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 (%s) (%s)",
890 GCCause::to_string(_g1->gc_cause()),
891 gcs_are_young() ? "young" : "mixed");
892 }
893
894 // We only need to do this here as the policy will only be applied
895 // to the GC we're about to start. so, no point is calculating this
896 // every time we calculate / recalculate the target young length.
897 update_survivors_policy();
898
899 assert(_g1->used() == _g1->recalculate_used(),
900 err_msg("sanity, used: "SIZE_FORMAT" recalculate_used: "SIZE_FORMAT,
901 _g1->used(), _g1->recalculate_used()));
902
903 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0;
904 _all_stop_world_times_ms->add(s_w_t_ms);
905 _stop_world_start = 0.0;
906
907 _cur_collection_start_sec = start_time_sec;
908 _cur_collection_pause_used_at_start_bytes = start_used;
909 _cur_collection_pause_used_regions_at_start = _g1->used_regions();
910 _pending_cards = _g1->pending_card_num();
911 _max_pending_cards = _g1->max_pending_card_num();
912
913 _bytes_in_collection_set_before_gc = 0;
914 _bytes_copied_during_gc = 0;
915
916 YoungList* young_list = _g1->young_list();
917 _eden_bytes_before_gc = young_list->eden_used_bytes();
918 _survivor_bytes_before_gc = young_list->survivor_used_bytes();
919 _capacity_before_gc = _g1->capacity();
920
921 #ifdef DEBUG
922 // initialise these to something well known so that we can spot
923 // if they are not set properly
924
925 for (int i = 0; i < _parallel_gc_threads; ++i) {
926 _par_last_gc_worker_start_times_ms[i] = -1234.0;
927 _par_last_ext_root_scan_times_ms[i] = -1234.0;
928 _par_last_satb_filtering_times_ms[i] = -1234.0;
929 _par_last_update_rs_times_ms[i] = -1234.0;
930 _par_last_update_rs_processed_buffers[i] = -1234.0;
931 _par_last_scan_rs_times_ms[i] = -1234.0;
932 _par_last_obj_copy_times_ms[i] = -1234.0;
933 _par_last_termination_times_ms[i] = -1234.0;
934 _par_last_termination_attempts[i] = -1234.0;
935 _par_last_gc_worker_end_times_ms[i] = -1234.0;
936 _par_last_gc_worker_times_ms[i] = -1234.0;
937 _par_last_gc_worker_other_times_ms[i] = -1234.0;
938 }
939 #endif
940
941 for (int i = 0; i < _aux_num; ++i) {
942 _cur_aux_times_ms[i] = 0.0;
943 _cur_aux_times_set[i] = false;
944 }
945
946 // This is initialized to zero here and is set during the evacuation
947 // pause if we actually waited for the root region scanning to finish.
948 _root_region_scan_wait_time_ms = 0.0;
949
950 _last_gc_was_young = false;
951
952 // do that for any other surv rate groups
953 _short_lived_surv_rate_group->stop_adding_regions();
954 _survivors_age_table.clear();
955
956 assert( verify_young_ages(), "region age verification" );
957 }
958
959 void G1CollectorPolicy::record_concurrent_mark_init_end(double
960 mark_init_elapsed_time_ms) {
961 _during_marking = true;
962 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now");
963 clear_during_initial_mark_pause();
964 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms;
965 }
966
967 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
968 _mark_remark_start_sec = os::elapsedTime();
969 _during_marking = false;
970 }
971
972 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
973 double end_time_sec = os::elapsedTime();
974 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
975 _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
976 _cur_mark_stop_world_time_ms += elapsed_time_ms;
977 _prev_collection_pause_end_ms += elapsed_time_ms;
978
979 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true);
980 }
981
982 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
983 _mark_cleanup_start_sec = os::elapsedTime();
984 }
985
986 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
987 _last_young_gc = true;
988 _in_marking_window = false;
989 }
990
991 void G1CollectorPolicy::record_concurrent_pause() {
992 if (_stop_world_start > 0.0) {
993 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0;
994 _all_yield_times_ms->add(yield_ms);
995 }
996 }
997
998 void G1CollectorPolicy::record_concurrent_pause_end() {
999 }
1000
1001 template<class T>
1002 T sum_of(T* sum_arr, int start, int n, int N) {
1003 T sum = (T)0;
1004 for (int i = 0; i < n; i++) {
1005 int j = (start + i) % N;
1006 sum += sum_arr[j];
1007 }
1008 return sum;
1009 }
1010
1011 void G1CollectorPolicy::print_par_stats(int level,
1012 const char* str,
1013 double* data) {
1014 double min = data[0], max = data[0];
1015 double total = 0.0;
1016 LineBuffer buf(level);
1017 buf.append("[%s (ms):", str);
1018 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1019 double val = data[i];
1020 if (val < min)
1021 min = val;
1022 if (val > max)
1023 max = val;
1024 total += val;
1025 if (G1Log::finest()) {
1026 buf.append(" %.1lf", val);
1027 }
1028 }
1029
1030 if (G1Log::finest()) {
1031 buf.append_and_print_cr("");
1032 }
1033 double avg = total / (double) no_of_gc_threads();
1034 buf.append_and_print_cr(" Avg: %.1lf Min: %.1lf Max: %.1lf Diff: %.1lf]",
1035 avg, min, max, max - min);
1036 }
1037
1038 void G1CollectorPolicy::print_par_sizes(int level,
1039 const char* str,
1040 double* data) {
1041 double min = data[0], max = data[0];
1042 double total = 0.0;
1043 LineBuffer buf(level);
1044 buf.append("[%s :", str);
1045 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1046 double val = data[i];
1047 if (val < min)
1048 min = val;
1049 if (val > max)
1050 max = val;
1051 total += val;
1052 buf.append(" %d", (int) val);
1053 }
1054 buf.append_and_print_cr("");
1055 double avg = total / (double) no_of_gc_threads();
1056 buf.append_and_print_cr(" Sum: %d, Avg: %d, Min: %d, Max: %d, Diff: %d]",
1057 (int)total, (int)avg, (int)min, (int)max, (int)max - (int)min);
1058 }
1059
1060 void G1CollectorPolicy::print_stats(int level,
1061 const char* str,
1062 double value) {
1063 LineBuffer(level).append_and_print_cr("[%s: %5.1lf ms]", str, value);
1064 }
1065
1066 void G1CollectorPolicy::print_stats(int level,
1067 const char* str,
1068 int value) {
1069 LineBuffer(level).append_and_print_cr("[%s: %d]", str, value);
1070 }
1071
1072 double G1CollectorPolicy::avg_value(double* data) {
1073 if (G1CollectedHeap::use_parallel_gc_threads()) {
1074 double ret = 0.0;
1075 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1076 ret += data[i];
1077 }
1078 return ret / (double) no_of_gc_threads();
1079 } else {
1080 return data[0];
1081 }
1082 }
1083
1084 double G1CollectorPolicy::max_value(double* data) {
1085 if (G1CollectedHeap::use_parallel_gc_threads()) {
1086 double ret = data[0];
1087 for (uint i = 1; i < no_of_gc_threads(); ++i) {
1088 if (data[i] > ret) {
1089 ret = data[i];
1090 }
1091 }
1092 return ret;
1093 } else {
1094 return data[0];
1095 }
1096 }
1097
1098 double G1CollectorPolicy::sum_of_values(double* data) {
1099 if (G1CollectedHeap::use_parallel_gc_threads()) {
1100 double sum = 0.0;
1101 for (uint i = 0; i < no_of_gc_threads(); i++) {
1102 sum += data[i];
1103 }
1104 return sum;
1105 } else {
1106 return data[0];
1107 }
1108 }
1109
1110 double G1CollectorPolicy::max_sum(double* data1, double* data2) {
1111 double ret = data1[0] + data2[0];
1112
1113 if (G1CollectedHeap::use_parallel_gc_threads()) {
1114 for (uint i = 1; i < no_of_gc_threads(); ++i) {
1115 double data = data1[i] + data2[i];
1116 if (data > ret) {
1117 ret = data;
1118 }
1119 }
1120 }
1121 return ret;
1122 }
1123
1124 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) {
1125 if (_g1->concurrent_mark()->cmThread()->during_cycle()) {
1126 return false;
1127 }
1128
1129 size_t marking_initiating_used_threshold =
1130 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent;
1131 size_t cur_used_bytes = _g1->non_young_capacity_bytes();
1132 size_t alloc_byte_size = alloc_word_size * HeapWordSize;
1133
1134 if ((cur_used_bytes + alloc_byte_size) > marking_initiating_used_threshold) {
1135 if (gcs_are_young()) {
1136 ergo_verbose5(ErgoConcCycles,
1137 "request concurrent cycle initiation",
1138 ergo_format_reason("occupancy higher than threshold")
1139 ergo_format_byte("occupancy")
1140 ergo_format_byte("allocation request")
1141 ergo_format_byte_perc("threshold")
1142 ergo_format_str("source"),
1143 cur_used_bytes,
1144 alloc_byte_size,
1145 marking_initiating_used_threshold,
1146 (double) InitiatingHeapOccupancyPercent,
1147 source);
1148 return true;
1149 } else {
1150 ergo_verbose5(ErgoConcCycles,
1151 "do not request concurrent cycle initiation",
1152 ergo_format_reason("still doing mixed collections")
1153 ergo_format_byte("occupancy")
1154 ergo_format_byte("allocation request")
1155 ergo_format_byte_perc("threshold")
1156 ergo_format_str("source"),
1157 cur_used_bytes,
1158 alloc_byte_size,
1159 marking_initiating_used_threshold,
1160 (double) InitiatingHeapOccupancyPercent,
1161 source);
1162 }
1163 }
1164
1165 return false;
1166 }
1167
1168 // Anything below that is considered to be zero
1169 #define MIN_TIMER_GRANULARITY 0.0000001
1170
1171 void G1CollectorPolicy::record_collection_pause_end(int no_of_gc_threads) {
1172 double end_time_sec = os::elapsedTime();
1173 double elapsed_ms = _last_pause_time_ms;
1174 bool parallel = G1CollectedHeap::use_parallel_gc_threads();
1175 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(),
1176 "otherwise, the subtraction below does not make sense");
1177 size_t rs_size =
1178 _cur_collection_pause_used_regions_at_start - cset_region_length();
1179 size_t cur_used_bytes = _g1->used();
1180 assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
1181 bool last_pause_included_initial_mark = false;
1182 bool update_stats = !_g1->evacuation_failed();
1183 set_no_of_gc_threads(no_of_gc_threads);
1184
1185 #ifndef PRODUCT
1186 if (G1YoungSurvRateVerbose) {
1187 gclog_or_tty->print_cr("");
1188 _short_lived_surv_rate_group->print();
1189 // do that for any other surv rate groups too
1190 }
1191 #endif // PRODUCT
1192
1193 last_pause_included_initial_mark = during_initial_mark_pause();
1194 if (last_pause_included_initial_mark) {
1195 record_concurrent_mark_init_end(0.0);
1196 } else if (!_last_young_gc && need_to_start_conc_mark("end of GC")) {
1197 // Note: this might have already been set, if during the last
1198 // pause we decided to start a cycle but at the beginning of
1199 // this pause we decided to postpone it. That's OK.
1200 set_initiate_conc_mark_if_possible();
1201 }
1202
1203 _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0,
1204 end_time_sec, false);
1205
1206 // This assert is exempted when we're doing parallel collection pauses,
1207 // because the fragmentation caused by the parallel GC allocation buffers
1208 // can lead to more memory being used during collection than was used
1209 // before. Best leave this out until the fragmentation problem is fixed.
1210 // Pauses in which evacuation failed can also lead to negative
1211 // collections, since no space is reclaimed from a region containing an
1212 // object whose evacuation failed.
1213 // Further, we're now always doing parallel collection. But I'm still
1214 // leaving this here as a placeholder for a more precise assertion later.
1215 // (DLD, 10/05.)
1216 assert((true || parallel) // Always using GC LABs now.
1217 || _g1->evacuation_failed()
1218 || _cur_collection_pause_used_at_start_bytes >= cur_used_bytes,
1219 "Negative collection");
1220
1221 size_t freed_bytes =
1222 _cur_collection_pause_used_at_start_bytes - cur_used_bytes;
1223 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes;
1224
1225 double survival_fraction =
1226 (double)surviving_bytes/
1227 (double)_collection_set_bytes_used_before;
1228
1229 // These values are used to update the summary information that is
1230 // displayed when TraceGen0Time is enabled, and are output as part
1231 // of the "finer" output, in the non-parallel case.
1232
1233 double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms);
1234 double satb_filtering_time = avg_value(_par_last_satb_filtering_times_ms);
1235 double update_rs_time = avg_value(_par_last_update_rs_times_ms);
1236 double update_rs_processed_buffers =
1237 sum_of_values(_par_last_update_rs_processed_buffers);
1238 double scan_rs_time = avg_value(_par_last_scan_rs_times_ms);
1239 double obj_copy_time = avg_value(_par_last_obj_copy_times_ms);
1240 double termination_time = avg_value(_par_last_termination_times_ms);
1241
1242 double known_time = ext_root_scan_time +
1243 satb_filtering_time +
1244 update_rs_time +
1245 scan_rs_time +
1246 obj_copy_time;
1247
1248 double other_time_ms = elapsed_ms;
1249
1250 // Subtract the root region scanning wait time. It's initialized to
1251 // zero at the start of the pause.
1252 other_time_ms -= _root_region_scan_wait_time_ms;
1253
1254 if (parallel) {
1255 other_time_ms -= _cur_collection_par_time_ms;
1256 } else {
1257 other_time_ms -= known_time;
1258 }
1259
1260 // Now subtract the time taken to fix up roots in generated code
1261 other_time_ms -= _cur_collection_code_root_fixup_time_ms;
1262
1263 // Subtract the time taken to clean the card table from the
1264 // current value of "other time"
1265 other_time_ms -= _cur_clear_ct_time_ms;
1266
1267 // TraceGen0Time and TraceGen1Time summary info updating.
1268 _all_pause_times_ms->add(elapsed_ms);
1269
1270 if (update_stats) {
1271 _summary->record_total_time_ms(elapsed_ms);
1272 _summary->record_other_time_ms(other_time_ms);
1273
1274 MainBodySummary* body_summary = _summary->main_body_summary();
1275 assert(body_summary != NULL, "should not be null!");
1276
1277 body_summary->record_root_region_scan_wait_time_ms(
1278 _root_region_scan_wait_time_ms);
1279 body_summary->record_ext_root_scan_time_ms(ext_root_scan_time);
1280 body_summary->record_satb_filtering_time_ms(satb_filtering_time);
1281 body_summary->record_update_rs_time_ms(update_rs_time);
1282 body_summary->record_scan_rs_time_ms(scan_rs_time);
1283 body_summary->record_obj_copy_time_ms(obj_copy_time);
1284
1285 if (parallel) {
1286 body_summary->record_parallel_time_ms(_cur_collection_par_time_ms);
1287 body_summary->record_termination_time_ms(termination_time);
1288
1289 double parallel_known_time = known_time + termination_time;
1290 double parallel_other_time = _cur_collection_par_time_ms - parallel_known_time;
1291 body_summary->record_parallel_other_time_ms(parallel_other_time);
1292 }
1293
1294 body_summary->record_clear_ct_time_ms(_cur_clear_ct_time_ms);
1295
1296 // We exempt parallel collection from this check because Alloc Buffer
1297 // fragmentation can produce negative collections. Same with evac
1298 // failure.
1299 // Further, we're now always doing parallel collection. But I'm still
1300 // leaving this here as a placeholder for a more precise assertion later.
1301 // (DLD, 10/05.
1302 assert((true || parallel)
1303 || _g1->evacuation_failed()
1304 || surviving_bytes <= _collection_set_bytes_used_before,
1305 "Or else negative collection!");
1306
1307 // this is where we update the allocation rate of the application
1308 double app_time_ms =
1309 (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms);
1310 if (app_time_ms < MIN_TIMER_GRANULARITY) {
1311 // This usually happens due to the timer not having the required
1312 // granularity. Some Linuxes are the usual culprits.
1313 // We'll just set it to something (arbitrarily) small.
1314 app_time_ms = 1.0;
1315 }
1316 // We maintain the invariant that all objects allocated by mutator
1317 // threads will be allocated out of eden regions. So, we can use
1318 // the eden region number allocated since the previous GC to
1319 // calculate the application's allocate rate. The only exception
1320 // to that is humongous objects that are allocated separately. But
1321 // given that humongous object allocations do not really affect
1322 // either the pause's duration nor when the next pause will take
1323 // place we can safely ignore them here.
1324 uint regions_allocated = eden_cset_region_length();
1325 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
1326 _alloc_rate_ms_seq->add(alloc_rate_ms);
1327
1328 double interval_ms =
1329 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
1330 update_recent_gc_times(end_time_sec, elapsed_ms);
1331 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
1332 if (recent_avg_pause_time_ratio() < 0.0 ||
1333 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) {
1334 #ifndef PRODUCT
1335 // Dump info to allow post-facto debugging
1336 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds");
1337 gclog_or_tty->print_cr("-------------------------------------------");
1338 gclog_or_tty->print_cr("Recent GC Times (ms):");
1339 _recent_gc_times_ms->dump();
1340 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec);
1341 _recent_prev_end_times_for_all_gcs_sec->dump();
1342 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f",
1343 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio());
1344 // In debug mode, terminate the JVM if the user wants to debug at this point.
1345 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above");
1346 #endif // !PRODUCT
1347 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
1348 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
1349 if (_recent_avg_pause_time_ratio < 0.0) {
1350 _recent_avg_pause_time_ratio = 0.0;
1351 } else {
1352 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
1353 _recent_avg_pause_time_ratio = 1.0;
1354 }
1355 }
1356 }
1357
1358 for (int i = 0; i < _aux_num; ++i) {
1359 if (_cur_aux_times_set[i]) {
1360 _all_aux_times_ms[i].add(_cur_aux_times_ms[i]);
1361 }
1362 }
1363
1364 if (G1Log::finer()) {
1365 bool print_marking_info =
1366 _g1->mark_in_progress() && !last_pause_included_initial_mark;
1367
1368 gclog_or_tty->print_cr("%s, %1.8lf secs]",
1369 (last_pause_included_initial_mark) ? " (initial-mark)" : "",
1370 elapsed_ms / 1000.0);
1371
1372 if (_root_region_scan_wait_time_ms > 0.0) {
1373 print_stats(1, "Root Region Scan Waiting", _root_region_scan_wait_time_ms);
1374 }
1375 if (parallel) {
1376 print_stats(1, "Parallel Time", _cur_collection_par_time_ms);
1377 print_par_stats(2, "GC Worker Start", _par_last_gc_worker_start_times_ms);
1378 print_par_stats(2, "Ext Root Scanning", _par_last_ext_root_scan_times_ms);
1379 if (print_marking_info) {
1380 print_par_stats(2, "SATB Filtering", _par_last_satb_filtering_times_ms);
1381 }
1382 print_par_stats(2, "Update RS", _par_last_update_rs_times_ms);
1383 if (G1Log::finest()) {
1384 print_par_sizes(3, "Processed Buffers", _par_last_update_rs_processed_buffers);
1385 }
1386 print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms);
1387 print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms);
1388 print_par_stats(2, "Termination", _par_last_termination_times_ms);
1389 if (G1Log::finest()) {
1390 print_par_sizes(3, "Termination Attempts", _par_last_termination_attempts);
1391 }
1392
1393 for (int i = 0; i < _parallel_gc_threads; i++) {
1394 _par_last_gc_worker_times_ms[i] = _par_last_gc_worker_end_times_ms[i] -
1395 _par_last_gc_worker_start_times_ms[i];
1396
1397 double worker_known_time = _par_last_ext_root_scan_times_ms[i] +
1398 _par_last_satb_filtering_times_ms[i] +
1399 _par_last_update_rs_times_ms[i] +
1400 _par_last_scan_rs_times_ms[i] +
1401 _par_last_obj_copy_times_ms[i] +
1402 _par_last_termination_times_ms[i];
1403
1404 _par_last_gc_worker_other_times_ms[i] = _par_last_gc_worker_times_ms[i] -
1405 worker_known_time;
1406 }
1407
1408 print_par_stats(2, "GC Worker Other", _par_last_gc_worker_other_times_ms);
1409 print_par_stats(2, "GC Worker Total", _par_last_gc_worker_times_ms);
1410 print_par_stats(2, "GC Worker End", _par_last_gc_worker_end_times_ms);
1411 } else {
1412 print_stats(1, "Ext Root Scanning", ext_root_scan_time);
1413 if (print_marking_info) {
1414 print_stats(1, "SATB Filtering", satb_filtering_time);
1415 }
1416 print_stats(1, "Update RS", update_rs_time);
1417 if (G1Log::finest()) {
1418 print_stats(2, "Processed Buffers", (int)update_rs_processed_buffers);
1419 }
1420 print_stats(1, "Scan RS", scan_rs_time);
1421 print_stats(1, "Object Copying", obj_copy_time);
1422 }
1423 print_stats(1, "Code Root Fixup", _cur_collection_code_root_fixup_time_ms);
1424 print_stats(1, "Clear CT", _cur_clear_ct_time_ms);
1425 #ifndef PRODUCT
1426 print_stats(1, "Cur Clear CC", _cur_clear_cc_time_ms);
1427 print_stats(1, "Cum Clear CC", _cum_clear_cc_time_ms);
1428 print_stats(1, "Min Clear CC", _min_clear_cc_time_ms);
1429 print_stats(1, "Max Clear CC", _max_clear_cc_time_ms);
1430 if (_num_cc_clears > 0) {
1431 print_stats(1, "Avg Clear CC", _cum_clear_cc_time_ms / ((double)_num_cc_clears));
1432 }
1433 #endif
1434 print_stats(1, "Other", other_time_ms);
1435 print_stats(2, "Choose CSet",
1436 (_recorded_young_cset_choice_time_ms +
1437 _recorded_non_young_cset_choice_time_ms));
1438 print_stats(2, "Ref Proc", _cur_ref_proc_time_ms);
1439 print_stats(2, "Ref Enq", _cur_ref_enq_time_ms);
1440 print_stats(2, "Free CSet",
1441 (_recorded_young_free_cset_time_ms +
1442 _recorded_non_young_free_cset_time_ms));
1443
1444 for (int i = 0; i < _aux_num; ++i) {
1445 if (_cur_aux_times_set[i]) {
1446 char buffer[96];
1447 sprintf(buffer, "Aux%d", i);
1448 print_stats(1, buffer, _cur_aux_times_ms[i]);
1449 }
1450 }
1451 }
1452
1453 bool new_in_marking_window = _in_marking_window;
1454 bool new_in_marking_window_im = false;
1455 if (during_initial_mark_pause()) {
1456 new_in_marking_window = true;
1457 new_in_marking_window_im = true;
1458 }
1459
1460 if (_last_young_gc) {
1461 // This is supposed to to be the "last young GC" before we start
1462 // doing mixed GCs. Here we decide whether to start mixed GCs or not.
1463
1464 if (!last_pause_included_initial_mark) {
1465 if (next_gc_should_be_mixed("start mixed GCs",
1466 "do not start mixed GCs")) {
1467 set_gcs_are_young(false);
1468 }
1469 } else {
1470 ergo_verbose0(ErgoMixedGCs,
1471 "do not start mixed GCs",
1472 ergo_format_reason("concurrent cycle is about to start"));
1473 }
1474 _last_young_gc = false;
1475 }
1476
1477 if (!_last_gc_was_young) {
1478 // This is a mixed GC. Here we decide whether to continue doing
1479 // mixed GCs or not.
1480
1481 if (!next_gc_should_be_mixed("continue mixed GCs",
1482 "do not continue mixed GCs")) {
1483 set_gcs_are_young(true);
1484 }
1485 }
1486
1487 _short_lived_surv_rate_group->start_adding_regions();
1488 // do that for any other surv rate groupsx
1489
1490 if (update_stats) {
1491 double pause_time_ms = elapsed_ms;
1492
1493 size_t diff = 0;
1494 if (_max_pending_cards >= _pending_cards) {
1495 diff = _max_pending_cards - _pending_cards;
1496 }
1497 _pending_card_diff_seq->add((double) diff);
1498
1499 double cost_per_card_ms = 0.0;
1500 if (_pending_cards > 0) {
1501 cost_per_card_ms = update_rs_time / (double) _pending_cards;
1502 _cost_per_card_ms_seq->add(cost_per_card_ms);
1503 }
1504
1505 size_t cards_scanned = _g1->cards_scanned();
1506
1507 double cost_per_entry_ms = 0.0;
1508 if (cards_scanned > 10) {
1509 cost_per_entry_ms = scan_rs_time / (double) cards_scanned;
1510 if (_last_gc_was_young) {
1511 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
1512 } else {
1513 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms);
1514 }
1515 }
1516
1517 if (_max_rs_lengths > 0) {
1518 double cards_per_entry_ratio =
1519 (double) cards_scanned / (double) _max_rs_lengths;
1520 if (_last_gc_was_young) {
1521 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1522 } else {
1523 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1524 }
1525 }
1526
1527 // This is defensive. For a while _max_rs_lengths could get
1528 // smaller than _recorded_rs_lengths which was causing
1529 // rs_length_diff to get very large and mess up the RSet length
1530 // predictions. The reason was unsafe concurrent updates to the
1531 // _inc_cset_recorded_rs_lengths field which the code below guards
1532 // against (see CR 7118202). This bug has now been fixed (see CR
1533 // 7119027). However, I'm still worried that
1534 // _inc_cset_recorded_rs_lengths might still end up somewhat
1535 // inaccurate. The concurrent refinement thread calculates an
1536 // RSet's length concurrently with other CR threads updating it
1537 // which might cause it to calculate the length incorrectly (if,
1538 // say, it's in mid-coarsening). So I'll leave in the defensive
1539 // conditional below just in case.
1540 size_t rs_length_diff = 0;
1541 if (_max_rs_lengths > _recorded_rs_lengths) {
1542 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths;
1543 }
1544 _rs_length_diff_seq->add((double) rs_length_diff);
1545
1546 size_t copied_bytes = surviving_bytes;
1547 double cost_per_byte_ms = 0.0;
1548 if (copied_bytes > 0) {
1549 cost_per_byte_ms = obj_copy_time / (double) copied_bytes;
1550 if (_in_marking_window) {
1551 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
1552 } else {
1553 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
1554 }
1555 }
1556
1557 double all_other_time_ms = pause_time_ms -
1558 (update_rs_time + scan_rs_time + obj_copy_time + termination_time);
1559
1560 double young_other_time_ms = 0.0;
1561 if (young_cset_region_length() > 0) {
1562 young_other_time_ms =
1563 _recorded_young_cset_choice_time_ms +
1564 _recorded_young_free_cset_time_ms;
1565 _young_other_cost_per_region_ms_seq->add(young_other_time_ms /
1566 (double) young_cset_region_length());
1567 }
1568 double non_young_other_time_ms = 0.0;
1569 if (old_cset_region_length() > 0) {
1570 non_young_other_time_ms =
1571 _recorded_non_young_cset_choice_time_ms +
1572 _recorded_non_young_free_cset_time_ms;
1573
1574 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms /
1575 (double) old_cset_region_length());
1576 }
1577
1578 double constant_other_time_ms = all_other_time_ms -
1579 (young_other_time_ms + non_young_other_time_ms);
1580 _constant_other_time_ms_seq->add(constant_other_time_ms);
1581
1582 double survival_ratio = 0.0;
1583 if (_bytes_in_collection_set_before_gc > 0) {
1584 survival_ratio = (double) _bytes_copied_during_gc /
1585 (double) _bytes_in_collection_set_before_gc;
1586 }
1587
1588 _pending_cards_seq->add((double) _pending_cards);
1589 _rs_lengths_seq->add((double) _max_rs_lengths);
1590 }
1591
1592 _in_marking_window = new_in_marking_window;
1593 _in_marking_window_im = new_in_marking_window_im;
1594 _free_regions_at_end_of_collection = _g1->free_regions();
1595 update_young_list_target_length();
1596
1597 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
1598 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
1599 adjust_concurrent_refinement(update_rs_time, update_rs_processed_buffers, update_rs_time_goal_ms);
1600
1601 _collectionSetChooser->verify();
1602 }
1603
1604 #define EXT_SIZE_FORMAT "%d%s"
1605 #define EXT_SIZE_PARAMS(bytes) \
1606 byte_size_in_proper_unit((bytes)), \
1607 proper_unit_for_byte_size((bytes))
1608
1609 void G1CollectorPolicy::print_heap_transition() {
1610 if (G1Log::finer()) {
1611 YoungList* young_list = _g1->young_list();
1612 size_t eden_bytes = young_list->eden_used_bytes();
1613 size_t survivor_bytes = young_list->survivor_used_bytes();
1614 size_t used_before_gc = _cur_collection_pause_used_at_start_bytes;
1615 size_t used = _g1->used();
1616 size_t capacity = _g1->capacity();
1617 size_t eden_capacity =
1618 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_bytes;
1619
1620 gclog_or_tty->print_cr(
1621 " [Eden: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT") "
1622 "Survivors: "EXT_SIZE_FORMAT"->"EXT_SIZE_FORMAT" "
1623 "Heap: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"
1624 EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")]",
1625 EXT_SIZE_PARAMS(_eden_bytes_before_gc),
1626 EXT_SIZE_PARAMS(_prev_eden_capacity),
1627 EXT_SIZE_PARAMS(eden_bytes),
1628 EXT_SIZE_PARAMS(eden_capacity),
1629 EXT_SIZE_PARAMS(_survivor_bytes_before_gc),
1630 EXT_SIZE_PARAMS(survivor_bytes),
1631 EXT_SIZE_PARAMS(used_before_gc),
1632 EXT_SIZE_PARAMS(_capacity_before_gc),
1633 EXT_SIZE_PARAMS(used),
1634 EXT_SIZE_PARAMS(capacity));
1635
1636 _prev_eden_capacity = eden_capacity;
1637 } else if (G1Log::fine()) {
1638 _g1->print_size_transition(gclog_or_tty,
1639 _cur_collection_pause_used_at_start_bytes,
1640 _g1->used(), _g1->capacity());
1641 }
1642 }
1643
1644 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1645 double update_rs_processed_buffers,
1646 double goal_ms) {
1647 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1648 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1649
1650 if (G1UseAdaptiveConcRefinement) {
1651 const int k_gy = 3, k_gr = 6;
1652 const double inc_k = 1.1, dec_k = 0.9;
1653
1654 int g = cg1r->green_zone();
1655 if (update_rs_time > goal_ms) {
1656 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1657 } else {
1658 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1659 g = (int)MAX2(g * inc_k, g + 1.0);
1660 }
1661 }
1662 // Change the refinement threads params
1663 cg1r->set_green_zone(g);
1664 cg1r->set_yellow_zone(g * k_gy);
1665 cg1r->set_red_zone(g * k_gr);
1666 cg1r->reinitialize_threads();
1667
1668 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1);
1669 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1670 cg1r->yellow_zone());
1671 // Change the barrier params
1672 dcqs.set_process_completed_threshold(processing_threshold);
1673 dcqs.set_max_completed_queue(cg1r->red_zone());
1674 }
1675
1676 int curr_queue_size = dcqs.completed_buffers_num();
1677 if (curr_queue_size >= cg1r->yellow_zone()) {
1678 dcqs.set_completed_queue_padding(curr_queue_size);
1679 } else {
1680 dcqs.set_completed_queue_padding(0);
1681 }
1682 dcqs.notify_if_necessary();
1683 }
1684
1685 double
1686 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) {
1687 size_t rs_length = predict_rs_length_diff();
1688 size_t card_num;
1689 if (gcs_are_young()) {
1690 card_num = predict_young_card_num(rs_length);
1691 } else {
1692 card_num = predict_non_young_card_num(rs_length);
1693 }
1694 return predict_base_elapsed_time_ms(pending_cards, card_num);
1695 }
1696
1697 double
1698 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
1699 size_t scanned_cards) {
1700 return
1701 predict_rs_update_time_ms(pending_cards) +
1702 predict_rs_scan_time_ms(scanned_cards) +
1703 predict_constant_other_time_ms();
1704 }
1705
1706 double
1707 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
1708 bool young) {
1709 size_t rs_length = hr->rem_set()->occupied();
1710 size_t card_num;
1711 if (gcs_are_young()) {
1712 card_num = predict_young_card_num(rs_length);
1713 } else {
1714 card_num = predict_non_young_card_num(rs_length);
1715 }
1716 size_t bytes_to_copy = predict_bytes_to_copy(hr);
1717
1718 double region_elapsed_time_ms =
1719 predict_rs_scan_time_ms(card_num) +
1720 predict_object_copy_time_ms(bytes_to_copy);
1721
1722 if (young)
1723 region_elapsed_time_ms += predict_young_other_time_ms(1);
1724 else
1725 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
1726
1727 return region_elapsed_time_ms;
1728 }
1729
1730 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) {
1731 size_t bytes_to_copy;
1732 if (hr->is_marked())
1733 bytes_to_copy = hr->max_live_bytes();
1734 else {
1735 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant");
1736 int age = hr->age_in_surv_rate_group();
1737 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
1738 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate);
1739 }
1740 return bytes_to_copy;
1741 }
1742
1743 void
1744 G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length,
1745 uint survivor_cset_region_length) {
1746 _eden_cset_region_length = eden_cset_region_length;
1747 _survivor_cset_region_length = survivor_cset_region_length;
1748 _old_cset_region_length = 0;
1749 }
1750
1751 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) {
1752 _recorded_rs_lengths = rs_lengths;
1753 }
1754
1755 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
1756 double elapsed_ms) {
1757 _recent_gc_times_ms->add(elapsed_ms);
1758 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
1759 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
1760 }
1761
1762 size_t G1CollectorPolicy::expansion_amount() {
1763 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0;
1764 double threshold = _gc_overhead_perc;
1765 if (recent_gc_overhead > threshold) {
1766 // We will double the existing space, or take
1767 // G1ExpandByPercentOfAvailable % of the available expansion
1768 // space, whichever is smaller, bounded below by a minimum
1769 // expansion (unless that's all that's left.)
1770 const size_t min_expand_bytes = 1*M;
1771 size_t reserved_bytes = _g1->max_capacity();
1772 size_t committed_bytes = _g1->capacity();
1773 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
1774 size_t expand_bytes;
1775 size_t expand_bytes_via_pct =
1776 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
1777 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
1778 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
1779 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
1780
1781 ergo_verbose5(ErgoHeapSizing,
1782 "attempt heap expansion",
1783 ergo_format_reason("recent GC overhead higher than "
1784 "threshold after GC")
1785 ergo_format_perc("recent GC overhead")
1786 ergo_format_perc("threshold")
1787 ergo_format_byte("uncommitted")
1788 ergo_format_byte_perc("calculated expansion amount"),
1789 recent_gc_overhead, threshold,
1790 uncommitted_bytes,
1791 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable);
1792
1793 return expand_bytes;
1794 } else {
1795 return 0;
1796 }
1797 }
1798
1799 class CountCSClosure: public HeapRegionClosure {
1800 G1CollectorPolicy* _g1_policy;
1801 public:
1802 CountCSClosure(G1CollectorPolicy* g1_policy) :
1803 _g1_policy(g1_policy) {}
1804 bool doHeapRegion(HeapRegion* r) {
1805 _g1_policy->_bytes_in_collection_set_before_gc += r->used();
1806 return false;
1807 }
1808 };
1809
1810 void G1CollectorPolicy::count_CS_bytes_used() {
1811 CountCSClosure cs_closure(this);
1812 _g1->collection_set_iterate(&cs_closure);
1813 }
1814
1815 void G1CollectorPolicy::print_summary(int level,
1816 const char* str,
1817 NumberSeq* seq) const {
1818 double sum = seq->sum();
1819 LineBuffer(level + 1).append_and_print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)",
1820 str, sum / 1000.0, seq->avg());
1821 }
1822
1823 void G1CollectorPolicy::print_summary_sd(int level,
1824 const char* str,
1825 NumberSeq* seq) const {
1826 print_summary(level, str, seq);
1827 LineBuffer(level + 6).append_and_print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
1828 seq->num(), seq->sd(), seq->maximum());
1829 }
1830
1831 void G1CollectorPolicy::check_other_times(int level,
1832 NumberSeq* other_times_ms,
1833 NumberSeq* calc_other_times_ms) const {
1834 bool should_print = false;
1835 LineBuffer buf(level + 2);
1836
1837 double max_sum = MAX2(fabs(other_times_ms->sum()),
1838 fabs(calc_other_times_ms->sum()));
1839 double min_sum = MIN2(fabs(other_times_ms->sum()),
1840 fabs(calc_other_times_ms->sum()));
1841 double sum_ratio = max_sum / min_sum;
1842 if (sum_ratio > 1.1) {
1843 should_print = true;
1844 buf.append_and_print_cr("## CALCULATED OTHER SUM DOESN'T MATCH RECORDED ###");
1845 }
1846
1847 double max_avg = MAX2(fabs(other_times_ms->avg()),
1848 fabs(calc_other_times_ms->avg()));
1849 double min_avg = MIN2(fabs(other_times_ms->avg()),
1850 fabs(calc_other_times_ms->avg()));
1851 double avg_ratio = max_avg / min_avg;
1852 if (avg_ratio > 1.1) {
1853 should_print = true;
1854 buf.append_and_print_cr("## CALCULATED OTHER AVG DOESN'T MATCH RECORDED ###");
1855 }
1856
1857 if (other_times_ms->sum() < -0.01) {
1858 buf.append_and_print_cr("## RECORDED OTHER SUM IS NEGATIVE ###");
1859 }
1860
1861 if (other_times_ms->avg() < -0.01) {
1862 buf.append_and_print_cr("## RECORDED OTHER AVG IS NEGATIVE ###");
1863 }
1864
1865 if (calc_other_times_ms->sum() < -0.01) {
1866 should_print = true;
1867 buf.append_and_print_cr("## CALCULATED OTHER SUM IS NEGATIVE ###");
1868 }
1869
1870 if (calc_other_times_ms->avg() < -0.01) {
1871 should_print = true;
1872 buf.append_and_print_cr("## CALCULATED OTHER AVG IS NEGATIVE ###");
1873 }
1874
1875 if (should_print)
1876 print_summary(level, "Other(Calc)", calc_other_times_ms);
1877 }
1878
1879 void G1CollectorPolicy::print_summary(PauseSummary* summary) const {
1880 bool parallel = G1CollectedHeap::use_parallel_gc_threads();
1881 MainBodySummary* body_summary = summary->main_body_summary();
1882 if (summary->get_total_seq()->num() > 0) {
1883 print_summary_sd(0, "Evacuation Pauses", summary->get_total_seq());
1884 if (body_summary != NULL) {
1885 print_summary(1, "Root Region Scan Wait", body_summary->get_root_region_scan_wait_seq());
1886 if (parallel) {
1887 print_summary(1, "Parallel Time", body_summary->get_parallel_seq());
1888 print_summary(2, "Ext Root Scanning", body_summary->get_ext_root_scan_seq());
1889 print_summary(2, "SATB Filtering", body_summary->get_satb_filtering_seq());
1890 print_summary(2, "Update RS", body_summary->get_update_rs_seq());
1891 print_summary(2, "Scan RS", body_summary->get_scan_rs_seq());
1892 print_summary(2, "Object Copy", body_summary->get_obj_copy_seq());
1893 print_summary(2, "Termination", body_summary->get_termination_seq());
1894 print_summary(2, "Parallel Other", body_summary->get_parallel_other_seq());
1895 {
1896 NumberSeq* other_parts[] = {
1897 body_summary->get_ext_root_scan_seq(),
1898 body_summary->get_satb_filtering_seq(),
1899 body_summary->get_update_rs_seq(),
1900 body_summary->get_scan_rs_seq(),
1901 body_summary->get_obj_copy_seq(),
1902 body_summary->get_termination_seq()
1903 };
1904 NumberSeq calc_other_times_ms(body_summary->get_parallel_seq(),
1905 6, other_parts);
1906 check_other_times(2, body_summary->get_parallel_other_seq(),
1907 &calc_other_times_ms);
1908 }
1909 } else {
1910 print_summary(1, "Ext Root Scanning", body_summary->get_ext_root_scan_seq());
1911 print_summary(1, "SATB Filtering", body_summary->get_satb_filtering_seq());
1912 print_summary(1, "Update RS", body_summary->get_update_rs_seq());
1913 print_summary(1, "Scan RS", body_summary->get_scan_rs_seq());
1914 print_summary(1, "Object Copy", body_summary->get_obj_copy_seq());
1915 }
1916 }
1917 print_summary(1, "Clear CT", body_summary->get_clear_ct_seq());
1918 print_summary(1, "Other", summary->get_other_seq());
1919 {
1920 if (body_summary != NULL) {
1921 NumberSeq calc_other_times_ms;
1922 if (parallel) {
1923 // parallel
1924 NumberSeq* other_parts[] = {
1925 body_summary->get_root_region_scan_wait_seq(),
1926 body_summary->get_parallel_seq(),
1927 body_summary->get_clear_ct_seq()
1928 };
1929 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
1930 3, other_parts);
1931 } else {
1932 // serial
1933 NumberSeq* other_parts[] = {
1934 body_summary->get_root_region_scan_wait_seq(),
1935 body_summary->get_update_rs_seq(),
1936 body_summary->get_ext_root_scan_seq(),
1937 body_summary->get_satb_filtering_seq(),
1938 body_summary->get_scan_rs_seq(),
1939 body_summary->get_obj_copy_seq()
1940 };
1941 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
1942 6, other_parts);
1943 }
1944 check_other_times(1, summary->get_other_seq(), &calc_other_times_ms);
1945 }
1946 }
1947 } else {
1948 LineBuffer(1).append_and_print_cr("none");
1949 }
1950 LineBuffer(0).append_and_print_cr("");
1951 }
1952
1953 void G1CollectorPolicy::print_tracing_info() const {
1954 if (TraceGen0Time) {
1955 gclog_or_tty->print_cr("ALL PAUSES");
1956 print_summary_sd(0, "Total", _all_pause_times_ms);
1957 gclog_or_tty->print_cr("");
1958 gclog_or_tty->print_cr("");
1959 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num);
1960 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num);
1961 gclog_or_tty->print_cr("");
1962
1963 gclog_or_tty->print_cr("EVACUATION PAUSES");
1964 print_summary(_summary);
1965
1966 gclog_or_tty->print_cr("MISC");
1967 print_summary_sd(0, "Stop World", _all_stop_world_times_ms);
1968 print_summary_sd(0, "Yields", _all_yield_times_ms);
1969 for (int i = 0; i < _aux_num; ++i) {
1970 if (_all_aux_times_ms[i].num() > 0) {
1971 char buffer[96];
1972 sprintf(buffer, "Aux%d", i);
1973 print_summary_sd(0, buffer, &_all_aux_times_ms[i]);
1974 }
1975 }
1976 }
1977 if (TraceGen1Time) {
1978 if (_all_full_gc_times_ms->num() > 0) {
1979 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s",
1980 _all_full_gc_times_ms->num(),
1981 _all_full_gc_times_ms->sum() / 1000.0);
1982 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times_ms->avg());
1983 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]",
1984 _all_full_gc_times_ms->sd(),
1985 _all_full_gc_times_ms->maximum());
1986 }
1987 }
1988 }
1989
1990 void G1CollectorPolicy::print_yg_surv_rate_info() const {
1991 #ifndef PRODUCT
1992 _short_lived_surv_rate_group->print_surv_rate_summary();
1993 // add this call for any other surv rate groups
1994 #endif // PRODUCT
1995 }
1996
1997 #ifndef PRODUCT
1998 // for debugging, bit of a hack...
1999 static char*
2000 region_num_to_mbs(int length) {
2001 static char buffer[64];
2002 double bytes = (double) (length * HeapRegion::GrainBytes);
2003 double mbs = bytes / (double) (1024 * 1024);
2004 sprintf(buffer, "%7.2lfMB", mbs);
2005 return buffer;
2006 }
2007 #endif // PRODUCT
2008
2009 uint G1CollectorPolicy::max_regions(int purpose) {
2010 switch (purpose) {
2011 case GCAllocForSurvived:
2012 return _max_survivor_regions;
2013 case GCAllocForTenured:
2014 return REGIONS_UNLIMITED;
2015 default:
2016 ShouldNotReachHere();
2017 return REGIONS_UNLIMITED;
2018 };
2019 }
2020
2021 void G1CollectorPolicy::update_max_gc_locker_expansion() {
2022 uint expansion_region_num = 0;
2023 if (GCLockerEdenExpansionPercent > 0) {
2024 double perc = (double) GCLockerEdenExpansionPercent / 100.0;
2025 double expansion_region_num_d = perc * (double) _young_list_target_length;
2026 // We use ceiling so that if expansion_region_num_d is > 0.0 (but
2027 // less than 1.0) we'll get 1.
2028 expansion_region_num = (uint) ceil(expansion_region_num_d);
2029 } else {
2030 assert(expansion_region_num == 0, "sanity");
2031 }
2032 _young_list_max_length = _young_list_target_length + expansion_region_num;
2033 assert(_young_list_target_length <= _young_list_max_length, "post-condition");
2034 }
2035
2036 // Calculates survivor space parameters.
2037 void G1CollectorPolicy::update_survivors_policy() {
2038 double max_survivor_regions_d =
2039 (double) _young_list_target_length / (double) SurvivorRatio;
2040 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but
2041 // smaller than 1.0) we'll get 1.
2042 _max_survivor_regions = (uint) ceil(max_survivor_regions_d);
2043
2044 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
2045 HeapRegion::GrainWords * _max_survivor_regions);
2046 }
2047
2048 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(
2049 GCCause::Cause gc_cause) {
2050 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2051 if (!during_cycle) {
2052 ergo_verbose1(ErgoConcCycles,
2053 "request concurrent cycle initiation",
2054 ergo_format_reason("requested by GC cause")
2055 ergo_format_str("GC cause"),
2056 GCCause::to_string(gc_cause));
2057 set_initiate_conc_mark_if_possible();
2058 return true;
2059 } else {
2060 ergo_verbose1(ErgoConcCycles,
2061 "do not request concurrent cycle initiation",
2062 ergo_format_reason("concurrent cycle already in progress")
2063 ergo_format_str("GC cause"),
2064 GCCause::to_string(gc_cause));
2065 return false;
2066 }
2067 }
2068
2069 void
2070 G1CollectorPolicy::decide_on_conc_mark_initiation() {
2071 // We are about to decide on whether this pause will be an
2072 // initial-mark pause.
2073
2074 // First, during_initial_mark_pause() should not be already set. We
2075 // will set it here if we have to. However, it should be cleared by
2076 // the end of the pause (it's only set for the duration of an
2077 // initial-mark pause).
2078 assert(!during_initial_mark_pause(), "pre-condition");
2079
2080 if (initiate_conc_mark_if_possible()) {
2081 // We had noticed on a previous pause that the heap occupancy has
2082 // gone over the initiating threshold and we should start a
2083 // concurrent marking cycle. So we might initiate one.
2084
2085 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2086 if (!during_cycle) {
2087 // The concurrent marking thread is not "during a cycle", i.e.,
2088 // it has completed the last one. So we can go ahead and
2089 // initiate a new cycle.
2090
2091 set_during_initial_mark_pause();
2092 // We do not allow mixed GCs during marking.
2093 if (!gcs_are_young()) {
2094 set_gcs_are_young(true);
2095 ergo_verbose0(ErgoMixedGCs,
2096 "end mixed GCs",
2097 ergo_format_reason("concurrent cycle is about to start"));
2098 }
2099
2100 // And we can now clear initiate_conc_mark_if_possible() as
2101 // we've already acted on it.
2102 clear_initiate_conc_mark_if_possible();
2103
2104 ergo_verbose0(ErgoConcCycles,
2105 "initiate concurrent cycle",
2106 ergo_format_reason("concurrent cycle initiation requested"));
2107 } else {
2108 // The concurrent marking thread is still finishing up the
2109 // previous cycle. If we start one right now the two cycles
2110 // overlap. In particular, the concurrent marking thread might
2111 // be in the process of clearing the next marking bitmap (which
2112 // we will use for the next cycle if we start one). Starting a
2113 // cycle now will be bad given that parts of the marking
2114 // information might get cleared by the marking thread. And we
2115 // cannot wait for the marking thread to finish the cycle as it
2116 // periodically yields while clearing the next marking bitmap
2117 // and, if it's in a yield point, it's waiting for us to
2118 // finish. So, at this point we will not start a cycle and we'll
2119 // let the concurrent marking thread complete the last one.
2120 ergo_verbose0(ErgoConcCycles,
2121 "do not initiate concurrent cycle",
2122 ergo_format_reason("concurrent cycle already in progress"));
2123 }
2124 }
2125 }
2126
2127 class KnownGarbageClosure: public HeapRegionClosure {
2128 G1CollectedHeap* _g1h;
2129 CollectionSetChooser* _hrSorted;
2130
2131 public:
2132 KnownGarbageClosure(CollectionSetChooser* hrSorted) :
2133 _g1h(G1CollectedHeap::heap()), _hrSorted(hrSorted) { }
2134
2135 bool doHeapRegion(HeapRegion* r) {
2136 // We only include humongous regions in collection
2137 // sets when concurrent mark shows that their contained object is
2138 // unreachable.
2139
2140 // Do we have any marking information for this region?
2141 if (r->is_marked()) {
2142 // We will skip any region that's currently used as an old GC
2143 // alloc region (we should not consider those for collection
2144 // before we fill them up).
2145 if (_hrSorted->should_add(r) && !_g1h->is_old_gc_alloc_region(r)) {
2146 _hrSorted->add_region(r);
2147 }
2148 }
2149 return false;
2150 }
2151 };
2152
2153 class ParKnownGarbageHRClosure: public HeapRegionClosure {
2154 G1CollectedHeap* _g1h;
2155 CollectionSetChooser* _hrSorted;
2156 uint _marked_regions_added;
2157 size_t _reclaimable_bytes_added;
2158 uint _chunk_size;
2159 uint _cur_chunk_idx;
2160 uint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end)
2161
2162 void get_new_chunk() {
2163 _cur_chunk_idx = _hrSorted->claim_array_chunk(_chunk_size);
2164 _cur_chunk_end = _cur_chunk_idx + _chunk_size;
2165 }
2166 void add_region(HeapRegion* r) {
2167 if (_cur_chunk_idx == _cur_chunk_end) {
2168 get_new_chunk();
2169 }
2170 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition");
2171 _hrSorted->set_region(_cur_chunk_idx, r);
2172 _marked_regions_added++;
2173 _reclaimable_bytes_added += r->reclaimable_bytes();
2174 _cur_chunk_idx++;
2175 }
2176
2177 public:
2178 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
2179 uint chunk_size) :
2180 _g1h(G1CollectedHeap::heap()),
2181 _hrSorted(hrSorted), _chunk_size(chunk_size),
2182 _marked_regions_added(0), _reclaimable_bytes_added(0),
2183 _cur_chunk_idx(0), _cur_chunk_end(0) { }
2184
2185 bool doHeapRegion(HeapRegion* r) {
2186 // Do we have any marking information for this region?
2187 if (r->is_marked()) {
2188 // We will skip any region that's currently used as an old GC
2189 // alloc region (we should not consider those for collection
2190 // before we fill them up).
2191 if (_hrSorted->should_add(r) && !_g1h->is_old_gc_alloc_region(r)) {
2192 add_region(r);
2193 }
2194 }
2195 return false;
2196 }
2197 uint marked_regions_added() { return _marked_regions_added; }
2198 size_t reclaimable_bytes_added() { return _reclaimable_bytes_added; }
2199 };
2200
2201 class ParKnownGarbageTask: public AbstractGangTask {
2202 CollectionSetChooser* _hrSorted;
2203 uint _chunk_size;
2204 G1CollectedHeap* _g1;
2205 public:
2206 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size) :
2207 AbstractGangTask("ParKnownGarbageTask"),
2208 _hrSorted(hrSorted), _chunk_size(chunk_size),
2209 _g1(G1CollectedHeap::heap()) { }
2210
2211 void work(uint worker_id) {
2212 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size);
2213
2214 // Back to zero for the claim value.
2215 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, worker_id,
2216 _g1->workers()->active_workers(),
2217 HeapRegion::InitialClaimValue);
2218 uint regions_added = parKnownGarbageCl.marked_regions_added();
2219 size_t reclaimable_bytes_added =
2220 parKnownGarbageCl.reclaimable_bytes_added();
2221 _hrSorted->update_totals(regions_added, reclaimable_bytes_added);
2222 }
2223 };
2224
2225 void
2226 G1CollectorPolicy::record_concurrent_mark_cleanup_end(int no_of_gc_threads) {
2227 _collectionSetChooser->clear();
2228
2229 uint region_num = _g1->n_regions();
2230 if (G1CollectedHeap::use_parallel_gc_threads()) {
2231 const uint OverpartitionFactor = 4;
2232 uint WorkUnit;
2233 // The use of MinChunkSize = 8 in the original code
2234 // causes some assertion failures when the total number of
2235 // region is less than 8. The code here tries to fix that.
2236 // Should the original code also be fixed?
2237 if (no_of_gc_threads > 0) {
2238 const uint MinWorkUnit = MAX2(region_num / no_of_gc_threads, 1U);
2239 WorkUnit = MAX2(region_num / (no_of_gc_threads * OverpartitionFactor),
2240 MinWorkUnit);
2241 } else {
2242 assert(no_of_gc_threads > 0,
2243 "The active gc workers should be greater than 0");
2244 // In a product build do something reasonable to avoid a crash.
2245 const uint MinWorkUnit = MAX2(region_num / (uint) ParallelGCThreads, 1U);
2246 WorkUnit =
2247 MAX2(region_num / (uint) (ParallelGCThreads * OverpartitionFactor),
2248 MinWorkUnit);
2249 }
2250 _collectionSetChooser->prepare_for_par_region_addition(_g1->n_regions(),
2251 WorkUnit);
2252 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser,
2253 (int) WorkUnit);
2254 _g1->workers()->run_task(&parKnownGarbageTask);
2255
2256 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2257 "sanity check");
2258 } else {
2259 KnownGarbageClosure knownGarbagecl(_collectionSetChooser);
2260 _g1->heap_region_iterate(&knownGarbagecl);
2261 }
2262
2263 _collectionSetChooser->sort_regions();
2264
2265 double end_sec = os::elapsedTime();
2266 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0;
2267 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
2268 _cur_mark_stop_world_time_ms += elapsed_time_ms;
2269 _prev_collection_pause_end_ms += elapsed_time_ms;
2270 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec, true);
2271 }
2272
2273 // Add the heap region at the head of the non-incremental collection set
2274 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) {
2275 assert(_inc_cset_build_state == Active, "Precondition");
2276 assert(!hr->is_young(), "non-incremental add of young region");
2277
2278 assert(!hr->in_collection_set(), "should not already be in the CSet");
2279 hr->set_in_collection_set(true);
2280 hr->set_next_in_collection_set(_collection_set);
2281 _collection_set = hr;
2282 _collection_set_bytes_used_before += hr->used();
2283 _g1->register_region_with_in_cset_fast_test(hr);
2284 size_t rs_length = hr->rem_set()->occupied();
2285 _recorded_rs_lengths += rs_length;
2286 _old_cset_region_length += 1;
2287 }
2288
2289 // Initialize the per-collection-set information
2290 void G1CollectorPolicy::start_incremental_cset_building() {
2291 assert(_inc_cset_build_state == Inactive, "Precondition");
2292
2293 _inc_cset_head = NULL;
2294 _inc_cset_tail = NULL;
2295 _inc_cset_bytes_used_before = 0;
2296
2297 _inc_cset_max_finger = 0;
2298 _inc_cset_recorded_rs_lengths = 0;
2299 _inc_cset_recorded_rs_lengths_diffs = 0;
2300 _inc_cset_predicted_elapsed_time_ms = 0.0;
2301 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
2302 _inc_cset_build_state = Active;
2303 }
2304
2305 void G1CollectorPolicy::finalize_incremental_cset_building() {
2306 assert(_inc_cset_build_state == Active, "Precondition");
2307 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
2308
2309 // The two "main" fields, _inc_cset_recorded_rs_lengths and
2310 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread
2311 // that adds a new region to the CSet. Further updates by the
2312 // concurrent refinement thread that samples the young RSet lengths
2313 // are accumulated in the *_diffs fields. Here we add the diffs to
2314 // the "main" fields.
2315
2316 if (_inc_cset_recorded_rs_lengths_diffs >= 0) {
2317 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs;
2318 } else {
2319 // This is defensive. The diff should in theory be always positive
2320 // as RSets can only grow between GCs. However, given that we
2321 // sample their size concurrently with other threads updating them
2322 // it's possible that we might get the wrong size back, which
2323 // could make the calculations somewhat inaccurate.
2324 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs);
2325 if (_inc_cset_recorded_rs_lengths >= diffs) {
2326 _inc_cset_recorded_rs_lengths -= diffs;
2327 } else {
2328 _inc_cset_recorded_rs_lengths = 0;
2329 }
2330 }
2331 _inc_cset_predicted_elapsed_time_ms +=
2332 _inc_cset_predicted_elapsed_time_ms_diffs;
2333
2334 _inc_cset_recorded_rs_lengths_diffs = 0;
2335 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
2336 }
2337
2338 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) {
2339 // This routine is used when:
2340 // * adding survivor regions to the incremental cset at the end of an
2341 // evacuation pause,
2342 // * adding the current allocation region to the incremental cset
2343 // when it is retired, and
2344 // * updating existing policy information for a region in the
2345 // incremental cset via young list RSet sampling.
2346 // Therefore this routine may be called at a safepoint by the
2347 // VM thread, or in-between safepoints by mutator threads (when
2348 // retiring the current allocation region) or a concurrent
2349 // refine thread (RSet sampling).
2350
2351 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true);
2352 size_t used_bytes = hr->used();
2353 _inc_cset_recorded_rs_lengths += rs_length;
2354 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms;
2355 _inc_cset_bytes_used_before += used_bytes;
2356
2357 // Cache the values we have added to the aggregated informtion
2358 // in the heap region in case we have to remove this region from
2359 // the incremental collection set, or it is updated by the
2360 // rset sampling code
2361 hr->set_recorded_rs_length(rs_length);
2362 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms);
2363 }
2364
2365 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr,
2366 size_t new_rs_length) {
2367 // Update the CSet information that is dependent on the new RS length
2368 assert(hr->is_young(), "Precondition");
2369 assert(!SafepointSynchronize::is_at_safepoint(),
2370 "should not be at a safepoint");
2371
2372 // We could have updated _inc_cset_recorded_rs_lengths and
2373 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do
2374 // that atomically, as this code is executed by a concurrent
2375 // refinement thread, potentially concurrently with a mutator thread
2376 // allocating a new region and also updating the same fields. To
2377 // avoid the atomic operations we accumulate these updates on two
2378 // separate fields (*_diffs) and we'll just add them to the "main"
2379 // fields at the start of a GC.
2380
2381 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length();
2382 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length;
2383 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff;
2384
2385 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms();
2386 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true);
2387 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms;
2388 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff;
2389
2390 hr->set_recorded_rs_length(new_rs_length);
2391 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms);
2392 }
2393
2394 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) {
2395 assert(hr->is_young(), "invariant");
2396 assert(hr->young_index_in_cset() > -1, "should have already been set");
2397 assert(_inc_cset_build_state == Active, "Precondition");
2398
2399 // We need to clear and set the cached recorded/cached collection set
2400 // information in the heap region here (before the region gets added
2401 // to the collection set). An individual heap region's cached values
2402 // are calculated, aggregated with the policy collection set info,
2403 // and cached in the heap region here (initially) and (subsequently)
2404 // by the Young List sampling code.
2405
2406 size_t rs_length = hr->rem_set()->occupied();
2407 add_to_incremental_cset_info(hr, rs_length);
2408
2409 HeapWord* hr_end = hr->end();
2410 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end);
2411
2412 assert(!hr->in_collection_set(), "invariant");
2413 hr->set_in_collection_set(true);
2414 assert( hr->next_in_collection_set() == NULL, "invariant");
2415
2416 _g1->register_region_with_in_cset_fast_test(hr);
2417 }
2418
2419 // Add the region at the RHS of the incremental cset
2420 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) {
2421 // We should only ever be appending survivors at the end of a pause
2422 assert( hr->is_survivor(), "Logic");
2423
2424 // Do the 'common' stuff
2425 add_region_to_incremental_cset_common(hr);
2426
2427 // Now add the region at the right hand side
2428 if (_inc_cset_tail == NULL) {
2429 assert(_inc_cset_head == NULL, "invariant");
2430 _inc_cset_head = hr;
2431 } else {
2432 _inc_cset_tail->set_next_in_collection_set(hr);
2433 }
2434 _inc_cset_tail = hr;
2435 }
2436
2437 // Add the region to the LHS of the incremental cset
2438 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) {
2439 // Survivors should be added to the RHS at the end of a pause
2440 assert(!hr->is_survivor(), "Logic");
2441
2442 // Do the 'common' stuff
2443 add_region_to_incremental_cset_common(hr);
2444
2445 // Add the region at the left hand side
2446 hr->set_next_in_collection_set(_inc_cset_head);
2447 if (_inc_cset_head == NULL) {
2448 assert(_inc_cset_tail == NULL, "Invariant");
2449 _inc_cset_tail = hr;
2450 }
2451 _inc_cset_head = hr;
2452 }
2453
2454 #ifndef PRODUCT
2455 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) {
2456 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be");
2457
2458 st->print_cr("\nCollection_set:");
2459 HeapRegion* csr = list_head;
2460 while (csr != NULL) {
2461 HeapRegion* next = csr->next_in_collection_set();
2462 assert(csr->in_collection_set(), "bad CS");
2463 st->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d",
2464 HR_FORMAT_PARAMS(csr),
2465 csr->prev_top_at_mark_start(), csr->next_top_at_mark_start(),
2466 csr->age_in_surv_rate_group_cond());
2467 csr = next;
2468 }
2469 }
2470 #endif // !PRODUCT
2471
2472 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str,
2473 const char* false_action_str) {
2474 CollectionSetChooser* cset_chooser = _collectionSetChooser;
2475 if (cset_chooser->is_empty()) {
2476 ergo_verbose0(ErgoMixedGCs,
2477 false_action_str,
2478 ergo_format_reason("candidate old regions not available"));
2479 return false;
2480 }
2481 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes();
2482 size_t capacity_bytes = _g1->capacity();
2483 double perc = (double) reclaimable_bytes * 100.0 / (double) capacity_bytes;
2484 double threshold = (double) G1HeapWastePercent;
2485 if (perc < threshold) {
2486 ergo_verbose4(ErgoMixedGCs,
2487 false_action_str,
2488 ergo_format_reason("reclaimable percentage lower than threshold")
2489 ergo_format_region("candidate old regions")
2490 ergo_format_byte_perc("reclaimable")
2491 ergo_format_perc("threshold"),
2492 cset_chooser->remaining_regions(),
2493 reclaimable_bytes, perc, threshold);
2494 return false;
2495 }
2496
2497 ergo_verbose4(ErgoMixedGCs,
2498 true_action_str,
2499 ergo_format_reason("candidate old regions available")
2500 ergo_format_region("candidate old regions")
2501 ergo_format_byte_perc("reclaimable")
2502 ergo_format_perc("threshold"),
2503 cset_chooser->remaining_regions(),
2504 reclaimable_bytes, perc, threshold);
2505 return true;
2506 }
2507
2508 void G1CollectorPolicy::finalize_cset(double target_pause_time_ms) {
2509 // Set this here - in case we're not doing young collections.
2510 double non_young_start_time_sec = os::elapsedTime();
2511
2512 YoungList* young_list = _g1->young_list();
2513 finalize_incremental_cset_building();
2514
2515 guarantee(target_pause_time_ms > 0.0,
2516 err_msg("target_pause_time_ms = %1.6lf should be positive",
2517 target_pause_time_ms));
2518 guarantee(_collection_set == NULL, "Precondition");
2519
2520 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
2521 double predicted_pause_time_ms = base_time_ms;
2522 double time_remaining_ms = target_pause_time_ms - base_time_ms;
2523
2524 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2525 "start choosing CSet",
2526 ergo_format_ms("predicted base time")
2527 ergo_format_ms("remaining time")
2528 ergo_format_ms("target pause time"),
2529 base_time_ms, time_remaining_ms, target_pause_time_ms);
2530
2531 HeapRegion* hr;
2532 double young_start_time_sec = os::elapsedTime();
2533
2534 _collection_set_bytes_used_before = 0;
2535 _last_gc_was_young = gcs_are_young() ? true : false;
2536
2537 if (_last_gc_was_young) {
2538 ++_young_pause_num;
2539 } else {
2540 ++_mixed_pause_num;
2541 }
2542
2543 // The young list is laid with the survivor regions from the previous
2544 // pause are appended to the RHS of the young list, i.e.
2545 // [Newly Young Regions ++ Survivors from last pause].
2546
2547 uint survivor_region_length = young_list->survivor_length();
2548 uint eden_region_length = young_list->length() - survivor_region_length;
2549 init_cset_region_lengths(eden_region_length, survivor_region_length);
2550 hr = young_list->first_survivor_region();
2551 while (hr != NULL) {
2552 assert(hr->is_survivor(), "badly formed young list");
2553 hr->set_young();
2554 hr = hr->get_next_young_region();
2555 }
2556
2557 // Clear the fields that point to the survivor list - they are all young now.
2558 young_list->clear_survivors();
2559
2560 _collection_set = _inc_cset_head;
2561 _collection_set_bytes_used_before = _inc_cset_bytes_used_before;
2562 time_remaining_ms -= _inc_cset_predicted_elapsed_time_ms;
2563 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms;
2564
2565 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2566 "add young regions to CSet",
2567 ergo_format_region("eden")
2568 ergo_format_region("survivors")
2569 ergo_format_ms("predicted young region time"),
2570 eden_region_length, survivor_region_length,
2571 _inc_cset_predicted_elapsed_time_ms);
2572
2573 // The number of recorded young regions is the incremental
2574 // collection set's current size
2575 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths);
2576
2577 double young_end_time_sec = os::elapsedTime();
2578 _recorded_young_cset_choice_time_ms =
2579 (young_end_time_sec - young_start_time_sec) * 1000.0;
2580
2581 // We are doing young collections so reset this.
2582 non_young_start_time_sec = young_end_time_sec;
2583
2584 if (!gcs_are_young()) {
2585 CollectionSetChooser* cset_chooser = _collectionSetChooser;
2586 cset_chooser->verify();
2587 const uint min_old_cset_length = cset_chooser->calc_min_old_cset_length();
2588 const uint max_old_cset_length = cset_chooser->calc_max_old_cset_length();
2589
2590 uint expensive_region_num = 0;
2591 bool check_time_remaining = adaptive_young_list_length();
2592 HeapRegion* hr = cset_chooser->peek();
2593 while (hr != NULL) {
2594 if (old_cset_region_length() >= max_old_cset_length) {
2595 // Added maximum number of old regions to the CSet.
2596 ergo_verbose2(ErgoCSetConstruction,
2597 "finish adding old regions to CSet",
2598 ergo_format_reason("old CSet region num reached max")
2599 ergo_format_region("old")
2600 ergo_format_region("max"),
2601 old_cset_region_length(), max_old_cset_length);
2602 break;
2603 }
2604
2605 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false);
2606 if (check_time_remaining) {
2607 if (predicted_time_ms > time_remaining_ms) {
2608 // Too expensive for the current CSet.
2609
2610 if (old_cset_region_length() >= min_old_cset_length) {
2611 // We have added the minimum number of old regions to the CSet,
2612 // we are done with this CSet.
2613 ergo_verbose4(ErgoCSetConstruction,
2614 "finish adding old regions to CSet",
2615 ergo_format_reason("predicted time is too high")
2616 ergo_format_ms("predicted time")
2617 ergo_format_ms("remaining time")
2618 ergo_format_region("old")
2619 ergo_format_region("min"),
2620 predicted_time_ms, time_remaining_ms,
2621 old_cset_region_length(), min_old_cset_length);
2622 break;
2623 }
2624
2625 // We'll add it anyway given that we haven't reached the
2626 // minimum number of old regions.
2627 expensive_region_num += 1;
2628 }
2629 } else {
2630 if (old_cset_region_length() >= min_old_cset_length) {
2631 // In the non-auto-tuning case, we'll finish adding regions
2632 // to the CSet if we reach the minimum.
2633 ergo_verbose2(ErgoCSetConstruction,
2634 "finish adding old regions to CSet",
2635 ergo_format_reason("old CSet region num reached min")
2636 ergo_format_region("old")
2637 ergo_format_region("min"),
2638 old_cset_region_length(), min_old_cset_length);
2639 break;
2640 }
2641 }
2642
2643 // We will add this region to the CSet.
2644 time_remaining_ms -= predicted_time_ms;
2645 predicted_pause_time_ms += predicted_time_ms;
2646 cset_chooser->remove_and_move_to_next(hr);
2647 _g1->old_set_remove(hr);
2648 add_old_region_to_cset(hr);
2649
2650 hr = cset_chooser->peek();
2651 }
2652 if (hr == NULL) {
2653 ergo_verbose0(ErgoCSetConstruction,
2654 "finish adding old regions to CSet",
2655 ergo_format_reason("candidate old regions not available"));
2656 }
2657
2658 if (expensive_region_num > 0) {
2659 // We print the information once here at the end, predicated on
2660 // whether we added any apparently expensive regions or not, to
2661 // avoid generating output per region.
2662 ergo_verbose4(ErgoCSetConstruction,
2663 "added expensive regions to CSet",
2664 ergo_format_reason("old CSet region num not reached min")
2665 ergo_format_region("old")
2666 ergo_format_region("expensive")
2667 ergo_format_region("min")
2668 ergo_format_ms("remaining time"),
2669 old_cset_region_length(),
2670 expensive_region_num,
2671 min_old_cset_length,
2672 time_remaining_ms);
2673 }
2674
2675 cset_chooser->verify();
2676 }
2677
2678 stop_incremental_cset_building();
2679
2680 count_CS_bytes_used();
2681
2682 ergo_verbose5(ErgoCSetConstruction,
2683 "finish choosing CSet",
2684 ergo_format_region("eden")
2685 ergo_format_region("survivors")
2686 ergo_format_region("old")
2687 ergo_format_ms("predicted pause time")
2688 ergo_format_ms("target pause time"),
2689 eden_region_length, survivor_region_length,
2690 old_cset_region_length(),
2691 predicted_pause_time_ms, target_pause_time_ms);
2692
2693 double non_young_end_time_sec = os::elapsedTime();
2694 _recorded_non_young_cset_choice_time_ms =
2695 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0;
2696 }