Fri, 16 Aug 2013 13:22:32 +0200
8007074: SIGSEGV at ParMarkBitMap::verify_clear()
Summary: Replace the broken large pages implementation on Linux. New flag: -XX:+UseTransparentHugePages - Linux specific flag to turn on transparent huge page hinting with madvise(..., MAP_HUGETLB). Changed behavior: -XX:+UseLargePages - tries to use -XX:+UseTransparentHugePages before trying other large pages implementations (on Linux). Changed behavior: -XX:+UseHugeTLBFS - Use upfront allocation of Large Pages instead of using the broken implementation to dynamically committing large pages. Changed behavior: -XX:LargePageSizeInBytes - Turned off the ability to use this flag on Linux and provides warning to user if set to a value different than the OS chosen large page size. Changed behavior: Setting no large page size - Now defaults to use -XX:UseTransparentHugePages if the OS supports it. Previously, -XX:+UseHugeTLBFS was chosen if the OS was configured to use large pages.
Reviewed-by: tschatzl, dcubed, brutisso
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.
171 HeapRegion::setup_heap_region_size(Arguments::min_heap_size());
172 HeapRegionRemSet::setup_remset_size();
174 G1ErgoVerbose::initialize();
175 if (PrintAdaptiveSizePolicy) {
176 // Currently, we only use a single switch for all the heuristics.
177 G1ErgoVerbose::set_enabled(true);
178 // Given that we don't currently have a verboseness level
179 // parameter, we'll hardcode this to high. This can be easily
180 // changed in the future.
181 G1ErgoVerbose::set_level(ErgoHigh);
182 } else {
183 G1ErgoVerbose::set_enabled(false);
184 }
186 // Verify PLAB sizes
187 const size_t region_size = HeapRegion::GrainWords;
188 if (YoungPLABSize > region_size || OldPLABSize > region_size) {
189 char buffer[128];
190 jio_snprintf(buffer, sizeof(buffer), "%sPLABSize should be at most "SIZE_FORMAT,
191 OldPLABSize > region_size ? "Old" : "Young", region_size);
192 vm_exit_during_initialization(buffer);
193 }
195 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime());
196 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0;
198 _phase_times = new G1GCPhaseTimes(_parallel_gc_threads);
200 int index = MIN2(_parallel_gc_threads - 1, 7);
202 _rs_length_diff_seq->add(rs_length_diff_defaults[index]);
203 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]);
204 _young_cards_per_entry_ratio_seq->add(
205 young_cards_per_entry_ratio_defaults[index]);
206 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]);
207 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]);
208 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]);
209 _young_other_cost_per_region_ms_seq->add(
210 young_other_cost_per_region_ms_defaults[index]);
211 _non_young_other_cost_per_region_ms_seq->add(
212 non_young_other_cost_per_region_ms_defaults[index]);
214 // Below, we might need to calculate the pause time target based on
215 // the pause interval. When we do so we are going to give G1 maximum
216 // flexibility and allow it to do pauses when it needs to. So, we'll
217 // arrange that the pause interval to be pause time target + 1 to
218 // ensure that a) the pause time target is maximized with respect to
219 // the pause interval and b) we maintain the invariant that pause
220 // time target < pause interval. If the user does not want this
221 // maximum flexibility, they will have to set the pause interval
222 // explicitly.
224 // First make sure that, if either parameter is set, its value is
225 // reasonable.
226 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
227 if (MaxGCPauseMillis < 1) {
228 vm_exit_during_initialization("MaxGCPauseMillis should be "
229 "greater than 0");
230 }
231 }
232 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
233 if (GCPauseIntervalMillis < 1) {
234 vm_exit_during_initialization("GCPauseIntervalMillis should be "
235 "greater than 0");
236 }
237 }
239 // Then, if the pause time target parameter was not set, set it to
240 // the default value.
241 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
242 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
243 // The default pause time target in G1 is 200ms
244 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200);
245 } else {
246 // We do not allow the pause interval to be set without the
247 // pause time target
248 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set "
249 "without setting MaxGCPauseMillis");
250 }
251 }
253 // Then, if the interval parameter was not set, set it according to
254 // the pause time target (this will also deal with the case when the
255 // pause time target is the default value).
256 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
257 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1);
258 }
260 // Finally, make sure that the two parameters are consistent.
261 if (MaxGCPauseMillis >= GCPauseIntervalMillis) {
262 char buffer[256];
263 jio_snprintf(buffer, 256,
264 "MaxGCPauseMillis (%u) should be less than "
265 "GCPauseIntervalMillis (%u)",
266 MaxGCPauseMillis, GCPauseIntervalMillis);
267 vm_exit_during_initialization(buffer);
268 }
270 double max_gc_time = (double) MaxGCPauseMillis / 1000.0;
271 double time_slice = (double) GCPauseIntervalMillis / 1000.0;
272 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time);
274 uintx confidence_perc = G1ConfidencePercent;
275 // Put an artificial ceiling on this so that it's not set to a silly value.
276 if (confidence_perc > 100) {
277 confidence_perc = 100;
278 warning("G1ConfidencePercent is set to a value that is too large, "
279 "it's been updated to %u", confidence_perc);
280 }
281 _sigma = (double) confidence_perc / 100.0;
283 // start conservatively (around 50ms is about right)
284 _concurrent_mark_remark_times_ms->add(0.05);
285 _concurrent_mark_cleanup_times_ms->add(0.20);
286 _tenuring_threshold = MaxTenuringThreshold;
287 // _max_survivor_regions will be calculated by
288 // update_young_list_target_length() during initialization.
289 _max_survivor_regions = 0;
291 assert(GCTimeRatio > 0,
292 "we should have set it to a default value set_g1_gc_flags() "
293 "if a user set it to 0");
294 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio));
296 uintx reserve_perc = G1ReservePercent;
297 // Put an artificial ceiling on this so that it's not set to a silly value.
298 if (reserve_perc > 50) {
299 reserve_perc = 50;
300 warning("G1ReservePercent is set to a value that is too large, "
301 "it's been updated to %u", reserve_perc);
302 }
303 _reserve_factor = (double) reserve_perc / 100.0;
304 // This will be set when the heap is expanded
305 // for the first time during initialization.
306 _reserve_regions = 0;
308 initialize_all();
309 _collectionSetChooser = new CollectionSetChooser();
310 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags
311 }
313 void G1CollectorPolicy::initialize_flags() {
314 set_min_alignment(HeapRegion::GrainBytes);
315 size_t card_table_alignment = GenRemSet::max_alignment_constraint(rem_set_name());
316 size_t page_size = UseLargePages ? os::large_page_size() : os::vm_page_size();
317 set_max_alignment(MAX3(card_table_alignment, min_alignment(), page_size));
318 if (SurvivorRatio < 1) {
319 vm_exit_during_initialization("Invalid survivor ratio specified");
320 }
321 CollectorPolicy::initialize_flags();
322 }
324 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true) {
325 assert(G1NewSizePercent <= G1MaxNewSizePercent, "Min larger than max");
326 assert(G1NewSizePercent > 0 && G1NewSizePercent < 100, "Min out of bounds");
327 assert(G1MaxNewSizePercent > 0 && G1MaxNewSizePercent < 100, "Max out of bounds");
329 if (FLAG_IS_CMDLINE(NewRatio)) {
330 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) {
331 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio");
332 } else {
333 _sizer_kind = SizerNewRatio;
334 _adaptive_size = false;
335 return;
336 }
337 }
339 if (FLAG_IS_CMDLINE(NewSize)) {
340 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes),
341 1U);
342 if (FLAG_IS_CMDLINE(MaxNewSize)) {
343 _max_desired_young_length =
344 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes),
345 1U);
346 _sizer_kind = SizerMaxAndNewSize;
347 _adaptive_size = _min_desired_young_length == _max_desired_young_length;
348 } else {
349 _sizer_kind = SizerNewSizeOnly;
350 }
351 } else if (FLAG_IS_CMDLINE(MaxNewSize)) {
352 _max_desired_young_length =
353 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes),
354 1U);
355 _sizer_kind = SizerMaxNewSizeOnly;
356 }
357 }
359 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) {
360 uint default_value = (new_number_of_heap_regions * G1NewSizePercent) / 100;
361 return MAX2(1U, default_value);
362 }
364 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) {
365 uint default_value = (new_number_of_heap_regions * G1MaxNewSizePercent) / 100;
366 return MAX2(1U, default_value);
367 }
369 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) {
370 assert(new_number_of_heap_regions > 0, "Heap must be initialized");
372 switch (_sizer_kind) {
373 case SizerDefaults:
374 _min_desired_young_length = calculate_default_min_length(new_number_of_heap_regions);
375 _max_desired_young_length = calculate_default_max_length(new_number_of_heap_regions);
376 break;
377 case SizerNewSizeOnly:
378 _max_desired_young_length = calculate_default_max_length(new_number_of_heap_regions);
379 _max_desired_young_length = MAX2(_min_desired_young_length, _max_desired_young_length);
380 break;
381 case SizerMaxNewSizeOnly:
382 _min_desired_young_length = calculate_default_min_length(new_number_of_heap_regions);
383 _min_desired_young_length = MIN2(_min_desired_young_length, _max_desired_young_length);
384 break;
385 case SizerMaxAndNewSize:
386 // Do nothing. Values set on the command line, don't update them at runtime.
387 break;
388 case SizerNewRatio:
389 _min_desired_young_length = new_number_of_heap_regions / (NewRatio + 1);
390 _max_desired_young_length = _min_desired_young_length;
391 break;
392 default:
393 ShouldNotReachHere();
394 }
396 assert(_min_desired_young_length <= _max_desired_young_length, "Invalid min/max young gen size values");
397 }
399 void G1CollectorPolicy::init() {
400 // Set aside an initial future to_space.
401 _g1 = G1CollectedHeap::heap();
403 assert(Heap_lock->owned_by_self(), "Locking discipline.");
405 initialize_gc_policy_counters();
407 if (adaptive_young_list_length()) {
408 _young_list_fixed_length = 0;
409 } else {
410 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length();
411 }
412 _free_regions_at_end_of_collection = _g1->free_regions();
413 update_young_list_target_length();
415 // We may immediately start allocating regions and placing them on the
416 // collection set list. Initialize the per-collection set info
417 start_incremental_cset_building();
418 }
420 // Create the jstat counters for the policy.
421 void G1CollectorPolicy::initialize_gc_policy_counters() {
422 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3);
423 }
425 bool G1CollectorPolicy::predict_will_fit(uint young_length,
426 double base_time_ms,
427 uint base_free_regions,
428 double target_pause_time_ms) {
429 if (young_length >= base_free_regions) {
430 // end condition 1: not enough space for the young regions
431 return false;
432 }
434 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1);
435 size_t bytes_to_copy =
436 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
437 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
438 double young_other_time_ms = predict_young_other_time_ms(young_length);
439 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms;
440 if (pause_time_ms > target_pause_time_ms) {
441 // end condition 2: prediction is over the target pause time
442 return false;
443 }
445 size_t free_bytes =
446 (base_free_regions - young_length) * HeapRegion::GrainBytes;
447 if ((2.0 * sigma()) * (double) bytes_to_copy > (double) free_bytes) {
448 // end condition 3: out-of-space (conservatively!)
449 return false;
450 }
452 // success!
453 return true;
454 }
456 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) {
457 // re-calculate the necessary reserve
458 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor;
459 // We use ceiling so that if reserve_regions_d is > 0.0 (but
460 // smaller than 1.0) we'll get 1.
461 _reserve_regions = (uint) ceil(reserve_regions_d);
463 _young_gen_sizer->heap_size_changed(new_number_of_regions);
464 }
466 uint G1CollectorPolicy::calculate_young_list_desired_min_length(
467 uint base_min_length) {
468 uint desired_min_length = 0;
469 if (adaptive_young_list_length()) {
470 if (_alloc_rate_ms_seq->num() > 3) {
471 double now_sec = os::elapsedTime();
472 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
473 double alloc_rate_ms = predict_alloc_rate_ms();
474 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms);
475 } else {
476 // otherwise we don't have enough info to make the prediction
477 }
478 }
479 desired_min_length += base_min_length;
480 // make sure we don't go below any user-defined minimum bound
481 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length);
482 }
484 uint G1CollectorPolicy::calculate_young_list_desired_max_length() {
485 // Here, we might want to also take into account any additional
486 // constraints (i.e., user-defined minimum bound). Currently, we
487 // effectively don't set this bound.
488 return _young_gen_sizer->max_desired_young_length();
489 }
491 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) {
492 if (rs_lengths == (size_t) -1) {
493 // if it's set to the default value (-1), we should predict it;
494 // otherwise, use the given value.
495 rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
496 }
498 // Calculate the absolute and desired min bounds.
500 // This is how many young regions we already have (currently: the survivors).
501 uint base_min_length = recorded_survivor_regions();
502 // This is the absolute minimum young length, which ensures that we
503 // can allocate one eden region in the worst-case.
504 uint absolute_min_length = base_min_length + 1;
505 uint desired_min_length =
506 calculate_young_list_desired_min_length(base_min_length);
507 if (desired_min_length < absolute_min_length) {
508 desired_min_length = absolute_min_length;
509 }
511 // Calculate the absolute and desired max bounds.
513 // We will try our best not to "eat" into the reserve.
514 uint absolute_max_length = 0;
515 if (_free_regions_at_end_of_collection > _reserve_regions) {
516 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions;
517 }
518 uint desired_max_length = calculate_young_list_desired_max_length();
519 if (desired_max_length > absolute_max_length) {
520 desired_max_length = absolute_max_length;
521 }
523 uint young_list_target_length = 0;
524 if (adaptive_young_list_length()) {
525 if (gcs_are_young()) {
526 young_list_target_length =
527 calculate_young_list_target_length(rs_lengths,
528 base_min_length,
529 desired_min_length,
530 desired_max_length);
531 _rs_lengths_prediction = rs_lengths;
532 } else {
533 // Don't calculate anything and let the code below bound it to
534 // the desired_min_length, i.e., do the next GC as soon as
535 // possible to maximize how many old regions we can add to it.
536 }
537 } else {
538 // The user asked for a fixed young gen so we'll fix the young gen
539 // whether the next GC is young or mixed.
540 young_list_target_length = _young_list_fixed_length;
541 }
543 // Make sure we don't go over the desired max length, nor under the
544 // desired min length. In case they clash, desired_min_length wins
545 // which is why that test is second.
546 if (young_list_target_length > desired_max_length) {
547 young_list_target_length = desired_max_length;
548 }
549 if (young_list_target_length < desired_min_length) {
550 young_list_target_length = desired_min_length;
551 }
553 assert(young_list_target_length > recorded_survivor_regions(),
554 "we should be able to allocate at least one eden region");
555 assert(young_list_target_length >= absolute_min_length, "post-condition");
556 _young_list_target_length = young_list_target_length;
558 update_max_gc_locker_expansion();
559 }
561 uint
562 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths,
563 uint base_min_length,
564 uint desired_min_length,
565 uint desired_max_length) {
566 assert(adaptive_young_list_length(), "pre-condition");
567 assert(gcs_are_young(), "only call this for young GCs");
569 // In case some edge-condition makes the desired max length too small...
570 if (desired_max_length <= desired_min_length) {
571 return desired_min_length;
572 }
574 // We'll adjust min_young_length and max_young_length not to include
575 // the already allocated young regions (i.e., so they reflect the
576 // min and max eden regions we'll allocate). The base_min_length
577 // will be reflected in the predictions by the
578 // survivor_regions_evac_time prediction.
579 assert(desired_min_length > base_min_length, "invariant");
580 uint min_young_length = desired_min_length - base_min_length;
581 assert(desired_max_length > base_min_length, "invariant");
582 uint max_young_length = desired_max_length - base_min_length;
584 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
585 double survivor_regions_evac_time = predict_survivor_regions_evac_time();
586 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
587 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
588 size_t scanned_cards = predict_young_card_num(adj_rs_lengths);
589 double base_time_ms =
590 predict_base_elapsed_time_ms(pending_cards, scanned_cards) +
591 survivor_regions_evac_time;
592 uint available_free_regions = _free_regions_at_end_of_collection;
593 uint base_free_regions = 0;
594 if (available_free_regions > _reserve_regions) {
595 base_free_regions = available_free_regions - _reserve_regions;
596 }
598 // Here, we will make sure that the shortest young length that
599 // makes sense fits within the target pause time.
601 if (predict_will_fit(min_young_length, base_time_ms,
602 base_free_regions, target_pause_time_ms)) {
603 // The shortest young length will fit into the target pause time;
604 // we'll now check whether the absolute maximum number of young
605 // regions will fit in the target pause time. If not, we'll do
606 // a binary search between min_young_length and max_young_length.
607 if (predict_will_fit(max_young_length, base_time_ms,
608 base_free_regions, target_pause_time_ms)) {
609 // The maximum young length will fit into the target pause time.
610 // We are done so set min young length to the maximum length (as
611 // the result is assumed to be returned in min_young_length).
612 min_young_length = max_young_length;
613 } else {
614 // The maximum possible number of young regions will not fit within
615 // the target pause time so we'll search for the optimal
616 // length. The loop invariants are:
617 //
618 // min_young_length < max_young_length
619 // min_young_length is known to fit into the target pause time
620 // max_young_length is known not to fit into the target pause time
621 //
622 // Going into the loop we know the above hold as we've just
623 // checked them. Every time around the loop we check whether
624 // the middle value between min_young_length and
625 // max_young_length fits into the target pause time. If it
626 // does, it becomes the new min. If it doesn't, it becomes
627 // the new max. This way we maintain the loop invariants.
629 assert(min_young_length < max_young_length, "invariant");
630 uint diff = (max_young_length - min_young_length) / 2;
631 while (diff > 0) {
632 uint young_length = min_young_length + diff;
633 if (predict_will_fit(young_length, base_time_ms,
634 base_free_regions, target_pause_time_ms)) {
635 min_young_length = young_length;
636 } else {
637 max_young_length = young_length;
638 }
639 assert(min_young_length < max_young_length, "invariant");
640 diff = (max_young_length - min_young_length) / 2;
641 }
642 // The results is min_young_length which, according to the
643 // loop invariants, should fit within the target pause time.
645 // These are the post-conditions of the binary search above:
646 assert(min_young_length < max_young_length,
647 "otherwise we should have discovered that max_young_length "
648 "fits into the pause target and not done the binary search");
649 assert(predict_will_fit(min_young_length, base_time_ms,
650 base_free_regions, target_pause_time_ms),
651 "min_young_length, the result of the binary search, should "
652 "fit into the pause target");
653 assert(!predict_will_fit(min_young_length + 1, base_time_ms,
654 base_free_regions, target_pause_time_ms),
655 "min_young_length, the result of the binary search, should be "
656 "optimal, so no larger length should fit into the pause target");
657 }
658 } else {
659 // Even the minimum length doesn't fit into the pause time
660 // target, return it as the result nevertheless.
661 }
662 return base_min_length + min_young_length;
663 }
665 double G1CollectorPolicy::predict_survivor_regions_evac_time() {
666 double survivor_regions_evac_time = 0.0;
667 for (HeapRegion * r = _recorded_survivor_head;
668 r != NULL && r != _recorded_survivor_tail->get_next_young_region();
669 r = r->get_next_young_region()) {
670 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, gcs_are_young());
671 }
672 return survivor_regions_evac_time;
673 }
675 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() {
676 guarantee( adaptive_young_list_length(), "should not call this otherwise" );
678 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths();
679 if (rs_lengths > _rs_lengths_prediction) {
680 // add 10% to avoid having to recalculate often
681 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000;
682 update_young_list_target_length(rs_lengths_prediction);
683 }
684 }
688 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size,
689 bool is_tlab,
690 bool* gc_overhead_limit_was_exceeded) {
691 guarantee(false, "Not using this policy feature yet.");
692 return NULL;
693 }
695 // This method controls how a collector handles one or more
696 // of its generations being fully allocated.
697 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size,
698 bool is_tlab) {
699 guarantee(false, "Not using this policy feature yet.");
700 return NULL;
701 }
704 #ifndef PRODUCT
705 bool G1CollectorPolicy::verify_young_ages() {
706 HeapRegion* head = _g1->young_list()->first_region();
707 return
708 verify_young_ages(head, _short_lived_surv_rate_group);
709 // also call verify_young_ages on any additional surv rate groups
710 }
712 bool
713 G1CollectorPolicy::verify_young_ages(HeapRegion* head,
714 SurvRateGroup *surv_rate_group) {
715 guarantee( surv_rate_group != NULL, "pre-condition" );
717 const char* name = surv_rate_group->name();
718 bool ret = true;
719 int prev_age = -1;
721 for (HeapRegion* curr = head;
722 curr != NULL;
723 curr = curr->get_next_young_region()) {
724 SurvRateGroup* group = curr->surv_rate_group();
725 if (group == NULL && !curr->is_survivor()) {
726 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name);
727 ret = false;
728 }
730 if (surv_rate_group == group) {
731 int age = curr->age_in_surv_rate_group();
733 if (age < 0) {
734 gclog_or_tty->print_cr("## %s: encountered negative age", name);
735 ret = false;
736 }
738 if (age <= prev_age) {
739 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing "
740 "(%d, %d)", name, age, prev_age);
741 ret = false;
742 }
743 prev_age = age;
744 }
745 }
747 return ret;
748 }
749 #endif // PRODUCT
751 void G1CollectorPolicy::record_full_collection_start() {
752 _full_collection_start_sec = os::elapsedTime();
753 record_heap_size_info_at_start(true /* full */);
754 // Release the future to-space so that it is available for compaction into.
755 _g1->set_full_collection();
756 }
758 void G1CollectorPolicy::record_full_collection_end() {
759 // Consider this like a collection pause for the purposes of allocation
760 // since last pause.
761 double end_sec = os::elapsedTime();
762 double full_gc_time_sec = end_sec - _full_collection_start_sec;
763 double full_gc_time_ms = full_gc_time_sec * 1000.0;
765 _trace_gen1_time_data.record_full_collection(full_gc_time_ms);
767 update_recent_gc_times(end_sec, full_gc_time_ms);
769 _g1->clear_full_collection();
771 // "Nuke" the heuristics that control the young/mixed GC
772 // transitions and make sure we start with young GCs after the Full GC.
773 set_gcs_are_young(true);
774 _last_young_gc = false;
775 clear_initiate_conc_mark_if_possible();
776 clear_during_initial_mark_pause();
777 _in_marking_window = false;
778 _in_marking_window_im = false;
780 _short_lived_surv_rate_group->start_adding_regions();
781 // also call this on any additional surv rate groups
783 record_survivor_regions(0, NULL, NULL);
785 _free_regions_at_end_of_collection = _g1->free_regions();
786 // Reset survivors SurvRateGroup.
787 _survivor_surv_rate_group->reset();
788 update_young_list_target_length();
789 _collectionSetChooser->clear();
790 }
792 void G1CollectorPolicy::record_stop_world_start() {
793 _stop_world_start = os::elapsedTime();
794 }
796 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec) {
797 // We only need to do this here as the policy will only be applied
798 // to the GC we're about to start. so, no point is calculating this
799 // every time we calculate / recalculate the target young length.
800 update_survivors_policy();
802 assert(_g1->used() == _g1->recalculate_used(),
803 err_msg("sanity, used: "SIZE_FORMAT" recalculate_used: "SIZE_FORMAT,
804 _g1->used(), _g1->recalculate_used()));
806 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0;
807 _trace_gen0_time_data.record_start_collection(s_w_t_ms);
808 _stop_world_start = 0.0;
810 record_heap_size_info_at_start(false /* full */);
812 phase_times()->record_cur_collection_start_sec(start_time_sec);
813 _pending_cards = _g1->pending_card_num();
815 _collection_set_bytes_used_before = 0;
816 _bytes_copied_during_gc = 0;
818 _last_gc_was_young = false;
820 // do that for any other surv rate groups
821 _short_lived_surv_rate_group->stop_adding_regions();
822 _survivors_age_table.clear();
824 assert( verify_young_ages(), "region age verification" );
825 }
827 void G1CollectorPolicy::record_concurrent_mark_init_end(double
828 mark_init_elapsed_time_ms) {
829 _during_marking = true;
830 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now");
831 clear_during_initial_mark_pause();
832 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms;
833 }
835 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
836 _mark_remark_start_sec = os::elapsedTime();
837 _during_marking = false;
838 }
840 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
841 double end_time_sec = os::elapsedTime();
842 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
843 _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
844 _cur_mark_stop_world_time_ms += elapsed_time_ms;
845 _prev_collection_pause_end_ms += elapsed_time_ms;
847 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true);
848 }
850 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
851 _mark_cleanup_start_sec = os::elapsedTime();
852 }
854 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
855 _last_young_gc = true;
856 _in_marking_window = false;
857 }
859 void G1CollectorPolicy::record_concurrent_pause() {
860 if (_stop_world_start > 0.0) {
861 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0;
862 _trace_gen0_time_data.record_yield_time(yield_ms);
863 }
864 }
866 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) {
867 if (_g1->concurrent_mark()->cmThread()->during_cycle()) {
868 return false;
869 }
871 size_t marking_initiating_used_threshold =
872 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent;
873 size_t cur_used_bytes = _g1->non_young_capacity_bytes();
874 size_t alloc_byte_size = alloc_word_size * HeapWordSize;
876 if ((cur_used_bytes + alloc_byte_size) > marking_initiating_used_threshold) {
877 if (gcs_are_young() && !_last_young_gc) {
878 ergo_verbose5(ErgoConcCycles,
879 "request concurrent cycle initiation",
880 ergo_format_reason("occupancy higher than threshold")
881 ergo_format_byte("occupancy")
882 ergo_format_byte("allocation request")
883 ergo_format_byte_perc("threshold")
884 ergo_format_str("source"),
885 cur_used_bytes,
886 alloc_byte_size,
887 marking_initiating_used_threshold,
888 (double) InitiatingHeapOccupancyPercent,
889 source);
890 return true;
891 } else {
892 ergo_verbose5(ErgoConcCycles,
893 "do not request concurrent cycle initiation",
894 ergo_format_reason("still doing mixed collections")
895 ergo_format_byte("occupancy")
896 ergo_format_byte("allocation request")
897 ergo_format_byte_perc("threshold")
898 ergo_format_str("source"),
899 cur_used_bytes,
900 alloc_byte_size,
901 marking_initiating_used_threshold,
902 (double) InitiatingHeapOccupancyPercent,
903 source);
904 }
905 }
907 return false;
908 }
910 // Anything below that is considered to be zero
911 #define MIN_TIMER_GRANULARITY 0.0000001
913 void G1CollectorPolicy::record_collection_pause_end(double pause_time_ms, EvacuationInfo& evacuation_info) {
914 double end_time_sec = os::elapsedTime();
915 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(),
916 "otherwise, the subtraction below does not make sense");
917 size_t rs_size =
918 _cur_collection_pause_used_regions_at_start - cset_region_length();
919 size_t cur_used_bytes = _g1->used();
920 assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
921 bool last_pause_included_initial_mark = false;
922 bool update_stats = !_g1->evacuation_failed();
924 #ifndef PRODUCT
925 if (G1YoungSurvRateVerbose) {
926 gclog_or_tty->print_cr("");
927 _short_lived_surv_rate_group->print();
928 // do that for any other surv rate groups too
929 }
930 #endif // PRODUCT
932 last_pause_included_initial_mark = during_initial_mark_pause();
933 if (last_pause_included_initial_mark) {
934 record_concurrent_mark_init_end(0.0);
935 } else if (need_to_start_conc_mark("end of GC")) {
936 // Note: this might have already been set, if during the last
937 // pause we decided to start a cycle but at the beginning of
938 // this pause we decided to postpone it. That's OK.
939 set_initiate_conc_mark_if_possible();
940 }
942 _mmu_tracker->add_pause(end_time_sec - pause_time_ms/1000.0,
943 end_time_sec, false);
945 evacuation_info.set_collectionset_used_before(_collection_set_bytes_used_before);
946 evacuation_info.set_bytes_copied(_bytes_copied_during_gc);
948 if (update_stats) {
949 _trace_gen0_time_data.record_end_collection(pause_time_ms, phase_times());
950 // this is where we update the allocation rate of the application
951 double app_time_ms =
952 (phase_times()->cur_collection_start_sec() * 1000.0 - _prev_collection_pause_end_ms);
953 if (app_time_ms < MIN_TIMER_GRANULARITY) {
954 // This usually happens due to the timer not having the required
955 // granularity. Some Linuxes are the usual culprits.
956 // We'll just set it to something (arbitrarily) small.
957 app_time_ms = 1.0;
958 }
959 // We maintain the invariant that all objects allocated by mutator
960 // threads will be allocated out of eden regions. So, we can use
961 // the eden region number allocated since the previous GC to
962 // calculate the application's allocate rate. The only exception
963 // to that is humongous objects that are allocated separately. But
964 // given that humongous object allocations do not really affect
965 // either the pause's duration nor when the next pause will take
966 // place we can safely ignore them here.
967 uint regions_allocated = eden_cset_region_length();
968 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
969 _alloc_rate_ms_seq->add(alloc_rate_ms);
971 double interval_ms =
972 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
973 update_recent_gc_times(end_time_sec, pause_time_ms);
974 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
975 if (recent_avg_pause_time_ratio() < 0.0 ||
976 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) {
977 #ifndef PRODUCT
978 // Dump info to allow post-facto debugging
979 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds");
980 gclog_or_tty->print_cr("-------------------------------------------");
981 gclog_or_tty->print_cr("Recent GC Times (ms):");
982 _recent_gc_times_ms->dump();
983 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec);
984 _recent_prev_end_times_for_all_gcs_sec->dump();
985 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f",
986 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio());
987 // In debug mode, terminate the JVM if the user wants to debug at this point.
988 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above");
989 #endif // !PRODUCT
990 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
991 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
992 if (_recent_avg_pause_time_ratio < 0.0) {
993 _recent_avg_pause_time_ratio = 0.0;
994 } else {
995 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
996 _recent_avg_pause_time_ratio = 1.0;
997 }
998 }
999 }
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 }
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.
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 }
1025 if (!_last_gc_was_young) {
1026 // This is a mixed GC. Here we decide whether to continue doing
1027 // mixed GCs or not.
1029 if (!next_gc_should_be_mixed("continue mixed GCs",
1030 "do not continue mixed GCs")) {
1031 set_gcs_are_young(true);
1032 }
1033 }
1035 _short_lived_surv_rate_group->start_adding_regions();
1036 // do that for any other surv rate groupsx
1038 if (update_stats) {
1039 double cost_per_card_ms = 0.0;
1040 if (_pending_cards > 0) {
1041 cost_per_card_ms = phase_times()->average_last_update_rs_time() / (double) _pending_cards;
1042 _cost_per_card_ms_seq->add(cost_per_card_ms);
1043 }
1045 size_t cards_scanned = _g1->cards_scanned();
1047 double cost_per_entry_ms = 0.0;
1048 if (cards_scanned > 10) {
1049 cost_per_entry_ms = phase_times()->average_last_scan_rs_time() / (double) cards_scanned;
1050 if (_last_gc_was_young) {
1051 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
1052 } else {
1053 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms);
1054 }
1055 }
1057 if (_max_rs_lengths > 0) {
1058 double cards_per_entry_ratio =
1059 (double) cards_scanned / (double) _max_rs_lengths;
1060 if (_last_gc_was_young) {
1061 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1062 } else {
1063 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1064 }
1065 }
1067 // This is defensive. For a while _max_rs_lengths could get
1068 // smaller than _recorded_rs_lengths which was causing
1069 // rs_length_diff to get very large and mess up the RSet length
1070 // predictions. The reason was unsafe concurrent updates to the
1071 // _inc_cset_recorded_rs_lengths field which the code below guards
1072 // against (see CR 7118202). This bug has now been fixed (see CR
1073 // 7119027). However, I'm still worried that
1074 // _inc_cset_recorded_rs_lengths might still end up somewhat
1075 // inaccurate. The concurrent refinement thread calculates an
1076 // RSet's length concurrently with other CR threads updating it
1077 // which might cause it to calculate the length incorrectly (if,
1078 // say, it's in mid-coarsening). So I'll leave in the defensive
1079 // conditional below just in case.
1080 size_t rs_length_diff = 0;
1081 if (_max_rs_lengths > _recorded_rs_lengths) {
1082 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths;
1083 }
1084 _rs_length_diff_seq->add((double) rs_length_diff);
1086 size_t freed_bytes = _heap_used_bytes_before_gc - cur_used_bytes;
1087 size_t copied_bytes = _collection_set_bytes_used_before - freed_bytes;
1088 double cost_per_byte_ms = 0.0;
1090 if (copied_bytes > 0) {
1091 cost_per_byte_ms = phase_times()->average_last_obj_copy_time() / (double) copied_bytes;
1092 if (_in_marking_window) {
1093 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
1094 } else {
1095 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
1096 }
1097 }
1099 double all_other_time_ms = pause_time_ms -
1100 (phase_times()->average_last_update_rs_time() + phase_times()->average_last_scan_rs_time()
1101 + phase_times()->average_last_obj_copy_time() + phase_times()->average_last_termination_time());
1103 double young_other_time_ms = 0.0;
1104 if (young_cset_region_length() > 0) {
1105 young_other_time_ms =
1106 phase_times()->young_cset_choice_time_ms() +
1107 phase_times()->young_free_cset_time_ms();
1108 _young_other_cost_per_region_ms_seq->add(young_other_time_ms /
1109 (double) young_cset_region_length());
1110 }
1111 double non_young_other_time_ms = 0.0;
1112 if (old_cset_region_length() > 0) {
1113 non_young_other_time_ms =
1114 phase_times()->non_young_cset_choice_time_ms() +
1115 phase_times()->non_young_free_cset_time_ms();
1117 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms /
1118 (double) old_cset_region_length());
1119 }
1121 double constant_other_time_ms = all_other_time_ms -
1122 (young_other_time_ms + non_young_other_time_ms);
1123 _constant_other_time_ms_seq->add(constant_other_time_ms);
1125 double survival_ratio = 0.0;
1126 if (_collection_set_bytes_used_before > 0) {
1127 survival_ratio = (double) _bytes_copied_during_gc /
1128 (double) _collection_set_bytes_used_before;
1129 }
1131 _pending_cards_seq->add((double) _pending_cards);
1132 _rs_lengths_seq->add((double) _max_rs_lengths);
1133 }
1135 _in_marking_window = new_in_marking_window;
1136 _in_marking_window_im = new_in_marking_window_im;
1137 _free_regions_at_end_of_collection = _g1->free_regions();
1138 update_young_list_target_length();
1140 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
1141 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
1142 adjust_concurrent_refinement(phase_times()->average_last_update_rs_time(),
1143 phase_times()->sum_last_update_rs_processed_buffers(), update_rs_time_goal_ms);
1145 _collectionSetChooser->verify();
1146 }
1148 #define EXT_SIZE_FORMAT "%.1f%s"
1149 #define EXT_SIZE_PARAMS(bytes) \
1150 byte_size_in_proper_unit((double)(bytes)), \
1151 proper_unit_for_byte_size((bytes))
1153 void G1CollectorPolicy::record_heap_size_info_at_start(bool full) {
1154 YoungList* young_list = _g1->young_list();
1155 _eden_used_bytes_before_gc = young_list->eden_used_bytes();
1156 _survivor_used_bytes_before_gc = young_list->survivor_used_bytes();
1157 _heap_capacity_bytes_before_gc = _g1->capacity();
1158 _heap_used_bytes_before_gc = _g1->used();
1159 _cur_collection_pause_used_regions_at_start = _g1->used_regions();
1161 _eden_capacity_bytes_before_gc =
1162 (_young_list_target_length * HeapRegion::GrainBytes) - _survivor_used_bytes_before_gc;
1164 if (full) {
1165 _metaspace_used_bytes_before_gc = MetaspaceAux::allocated_used_bytes();
1166 }
1167 }
1169 void G1CollectorPolicy::print_heap_transition() {
1170 _g1->print_size_transition(gclog_or_tty,
1171 _heap_used_bytes_before_gc,
1172 _g1->used(),
1173 _g1->capacity());
1174 }
1176 void G1CollectorPolicy::print_detailed_heap_transition(bool full) {
1177 YoungList* young_list = _g1->young_list();
1179 size_t eden_used_bytes_after_gc = young_list->eden_used_bytes();
1180 size_t survivor_used_bytes_after_gc = young_list->survivor_used_bytes();
1181 size_t heap_used_bytes_after_gc = _g1->used();
1183 size_t heap_capacity_bytes_after_gc = _g1->capacity();
1184 size_t eden_capacity_bytes_after_gc =
1185 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_used_bytes_after_gc;
1187 gclog_or_tty->print(
1188 " [Eden: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT") "
1189 "Survivors: "EXT_SIZE_FORMAT"->"EXT_SIZE_FORMAT" "
1190 "Heap: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"
1191 EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")]",
1192 EXT_SIZE_PARAMS(_eden_used_bytes_before_gc),
1193 EXT_SIZE_PARAMS(_eden_capacity_bytes_before_gc),
1194 EXT_SIZE_PARAMS(eden_used_bytes_after_gc),
1195 EXT_SIZE_PARAMS(eden_capacity_bytes_after_gc),
1196 EXT_SIZE_PARAMS(_survivor_used_bytes_before_gc),
1197 EXT_SIZE_PARAMS(survivor_used_bytes_after_gc),
1198 EXT_SIZE_PARAMS(_heap_used_bytes_before_gc),
1199 EXT_SIZE_PARAMS(_heap_capacity_bytes_before_gc),
1200 EXT_SIZE_PARAMS(heap_used_bytes_after_gc),
1201 EXT_SIZE_PARAMS(heap_capacity_bytes_after_gc));
1203 if (full) {
1204 MetaspaceAux::print_metaspace_change(_metaspace_used_bytes_before_gc);
1205 }
1207 gclog_or_tty->cr();
1208 }
1210 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1211 double update_rs_processed_buffers,
1212 double goal_ms) {
1213 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1214 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1216 if (G1UseAdaptiveConcRefinement) {
1217 const int k_gy = 3, k_gr = 6;
1218 const double inc_k = 1.1, dec_k = 0.9;
1220 int g = cg1r->green_zone();
1221 if (update_rs_time > goal_ms) {
1222 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1223 } else {
1224 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1225 g = (int)MAX2(g * inc_k, g + 1.0);
1226 }
1227 }
1228 // Change the refinement threads params
1229 cg1r->set_green_zone(g);
1230 cg1r->set_yellow_zone(g * k_gy);
1231 cg1r->set_red_zone(g * k_gr);
1232 cg1r->reinitialize_threads();
1234 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1);
1235 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1236 cg1r->yellow_zone());
1237 // Change the barrier params
1238 dcqs.set_process_completed_threshold(processing_threshold);
1239 dcqs.set_max_completed_queue(cg1r->red_zone());
1240 }
1242 int curr_queue_size = dcqs.completed_buffers_num();
1243 if (curr_queue_size >= cg1r->yellow_zone()) {
1244 dcqs.set_completed_queue_padding(curr_queue_size);
1245 } else {
1246 dcqs.set_completed_queue_padding(0);
1247 }
1248 dcqs.notify_if_necessary();
1249 }
1251 double
1252 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
1253 size_t scanned_cards) {
1254 return
1255 predict_rs_update_time_ms(pending_cards) +
1256 predict_rs_scan_time_ms(scanned_cards) +
1257 predict_constant_other_time_ms();
1258 }
1260 double
1261 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) {
1262 size_t rs_length = predict_rs_length_diff();
1263 size_t card_num;
1264 if (gcs_are_young()) {
1265 card_num = predict_young_card_num(rs_length);
1266 } else {
1267 card_num = predict_non_young_card_num(rs_length);
1268 }
1269 return predict_base_elapsed_time_ms(pending_cards, card_num);
1270 }
1272 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) {
1273 size_t bytes_to_copy;
1274 if (hr->is_marked())
1275 bytes_to_copy = hr->max_live_bytes();
1276 else {
1277 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant");
1278 int age = hr->age_in_surv_rate_group();
1279 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
1280 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate);
1281 }
1282 return bytes_to_copy;
1283 }
1285 double
1286 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
1287 bool for_young_gc) {
1288 size_t rs_length = hr->rem_set()->occupied();
1289 size_t card_num;
1291 // Predicting the number of cards is based on which type of GC
1292 // we're predicting for.
1293 if (for_young_gc) {
1294 card_num = predict_young_card_num(rs_length);
1295 } else {
1296 card_num = predict_non_young_card_num(rs_length);
1297 }
1298 size_t bytes_to_copy = predict_bytes_to_copy(hr);
1300 double region_elapsed_time_ms =
1301 predict_rs_scan_time_ms(card_num) +
1302 predict_object_copy_time_ms(bytes_to_copy);
1304 // The prediction of the "other" time for this region is based
1305 // upon the region type and NOT the GC type.
1306 if (hr->is_young()) {
1307 region_elapsed_time_ms += predict_young_other_time_ms(1);
1308 } else {
1309 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
1310 }
1311 return region_elapsed_time_ms;
1312 }
1314 void
1315 G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length,
1316 uint survivor_cset_region_length) {
1317 _eden_cset_region_length = eden_cset_region_length;
1318 _survivor_cset_region_length = survivor_cset_region_length;
1319 _old_cset_region_length = 0;
1320 }
1322 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) {
1323 _recorded_rs_lengths = rs_lengths;
1324 }
1326 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
1327 double elapsed_ms) {
1328 _recent_gc_times_ms->add(elapsed_ms);
1329 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
1330 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
1331 }
1333 size_t G1CollectorPolicy::expansion_amount() {
1334 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0;
1335 double threshold = _gc_overhead_perc;
1336 if (recent_gc_overhead > threshold) {
1337 // We will double the existing space, or take
1338 // G1ExpandByPercentOfAvailable % of the available expansion
1339 // space, whichever is smaller, bounded below by a minimum
1340 // expansion (unless that's all that's left.)
1341 const size_t min_expand_bytes = 1*M;
1342 size_t reserved_bytes = _g1->max_capacity();
1343 size_t committed_bytes = _g1->capacity();
1344 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
1345 size_t expand_bytes;
1346 size_t expand_bytes_via_pct =
1347 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
1348 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
1349 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
1350 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
1352 ergo_verbose5(ErgoHeapSizing,
1353 "attempt heap expansion",
1354 ergo_format_reason("recent GC overhead higher than "
1355 "threshold after GC")
1356 ergo_format_perc("recent GC overhead")
1357 ergo_format_perc("threshold")
1358 ergo_format_byte("uncommitted")
1359 ergo_format_byte_perc("calculated expansion amount"),
1360 recent_gc_overhead, threshold,
1361 uncommitted_bytes,
1362 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable);
1364 return expand_bytes;
1365 } else {
1366 return 0;
1367 }
1368 }
1370 void G1CollectorPolicy::print_tracing_info() const {
1371 _trace_gen0_time_data.print();
1372 _trace_gen1_time_data.print();
1373 }
1375 void G1CollectorPolicy::print_yg_surv_rate_info() const {
1376 #ifndef PRODUCT
1377 _short_lived_surv_rate_group->print_surv_rate_summary();
1378 // add this call for any other surv rate groups
1379 #endif // PRODUCT
1380 }
1382 uint G1CollectorPolicy::max_regions(int purpose) {
1383 switch (purpose) {
1384 case GCAllocForSurvived:
1385 return _max_survivor_regions;
1386 case GCAllocForTenured:
1387 return REGIONS_UNLIMITED;
1388 default:
1389 ShouldNotReachHere();
1390 return REGIONS_UNLIMITED;
1391 };
1392 }
1394 void G1CollectorPolicy::update_max_gc_locker_expansion() {
1395 uint expansion_region_num = 0;
1396 if (GCLockerEdenExpansionPercent > 0) {
1397 double perc = (double) GCLockerEdenExpansionPercent / 100.0;
1398 double expansion_region_num_d = perc * (double) _young_list_target_length;
1399 // We use ceiling so that if expansion_region_num_d is > 0.0 (but
1400 // less than 1.0) we'll get 1.
1401 expansion_region_num = (uint) ceil(expansion_region_num_d);
1402 } else {
1403 assert(expansion_region_num == 0, "sanity");
1404 }
1405 _young_list_max_length = _young_list_target_length + expansion_region_num;
1406 assert(_young_list_target_length <= _young_list_max_length, "post-condition");
1407 }
1409 // Calculates survivor space parameters.
1410 void G1CollectorPolicy::update_survivors_policy() {
1411 double max_survivor_regions_d =
1412 (double) _young_list_target_length / (double) SurvivorRatio;
1413 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but
1414 // smaller than 1.0) we'll get 1.
1415 _max_survivor_regions = (uint) ceil(max_survivor_regions_d);
1417 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
1418 HeapRegion::GrainWords * _max_survivor_regions);
1419 }
1421 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(
1422 GCCause::Cause gc_cause) {
1423 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
1424 if (!during_cycle) {
1425 ergo_verbose1(ErgoConcCycles,
1426 "request concurrent cycle initiation",
1427 ergo_format_reason("requested by GC cause")
1428 ergo_format_str("GC cause"),
1429 GCCause::to_string(gc_cause));
1430 set_initiate_conc_mark_if_possible();
1431 return true;
1432 } else {
1433 ergo_verbose1(ErgoConcCycles,
1434 "do not request concurrent cycle initiation",
1435 ergo_format_reason("concurrent cycle already in progress")
1436 ergo_format_str("GC cause"),
1437 GCCause::to_string(gc_cause));
1438 return false;
1439 }
1440 }
1442 void
1443 G1CollectorPolicy::decide_on_conc_mark_initiation() {
1444 // We are about to decide on whether this pause will be an
1445 // initial-mark pause.
1447 // First, during_initial_mark_pause() should not be already set. We
1448 // will set it here if we have to. However, it should be cleared by
1449 // the end of the pause (it's only set for the duration of an
1450 // initial-mark pause).
1451 assert(!during_initial_mark_pause(), "pre-condition");
1453 if (initiate_conc_mark_if_possible()) {
1454 // We had noticed on a previous pause that the heap occupancy has
1455 // gone over the initiating threshold and we should start a
1456 // concurrent marking cycle. So we might initiate one.
1458 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
1459 if (!during_cycle) {
1460 // The concurrent marking thread is not "during a cycle", i.e.,
1461 // it has completed the last one. So we can go ahead and
1462 // initiate a new cycle.
1464 set_during_initial_mark_pause();
1465 // We do not allow mixed GCs during marking.
1466 if (!gcs_are_young()) {
1467 set_gcs_are_young(true);
1468 ergo_verbose0(ErgoMixedGCs,
1469 "end mixed GCs",
1470 ergo_format_reason("concurrent cycle is about to start"));
1471 }
1473 // And we can now clear initiate_conc_mark_if_possible() as
1474 // we've already acted on it.
1475 clear_initiate_conc_mark_if_possible();
1477 ergo_verbose0(ErgoConcCycles,
1478 "initiate concurrent cycle",
1479 ergo_format_reason("concurrent cycle initiation requested"));
1480 } else {
1481 // The concurrent marking thread is still finishing up the
1482 // previous cycle. If we start one right now the two cycles
1483 // overlap. In particular, the concurrent marking thread might
1484 // be in the process of clearing the next marking bitmap (which
1485 // we will use for the next cycle if we start one). Starting a
1486 // cycle now will be bad given that parts of the marking
1487 // information might get cleared by the marking thread. And we
1488 // cannot wait for the marking thread to finish the cycle as it
1489 // periodically yields while clearing the next marking bitmap
1490 // and, if it's in a yield point, it's waiting for us to
1491 // finish. So, at this point we will not start a cycle and we'll
1492 // let the concurrent marking thread complete the last one.
1493 ergo_verbose0(ErgoConcCycles,
1494 "do not initiate concurrent cycle",
1495 ergo_format_reason("concurrent cycle already in progress"));
1496 }
1497 }
1498 }
1500 class KnownGarbageClosure: public HeapRegionClosure {
1501 G1CollectedHeap* _g1h;
1502 CollectionSetChooser* _hrSorted;
1504 public:
1505 KnownGarbageClosure(CollectionSetChooser* hrSorted) :
1506 _g1h(G1CollectedHeap::heap()), _hrSorted(hrSorted) { }
1508 bool doHeapRegion(HeapRegion* r) {
1509 // We only include humongous regions in collection
1510 // sets when concurrent mark shows that their contained object is
1511 // unreachable.
1513 // Do we have any marking information for this region?
1514 if (r->is_marked()) {
1515 // We will skip any region that's currently used as an old GC
1516 // alloc region (we should not consider those for collection
1517 // before we fill them up).
1518 if (_hrSorted->should_add(r) && !_g1h->is_old_gc_alloc_region(r)) {
1519 _hrSorted->add_region(r);
1520 }
1521 }
1522 return false;
1523 }
1524 };
1526 class ParKnownGarbageHRClosure: public HeapRegionClosure {
1527 G1CollectedHeap* _g1h;
1528 CSetChooserParUpdater _cset_updater;
1530 public:
1531 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
1532 uint chunk_size) :
1533 _g1h(G1CollectedHeap::heap()),
1534 _cset_updater(hrSorted, true /* parallel */, chunk_size) { }
1536 bool doHeapRegion(HeapRegion* r) {
1537 // Do we have any marking information for this region?
1538 if (r->is_marked()) {
1539 // We will skip any region that's currently used as an old GC
1540 // alloc region (we should not consider those for collection
1541 // before we fill them up).
1542 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) {
1543 _cset_updater.add_region(r);
1544 }
1545 }
1546 return false;
1547 }
1548 };
1550 class ParKnownGarbageTask: public AbstractGangTask {
1551 CollectionSetChooser* _hrSorted;
1552 uint _chunk_size;
1553 G1CollectedHeap* _g1;
1554 public:
1555 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size) :
1556 AbstractGangTask("ParKnownGarbageTask"),
1557 _hrSorted(hrSorted), _chunk_size(chunk_size),
1558 _g1(G1CollectedHeap::heap()) { }
1560 void work(uint worker_id) {
1561 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size);
1563 // Back to zero for the claim value.
1564 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, worker_id,
1565 _g1->workers()->active_workers(),
1566 HeapRegion::InitialClaimValue);
1567 }
1568 };
1570 void
1571 G1CollectorPolicy::record_concurrent_mark_cleanup_end(int no_of_gc_threads) {
1572 _collectionSetChooser->clear();
1574 uint region_num = _g1->n_regions();
1575 if (G1CollectedHeap::use_parallel_gc_threads()) {
1576 const uint OverpartitionFactor = 4;
1577 uint WorkUnit;
1578 // The use of MinChunkSize = 8 in the original code
1579 // causes some assertion failures when the total number of
1580 // region is less than 8. The code here tries to fix that.
1581 // Should the original code also be fixed?
1582 if (no_of_gc_threads > 0) {
1583 const uint MinWorkUnit = MAX2(region_num / no_of_gc_threads, 1U);
1584 WorkUnit = MAX2(region_num / (no_of_gc_threads * OverpartitionFactor),
1585 MinWorkUnit);
1586 } else {
1587 assert(no_of_gc_threads > 0,
1588 "The active gc workers should be greater than 0");
1589 // In a product build do something reasonable to avoid a crash.
1590 const uint MinWorkUnit = MAX2(region_num / (uint) ParallelGCThreads, 1U);
1591 WorkUnit =
1592 MAX2(region_num / (uint) (ParallelGCThreads * OverpartitionFactor),
1593 MinWorkUnit);
1594 }
1595 _collectionSetChooser->prepare_for_par_region_addition(_g1->n_regions(),
1596 WorkUnit);
1597 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser,
1598 (int) WorkUnit);
1599 _g1->workers()->run_task(&parKnownGarbageTask);
1601 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
1602 "sanity check");
1603 } else {
1604 KnownGarbageClosure knownGarbagecl(_collectionSetChooser);
1605 _g1->heap_region_iterate(&knownGarbagecl);
1606 }
1608 _collectionSetChooser->sort_regions();
1610 double end_sec = os::elapsedTime();
1611 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0;
1612 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
1613 _cur_mark_stop_world_time_ms += elapsed_time_ms;
1614 _prev_collection_pause_end_ms += elapsed_time_ms;
1615 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec, true);
1616 }
1618 // Add the heap region at the head of the non-incremental collection set
1619 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) {
1620 assert(_inc_cset_build_state == Active, "Precondition");
1621 assert(!hr->is_young(), "non-incremental add of young region");
1623 assert(!hr->in_collection_set(), "should not already be in the CSet");
1624 hr->set_in_collection_set(true);
1625 hr->set_next_in_collection_set(_collection_set);
1626 _collection_set = hr;
1627 _collection_set_bytes_used_before += hr->used();
1628 _g1->register_region_with_in_cset_fast_test(hr);
1629 size_t rs_length = hr->rem_set()->occupied();
1630 _recorded_rs_lengths += rs_length;
1631 _old_cset_region_length += 1;
1632 }
1634 // Initialize the per-collection-set information
1635 void G1CollectorPolicy::start_incremental_cset_building() {
1636 assert(_inc_cset_build_state == Inactive, "Precondition");
1638 _inc_cset_head = NULL;
1639 _inc_cset_tail = NULL;
1640 _inc_cset_bytes_used_before = 0;
1642 _inc_cset_max_finger = 0;
1643 _inc_cset_recorded_rs_lengths = 0;
1644 _inc_cset_recorded_rs_lengths_diffs = 0;
1645 _inc_cset_predicted_elapsed_time_ms = 0.0;
1646 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
1647 _inc_cset_build_state = Active;
1648 }
1650 void G1CollectorPolicy::finalize_incremental_cset_building() {
1651 assert(_inc_cset_build_state == Active, "Precondition");
1652 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
1654 // The two "main" fields, _inc_cset_recorded_rs_lengths and
1655 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread
1656 // that adds a new region to the CSet. Further updates by the
1657 // concurrent refinement thread that samples the young RSet lengths
1658 // are accumulated in the *_diffs fields. Here we add the diffs to
1659 // the "main" fields.
1661 if (_inc_cset_recorded_rs_lengths_diffs >= 0) {
1662 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs;
1663 } else {
1664 // This is defensive. The diff should in theory be always positive
1665 // as RSets can only grow between GCs. However, given that we
1666 // sample their size concurrently with other threads updating them
1667 // it's possible that we might get the wrong size back, which
1668 // could make the calculations somewhat inaccurate.
1669 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs);
1670 if (_inc_cset_recorded_rs_lengths >= diffs) {
1671 _inc_cset_recorded_rs_lengths -= diffs;
1672 } else {
1673 _inc_cset_recorded_rs_lengths = 0;
1674 }
1675 }
1676 _inc_cset_predicted_elapsed_time_ms +=
1677 _inc_cset_predicted_elapsed_time_ms_diffs;
1679 _inc_cset_recorded_rs_lengths_diffs = 0;
1680 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
1681 }
1683 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) {
1684 // This routine is used when:
1685 // * adding survivor regions to the incremental cset at the end of an
1686 // evacuation pause,
1687 // * adding the current allocation region to the incremental cset
1688 // when it is retired, and
1689 // * updating existing policy information for a region in the
1690 // incremental cset via young list RSet sampling.
1691 // Therefore this routine may be called at a safepoint by the
1692 // VM thread, or in-between safepoints by mutator threads (when
1693 // retiring the current allocation region) or a concurrent
1694 // refine thread (RSet sampling).
1696 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, gcs_are_young());
1697 size_t used_bytes = hr->used();
1698 _inc_cset_recorded_rs_lengths += rs_length;
1699 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms;
1700 _inc_cset_bytes_used_before += used_bytes;
1702 // Cache the values we have added to the aggregated informtion
1703 // in the heap region in case we have to remove this region from
1704 // the incremental collection set, or it is updated by the
1705 // rset sampling code
1706 hr->set_recorded_rs_length(rs_length);
1707 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms);
1708 }
1710 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr,
1711 size_t new_rs_length) {
1712 // Update the CSet information that is dependent on the new RS length
1713 assert(hr->is_young(), "Precondition");
1714 assert(!SafepointSynchronize::is_at_safepoint(),
1715 "should not be at a safepoint");
1717 // We could have updated _inc_cset_recorded_rs_lengths and
1718 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do
1719 // that atomically, as this code is executed by a concurrent
1720 // refinement thread, potentially concurrently with a mutator thread
1721 // allocating a new region and also updating the same fields. To
1722 // avoid the atomic operations we accumulate these updates on two
1723 // separate fields (*_diffs) and we'll just add them to the "main"
1724 // fields at the start of a GC.
1726 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length();
1727 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length;
1728 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff;
1730 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms();
1731 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, gcs_are_young());
1732 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms;
1733 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff;
1735 hr->set_recorded_rs_length(new_rs_length);
1736 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms);
1737 }
1739 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) {
1740 assert(hr->is_young(), "invariant");
1741 assert(hr->young_index_in_cset() > -1, "should have already been set");
1742 assert(_inc_cset_build_state == Active, "Precondition");
1744 // We need to clear and set the cached recorded/cached collection set
1745 // information in the heap region here (before the region gets added
1746 // to the collection set). An individual heap region's cached values
1747 // are calculated, aggregated with the policy collection set info,
1748 // and cached in the heap region here (initially) and (subsequently)
1749 // by the Young List sampling code.
1751 size_t rs_length = hr->rem_set()->occupied();
1752 add_to_incremental_cset_info(hr, rs_length);
1754 HeapWord* hr_end = hr->end();
1755 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end);
1757 assert(!hr->in_collection_set(), "invariant");
1758 hr->set_in_collection_set(true);
1759 assert( hr->next_in_collection_set() == NULL, "invariant");
1761 _g1->register_region_with_in_cset_fast_test(hr);
1762 }
1764 // Add the region at the RHS of the incremental cset
1765 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) {
1766 // We should only ever be appending survivors at the end of a pause
1767 assert( hr->is_survivor(), "Logic");
1769 // Do the 'common' stuff
1770 add_region_to_incremental_cset_common(hr);
1772 // Now add the region at the right hand side
1773 if (_inc_cset_tail == NULL) {
1774 assert(_inc_cset_head == NULL, "invariant");
1775 _inc_cset_head = hr;
1776 } else {
1777 _inc_cset_tail->set_next_in_collection_set(hr);
1778 }
1779 _inc_cset_tail = hr;
1780 }
1782 // Add the region to the LHS of the incremental cset
1783 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) {
1784 // Survivors should be added to the RHS at the end of a pause
1785 assert(!hr->is_survivor(), "Logic");
1787 // Do the 'common' stuff
1788 add_region_to_incremental_cset_common(hr);
1790 // Add the region at the left hand side
1791 hr->set_next_in_collection_set(_inc_cset_head);
1792 if (_inc_cset_head == NULL) {
1793 assert(_inc_cset_tail == NULL, "Invariant");
1794 _inc_cset_tail = hr;
1795 }
1796 _inc_cset_head = hr;
1797 }
1799 #ifndef PRODUCT
1800 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) {
1801 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be");
1803 st->print_cr("\nCollection_set:");
1804 HeapRegion* csr = list_head;
1805 while (csr != NULL) {
1806 HeapRegion* next = csr->next_in_collection_set();
1807 assert(csr->in_collection_set(), "bad CS");
1808 st->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d",
1809 HR_FORMAT_PARAMS(csr),
1810 csr->prev_top_at_mark_start(), csr->next_top_at_mark_start(),
1811 csr->age_in_surv_rate_group_cond());
1812 csr = next;
1813 }
1814 }
1815 #endif // !PRODUCT
1817 double G1CollectorPolicy::reclaimable_bytes_perc(size_t reclaimable_bytes) {
1818 // Returns the given amount of reclaimable bytes (that represents
1819 // the amount of reclaimable space still to be collected) as a
1820 // percentage of the current heap capacity.
1821 size_t capacity_bytes = _g1->capacity();
1822 return (double) reclaimable_bytes * 100.0 / (double) capacity_bytes;
1823 }
1825 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str,
1826 const char* false_action_str) {
1827 CollectionSetChooser* cset_chooser = _collectionSetChooser;
1828 if (cset_chooser->is_empty()) {
1829 ergo_verbose0(ErgoMixedGCs,
1830 false_action_str,
1831 ergo_format_reason("candidate old regions not available"));
1832 return false;
1833 }
1835 // Is the amount of uncollected reclaimable space above G1HeapWastePercent?
1836 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes();
1837 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes);
1838 double threshold = (double) G1HeapWastePercent;
1839 if (reclaimable_perc <= threshold) {
1840 ergo_verbose4(ErgoMixedGCs,
1841 false_action_str,
1842 ergo_format_reason("reclaimable percentage not over threshold")
1843 ergo_format_region("candidate old regions")
1844 ergo_format_byte_perc("reclaimable")
1845 ergo_format_perc("threshold"),
1846 cset_chooser->remaining_regions(),
1847 reclaimable_bytes,
1848 reclaimable_perc, threshold);
1849 return false;
1850 }
1852 ergo_verbose4(ErgoMixedGCs,
1853 true_action_str,
1854 ergo_format_reason("candidate old regions available")
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 true;
1862 }
1864 uint G1CollectorPolicy::calc_min_old_cset_length() {
1865 // The min old CSet region bound is based on the maximum desired
1866 // number of mixed GCs after a cycle. I.e., even if some old regions
1867 // look expensive, we should add them to the CSet anyway to make
1868 // sure we go through the available old regions in no more than the
1869 // maximum desired number of mixed GCs.
1870 //
1871 // The calculation is based on the number of marked regions we added
1872 // to the CSet chooser in the first place, not how many remain, so
1873 // that the result is the same during all mixed GCs that follow a cycle.
1875 const size_t region_num = (size_t) _collectionSetChooser->length();
1876 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1);
1877 size_t result = region_num / gc_num;
1878 // emulate ceiling
1879 if (result * gc_num < region_num) {
1880 result += 1;
1881 }
1882 return (uint) result;
1883 }
1885 uint G1CollectorPolicy::calc_max_old_cset_length() {
1886 // The max old CSet region bound is based on the threshold expressed
1887 // as a percentage of the heap size. I.e., it should bound the
1888 // number of old regions added to the CSet irrespective of how many
1889 // of them are available.
1891 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1892 const size_t region_num = g1h->n_regions();
1893 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent;
1894 size_t result = region_num * perc / 100;
1895 // emulate ceiling
1896 if (100 * result < region_num * perc) {
1897 result += 1;
1898 }
1899 return (uint) result;
1900 }
1903 void G1CollectorPolicy::finalize_cset(double target_pause_time_ms, EvacuationInfo& evacuation_info) {
1904 double young_start_time_sec = os::elapsedTime();
1906 YoungList* young_list = _g1->young_list();
1907 finalize_incremental_cset_building();
1909 guarantee(target_pause_time_ms > 0.0,
1910 err_msg("target_pause_time_ms = %1.6lf should be positive",
1911 target_pause_time_ms));
1912 guarantee(_collection_set == NULL, "Precondition");
1914 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
1915 double predicted_pause_time_ms = base_time_ms;
1916 double time_remaining_ms = MAX2(target_pause_time_ms - base_time_ms, 0.0);
1918 ergo_verbose4(ErgoCSetConstruction | ErgoHigh,
1919 "start choosing CSet",
1920 ergo_format_size("_pending_cards")
1921 ergo_format_ms("predicted base time")
1922 ergo_format_ms("remaining time")
1923 ergo_format_ms("target pause time"),
1924 _pending_cards, base_time_ms, time_remaining_ms, target_pause_time_ms);
1926 _last_gc_was_young = gcs_are_young() ? true : false;
1928 if (_last_gc_was_young) {
1929 _trace_gen0_time_data.increment_young_collection_count();
1930 } else {
1931 _trace_gen0_time_data.increment_mixed_collection_count();
1932 }
1934 // The young list is laid with the survivor regions from the previous
1935 // pause are appended to the RHS of the young list, i.e.
1936 // [Newly Young Regions ++ Survivors from last pause].
1938 uint survivor_region_length = young_list->survivor_length();
1939 uint eden_region_length = young_list->length() - survivor_region_length;
1940 init_cset_region_lengths(eden_region_length, survivor_region_length);
1942 HeapRegion* hr = young_list->first_survivor_region();
1943 while (hr != NULL) {
1944 assert(hr->is_survivor(), "badly formed young list");
1945 hr->set_young();
1946 hr = hr->get_next_young_region();
1947 }
1949 // Clear the fields that point to the survivor list - they are all young now.
1950 young_list->clear_survivors();
1952 _collection_set = _inc_cset_head;
1953 _collection_set_bytes_used_before = _inc_cset_bytes_used_before;
1954 time_remaining_ms = MAX2(time_remaining_ms - _inc_cset_predicted_elapsed_time_ms, 0.0);
1955 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms;
1957 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
1958 "add young regions to CSet",
1959 ergo_format_region("eden")
1960 ergo_format_region("survivors")
1961 ergo_format_ms("predicted young region time"),
1962 eden_region_length, survivor_region_length,
1963 _inc_cset_predicted_elapsed_time_ms);
1965 // The number of recorded young regions is the incremental
1966 // collection set's current size
1967 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths);
1969 double young_end_time_sec = os::elapsedTime();
1970 phase_times()->record_young_cset_choice_time_ms((young_end_time_sec - young_start_time_sec) * 1000.0);
1972 // Set the start of the non-young choice time.
1973 double non_young_start_time_sec = young_end_time_sec;
1975 if (!gcs_are_young()) {
1976 CollectionSetChooser* cset_chooser = _collectionSetChooser;
1977 cset_chooser->verify();
1978 const uint min_old_cset_length = calc_min_old_cset_length();
1979 const uint max_old_cset_length = calc_max_old_cset_length();
1981 uint expensive_region_num = 0;
1982 bool check_time_remaining = adaptive_young_list_length();
1984 HeapRegion* hr = cset_chooser->peek();
1985 while (hr != NULL) {
1986 if (old_cset_region_length() >= max_old_cset_length) {
1987 // Added maximum number of old regions to the CSet.
1988 ergo_verbose2(ErgoCSetConstruction,
1989 "finish adding old regions to CSet",
1990 ergo_format_reason("old CSet region num reached max")
1991 ergo_format_region("old")
1992 ergo_format_region("max"),
1993 old_cset_region_length(), max_old_cset_length);
1994 break;
1995 }
1998 // Stop adding regions if the remaining reclaimable space is
1999 // not above G1HeapWastePercent.
2000 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes();
2001 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes);
2002 double threshold = (double) G1HeapWastePercent;
2003 if (reclaimable_perc <= threshold) {
2004 // We've added enough old regions that the amount of uncollected
2005 // reclaimable space is at or below the waste threshold. Stop
2006 // adding old regions to the CSet.
2007 ergo_verbose5(ErgoCSetConstruction,
2008 "finish adding old regions to CSet",
2009 ergo_format_reason("reclaimable percentage not over threshold")
2010 ergo_format_region("old")
2011 ergo_format_region("max")
2012 ergo_format_byte_perc("reclaimable")
2013 ergo_format_perc("threshold"),
2014 old_cset_region_length(),
2015 max_old_cset_length,
2016 reclaimable_bytes,
2017 reclaimable_perc, threshold);
2018 break;
2019 }
2021 double predicted_time_ms = predict_region_elapsed_time_ms(hr, gcs_are_young());
2022 if (check_time_remaining) {
2023 if (predicted_time_ms > time_remaining_ms) {
2024 // Too expensive for the current CSet.
2026 if (old_cset_region_length() >= min_old_cset_length) {
2027 // We have added the minimum number of old regions to the CSet,
2028 // we are done with this CSet.
2029 ergo_verbose4(ErgoCSetConstruction,
2030 "finish adding old regions to CSet",
2031 ergo_format_reason("predicted time is too high")
2032 ergo_format_ms("predicted time")
2033 ergo_format_ms("remaining time")
2034 ergo_format_region("old")
2035 ergo_format_region("min"),
2036 predicted_time_ms, time_remaining_ms,
2037 old_cset_region_length(), min_old_cset_length);
2038 break;
2039 }
2041 // We'll add it anyway given that we haven't reached the
2042 // minimum number of old regions.
2043 expensive_region_num += 1;
2044 }
2045 } else {
2046 if (old_cset_region_length() >= min_old_cset_length) {
2047 // In the non-auto-tuning case, we'll finish adding regions
2048 // to the CSet if we reach the minimum.
2049 ergo_verbose2(ErgoCSetConstruction,
2050 "finish adding old regions to CSet",
2051 ergo_format_reason("old CSet region num reached min")
2052 ergo_format_region("old")
2053 ergo_format_region("min"),
2054 old_cset_region_length(), min_old_cset_length);
2055 break;
2056 }
2057 }
2059 // We will add this region to the CSet.
2060 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0);
2061 predicted_pause_time_ms += predicted_time_ms;
2062 cset_chooser->remove_and_move_to_next(hr);
2063 _g1->old_set_remove(hr);
2064 add_old_region_to_cset(hr);
2066 hr = cset_chooser->peek();
2067 }
2068 if (hr == NULL) {
2069 ergo_verbose0(ErgoCSetConstruction,
2070 "finish adding old regions to CSet",
2071 ergo_format_reason("candidate old regions not available"));
2072 }
2074 if (expensive_region_num > 0) {
2075 // We print the information once here at the end, predicated on
2076 // whether we added any apparently expensive regions or not, to
2077 // avoid generating output per region.
2078 ergo_verbose4(ErgoCSetConstruction,
2079 "added expensive regions to CSet",
2080 ergo_format_reason("old CSet region num not reached min")
2081 ergo_format_region("old")
2082 ergo_format_region("expensive")
2083 ergo_format_region("min")
2084 ergo_format_ms("remaining time"),
2085 old_cset_region_length(),
2086 expensive_region_num,
2087 min_old_cset_length,
2088 time_remaining_ms);
2089 }
2091 cset_chooser->verify();
2092 }
2094 stop_incremental_cset_building();
2096 ergo_verbose5(ErgoCSetConstruction,
2097 "finish choosing CSet",
2098 ergo_format_region("eden")
2099 ergo_format_region("survivors")
2100 ergo_format_region("old")
2101 ergo_format_ms("predicted pause time")
2102 ergo_format_ms("target pause time"),
2103 eden_region_length, survivor_region_length,
2104 old_cset_region_length(),
2105 predicted_pause_time_ms, target_pause_time_ms);
2107 double non_young_end_time_sec = os::elapsedTime();
2108 phase_times()->record_non_young_cset_choice_time_ms((non_young_end_time_sec - non_young_start_time_sec) * 1000.0);
2109 evacuation_info.set_collectionset_regions(cset_region_length());
2110 }
2112 void TraceGen0TimeData::record_start_collection(double time_to_stop_the_world_ms) {
2113 if(TraceGen0Time) {
2114 _all_stop_world_times_ms.add(time_to_stop_the_world_ms);
2115 }
2116 }
2118 void TraceGen0TimeData::record_yield_time(double yield_time_ms) {
2119 if(TraceGen0Time) {
2120 _all_yield_times_ms.add(yield_time_ms);
2121 }
2122 }
2124 void TraceGen0TimeData::record_end_collection(double pause_time_ms, G1GCPhaseTimes* phase_times) {
2125 if(TraceGen0Time) {
2126 _total.add(pause_time_ms);
2127 _other.add(pause_time_ms - phase_times->accounted_time_ms());
2128 _root_region_scan_wait.add(phase_times->root_region_scan_wait_time_ms());
2129 _parallel.add(phase_times->cur_collection_par_time_ms());
2130 _ext_root_scan.add(phase_times->average_last_ext_root_scan_time());
2131 _satb_filtering.add(phase_times->average_last_satb_filtering_times_ms());
2132 _update_rs.add(phase_times->average_last_update_rs_time());
2133 _scan_rs.add(phase_times->average_last_scan_rs_time());
2134 _obj_copy.add(phase_times->average_last_obj_copy_time());
2135 _termination.add(phase_times->average_last_termination_time());
2137 double parallel_known_time = phase_times->average_last_ext_root_scan_time() +
2138 phase_times->average_last_satb_filtering_times_ms() +
2139 phase_times->average_last_update_rs_time() +
2140 phase_times->average_last_scan_rs_time() +
2141 phase_times->average_last_obj_copy_time() +
2142 + phase_times->average_last_termination_time();
2144 double parallel_other_time = phase_times->cur_collection_par_time_ms() - parallel_known_time;
2145 _parallel_other.add(parallel_other_time);
2146 _clear_ct.add(phase_times->cur_clear_ct_time_ms());
2147 }
2148 }
2150 void TraceGen0TimeData::increment_young_collection_count() {
2151 if(TraceGen0Time) {
2152 ++_young_pause_num;
2153 }
2154 }
2156 void TraceGen0TimeData::increment_mixed_collection_count() {
2157 if(TraceGen0Time) {
2158 ++_mixed_pause_num;
2159 }
2160 }
2162 void TraceGen0TimeData::print_summary(const char* str,
2163 const NumberSeq* seq) const {
2164 double sum = seq->sum();
2165 gclog_or_tty->print_cr("%-27s = %8.2lf s (avg = %8.2lf ms)",
2166 str, sum / 1000.0, seq->avg());
2167 }
2169 void TraceGen0TimeData::print_summary_sd(const char* str,
2170 const NumberSeq* seq) const {
2171 print_summary(str, seq);
2172 gclog_or_tty->print_cr("%+45s = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
2173 "(num", seq->num(), seq->sd(), seq->maximum());
2174 }
2176 void TraceGen0TimeData::print() const {
2177 if (!TraceGen0Time) {
2178 return;
2179 }
2181 gclog_or_tty->print_cr("ALL PAUSES");
2182 print_summary_sd(" Total", &_total);
2183 gclog_or_tty->print_cr("");
2184 gclog_or_tty->print_cr("");
2185 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num);
2186 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num);
2187 gclog_or_tty->print_cr("");
2189 gclog_or_tty->print_cr("EVACUATION PAUSES");
2191 if (_young_pause_num == 0 && _mixed_pause_num == 0) {
2192 gclog_or_tty->print_cr("none");
2193 } else {
2194 print_summary_sd(" Evacuation Pauses", &_total);
2195 print_summary(" Root Region Scan Wait", &_root_region_scan_wait);
2196 print_summary(" Parallel Time", &_parallel);
2197 print_summary(" Ext Root Scanning", &_ext_root_scan);
2198 print_summary(" SATB Filtering", &_satb_filtering);
2199 print_summary(" Update RS", &_update_rs);
2200 print_summary(" Scan RS", &_scan_rs);
2201 print_summary(" Object Copy", &_obj_copy);
2202 print_summary(" Termination", &_termination);
2203 print_summary(" Parallel Other", &_parallel_other);
2204 print_summary(" Clear CT", &_clear_ct);
2205 print_summary(" Other", &_other);
2206 }
2207 gclog_or_tty->print_cr("");
2209 gclog_or_tty->print_cr("MISC");
2210 print_summary_sd(" Stop World", &_all_stop_world_times_ms);
2211 print_summary_sd(" Yields", &_all_yield_times_ms);
2212 }
2214 void TraceGen1TimeData::record_full_collection(double full_gc_time_ms) {
2215 if (TraceGen1Time) {
2216 _all_full_gc_times.add(full_gc_time_ms);
2217 }
2218 }
2220 void TraceGen1TimeData::print() const {
2221 if (!TraceGen1Time) {
2222 return;
2223 }
2225 if (_all_full_gc_times.num() > 0) {
2226 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s",
2227 _all_full_gc_times.num(),
2228 _all_full_gc_times.sum() / 1000.0);
2229 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg());
2230 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]",
2231 _all_full_gc_times.sd(),
2232 _all_full_gc_times.maximum());
2233 }
2234 }