max_survivor_regions_d =
1415 (double) _young_list_target_length / (double) SurvivorRatio;
1416 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but
1417 // smaller than 1.0) we'll get 1.
1418 _max_survivor_regions = (uint) ceil(max_survivor_regions_d);
1419
1420 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
1421 HeapRegion::GrainWords * _max_survivor_regions);
1422 }
1423
1424 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(
1425 GCCause::Cause gc_cause) {
1426 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
1427 if (!during_cycle) {
1428 ergo_verbose1(ErgoConcCycles,
1429 "request concurrent cycle initiation",
1430 ergo_format_reason("requested by GC cause")
1431 ergo_format_str("GC cause"),
1432 GCCause::to_string(gc_cause));
1433 set_initiate_conc_mark_if_possible();
1434 return true;
1435 } else {
1436 ergo_verbose1(ErgoConcCycles,
1437 "do not request concurrent cycle initiation",
1438 ergo_format_reason("concurrent cycle already in progress")
1439 ergo_format_str("GC cause"),
1440 GCCause::to_string(gc_cause));
1441 return false;
1442 }
1443 }
1444
1445 void
1446 G1CollectorPolicy::decide_on_conc_mark_initiation() {
1447 // We are about to decide on whether this pause will be an
1448 // initial-mark pause.
1449
1450 // First, during_initial_mark_pause() should not be already set. We
1451 // will set it here if we have to. However, it should be cleared by
1452 // the end of the pause (it's only set for the duration of an
1453 // initial-mark pause).
1454 assert(!during_initial_mark_pause(), "pre-condition");
1455
1456 if (initiate_conc_mark_if_possible()) {
1457 // We had noticed on a previous pause that the heap occupancy has
1458 // gone over the initiating threshold and we should start a
1459 // concurrent marking cycle. So we might initiate one.
1460
1461 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
1462 if (!during_cycle) {
1463 // The concurrent marking thread is not "during a cycle", i.e.,
1464 // it has completed the last one. So we can go ahead and
1465 // initiate a new cycle.
1466
1467 set_during_initial_mark_pause();
1468 // We do not allow mixed GCs during marking.
1469 if (!gcs_are_young()) {
1470 set_gcs_are_young(true);
1471 ergo_verbose0(ErgoMixedGCs,
1472 "end mixed GCs",
1473 ergo_format_reason("concurrent cycle is about to start"));
1474 }
1475
1476 // And we can now clear initiate_conc_mark_if_possible() as
1477 // we've already acted on it.
1478 clear_initiate_conc_mark_if_possible();
1479
1480 ergo_verbose0(ErgoConcCycles,
1481 "initiate concurrent cycle",
1482 ergo_format_reason("concurrent cycle initiation requested"));
1483 } else {
1484 // The concurrent marking thread is still finishing up the
1485 // previous cycle. If we start one right now the two cycles
1486 // overlap. In particular, the concurrent marking thread might
1487 // be in the process of clearing the next marking bitmap (which
1488 // we will use for the next cycle if we start one). Starting a
1489 // cycle now will be bad given that parts of the marking
1490 // information might get cleared by the marking thread. And we
1491 // cannot wait for the marking thread to finish the cycle as it
1492 // periodically yields while clearing the next marking bitmap
1493 // and, if it's in a yield point, it's waiting for us to
1494 // finish. So, at this point we will not start a cycle and we'll
1495 // let the concurrent marking thread complete the last one.
1496 ergo_verbose0(ErgoConcCycles,
1497 "do not initiate concurrent cycle",
1498 ergo_format_reason("concurrent cycle already in progress"));
1499 }
1500 }
1501 }
1502
1503 class KnownGarbageClosure: public HeapRegionClosure {
1504 G1CollectedHeap* _g1h;
1505 CollectionSetChooser* _hrSorted;
1506
1507 public:
1508 KnownGarbageClosure(CollectionSetChooser* hrSorted) :
1509 _g1h(G1CollectedHeap::heap()), _hrSorted(hrSorted) { }
1510
1511 bool doHeapRegion(HeapRegion* r) {
1512 // We only include humongous regions in collection
1513 // sets when concurrent mark shows that their contained object is
1514 // unreachable.
1515
1516 // Do we have any marking information for this region?
1517 if (r->is_marked()) {
1518 // We will skip any region that's currently used as an old GC
1519 // alloc region (we should not consider those for collection
1520 // before we fill them up).
1521 if (_hrSorted->should_add(r) && !_g1h->is_old_gc_alloc_region(r)) {
1522 _hrSorted->add_region(r);
1523 }
1524 }
1525 return false;
1526 }
1527 };
1528
1529 class ParKnownGarbageHRClosure: public HeapRegionClosure {
1530 G1CollectedHeap* _g1h;
1531 CollectionSetChooser* _hrSorted;
1532 uint _marked_regions_added;
1533 size_t _reclaimable_bytes_added;
1534 uint _chunk_size;
1535 uint _cur_chunk_idx;
1536 uint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end)
1537
1538 void get_new_chunk() {
1539 _cur_chunk_idx = _hrSorted->claim_array_chunk(_chunk_size);
1540 _cur_chunk_end = _cur_chunk_idx + _chunk_size;
1541 }
1542 void add_region(HeapRegion* r) {
1543 if (_cur_chunk_idx == _cur_chunk_end) {
1544 get_new_chunk();
1545 }
1546 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition");
1547 _hrSorted->set_region(_cur_chunk_idx, r);
1548 _marked_regions_added++;
1549 _reclaimable_bytes_added += r->reclaimable_bytes();
1550 _cur_chunk_idx++;
1551 }
1552
1553 public:
1554 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
1555 uint chunk_size) :
1556 _g1h(G1CollectedHeap::heap()),
1557 _hrSorted(hrSorted), _chunk_size(chunk_size),
1558 _marked_regions_added(0), _reclaimable_bytes_added(0),
1559 _cur_chunk_idx(0), _cur_chunk_end(0) { }
1560
1561 bool doHeapRegion(HeapRegion* r) {
1562 // Do we have any marking information for this region?
1563 if (r->is_marked()) {
1564 // We will skip any region that's currently used as an old GC
1565 // alloc region (we should not consider those for collection
1566 // before we fill them up).
1567 if (_hrSorted->should_add(r) && !_g1h->is_old_gc_alloc_region(r)) {
1568 add_region(r);
1569 }
1570 }
1571 return false;
1572 }
1573 uint marked_regions_added() { return _marked_regions_added; }
1574 size_t reclaimable_bytes_added() { return _reclaimable_bytes_added; }
1575 };
1576
1577 class ParKnownGarbageTask: public AbstractGangTask {
1578 CollectionSetChooser* _hrSorted;
1579 uint _chunk_size;
1580 G1CollectedHeap* _g1;
1581 public:
1582 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size) :
1583 AbstractGangTask("ParKnownGarbageTask"),
1584 _hrSorted(hrSorted), _chunk_size(chunk_size),
1585 _g1(G1CollectedHeap::heap()) { }
1586
1587 void work(uint worker_id) {
1588 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size);
1589
1590 // Back to zero for the claim value.
1591 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, worker_id,
1592 _g1->workers()->active_workers(),
1593 HeapRegion::InitialClaimValue);
1594 uint regions_added = parKnownGarbageCl.marked_regions_added();
1595 size_t reclaimable_bytes_added =
1596 parKnownGarbageCl.reclaimable_bytes_added();
1597 _hrSorted->update_totals(regions_added, reclaimable_bytes_added);
1598 }
1599 };
1600
1601 void
1602 G1CollectorPolicy::record_concurrent_mark_cleanup_end(int no_of_gc_threads) {
1603 _collectionSetChooser->clear();
1604
1605 uint region_num = _g1->n_regions();
1606 if (G1CollectedHeap::use_parallel_gc_threads()) {
1607 const uint OverpartitionFactor = 4;
1608 uint WorkUnit;
1609 // The use of MinChunkSize = 8 in the original code
1610 // causes some assertion failures when the total number of
1611 // region is less than 8. The code here tries to fix that.
1612 // Should the original code also be fixed?
1613 if (no_of_gc_threads > 0) {
1614 const uint MinWorkUnit = MAX2(region_num / no_of_gc_threads, 1U);
1615 WorkUnit = MAX2(region_num / (no_of_gc_threads * OverpartitionFactor),
1616 MinWorkUnit);
1617 } else {
1618 assert(no_of_gc_threads > 0,
1619 "The active gc workers should be greater than 0");
1620 // In a product build do something reasonable to avoid a crash.
1621 const uint MinWorkUnit = MAX2(region_num / (uint) ParallelGCThreads, 1U);
1622 WorkUnit =
1623 MAX2(region_num / (uint) (ParallelGCThreads * OverpartitionFactor),
1624 MinWorkUnit);
1625 }
1626 _collectionSetChooser->prepare_for_par_region_addition(_g1->n_regions(),
1627 WorkUnit);
1628 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser,
1629 (int) WorkUnit);
1630 _g1->workers()->run_task(&parKnownGarbageTask);
1631
1632 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
1633 "sanity check");
1634 } else {
1635 KnownGarbageClosure knownGarbagecl(_collectionSetChooser);
1636 _g1->heap_region_iterate(&knownGarbagecl);
1637 }
1638
1639 _collectionSetChooser->sort_regions();
1640
1641 double end_sec = os::elapsedTime();
1642 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0;
1643 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
1644 _cur_mark_stop_world_time_ms += elapsed_time_ms;
1645 _prev_collection_pause_end_ms += elapsed_time_ms;
1646 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec, true);
1647 }
1648
1649 // Add the heap region at the head of the non-incremental collection set
1650 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) {
1651 assert(_inc_cset_build_state == Active, "Precondition");
1652 assert(!hr->is_young(), "non-incremental add of young region");
1653
1654 assert(!hr->in_collection_set(), "should not already be in the CSet");
1655 hr->set_in_collection_set(true);
1656 hr->set_next_in_collection_set(_collection_set);
1657 _collection_set = hr;
1658 _collection_set_bytes_used_before += hr->used();
1659 _g1->register_region_with_in_cset_fast_test(hr);
1660 size_t rs_length = hr->rem_set()->occupied();
1661 _recorded_rs_lengths += rs_length;
1662 _old_cset_region_length += 1;
1663 }
1664
1665 // Initialize the per-collection-set information
1666 void G1CollectorPolicy::start_incremental_cset_building() {
1667 assert(_inc_cset_build_state == Inactive, "Precondition");
1668
1669 _inc_cset_head = NULL;
1670 _inc_cset_tail = NULL;
1671 _inc_cset_bytes_used_before = 0;
1672
1673 _inc_cset_max_finger = 0;
1674 _inc_cset_recorded_rs_lengths = 0;
1675 _inc_cset_recorded_rs_lengths_diffs = 0;
1676 _inc_cset_predicted_elapsed_time_ms = 0.0;
1677 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
1678 _inc_cset_build_state = Active;
1679 }
1680
1681 void G1CollectorPolicy::finalize_incremental_cset_building() {
1682 assert(_inc_cset_build_state == Active, "Precondition");
1683 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
1684
1685 // The two "main" fields, _inc_cset_recorded_rs_lengths and
1686 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread
1687 // that adds a new region to the CSet. Further updates by the
1688 // concurrent refinement thread that samples the young RSet lengths
1689 // are accumulated in the *_diffs fields. Here we add the diffs to
1690 // the "main" fields.
1691
1692 if (_inc_cset_recorded_rs_lengths_diffs >= 0) {
1693 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs;
1694 } else {
1695 // This is defensive. The diff should in theory be always positive
1696 // as RSets can only grow between GCs. However, given that we
1697 // sample their size concurrently with other threads updating them
1698 // it's possible that we might get the wrong size back, which
1699 // could make the calculations somewhat inaccurate.
1700 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs);
1701 if (_inc_cset_recorded_rs_lengths >= diffs) {
1702 _inc_cset_recorded_rs_lengths -= diffs;
1703 } else {
1704 _inc_cset_recorded_rs_lengths = 0;
1705 }
1706 }
1707 _inc_cset_predicted_elapsed_time_ms +=
1708 _inc_cset_predicted_elapsed_time_ms_diffs;
1709
1710 _inc_cset_recorded_rs_lengths_diffs = 0;
1711 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
1712 }
1713
1714 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) {
1715 // This routine is used when:
1716 // * adding survivor regions to the incremental cset at the end of an
1717 // evacuation pause,
1718 // * adding the current allocation region to the incremental cset
1719 // when it is retired, and
1720 // * updating existing policy information for a region in the
1721 // incremental cset via young list RSet sampling.
1722 // Therefore this routine may be called at a safepoint by the
1723 // VM thread, or in-between safepoints by mutator threads (when
1724 // retiring the current allocation region) or a concurrent
1725 // refine thread (RSet sampling).
1726
1727 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true);
1728 size_t used_bytes = hr->used();
1729 _inc_cset_recorded_rs_lengths += rs_length;
1730 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms;
1731 _inc_cset_bytes_used_before += used_bytes;
1732
1733 // Cache the values we have added to the aggregated informtion
1734 // in the heap region in case we have to remove this region from
1735 // the incremental collection set, or it is updated by the
1736 // rset sampling code
1737 hr->set_recorded_rs_length(rs_length);
1738 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms);
1739 }
1740
1741 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr,
1742 size_t new_rs_length) {
1743 // Update the CSet information that is dependent on the new RS length
1744 assert(hr->is_young(), "Precondition");
1745 assert(!SafepointSynchronize::is_at_safepoint(),
1746 "should not be at a safepoint");
1747
1748 // We could have updated _inc_cset_recorded_rs_lengths and
1749 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do
1750 // that atomically, as this code is executed by a concurrent
1751 // refinement thread, potentially concurrently with a mutator thread
1752 // allocating a new region and also updating the same fields. To
1753 // avoid the atomic operations we accumulate these updates on two
1754 // separate fields (*_diffs) and we'll just add them to the "main"
1755 // fields at the start of a GC.
1756
1757 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length();
1758 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length;
1759 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff;
1760
1761 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms();
1762 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true);
1763 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms;
1764 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff;
1765
1766 hr->set_recorded_rs_length(new_rs_length);
1767 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms);
1768 }
1769
1770 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) {
1771 assert(hr->is_young(), "invariant");
1772 assert(hr->young_index_in_cset() > -1, "should have already been set");
1773 assert(_inc_cset_build_state == Active, "Precondition");
1774
1775 // We need to clear and set the cached recorded/cached collection set
1776 // information in the heap region here (before the region gets added
1777 // to the collection set). An individual heap region's cached values
1778 // are calculated, aggregated with the policy collection set info,
1779 // and cached in the heap region here (initially) and (subsequently)
1780 // by the Young List sampling code.
1781
1782 size_t rs_length = hr->rem_set()->occupied();
1783 add_to_incremental_cset_info(hr, rs_length);
1784
1785 HeapWord* hr_end = hr->end();
1786 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end);
1787
1788 assert(!hr->in_collection_set(), "invariant");
1789 hr->set_in_collection_set(true);
1790 assert( hr->next_in_collection_set() == NULL, "invariant");
1791
1792 _g1->register_region_with_in_cset_fast_test(hr);
1793 }
1794
1795 // Add the region at the RHS of the incremental cset
1796 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) {
1797 // We should only ever be appending survivors at the end of a pause
1798 assert( hr->is_survivor(), "Logic");
1799
1800 // Do the 'common' stuff
1801 add_region_to_incremental_cset_common(hr);
1802
1803 // Now add the region at the right hand side
1804 if (_inc_cset_tail == NULL) {
1805 assert(_inc_cset_head == NULL, "invariant");
1806 _inc_cset_head = hr;
1807 } else {
1808 _inc_cset_tail->set_next_in_collection_set(hr);
1809 }
1810 _inc_cset_tail = hr;
1811 }
1812
1813 // Add the region to the LHS of the incremental cset
1814 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) {
1815 // Survivors should be added to the RHS at the end of a pause
1816 assert(!hr->is_survivor(), "Logic");
1817
1818 // Do the 'common' stuff
1819 add_region_to_incremental_cset_common(hr);
1820
1821 // Add the region at the left hand side
1822 hr->set_next_in_collection_set(_inc_cset_head);
1823 if (_inc_cset_head == NULL) {
1824 assert(_inc_cset_tail == NULL, "Invariant");
1825 _inc_cset_tail = hr;
1826 }
1827 _inc_cset_head = hr;
1828 }
1829
1830 #ifndef PRODUCT
1831 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) {
1832 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be");
1833
1834 st->print_cr("\nCollection_set:");
1835 HeapRegion* csr = list_head;
1836 while (csr != NULL) {
1837 HeapRegion* next = csr->next_in_collection_set();
1838 assert(csr->in_collection_set(), "bad CS");
1839 st->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d",
1840 HR_FORMAT_PARAMS(csr),
1841 csr->prev_top_at_mark_start(), csr->next_top_at_mark_start(),
1842 csr->age_in_surv_rate_group_cond());
1843 csr = next;
1844 }
1845 }
1846 #endif // !PRODUCT
1847
1848 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str,
1849 const char* false_action_str) {
1850 CollectionSetChooser* cset_chooser = _collectionSetChooser;
1851 if (cset_chooser->is_empty()) {
1852 ergo_verbose0(ErgoMixedGCs,
1853 false_action_str,
1854 ergo_format_reason("candidate old regions not available"));
1855 return false;
1856 }
1857 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes();
1858 size_t capacity_bytes = _g1->capacity();
1859 double perc = (double) reclaimable_bytes * 100.0 / (double) capacity_bytes;
1860 double threshold = (double) G1HeapWastePercent;
1861 if (perc < threshold) {
1862 ergo_verbose4(ErgoMixedGCs,
1863 false_action_str,
1864 ergo_format_reason("reclaimable percentage lower than threshold")
1865 ergo_format_region("candidate old regions")
1866 ergo_format_byte_perc("reclaimable")
1867 ergo_format_perc("threshold"),
1868 cset_chooser->remaining_regions(),
1869 reclaimable_bytes, perc, threshold);
1870 return false;
1871 }
1872
1873 ergo_verbose4(ErgoMixedGCs,
1874 true_action_str,
1875 ergo_format_reason("candidate old regions available")
1876 ergo_format_region("candidate old regions")
1877 ergo_format_byte_perc("reclaimable")
1878 ergo_format_perc("threshold"),
1879 cset_chooser->remaining_regions(),
1880 reclaimable_bytes, perc, threshold);
1881 return true;
1882 }
1883
1884 void G1CollectorPolicy::finalize_cset(double target_pause_time_ms) {
1885 // Set this here - in case we're not doing young collections.
1886 double non_young_start_time_sec = os::elapsedTime();
1887
1888 YoungList* young_list = _g1->young_list();
1889 finalize_incremental_cset_building();
1890
1891 guarantee(target_pause_time_ms > 0.0,
1892 err_msg("target_pause_time_ms = %1.6lf should be positive",
1893 target_pause_time_ms));
1894 guarantee(_collection_set == NULL, "Precondition");
1895
1896 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
1897 double predicted_pause_time_ms = base_time_ms;
1898 double time_remaining_ms = target_pause_time_ms - base_time_ms;
1899
1900 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
1901 "start choosing CSet",
1902 ergo_format_ms("predicted base time")
1903 ergo_format_ms("remaining time")
1904 ergo_format_ms("target pause time"),
1905 base_time_ms, time_remaining_ms, target_pause_time_ms);
1906
1907 HeapRegion* hr;
1908 double young_start_time_sec = os::elapsedTime();
1909
1910 _collection_set_bytes_used_before = 0;
1911 _last_gc_was_young = gcs_are_young() ? true : false;
1912
1913 if (_last_gc_was_young) {
1914 _trace_gen0_time_data.increment_young_collection_count();
1915 } else {
1916 _trace_gen0_time_data.increment_mixed_collection_count();
1917 }
1918
1919 // The young list is laid with the survivor regions from the previous
1920 // pause are appended to the RHS of the young list, i.e.
1921 // [Newly Young Regions ++ Survivors from last pause].
1922
1923 uint survivor_region_length = young_list->survivor_length();
1924 uint eden_region_length = young_list->length() - survivor_region_length;
1925 init_cset_region_lengths(eden_region_length, survivor_region_length);
1926 hr = young_list->first_survivor_region();
1927 while (hr != NULL) {
1928 assert(hr->is_survivor(), "badly formed young list");
1929 hr->set_young();
1930 hr = hr->get_next_young_region();
1931 }
1932
1933 // Clear the fields that point to the survivor list - they are all young now.
1934 young_list->clear_survivors();
1935
1936 _collection_set = _inc_cset_head;
1937 _collection_set_bytes_used_before = _inc_cset_bytes_used_before;
1938 time_remaining_ms -= _inc_cset_predicted_elapsed_time_ms;
1939 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms;
1940
1941 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
1942 "add young regions to CSet",
1943 ergo_format_region("eden")
1944 ergo_format_region("survivors")
1945 ergo_format_ms("predicted young region time"),
1946 eden_region_length, survivor_region_length,
1947 _inc_cset_predicted_elapsed_time_ms);
1948
1949 // The number of recorded young regions is the incremental
1950 // collection set's current size
1951 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths);
1952
1953 double young_end_time_sec = os::elapsedTime();
1954 phase_times()->_recorded_young_cset_choice_time_ms =
1955 (young_end_time_sec - young_start_time_sec) * 1000.0;
1956
1957 // We are doing young collections so reset this.
1958 non_young_start_time_sec = young_end_time_sec;
1959
1960 if (!gcs_are_young()) {
1961 CollectionSetChooser* cset_chooser = _collectionSetChooser;
1962 cset_chooser->verify();
1963 const uint min_old_cset_length = cset_chooser->calc_min_old_cset_length();
1964 const uint max_old_cset_length = cset_chooser->calc_max_old_cset_length();
1965
1966 uint expensive_region_num = 0;
1967 bool check_time_remaining = adaptive_young_list_length();
1968 HeapRegion* hr = cset_chooser->peek();
1969 while (hr != NULL) {
1970 if (old_cset_region_length() >= max_old_cset_length) {
1971 // Added maximum number of old regions to the CSet.
1972 ergo_verbose2(ErgoCSetConstruction,
1973 "finish adding old regions to CSet",
1974 ergo_format_reason("old CSet region num reached max")
1975 ergo_format_region("old")
1976 ergo_format_region("max"),
1977 old_cset_region_length(), max_old_cset_length);
1978 break;
1979 }
1980
1981 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false);
1982 if (check_time_remaining) {
1983 if (predicted_time_ms > time_remaining_ms) {
1984 // Too expensive for the current CSet.
1985
1986 if (old_cset_region_length() >= min_old_cset_length) {
1987 // We have added the minimum number of old regions to the CSet,
1988 // we are done with this CSet.
1989 ergo_verbose4(ErgoCSetConstruction,
1990 "finish adding old regions to CSet",
1991 ergo_format_reason("predicted time is too high")
1992 ergo_format_ms("predicted time")
1993 ergo_format_ms("remaining time")
1994 ergo_format_region("old")
1995 ergo_format_region("min"),
1996 predicted_time_ms, time_remaining_ms,
1997 old_cset_region_length(), min_old_cset_length);
1998 break;
1999 }
2000
2001 // We'll add it anyway given that we haven't reached the
2002 // minimum number of old regions.
2003 expensive_region_num += 1;
2004 }
2005 } else {
2006 if (old_cset_region_length() >= min_old_cset_length) {
2007 // In the non-auto-tuning case, we'll finish adding regions
2008 // to the CSet if we reach the minimum.
2009 ergo_verbose2(ErgoCSetConstruction,
2010 "finish adding old regions to CSet",
2011 ergo_format_reason("old CSet region num reached min")
2012 ergo_format_region("old")
2013 ergo_format_region("min"),
2014 old_cset_region_length(), min_old_cset_length);
2015 break;
2016 }
2017 }
2018
2019 // We will add this region to the CSet.
2020 time_remaining_ms -= predicted_time_ms;
2021 predicted_pause_time_ms += predicted_time_ms;
2022 cset_chooser->remove_and_move_to_next(hr);
2023 _g1->old_set_remove(hr);
2024 add_old_region_to_cset(hr);
2025
2026 hr = cset_chooser->peek();
2027 }
2028 if (hr == NULL) {
2029 ergo_verbose0(ErgoCSetConstruction,
2030 "finish adding old regions to CSet",
2031 ergo_format_reason("candidate old regions not available"));
2032 }
2033
2034 if (expensive_region_num > 0) {
2035 // We print the information once here at the end, predicated on
2036 // whether we added any apparently expensive regions or not, to
2037 // avoid generating output per region.
2038 ergo_verbose4(ErgoCSetConstruction,
2039 "added expensive regions to CSet",
2040 ergo_format_reason("old CSet region num not reached min")
2041 ergo_format_region("old")
2042 ergo_format_region("expensive")
2043 ergo_format_region("min")
2044 ergo_format_ms("remaining time"),
2045 old_cset_region_length(),
2046 expensive_region_num,
2047 min_old_cset_length,
2048 time_remaining_ms);
2049 }
2050
2051 cset_chooser->verify();
2052 }
2053
2054 stop_incremental_cset_building();
2055
2056 count_CS_bytes_used();
2057
2058 ergo_verbose5(ErgoCSetConstruction,
2059 "finish choosing CSet",
2060 ergo_format_region("eden")
2061 ergo_format_region("survivors")
2062 ergo_format_region("old")
2063 ergo_format_ms("predicted pause time")
2064 ergo_format_ms("target pause time"),
2065 eden_region_length, survivor_region_length,
2066 old_cset_region_length(),
2067 predicted_pause_time_ms, target_pause_time_ms);
2068
2069 double non_young_end_time_sec = os::elapsedTime();
2070 phase_times()->_recorded_non_young_cset_choice_time_ms =
2071 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0;
2072 }
2073
2074 void TraceGen0TimeData::record_start_collection(double time_to_stop_the_world_ms) {
2075 if(TraceGen0Time) {
2076 _all_stop_world_times_ms.add(time_to_stop_the_world_ms);
2077 }
2078 }
2079
2080 void TraceGen0TimeData::record_yield_time(double yield_time_ms) {
2081 if(TraceGen0Time) {
2082 _all_yield_times_ms.add(yield_time_ms);
2083 }
2084 }
2085
2086 void TraceGen0TimeData::record_end_collection(double pause_time_ms, G1GCPhaseTimes* phase_times) {
2087 if(TraceGen0Time) {
2088 _total.add(pause_time_ms);
2089 _other.add(pause_time_ms - phase_times->accounted_time_ms());
2090 _root_region_scan_wait.add(phase_times->_root_region_scan_wait_time_ms);
2091 _parallel.add(phase_times->_cur_collection_par_time_ms);
2092 _ext_root_scan.add(phase_times->_ext_root_scan_time);
2093 _satb_filtering.add(phase_times->_satb_filtering_time);
2094 _update_rs.add(phase_times->_update_rs_time);
2095 _scan_rs.add(phase_times->_scan_rs_time);
2096 _obj_copy.add(phase_times->_obj_copy_time);
2097 _termination.add(phase_times->_termination_time);
2098
2099 double parallel_known_time = phase_times->_ext_root_scan_time +
2100 phase_times->_satb_filtering_time +
2101 phase_times->_update_rs_time +
2102 phase_times->_scan_rs_time +
2103 phase_times->_obj_copy_time +
2104 + phase_times->_termination_time;
2105
2106 double parallel_other_time = phase_times->_cur_collection_par_time_ms - parallel_known_time;
2107 _parallel_other.add(parallel_other_time);
2108 _clear_ct.add(phase_times->_cur_clear_ct_time_ms);
2109 }
2110 }
2111
2112 void TraceGen0TimeData::increment_young_collection_count() {
2113 if(TraceGen0Time) {
2114 ++_young_pause_num;
2115 }
2116 }
2117
2118 void TraceGen0TimeData::increment_mixed_collection_count() {
2119 if(TraceGen0Time) {
2120 ++_mixed_pause_num;
2121 }
2122 }
2123
2124 void TraceGen0TimeData::print_summary(const char* str,
2125 const NumberSeq* seq) const {
2126 double sum = seq->sum();
2127 gclog_or_tty->print_cr("%-27s = %8.2lf s (avg = %8.2lf ms)",
2128 str, sum / 1000.0, seq->avg());
2129 }
2130
2131 void TraceGen0TimeData::print_summary_sd(const char* str,
2132 const NumberSeq* seq) const {
2133 print_summary(str, seq);
2134 gclog_or_tty->print_cr("%+45s = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
2135 "(num", seq->num(), seq->sd(), seq->maximum());
2136 }
2137
2138 void TraceGen0TimeData::print() const {
2139 if (!TraceGen0Time) {
2140 return;
2141 }
2142
2143 gclog_or_tty->print_cr("ALL PAUSES");
2144 print_summary_sd(" Total", &_total);
2145 gclog_or_tty->print_cr("");
2146 gclog_or_tty->print_cr("");
2147 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num);
2148 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num);
2149 gclog_or_tty->print_cr("");
2150
2151 gclog_or_tty->print_cr("EVACUATION PAUSES");
2152
2153 if (_young_pause_num == 0 && _mixed_pause_num == 0) {
2154 gclog_or_tty->print_cr("none");
2155 } else {
2156 print_summary_sd(" Evacuation Pauses", &_total);
2157 print_summary(" Root Region Scan Wait", &_root_region_scan_wait);
2158 print_summary(" Parallel Time", &_parallel);
2159 print_summary(" Ext Root Scanning", &_ext_root_scan);
2160 print_summary(" SATB Filtering", &_satb_filtering);
2161 print_summary(" Update RS", &_update_rs);
2162 print_summary(" Scan RS", &_scan_rs);
2163 print_summary(" Object Copy", &_obj_copy);
2164 print_summary(" Termination", &_termination);
2165 print_summary(" Parallel Other", &_parallel_other);
2166 print_summary(" Clear CT", &_clear_ct);
2167 print_summary(" Other", &_other);
2168 }
2169 gclog_or_tty->print_cr("");
2170
2171 gclog_or_tty->print_cr("MISC");
2172 print_summary_sd(" Stop World", &_all_stop_world_times_ms);
2173 print_summary_sd(" Yields", &_all_yield_times_ms);
2174 }
2175
2176 void TraceGen1TimeData::record_full_collection(double full_gc_time_ms) {
2177 if (TraceGen1Time) {
2178 _all_full_gc_times.add(full_gc_time_ms);
2179 }
2180 }
2181
2182 void TraceGen1TimeData::print() const {
2183 if (!TraceGen1Time) {
2184 return;
2185 }
2186
2187 if (_all_full_gc_times.num() > 0) {
2188 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s",
2189 _all_full_gc_times.num(),
2190 _all_full_gc_times.sum() / 1000.0);
2191 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg());
2192 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]",
2193 _all_full_gc_times.sd(),
2194 _all_full_gc_times.maximum());
2195 }
2196 }
--- EOF ---