Sat, 12 Oct 2013 00:49:19 +0200
8023643: G1 assert failed when NewSize was specified greater than MaxNewSize
Summary: Exit with an error if incompatible NewSize and MaxNeSize are set
Reviewed-by: brutisso, tschatzl
1 /*
2 * Copyright (c) 2001, 2013, 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 */
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"
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
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 };
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 };
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 };
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 };
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 };
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 };
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 };
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 };
81 G1CollectorPolicy::G1CollectorPolicy() :
82 _parallel_gc_threads(G1CollectedHeap::use_parallel_gc_threads()
83 ? ParallelGCThreads : 1),
85 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
86 _stop_world_start(0.0),
88 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
89 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
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)),
106 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)),
107 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)),
109 _pause_time_target_ms((double) MaxGCPauseMillis),
111 _gcs_are_young(true),
113 _during_marking(false),
114 _in_marking_window(false),
115 _in_marking_window_im(false),
117 _recent_prev_end_times_for_all_gcs_sec(
118 new TruncatedSeq(NumPrevPausesForHeuristics)),
120 _recent_avg_pause_time_ratio(0.0),
122 _initiate_conc_mark_if_possible(false),
123 _during_initial_mark_pause(false),
124 _last_young_gc(false),
125 _last_gc_was_young(false),
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),
134 _eden_cset_region_length(0),
135 _survivor_cset_region_length(0),
136 _old_cset_region_length(0),
138 _collection_set(NULL),
139 _collection_set_bytes_used_before(0),
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),
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
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),
166 _gc_overhead_perc(0.0) {
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.
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();
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 }
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 }
203 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime());
204 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0;
206 _phase_times = new G1GCPhaseTimes(_parallel_gc_threads);
208 int index = MIN2(_parallel_gc_threads - 1, 7);
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]);
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.
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 }
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 }
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 }
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 }
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);
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 %u", confidence_perc);
288 }
289 _sigma = (double) confidence_perc / 100.0;
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;
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));
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 %u", 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;
316 initialize_all();
317 _collectionSetChooser = new CollectionSetChooser();
318 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags
319 }
321 void G1CollectorPolicy::initialize_flags() {
322 _min_alignment = HeapRegion::GrainBytes;
323 size_t card_table_alignment = GenRemSet::max_alignment_constraint(rem_set_name());
324 size_t page_size = UseLargePages ? os::large_page_size() : os::vm_page_size();
325 _max_alignment = MAX3(card_table_alignment, _min_alignment, page_size);
326 if (SurvivorRatio < 1) {
327 vm_exit_during_initialization("Invalid survivor ratio specified");
328 }
329 CollectorPolicy::initialize_flags();
330 }
332 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true) {
333 assert(G1NewSizePercent <= G1MaxNewSizePercent, "Min larger than max");
334 assert(G1NewSizePercent > 0 && G1NewSizePercent < 100, "Min out of bounds");
335 assert(G1MaxNewSizePercent > 0 && G1MaxNewSizePercent < 100, "Max out of bounds");
337 if (FLAG_IS_CMDLINE(NewRatio)) {
338 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) {
339 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio");
340 } else {
341 _sizer_kind = SizerNewRatio;
342 _adaptive_size = false;
343 return;
344 }
345 }
347 if (FLAG_IS_CMDLINE(NewSize) && FLAG_IS_CMDLINE(MaxNewSize) && NewSize > MaxNewSize) {
348 vm_exit_during_initialization("Initial young gen size set larger than the maximum young gen size");
349 }
351 if (FLAG_IS_CMDLINE(NewSize)) {
352 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes),
353 1U);
354 if (FLAG_IS_CMDLINE(MaxNewSize)) {
355 _max_desired_young_length =
356 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes),
357 1U);
358 _sizer_kind = SizerMaxAndNewSize;
359 _adaptive_size = _min_desired_young_length == _max_desired_young_length;
360 } else {
361 _sizer_kind = SizerNewSizeOnly;
362 }
363 } else if (FLAG_IS_CMDLINE(MaxNewSize)) {
364 _max_desired_young_length =
365 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes),
366 1U);
367 _sizer_kind = SizerMaxNewSizeOnly;
368 }
369 }
371 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) {
372 uint default_value = (new_number_of_heap_regions * G1NewSizePercent) / 100;
373 return MAX2(1U, default_value);
374 }
376 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) {
377 uint default_value = (new_number_of_heap_regions * G1MaxNewSizePercent) / 100;
378 return MAX2(1U, default_value);
379 }
381 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) {
382 assert(new_number_of_heap_regions > 0, "Heap must be initialized");
384 switch (_sizer_kind) {
385 case SizerDefaults:
386 _min_desired_young_length = calculate_default_min_length(new_number_of_heap_regions);
387 _max_desired_young_length = calculate_default_max_length(new_number_of_heap_regions);
388 break;
389 case SizerNewSizeOnly:
390 _max_desired_young_length = calculate_default_max_length(new_number_of_heap_regions);
391 _max_desired_young_length = MAX2(_min_desired_young_length, _max_desired_young_length);
392 break;
393 case SizerMaxNewSizeOnly:
394 _min_desired_young_length = calculate_default_min_length(new_number_of_heap_regions);
395 _min_desired_young_length = MIN2(_min_desired_young_length, _max_desired_young_length);
396 break;
397 case SizerMaxAndNewSize:
398 // Do nothing. Values set on the command line, don't update them at runtime.
399 break;
400 case SizerNewRatio:
401 _min_desired_young_length = new_number_of_heap_regions / (NewRatio + 1);
402 _max_desired_young_length = _min_desired_young_length;
403 break;
404 default:
405 ShouldNotReachHere();
406 }
408 assert(_min_desired_young_length <= _max_desired_young_length, "Invalid min/max young gen size values");
409 }
411 void G1CollectorPolicy::init() {
412 // Set aside an initial future to_space.
413 _g1 = G1CollectedHeap::heap();
415 assert(Heap_lock->owned_by_self(), "Locking discipline.");
417 initialize_gc_policy_counters();
419 if (adaptive_young_list_length()) {
420 _young_list_fixed_length = 0;
421 } else {
422 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length();
423 }
424 _free_regions_at_end_of_collection = _g1->free_regions();
425 update_young_list_target_length();
427 // We may immediately start allocating regions and placing them on the
428 // collection set list. Initialize the per-collection set info
429 start_incremental_cset_building();
430 }
432 // Create the jstat counters for the policy.
433 void G1CollectorPolicy::initialize_gc_policy_counters() {
434 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3);
435 }
437 bool G1CollectorPolicy::predict_will_fit(uint young_length,
438 double base_time_ms,
439 uint base_free_regions,
440 double target_pause_time_ms) {
441 if (young_length >= base_free_regions) {
442 // end condition 1: not enough space for the young regions
443 return false;
444 }
446 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1);
447 size_t bytes_to_copy =
448 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
449 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
450 double young_other_time_ms = predict_young_other_time_ms(young_length);
451 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms;
452 if (pause_time_ms > target_pause_time_ms) {
453 // end condition 2: prediction is over the target pause time
454 return false;
455 }
457 size_t free_bytes =
458 (base_free_regions - young_length) * HeapRegion::GrainBytes;
459 if ((2.0 * sigma()) * (double) bytes_to_copy > (double) free_bytes) {
460 // end condition 3: out-of-space (conservatively!)
461 return false;
462 }
464 // success!
465 return true;
466 }
468 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) {
469 // re-calculate the necessary reserve
470 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor;
471 // We use ceiling so that if reserve_regions_d is > 0.0 (but
472 // smaller than 1.0) we'll get 1.
473 _reserve_regions = (uint) ceil(reserve_regions_d);
475 _young_gen_sizer->heap_size_changed(new_number_of_regions);
476 }
478 uint G1CollectorPolicy::calculate_young_list_desired_min_length(
479 uint base_min_length) {
480 uint desired_min_length = 0;
481 if (adaptive_young_list_length()) {
482 if (_alloc_rate_ms_seq->num() > 3) {
483 double now_sec = os::elapsedTime();
484 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
485 double alloc_rate_ms = predict_alloc_rate_ms();
486 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms);
487 } else {
488 // otherwise we don't have enough info to make the prediction
489 }
490 }
491 desired_min_length += base_min_length;
492 // make sure we don't go below any user-defined minimum bound
493 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length);
494 }
496 uint G1CollectorPolicy::calculate_young_list_desired_max_length() {
497 // Here, we might want to also take into account any additional
498 // constraints (i.e., user-defined minimum bound). Currently, we
499 // effectively don't set this bound.
500 return _young_gen_sizer->max_desired_young_length();
501 }
503 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) {
504 if (rs_lengths == (size_t) -1) {
505 // if it's set to the default value (-1), we should predict it;
506 // otherwise, use the given value.
507 rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
508 }
510 // Calculate the absolute and desired min bounds.
512 // This is how many young regions we already have (currently: the survivors).
513 uint base_min_length = recorded_survivor_regions();
514 // This is the absolute minimum young length, which ensures that we
515 // can allocate one eden region in the worst-case.
516 uint absolute_min_length = base_min_length + 1;
517 uint desired_min_length =
518 calculate_young_list_desired_min_length(base_min_length);
519 if (desired_min_length < absolute_min_length) {
520 desired_min_length = absolute_min_length;
521 }
523 // Calculate the absolute and desired max bounds.
525 // We will try our best not to "eat" into the reserve.
526 uint absolute_max_length = 0;
527 if (_free_regions_at_end_of_collection > _reserve_regions) {
528 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions;
529 }
530 uint desired_max_length = calculate_young_list_desired_max_length();
531 if (desired_max_length > absolute_max_length) {
532 desired_max_length = absolute_max_length;
533 }
535 uint young_list_target_length = 0;
536 if (adaptive_young_list_length()) {
537 if (gcs_are_young()) {
538 young_list_target_length =
539 calculate_young_list_target_length(rs_lengths,
540 base_min_length,
541 desired_min_length,
542 desired_max_length);
543 _rs_lengths_prediction = rs_lengths;
544 } else {
545 // Don't calculate anything and let the code below bound it to
546 // the desired_min_length, i.e., do the next GC as soon as
547 // possible to maximize how many old regions we can add to it.
548 }
549 } else {
550 // The user asked for a fixed young gen so we'll fix the young gen
551 // whether the next GC is young or mixed.
552 young_list_target_length = _young_list_fixed_length;
553 }
555 // Make sure we don't go over the desired max length, nor under the
556 // desired min length. In case they clash, desired_min_length wins
557 // which is why that test is second.
558 if (young_list_target_length > desired_max_length) {
559 young_list_target_length = desired_max_length;
560 }
561 if (young_list_target_length < desired_min_length) {
562 young_list_target_length = desired_min_length;
563 }
565 assert(young_list_target_length > recorded_survivor_regions(),
566 "we should be able to allocate at least one eden region");
567 assert(young_list_target_length >= absolute_min_length, "post-condition");
568 _young_list_target_length = young_list_target_length;
570 update_max_gc_locker_expansion();
571 }
573 uint
574 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths,
575 uint base_min_length,
576 uint desired_min_length,
577 uint desired_max_length) {
578 assert(adaptive_young_list_length(), "pre-condition");
579 assert(gcs_are_young(), "only call this for young GCs");
581 // In case some edge-condition makes the desired max length too small...
582 if (desired_max_length <= desired_min_length) {
583 return desired_min_length;
584 }
586 // We'll adjust min_young_length and max_young_length not to include
587 // the already allocated young regions (i.e., so they reflect the
588 // min and max eden regions we'll allocate). The base_min_length
589 // will be reflected in the predictions by the
590 // survivor_regions_evac_time prediction.
591 assert(desired_min_length > base_min_length, "invariant");
592 uint min_young_length = desired_min_length - base_min_length;
593 assert(desired_max_length > base_min_length, "invariant");
594 uint max_young_length = desired_max_length - base_min_length;
596 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
597 double survivor_regions_evac_time = predict_survivor_regions_evac_time();
598 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
599 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
600 size_t scanned_cards = predict_young_card_num(adj_rs_lengths);
601 double base_time_ms =
602 predict_base_elapsed_time_ms(pending_cards, scanned_cards) +
603 survivor_regions_evac_time;
604 uint available_free_regions = _free_regions_at_end_of_collection;
605 uint base_free_regions = 0;
606 if (available_free_regions > _reserve_regions) {
607 base_free_regions = available_free_regions - _reserve_regions;
608 }
610 // Here, we will make sure that the shortest young length that
611 // makes sense fits within the target pause time.
613 if (predict_will_fit(min_young_length, base_time_ms,
614 base_free_regions, target_pause_time_ms)) {
615 // The shortest young length will fit into the target pause time;
616 // we'll now check whether the absolute maximum number of young
617 // regions will fit in the target pause time. If not, we'll do
618 // a binary search between min_young_length and max_young_length.
619 if (predict_will_fit(max_young_length, base_time_ms,
620 base_free_regions, target_pause_time_ms)) {
621 // The maximum young length will fit into the target pause time.
622 // We are done so set min young length to the maximum length (as
623 // the result is assumed to be returned in min_young_length).
624 min_young_length = max_young_length;
625 } else {
626 // The maximum possible number of young regions will not fit within
627 // the target pause time so we'll search for the optimal
628 // length. The loop invariants are:
629 //
630 // min_young_length < max_young_length
631 // min_young_length is known to fit into the target pause time
632 // max_young_length is known not to fit into the target pause time
633 //
634 // Going into the loop we know the above hold as we've just
635 // checked them. Every time around the loop we check whether
636 // the middle value between min_young_length and
637 // max_young_length fits into the target pause time. If it
638 // does, it becomes the new min. If it doesn't, it becomes
639 // the new max. This way we maintain the loop invariants.
641 assert(min_young_length < max_young_length, "invariant");
642 uint diff = (max_young_length - min_young_length) / 2;
643 while (diff > 0) {
644 uint young_length = min_young_length + diff;
645 if (predict_will_fit(young_length, base_time_ms,
646 base_free_regions, target_pause_time_ms)) {
647 min_young_length = young_length;
648 } else {
649 max_young_length = young_length;
650 }
651 assert(min_young_length < max_young_length, "invariant");
652 diff = (max_young_length - min_young_length) / 2;
653 }
654 // The results is min_young_length which, according to the
655 // loop invariants, should fit within the target pause time.
657 // These are the post-conditions of the binary search above:
658 assert(min_young_length < max_young_length,
659 "otherwise we should have discovered that max_young_length "
660 "fits into the pause target and not done the binary search");
661 assert(predict_will_fit(min_young_length, base_time_ms,
662 base_free_regions, target_pause_time_ms),
663 "min_young_length, the result of the binary search, should "
664 "fit into the pause target");
665 assert(!predict_will_fit(min_young_length + 1, base_time_ms,
666 base_free_regions, target_pause_time_ms),
667 "min_young_length, the result of the binary search, should be "
668 "optimal, so no larger length should fit into the pause target");
669 }
670 } else {
671 // Even the minimum length doesn't fit into the pause time
672 // target, return it as the result nevertheless.
673 }
674 return base_min_length + min_young_length;
675 }
677 double G1CollectorPolicy::predict_survivor_regions_evac_time() {
678 double survivor_regions_evac_time = 0.0;
679 for (HeapRegion * r = _recorded_survivor_head;
680 r != NULL && r != _recorded_survivor_tail->get_next_young_region();
681 r = r->get_next_young_region()) {
682 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, gcs_are_young());
683 }
684 return survivor_regions_evac_time;
685 }
687 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() {
688 guarantee( adaptive_young_list_length(), "should not call this otherwise" );
690 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths();
691 if (rs_lengths > _rs_lengths_prediction) {
692 // add 10% to avoid having to recalculate often
693 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000;
694 update_young_list_target_length(rs_lengths_prediction);
695 }
696 }
700 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size,
701 bool is_tlab,
702 bool* gc_overhead_limit_was_exceeded) {
703 guarantee(false, "Not using this policy feature yet.");
704 return NULL;
705 }
707 // This method controls how a collector handles one or more
708 // of its generations being fully allocated.
709 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size,
710 bool is_tlab) {
711 guarantee(false, "Not using this policy feature yet.");
712 return NULL;
713 }
716 #ifndef PRODUCT
717 bool G1CollectorPolicy::verify_young_ages() {
718 HeapRegion* head = _g1->young_list()->first_region();
719 return
720 verify_young_ages(head, _short_lived_surv_rate_group);
721 // also call verify_young_ages on any additional surv rate groups
722 }
724 bool
725 G1CollectorPolicy::verify_young_ages(HeapRegion* head,
726 SurvRateGroup *surv_rate_group) {
727 guarantee( surv_rate_group != NULL, "pre-condition" );
729 const char* name = surv_rate_group->name();
730 bool ret = true;
731 int prev_age = -1;
733 for (HeapRegion* curr = head;
734 curr != NULL;
735 curr = curr->get_next_young_region()) {
736 SurvRateGroup* group = curr->surv_rate_group();
737 if (group == NULL && !curr->is_survivor()) {
738 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name);
739 ret = false;
740 }
742 if (surv_rate_group == group) {
743 int age = curr->age_in_surv_rate_group();
745 if (age < 0) {
746 gclog_or_tty->print_cr("## %s: encountered negative age", name);
747 ret = false;
748 }
750 if (age <= prev_age) {
751 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing "
752 "(%d, %d)", name, age, prev_age);
753 ret = false;
754 }
755 prev_age = age;
756 }
757 }
759 return ret;
760 }
761 #endif // PRODUCT
763 void G1CollectorPolicy::record_full_collection_start() {
764 _full_collection_start_sec = os::elapsedTime();
765 record_heap_size_info_at_start(true /* full */);
766 // Release the future to-space so that it is available for compaction into.
767 _g1->set_full_collection();
768 }
770 void G1CollectorPolicy::record_full_collection_end() {
771 // Consider this like a collection pause for the purposes of allocation
772 // since last pause.
773 double end_sec = os::elapsedTime();
774 double full_gc_time_sec = end_sec - _full_collection_start_sec;
775 double full_gc_time_ms = full_gc_time_sec * 1000.0;
777 _trace_gen1_time_data.record_full_collection(full_gc_time_ms);
779 update_recent_gc_times(end_sec, full_gc_time_ms);
781 _g1->clear_full_collection();
783 // "Nuke" the heuristics that control the young/mixed GC
784 // transitions and make sure we start with young GCs after the Full GC.
785 set_gcs_are_young(true);
786 _last_young_gc = false;
787 clear_initiate_conc_mark_if_possible();
788 clear_during_initial_mark_pause();
789 _in_marking_window = false;
790 _in_marking_window_im = false;
792 _short_lived_surv_rate_group->start_adding_regions();
793 // also call this on any additional surv rate groups
795 record_survivor_regions(0, NULL, NULL);
797 _free_regions_at_end_of_collection = _g1->free_regions();
798 // Reset survivors SurvRateGroup.
799 _survivor_surv_rate_group->reset();
800 update_young_list_target_length();
801 _collectionSetChooser->clear();
802 }
804 void G1CollectorPolicy::record_stop_world_start() {
805 _stop_world_start = os::elapsedTime();
806 }
808 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec) {
809 // We only need to do this here as the policy will only be applied
810 // to the GC we're about to start. so, no point is calculating this
811 // every time we calculate / recalculate the target young length.
812 update_survivors_policy();
814 assert(_g1->used() == _g1->recalculate_used(),
815 err_msg("sanity, used: "SIZE_FORMAT" recalculate_used: "SIZE_FORMAT,
816 _g1->used(), _g1->recalculate_used()));
818 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0;
819 _trace_gen0_time_data.record_start_collection(s_w_t_ms);
820 _stop_world_start = 0.0;
822 record_heap_size_info_at_start(false /* full */);
824 phase_times()->record_cur_collection_start_sec(start_time_sec);
825 _pending_cards = _g1->pending_card_num();
827 _collection_set_bytes_used_before = 0;
828 _bytes_copied_during_gc = 0;
830 _last_gc_was_young = false;
832 // do that for any other surv rate groups
833 _short_lived_surv_rate_group->stop_adding_regions();
834 _survivors_age_table.clear();
836 assert( verify_young_ages(), "region age verification" );
837 }
839 void G1CollectorPolicy::record_concurrent_mark_init_end(double
840 mark_init_elapsed_time_ms) {
841 _during_marking = true;
842 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now");
843 clear_during_initial_mark_pause();
844 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms;
845 }
847 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
848 _mark_remark_start_sec = os::elapsedTime();
849 _during_marking = false;
850 }
852 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
853 double end_time_sec = os::elapsedTime();
854 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
855 _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
856 _cur_mark_stop_world_time_ms += elapsed_time_ms;
857 _prev_collection_pause_end_ms += elapsed_time_ms;
859 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true);
860 }
862 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
863 _mark_cleanup_start_sec = os::elapsedTime();
864 }
866 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
867 _last_young_gc = true;
868 _in_marking_window = false;
869 }
871 void G1CollectorPolicy::record_concurrent_pause() {
872 if (_stop_world_start > 0.0) {
873 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0;
874 _trace_gen0_time_data.record_yield_time(yield_ms);
875 }
876 }
878 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) {
879 if (_g1->concurrent_mark()->cmThread()->during_cycle()) {
880 return false;
881 }
883 size_t marking_initiating_used_threshold =
884 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent;
885 size_t cur_used_bytes = _g1->non_young_capacity_bytes();
886 size_t alloc_byte_size = alloc_word_size * HeapWordSize;
888 if ((cur_used_bytes + alloc_byte_size) > marking_initiating_used_threshold) {
889 if (gcs_are_young() && !_last_young_gc) {
890 ergo_verbose5(ErgoConcCycles,
891 "request concurrent cycle initiation",
892 ergo_format_reason("occupancy higher than threshold")
893 ergo_format_byte("occupancy")
894 ergo_format_byte("allocation request")
895 ergo_format_byte_perc("threshold")
896 ergo_format_str("source"),
897 cur_used_bytes,
898 alloc_byte_size,
899 marking_initiating_used_threshold,
900 (double) InitiatingHeapOccupancyPercent,
901 source);
902 return true;
903 } else {
904 ergo_verbose5(ErgoConcCycles,
905 "do not request concurrent cycle initiation",
906 ergo_format_reason("still doing mixed collections")
907 ergo_format_byte("occupancy")
908 ergo_format_byte("allocation request")
909 ergo_format_byte_perc("threshold")
910 ergo_format_str("source"),
911 cur_used_bytes,
912 alloc_byte_size,
913 marking_initiating_used_threshold,
914 (double) InitiatingHeapOccupancyPercent,
915 source);
916 }
917 }
919 return false;
920 }
922 // Anything below that is considered to be zero
923 #define MIN_TIMER_GRANULARITY 0.0000001
925 void G1CollectorPolicy::record_collection_pause_end(double pause_time_ms, EvacuationInfo& evacuation_info) {
926 double end_time_sec = os::elapsedTime();
927 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(),
928 "otherwise, the subtraction below does not make sense");
929 size_t rs_size =
930 _cur_collection_pause_used_regions_at_start - cset_region_length();
931 size_t cur_used_bytes = _g1->used();
932 assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
933 bool last_pause_included_initial_mark = false;
934 bool update_stats = !_g1->evacuation_failed();
936 #ifndef PRODUCT
937 if (G1YoungSurvRateVerbose) {
938 gclog_or_tty->print_cr("");
939 _short_lived_surv_rate_group->print();
940 // do that for any other surv rate groups too
941 }
942 #endif // PRODUCT
944 last_pause_included_initial_mark = during_initial_mark_pause();
945 if (last_pause_included_initial_mark) {
946 record_concurrent_mark_init_end(0.0);
947 } else if (need_to_start_conc_mark("end of GC")) {
948 // Note: this might have already been set, if during the last
949 // pause we decided to start a cycle but at the beginning of
950 // this pause we decided to postpone it. That's OK.
951 set_initiate_conc_mark_if_possible();
952 }
954 _mmu_tracker->add_pause(end_time_sec - pause_time_ms/1000.0,
955 end_time_sec, false);
957 evacuation_info.set_collectionset_used_before(_collection_set_bytes_used_before);
958 evacuation_info.set_bytes_copied(_bytes_copied_during_gc);
960 if (update_stats) {
961 _trace_gen0_time_data.record_end_collection(pause_time_ms, phase_times());
962 // this is where we update the allocation rate of the application
963 double app_time_ms =
964 (phase_times()->cur_collection_start_sec() * 1000.0 - _prev_collection_pause_end_ms);
965 if (app_time_ms < MIN_TIMER_GRANULARITY) {
966 // This usually happens due to the timer not having the required
967 // granularity. Some Linuxes are the usual culprits.
968 // We'll just set it to something (arbitrarily) small.
969 app_time_ms = 1.0;
970 }
971 // We maintain the invariant that all objects allocated by mutator
972 // threads will be allocated out of eden regions. So, we can use
973 // the eden region number allocated since the previous GC to
974 // calculate the application's allocate rate. The only exception
975 // to that is humongous objects that are allocated separately. But
976 // given that humongous object allocations do not really affect
977 // either the pause's duration nor when the next pause will take
978 // place we can safely ignore them here.
979 uint regions_allocated = eden_cset_region_length();
980 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
981 _alloc_rate_ms_seq->add(alloc_rate_ms);
983 double interval_ms =
984 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
985 update_recent_gc_times(end_time_sec, pause_time_ms);
986 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
987 if (recent_avg_pause_time_ratio() < 0.0 ||
988 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) {
989 #ifndef PRODUCT
990 // Dump info to allow post-facto debugging
991 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds");
992 gclog_or_tty->print_cr("-------------------------------------------");
993 gclog_or_tty->print_cr("Recent GC Times (ms):");
994 _recent_gc_times_ms->dump();
995 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec);
996 _recent_prev_end_times_for_all_gcs_sec->dump();
997 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f",
998 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio());
999 // In debug mode, terminate the JVM if the user wants to debug at this point.
1000 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above");
1001 #endif // !PRODUCT
1002 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
1003 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
1004 if (_recent_avg_pause_time_ratio < 0.0) {
1005 _recent_avg_pause_time_ratio = 0.0;
1006 } else {
1007 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
1008 _recent_avg_pause_time_ratio = 1.0;
1009 }
1010 }
1011 }
1013 bool new_in_marking_window = _in_marking_window;
1014 bool new_in_marking_window_im = false;
1015 if (during_initial_mark_pause()) {
1016 new_in_marking_window = true;
1017 new_in_marking_window_im = true;
1018 }
1020 if (_last_young_gc) {
1021 // This is supposed to to be the "last young GC" before we start
1022 // doing mixed GCs. Here we decide whether to start mixed GCs or not.
1024 if (!last_pause_included_initial_mark) {
1025 if (next_gc_should_be_mixed("start mixed GCs",
1026 "do not start mixed GCs")) {
1027 set_gcs_are_young(false);
1028 }
1029 } else {
1030 ergo_verbose0(ErgoMixedGCs,
1031 "do not start mixed GCs",
1032 ergo_format_reason("concurrent cycle is about to start"));
1033 }
1034 _last_young_gc = false;
1035 }
1037 if (!_last_gc_was_young) {
1038 // This is a mixed GC. Here we decide whether to continue doing
1039 // mixed GCs or not.
1041 if (!next_gc_should_be_mixed("continue mixed GCs",
1042 "do not continue mixed GCs")) {
1043 set_gcs_are_young(true);
1044 }
1045 }
1047 _short_lived_surv_rate_group->start_adding_regions();
1048 // do that for any other surv rate groupsx
1050 if (update_stats) {
1051 double cost_per_card_ms = 0.0;
1052 if (_pending_cards > 0) {
1053 cost_per_card_ms = phase_times()->average_last_update_rs_time() / (double) _pending_cards;
1054 _cost_per_card_ms_seq->add(cost_per_card_ms);
1055 }
1057 size_t cards_scanned = _g1->cards_scanned();
1059 double cost_per_entry_ms = 0.0;
1060 if (cards_scanned > 10) {
1061 cost_per_entry_ms = phase_times()->average_last_scan_rs_time() / (double) cards_scanned;
1062 if (_last_gc_was_young) {
1063 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
1064 } else {
1065 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms);
1066 }
1067 }
1069 if (_max_rs_lengths > 0) {
1070 double cards_per_entry_ratio =
1071 (double) cards_scanned / (double) _max_rs_lengths;
1072 if (_last_gc_was_young) {
1073 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1074 } else {
1075 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1076 }
1077 }
1079 // This is defensive. For a while _max_rs_lengths could get
1080 // smaller than _recorded_rs_lengths which was causing
1081 // rs_length_diff to get very large and mess up the RSet length
1082 // predictions. The reason was unsafe concurrent updates to the
1083 // _inc_cset_recorded_rs_lengths field which the code below guards
1084 // against (see CR 7118202). This bug has now been fixed (see CR
1085 // 7119027). However, I'm still worried that
1086 // _inc_cset_recorded_rs_lengths might still end up somewhat
1087 // inaccurate. The concurrent refinement thread calculates an
1088 // RSet's length concurrently with other CR threads updating it
1089 // which might cause it to calculate the length incorrectly (if,
1090 // say, it's in mid-coarsening). So I'll leave in the defensive
1091 // conditional below just in case.
1092 size_t rs_length_diff = 0;
1093 if (_max_rs_lengths > _recorded_rs_lengths) {
1094 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths;
1095 }
1096 _rs_length_diff_seq->add((double) rs_length_diff);
1098 size_t freed_bytes = _heap_used_bytes_before_gc - cur_used_bytes;
1099 size_t copied_bytes = _collection_set_bytes_used_before - freed_bytes;
1100 double cost_per_byte_ms = 0.0;
1102 if (copied_bytes > 0) {
1103 cost_per_byte_ms = phase_times()->average_last_obj_copy_time() / (double) copied_bytes;
1104 if (_in_marking_window) {
1105 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
1106 } else {
1107 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
1108 }
1109 }
1111 double all_other_time_ms = pause_time_ms -
1112 (phase_times()->average_last_update_rs_time() + phase_times()->average_last_scan_rs_time()
1113 + phase_times()->average_last_obj_copy_time() + phase_times()->average_last_termination_time());
1115 double young_other_time_ms = 0.0;
1116 if (young_cset_region_length() > 0) {
1117 young_other_time_ms =
1118 phase_times()->young_cset_choice_time_ms() +
1119 phase_times()->young_free_cset_time_ms();
1120 _young_other_cost_per_region_ms_seq->add(young_other_time_ms /
1121 (double) young_cset_region_length());
1122 }
1123 double non_young_other_time_ms = 0.0;
1124 if (old_cset_region_length() > 0) {
1125 non_young_other_time_ms =
1126 phase_times()->non_young_cset_choice_time_ms() +
1127 phase_times()->non_young_free_cset_time_ms();
1129 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms /
1130 (double) old_cset_region_length());
1131 }
1133 double constant_other_time_ms = all_other_time_ms -
1134 (young_other_time_ms + non_young_other_time_ms);
1135 _constant_other_time_ms_seq->add(constant_other_time_ms);
1137 double survival_ratio = 0.0;
1138 if (_collection_set_bytes_used_before > 0) {
1139 survival_ratio = (double) _bytes_copied_during_gc /
1140 (double) _collection_set_bytes_used_before;
1141 }
1143 _pending_cards_seq->add((double) _pending_cards);
1144 _rs_lengths_seq->add((double) _max_rs_lengths);
1145 }
1147 _in_marking_window = new_in_marking_window;
1148 _in_marking_window_im = new_in_marking_window_im;
1149 _free_regions_at_end_of_collection = _g1->free_regions();
1150 update_young_list_target_length();
1152 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
1153 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
1154 adjust_concurrent_refinement(phase_times()->average_last_update_rs_time(),
1155 phase_times()->sum_last_update_rs_processed_buffers(), update_rs_time_goal_ms);
1157 _collectionSetChooser->verify();
1158 }
1160 #define EXT_SIZE_FORMAT "%.1f%s"
1161 #define EXT_SIZE_PARAMS(bytes) \
1162 byte_size_in_proper_unit((double)(bytes)), \
1163 proper_unit_for_byte_size((bytes))
1165 void G1CollectorPolicy::record_heap_size_info_at_start(bool full) {
1166 YoungList* young_list = _g1->young_list();
1167 _eden_used_bytes_before_gc = young_list->eden_used_bytes();
1168 _survivor_used_bytes_before_gc = young_list->survivor_used_bytes();
1169 _heap_capacity_bytes_before_gc = _g1->capacity();
1170 _heap_used_bytes_before_gc = _g1->used();
1171 _cur_collection_pause_used_regions_at_start = _g1->used_regions();
1173 _eden_capacity_bytes_before_gc =
1174 (_young_list_target_length * HeapRegion::GrainBytes) - _survivor_used_bytes_before_gc;
1176 if (full) {
1177 _metaspace_used_bytes_before_gc = MetaspaceAux::allocated_used_bytes();
1178 }
1179 }
1181 void G1CollectorPolicy::print_heap_transition() {
1182 _g1->print_size_transition(gclog_or_tty,
1183 _heap_used_bytes_before_gc,
1184 _g1->used(),
1185 _g1->capacity());
1186 }
1188 void G1CollectorPolicy::print_detailed_heap_transition(bool full) {
1189 YoungList* young_list = _g1->young_list();
1191 size_t eden_used_bytes_after_gc = young_list->eden_used_bytes();
1192 size_t survivor_used_bytes_after_gc = young_list->survivor_used_bytes();
1193 size_t heap_used_bytes_after_gc = _g1->used();
1195 size_t heap_capacity_bytes_after_gc = _g1->capacity();
1196 size_t eden_capacity_bytes_after_gc =
1197 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_used_bytes_after_gc;
1199 gclog_or_tty->print(
1200 " [Eden: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT") "
1201 "Survivors: "EXT_SIZE_FORMAT"->"EXT_SIZE_FORMAT" "
1202 "Heap: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"
1203 EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")]",
1204 EXT_SIZE_PARAMS(_eden_used_bytes_before_gc),
1205 EXT_SIZE_PARAMS(_eden_capacity_bytes_before_gc),
1206 EXT_SIZE_PARAMS(eden_used_bytes_after_gc),
1207 EXT_SIZE_PARAMS(eden_capacity_bytes_after_gc),
1208 EXT_SIZE_PARAMS(_survivor_used_bytes_before_gc),
1209 EXT_SIZE_PARAMS(survivor_used_bytes_after_gc),
1210 EXT_SIZE_PARAMS(_heap_used_bytes_before_gc),
1211 EXT_SIZE_PARAMS(_heap_capacity_bytes_before_gc),
1212 EXT_SIZE_PARAMS(heap_used_bytes_after_gc),
1213 EXT_SIZE_PARAMS(heap_capacity_bytes_after_gc));
1215 if (full) {
1216 MetaspaceAux::print_metaspace_change(_metaspace_used_bytes_before_gc);
1217 }
1219 gclog_or_tty->cr();
1220 }
1222 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1223 double update_rs_processed_buffers,
1224 double goal_ms) {
1225 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1226 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1228 if (G1UseAdaptiveConcRefinement) {
1229 const int k_gy = 3, k_gr = 6;
1230 const double inc_k = 1.1, dec_k = 0.9;
1232 int g = cg1r->green_zone();
1233 if (update_rs_time > goal_ms) {
1234 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1235 } else {
1236 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1237 g = (int)MAX2(g * inc_k, g + 1.0);
1238 }
1239 }
1240 // Change the refinement threads params
1241 cg1r->set_green_zone(g);
1242 cg1r->set_yellow_zone(g * k_gy);
1243 cg1r->set_red_zone(g * k_gr);
1244 cg1r->reinitialize_threads();
1246 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1);
1247 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1248 cg1r->yellow_zone());
1249 // Change the barrier params
1250 dcqs.set_process_completed_threshold(processing_threshold);
1251 dcqs.set_max_completed_queue(cg1r->red_zone());
1252 }
1254 int curr_queue_size = dcqs.completed_buffers_num();
1255 if (curr_queue_size >= cg1r->yellow_zone()) {
1256 dcqs.set_completed_queue_padding(curr_queue_size);
1257 } else {
1258 dcqs.set_completed_queue_padding(0);
1259 }
1260 dcqs.notify_if_necessary();
1261 }
1263 double
1264 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
1265 size_t scanned_cards) {
1266 return
1267 predict_rs_update_time_ms(pending_cards) +
1268 predict_rs_scan_time_ms(scanned_cards) +
1269 predict_constant_other_time_ms();
1270 }
1272 double
1273 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) {
1274 size_t rs_length = predict_rs_length_diff();
1275 size_t card_num;
1276 if (gcs_are_young()) {
1277 card_num = predict_young_card_num(rs_length);
1278 } else {
1279 card_num = predict_non_young_card_num(rs_length);
1280 }
1281 return predict_base_elapsed_time_ms(pending_cards, card_num);
1282 }
1284 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) {
1285 size_t bytes_to_copy;
1286 if (hr->is_marked())
1287 bytes_to_copy = hr->max_live_bytes();
1288 else {
1289 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant");
1290 int age = hr->age_in_surv_rate_group();
1291 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
1292 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate);
1293 }
1294 return bytes_to_copy;
1295 }
1297 double
1298 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
1299 bool for_young_gc) {
1300 size_t rs_length = hr->rem_set()->occupied();
1301 size_t card_num;
1303 // Predicting the number of cards is based on which type of GC
1304 // we're predicting for.
1305 if (for_young_gc) {
1306 card_num = predict_young_card_num(rs_length);
1307 } else {
1308 card_num = predict_non_young_card_num(rs_length);
1309 }
1310 size_t bytes_to_copy = predict_bytes_to_copy(hr);
1312 double region_elapsed_time_ms =
1313 predict_rs_scan_time_ms(card_num) +
1314 predict_object_copy_time_ms(bytes_to_copy);
1316 // The prediction of the "other" time for this region is based
1317 // upon the region type and NOT the GC type.
1318 if (hr->is_young()) {
1319 region_elapsed_time_ms += predict_young_other_time_ms(1);
1320 } else {
1321 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
1322 }
1323 return region_elapsed_time_ms;
1324 }
1326 void
1327 G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length,
1328 uint survivor_cset_region_length) {
1329 _eden_cset_region_length = eden_cset_region_length;
1330 _survivor_cset_region_length = survivor_cset_region_length;
1331 _old_cset_region_length = 0;
1332 }
1334 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) {
1335 _recorded_rs_lengths = rs_lengths;
1336 }
1338 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
1339 double elapsed_ms) {
1340 _recent_gc_times_ms->add(elapsed_ms);
1341 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
1342 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
1343 }
1345 size_t G1CollectorPolicy::expansion_amount() {
1346 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0;
1347 double threshold = _gc_overhead_perc;
1348 if (recent_gc_overhead > threshold) {
1349 // We will double the existing space, or take
1350 // G1ExpandByPercentOfAvailable % of the available expansion
1351 // space, whichever is smaller, bounded below by a minimum
1352 // expansion (unless that's all that's left.)
1353 const size_t min_expand_bytes = 1*M;
1354 size_t reserved_bytes = _g1->max_capacity();
1355 size_t committed_bytes = _g1->capacity();
1356 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
1357 size_t expand_bytes;
1358 size_t expand_bytes_via_pct =
1359 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
1360 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
1361 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
1362 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
1364 ergo_verbose5(ErgoHeapSizing,
1365 "attempt heap expansion",
1366 ergo_format_reason("recent GC overhead higher than "
1367 "threshold after GC")
1368 ergo_format_perc("recent GC overhead")
1369 ergo_format_perc("threshold")
1370 ergo_format_byte("uncommitted")
1371 ergo_format_byte_perc("calculated expansion amount"),
1372 recent_gc_overhead, threshold,
1373 uncommitted_bytes,
1374 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable);
1376 return expand_bytes;
1377 } else {
1378 return 0;
1379 }
1380 }
1382 void G1CollectorPolicy::print_tracing_info() const {
1383 _trace_gen0_time_data.print();
1384 _trace_gen1_time_data.print();
1385 }
1387 void G1CollectorPolicy::print_yg_surv_rate_info() const {
1388 #ifndef PRODUCT
1389 _short_lived_surv_rate_group->print_surv_rate_summary();
1390 // add this call for any other surv rate groups
1391 #endif // PRODUCT
1392 }
1394 uint G1CollectorPolicy::max_regions(int purpose) {
1395 switch (purpose) {
1396 case GCAllocForSurvived:
1397 return _max_survivor_regions;
1398 case GCAllocForTenured:
1399 return REGIONS_UNLIMITED;
1400 default:
1401 ShouldNotReachHere();
1402 return REGIONS_UNLIMITED;
1403 };
1404 }
1406 void G1CollectorPolicy::update_max_gc_locker_expansion() {
1407 uint expansion_region_num = 0;
1408 if (GCLockerEdenExpansionPercent > 0) {
1409 double perc = (double) GCLockerEdenExpansionPercent / 100.0;
1410 double expansion_region_num_d = perc * (double) _young_list_target_length;
1411 // We use ceiling so that if expansion_region_num_d is > 0.0 (but
1412 // less than 1.0) we'll get 1.
1413 expansion_region_num = (uint) ceil(expansion_region_num_d);
1414 } else {
1415 assert(expansion_region_num == 0, "sanity");
1416 }
1417 _young_list_max_length = _young_list_target_length + expansion_region_num;
1418 assert(_young_list_target_length <= _young_list_max_length, "post-condition");
1419 }
1421 // Calculates survivor space parameters.
1422 void G1CollectorPolicy::update_survivors_policy() {
1423 double max_survivor_regions_d =
1424 (double) _young_list_target_length / (double) SurvivorRatio;
1425 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but
1426 // smaller than 1.0) we'll get 1.
1427 _max_survivor_regions = (uint) ceil(max_survivor_regions_d);
1429 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
1430 HeapRegion::GrainWords * _max_survivor_regions);
1431 }
1433 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(
1434 GCCause::Cause gc_cause) {
1435 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
1436 if (!during_cycle) {
1437 ergo_verbose1(ErgoConcCycles,
1438 "request concurrent cycle initiation",
1439 ergo_format_reason("requested by GC cause")
1440 ergo_format_str("GC cause"),
1441 GCCause::to_string(gc_cause));
1442 set_initiate_conc_mark_if_possible();
1443 return true;
1444 } else {
1445 ergo_verbose1(ErgoConcCycles,
1446 "do not request concurrent cycle initiation",
1447 ergo_format_reason("concurrent cycle already in progress")
1448 ergo_format_str("GC cause"),
1449 GCCause::to_string(gc_cause));
1450 return false;
1451 }
1452 }
1454 void
1455 G1CollectorPolicy::decide_on_conc_mark_initiation() {
1456 // We are about to decide on whether this pause will be an
1457 // initial-mark pause.
1459 // First, during_initial_mark_pause() should not be already set. We
1460 // will set it here if we have to. However, it should be cleared by
1461 // the end of the pause (it's only set for the duration of an
1462 // initial-mark pause).
1463 assert(!during_initial_mark_pause(), "pre-condition");
1465 if (initiate_conc_mark_if_possible()) {
1466 // We had noticed on a previous pause that the heap occupancy has
1467 // gone over the initiating threshold and we should start a
1468 // concurrent marking cycle. So we might initiate one.
1470 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
1471 if (!during_cycle) {
1472 // The concurrent marking thread is not "during a cycle", i.e.,
1473 // it has completed the last one. So we can go ahead and
1474 // initiate a new cycle.
1476 set_during_initial_mark_pause();
1477 // We do not allow mixed GCs during marking.
1478 if (!gcs_are_young()) {
1479 set_gcs_are_young(true);
1480 ergo_verbose0(ErgoMixedGCs,
1481 "end mixed GCs",
1482 ergo_format_reason("concurrent cycle is about to start"));
1483 }
1485 // And we can now clear initiate_conc_mark_if_possible() as
1486 // we've already acted on it.
1487 clear_initiate_conc_mark_if_possible();
1489 ergo_verbose0(ErgoConcCycles,
1490 "initiate concurrent cycle",
1491 ergo_format_reason("concurrent cycle initiation requested"));
1492 } else {
1493 // The concurrent marking thread is still finishing up the
1494 // previous cycle. If we start one right now the two cycles
1495 // overlap. In particular, the concurrent marking thread might
1496 // be in the process of clearing the next marking bitmap (which
1497 // we will use for the next cycle if we start one). Starting a
1498 // cycle now will be bad given that parts of the marking
1499 // information might get cleared by the marking thread. And we
1500 // cannot wait for the marking thread to finish the cycle as it
1501 // periodically yields while clearing the next marking bitmap
1502 // and, if it's in a yield point, it's waiting for us to
1503 // finish. So, at this point we will not start a cycle and we'll
1504 // let the concurrent marking thread complete the last one.
1505 ergo_verbose0(ErgoConcCycles,
1506 "do not initiate concurrent cycle",
1507 ergo_format_reason("concurrent cycle already in progress"));
1508 }
1509 }
1510 }
1512 class KnownGarbageClosure: public HeapRegionClosure {
1513 G1CollectedHeap* _g1h;
1514 CollectionSetChooser* _hrSorted;
1516 public:
1517 KnownGarbageClosure(CollectionSetChooser* hrSorted) :
1518 _g1h(G1CollectedHeap::heap()), _hrSorted(hrSorted) { }
1520 bool doHeapRegion(HeapRegion* r) {
1521 // We only include humongous regions in collection
1522 // sets when concurrent mark shows that their contained object is
1523 // unreachable.
1525 // Do we have any marking information for this region?
1526 if (r->is_marked()) {
1527 // We will skip any region that's currently used as an old GC
1528 // alloc region (we should not consider those for collection
1529 // before we fill them up).
1530 if (_hrSorted->should_add(r) && !_g1h->is_old_gc_alloc_region(r)) {
1531 _hrSorted->add_region(r);
1532 }
1533 }
1534 return false;
1535 }
1536 };
1538 class ParKnownGarbageHRClosure: public HeapRegionClosure {
1539 G1CollectedHeap* _g1h;
1540 CSetChooserParUpdater _cset_updater;
1542 public:
1543 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
1544 uint chunk_size) :
1545 _g1h(G1CollectedHeap::heap()),
1546 _cset_updater(hrSorted, true /* parallel */, chunk_size) { }
1548 bool doHeapRegion(HeapRegion* r) {
1549 // Do we have any marking information for this region?
1550 if (r->is_marked()) {
1551 // We will skip any region that's currently used as an old GC
1552 // alloc region (we should not consider those for collection
1553 // before we fill them up).
1554 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) {
1555 _cset_updater.add_region(r);
1556 }
1557 }
1558 return false;
1559 }
1560 };
1562 class ParKnownGarbageTask: public AbstractGangTask {
1563 CollectionSetChooser* _hrSorted;
1564 uint _chunk_size;
1565 G1CollectedHeap* _g1;
1566 public:
1567 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size) :
1568 AbstractGangTask("ParKnownGarbageTask"),
1569 _hrSorted(hrSorted), _chunk_size(chunk_size),
1570 _g1(G1CollectedHeap::heap()) { }
1572 void work(uint worker_id) {
1573 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size);
1575 // Back to zero for the claim value.
1576 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, worker_id,
1577 _g1->workers()->active_workers(),
1578 HeapRegion::InitialClaimValue);
1579 }
1580 };
1582 void
1583 G1CollectorPolicy::record_concurrent_mark_cleanup_end(int no_of_gc_threads) {
1584 _collectionSetChooser->clear();
1586 uint region_num = _g1->n_regions();
1587 if (G1CollectedHeap::use_parallel_gc_threads()) {
1588 const uint OverpartitionFactor = 4;
1589 uint WorkUnit;
1590 // The use of MinChunkSize = 8 in the original code
1591 // causes some assertion failures when the total number of
1592 // region is less than 8. The code here tries to fix that.
1593 // Should the original code also be fixed?
1594 if (no_of_gc_threads > 0) {
1595 const uint MinWorkUnit = MAX2(region_num / no_of_gc_threads, 1U);
1596 WorkUnit = MAX2(region_num / (no_of_gc_threads * OverpartitionFactor),
1597 MinWorkUnit);
1598 } else {
1599 assert(no_of_gc_threads > 0,
1600 "The active gc workers should be greater than 0");
1601 // In a product build do something reasonable to avoid a crash.
1602 const uint MinWorkUnit = MAX2(region_num / (uint) ParallelGCThreads, 1U);
1603 WorkUnit =
1604 MAX2(region_num / (uint) (ParallelGCThreads * OverpartitionFactor),
1605 MinWorkUnit);
1606 }
1607 _collectionSetChooser->prepare_for_par_region_addition(_g1->n_regions(),
1608 WorkUnit);
1609 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser,
1610 (int) WorkUnit);
1611 _g1->workers()->run_task(&parKnownGarbageTask);
1613 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
1614 "sanity check");
1615 } else {
1616 KnownGarbageClosure knownGarbagecl(_collectionSetChooser);
1617 _g1->heap_region_iterate(&knownGarbagecl);
1618 }
1620 _collectionSetChooser->sort_regions();
1622 double end_sec = os::elapsedTime();
1623 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0;
1624 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
1625 _cur_mark_stop_world_time_ms += elapsed_time_ms;
1626 _prev_collection_pause_end_ms += elapsed_time_ms;
1627 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec, true);
1628 }
1630 // Add the heap region at the head of the non-incremental collection set
1631 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) {
1632 assert(_inc_cset_build_state == Active, "Precondition");
1633 assert(!hr->is_young(), "non-incremental add of young region");
1635 assert(!hr->in_collection_set(), "should not already be in the CSet");
1636 hr->set_in_collection_set(true);
1637 hr->set_next_in_collection_set(_collection_set);
1638 _collection_set = hr;
1639 _collection_set_bytes_used_before += hr->used();
1640 _g1->register_region_with_in_cset_fast_test(hr);
1641 size_t rs_length = hr->rem_set()->occupied();
1642 _recorded_rs_lengths += rs_length;
1643 _old_cset_region_length += 1;
1644 }
1646 // Initialize the per-collection-set information
1647 void G1CollectorPolicy::start_incremental_cset_building() {
1648 assert(_inc_cset_build_state == Inactive, "Precondition");
1650 _inc_cset_head = NULL;
1651 _inc_cset_tail = NULL;
1652 _inc_cset_bytes_used_before = 0;
1654 _inc_cset_max_finger = 0;
1655 _inc_cset_recorded_rs_lengths = 0;
1656 _inc_cset_recorded_rs_lengths_diffs = 0;
1657 _inc_cset_predicted_elapsed_time_ms = 0.0;
1658 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
1659 _inc_cset_build_state = Active;
1660 }
1662 void G1CollectorPolicy::finalize_incremental_cset_building() {
1663 assert(_inc_cset_build_state == Active, "Precondition");
1664 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
1666 // The two "main" fields, _inc_cset_recorded_rs_lengths and
1667 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread
1668 // that adds a new region to the CSet. Further updates by the
1669 // concurrent refinement thread that samples the young RSet lengths
1670 // are accumulated in the *_diffs fields. Here we add the diffs to
1671 // the "main" fields.
1673 if (_inc_cset_recorded_rs_lengths_diffs >= 0) {
1674 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs;
1675 } else {
1676 // This is defensive. The diff should in theory be always positive
1677 // as RSets can only grow between GCs. However, given that we
1678 // sample their size concurrently with other threads updating them
1679 // it's possible that we might get the wrong size back, which
1680 // could make the calculations somewhat inaccurate.
1681 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs);
1682 if (_inc_cset_recorded_rs_lengths >= diffs) {
1683 _inc_cset_recorded_rs_lengths -= diffs;
1684 } else {
1685 _inc_cset_recorded_rs_lengths = 0;
1686 }
1687 }
1688 _inc_cset_predicted_elapsed_time_ms +=
1689 _inc_cset_predicted_elapsed_time_ms_diffs;
1691 _inc_cset_recorded_rs_lengths_diffs = 0;
1692 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
1693 }
1695 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) {
1696 // This routine is used when:
1697 // * adding survivor regions to the incremental cset at the end of an
1698 // evacuation pause,
1699 // * adding the current allocation region to the incremental cset
1700 // when it is retired, and
1701 // * updating existing policy information for a region in the
1702 // incremental cset via young list RSet sampling.
1703 // Therefore this routine may be called at a safepoint by the
1704 // VM thread, or in-between safepoints by mutator threads (when
1705 // retiring the current allocation region) or a concurrent
1706 // refine thread (RSet sampling).
1708 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, gcs_are_young());
1709 size_t used_bytes = hr->used();
1710 _inc_cset_recorded_rs_lengths += rs_length;
1711 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms;
1712 _inc_cset_bytes_used_before += used_bytes;
1714 // Cache the values we have added to the aggregated informtion
1715 // in the heap region in case we have to remove this region from
1716 // the incremental collection set, or it is updated by the
1717 // rset sampling code
1718 hr->set_recorded_rs_length(rs_length);
1719 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms);
1720 }
1722 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr,
1723 size_t new_rs_length) {
1724 // Update the CSet information that is dependent on the new RS length
1725 assert(hr->is_young(), "Precondition");
1726 assert(!SafepointSynchronize::is_at_safepoint(),
1727 "should not be at a safepoint");
1729 // We could have updated _inc_cset_recorded_rs_lengths and
1730 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do
1731 // that atomically, as this code is executed by a concurrent
1732 // refinement thread, potentially concurrently with a mutator thread
1733 // allocating a new region and also updating the same fields. To
1734 // avoid the atomic operations we accumulate these updates on two
1735 // separate fields (*_diffs) and we'll just add them to the "main"
1736 // fields at the start of a GC.
1738 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length();
1739 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length;
1740 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff;
1742 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms();
1743 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, gcs_are_young());
1744 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms;
1745 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff;
1747 hr->set_recorded_rs_length(new_rs_length);
1748 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms);
1749 }
1751 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) {
1752 assert(hr->is_young(), "invariant");
1753 assert(hr->young_index_in_cset() > -1, "should have already been set");
1754 assert(_inc_cset_build_state == Active, "Precondition");
1756 // We need to clear and set the cached recorded/cached collection set
1757 // information in the heap region here (before the region gets added
1758 // to the collection set). An individual heap region's cached values
1759 // are calculated, aggregated with the policy collection set info,
1760 // and cached in the heap region here (initially) and (subsequently)
1761 // by the Young List sampling code.
1763 size_t rs_length = hr->rem_set()->occupied();
1764 add_to_incremental_cset_info(hr, rs_length);
1766 HeapWord* hr_end = hr->end();
1767 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end);
1769 assert(!hr->in_collection_set(), "invariant");
1770 hr->set_in_collection_set(true);
1771 assert( hr->next_in_collection_set() == NULL, "invariant");
1773 _g1->register_region_with_in_cset_fast_test(hr);
1774 }
1776 // Add the region at the RHS of the incremental cset
1777 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) {
1778 // We should only ever be appending survivors at the end of a pause
1779 assert( hr->is_survivor(), "Logic");
1781 // Do the 'common' stuff
1782 add_region_to_incremental_cset_common(hr);
1784 // Now add the region at the right hand side
1785 if (_inc_cset_tail == NULL) {
1786 assert(_inc_cset_head == NULL, "invariant");
1787 _inc_cset_head = hr;
1788 } else {
1789 _inc_cset_tail->set_next_in_collection_set(hr);
1790 }
1791 _inc_cset_tail = hr;
1792 }
1794 // Add the region to the LHS of the incremental cset
1795 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) {
1796 // Survivors should be added to the RHS at the end of a pause
1797 assert(!hr->is_survivor(), "Logic");
1799 // Do the 'common' stuff
1800 add_region_to_incremental_cset_common(hr);
1802 // Add the region at the left hand side
1803 hr->set_next_in_collection_set(_inc_cset_head);
1804 if (_inc_cset_head == NULL) {
1805 assert(_inc_cset_tail == NULL, "Invariant");
1806 _inc_cset_tail = hr;
1807 }
1808 _inc_cset_head = hr;
1809 }
1811 #ifndef PRODUCT
1812 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) {
1813 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be");
1815 st->print_cr("\nCollection_set:");
1816 HeapRegion* csr = list_head;
1817 while (csr != NULL) {
1818 HeapRegion* next = csr->next_in_collection_set();
1819 assert(csr->in_collection_set(), "bad CS");
1820 st->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d",
1821 HR_FORMAT_PARAMS(csr),
1822 csr->prev_top_at_mark_start(), csr->next_top_at_mark_start(),
1823 csr->age_in_surv_rate_group_cond());
1824 csr = next;
1825 }
1826 }
1827 #endif // !PRODUCT
1829 double G1CollectorPolicy::reclaimable_bytes_perc(size_t reclaimable_bytes) {
1830 // Returns the given amount of reclaimable bytes (that represents
1831 // the amount of reclaimable space still to be collected) as a
1832 // percentage of the current heap capacity.
1833 size_t capacity_bytes = _g1->capacity();
1834 return (double) reclaimable_bytes * 100.0 / (double) capacity_bytes;
1835 }
1837 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str,
1838 const char* false_action_str) {
1839 CollectionSetChooser* cset_chooser = _collectionSetChooser;
1840 if (cset_chooser->is_empty()) {
1841 ergo_verbose0(ErgoMixedGCs,
1842 false_action_str,
1843 ergo_format_reason("candidate old regions not available"));
1844 return false;
1845 }
1847 // Is the amount of uncollected reclaimable space above G1HeapWastePercent?
1848 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes();
1849 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes);
1850 double threshold = (double) G1HeapWastePercent;
1851 if (reclaimable_perc <= threshold) {
1852 ergo_verbose4(ErgoMixedGCs,
1853 false_action_str,
1854 ergo_format_reason("reclaimable percentage not over threshold")
1855 ergo_format_region("candidate old regions")
1856 ergo_format_byte_perc("reclaimable")
1857 ergo_format_perc("threshold"),
1858 cset_chooser->remaining_regions(),
1859 reclaimable_bytes,
1860 reclaimable_perc, threshold);
1861 return false;
1862 }
1864 ergo_verbose4(ErgoMixedGCs,
1865 true_action_str,
1866 ergo_format_reason("candidate old regions available")
1867 ergo_format_region("candidate old regions")
1868 ergo_format_byte_perc("reclaimable")
1869 ergo_format_perc("threshold"),
1870 cset_chooser->remaining_regions(),
1871 reclaimable_bytes,
1872 reclaimable_perc, threshold);
1873 return true;
1874 }
1876 uint G1CollectorPolicy::calc_min_old_cset_length() {
1877 // The min old CSet region bound is based on the maximum desired
1878 // number of mixed GCs after a cycle. I.e., even if some old regions
1879 // look expensive, we should add them to the CSet anyway to make
1880 // sure we go through the available old regions in no more than the
1881 // maximum desired number of mixed GCs.
1882 //
1883 // The calculation is based on the number of marked regions we added
1884 // to the CSet chooser in the first place, not how many remain, so
1885 // that the result is the same during all mixed GCs that follow a cycle.
1887 const size_t region_num = (size_t) _collectionSetChooser->length();
1888 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1);
1889 size_t result = region_num / gc_num;
1890 // emulate ceiling
1891 if (result * gc_num < region_num) {
1892 result += 1;
1893 }
1894 return (uint) result;
1895 }
1897 uint G1CollectorPolicy::calc_max_old_cset_length() {
1898 // The max old CSet region bound is based on the threshold expressed
1899 // as a percentage of the heap size. I.e., it should bound the
1900 // number of old regions added to the CSet irrespective of how many
1901 // of them are available.
1903 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1904 const size_t region_num = g1h->n_regions();
1905 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent;
1906 size_t result = region_num * perc / 100;
1907 // emulate ceiling
1908 if (100 * result < region_num * perc) {
1909 result += 1;
1910 }
1911 return (uint) result;
1912 }
1915 void G1CollectorPolicy::finalize_cset(double target_pause_time_ms, EvacuationInfo& evacuation_info) {
1916 double young_start_time_sec = os::elapsedTime();
1918 YoungList* young_list = _g1->young_list();
1919 finalize_incremental_cset_building();
1921 guarantee(target_pause_time_ms > 0.0,
1922 err_msg("target_pause_time_ms = %1.6lf should be positive",
1923 target_pause_time_ms));
1924 guarantee(_collection_set == NULL, "Precondition");
1926 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
1927 double predicted_pause_time_ms = base_time_ms;
1928 double time_remaining_ms = MAX2(target_pause_time_ms - base_time_ms, 0.0);
1930 ergo_verbose4(ErgoCSetConstruction | ErgoHigh,
1931 "start choosing CSet",
1932 ergo_format_size("_pending_cards")
1933 ergo_format_ms("predicted base time")
1934 ergo_format_ms("remaining time")
1935 ergo_format_ms("target pause time"),
1936 _pending_cards, base_time_ms, time_remaining_ms, target_pause_time_ms);
1938 _last_gc_was_young = gcs_are_young() ? true : false;
1940 if (_last_gc_was_young) {
1941 _trace_gen0_time_data.increment_young_collection_count();
1942 } else {
1943 _trace_gen0_time_data.increment_mixed_collection_count();
1944 }
1946 // The young list is laid with the survivor regions from the previous
1947 // pause are appended to the RHS of the young list, i.e.
1948 // [Newly Young Regions ++ Survivors from last pause].
1950 uint survivor_region_length = young_list->survivor_length();
1951 uint eden_region_length = young_list->length() - survivor_region_length;
1952 init_cset_region_lengths(eden_region_length, survivor_region_length);
1954 HeapRegion* hr = young_list->first_survivor_region();
1955 while (hr != NULL) {
1956 assert(hr->is_survivor(), "badly formed young list");
1957 hr->set_young();
1958 hr = hr->get_next_young_region();
1959 }
1961 // Clear the fields that point to the survivor list - they are all young now.
1962 young_list->clear_survivors();
1964 _collection_set = _inc_cset_head;
1965 _collection_set_bytes_used_before = _inc_cset_bytes_used_before;
1966 time_remaining_ms = MAX2(time_remaining_ms - _inc_cset_predicted_elapsed_time_ms, 0.0);
1967 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms;
1969 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
1970 "add young regions to CSet",
1971 ergo_format_region("eden")
1972 ergo_format_region("survivors")
1973 ergo_format_ms("predicted young region time"),
1974 eden_region_length, survivor_region_length,
1975 _inc_cset_predicted_elapsed_time_ms);
1977 // The number of recorded young regions is the incremental
1978 // collection set's current size
1979 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths);
1981 double young_end_time_sec = os::elapsedTime();
1982 phase_times()->record_young_cset_choice_time_ms((young_end_time_sec - young_start_time_sec) * 1000.0);
1984 // Set the start of the non-young choice time.
1985 double non_young_start_time_sec = young_end_time_sec;
1987 if (!gcs_are_young()) {
1988 CollectionSetChooser* cset_chooser = _collectionSetChooser;
1989 cset_chooser->verify();
1990 const uint min_old_cset_length = calc_min_old_cset_length();
1991 const uint max_old_cset_length = calc_max_old_cset_length();
1993 uint expensive_region_num = 0;
1994 bool check_time_remaining = adaptive_young_list_length();
1996 HeapRegion* hr = cset_chooser->peek();
1997 while (hr != NULL) {
1998 if (old_cset_region_length() >= max_old_cset_length) {
1999 // Added maximum number of old regions to the CSet.
2000 ergo_verbose2(ErgoCSetConstruction,
2001 "finish adding old regions to CSet",
2002 ergo_format_reason("old CSet region num reached max")
2003 ergo_format_region("old")
2004 ergo_format_region("max"),
2005 old_cset_region_length(), max_old_cset_length);
2006 break;
2007 }
2010 // Stop adding regions if the remaining reclaimable space is
2011 // not above G1HeapWastePercent.
2012 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes();
2013 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes);
2014 double threshold = (double) G1HeapWastePercent;
2015 if (reclaimable_perc <= threshold) {
2016 // We've added enough old regions that the amount of uncollected
2017 // reclaimable space is at or below the waste threshold. Stop
2018 // adding old regions to the CSet.
2019 ergo_verbose5(ErgoCSetConstruction,
2020 "finish adding old regions to CSet",
2021 ergo_format_reason("reclaimable percentage not over threshold")
2022 ergo_format_region("old")
2023 ergo_format_region("max")
2024 ergo_format_byte_perc("reclaimable")
2025 ergo_format_perc("threshold"),
2026 old_cset_region_length(),
2027 max_old_cset_length,
2028 reclaimable_bytes,
2029 reclaimable_perc, threshold);
2030 break;
2031 }
2033 double predicted_time_ms = predict_region_elapsed_time_ms(hr, gcs_are_young());
2034 if (check_time_remaining) {
2035 if (predicted_time_ms > time_remaining_ms) {
2036 // Too expensive for the current CSet.
2038 if (old_cset_region_length() >= min_old_cset_length) {
2039 // We have added the minimum number of old regions to the CSet,
2040 // we are done with this CSet.
2041 ergo_verbose4(ErgoCSetConstruction,
2042 "finish adding old regions to CSet",
2043 ergo_format_reason("predicted time is too high")
2044 ergo_format_ms("predicted time")
2045 ergo_format_ms("remaining time")
2046 ergo_format_region("old")
2047 ergo_format_region("min"),
2048 predicted_time_ms, time_remaining_ms,
2049 old_cset_region_length(), min_old_cset_length);
2050 break;
2051 }
2053 // We'll add it anyway given that we haven't reached the
2054 // minimum number of old regions.
2055 expensive_region_num += 1;
2056 }
2057 } else {
2058 if (old_cset_region_length() >= min_old_cset_length) {
2059 // In the non-auto-tuning case, we'll finish adding regions
2060 // to the CSet if we reach the minimum.
2061 ergo_verbose2(ErgoCSetConstruction,
2062 "finish adding old regions to CSet",
2063 ergo_format_reason("old CSet region num reached min")
2064 ergo_format_region("old")
2065 ergo_format_region("min"),
2066 old_cset_region_length(), min_old_cset_length);
2067 break;
2068 }
2069 }
2071 // We will add this region to the CSet.
2072 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0);
2073 predicted_pause_time_ms += predicted_time_ms;
2074 cset_chooser->remove_and_move_to_next(hr);
2075 _g1->old_set_remove(hr);
2076 add_old_region_to_cset(hr);
2078 hr = cset_chooser->peek();
2079 }
2080 if (hr == NULL) {
2081 ergo_verbose0(ErgoCSetConstruction,
2082 "finish adding old regions to CSet",
2083 ergo_format_reason("candidate old regions not available"));
2084 }
2086 if (expensive_region_num > 0) {
2087 // We print the information once here at the end, predicated on
2088 // whether we added any apparently expensive regions or not, to
2089 // avoid generating output per region.
2090 ergo_verbose4(ErgoCSetConstruction,
2091 "added expensive regions to CSet",
2092 ergo_format_reason("old CSet region num not reached min")
2093 ergo_format_region("old")
2094 ergo_format_region("expensive")
2095 ergo_format_region("min")
2096 ergo_format_ms("remaining time"),
2097 old_cset_region_length(),
2098 expensive_region_num,
2099 min_old_cset_length,
2100 time_remaining_ms);
2101 }
2103 cset_chooser->verify();
2104 }
2106 stop_incremental_cset_building();
2108 ergo_verbose5(ErgoCSetConstruction,
2109 "finish choosing CSet",
2110 ergo_format_region("eden")
2111 ergo_format_region("survivors")
2112 ergo_format_region("old")
2113 ergo_format_ms("predicted pause time")
2114 ergo_format_ms("target pause time"),
2115 eden_region_length, survivor_region_length,
2116 old_cset_region_length(),
2117 predicted_pause_time_ms, target_pause_time_ms);
2119 double non_young_end_time_sec = os::elapsedTime();
2120 phase_times()->record_non_young_cset_choice_time_ms((non_young_end_time_sec - non_young_start_time_sec) * 1000.0);
2121 evacuation_info.set_collectionset_regions(cset_region_length());
2122 }
2124 void TraceGen0TimeData::record_start_collection(double time_to_stop_the_world_ms) {
2125 if(TraceGen0Time) {
2126 _all_stop_world_times_ms.add(time_to_stop_the_world_ms);
2127 }
2128 }
2130 void TraceGen0TimeData::record_yield_time(double yield_time_ms) {
2131 if(TraceGen0Time) {
2132 _all_yield_times_ms.add(yield_time_ms);
2133 }
2134 }
2136 void TraceGen0TimeData::record_end_collection(double pause_time_ms, G1GCPhaseTimes* phase_times) {
2137 if(TraceGen0Time) {
2138 _total.add(pause_time_ms);
2139 _other.add(pause_time_ms - phase_times->accounted_time_ms());
2140 _root_region_scan_wait.add(phase_times->root_region_scan_wait_time_ms());
2141 _parallel.add(phase_times->cur_collection_par_time_ms());
2142 _ext_root_scan.add(phase_times->average_last_ext_root_scan_time());
2143 _satb_filtering.add(phase_times->average_last_satb_filtering_times_ms());
2144 _update_rs.add(phase_times->average_last_update_rs_time());
2145 _scan_rs.add(phase_times->average_last_scan_rs_time());
2146 _obj_copy.add(phase_times->average_last_obj_copy_time());
2147 _termination.add(phase_times->average_last_termination_time());
2149 double parallel_known_time = phase_times->average_last_ext_root_scan_time() +
2150 phase_times->average_last_satb_filtering_times_ms() +
2151 phase_times->average_last_update_rs_time() +
2152 phase_times->average_last_scan_rs_time() +
2153 phase_times->average_last_obj_copy_time() +
2154 + phase_times->average_last_termination_time();
2156 double parallel_other_time = phase_times->cur_collection_par_time_ms() - parallel_known_time;
2157 _parallel_other.add(parallel_other_time);
2158 _clear_ct.add(phase_times->cur_clear_ct_time_ms());
2159 }
2160 }
2162 void TraceGen0TimeData::increment_young_collection_count() {
2163 if(TraceGen0Time) {
2164 ++_young_pause_num;
2165 }
2166 }
2168 void TraceGen0TimeData::increment_mixed_collection_count() {
2169 if(TraceGen0Time) {
2170 ++_mixed_pause_num;
2171 }
2172 }
2174 void TraceGen0TimeData::print_summary(const char* str,
2175 const NumberSeq* seq) const {
2176 double sum = seq->sum();
2177 gclog_or_tty->print_cr("%-27s = %8.2lf s (avg = %8.2lf ms)",
2178 str, sum / 1000.0, seq->avg());
2179 }
2181 void TraceGen0TimeData::print_summary_sd(const char* str,
2182 const NumberSeq* seq) const {
2183 print_summary(str, seq);
2184 gclog_or_tty->print_cr("%+45s = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
2185 "(num", seq->num(), seq->sd(), seq->maximum());
2186 }
2188 void TraceGen0TimeData::print() const {
2189 if (!TraceGen0Time) {
2190 return;
2191 }
2193 gclog_or_tty->print_cr("ALL PAUSES");
2194 print_summary_sd(" Total", &_total);
2195 gclog_or_tty->print_cr("");
2196 gclog_or_tty->print_cr("");
2197 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num);
2198 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num);
2199 gclog_or_tty->print_cr("");
2201 gclog_or_tty->print_cr("EVACUATION PAUSES");
2203 if (_young_pause_num == 0 && _mixed_pause_num == 0) {
2204 gclog_or_tty->print_cr("none");
2205 } else {
2206 print_summary_sd(" Evacuation Pauses", &_total);
2207 print_summary(" Root Region Scan Wait", &_root_region_scan_wait);
2208 print_summary(" Parallel Time", &_parallel);
2209 print_summary(" Ext Root Scanning", &_ext_root_scan);
2210 print_summary(" SATB Filtering", &_satb_filtering);
2211 print_summary(" Update RS", &_update_rs);
2212 print_summary(" Scan RS", &_scan_rs);
2213 print_summary(" Object Copy", &_obj_copy);
2214 print_summary(" Termination", &_termination);
2215 print_summary(" Parallel Other", &_parallel_other);
2216 print_summary(" Clear CT", &_clear_ct);
2217 print_summary(" Other", &_other);
2218 }
2219 gclog_or_tty->print_cr("");
2221 gclog_or_tty->print_cr("MISC");
2222 print_summary_sd(" Stop World", &_all_stop_world_times_ms);
2223 print_summary_sd(" Yields", &_all_yield_times_ms);
2224 }
2226 void TraceGen1TimeData::record_full_collection(double full_gc_time_ms) {
2227 if (TraceGen1Time) {
2228 _all_full_gc_times.add(full_gc_time_ms);
2229 }
2230 }
2232 void TraceGen1TimeData::print() const {
2233 if (!TraceGen1Time) {
2234 return;
2235 }
2237 if (_all_full_gc_times.num() > 0) {
2238 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s",
2239 _all_full_gc_times.num(),
2240 _all_full_gc_times.sum() / 1000.0);
2241 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg());
2242 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]",
2243 _all_full_gc_times.sd(),
2244 _all_full_gc_times.maximum());
2245 }
2246 }