1 /*
2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc_implementation/g1/concurrentG1Refine.hpp"
27 #include "gc_implementation/g1/concurrentMark.hpp"
28 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
29 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
30 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
31 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
32 #include "gc_implementation/g1/g1Log.hpp"
33 #include "gc_implementation/g1/heapRegionRemSet.hpp"
34 #include "gc_implementation/shared/gcPolicyCounters.hpp"
35 #include "runtime/arguments.hpp"
36 #include "runtime/java.hpp"
37 #include "runtime/mutexLocker.hpp"
38 #include "utilities/debug.hpp"
39
40 // Different defaults for different number of GC threads
41 // They were chosen by running GCOld and SPECjbb on debris with different
42 // numbers of GC threads and choosing them based on the results
43
44 // all the same
45 static double rs_length_diff_defaults[] = {
46 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
47 };
48
49 static double cost_per_card_ms_defaults[] = {
50 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015
51 };
52
53 // all the same
54 static double young_cards_per_entry_ratio_defaults[] = {
55 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0
56 };
57
58 static double cost_per_entry_ms_defaults[] = {
59 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005
60 };
61
62 static double cost_per_byte_ms_defaults[] = {
63 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009
64 };
65
66 // these should be pretty consistent
67 static double constant_other_time_ms_defaults[] = {
68 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0
69 };
70
71
72 static double young_other_cost_per_region_ms_defaults[] = {
73 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1
74 };
75
76 static double non_young_other_cost_per_region_ms_defaults[] = {
77 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30
78 };
79
80 // Help class for avoiding interleaved logging
81 class LineBuffer: public StackObj {
82
83 private:
84 static const int BUFFER_LEN = 1024;
85 static const int INDENT_CHARS = 3;
86 char _buffer[BUFFER_LEN];
87 int _indent_level;
88 int _cur;
89
90 void vappend(const char* format, va_list ap) {
91 int res = vsnprintf(&_buffer[_cur], BUFFER_LEN - _cur, format, ap);
92 if (res != -1) {
93 _cur += res;
94 } else {
95 DEBUG_ONLY(warning("buffer too small in LineBuffer");)
96 _buffer[BUFFER_LEN -1] = 0;
97 _cur = BUFFER_LEN; // vsnprintf above should not add to _buffer if we are called again
98 }
99 }
100
101 public:
102 explicit LineBuffer(int indent_level): _indent_level(indent_level), _cur(0) {
103 for (; (_cur < BUFFER_LEN && _cur < (_indent_level * INDENT_CHARS)); _cur++) {
104 _buffer[_cur] = ' ';
105 }
106 }
107
108 #ifndef PRODUCT
109 ~LineBuffer() {
110 assert(_cur == _indent_level * INDENT_CHARS, "pending data in buffer - append_and_print_cr() not called?");
111 }
112 #endif
113
114 void append(const char* format, ...) {
115 va_list ap;
116 va_start(ap, format);
117 vappend(format, ap);
118 va_end(ap);
119 }
120
121 void append_and_print_cr(const char* format, ...) {
122 va_list ap;
123 va_start(ap, format);
124 vappend(format, ap);
125 va_end(ap);
126 gclog_or_tty->print_cr("%s", _buffer);
127 _cur = _indent_level * INDENT_CHARS;
128 }
129 };
130
131 G1CollectorPolicy::G1CollectorPolicy() :
132 _parallel_gc_threads(G1CollectedHeap::use_parallel_gc_threads()
133 ? ParallelGCThreads : 1),
134
135 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
136 _all_pause_times_ms(new NumberSeq()),
137 _stop_world_start(0.0),
138 _all_stop_world_times_ms(new NumberSeq()),
139 _all_yield_times_ms(new NumberSeq()),
140
141 _summary(new Summary()),
142
143 _cur_clear_ct_time_ms(0.0),
144 _root_region_scan_wait_time_ms(0.0),
145
146 _cur_ref_proc_time_ms(0.0),
147 _cur_ref_enq_time_ms(0.0),
148
149 #ifndef PRODUCT
150 _min_clear_cc_time_ms(-1.0),
151 _max_clear_cc_time_ms(-1.0),
152 _cur_clear_cc_time_ms(0.0),
153 _cum_clear_cc_time_ms(0.0),
154 _num_cc_clears(0L),
155 #endif
156
157 _aux_num(10),
158 _all_aux_times_ms(new NumberSeq[_aux_num]),
159 _cur_aux_start_times_ms(new double[_aux_num]),
160 _cur_aux_times_ms(new double[_aux_num]),
161 _cur_aux_times_set(new bool[_aux_num]),
162
163 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
164 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
165
166 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
167 _prev_collection_pause_end_ms(0.0),
168 _pending_card_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
169 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
170 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
171 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
172 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
173 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
174 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
175 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
176 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)),
177 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
178 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
179 _non_young_other_cost_per_region_ms_seq(
180 new TruncatedSeq(TruncatedSeqLength)),
181
182 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)),
183 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)),
184
185 _pause_time_target_ms((double) MaxGCPauseMillis),
186
187 _gcs_are_young(true),
188 _young_pause_num(0),
189 _mixed_pause_num(0),
190
191 _during_marking(false),
192 _in_marking_window(false),
193 _in_marking_window_im(false),
194
195 _recent_prev_end_times_for_all_gcs_sec(
196 new TruncatedSeq(NumPrevPausesForHeuristics)),
197
198 _recent_avg_pause_time_ratio(0.0),
199
200 _all_full_gc_times_ms(new NumberSeq()),
201
202 _initiate_conc_mark_if_possible(false),
203 _during_initial_mark_pause(false),
204 _last_young_gc(false),
205 _last_gc_was_young(false),
206
207 _eden_bytes_before_gc(0),
208 _survivor_bytes_before_gc(0),
209 _capacity_before_gc(0),
210
211 _eden_cset_region_length(0),
212 _survivor_cset_region_length(0),
213 _old_cset_region_length(0),
214
215 _collection_set(NULL),
216 _collection_set_bytes_used_before(0),
217
218 // Incremental CSet attributes
219 _inc_cset_build_state(Inactive),
220 _inc_cset_head(NULL),
221 _inc_cset_tail(NULL),
222 _inc_cset_bytes_used_before(0),
223 _inc_cset_max_finger(NULL),
224 _inc_cset_recorded_rs_lengths(0),
225 _inc_cset_recorded_rs_lengths_diffs(0),
226 _inc_cset_predicted_elapsed_time_ms(0.0),
227 _inc_cset_predicted_elapsed_time_ms_diffs(0.0),
228
229 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
230 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
231 #endif // _MSC_VER
232
233 _short_lived_surv_rate_group(new SurvRateGroup(this, "Short Lived",
234 G1YoungSurvRateNumRegionsSummary)),
235 _survivor_surv_rate_group(new SurvRateGroup(this, "Survivor",
236 G1YoungSurvRateNumRegionsSummary)),
237 // add here any more surv rate groups
238 _recorded_survivor_regions(0),
239 _recorded_survivor_head(NULL),
240 _recorded_survivor_tail(NULL),
241 _survivors_age_table(true),
242
243 _gc_overhead_perc(0.0) {
244
245 // Set up the region size and associated fields. Given that the
246 // policy is created before the heap, we have to set this up here,
247 // so it's done as soon as possible.
248 HeapRegion::setup_heap_region_size(Arguments::min_heap_size());
249 HeapRegionRemSet::setup_remset_size();
250
251 G1ErgoVerbose::initialize();
252 if (PrintAdaptiveSizePolicy) {
253 // Currently, we only use a single switch for all the heuristics.
254 G1ErgoVerbose::set_enabled(true);
255 // Given that we don't currently have a verboseness level
256 // parameter, we'll hardcode this to high. This can be easily
257 // changed in the future.
258 G1ErgoVerbose::set_level(ErgoHigh);
259 } else {
260 G1ErgoVerbose::set_enabled(false);
261 }
262
263 // Verify PLAB sizes
264 const size_t region_size = HeapRegion::GrainWords;
265 if (YoungPLABSize > region_size || OldPLABSize > region_size) {
266 char buffer[128];
267 jio_snprintf(buffer, sizeof(buffer), "%sPLABSize should be at most "SIZE_FORMAT,
268 OldPLABSize > region_size ? "Old" : "Young", region_size);
269 vm_exit_during_initialization(buffer);
270 }
271
272 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime());
273 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0;
274
275 _par_last_gc_worker_start_times_ms = new double[_parallel_gc_threads];
276 _par_last_ext_root_scan_times_ms = new double[_parallel_gc_threads];
277 _par_last_satb_filtering_times_ms = new double[_parallel_gc_threads];
278
279 _par_last_update_rs_times_ms = new double[_parallel_gc_threads];
280 _par_last_update_rs_processed_buffers = new double[_parallel_gc_threads];
281
282 _par_last_scan_rs_times_ms = new double[_parallel_gc_threads];
283
284 _par_last_obj_copy_times_ms = new double[_parallel_gc_threads];
285
286 _par_last_termination_times_ms = new double[_parallel_gc_threads];
287 _par_last_termination_attempts = new double[_parallel_gc_threads];
288 _par_last_gc_worker_end_times_ms = new double[_parallel_gc_threads];
289 _par_last_gc_worker_times_ms = new double[_parallel_gc_threads];
290 _par_last_gc_worker_other_times_ms = new double[_parallel_gc_threads];
291
292 int index;
293 if (ParallelGCThreads == 0)
294 index = 0;
295 else if (ParallelGCThreads > 8)
296 index = 7;
297 else
298 index = ParallelGCThreads - 1;
299
300 _pending_card_diff_seq->add(0.0);
301 _rs_length_diff_seq->add(rs_length_diff_defaults[index]);
302 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]);
303 _young_cards_per_entry_ratio_seq->add(
304 young_cards_per_entry_ratio_defaults[index]);
305 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]);
306 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]);
307 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]);
308 _young_other_cost_per_region_ms_seq->add(
309 young_other_cost_per_region_ms_defaults[index]);
310 _non_young_other_cost_per_region_ms_seq->add(
311 non_young_other_cost_per_region_ms_defaults[index]);
312
313 // Below, we might need to calculate the pause time target based on
314 // the pause interval. When we do so we are going to give G1 maximum
315 // flexibility and allow it to do pauses when it needs to. So, we'll
316 // arrange that the pause interval to be pause time target + 1 to
317 // ensure that a) the pause time target is maximized with respect to
318 // the pause interval and b) we maintain the invariant that pause
319 // time target < pause interval. If the user does not want this
320 // maximum flexibility, they will have to set the pause interval
321 // explicitly.
322
323 // First make sure that, if either parameter is set, its value is
324 // reasonable.
325 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
326 if (MaxGCPauseMillis < 1) {
327 vm_exit_during_initialization("MaxGCPauseMillis should be "
328 "greater than 0");
329 }
330 }
331 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
332 if (GCPauseIntervalMillis < 1) {
333 vm_exit_during_initialization("GCPauseIntervalMillis should be "
334 "greater than 0");
335 }
336 }
337
338 // Then, if the pause time target parameter was not set, set it to
339 // the default value.
340 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
341 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
342 // The default pause time target in G1 is 200ms
343 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200);
344 } else {
345 // We do not allow the pause interval to be set without the
346 // pause time target
347 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set "
348 "without setting MaxGCPauseMillis");
349 }
350 }
351
352 // Then, if the interval parameter was not set, set it according to
353 // the pause time target (this will also deal with the case when the
354 // pause time target is the default value).
355 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
356 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1);
357 }
358
359 // Finally, make sure that the two parameters are consistent.
360 if (MaxGCPauseMillis >= GCPauseIntervalMillis) {
361 char buffer[256];
362 jio_snprintf(buffer, 256,
363 "MaxGCPauseMillis (%u) should be less than "
364 "GCPauseIntervalMillis (%u)",
365 MaxGCPauseMillis, GCPauseIntervalMillis);
366 vm_exit_during_initialization(buffer);
367 }
368
369 double max_gc_time = (double) MaxGCPauseMillis / 1000.0;
370 double time_slice = (double) GCPauseIntervalMillis / 1000.0;
371 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time);
372 _sigma = (double) G1ConfidencePercent / 100.0;
373
374 // start conservatively (around 50ms is about right)
375 _concurrent_mark_remark_times_ms->add(0.05);
376 _concurrent_mark_cleanup_times_ms->add(0.20);
377 _tenuring_threshold = MaxTenuringThreshold;
378 // _max_survivor_regions will be calculated by
379 // update_young_list_target_length() during initialization.
380 _max_survivor_regions = 0;
381
382 assert(GCTimeRatio > 0,
383 "we should have set it to a default value set_g1_gc_flags() "
384 "if a user set it to 0");
385 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio));
386
387 uintx reserve_perc = G1ReservePercent;
388 // Put an artificial ceiling on this so that it's not set to a silly value.
389 if (reserve_perc > 50) {
390 reserve_perc = 50;
391 warning("G1ReservePercent is set to a value that is too large, "
392 "it's been updated to %u", reserve_perc);
393 }
394 _reserve_factor = (double) reserve_perc / 100.0;
395 // This will be set when the heap is expanded
396 // for the first time during initialization.
397 _reserve_regions = 0;
398
399 initialize_all();
400 _collectionSetChooser = new CollectionSetChooser();
401 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags
402 }
403
404 void G1CollectorPolicy::initialize_flags() {
405 set_min_alignment(HeapRegion::GrainBytes);
406 set_max_alignment(GenRemSet::max_alignment_constraint(rem_set_name()));
407 if (SurvivorRatio < 1) {
408 vm_exit_during_initialization("Invalid survivor ratio specified");
409 }
410 CollectorPolicy::initialize_flags();
411 }
412
413 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true) {
414 assert(G1DefaultMinNewGenPercent <= G1DefaultMaxNewGenPercent, "Min larger than max");
415 assert(G1DefaultMinNewGenPercent > 0 && G1DefaultMinNewGenPercent < 100, "Min out of bounds");
416 assert(G1DefaultMaxNewGenPercent > 0 && G1DefaultMaxNewGenPercent < 100, "Max out of bounds");
417
418 if (FLAG_IS_CMDLINE(NewRatio)) {
419 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) {
420 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio");
421 } else {
422 _sizer_kind = SizerNewRatio;
423 _adaptive_size = false;
424 return;
425 }
426 }
427
428 if (FLAG_IS_CMDLINE(NewSize)) {
429 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes),
430 1U);
431 if (FLAG_IS_CMDLINE(MaxNewSize)) {
432 _max_desired_young_length =
433 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes),
434 1U);
435 _sizer_kind = SizerMaxAndNewSize;
436 _adaptive_size = _min_desired_young_length == _max_desired_young_length;
437 } else {
438 _sizer_kind = SizerNewSizeOnly;
439 }
440 } else if (FLAG_IS_CMDLINE(MaxNewSize)) {
441 _max_desired_young_length =
442 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes),
443 1U);
444 _sizer_kind = SizerMaxNewSizeOnly;
445 }
446 }
447
448 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) {
449 uint default_value = (new_number_of_heap_regions * G1DefaultMinNewGenPercent) / 100;
450 return MAX2(1U, default_value);
451 }
452
453 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) {
454 uint default_value = (new_number_of_heap_regions * G1DefaultMaxNewGenPercent) / 100;
455 return MAX2(1U, default_value);
456 }
457
458 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) {
459 assert(new_number_of_heap_regions > 0, "Heap must be initialized");
460
461 switch (_sizer_kind) {
462 case SizerDefaults:
463 _min_desired_young_length = calculate_default_min_length(new_number_of_heap_regions);
464 _max_desired_young_length = calculate_default_max_length(new_number_of_heap_regions);
465 break;
466 case SizerNewSizeOnly:
467 _max_desired_young_length = calculate_default_max_length(new_number_of_heap_regions);
468 _max_desired_young_length = MAX2(_min_desired_young_length, _max_desired_young_length);
469 break;
470 case SizerMaxNewSizeOnly:
471 _min_desired_young_length = calculate_default_min_length(new_number_of_heap_regions);
472 _min_desired_young_length = MIN2(_min_desired_young_length, _max_desired_young_length);
473 break;
474 case SizerMaxAndNewSize:
475 // Do nothing. Values set on the command line, don't update them at runtime.
476 break;
477 case SizerNewRatio:
478 _min_desired_young_length = new_number_of_heap_regions / (NewRatio + 1);
479 _max_desired_young_length = _min_desired_young_length;
480 break;
481 default:
482 ShouldNotReachHere();
483 }
484
485 assert(_min_desired_young_length <= _max_desired_young_length, "Invalid min/max young gen size values");
486 }
487
488 void G1CollectorPolicy::init() {
489 // Set aside an initial future to_space.
490 _g1 = G1CollectedHeap::heap();
491
492 assert(Heap_lock->owned_by_self(), "Locking discipline.");
493
494 initialize_gc_policy_counters();
495
496 if (adaptive_young_list_length()) {
497 _young_list_fixed_length = 0;
498 } else {
499 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length();
500 }
501 _free_regions_at_end_of_collection = _g1->free_regions();
502 update_young_list_target_length();
503 _prev_eden_capacity = _young_list_target_length * HeapRegion::GrainBytes;
504
505 // We may immediately start allocating regions and placing them on the
506 // collection set list. Initialize the per-collection set info
507 start_incremental_cset_building();
508 }
509
510 // Create the jstat counters for the policy.
511 void G1CollectorPolicy::initialize_gc_policy_counters() {
512 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3);
513 }
514
515 bool G1CollectorPolicy::predict_will_fit(uint young_length,
516 double base_time_ms,
517 uint base_free_regions,
518 double target_pause_time_ms) {
519 if (young_length >= base_free_regions) {
520 // end condition 1: not enough space for the young regions
521 return false;
522 }
523
524 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1);
525 size_t bytes_to_copy =
526 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
527 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
528 double young_other_time_ms = predict_young_other_time_ms(young_length);
529 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms;
530 if (pause_time_ms > target_pause_time_ms) {
531 // end condition 2: prediction is over the target pause time
532 return false;
533 }
534
535 size_t free_bytes =
536 (base_free_regions - young_length) * HeapRegion::GrainBytes;
537 if ((2.0 * sigma()) * (double) bytes_to_copy > (double) free_bytes) {
538 // end condition 3: out-of-space (conservatively!)
539 return false;
540 }
541
542 // success!
543 return true;
544 }
545
546 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) {
547 // re-calculate the necessary reserve
548 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor;
549 // We use ceiling so that if reserve_regions_d is > 0.0 (but
550 // smaller than 1.0) we'll get 1.
551 _reserve_regions = (uint) ceil(reserve_regions_d);
552
553 _young_gen_sizer->heap_size_changed(new_number_of_regions);
554 }
555
556 uint G1CollectorPolicy::calculate_young_list_desired_min_length(
557 uint base_min_length) {
558 uint desired_min_length = 0;
559 if (adaptive_young_list_length()) {
560 if (_alloc_rate_ms_seq->num() > 3) {
561 double now_sec = os::elapsedTime();
562 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
563 double alloc_rate_ms = predict_alloc_rate_ms();
564 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms);
565 } else {
566 // otherwise we don't have enough info to make the prediction
567 }
568 }
569 desired_min_length += base_min_length;
570 // make sure we don't go below any user-defined minimum bound
571 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length);
572 }
573
574 uint G1CollectorPolicy::calculate_young_list_desired_max_length() {
575 // Here, we might want to also take into account any additional
576 // constraints (i.e., user-defined minimum bound). Currently, we
577 // effectively don't set this bound.
578 return _young_gen_sizer->max_desired_young_length();
579 }
580
581 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) {
582 if (rs_lengths == (size_t) -1) {
583 // if it's set to the default value (-1), we should predict it;
584 // otherwise, use the given value.
585 rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
586 }
587
588 // Calculate the absolute and desired min bounds.
589
590 // This is how many young regions we already have (currently: the survivors).
591 uint base_min_length = recorded_survivor_regions();
592 // This is the absolute minimum young length, which ensures that we
593 // can allocate one eden region in the worst-case.
594 uint absolute_min_length = base_min_length + 1;
595 uint desired_min_length =
596 calculate_young_list_desired_min_length(base_min_length);
597 if (desired_min_length < absolute_min_length) {
598 desired_min_length = absolute_min_length;
599 }
600
601 // Calculate the absolute and desired max bounds.
602
603 // We will try our best not to "eat" into the reserve.
604 uint absolute_max_length = 0;
605 if (_free_regions_at_end_of_collection > _reserve_regions) {
606 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions;
607 }
608 uint desired_max_length = calculate_young_list_desired_max_length();
609 if (desired_max_length > absolute_max_length) {
610 desired_max_length = absolute_max_length;
611 }
612
613 uint young_list_target_length = 0;
614 if (adaptive_young_list_length()) {
615 if (gcs_are_young()) {
616 young_list_target_length =
617 calculate_young_list_target_length(rs_lengths,
618 base_min_length,
619 desired_min_length,
620 desired_max_length);
621 _rs_lengths_prediction = rs_lengths;
622 } else {
623 // Don't calculate anything and let the code below bound it to
624 // the desired_min_length, i.e., do the next GC as soon as
625 // possible to maximize how many old regions we can add to it.
626 }
627 } else {
628 // The user asked for a fixed young gen so we'll fix the young gen
629 // whether the next GC is young or mixed.
630 young_list_target_length = _young_list_fixed_length;
631 }
632
633 // Make sure we don't go over the desired max length, nor under the
634 // desired min length. In case they clash, desired_min_length wins
635 // which is why that test is second.
636 if (young_list_target_length > desired_max_length) {
637 young_list_target_length = desired_max_length;
638 }
639 if (young_list_target_length < desired_min_length) {
640 young_list_target_length = desired_min_length;
641 }
642
643 assert(young_list_target_length > recorded_survivor_regions(),
644 "we should be able to allocate at least one eden region");
645 assert(young_list_target_length >= absolute_min_length, "post-condition");
646 _young_list_target_length = young_list_target_length;
647
648 update_max_gc_locker_expansion();
649 }
650
651 uint
652 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths,
653 uint base_min_length,
654 uint desired_min_length,
655 uint desired_max_length) {
656 assert(adaptive_young_list_length(), "pre-condition");
657 assert(gcs_are_young(), "only call this for young GCs");
658
659 // In case some edge-condition makes the desired max length too small...
660 if (desired_max_length <= desired_min_length) {
661 return desired_min_length;
662 }
663
664 // We'll adjust min_young_length and max_young_length not to include
665 // the already allocated young regions (i.e., so they reflect the
666 // min and max eden regions we'll allocate). The base_min_length
667 // will be reflected in the predictions by the
668 // survivor_regions_evac_time prediction.
669 assert(desired_min_length > base_min_length, "invariant");
670 uint min_young_length = desired_min_length - base_min_length;
671 assert(desired_max_length > base_min_length, "invariant");
672 uint max_young_length = desired_max_length - base_min_length;
673
674 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
675 double survivor_regions_evac_time = predict_survivor_regions_evac_time();
676 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
677 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
678 size_t scanned_cards = predict_young_card_num(adj_rs_lengths);
679 double base_time_ms =
680 predict_base_elapsed_time_ms(pending_cards, scanned_cards) +
681 survivor_regions_evac_time;
682 uint available_free_regions = _free_regions_at_end_of_collection;
683 uint base_free_regions = 0;
684 if (available_free_regions > _reserve_regions) {
685 base_free_regions = available_free_regions - _reserve_regions;
686 }
687
688 // Here, we will make sure that the shortest young length that
689 // makes sense fits within the target pause time.
690
691 if (predict_will_fit(min_young_length, base_time_ms,
692 base_free_regions, target_pause_time_ms)) {
693 // The shortest young length will fit into the target pause time;
694 // we'll now check whether the absolute maximum number of young
695 // regions will fit in the target pause time. If not, we'll do
696 // a binary search between min_young_length and max_young_length.
697 if (predict_will_fit(max_young_length, base_time_ms,
698 base_free_regions, target_pause_time_ms)) {
699 // The maximum young length will fit into the target pause time.
700 // We are done so set min young length to the maximum length (as
701 // the result is assumed to be returned in min_young_length).
702 min_young_length = max_young_length;
703 } else {
704 // The maximum possible number of young regions will not fit within
705 // the target pause time so we'll search for the optimal
706 // length. The loop invariants are:
707 //
708 // min_young_length < max_young_length
709 // min_young_length is known to fit into the target pause time
710 // max_young_length is known not to fit into the target pause time
711 //
712 // Going into the loop we know the above hold as we've just
713 // checked them. Every time around the loop we check whether
714 // the middle value between min_young_length and
715 // max_young_length fits into the target pause time. If it
716 // does, it becomes the new min. If it doesn't, it becomes
717 // the new max. This way we maintain the loop invariants.
718
719 assert(min_young_length < max_young_length, "invariant");
720 uint diff = (max_young_length - min_young_length) / 2;
721 while (diff > 0) {
722 uint young_length = min_young_length + diff;
723 if (predict_will_fit(young_length, base_time_ms,
724 base_free_regions, target_pause_time_ms)) {
725 min_young_length = young_length;
726 } else {
727 max_young_length = young_length;
728 }
729 assert(min_young_length < max_young_length, "invariant");
730 diff = (max_young_length - min_young_length) / 2;
731 }
732 // The results is min_young_length which, according to the
733 // loop invariants, should fit within the target pause time.
734
735 // These are the post-conditions of the binary search above:
736 assert(min_young_length < max_young_length,
737 "otherwise we should have discovered that max_young_length "
738 "fits into the pause target and not done the binary search");
739 assert(predict_will_fit(min_young_length, base_time_ms,
740 base_free_regions, target_pause_time_ms),
741 "min_young_length, the result of the binary search, should "
742 "fit into the pause target");
743 assert(!predict_will_fit(min_young_length + 1, base_time_ms,
744 base_free_regions, target_pause_time_ms),
745 "min_young_length, the result of the binary search, should be "
746 "optimal, so no larger length should fit into the pause target");
747 }
748 } else {
749 // Even the minimum length doesn't fit into the pause time
750 // target, return it as the result nevertheless.
751 }
752 return base_min_length + min_young_length;
753 }
754
755 double G1CollectorPolicy::predict_survivor_regions_evac_time() {
756 double survivor_regions_evac_time = 0.0;
757 for (HeapRegion * r = _recorded_survivor_head;
758 r != NULL && r != _recorded_survivor_tail->get_next_young_region();
759 r = r->get_next_young_region()) {
760 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, true);
761 }
762 return survivor_regions_evac_time;
763 }
764
765 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() {
766 guarantee( adaptive_young_list_length(), "should not call this otherwise" );
767
768 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths();
769 if (rs_lengths > _rs_lengths_prediction) {
770 // add 10% to avoid having to recalculate often
771 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000;
772 update_young_list_target_length(rs_lengths_prediction);
773 }
774 }
775
776
777
778 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size,
779 bool is_tlab,
780 bool* gc_overhead_limit_was_exceeded) {
781 guarantee(false, "Not using this policy feature yet.");
782 return NULL;
783 }
784
785 // This method controls how a collector handles one or more
786 // of its generations being fully allocated.
787 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size,
788 bool is_tlab) {
789 guarantee(false, "Not using this policy feature yet.");
790 return NULL;
791 }
792
793
794 #ifndef PRODUCT
795 bool G1CollectorPolicy::verify_young_ages() {
796 HeapRegion* head = _g1->young_list()->first_region();
797 return
798 verify_young_ages(head, _short_lived_surv_rate_group);
799 // also call verify_young_ages on any additional surv rate groups
800 }
801
802 bool
803 G1CollectorPolicy::verify_young_ages(HeapRegion* head,
804 SurvRateGroup *surv_rate_group) {
805 guarantee( surv_rate_group != NULL, "pre-condition" );
806
807 const char* name = surv_rate_group->name();
808 bool ret = true;
809 int prev_age = -1;
810
811 for (HeapRegion* curr = head;
812 curr != NULL;
813 curr = curr->get_next_young_region()) {
814 SurvRateGroup* group = curr->surv_rate_group();
815 if (group == NULL && !curr->is_survivor()) {
816 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name);
817 ret = false;
818 }
819
820 if (surv_rate_group == group) {
821 int age = curr->age_in_surv_rate_group();
822
823 if (age < 0) {
824 gclog_or_tty->print_cr("## %s: encountered negative age", name);
825 ret = false;
826 }
827
828 if (age <= prev_age) {
829 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing "
830 "(%d, %d)", name, age, prev_age);
831 ret = false;
832 }
833 prev_age = age;
834 }
835 }
836
837 return ret;
838 }
839 #endif // PRODUCT
840
841 void G1CollectorPolicy::record_full_collection_start() {
842 _cur_collection_start_sec = os::elapsedTime();
843 // Release the future to-space so that it is available for compaction into.
844 _g1->set_full_collection();
845 }
846
847 void G1CollectorPolicy::record_full_collection_end() {
848 // Consider this like a collection pause for the purposes of allocation
849 // since last pause.
850 double end_sec = os::elapsedTime();
851 double full_gc_time_sec = end_sec - _cur_collection_start_sec;
852 double full_gc_time_ms = full_gc_time_sec * 1000.0;
853
854 _all_full_gc_times_ms->add(full_gc_time_ms);
855
856 update_recent_gc_times(end_sec, full_gc_time_ms);
857
858 _g1->clear_full_collection();
859
860 // "Nuke" the heuristics that control the young/mixed GC
861 // transitions and make sure we start with young GCs after the Full GC.
862 set_gcs_are_young(true);
863 _last_young_gc = false;
864 clear_initiate_conc_mark_if_possible();
865 clear_during_initial_mark_pause();
866 _in_marking_window = false;
867 _in_marking_window_im = false;
868
869 _short_lived_surv_rate_group->start_adding_regions();
870 // also call this on any additional surv rate groups
871
872 record_survivor_regions(0, NULL, NULL);
873
874 _free_regions_at_end_of_collection = _g1->free_regions();
875 // Reset survivors SurvRateGroup.
876 _survivor_surv_rate_group->reset();
877 update_young_list_target_length();
878 _collectionSetChooser->clear();
879 }
880
881 void G1CollectorPolicy::record_stop_world_start() {
882 _stop_world_start = os::elapsedTime();
883 }
884
885 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec,
886 size_t start_used) {
887 if (G1Log::finer()) {
888 gclog_or_tty->stamp(PrintGCTimeStamps);
889 gclog_or_tty->print("[GC pause");
890 gclog_or_tty->print(" (%s)", gcs_are_young() ? "young" : "mixed");
891 }
892
893 // We only need to do this here as the policy will only be applied
894 // to the GC we're about to start. so, no point is calculating this
895 // every time we calculate / recalculate the target young length.
896 update_survivors_policy();
897
898 assert(_g1->used() == _g1->recalculate_used(),
899 err_msg("sanity, used: "SIZE_FORMAT" recalculate_used: "SIZE_FORMAT,
900 _g1->used(), _g1->recalculate_used()));
901
902 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0;
903 _all_stop_world_times_ms->add(s_w_t_ms);
904 _stop_world_start = 0.0;
905
906 _cur_collection_start_sec = start_time_sec;
907 _cur_collection_pause_used_at_start_bytes = start_used;
908 _cur_collection_pause_used_regions_at_start = _g1->used_regions();
909 _pending_cards = _g1->pending_card_num();
910 _max_pending_cards = _g1->max_pending_card_num();
911
912 _bytes_in_collection_set_before_gc = 0;
913 _bytes_copied_during_gc = 0;
914
915 YoungList* young_list = _g1->young_list();
916 _eden_bytes_before_gc = young_list->eden_used_bytes();
917 _survivor_bytes_before_gc = young_list->survivor_used_bytes();
918 _capacity_before_gc = _g1->capacity();
919
920 #ifdef DEBUG
921 // initialise these to something well known so that we can spot
922 // if they are not set properly
923
924 for (int i = 0; i < _parallel_gc_threads; ++i) {
925 _par_last_gc_worker_start_times_ms[i] = -1234.0;
926 _par_last_ext_root_scan_times_ms[i] = -1234.0;
927 _par_last_satb_filtering_times_ms[i] = -1234.0;
928 _par_last_update_rs_times_ms[i] = -1234.0;
929 _par_last_update_rs_processed_buffers[i] = -1234.0;
930 _par_last_scan_rs_times_ms[i] = -1234.0;
931 _par_last_obj_copy_times_ms[i] = -1234.0;
932 _par_last_termination_times_ms[i] = -1234.0;
933 _par_last_termination_attempts[i] = -1234.0;
934 _par_last_gc_worker_end_times_ms[i] = -1234.0;
935 _par_last_gc_worker_times_ms[i] = -1234.0;
936 _par_last_gc_worker_other_times_ms[i] = -1234.0;
937 }
938 #endif
939
940 for (int i = 0; i < _aux_num; ++i) {
941 _cur_aux_times_ms[i] = 0.0;
942 _cur_aux_times_set[i] = false;
943 }
944
945 // This is initialized to zero here and is set during the evacuation
946 // pause if we actually waited for the root region scanning to finish.
947 _root_region_scan_wait_time_ms = 0.0;
948
949 _last_gc_was_young = false;
950
951 // do that for any other surv rate groups
952 _short_lived_surv_rate_group->stop_adding_regions();
953 _survivors_age_table.clear();
954
955 assert( verify_young_ages(), "region age verification" );
956 }
957
958 void G1CollectorPolicy::record_concurrent_mark_init_end(double
959 mark_init_elapsed_time_ms) {
960 _during_marking = true;
961 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now");
962 clear_during_initial_mark_pause();
963 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms;
964 }
965
966 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
967 _mark_remark_start_sec = os::elapsedTime();
968 _during_marking = false;
969 }
970
971 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
972 double end_time_sec = os::elapsedTime();
973 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
974 _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
975 _cur_mark_stop_world_time_ms += elapsed_time_ms;
976 _prev_collection_pause_end_ms += elapsed_time_ms;
977
978 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true);
979 }
980
981 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
982 _mark_cleanup_start_sec = os::elapsedTime();
983 }
984
985 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
986 _last_young_gc = true;
987 _in_marking_window = false;
988 }
989
990 void G1CollectorPolicy::record_concurrent_pause() {
991 if (_stop_world_start > 0.0) {
992 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0;
993 _all_yield_times_ms->add(yield_ms);
994 }
995 }
996
997 void G1CollectorPolicy::record_concurrent_pause_end() {
998 }
999
1000 template<class T>
1001 T sum_of(T* sum_arr, int start, int n, int N) {
1002 T sum = (T)0;
1003 for (int i = 0; i < n; i++) {
1004 int j = (start + i) % N;
1005 sum += sum_arr[j];
1006 }
1007 return sum;
1008 }
1009
1010 void G1CollectorPolicy::print_par_stats(int level,
1011 const char* str,
1012 double* data) {
1013 double min = data[0], max = data[0];
1014 double total = 0.0;
1015 LineBuffer buf(level);
1016 buf.append("[%s (ms):", str);
1017 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1018 double val = data[i];
1019 if (val < min)
1020 min = val;
1021 if (val > max)
1022 max = val;
1023 total += val;
1024 if (G1Log::finest()) {
1025 buf.append(" %.1lf", val);
1026 }
1027 }
1028
1029 if (G1Log::finest()) {
1030 buf.append_and_print_cr("");
1031 }
1032 double avg = total / (double) no_of_gc_threads();
1033 buf.append_and_print_cr(" Avg: %.1lf Min: %.1lf Max: %.1lf Diff: %.1lf]",
1034 avg, min, max, max - min);
1035 }
1036
1037 void G1CollectorPolicy::print_par_sizes(int level,
1038 const char* str,
1039 double* data) {
1040 double min = data[0], max = data[0];
1041 double total = 0.0;
1042 LineBuffer buf(level);
1043 buf.append("[%s :", str);
1044 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1045 double val = data[i];
1046 if (val < min)
1047 min = val;
1048 if (val > max)
1049 max = val;
1050 total += val;
1051 buf.append(" %d", (int) val);
1052 }
1053 buf.append_and_print_cr("");
1054 double avg = total / (double) no_of_gc_threads();
1055 buf.append_and_print_cr(" Sum: %d, Avg: %d, Min: %d, Max: %d, Diff: %d]",
1056 (int)total, (int)avg, (int)min, (int)max, (int)max - (int)min);
1057 }
1058
1059 void G1CollectorPolicy::print_stats(int level,
1060 const char* str,
1061 double value) {
1062 LineBuffer(level).append_and_print_cr("[%s: %5.1lf ms]", str, value);
1063 }
1064
1065 void G1CollectorPolicy::print_stats(int level,
1066 const char* str,
1067 int value) {
1068 LineBuffer(level).append_and_print_cr("[%s: %d]", str, value);
1069 }
1070
1071 double G1CollectorPolicy::avg_value(double* data) {
1072 if (G1CollectedHeap::use_parallel_gc_threads()) {
1073 double ret = 0.0;
1074 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1075 ret += data[i];
1076 }
1077 return ret / (double) no_of_gc_threads();
1078 } else {
1079 return data[0];
1080 }
1081 }
1082
1083 double G1CollectorPolicy::max_value(double* data) {
1084 if (G1CollectedHeap::use_parallel_gc_threads()) {
1085 double ret = data[0];
1086 for (uint i = 1; i < no_of_gc_threads(); ++i) {
1087 if (data[i] > ret) {
1088 ret = data[i];
1089 }
1090 }
1091 return ret;
1092 } else {
1093 return data[0];
1094 }
1095 }
1096
1097 double G1CollectorPolicy::sum_of_values(double* data) {
1098 if (G1CollectedHeap::use_parallel_gc_threads()) {
1099 double sum = 0.0;
1100 for (uint i = 0; i < no_of_gc_threads(); i++) {
1101 sum += data[i];
1102 }
1103 return sum;
1104 } else {
1105 return data[0];
1106 }
1107 }
1108
1109 double G1CollectorPolicy::max_sum(double* data1, double* data2) {
1110 double ret = data1[0] + data2[0];
1111
1112 if (G1CollectedHeap::use_parallel_gc_threads()) {
1113 for (uint i = 1; i < no_of_gc_threads(); ++i) {
1114 double data = data1[i] + data2[i];
1115 if (data > ret) {
1116 ret = data;
1117 }
1118 }
1119 }
1120 return ret;
1121 }
1122
1123 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) {
1124 if (_g1->concurrent_mark()->cmThread()->during_cycle()) {
1125 return false;
1126 }
1127
1128 size_t marking_initiating_used_threshold =
1129 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent;
1130 size_t cur_used_bytes = _g1->non_young_capacity_bytes();
1131 size_t alloc_byte_size = alloc_word_size * HeapWordSize;
1132
1133 if ((cur_used_bytes + alloc_byte_size) > marking_initiating_used_threshold) {
1134 if (gcs_are_young()) {
1135 ergo_verbose5(ErgoConcCycles,
1136 "request concurrent cycle initiation",
1137 ergo_format_reason("occupancy higher than threshold")
1138 ergo_format_byte("occupancy")
1139 ergo_format_byte("allocation request")
1140 ergo_format_byte_perc("threshold")
1141 ergo_format_str("source"),
1142 cur_used_bytes,
1143 alloc_byte_size,
1144 marking_initiating_used_threshold,
1145 (double) InitiatingHeapOccupancyPercent,
1146 source);
1147 return true;
1148 } else {
1149 ergo_verbose5(ErgoConcCycles,
1150 "do not request concurrent cycle initiation",
1151 ergo_format_reason("still doing mixed collections")
1152 ergo_format_byte("occupancy")
1153 ergo_format_byte("allocation request")
1154 ergo_format_byte_perc("threshold")
1155 ergo_format_str("source"),
1156 cur_used_bytes,
1157 alloc_byte_size,
1158 marking_initiating_used_threshold,
1159 (double) InitiatingHeapOccupancyPercent,
1160 source);
1161 }
1162 }
1163
1164 return false;
1165 }
1166
1167 // Anything below that is considered to be zero
1168 #define MIN_TIMER_GRANULARITY 0.0000001
1169
1170 void G1CollectorPolicy::record_collection_pause_end(int no_of_gc_threads) {
1171 double end_time_sec = os::elapsedTime();
1172 double elapsed_ms = _last_pause_time_ms;
1173 bool parallel = G1CollectedHeap::use_parallel_gc_threads();
1174 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(),
1175 "otherwise, the subtraction below does not make sense");
1176 size_t rs_size =
1177 _cur_collection_pause_used_regions_at_start - cset_region_length();
1178 size_t cur_used_bytes = _g1->used();
1179 assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
1180 bool last_pause_included_initial_mark = false;
1181 bool update_stats = !_g1->evacuation_failed();
1182 set_no_of_gc_threads(no_of_gc_threads);
1183
1184 #ifndef PRODUCT
1185 if (G1YoungSurvRateVerbose) {
1186 gclog_or_tty->print_cr("");
1187 _short_lived_surv_rate_group->print();
1188 // do that for any other surv rate groups too
1189 }
1190 #endif // PRODUCT
1191
1192 last_pause_included_initial_mark = during_initial_mark_pause();
1193 if (last_pause_included_initial_mark) {
1194 record_concurrent_mark_init_end(0.0);
1195 } else if (!_last_young_gc && need_to_start_conc_mark("end of GC")) {
1196 // Note: this might have already been set, if during the last
1197 // pause we decided to start a cycle but at the beginning of
1198 // this pause we decided to postpone it. That's OK.
1199 set_initiate_conc_mark_if_possible();
1200 }
1201
1202 _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0,
1203 end_time_sec, false);
1204
1205 // This assert is exempted when we're doing parallel collection pauses,
1206 // because the fragmentation caused by the parallel GC allocation buffers
1207 // can lead to more memory being used during collection than was used
1208 // before. Best leave this out until the fragmentation problem is fixed.
1209 // Pauses in which evacuation failed can also lead to negative
1210 // collections, since no space is reclaimed from a region containing an
1211 // object whose evacuation failed.
1212 // Further, we're now always doing parallel collection. But I'm still
1213 // leaving this here as a placeholder for a more precise assertion later.
1214 // (DLD, 10/05.)
1215 assert((true || parallel) // Always using GC LABs now.
1216 || _g1->evacuation_failed()
1217 || _cur_collection_pause_used_at_start_bytes >= cur_used_bytes,
1218 "Negative collection");
1219
1220 size_t freed_bytes =
1221 _cur_collection_pause_used_at_start_bytes - cur_used_bytes;
1222 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes;
1223
1224 double survival_fraction =
1225 (double)surviving_bytes/
1226 (double)_collection_set_bytes_used_before;
1227
1228 // These values are used to update the summary information that is
1229 // displayed when TraceGen0Time is enabled, and are output as part
1230 // of the "finer" output, in the non-parallel case.
1231
1232 double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms);
1233 double satb_filtering_time = avg_value(_par_last_satb_filtering_times_ms);
1234 double update_rs_time = avg_value(_par_last_update_rs_times_ms);
1235 double update_rs_processed_buffers =
1236 sum_of_values(_par_last_update_rs_processed_buffers);
1237 double scan_rs_time = avg_value(_par_last_scan_rs_times_ms);
1238 double obj_copy_time = avg_value(_par_last_obj_copy_times_ms);
1239 double termination_time = avg_value(_par_last_termination_times_ms);
1240
1241 double known_time = ext_root_scan_time +
1242 satb_filtering_time +
1243 update_rs_time +
1244 scan_rs_time +
1245 obj_copy_time;
1246
1247 double other_time_ms = elapsed_ms;
1248
1249 // Subtract the root region scanning wait time. It's initialized to
1250 // zero at the start of the pause.
1251 other_time_ms -= _root_region_scan_wait_time_ms;
1252
1253 if (parallel) {
1254 other_time_ms -= _cur_collection_par_time_ms;
1255 } else {
1256 other_time_ms -= known_time;
1257 }
1258
1259 // Now subtract the time taken to fix up roots in generated code
1260 other_time_ms -= _cur_collection_code_root_fixup_time_ms;
1261
1262 // Subtract the time taken to clean the card table from the
1263 // current value of "other time"
1264 other_time_ms -= _cur_clear_ct_time_ms;
1265
1266 // TraceGen0Time and TraceGen1Time summary info updating.
1267 _all_pause_times_ms->add(elapsed_ms);
1268
1269 if (update_stats) {
1270 _summary->record_total_time_ms(elapsed_ms);
1271 _summary->record_other_time_ms(other_time_ms);
1272
1273 MainBodySummary* body_summary = _summary->main_body_summary();
1274 assert(body_summary != NULL, "should not be null!");
1275
1276 body_summary->record_root_region_scan_wait_time_ms(
1277 _root_region_scan_wait_time_ms);
1278 body_summary->record_ext_root_scan_time_ms(ext_root_scan_time);
1279 body_summary->record_satb_filtering_time_ms(satb_filtering_time);
1280 body_summary->record_update_rs_time_ms(update_rs_time);
1281 body_summary->record_scan_rs_time_ms(scan_rs_time);
1282 body_summary->record_obj_copy_time_ms(obj_copy_time);
1283
1284 if (parallel) {
1285 body_summary->record_parallel_time_ms(_cur_collection_par_time_ms);
1286 body_summary->record_termination_time_ms(termination_time);
1287
1288 double parallel_known_time = known_time + termination_time;
1289 double parallel_other_time = _cur_collection_par_time_ms - parallel_known_time;
1290 body_summary->record_parallel_other_time_ms(parallel_other_time);
1291 }
1292
1293 body_summary->record_clear_ct_time_ms(_cur_clear_ct_time_ms);
1294
1295 // We exempt parallel collection from this check because Alloc Buffer
1296 // fragmentation can produce negative collections. Same with evac
1297 // failure.
1298 // Further, we're now always doing parallel collection. But I'm still
1299 // leaving this here as a placeholder for a more precise assertion later.
1300 // (DLD, 10/05.
1301 assert((true || parallel)
1302 || _g1->evacuation_failed()
1303 || surviving_bytes <= _collection_set_bytes_used_before,
1304 "Or else negative collection!");
1305
1306 // this is where we update the allocation rate of the application
1307 double app_time_ms =
1308 (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms);
1309 if (app_time_ms < MIN_TIMER_GRANULARITY) {
1310 // This usually happens due to the timer not having the required
1311 // granularity. Some Linuxes are the usual culprits.
1312 // We'll just set it to something (arbitrarily) small.
1313 app_time_ms = 1.0;
1314 }
1315 // We maintain the invariant that all objects allocated by mutator
1316 // threads will be allocated out of eden regions. So, we can use
1317 // the eden region number allocated since the previous GC to
1318 // calculate the application's allocate rate. The only exception
1319 // to that is humongous objects that are allocated separately. But
1320 // given that humongous object allocations do not really affect
1321 // either the pause's duration nor when the next pause will take
1322 // place we can safely ignore them here.
1323 uint regions_allocated = eden_cset_region_length();
1324 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
1325 _alloc_rate_ms_seq->add(alloc_rate_ms);
1326
1327 double interval_ms =
1328 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
1329 update_recent_gc_times(end_time_sec, elapsed_ms);
1330 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
1331 if (recent_avg_pause_time_ratio() < 0.0 ||
1332 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) {
1333 #ifndef PRODUCT
1334 // Dump info to allow post-facto debugging
1335 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds");
1336 gclog_or_tty->print_cr("-------------------------------------------");
1337 gclog_or_tty->print_cr("Recent GC Times (ms):");
1338 _recent_gc_times_ms->dump();
1339 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec);
1340 _recent_prev_end_times_for_all_gcs_sec->dump();
1341 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f",
1342 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio());
1343 // In debug mode, terminate the JVM if the user wants to debug at this point.
1344 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above");
1345 #endif // !PRODUCT
1346 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
1347 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
1348 if (_recent_avg_pause_time_ratio < 0.0) {
1349 _recent_avg_pause_time_ratio = 0.0;
1350 } else {
1351 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
1352 _recent_avg_pause_time_ratio = 1.0;
1353 }
1354 }
1355 }
1356
1357 for (int i = 0; i < _aux_num; ++i) {
1358 if (_cur_aux_times_set[i]) {
1359 _all_aux_times_ms[i].add(_cur_aux_times_ms[i]);
1360 }
1361 }
1362
1363 if (G1Log::finer()) {
1364 bool print_marking_info =
1365 _g1->mark_in_progress() && !last_pause_included_initial_mark;
1366
1367 gclog_or_tty->print_cr("%s, %1.8lf secs]",
1368 (last_pause_included_initial_mark) ? " (initial-mark)" : "",
1369 elapsed_ms / 1000.0);
1370
1371 if (_root_region_scan_wait_time_ms > 0.0) {
1372 print_stats(1, "Root Region Scan Waiting", _root_region_scan_wait_time_ms);
1373 }
1374 if (parallel) {
1375 print_stats(1, "Parallel Time", _cur_collection_par_time_ms);
1376 print_par_stats(2, "GC Worker Start", _par_last_gc_worker_start_times_ms);
1377 print_par_stats(2, "Ext Root Scanning", _par_last_ext_root_scan_times_ms);
1378 if (print_marking_info) {
1379 print_par_stats(2, "SATB Filtering", _par_last_satb_filtering_times_ms);
1380 }
1381 print_par_stats(2, "Update RS", _par_last_update_rs_times_ms);
1382 if (G1Log::finest()) {
1383 print_par_sizes(3, "Processed Buffers", _par_last_update_rs_processed_buffers);
1384 }
1385 print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms);
1386 print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms);
1387 print_par_stats(2, "Termination", _par_last_termination_times_ms);
1388 if (G1Log::finest()) {
1389 print_par_sizes(3, "Termination Attempts", _par_last_termination_attempts);
1390 }
1391
1392 for (int i = 0; i < _parallel_gc_threads; i++) {
1393 _par_last_gc_worker_times_ms[i] = _par_last_gc_worker_end_times_ms[i] -
1394 _par_last_gc_worker_start_times_ms[i];
1395
1396 double worker_known_time = _par_last_ext_root_scan_times_ms[i] +
1397 _par_last_satb_filtering_times_ms[i] +
1398 _par_last_update_rs_times_ms[i] +
1399 _par_last_scan_rs_times_ms[i] +
1400 _par_last_obj_copy_times_ms[i] +
1401 _par_last_termination_times_ms[i];
1402
1403 _par_last_gc_worker_other_times_ms[i] = _par_last_gc_worker_times_ms[i] -
1404 worker_known_time;
1405 }
1406
1407 print_par_stats(2, "GC Worker Other", _par_last_gc_worker_other_times_ms);
1408 print_par_stats(2, "GC Worker Total", _par_last_gc_worker_times_ms);
1409 print_par_stats(2, "GC Worker End", _par_last_gc_worker_end_times_ms);
1410 } else {
1411 print_stats(1, "Ext Root Scanning", ext_root_scan_time);
1412 if (print_marking_info) {
1413 print_stats(1, "SATB Filtering", satb_filtering_time);
1414 }
1415 print_stats(1, "Update RS", update_rs_time);
1416 if (G1Log::finest()) {
1417 print_stats(2, "Processed Buffers", (int)update_rs_processed_buffers);
1418 }
1419 print_stats(1, "Scan RS", scan_rs_time);
1420 print_stats(1, "Object Copying", obj_copy_time);
1421 }
1422 print_stats(1, "Code Root Fixup", _cur_collection_code_root_fixup_time_ms);
1423 print_stats(1, "Clear CT", _cur_clear_ct_time_ms);
1424 #ifndef PRODUCT
1425 print_stats(1, "Cur Clear CC", _cur_clear_cc_time_ms);
1426 print_stats(1, "Cum Clear CC", _cum_clear_cc_time_ms);
1427 print_stats(1, "Min Clear CC", _min_clear_cc_time_ms);
1428 print_stats(1, "Max Clear CC", _max_clear_cc_time_ms);
1429 if (_num_cc_clears > 0) {
1430 print_stats(1, "Avg Clear CC", _cum_clear_cc_time_ms / ((double)_num_cc_clears));
1431 }
1432 #endif
1433 print_stats(1, "Other", other_time_ms);
1434 print_stats(2, "Choose CSet",
1435 (_recorded_young_cset_choice_time_ms +
1436 _recorded_non_young_cset_choice_time_ms));
1437 print_stats(2, "Ref Proc", _cur_ref_proc_time_ms);
1438 print_stats(2, "Ref Enq", _cur_ref_enq_time_ms);
1439 print_stats(2, "Free CSet",
1440 (_recorded_young_free_cset_time_ms +
1441 _recorded_non_young_free_cset_time_ms));
1442
1443 for (int i = 0; i < _aux_num; ++i) {
1444 if (_cur_aux_times_set[i]) {
1445 char buffer[96];
1446 sprintf(buffer, "Aux%d", i);
1447 print_stats(1, buffer, _cur_aux_times_ms[i]);
1448 }
1449 }
1450 }
1451
1452 bool new_in_marking_window = _in_marking_window;
1453 bool new_in_marking_window_im = false;
1454 if (during_initial_mark_pause()) {
1455 new_in_marking_window = true;
1456 new_in_marking_window_im = true;
1457 }
1458
1459 if (_last_young_gc) {
1460 // This is supposed to to be the "last young GC" before we start
1461 // doing mixed GCs. Here we decide whether to start mixed GCs or not.
1462
1463 if (!last_pause_included_initial_mark) {
1464 if (next_gc_should_be_mixed("start mixed GCs",
1465 "do not start mixed GCs")) {
1466 set_gcs_are_young(false);
1467 }
1468 } else {
1469 ergo_verbose0(ErgoMixedGCs,
1470 "do not start mixed GCs",
1471 ergo_format_reason("concurrent cycle is about to start"));
1472 }
1473 _last_young_gc = false;
1474 }
1475
1476 if (!_last_gc_was_young) {
1477 // This is a mixed GC. Here we decide whether to continue doing
1478 // mixed GCs or not.
1479
1480 if (!next_gc_should_be_mixed("continue mixed GCs",
1481 "do not continue mixed GCs")) {
1482 set_gcs_are_young(true);
1483 }
1484 }
1485
1486 _short_lived_surv_rate_group->start_adding_regions();
1487 // do that for any other surv rate groupsx
1488
1489 if (update_stats) {
1490 double pause_time_ms = elapsed_ms;
1491
1492 size_t diff = 0;
1493 if (_max_pending_cards >= _pending_cards) {
1494 diff = _max_pending_cards - _pending_cards;
1495 }
1496 _pending_card_diff_seq->add((double) diff);
1497
1498 double cost_per_card_ms = 0.0;
1499 if (_pending_cards > 0) {
1500 cost_per_card_ms = update_rs_time / (double) _pending_cards;
1501 _cost_per_card_ms_seq->add(cost_per_card_ms);
1502 }
1503
1504 size_t cards_scanned = _g1->cards_scanned();
1505
1506 double cost_per_entry_ms = 0.0;
1507 if (cards_scanned > 10) {
1508 cost_per_entry_ms = scan_rs_time / (double) cards_scanned;
1509 if (_last_gc_was_young) {
1510 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
1511 } else {
1512 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms);
1513 }
1514 }
1515
1516 if (_max_rs_lengths > 0) {
1517 double cards_per_entry_ratio =
1518 (double) cards_scanned / (double) _max_rs_lengths;
1519 if (_last_gc_was_young) {
1520 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1521 } else {
1522 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1523 }
1524 }
1525
1526 // This is defensive. For a while _max_rs_lengths could get
1527 // smaller than _recorded_rs_lengths which was causing
1528 // rs_length_diff to get very large and mess up the RSet length
1529 // predictions. The reason was unsafe concurrent updates to the
1530 // _inc_cset_recorded_rs_lengths field which the code below guards
1531 // against (see CR 7118202). This bug has now been fixed (see CR
1532 // 7119027). However, I'm still worried that
1533 // _inc_cset_recorded_rs_lengths might still end up somewhat
1534 // inaccurate. The concurrent refinement thread calculates an
1535 // RSet's length concurrently with other CR threads updating it
1536 // which might cause it to calculate the length incorrectly (if,
1537 // say, it's in mid-coarsening). So I'll leave in the defensive
1538 // conditional below just in case.
1539 size_t rs_length_diff = 0;
1540 if (_max_rs_lengths > _recorded_rs_lengths) {
1541 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths;
1542 }
1543 _rs_length_diff_seq->add((double) rs_length_diff);
1544
1545 size_t copied_bytes = surviving_bytes;
1546 double cost_per_byte_ms = 0.0;
1547 if (copied_bytes > 0) {
1548 cost_per_byte_ms = obj_copy_time / (double) copied_bytes;
1549 if (_in_marking_window) {
1550 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
1551 } else {
1552 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
1553 }
1554 }
1555
1556 double all_other_time_ms = pause_time_ms -
1557 (update_rs_time + scan_rs_time + obj_copy_time + termination_time);
1558
1559 double young_other_time_ms = 0.0;
1560 if (young_cset_region_length() > 0) {
1561 young_other_time_ms =
1562 _recorded_young_cset_choice_time_ms +
1563 _recorded_young_free_cset_time_ms;
1564 _young_other_cost_per_region_ms_seq->add(young_other_time_ms /
1565 (double) young_cset_region_length());
1566 }
1567 double non_young_other_time_ms = 0.0;
1568 if (old_cset_region_length() > 0) {
1569 non_young_other_time_ms =
1570 _recorded_non_young_cset_choice_time_ms +
1571 _recorded_non_young_free_cset_time_ms;
1572
1573 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms /
1574 (double) old_cset_region_length());
1575 }
1576
1577 double constant_other_time_ms = all_other_time_ms -
1578 (young_other_time_ms + non_young_other_time_ms);
1579 _constant_other_time_ms_seq->add(constant_other_time_ms);
1580
1581 double survival_ratio = 0.0;
1582 if (_bytes_in_collection_set_before_gc > 0) {
1583 survival_ratio = (double) _bytes_copied_during_gc /
1584 (double) _bytes_in_collection_set_before_gc;
1585 }
1586
1587 _pending_cards_seq->add((double) _pending_cards);
1588 _rs_lengths_seq->add((double) _max_rs_lengths);
1589 }
1590
1591 _in_marking_window = new_in_marking_window;
1592 _in_marking_window_im = new_in_marking_window_im;
1593 _free_regions_at_end_of_collection = _g1->free_regions();
1594 update_young_list_target_length();
1595
1596 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
1597 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
1598 adjust_concurrent_refinement(update_rs_time, update_rs_processed_buffers, update_rs_time_goal_ms);
1599
1600 _collectionSetChooser->verify();
1601 }
1602
1603 #define EXT_SIZE_FORMAT "%d%s"
1604 #define EXT_SIZE_PARAMS(bytes) \
1605 byte_size_in_proper_unit((bytes)), \
1606 proper_unit_for_byte_size((bytes))
1607
1608 void G1CollectorPolicy::print_heap_transition() {
1609 if (G1Log::finer()) {
1610 YoungList* young_list = _g1->young_list();
1611 size_t eden_bytes = young_list->eden_used_bytes();
1612 size_t survivor_bytes = young_list->survivor_used_bytes();
1613 size_t used_before_gc = _cur_collection_pause_used_at_start_bytes;
1614 size_t used = _g1->used();
1615 size_t capacity = _g1->capacity();
1616 size_t eden_capacity =
1617 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_bytes;
1618
1619 gclog_or_tty->print_cr(
1620 " [Eden: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT") "
1621 "Survivors: "EXT_SIZE_FORMAT"->"EXT_SIZE_FORMAT" "
1622 "Heap: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"
1623 EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")]",
1624 EXT_SIZE_PARAMS(_eden_bytes_before_gc),
1625 EXT_SIZE_PARAMS(_prev_eden_capacity),
1626 EXT_SIZE_PARAMS(eden_bytes),
1627 EXT_SIZE_PARAMS(eden_capacity),
1628 EXT_SIZE_PARAMS(_survivor_bytes_before_gc),
1629 EXT_SIZE_PARAMS(survivor_bytes),
1630 EXT_SIZE_PARAMS(used_before_gc),
1631 EXT_SIZE_PARAMS(_capacity_before_gc),
1632 EXT_SIZE_PARAMS(used),
1633 EXT_SIZE_PARAMS(capacity));
1634
1635 _prev_eden_capacity = eden_capacity;
1636 } else if (G1Log::fine()) {
1637 _g1->print_size_transition(gclog_or_tty,
1638 _cur_collection_pause_used_at_start_bytes,
1639 _g1->used(), _g1->capacity());
1640 }
1641 }
1642
1643 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1644 double update_rs_processed_buffers,
1645 double goal_ms) {
1646 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1647 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1648
1649 if (G1UseAdaptiveConcRefinement) {
1650 const int k_gy = 3, k_gr = 6;
1651 const double inc_k = 1.1, dec_k = 0.9;
1652
1653 int g = cg1r->green_zone();
1654 if (update_rs_time > goal_ms) {
1655 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1656 } else {
1657 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1658 g = (int)MAX2(g * inc_k, g + 1.0);
1659 }
1660 }
1661 // Change the refinement threads params
1662 cg1r->set_green_zone(g);
1663 cg1r->set_yellow_zone(g * k_gy);
1664 cg1r->set_red_zone(g * k_gr);
1665 cg1r->reinitialize_threads();
1666
1667 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1);
1668 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1669 cg1r->yellow_zone());
1670 // Change the barrier params
1671 dcqs.set_process_completed_threshold(processing_threshold);
1672 dcqs.set_max_completed_queue(cg1r->red_zone());
1673 }
1674
1675 int curr_queue_size = dcqs.completed_buffers_num();
1676 if (curr_queue_size >= cg1r->yellow_zone()) {
1677 dcqs.set_completed_queue_padding(curr_queue_size);
1678 } else {
1679 dcqs.set_completed_queue_padding(0);
1680 }
1681 dcqs.notify_if_necessary();
1682 }
1683
1684 double
1685 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) {
1686 size_t rs_length = predict_rs_length_diff();
1687 size_t card_num;
1688 if (gcs_are_young()) {
1689 card_num = predict_young_card_num(rs_length);
1690 } else {
1691 card_num = predict_non_young_card_num(rs_length);
1692 }
1693 return predict_base_elapsed_time_ms(pending_cards, card_num);
1694 }
1695
1696 double
1697 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
1698 size_t scanned_cards) {
1699 return
1700 predict_rs_update_time_ms(pending_cards) +
1701 predict_rs_scan_time_ms(scanned_cards) +
1702 predict_constant_other_time_ms();
1703 }
1704
1705 double
1706 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
1707 bool young) {
1708 size_t rs_length = hr->rem_set()->occupied();
1709 size_t card_num;
1710 if (gcs_are_young()) {
1711 card_num = predict_young_card_num(rs_length);
1712 } else {
1713 card_num = predict_non_young_card_num(rs_length);
1714 }
1715 size_t bytes_to_copy = predict_bytes_to_copy(hr);
1716
1717 double region_elapsed_time_ms =
1718 predict_rs_scan_time_ms(card_num) +
1719 predict_object_copy_time_ms(bytes_to_copy);
1720
1721 if (young)
1722 region_elapsed_time_ms += predict_young_other_time_ms(1);
1723 else
1724 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
1725
1726 return region_elapsed_time_ms;
1727 }
1728
1729 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) {
1730 size_t bytes_to_copy;
1731 if (hr->is_marked())
1732 bytes_to_copy = hr->max_live_bytes();
1733 else {
1734 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant");
1735 int age = hr->age_in_surv_rate_group();
1736 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
1737 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate);
1738 }
1739 return bytes_to_copy;
1740 }
1741
1742 void
1743 G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length,
1744 uint survivor_cset_region_length) {
1745 _eden_cset_region_length = eden_cset_region_length;
1746 _survivor_cset_region_length = survivor_cset_region_length;
1747 _old_cset_region_length = 0;
1748 }
1749
1750 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) {
1751 _recorded_rs_lengths = rs_lengths;
1752 }
1753
1754 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
1755 double elapsed_ms) {
1756 _recent_gc_times_ms->add(elapsed_ms);
1757 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
1758 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
1759 }
1760
1761 size_t G1CollectorPolicy::expansion_amount() {
1762 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0;
1763 double threshold = _gc_overhead_perc;
1764 if (recent_gc_overhead > threshold) {
1765 // We will double the existing space, or take
1766 // G1ExpandByPercentOfAvailable % of the available expansion
1767 // space, whichever is smaller, bounded below by a minimum
1768 // expansion (unless that's all that's left.)
1769 const size_t min_expand_bytes = 1*M;
1770 size_t reserved_bytes = _g1->max_capacity();
1771 size_t committed_bytes = _g1->capacity();
1772 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
1773 size_t expand_bytes;
1774 size_t expand_bytes_via_pct =
1775 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
1776 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
1777 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
1778 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
1779
1780 ergo_verbose5(ErgoHeapSizing,
1781 "attempt heap expansion",
1782 ergo_format_reason("recent GC overhead higher than "
1783 "threshold after GC")
1784 ergo_format_perc("recent GC overhead")
1785 ergo_format_perc("threshold")
1786 ergo_format_byte("uncommitted")
1787 ergo_format_byte_perc("calculated expansion amount"),
1788 recent_gc_overhead, threshold,
1789 uncommitted_bytes,
1790 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable);
1791
1792 return expand_bytes;
1793 } else {
1794 return 0;
1795 }
1796 }
1797
1798 class CountCSClosure: public HeapRegionClosure {
1799 G1CollectorPolicy* _g1_policy;
1800 public:
1801 CountCSClosure(G1CollectorPolicy* g1_policy) :
1802 _g1_policy(g1_policy) {}
1803 bool doHeapRegion(HeapRegion* r) {
1804 _g1_policy->_bytes_in_collection_set_before_gc += r->used();
1805 return false;
1806 }
1807 };
1808
1809 void G1CollectorPolicy::count_CS_bytes_used() {
1810 CountCSClosure cs_closure(this);
1811 _g1->collection_set_iterate(&cs_closure);
1812 }
1813
1814 void G1CollectorPolicy::print_summary(int level,
1815 const char* str,
1816 NumberSeq* seq) const {
1817 double sum = seq->sum();
1818 LineBuffer(level + 1).append_and_print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)",
1819 str, sum / 1000.0, seq->avg());
1820 }
1821
1822 void G1CollectorPolicy::print_summary_sd(int level,
1823 const char* str,
1824 NumberSeq* seq) const {
1825 print_summary(level, str, seq);
1826 LineBuffer(level + 6).append_and_print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
1827 seq->num(), seq->sd(), seq->maximum());
1828 }
1829
1830 void G1CollectorPolicy::check_other_times(int level,
1831 NumberSeq* other_times_ms,
1832 NumberSeq* calc_other_times_ms) const {
1833 bool should_print = false;
1834 LineBuffer buf(level + 2);
1835
1836 double max_sum = MAX2(fabs(other_times_ms->sum()),
1837 fabs(calc_other_times_ms->sum()));
1838 double min_sum = MIN2(fabs(other_times_ms->sum()),
1839 fabs(calc_other_times_ms->sum()));
1840 double sum_ratio = max_sum / min_sum;
1841 if (sum_ratio > 1.1) {
1842 should_print = true;
1843 buf.append_and_print_cr("## CALCULATED OTHER SUM DOESN'T MATCH RECORDED ###");
1844 }
1845
1846 double max_avg = MAX2(fabs(other_times_ms->avg()),
1847 fabs(calc_other_times_ms->avg()));
1848 double min_avg = MIN2(fabs(other_times_ms->avg()),
1849 fabs(calc_other_times_ms->avg()));
1850 double avg_ratio = max_avg / min_avg;
1851 if (avg_ratio > 1.1) {
1852 should_print = true;
1853 buf.append_and_print_cr("## CALCULATED OTHER AVG DOESN'T MATCH RECORDED ###");
1854 }
1855
1856 if (other_times_ms->sum() < -0.01) {
1857 buf.append_and_print_cr("## RECORDED OTHER SUM IS NEGATIVE ###");
1858 }
1859
1860 if (other_times_ms->avg() < -0.01) {
1861 buf.append_and_print_cr("## RECORDED OTHER AVG IS NEGATIVE ###");
1862 }
1863
1864 if (calc_other_times_ms->sum() < -0.01) {
1865 should_print = true;
1866 buf.append_and_print_cr("## CALCULATED OTHER SUM IS NEGATIVE ###");
1867 }
1868
1869 if (calc_other_times_ms->avg() < -0.01) {
1870 should_print = true;
1871 buf.append_and_print_cr("## CALCULATED OTHER AVG IS NEGATIVE ###");
1872 }
1873
1874 if (should_print)
1875 print_summary(level, "Other(Calc)", calc_other_times_ms);
1876 }
1877
1878 void G1CollectorPolicy::print_summary(PauseSummary* summary) const {
1879 bool parallel = G1CollectedHeap::use_parallel_gc_threads();
1880 MainBodySummary* body_summary = summary->main_body_summary();
1881 if (summary->get_total_seq()->num() > 0) {
1882 print_summary_sd(0, "Evacuation Pauses", summary->get_total_seq());
1883 if (body_summary != NULL) {
1884 print_summary(1, "Root Region Scan Wait", body_summary->get_root_region_scan_wait_seq());
1885 if (parallel) {
1886 print_summary(1, "Parallel Time", body_summary->get_parallel_seq());
1887 print_summary(2, "Ext Root Scanning", body_summary->get_ext_root_scan_seq());
1888 print_summary(2, "SATB Filtering", body_summary->get_satb_filtering_seq());
1889 print_summary(2, "Update RS", body_summary->get_update_rs_seq());
1890 print_summary(2, "Scan RS", body_summary->get_scan_rs_seq());
1891 print_summary(2, "Object Copy", body_summary->get_obj_copy_seq());
1892 print_summary(2, "Termination", body_summary->get_termination_seq());
1893 print_summary(2, "Parallel Other", body_summary->get_parallel_other_seq());
1894 {
1895 NumberSeq* other_parts[] = {
1896 body_summary->get_ext_root_scan_seq(),
1897 body_summary->get_satb_filtering_seq(),
1898 body_summary->get_update_rs_seq(),
1899 body_summary->get_scan_rs_seq(),
1900 body_summary->get_obj_copy_seq(),
1901 body_summary->get_termination_seq()
1902 };
1903 NumberSeq calc_other_times_ms(body_summary->get_parallel_seq(),
1904 6, other_parts);
1905 check_other_times(2, body_summary->get_parallel_other_seq(),
1906 &calc_other_times_ms);
1907 }
1908 } else {
1909 print_summary(1, "Ext Root Scanning", body_summary->get_ext_root_scan_seq());
1910 print_summary(1, "SATB Filtering", body_summary->get_satb_filtering_seq());
1911 print_summary(1, "Update RS", body_summary->get_update_rs_seq());
1912 print_summary(1, "Scan RS", body_summary->get_scan_rs_seq());
1913 print_summary(1, "Object Copy", body_summary->get_obj_copy_seq());
1914 }
1915 }
1916 print_summary(1, "Clear CT", body_summary->get_clear_ct_seq());
1917 print_summary(1, "Other", summary->get_other_seq());
1918 {
1919 if (body_summary != NULL) {
1920 NumberSeq calc_other_times_ms;
1921 if (parallel) {
1922 // parallel
1923 NumberSeq* other_parts[] = {
1924 body_summary->get_root_region_scan_wait_seq(),
1925 body_summary->get_parallel_seq(),
1926 body_summary->get_clear_ct_seq()
1927 };
1928 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
1929 3, other_parts);
1930 } else {
1931 // serial
1932 NumberSeq* other_parts[] = {
1933 body_summary->get_root_region_scan_wait_seq(),
1934 body_summary->get_update_rs_seq(),
1935 body_summary->get_ext_root_scan_seq(),
1936 body_summary->get_satb_filtering_seq(),
1937 body_summary->get_scan_rs_seq(),
1938 body_summary->get_obj_copy_seq()
1939 };
1940 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
1941 6, other_parts);
1942 }
1943 check_other_times(1, summary->get_other_seq(), &calc_other_times_ms);
1944 }
1945 }
1946 } else {
1947 LineBuffer(1).append_and_print_cr("none");
1948 }
1949 LineBuffer(0).append_and_print_cr("");
1950 }
1951
1952 void G1CollectorPolicy::print_tracing_info() const {
1953 if (TraceGen0Time) {
1954 gclog_or_tty->print_cr("ALL PAUSES");
1955 print_summary_sd(0, "Total", _all_pause_times_ms);
1956 gclog_or_tty->print_cr("");
1957 gclog_or_tty->print_cr("");
1958 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num);
1959 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num);
1960 gclog_or_tty->print_cr("");
1961
1962 gclog_or_tty->print_cr("EVACUATION PAUSES");
1963 print_summary(_summary);
1964
1965 gclog_or_tty->print_cr("MISC");
1966 print_summary_sd(0, "Stop World", _all_stop_world_times_ms);
1967 print_summary_sd(0, "Yields", _all_yield_times_ms);
1968 for (int i = 0; i < _aux_num; ++i) {
1969 if (_all_aux_times_ms[i].num() > 0) {
1970 char buffer[96];
1971 sprintf(buffer, "Aux%d", i);
1972 print_summary_sd(0, buffer, &_all_aux_times_ms[i]);
1973 }
1974 }
1975 }
1976 if (TraceGen1Time) {
1977 if (_all_full_gc_times_ms->num() > 0) {
1978 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s",
1979 _all_full_gc_times_ms->num(),
1980 _all_full_gc_times_ms->sum() / 1000.0);
1981 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times_ms->avg());
1982 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]",
1983 _all_full_gc_times_ms->sd(),
1984 _all_full_gc_times_ms->maximum());
1985 }
1986 }
1987 }
1988
1989 void G1CollectorPolicy::print_yg_surv_rate_info() const {
1990 #ifndef PRODUCT
1991 _short_lived_surv_rate_group->print_surv_rate_summary();
1992 // add this call for any other surv rate groups
1993 #endif // PRODUCT
1994 }
1995
1996 #ifndef PRODUCT
1997 // for debugging, bit of a hack...
1998 static char*
1999 region_num_to_mbs(int length) {
2000 static char buffer[64];
2001 double bytes = (double) (length * HeapRegion::GrainBytes);
2002 double mbs = bytes / (double) (1024 * 1024);
2003 sprintf(buffer, "%7.2lfMB", mbs);
2004 return buffer;
2005 }
2006 #endif // PRODUCT
2007
2008 uint G1CollectorPolicy::max_regions(int purpose) {
2009 switch (purpose) {
2010 case GCAllocForSurvived:
2011 return _max_survivor_regions;
2012 case GCAllocForTenured:
2013 return REGIONS_UNLIMITED;
2014 default:
2015 ShouldNotReachHere();
2016 return REGIONS_UNLIMITED;
2017 };
2018 }
2019
2020 void G1CollectorPolicy::update_max_gc_locker_expansion() {
2021 uint expansion_region_num = 0;
2022 if (GCLockerEdenExpansionPercent > 0) {
2023 double perc = (double) GCLockerEdenExpansionPercent / 100.0;
2024 double expansion_region_num_d = perc * (double) _young_list_target_length;
2025 // We use ceiling so that if expansion_region_num_d is > 0.0 (but
2026 // less than 1.0) we'll get 1.
2027 expansion_region_num = (uint) ceil(expansion_region_num_d);
2028 } else {
2029 assert(expansion_region_num == 0, "sanity");
2030 }
2031 _young_list_max_length = _young_list_target_length + expansion_region_num;
2032 assert(_young_list_target_length <= _young_list_max_length, "post-condition");
2033 }
2034
2035 // Calculates survivor space parameters.
2036 void G1CollectorPolicy::update_survivors_policy() {
2037 double max_survivor_regions_d =
2038 (double) _young_list_target_length / (double) SurvivorRatio;
2039 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but
2040 // smaller than 1.0) we'll get 1.
2041 _max_survivor_regions = (uint) ceil(max_survivor_regions_d);
2042
2043 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
2044 HeapRegion::GrainWords * _max_survivor_regions);
2045 }
2046
2047 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(
2048 GCCause::Cause gc_cause) {
2049 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2050 if (!during_cycle) {
2051 ergo_verbose1(ErgoConcCycles,
2052 "request concurrent cycle initiation",
2053 ergo_format_reason("requested by GC cause")
2054 ergo_format_str("GC cause"),
2055 GCCause::to_string(gc_cause));
2056 set_initiate_conc_mark_if_possible();
2057 return true;
2058 } else {
2059 ergo_verbose1(ErgoConcCycles,
2060 "do not request concurrent cycle initiation",
2061 ergo_format_reason("concurrent cycle already in progress")
2062 ergo_format_str("GC cause"),
2063 GCCause::to_string(gc_cause));
2064 return false;
2065 }
2066 }
2067
2068 void
2069 G1CollectorPolicy::decide_on_conc_mark_initiation() {
2070 // We are about to decide on whether this pause will be an
2071 // initial-mark pause.
2072
2073 // First, during_initial_mark_pause() should not be already set. We
2074 // will set it here if we have to. However, it should be cleared by
2075 // the end of the pause (it's only set for the duration of an
2076 // initial-mark pause).
2077 assert(!during_initial_mark_pause(), "pre-condition");
2078
2079 if (initiate_conc_mark_if_possible()) {
2080 // We had noticed on a previous pause that the heap occupancy has
2081 // gone over the initiating threshold and we should start a
2082 // concurrent marking cycle. So we might initiate one.
2083
2084 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2085 if (!during_cycle) {
2086 // The concurrent marking thread is not "during a cycle", i.e.,
2087 // it has completed the last one. So we can go ahead and
2088 // initiate a new cycle.
2089
2090 set_during_initial_mark_pause();
2091 // We do not allow mixed GCs during marking.
2092 if (!gcs_are_young()) {
2093 set_gcs_are_young(true);
2094 ergo_verbose0(ErgoMixedGCs,
2095 "end mixed GCs",
2096 ergo_format_reason("concurrent cycle is about to start"));
2097 }
2098
2099 // And we can now clear initiate_conc_mark_if_possible() as
2100 // we've already acted on it.
2101 clear_initiate_conc_mark_if_possible();
2102
2103 ergo_verbose0(ErgoConcCycles,
2104 "initiate concurrent cycle",
2105 ergo_format_reason("concurrent cycle initiation requested"));
2106 } else {
2107 // The concurrent marking thread is still finishing up the
2108 // previous cycle. If we start one right now the two cycles
2109 // overlap. In particular, the concurrent marking thread might
2110 // be in the process of clearing the next marking bitmap (which
2111 // we will use for the next cycle if we start one). Starting a
2112 // cycle now will be bad given that parts of the marking
2113 // information might get cleared by the marking thread. And we
2114 // cannot wait for the marking thread to finish the cycle as it
2115 // periodically yields while clearing the next marking bitmap
2116 // and, if it's in a yield point, it's waiting for us to
2117 // finish. So, at this point we will not start a cycle and we'll
2118 // let the concurrent marking thread complete the last one.
2119 ergo_verbose0(ErgoConcCycles,
2120 "do not initiate concurrent cycle",
2121 ergo_format_reason("concurrent cycle already in progress"));
2122 }
2123 }
2124 }
2125
2126 class KnownGarbageClosure: public HeapRegionClosure {
2127 G1CollectedHeap* _g1h;
2128 CollectionSetChooser* _hrSorted;
2129
2130 public:
2131 KnownGarbageClosure(CollectionSetChooser* hrSorted) :
2132 _g1h(G1CollectedHeap::heap()), _hrSorted(hrSorted) { }
2133
2134 bool doHeapRegion(HeapRegion* r) {
2135 // We only include humongous regions in collection
2136 // sets when concurrent mark shows that their contained object is
2137 // unreachable.
2138
2139 // Do we have any marking information for this region?
2140 if (r->is_marked()) {
2141 // We will skip any region that's currently used as an old GC
2142 // alloc region (we should not consider those for collection
2143 // before we fill them up).
2144 if (_hrSorted->should_add(r) && !_g1h->is_old_gc_alloc_region(r)) {
2145 _hrSorted->add_region(r);
2146 }
2147 }
2148 return false;
2149 }
2150 };
2151
2152 class ParKnownGarbageHRClosure: public HeapRegionClosure {
2153 G1CollectedHeap* _g1h;
2154 CollectionSetChooser* _hrSorted;
2155 uint _marked_regions_added;
2156 size_t _reclaimable_bytes_added;
2157 uint _chunk_size;
2158 uint _cur_chunk_idx;
2159 uint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end)
2160
2161 void get_new_chunk() {
2162 _cur_chunk_idx = _hrSorted->claim_array_chunk(_chunk_size);
2163 _cur_chunk_end = _cur_chunk_idx + _chunk_size;
2164 }
2165 void add_region(HeapRegion* r) {
2166 if (_cur_chunk_idx == _cur_chunk_end) {
2167 get_new_chunk();
2168 }
2169 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition");
2170 _hrSorted->set_region(_cur_chunk_idx, r);
2171 _marked_regions_added++;
2172 _reclaimable_bytes_added += r->reclaimable_bytes();
2173 _cur_chunk_idx++;
2174 }
2175
2176 public:
2177 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
2178 uint chunk_size) :
2179 _g1h(G1CollectedHeap::heap()),
2180 _hrSorted(hrSorted), _chunk_size(chunk_size),
2181 _marked_regions_added(0), _reclaimable_bytes_added(0),
2182 _cur_chunk_idx(0), _cur_chunk_end(0) { }
2183
2184 bool doHeapRegion(HeapRegion* r) {
2185 // Do we have any marking information for this region?
2186 if (r->is_marked()) {
2187 // We will skip any region that's currently used as an old GC
2188 // alloc region (we should not consider those for collection
2189 // before we fill them up).
2190 if (_hrSorted->should_add(r) && !_g1h->is_old_gc_alloc_region(r)) {
2191 add_region(r);
2192 }
2193 }
2194 return false;
2195 }
2196 uint marked_regions_added() { return _marked_regions_added; }
2197 size_t reclaimable_bytes_added() { return _reclaimable_bytes_added; }
2198 };
2199
2200 class ParKnownGarbageTask: public AbstractGangTask {
2201 CollectionSetChooser* _hrSorted;
2202 uint _chunk_size;
2203 G1CollectedHeap* _g1;
2204 public:
2205 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size) :
2206 AbstractGangTask("ParKnownGarbageTask"),
2207 _hrSorted(hrSorted), _chunk_size(chunk_size),
2208 _g1(G1CollectedHeap::heap()) { }
2209
2210 void work(uint worker_id) {
2211 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size);
2212
2213 // Back to zero for the claim value.
2214 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, worker_id,
2215 _g1->workers()->active_workers(),
2216 HeapRegion::InitialClaimValue);
2217 uint regions_added = parKnownGarbageCl.marked_regions_added();
2218 size_t reclaimable_bytes_added =
2219 parKnownGarbageCl.reclaimable_bytes_added();
2220 _hrSorted->update_totals(regions_added, reclaimable_bytes_added);
2221 }
2222 };
2223
2224 void
2225 G1CollectorPolicy::record_concurrent_mark_cleanup_end(int no_of_gc_threads) {
2226 _collectionSetChooser->clear();
2227
2228 uint region_num = _g1->n_regions();
2229 if (G1CollectedHeap::use_parallel_gc_threads()) {
2230 const uint OverpartitionFactor = 4;
2231 uint WorkUnit;
2232 // The use of MinChunkSize = 8 in the original code
2233 // causes some assertion failures when the total number of
2234 // region is less than 8. The code here tries to fix that.
2235 // Should the original code also be fixed?
2236 if (no_of_gc_threads > 0) {
2237 const uint MinWorkUnit = MAX2(region_num / no_of_gc_threads, 1U);
2238 WorkUnit = MAX2(region_num / (no_of_gc_threads * OverpartitionFactor),
2239 MinWorkUnit);
2240 } else {
2241 assert(no_of_gc_threads > 0,
2242 "The active gc workers should be greater than 0");
2243 // In a product build do something reasonable to avoid a crash.
2244 const uint MinWorkUnit = MAX2(region_num / (uint) ParallelGCThreads, 1U);
2245 WorkUnit =
2246 MAX2(region_num / (uint) (ParallelGCThreads * OverpartitionFactor),
2247 MinWorkUnit);
2248 }
2249 _collectionSetChooser->prepare_for_par_region_addition(_g1->n_regions(),
2250 WorkUnit);
2251 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser,
2252 (int) WorkUnit);
2253 _g1->workers()->run_task(&parKnownGarbageTask);
2254
2255 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2256 "sanity check");
2257 } else {
2258 KnownGarbageClosure knownGarbagecl(_collectionSetChooser);
2259 _g1->heap_region_iterate(&knownGarbagecl);
2260 }
2261
2262 _collectionSetChooser->sort_regions();
2263
2264 double end_sec = os::elapsedTime();
2265 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0;
2266 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
2267 _cur_mark_stop_world_time_ms += elapsed_time_ms;
2268 _prev_collection_pause_end_ms += elapsed_time_ms;
2269 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec, true);
2270 }
2271
2272 // Add the heap region at the head of the non-incremental collection set
2273 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) {
2274 assert(_inc_cset_build_state == Active, "Precondition");
2275 assert(!hr->is_young(), "non-incremental add of young region");
2276
2277 assert(!hr->in_collection_set(), "should not already be in the CSet");
2278 hr->set_in_collection_set(true);
2279 hr->set_next_in_collection_set(_collection_set);
2280 _collection_set = hr;
2281 _collection_set_bytes_used_before += hr->used();
2282 _g1->register_region_with_in_cset_fast_test(hr);
2283 size_t rs_length = hr->rem_set()->occupied();
2284 _recorded_rs_lengths += rs_length;
2285 _old_cset_region_length += 1;
2286 }
2287
2288 // Initialize the per-collection-set information
2289 void G1CollectorPolicy::start_incremental_cset_building() {
2290 assert(_inc_cset_build_state == Inactive, "Precondition");
2291
2292 _inc_cset_head = NULL;
2293 _inc_cset_tail = NULL;
2294 _inc_cset_bytes_used_before = 0;
2295
2296 _inc_cset_max_finger = 0;
2297 _inc_cset_recorded_rs_lengths = 0;
2298 _inc_cset_recorded_rs_lengths_diffs = 0;
2299 _inc_cset_predicted_elapsed_time_ms = 0.0;
2300 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
2301 _inc_cset_build_state = Active;
2302 }
2303
2304 void G1CollectorPolicy::finalize_incremental_cset_building() {
2305 assert(_inc_cset_build_state == Active, "Precondition");
2306 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
2307
2308 // The two "main" fields, _inc_cset_recorded_rs_lengths and
2309 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread
2310 // that adds a new region to the CSet. Further updates by the
2311 // concurrent refinement thread that samples the young RSet lengths
2312 // are accumulated in the *_diffs fields. Here we add the diffs to
2313 // the "main" fields.
2314
2315 if (_inc_cset_recorded_rs_lengths_diffs >= 0) {
2316 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs;
2317 } else {
2318 // This is defensive. The diff should in theory be always positive
2319 // as RSets can only grow between GCs. However, given that we
2320 // sample their size concurrently with other threads updating them
2321 // it's possible that we might get the wrong size back, which
2322 // could make the calculations somewhat inaccurate.
2323 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs);
2324 if (_inc_cset_recorded_rs_lengths >= diffs) {
2325 _inc_cset_recorded_rs_lengths -= diffs;
2326 } else {
2327 _inc_cset_recorded_rs_lengths = 0;
2328 }
2329 }
2330 _inc_cset_predicted_elapsed_time_ms +=
2331 _inc_cset_predicted_elapsed_time_ms_diffs;
2332
2333 _inc_cset_recorded_rs_lengths_diffs = 0;
2334 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
2335 }
2336
2337 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) {
2338 // This routine is used when:
2339 // * adding survivor regions to the incremental cset at the end of an
2340 // evacuation pause,
2341 // * adding the current allocation region to the incremental cset
2342 // when it is retired, and
2343 // * updating existing policy information for a region in the
2344 // incremental cset via young list RSet sampling.
2345 // Therefore this routine may be called at a safepoint by the
2346 // VM thread, or in-between safepoints by mutator threads (when
2347 // retiring the current allocation region) or a concurrent
2348 // refine thread (RSet sampling).
2349
2350 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true);
2351 size_t used_bytes = hr->used();
2352 _inc_cset_recorded_rs_lengths += rs_length;
2353 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms;
2354 _inc_cset_bytes_used_before += used_bytes;
2355
2356 // Cache the values we have added to the aggregated informtion
2357 // in the heap region in case we have to remove this region from
2358 // the incremental collection set, or it is updated by the
2359 // rset sampling code
2360 hr->set_recorded_rs_length(rs_length);
2361 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms);
2362 }
2363
2364 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr,
2365 size_t new_rs_length) {
2366 // Update the CSet information that is dependent on the new RS length
2367 assert(hr->is_young(), "Precondition");
2368 assert(!SafepointSynchronize::is_at_safepoint(),
2369 "should not be at a safepoint");
2370
2371 // We could have updated _inc_cset_recorded_rs_lengths and
2372 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do
2373 // that atomically, as this code is executed by a concurrent
2374 // refinement thread, potentially concurrently with a mutator thread
2375 // allocating a new region and also updating the same fields. To
2376 // avoid the atomic operations we accumulate these updates on two
2377 // separate fields (*_diffs) and we'll just add them to the "main"
2378 // fields at the start of a GC.
2379
2380 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length();
2381 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length;
2382 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff;
2383
2384 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms();
2385 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true);
2386 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms;
2387 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff;
2388
2389 hr->set_recorded_rs_length(new_rs_length);
2390 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms);
2391 }
2392
2393 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) {
2394 assert(hr->is_young(), "invariant");
2395 assert(hr->young_index_in_cset() > -1, "should have already been set");
2396 assert(_inc_cset_build_state == Active, "Precondition");
2397
2398 // We need to clear and set the cached recorded/cached collection set
2399 // information in the heap region here (before the region gets added
2400 // to the collection set). An individual heap region's cached values
2401 // are calculated, aggregated with the policy collection set info,
2402 // and cached in the heap region here (initially) and (subsequently)
2403 // by the Young List sampling code.
2404
2405 size_t rs_length = hr->rem_set()->occupied();
2406 add_to_incremental_cset_info(hr, rs_length);
2407
2408 HeapWord* hr_end = hr->end();
2409 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end);
2410
2411 assert(!hr->in_collection_set(), "invariant");
2412 hr->set_in_collection_set(true);
2413 assert( hr->next_in_collection_set() == NULL, "invariant");
2414
2415 _g1->register_region_with_in_cset_fast_test(hr);
2416 }
2417
2418 // Add the region at the RHS of the incremental cset
2419 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) {
2420 // We should only ever be appending survivors at the end of a pause
2421 assert( hr->is_survivor(), "Logic");
2422
2423 // Do the 'common' stuff
2424 add_region_to_incremental_cset_common(hr);
2425
2426 // Now add the region at the right hand side
2427 if (_inc_cset_tail == NULL) {
2428 assert(_inc_cset_head == NULL, "invariant");
2429 _inc_cset_head = hr;
2430 } else {
2431 _inc_cset_tail->set_next_in_collection_set(hr);
2432 }
2433 _inc_cset_tail = hr;
2434 }
2435
2436 // Add the region to the LHS of the incremental cset
2437 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) {
2438 // Survivors should be added to the RHS at the end of a pause
2439 assert(!hr->is_survivor(), "Logic");
2440
2441 // Do the 'common' stuff
2442 add_region_to_incremental_cset_common(hr);
2443
2444 // Add the region at the left hand side
2445 hr->set_next_in_collection_set(_inc_cset_head);
2446 if (_inc_cset_head == NULL) {
2447 assert(_inc_cset_tail == NULL, "Invariant");
2448 _inc_cset_tail = hr;
2449 }
2450 _inc_cset_head = hr;
2451 }
2452
2453 #ifndef PRODUCT
2454 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) {
2455 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be");
2456
2457 st->print_cr("\nCollection_set:");
2458 HeapRegion* csr = list_head;
2459 while (csr != NULL) {
2460 HeapRegion* next = csr->next_in_collection_set();
2461 assert(csr->in_collection_set(), "bad CS");
2462 st->print_cr(" [%08x-%08x], t: %08x, P: %08x, N: %08x, C: %08x, "
2463 "age: %4d, y: %d, surv: %d",
2464 csr->bottom(), csr->end(),
2465 csr->top(),
2466 csr->prev_top_at_mark_start(),
2467 csr->next_top_at_mark_start(),
2468 csr->top_at_conc_mark_count(),
2469 csr->age_in_surv_rate_group_cond(),
2470 csr->is_young(),
2471 csr->is_survivor());
2472 csr = next;
2473 }
2474 }
2475 #endif // !PRODUCT
2476
2477 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str,
2478 const char* false_action_str) {
2479 CollectionSetChooser* cset_chooser = _collectionSetChooser;
2480 if (cset_chooser->is_empty()) {
2481 ergo_verbose0(ErgoMixedGCs,
2482 false_action_str,
2483 ergo_format_reason("candidate old regions not available"));
2484 return false;
2485 }
2486 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes();
2487 size_t capacity_bytes = _g1->capacity();
2488 double perc = (double) reclaimable_bytes * 100.0 / (double) capacity_bytes;
2489 double threshold = (double) G1HeapWastePercent;
2490 if (perc < threshold) {
2491 ergo_verbose4(ErgoMixedGCs,
2492 false_action_str,
2493 ergo_format_reason("reclaimable percentage lower than threshold")
2494 ergo_format_region("candidate old regions")
2495 ergo_format_byte_perc("reclaimable")
2496 ergo_format_perc("threshold"),
2497 cset_chooser->remaining_regions(),
2498 reclaimable_bytes, perc, threshold);
2499 return false;
2500 }
2501
2502 ergo_verbose4(ErgoMixedGCs,
2503 true_action_str,
2504 ergo_format_reason("candidate old regions available")
2505 ergo_format_region("candidate old regions")
2506 ergo_format_byte_perc("reclaimable")
2507 ergo_format_perc("threshold"),
2508 cset_chooser->remaining_regions(),
2509 reclaimable_bytes, perc, threshold);
2510 return true;
2511 }
2512
2513 void G1CollectorPolicy::finalize_cset(double target_pause_time_ms) {
2514 // Set this here - in case we're not doing young collections.
2515 double non_young_start_time_sec = os::elapsedTime();
2516
2517 YoungList* young_list = _g1->young_list();
2518 finalize_incremental_cset_building();
2519
2520 guarantee(target_pause_time_ms > 0.0,
2521 err_msg("target_pause_time_ms = %1.6lf should be positive",
2522 target_pause_time_ms));
2523 guarantee(_collection_set == NULL, "Precondition");
2524
2525 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
2526 double predicted_pause_time_ms = base_time_ms;
2527 double time_remaining_ms = target_pause_time_ms - base_time_ms;
2528
2529 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2530 "start choosing CSet",
2531 ergo_format_ms("predicted base time")
2532 ergo_format_ms("remaining time")
2533 ergo_format_ms("target pause time"),
2534 base_time_ms, time_remaining_ms, target_pause_time_ms);
2535
2536 HeapRegion* hr;
2537 double young_start_time_sec = os::elapsedTime();
2538
2539 _collection_set_bytes_used_before = 0;
2540 _last_gc_was_young = gcs_are_young() ? true : false;
2541
2542 if (_last_gc_was_young) {
2543 ++_young_pause_num;
2544 } else {
2545 ++_mixed_pause_num;
2546 }
2547
2548 // The young list is laid with the survivor regions from the previous
2549 // pause are appended to the RHS of the young list, i.e.
2550 // [Newly Young Regions ++ Survivors from last pause].
2551
2552 uint survivor_region_length = young_list->survivor_length();
2553 uint eden_region_length = young_list->length() - survivor_region_length;
2554 init_cset_region_lengths(eden_region_length, survivor_region_length);
2555 hr = young_list->first_survivor_region();
2556 while (hr != NULL) {
2557 assert(hr->is_survivor(), "badly formed young list");
2558 hr->set_young();
2559 hr = hr->get_next_young_region();
2560 }
2561
2562 // Clear the fields that point to the survivor list - they are all young now.
2563 young_list->clear_survivors();
2564
2565 _collection_set = _inc_cset_head;
2566 _collection_set_bytes_used_before = _inc_cset_bytes_used_before;
2567 time_remaining_ms -= _inc_cset_predicted_elapsed_time_ms;
2568 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms;
2569
2570 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2571 "add young regions to CSet",
2572 ergo_format_region("eden")
2573 ergo_format_region("survivors")
2574 ergo_format_ms("predicted young region time"),
2575 eden_region_length, survivor_region_length,
2576 _inc_cset_predicted_elapsed_time_ms);
2577
2578 // The number of recorded young regions is the incremental
2579 // collection set's current size
2580 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths);
2581
2582 double young_end_time_sec = os::elapsedTime();
2583 _recorded_young_cset_choice_time_ms =
2584 (young_end_time_sec - young_start_time_sec) * 1000.0;
2585
2586 // We are doing young collections so reset this.
2587 non_young_start_time_sec = young_end_time_sec;
2588
2589 if (!gcs_are_young()) {
2590 CollectionSetChooser* cset_chooser = _collectionSetChooser;
2591 cset_chooser->verify();
2592 const uint min_old_cset_length = cset_chooser->calc_min_old_cset_length();
2593 const uint max_old_cset_length = cset_chooser->calc_max_old_cset_length();
2594
2595 uint expensive_region_num = 0;
2596 bool check_time_remaining = adaptive_young_list_length();
2597 HeapRegion* hr = cset_chooser->peek();
2598 while (hr != NULL) {
2599 if (old_cset_region_length() >= max_old_cset_length) {
2600 // Added maximum number of old regions to the CSet.
2601 ergo_verbose2(ErgoCSetConstruction,
2602 "finish adding old regions to CSet",
2603 ergo_format_reason("old CSet region num reached max")
2604 ergo_format_region("old")
2605 ergo_format_region("max"),
2606 old_cset_region_length(), max_old_cset_length);
2607 break;
2608 }
2609
2610 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false);
2611 if (check_time_remaining) {
2612 if (predicted_time_ms > time_remaining_ms) {
2613 // Too expensive for the current CSet.
2614
2615 if (old_cset_region_length() >= min_old_cset_length) {
2616 // We have added the minimum number of old regions to the CSet,
2617 // we are done with this CSet.
2618 ergo_verbose4(ErgoCSetConstruction,
2619 "finish adding old regions to CSet",
2620 ergo_format_reason("predicted time is too high")
2621 ergo_format_ms("predicted time")
2622 ergo_format_ms("remaining time")
2623 ergo_format_region("old")
2624 ergo_format_region("min"),
2625 predicted_time_ms, time_remaining_ms,
2626 old_cset_region_length(), min_old_cset_length);
2627 break;
2628 }
2629
2630 // We'll add it anyway given that we haven't reached the
2631 // minimum number of old regions.
2632 expensive_region_num += 1;
2633 }
2634 } else {
2635 if (old_cset_region_length() >= min_old_cset_length) {
2636 // In the non-auto-tuning case, we'll finish adding regions
2637 // to the CSet if we reach the minimum.
2638 ergo_verbose2(ErgoCSetConstruction,
2639 "finish adding old regions to CSet",
2640 ergo_format_reason("old CSet region num reached min")
2641 ergo_format_region("old")
2642 ergo_format_region("min"),
2643 old_cset_region_length(), min_old_cset_length);
2644 break;
2645 }
2646 }
2647
2648 // We will add this region to the CSet.
2649 time_remaining_ms -= predicted_time_ms;
2650 predicted_pause_time_ms += predicted_time_ms;
2651 cset_chooser->remove_and_move_to_next(hr);
2652 _g1->old_set_remove(hr);
2653 add_old_region_to_cset(hr);
2654
2655 hr = cset_chooser->peek();
2656 }
2657 if (hr == NULL) {
2658 ergo_verbose0(ErgoCSetConstruction,
2659 "finish adding old regions to CSet",
2660 ergo_format_reason("candidate old regions not available"));
2661 }
2662
2663 if (expensive_region_num > 0) {
2664 // We print the information once here at the end, predicated on
2665 // whether we added any apparently expensive regions or not, to
2666 // avoid generating output per region.
2667 ergo_verbose4(ErgoCSetConstruction,
2668 "added expensive regions to CSet",
2669 ergo_format_reason("old CSet region num not reached min")
2670 ergo_format_region("old")
2671 ergo_format_region("expensive")
2672 ergo_format_region("min")
2673 ergo_format_ms("remaining time"),
2674 old_cset_region_length(),
2675 expensive_region_num,
2676 min_old_cset_length,
2677 time_remaining_ms);
2678 }
2679
2680 cset_chooser->verify();
2681 }
2682
2683 stop_incremental_cset_building();
2684
2685 count_CS_bytes_used();
2686
2687 ergo_verbose5(ErgoCSetConstruction,
2688 "finish choosing CSet",
2689 ergo_format_region("eden")
2690 ergo_format_region("survivors")
2691 ergo_format_region("old")
2692 ergo_format_ms("predicted pause time")
2693 ergo_format_ms("target pause time"),
2694 eden_region_length, survivor_region_length,
2695 old_cset_region_length(),
2696 predicted_pause_time_ms, target_pause_time_ms);
2697
2698 double non_young_end_time_sec = os::elapsedTime();
2699 _recorded_non_young_cset_choice_time_ms =
2700 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0;
2701 }