1 /*
2 * Copyright (c) 2001, 2011, 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/heapRegionRemSet.hpp"
33 #include "gc_implementation/shared/gcPolicyCounters.hpp"
34 #include "runtime/arguments.hpp"
35 #include "runtime/java.hpp"
36 #include "runtime/mutexLocker.hpp"
37 #include "utilities/debug.hpp"
38
39 // Different defaults for different number of GC threads
40 // They were chosen by running GCOld and SPECjbb on debris with different
41 // numbers of GC threads and choosing them based on the results
42
43 // all the same
44 static double rs_length_diff_defaults[] = {
45 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
46 };
47
48 static double cost_per_card_ms_defaults[] = {
49 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015
50 };
51
52 // all the same
53 static double young_cards_per_entry_ratio_defaults[] = {
54 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0
55 };
56
57 static double cost_per_entry_ms_defaults[] = {
58 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005
59 };
60
61 static double cost_per_byte_ms_defaults[] = {
62 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009
63 };
64
65 // these should be pretty consistent
66 static double constant_other_time_ms_defaults[] = {
67 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0
68 };
69
70
71 static double young_other_cost_per_region_ms_defaults[] = {
72 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1
73 };
74
75 static double non_young_other_cost_per_region_ms_defaults[] = {
76 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30
77 };
78
79 // Help class for avoiding interleaved logging
80 class LineBuffer: public StackObj {
81
82 private:
83 static const int BUFFER_LEN = 1024;
84 static const int INDENT_CHARS = 3;
85 char _buffer[BUFFER_LEN];
86 int _indent_level;
87 int _cur;
88
89 void vappend(const char* format, va_list ap) {
90 int res = vsnprintf(&_buffer[_cur], BUFFER_LEN - _cur, format, ap);
91 if (res != -1) {
92 _cur += res;
93 } else {
94 DEBUG_ONLY(warning("buffer too small in LineBuffer");)
95 _buffer[BUFFER_LEN -1] = 0;
96 _cur = BUFFER_LEN; // vsnprintf above should not add to _buffer if we are called again
97 }
98 }
99
100 public:
101 explicit LineBuffer(int indent_level): _indent_level(indent_level), _cur(0) {
102 for (; (_cur < BUFFER_LEN && _cur < (_indent_level * INDENT_CHARS)); _cur++) {
103 _buffer[_cur] = ' ';
104 }
105 }
106
107 #ifndef PRODUCT
108 ~LineBuffer() {
109 assert(_cur == _indent_level * INDENT_CHARS, "pending data in buffer - append_and_print_cr() not called?");
110 }
111 #endif
112
113 void append(const char* format, ...) {
114 va_list ap;
115 va_start(ap, format);
116 vappend(format, ap);
117 va_end(ap);
118 }
119
120 void append_and_print_cr(const char* format, ...) {
121 va_list ap;
122 va_start(ap, format);
123 vappend(format, ap);
124 va_end(ap);
125 gclog_or_tty->print_cr("%s", _buffer);
126 _cur = _indent_level * INDENT_CHARS;
127 }
128 };
129
130 G1CollectorPolicy::G1CollectorPolicy() :
131 _parallel_gc_threads(G1CollectedHeap::use_parallel_gc_threads()
132 ? ParallelGCThreads : 1),
133
134 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
135 _all_pause_times_ms(new NumberSeq()),
136 _stop_world_start(0.0),
137 _all_stop_world_times_ms(new NumberSeq()),
138 _all_yield_times_ms(new NumberSeq()),
139
140 _summary(new Summary()),
141
142 _cur_clear_ct_time_ms(0.0),
143 _mark_closure_time_ms(0.0),
144
145 _cur_ref_proc_time_ms(0.0),
146 _cur_ref_enq_time_ms(0.0),
147
148 #ifndef PRODUCT
149 _min_clear_cc_time_ms(-1.0),
150 _max_clear_cc_time_ms(-1.0),
151 _cur_clear_cc_time_ms(0.0),
152 _cum_clear_cc_time_ms(0.0),
153 _num_cc_clears(0L),
154 #endif
155
156 _aux_num(10),
157 _all_aux_times_ms(new NumberSeq[_aux_num]),
158 _cur_aux_start_times_ms(new double[_aux_num]),
159 _cur_aux_times_ms(new double[_aux_num]),
160 _cur_aux_times_set(new bool[_aux_num]),
161
162 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
163 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
164
165 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
166 _prev_collection_pause_end_ms(0.0),
167 _pending_card_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
168 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
169 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
170 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
171 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
172 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
173 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
174 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
175 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)),
176 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
177 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
178 _non_young_other_cost_per_region_ms_seq(
179 new TruncatedSeq(TruncatedSeqLength)),
180
181 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)),
182 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)),
183
184 _pause_time_target_ms((double) MaxGCPauseMillis),
185
186 _gcs_are_young(true),
187 _young_pause_num(0),
188 _mixed_pause_num(0),
189
190 _during_marking(false),
191 _in_marking_window(false),
192 _in_marking_window_im(false),
193
194 _known_garbage_ratio(0.0),
195 _known_garbage_bytes(0),
196
197 _young_gc_eff_seq(new TruncatedSeq(TruncatedSeqLength)),
198
199 _recent_prev_end_times_for_all_gcs_sec(
200 new TruncatedSeq(NumPrevPausesForHeuristics)),
201
202 _recent_avg_pause_time_ratio(0.0),
203
204 _all_full_gc_times_ms(new NumberSeq()),
205
206 _initiate_conc_mark_if_possible(false),
207 _during_initial_mark_pause(false),
208 _should_revert_to_young_gcs(false),
209 _last_young_gc(false),
210 _last_gc_was_young(false),
211
212 _eden_bytes_before_gc(0),
213 _survivor_bytes_before_gc(0),
214 _capacity_before_gc(0),
215
216 _prev_collection_pause_used_at_end_bytes(0),
217
218 _eden_cset_region_length(0),
219 _survivor_cset_region_length(0),
220 _old_cset_region_length(0),
221
222 _collection_set(NULL),
223 _collection_set_bytes_used_before(0),
224
225 // Incremental CSet attributes
226 _inc_cset_build_state(Inactive),
227 _inc_cset_head(NULL),
228 _inc_cset_tail(NULL),
229 _inc_cset_bytes_used_before(0),
230 _inc_cset_max_finger(NULL),
231 _inc_cset_recorded_rs_lengths(0),
232 _inc_cset_predicted_elapsed_time_ms(0.0),
233
234 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
235 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
236 #endif // _MSC_VER
237
238 _short_lived_surv_rate_group(new SurvRateGroup(this, "Short Lived",
239 G1YoungSurvRateNumRegionsSummary)),
240 _survivor_surv_rate_group(new SurvRateGroup(this, "Survivor",
241 G1YoungSurvRateNumRegionsSummary)),
242 // add here any more surv rate groups
243 _recorded_survivor_regions(0),
244 _recorded_survivor_head(NULL),
245 _recorded_survivor_tail(NULL),
246 _survivors_age_table(true),
247
248 _gc_overhead_perc(0.0) {
249
250 // Set up the region size and associated fields. Given that the
251 // policy is created before the heap, we have to set this up here,
252 // so it's done as soon as possible.
253 HeapRegion::setup_heap_region_size(Arguments::min_heap_size());
254 HeapRegionRemSet::setup_remset_size();
255
256 G1ErgoVerbose::initialize();
257 if (PrintAdaptiveSizePolicy) {
258 // Currently, we only use a single switch for all the heuristics.
259 G1ErgoVerbose::set_enabled(true);
260 // Given that we don't currently have a verboseness level
261 // parameter, we'll hardcode this to high. This can be easily
262 // changed in the future.
263 G1ErgoVerbose::set_level(ErgoHigh);
264 } else {
265 G1ErgoVerbose::set_enabled(false);
266 }
267
268 // Verify PLAB sizes
269 const size_t region_size = HeapRegion::GrainWords;
270 if (YoungPLABSize > region_size || OldPLABSize > region_size) {
271 char buffer[128];
272 jio_snprintf(buffer, sizeof(buffer), "%sPLABSize should be at most "SIZE_FORMAT,
273 OldPLABSize > region_size ? "Old" : "Young", region_size);
274 vm_exit_during_initialization(buffer);
275 }
276
277 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime());
278 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0;
279
280 _par_last_gc_worker_start_times_ms = new double[_parallel_gc_threads];
281 _par_last_ext_root_scan_times_ms = new double[_parallel_gc_threads];
282 _par_last_mark_stack_scan_times_ms = new double[_parallel_gc_threads];
283
284 _par_last_update_rs_times_ms = new double[_parallel_gc_threads];
285 _par_last_update_rs_processed_buffers = new double[_parallel_gc_threads];
286
287 _par_last_scan_rs_times_ms = new double[_parallel_gc_threads];
288
289 _par_last_obj_copy_times_ms = new double[_parallel_gc_threads];
290
291 _par_last_termination_times_ms = new double[_parallel_gc_threads];
292 _par_last_termination_attempts = new double[_parallel_gc_threads];
293 _par_last_gc_worker_end_times_ms = new double[_parallel_gc_threads];
294 _par_last_gc_worker_times_ms = new double[_parallel_gc_threads];
295 _par_last_gc_worker_other_times_ms = new double[_parallel_gc_threads];
296
297 // start conservatively
298 _expensive_region_limit_ms = 0.5 * (double) MaxGCPauseMillis;
299
300 int index;
301 if (ParallelGCThreads == 0)
302 index = 0;
303 else if (ParallelGCThreads > 8)
304 index = 7;
305 else
306 index = ParallelGCThreads - 1;
307
308 _pending_card_diff_seq->add(0.0);
309 _rs_length_diff_seq->add(rs_length_diff_defaults[index]);
310 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]);
311 _young_cards_per_entry_ratio_seq->add(
312 young_cards_per_entry_ratio_defaults[index]);
313 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]);
314 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]);
315 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]);
316 _young_other_cost_per_region_ms_seq->add(
317 young_other_cost_per_region_ms_defaults[index]);
318 _non_young_other_cost_per_region_ms_seq->add(
319 non_young_other_cost_per_region_ms_defaults[index]);
320
321 // Below, we might need to calculate the pause time target based on
322 // the pause interval. When we do so we are going to give G1 maximum
323 // flexibility and allow it to do pauses when it needs to. So, we'll
324 // arrange that the pause interval to be pause time target + 1 to
325 // ensure that a) the pause time target is maximized with respect to
326 // the pause interval and b) we maintain the invariant that pause
327 // time target < pause interval. If the user does not want this
328 // maximum flexibility, they will have to set the pause interval
329 // explicitly.
330
331 // First make sure that, if either parameter is set, its value is
332 // reasonable.
333 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
334 if (MaxGCPauseMillis < 1) {
335 vm_exit_during_initialization("MaxGCPauseMillis should be "
336 "greater than 0");
337 }
338 }
339 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
340 if (GCPauseIntervalMillis < 1) {
341 vm_exit_during_initialization("GCPauseIntervalMillis should be "
342 "greater than 0");
343 }
344 }
345
346 // Then, if the pause time target parameter was not set, set it to
347 // the default value.
348 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
349 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
350 // The default pause time target in G1 is 200ms
351 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200);
352 } else {
353 // We do not allow the pause interval to be set without the
354 // pause time target
355 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set "
356 "without setting MaxGCPauseMillis");
357 }
358 }
359
360 // Then, if the interval parameter was not set, set it according to
361 // the pause time target (this will also deal with the case when the
362 // pause time target is the default value).
363 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
364 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1);
365 }
366
367 // Finally, make sure that the two parameters are consistent.
368 if (MaxGCPauseMillis >= GCPauseIntervalMillis) {
369 char buffer[256];
370 jio_snprintf(buffer, 256,
371 "MaxGCPauseMillis (%u) should be less than "
372 "GCPauseIntervalMillis (%u)",
373 MaxGCPauseMillis, GCPauseIntervalMillis);
374 vm_exit_during_initialization(buffer);
375 }
376
377 double max_gc_time = (double) MaxGCPauseMillis / 1000.0;
378 double time_slice = (double) GCPauseIntervalMillis / 1000.0;
379 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time);
380 _sigma = (double) G1ConfidencePercent / 100.0;
381
382 // start conservatively (around 50ms is about right)
383 _concurrent_mark_remark_times_ms->add(0.05);
384 _concurrent_mark_cleanup_times_ms->add(0.20);
385 _tenuring_threshold = MaxTenuringThreshold;
386 // _max_survivor_regions will be calculated by
387 // update_young_list_target_length() during initialization.
388 _max_survivor_regions = 0;
389
390 assert(GCTimeRatio > 0,
391 "we should have set it to a default value set_g1_gc_flags() "
392 "if a user set it to 0");
393 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio));
394
395 uintx reserve_perc = G1ReservePercent;
396 // Put an artificial ceiling on this so that it's not set to a silly value.
397 if (reserve_perc > 50) {
398 reserve_perc = 50;
399 warning("G1ReservePercent is set to a value that is too large, "
400 "it's been updated to %u", reserve_perc);
401 }
402 _reserve_factor = (double) reserve_perc / 100.0;
403 // This will be set when the heap is expanded
404 // for the first time during initialization.
405 _reserve_regions = 0;
406
407 initialize_all();
408 _collectionSetChooser = new CollectionSetChooser();
409 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags
410 }
411
412 void G1CollectorPolicy::initialize_flags() {
413 set_min_alignment(HeapRegion::GrainBytes);
414 set_max_alignment(GenRemSet::max_alignment_constraint(rem_set_name()));
415 if (SurvivorRatio < 1) {
416 vm_exit_during_initialization("Invalid survivor ratio specified");
417 }
418 CollectorPolicy::initialize_flags();
419 }
420
421 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true) {
422 assert(G1DefaultMinNewGenPercent <= G1DefaultMaxNewGenPercent, "Min larger than max");
423 assert(G1DefaultMinNewGenPercent > 0 && G1DefaultMinNewGenPercent < 100, "Min out of bounds");
424 assert(G1DefaultMaxNewGenPercent > 0 && G1DefaultMaxNewGenPercent < 100, "Max out of bounds");
425
426 if (FLAG_IS_CMDLINE(NewRatio)) {
427 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) {
428 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio");
429 } else {
430 _sizer_kind = SizerNewRatio;
431 _adaptive_size = false;
432 return;
433 }
434 }
435
436 if (FLAG_IS_CMDLINE(NewSize)) {
437 _min_desired_young_length = MAX2((size_t) 1, NewSize / HeapRegion::GrainBytes);
438 if (FLAG_IS_CMDLINE(MaxNewSize)) {
439 _max_desired_young_length = MAX2((size_t) 1, MaxNewSize / HeapRegion::GrainBytes);
440 _sizer_kind = SizerMaxAndNewSize;
441 _adaptive_size = _min_desired_young_length == _max_desired_young_length;
442 } else {
443 _sizer_kind = SizerNewSizeOnly;
444 }
445 } else if (FLAG_IS_CMDLINE(MaxNewSize)) {
446 _max_desired_young_length = MAX2((size_t) 1, MaxNewSize / HeapRegion::GrainBytes);
447 _sizer_kind = SizerMaxNewSizeOnly;
448 }
449 }
450
451 size_t G1YoungGenSizer::calculate_default_min_size(size_t new_number_of_heap_regions) {
452 size_t default_value = (new_number_of_heap_regions * G1DefaultMinNewGenPercent) / 100;
453 return MAX2((size_t)1, default_value);
454 }
455
456 size_t G1YoungGenSizer::calculate_default_max_size(size_t new_number_of_heap_regions) {
457 size_t default_value = (new_number_of_heap_regions * G1DefaultMaxNewGenPercent) / 100;
458 return MAX2((size_t)1, default_value);
459 }
460
461 void G1YoungGenSizer::heap_size_changed(size_t new_number_of_heap_regions) {
462 assert(new_number_of_heap_regions > 0, "Heap must be initialized");
463
464 switch (_sizer_kind) {
465 case SizerDefaults:
466 _min_desired_young_length = calculate_default_min_size(new_number_of_heap_regions);
467 _max_desired_young_length = calculate_default_max_size(new_number_of_heap_regions);
468 break;
469 case SizerNewSizeOnly:
470 _max_desired_young_length = calculate_default_max_size(new_number_of_heap_regions);
471 _max_desired_young_length = MAX2(_min_desired_young_length, _max_desired_young_length);
472 break;
473 case SizerMaxNewSizeOnly:
474 _min_desired_young_length = calculate_default_min_size(new_number_of_heap_regions);
475 _min_desired_young_length = MIN2(_min_desired_young_length, _max_desired_young_length);
476 break;
477 case SizerMaxAndNewSize:
478 // Do nothing. Values set on the command line, don't update them at runtime.
479 break;
480 case SizerNewRatio:
481 _min_desired_young_length = new_number_of_heap_regions / (NewRatio + 1);
482 _max_desired_young_length = _min_desired_young_length;
483 break;
484 default:
485 ShouldNotReachHere();
486 }
487
488 assert(_min_desired_young_length <= _max_desired_young_length, "Invalid min/max young gen size values");
489 }
490
491 void G1CollectorPolicy::init() {
492 // Set aside an initial future to_space.
493 _g1 = G1CollectedHeap::heap();
494
495 assert(Heap_lock->owned_by_self(), "Locking discipline.");
496
497 initialize_gc_policy_counters();
498
499 if (adaptive_young_list_length()) {
500 _young_list_fixed_length = 0;
501 } else {
502 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length();
503 }
504 _free_regions_at_end_of_collection = _g1->free_regions();
505 update_young_list_target_length();
506 _prev_eden_capacity = _young_list_target_length * HeapRegion::GrainBytes;
507
508 // We may immediately start allocating regions and placing them on the
509 // collection set list. Initialize the per-collection set info
510 start_incremental_cset_building();
511 }
512
513 // Create the jstat counters for the policy.
514 void G1CollectorPolicy::initialize_gc_policy_counters() {
515 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3);
516 }
517
518 bool G1CollectorPolicy::predict_will_fit(size_t young_length,
519 double base_time_ms,
520 size_t base_free_regions,
521 double target_pause_time_ms) {
522 if (young_length >= base_free_regions) {
523 // end condition 1: not enough space for the young regions
524 return false;
525 }
526
527 double accum_surv_rate = accum_yg_surv_rate_pred((int)(young_length - 1));
528 size_t bytes_to_copy =
529 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
530 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
531 double young_other_time_ms = predict_young_other_time_ms(young_length);
532 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms;
533 if (pause_time_ms > target_pause_time_ms) {
534 // end condition 2: prediction is over the target pause time
535 return false;
536 }
537
538 size_t free_bytes =
539 (base_free_regions - young_length) * HeapRegion::GrainBytes;
540 if ((2.0 * sigma()) * (double) bytes_to_copy > (double) free_bytes) {
541 // end condition 3: out-of-space (conservatively!)
542 return false;
543 }
544
545 // success!
546 return true;
547 }
548
549 void G1CollectorPolicy::record_new_heap_size(size_t new_number_of_regions) {
550 // re-calculate the necessary reserve
551 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor;
552 // We use ceiling so that if reserve_regions_d is > 0.0 (but
553 // smaller than 1.0) we'll get 1.
554 _reserve_regions = (size_t) ceil(reserve_regions_d);
555
556 _young_gen_sizer->heap_size_changed(new_number_of_regions);
557 }
558
559 size_t G1CollectorPolicy::calculate_young_list_desired_min_length(
560 size_t base_min_length) {
561 size_t desired_min_length = 0;
562 if (adaptive_young_list_length()) {
563 if (_alloc_rate_ms_seq->num() > 3) {
564 double now_sec = os::elapsedTime();
565 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
566 double alloc_rate_ms = predict_alloc_rate_ms();
567 desired_min_length = (size_t) ceil(alloc_rate_ms * when_ms);
568 } else {
569 // otherwise we don't have enough info to make the prediction
570 }
571 }
572 desired_min_length += base_min_length;
573 // make sure we don't go below any user-defined minimum bound
574 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length);
575 }
576
577 size_t G1CollectorPolicy::calculate_young_list_desired_max_length() {
578 // Here, we might want to also take into account any additional
579 // constraints (i.e., user-defined minimum bound). Currently, we
580 // effectively don't set this bound.
581 return _young_gen_sizer->max_desired_young_length();
582 }
583
584 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) {
585 if (rs_lengths == (size_t) -1) {
586 // if it's set to the default value (-1), we should predict it;
587 // otherwise, use the given value.
588 rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
589 }
590
591 // Calculate the absolute and desired min bounds.
592
593 // This is how many young regions we already have (currently: the survivors).
594 size_t base_min_length = recorded_survivor_regions();
595 // This is the absolute minimum young length, which ensures that we
596 // can allocate one eden region in the worst-case.
597 size_t absolute_min_length = base_min_length + 1;
598 size_t desired_min_length =
599 calculate_young_list_desired_min_length(base_min_length);
600 if (desired_min_length < absolute_min_length) {
601 desired_min_length = absolute_min_length;
602 }
603
604 // Calculate the absolute and desired max bounds.
605
606 // We will try our best not to "eat" into the reserve.
607 size_t absolute_max_length = 0;
608 if (_free_regions_at_end_of_collection > _reserve_regions) {
609 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions;
610 }
611 size_t desired_max_length = calculate_young_list_desired_max_length();
612 if (desired_max_length > absolute_max_length) {
613 desired_max_length = absolute_max_length;
614 }
615
616 size_t young_list_target_length = 0;
617 if (adaptive_young_list_length()) {
618 if (gcs_are_young()) {
619 young_list_target_length =
620 calculate_young_list_target_length(rs_lengths,
621 base_min_length,
622 desired_min_length,
623 desired_max_length);
624 _rs_lengths_prediction = rs_lengths;
625 } else {
626 // Don't calculate anything and let the code below bound it to
627 // the desired_min_length, i.e., do the next GC as soon as
628 // possible to maximize how many old regions we can add to it.
629 }
630 } else {
631 if (gcs_are_young()) {
632 young_list_target_length = _young_list_fixed_length;
633 } else {
634 // A bit arbitrary: during mixed GCs we allocate half
635 // the young regions to try to add old regions to the CSet.
636 young_list_target_length = _young_list_fixed_length / 2;
637 // We choose to accept that we might go under the desired min
638 // length given that we intentionally ask for a smaller young gen.
639 desired_min_length = absolute_min_length;
640 }
641 }
642
643 // Make sure we don't go over the desired max length, nor under the
644 // desired min length. In case they clash, desired_min_length wins
645 // which is why that test is second.
646 if (young_list_target_length > desired_max_length) {
647 young_list_target_length = desired_max_length;
648 }
649 if (young_list_target_length < desired_min_length) {
650 young_list_target_length = desired_min_length;
651 }
652
653 assert(young_list_target_length > recorded_survivor_regions(),
654 "we should be able to allocate at least one eden region");
655 assert(young_list_target_length >= absolute_min_length, "post-condition");
656 _young_list_target_length = young_list_target_length;
657
658 update_max_gc_locker_expansion();
659 }
660
661 size_t
662 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths,
663 size_t base_min_length,
664 size_t desired_min_length,
665 size_t desired_max_length) {
666 assert(adaptive_young_list_length(), "pre-condition");
667 assert(gcs_are_young(), "only call this for young GCs");
668
669 // In case some edge-condition makes the desired max length too small...
670 if (desired_max_length <= desired_min_length) {
671 return desired_min_length;
672 }
673
674 // We'll adjust min_young_length and max_young_length not to include
675 // the already allocated young regions (i.e., so they reflect the
676 // min and max eden regions we'll allocate). The base_min_length
677 // will be reflected in the predictions by the
678 // survivor_regions_evac_time prediction.
679 assert(desired_min_length > base_min_length, "invariant");
680 size_t min_young_length = desired_min_length - base_min_length;
681 assert(desired_max_length > base_min_length, "invariant");
682 size_t max_young_length = desired_max_length - base_min_length;
683
684 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
685 double survivor_regions_evac_time = predict_survivor_regions_evac_time();
686 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
687 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
688 size_t scanned_cards = predict_young_card_num(adj_rs_lengths);
689 double base_time_ms =
690 predict_base_elapsed_time_ms(pending_cards, scanned_cards) +
691 survivor_regions_evac_time;
692 size_t available_free_regions = _free_regions_at_end_of_collection;
693 size_t base_free_regions = 0;
694 if (available_free_regions > _reserve_regions) {
695 base_free_regions = available_free_regions - _reserve_regions;
696 }
697
698 // Here, we will make sure that the shortest young length that
699 // makes sense fits within the target pause time.
700
701 if (predict_will_fit(min_young_length, base_time_ms,
702 base_free_regions, target_pause_time_ms)) {
703 // The shortest young length will fit into the target pause time;
704 // we'll now check whether the absolute maximum number of young
705 // regions will fit in the target pause time. If not, we'll do
706 // a binary search between min_young_length and max_young_length.
707 if (predict_will_fit(max_young_length, base_time_ms,
708 base_free_regions, target_pause_time_ms)) {
709 // The maximum young length will fit into the target pause time.
710 // We are done so set min young length to the maximum length (as
711 // the result is assumed to be returned in min_young_length).
712 min_young_length = max_young_length;
713 } else {
714 // The maximum possible number of young regions will not fit within
715 // the target pause time so we'll search for the optimal
716 // length. The loop invariants are:
717 //
718 // min_young_length < max_young_length
719 // min_young_length is known to fit into the target pause time
720 // max_young_length is known not to fit into the target pause time
721 //
722 // Going into the loop we know the above hold as we've just
723 // checked them. Every time around the loop we check whether
724 // the middle value between min_young_length and
725 // max_young_length fits into the target pause time. If it
726 // does, it becomes the new min. If it doesn't, it becomes
727 // the new max. This way we maintain the loop invariants.
728
729 assert(min_young_length < max_young_length, "invariant");
730 size_t diff = (max_young_length - min_young_length) / 2;
731 while (diff > 0) {
732 size_t young_length = min_young_length + diff;
733 if (predict_will_fit(young_length, base_time_ms,
734 base_free_regions, target_pause_time_ms)) {
735 min_young_length = young_length;
736 } else {
737 max_young_length = young_length;
738 }
739 assert(min_young_length < max_young_length, "invariant");
740 diff = (max_young_length - min_young_length) / 2;
741 }
742 // The results is min_young_length which, according to the
743 // loop invariants, should fit within the target pause time.
744
745 // These are the post-conditions of the binary search above:
746 assert(min_young_length < max_young_length,
747 "otherwise we should have discovered that max_young_length "
748 "fits into the pause target and not done the binary search");
749 assert(predict_will_fit(min_young_length, base_time_ms,
750 base_free_regions, target_pause_time_ms),
751 "min_young_length, the result of the binary search, should "
752 "fit into the pause target");
753 assert(!predict_will_fit(min_young_length + 1, base_time_ms,
754 base_free_regions, target_pause_time_ms),
755 "min_young_length, the result of the binary search, should be "
756 "optimal, so no larger length should fit into the pause target");
757 }
758 } else {
759 // Even the minimum length doesn't fit into the pause time
760 // target, return it as the result nevertheless.
761 }
762 return base_min_length + min_young_length;
763 }
764
765 double G1CollectorPolicy::predict_survivor_regions_evac_time() {
766 double survivor_regions_evac_time = 0.0;
767 for (HeapRegion * r = _recorded_survivor_head;
768 r != NULL && r != _recorded_survivor_tail->get_next_young_region();
769 r = r->get_next_young_region()) {
770 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, true);
771 }
772 return survivor_regions_evac_time;
773 }
774
775 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() {
776 guarantee( adaptive_young_list_length(), "should not call this otherwise" );
777
778 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths();
779 if (rs_lengths > _rs_lengths_prediction) {
780 // add 10% to avoid having to recalculate often
781 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000;
782 update_young_list_target_length(rs_lengths_prediction);
783 }
784 }
785
786
787
788 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size,
789 bool is_tlab,
790 bool* gc_overhead_limit_was_exceeded) {
791 guarantee(false, "Not using this policy feature yet.");
792 return NULL;
793 }
794
795 // This method controls how a collector handles one or more
796 // of its generations being fully allocated.
797 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size,
798 bool is_tlab) {
799 guarantee(false, "Not using this policy feature yet.");
800 return NULL;
801 }
802
803
804 #ifndef PRODUCT
805 bool G1CollectorPolicy::verify_young_ages() {
806 HeapRegion* head = _g1->young_list()->first_region();
807 return
808 verify_young_ages(head, _short_lived_surv_rate_group);
809 // also call verify_young_ages on any additional surv rate groups
810 }
811
812 bool
813 G1CollectorPolicy::verify_young_ages(HeapRegion* head,
814 SurvRateGroup *surv_rate_group) {
815 guarantee( surv_rate_group != NULL, "pre-condition" );
816
817 const char* name = surv_rate_group->name();
818 bool ret = true;
819 int prev_age = -1;
820
821 for (HeapRegion* curr = head;
822 curr != NULL;
823 curr = curr->get_next_young_region()) {
824 SurvRateGroup* group = curr->surv_rate_group();
825 if (group == NULL && !curr->is_survivor()) {
826 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name);
827 ret = false;
828 }
829
830 if (surv_rate_group == group) {
831 int age = curr->age_in_surv_rate_group();
832
833 if (age < 0) {
834 gclog_or_tty->print_cr("## %s: encountered negative age", name);
835 ret = false;
836 }
837
838 if (age <= prev_age) {
839 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing "
840 "(%d, %d)", name, age, prev_age);
841 ret = false;
842 }
843 prev_age = age;
844 }
845 }
846
847 return ret;
848 }
849 #endif // PRODUCT
850
851 void G1CollectorPolicy::record_full_collection_start() {
852 _cur_collection_start_sec = os::elapsedTime();
853 // Release the future to-space so that it is available for compaction into.
854 _g1->set_full_collection();
855 }
856
857 void G1CollectorPolicy::record_full_collection_end() {
858 // Consider this like a collection pause for the purposes of allocation
859 // since last pause.
860 double end_sec = os::elapsedTime();
861 double full_gc_time_sec = end_sec - _cur_collection_start_sec;
862 double full_gc_time_ms = full_gc_time_sec * 1000.0;
863
864 _all_full_gc_times_ms->add(full_gc_time_ms);
865
866 update_recent_gc_times(end_sec, full_gc_time_ms);
867
868 _g1->clear_full_collection();
869
870 // "Nuke" the heuristics that control the young/mixed GC
871 // transitions and make sure we start with young GCs after the Full GC.
872 set_gcs_are_young(true);
873 _last_young_gc = false;
874 _should_revert_to_young_gcs = false;
875 clear_initiate_conc_mark_if_possible();
876 clear_during_initial_mark_pause();
877 _known_garbage_bytes = 0;
878 _known_garbage_ratio = 0.0;
879 _in_marking_window = false;
880 _in_marking_window_im = false;
881
882 _short_lived_surv_rate_group->start_adding_regions();
883 // also call this on any additional surv rate groups
884
885 record_survivor_regions(0, NULL, NULL);
886
887 _free_regions_at_end_of_collection = _g1->free_regions();
888 // Reset survivors SurvRateGroup.
889 _survivor_surv_rate_group->reset();
890 update_young_list_target_length();
891 _collectionSetChooser->updateAfterFullCollection();
892 }
893
894 void G1CollectorPolicy::record_stop_world_start() {
895 _stop_world_start = os::elapsedTime();
896 }
897
898 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec,
899 size_t start_used) {
900 if (PrintGCDetails) {
901 gclog_or_tty->stamp(PrintGCTimeStamps);
902 gclog_or_tty->print("[GC pause");
903 gclog_or_tty->print(" (%s)", gcs_are_young() ? "young" : "mixed");
904 }
905
906 // We only need to do this here as the policy will only be applied
907 // to the GC we're about to start. so, no point is calculating this
908 // every time we calculate / recalculate the target young length.
909 update_survivors_policy();
910
911 assert(_g1->used() == _g1->recalculate_used(),
912 err_msg("sanity, used: "SIZE_FORMAT" recalculate_used: "SIZE_FORMAT,
913 _g1->used(), _g1->recalculate_used()));
914
915 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0;
916 _all_stop_world_times_ms->add(s_w_t_ms);
917 _stop_world_start = 0.0;
918
919 _cur_collection_start_sec = start_time_sec;
920 _cur_collection_pause_used_at_start_bytes = start_used;
921 _cur_collection_pause_used_regions_at_start = _g1->used_regions();
922 _pending_cards = _g1->pending_card_num();
923 _max_pending_cards = _g1->max_pending_card_num();
924
925 _bytes_in_collection_set_before_gc = 0;
926 _bytes_copied_during_gc = 0;
927
928 YoungList* young_list = _g1->young_list();
929 _eden_bytes_before_gc = young_list->eden_used_bytes();
930 _survivor_bytes_before_gc = young_list->survivor_used_bytes();
931 _capacity_before_gc = _g1->capacity();
932
933 #ifdef DEBUG
934 // initialise these to something well known so that we can spot
935 // if they are not set properly
936
937 for (int i = 0; i < _parallel_gc_threads; ++i) {
938 _par_last_gc_worker_start_times_ms[i] = -1234.0;
939 _par_last_ext_root_scan_times_ms[i] = -1234.0;
940 _par_last_mark_stack_scan_times_ms[i] = -1234.0;
941 _par_last_update_rs_times_ms[i] = -1234.0;
942 _par_last_update_rs_processed_buffers[i] = -1234.0;
943 _par_last_scan_rs_times_ms[i] = -1234.0;
944 _par_last_obj_copy_times_ms[i] = -1234.0;
945 _par_last_termination_times_ms[i] = -1234.0;
946 _par_last_termination_attempts[i] = -1234.0;
947 _par_last_gc_worker_end_times_ms[i] = -1234.0;
948 _par_last_gc_worker_times_ms[i] = -1234.0;
949 _par_last_gc_worker_other_times_ms[i] = -1234.0;
950 }
951 #endif
952
953 for (int i = 0; i < _aux_num; ++i) {
954 _cur_aux_times_ms[i] = 0.0;
955 _cur_aux_times_set[i] = false;
956 }
957
958 // This is initialized to zero here and is set during
959 // the evacuation pause if marking is in progress.
960 _cur_satb_drain_time_ms = 0.0;
961
962 _last_gc_was_young = false;
963
964 // do that for any other surv rate groups
965 _short_lived_surv_rate_group->stop_adding_regions();
966 _survivors_age_table.clear();
967
968 assert( verify_young_ages(), "region age verification" );
969 }
970
971 void G1CollectorPolicy::record_concurrent_mark_init_end(double
972 mark_init_elapsed_time_ms) {
973 _during_marking = true;
974 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now");
975 clear_during_initial_mark_pause();
976 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms;
977 }
978
979 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
980 _mark_remark_start_sec = os::elapsedTime();
981 _during_marking = false;
982 }
983
984 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
985 double end_time_sec = os::elapsedTime();
986 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
987 _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
988 _cur_mark_stop_world_time_ms += elapsed_time_ms;
989 _prev_collection_pause_end_ms += elapsed_time_ms;
990
991 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true);
992 }
993
994 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
995 _mark_cleanup_start_sec = os::elapsedTime();
996 }
997
998 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
999 _should_revert_to_young_gcs = false;
1000 _last_young_gc = true;
1001 _in_marking_window = false;
1002 }
1003
1004 void G1CollectorPolicy::record_concurrent_pause() {
1005 if (_stop_world_start > 0.0) {
1006 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0;
1007 _all_yield_times_ms->add(yield_ms);
1008 }
1009 }
1010
1011 void G1CollectorPolicy::record_concurrent_pause_end() {
1012 }
1013
1014 template<class T>
1015 T sum_of(T* sum_arr, int start, int n, int N) {
1016 T sum = (T)0;
1017 for (int i = 0; i < n; i++) {
1018 int j = (start + i) % N;
1019 sum += sum_arr[j];
1020 }
1021 return sum;
1022 }
1023
1024 void G1CollectorPolicy::print_par_stats(int level,
1025 const char* str,
1026 double* data) {
1027 double min = data[0], max = data[0];
1028 double total = 0.0;
1029 LineBuffer buf(level);
1030 buf.append("[%s (ms):", str);
1031 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1032 double val = data[i];
1033 if (val < min)
1034 min = val;
1035 if (val > max)
1036 max = val;
1037 total += val;
1038 buf.append(" %3.1lf", val);
1039 }
1040 buf.append_and_print_cr("");
1041 double avg = total / (double) no_of_gc_threads();
1042 buf.append_and_print_cr(" Avg: %5.1lf, Min: %5.1lf, Max: %5.1lf, Diff: %5.1lf]",
1043 avg, min, max, max - min);
1044 }
1045
1046 void G1CollectorPolicy::print_par_sizes(int level,
1047 const char* str,
1048 double* data) {
1049 double min = data[0], max = data[0];
1050 double total = 0.0;
1051 LineBuffer buf(level);
1052 buf.append("[%s :", str);
1053 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1054 double val = data[i];
1055 if (val < min)
1056 min = val;
1057 if (val > max)
1058 max = val;
1059 total += val;
1060 buf.append(" %d", (int) val);
1061 }
1062 buf.append_and_print_cr("");
1063 double avg = total / (double) no_of_gc_threads();
1064 buf.append_and_print_cr(" Sum: %d, Avg: %d, Min: %d, Max: %d, Diff: %d]",
1065 (int)total, (int)avg, (int)min, (int)max, (int)max - (int)min);
1066 }
1067
1068 void G1CollectorPolicy::print_stats(int level,
1069 const char* str,
1070 double value) {
1071 LineBuffer(level).append_and_print_cr("[%s: %5.1lf ms]", str, value);
1072 }
1073
1074 void G1CollectorPolicy::print_stats(int level,
1075 const char* str,
1076 int value) {
1077 LineBuffer(level).append_and_print_cr("[%s: %d]", str, value);
1078 }
1079
1080 double G1CollectorPolicy::avg_value(double* data) {
1081 if (G1CollectedHeap::use_parallel_gc_threads()) {
1082 double ret = 0.0;
1083 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1084 ret += data[i];
1085 }
1086 return ret / (double) no_of_gc_threads();
1087 } else {
1088 return data[0];
1089 }
1090 }
1091
1092 double G1CollectorPolicy::max_value(double* data) {
1093 if (G1CollectedHeap::use_parallel_gc_threads()) {
1094 double ret = data[0];
1095 for (uint i = 1; i < no_of_gc_threads(); ++i) {
1096 if (data[i] > ret) {
1097 ret = data[i];
1098 }
1099 }
1100 return ret;
1101 } else {
1102 return data[0];
1103 }
1104 }
1105
1106 double G1CollectorPolicy::sum_of_values(double* data) {
1107 if (G1CollectedHeap::use_parallel_gc_threads()) {
1108 double sum = 0.0;
1109 for (uint i = 0; i < no_of_gc_threads(); i++) {
1110 sum += data[i];
1111 }
1112 return sum;
1113 } else {
1114 return data[0];
1115 }
1116 }
1117
1118 double G1CollectorPolicy::max_sum(double* data1, double* data2) {
1119 double ret = data1[0] + data2[0];
1120
1121 if (G1CollectedHeap::use_parallel_gc_threads()) {
1122 for (uint i = 1; i < no_of_gc_threads(); ++i) {
1123 double data = data1[i] + data2[i];
1124 if (data > ret) {
1125 ret = data;
1126 }
1127 }
1128 }
1129 return ret;
1130 }
1131
1132 // Anything below that is considered to be zero
1133 #define MIN_TIMER_GRANULARITY 0.0000001
1134
1135 void G1CollectorPolicy::record_collection_pause_end(int no_of_gc_threads) {
1136 double end_time_sec = os::elapsedTime();
1137 double elapsed_ms = _last_pause_time_ms;
1138 bool parallel = G1CollectedHeap::use_parallel_gc_threads();
1139 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(),
1140 "otherwise, the subtraction below does not make sense");
1141 size_t rs_size =
1142 _cur_collection_pause_used_regions_at_start - cset_region_length();
1143 size_t cur_used_bytes = _g1->used();
1144 assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
1145 bool last_pause_included_initial_mark = false;
1146 bool update_stats = !_g1->evacuation_failed();
1147 set_no_of_gc_threads(no_of_gc_threads);
1148
1149 #ifndef PRODUCT
1150 if (G1YoungSurvRateVerbose) {
1151 gclog_or_tty->print_cr("");
1152 _short_lived_surv_rate_group->print();
1153 // do that for any other surv rate groups too
1154 }
1155 #endif // PRODUCT
1156
1157 last_pause_included_initial_mark = during_initial_mark_pause();
1158 if (last_pause_included_initial_mark)
1159 record_concurrent_mark_init_end(0.0);
1160
1161 size_t marking_initiating_used_threshold =
1162 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent;
1163
1164 if (!_g1->mark_in_progress() && !_last_young_gc) {
1165 assert(!last_pause_included_initial_mark, "invariant");
1166 if (cur_used_bytes > marking_initiating_used_threshold) {
1167 if (cur_used_bytes > _prev_collection_pause_used_at_end_bytes) {
1168 assert(!during_initial_mark_pause(), "we should not see this here");
1169
1170 ergo_verbose3(ErgoConcCycles,
1171 "request concurrent cycle initiation",
1172 ergo_format_reason("occupancy higher than threshold")
1173 ergo_format_byte("occupancy")
1174 ergo_format_byte_perc("threshold"),
1175 cur_used_bytes,
1176 marking_initiating_used_threshold,
1177 (double) InitiatingHeapOccupancyPercent);
1178
1179 // Note: this might have already been set, if during the last
1180 // pause we decided to start a cycle but at the beginning of
1181 // this pause we decided to postpone it. That's OK.
1182 set_initiate_conc_mark_if_possible();
1183 } else {
1184 ergo_verbose2(ErgoConcCycles,
1185 "do not request concurrent cycle initiation",
1186 ergo_format_reason("occupancy lower than previous occupancy")
1187 ergo_format_byte("occupancy")
1188 ergo_format_byte("previous occupancy"),
1189 cur_used_bytes,
1190 _prev_collection_pause_used_at_end_bytes);
1191 }
1192 }
1193 }
1194
1195 _prev_collection_pause_used_at_end_bytes = cur_used_bytes;
1196
1197 _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0,
1198 end_time_sec, false);
1199
1200 // This assert is exempted when we're doing parallel collection pauses,
1201 // because the fragmentation caused by the parallel GC allocation buffers
1202 // can lead to more memory being used during collection than was used
1203 // before. Best leave this out until the fragmentation problem is fixed.
1204 // Pauses in which evacuation failed can also lead to negative
1205 // collections, since no space is reclaimed from a region containing an
1206 // object whose evacuation failed.
1207 // Further, we're now always doing parallel collection. But I'm still
1208 // leaving this here as a placeholder for a more precise assertion later.
1209 // (DLD, 10/05.)
1210 assert((true || parallel) // Always using GC LABs now.
1211 || _g1->evacuation_failed()
1212 || _cur_collection_pause_used_at_start_bytes >= cur_used_bytes,
1213 "Negative collection");
1214
1215 size_t freed_bytes =
1216 _cur_collection_pause_used_at_start_bytes - cur_used_bytes;
1217 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes;
1218
1219 double survival_fraction =
1220 (double)surviving_bytes/
1221 (double)_collection_set_bytes_used_before;
1222
1223 // These values are used to update the summary information that is
1224 // displayed when TraceGen0Time is enabled, and are output as part
1225 // of the PrintGCDetails output, in the non-parallel case.
1226
1227 double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms);
1228 double mark_stack_scan_time = avg_value(_par_last_mark_stack_scan_times_ms);
1229 double update_rs_time = avg_value(_par_last_update_rs_times_ms);
1230 double update_rs_processed_buffers =
1231 sum_of_values(_par_last_update_rs_processed_buffers);
1232 double scan_rs_time = avg_value(_par_last_scan_rs_times_ms);
1233 double obj_copy_time = avg_value(_par_last_obj_copy_times_ms);
1234 double termination_time = avg_value(_par_last_termination_times_ms);
1235
1236 double known_time = ext_root_scan_time +
1237 mark_stack_scan_time +
1238 update_rs_time +
1239 scan_rs_time +
1240 obj_copy_time;
1241
1242 double other_time_ms = elapsed_ms;
1243
1244 // Subtract the SATB drain time. It's initialized to zero at the
1245 // start of the pause and is updated during the pause if marking
1246 // is in progress.
1247 other_time_ms -= _cur_satb_drain_time_ms;
1248
1249 if (parallel) {
1250 other_time_ms -= _cur_collection_par_time_ms;
1251 } else {
1252 other_time_ms -= known_time;
1253 }
1254
1255 // Subtract the time taken to clean the card table from the
1256 // current value of "other time"
1257 other_time_ms -= _cur_clear_ct_time_ms;
1258
1259 // Subtract the time spent completing marking in the collection
1260 // set. Note if marking is not in progress during the pause
1261 // the value of _mark_closure_time_ms will be zero.
1262 other_time_ms -= _mark_closure_time_ms;
1263
1264 // TraceGen0Time and TraceGen1Time summary info updating.
1265 _all_pause_times_ms->add(elapsed_ms);
1266
1267 if (update_stats) {
1268 _summary->record_total_time_ms(elapsed_ms);
1269 _summary->record_other_time_ms(other_time_ms);
1270
1271 MainBodySummary* body_summary = _summary->main_body_summary();
1272 assert(body_summary != NULL, "should not be null!");
1273
1274 // This will be non-zero iff marking is currently in progress (i.e.
1275 // _g1->mark_in_progress() == true) and the currrent pause was not
1276 // an initial mark pause. Since the body_summary items are NumberSeqs,
1277 // however, they have to be consistent and updated in lock-step with
1278 // each other. Therefore we unconditionally record the SATB drain
1279 // time - even if it's zero.
1280 body_summary->record_satb_drain_time_ms(_cur_satb_drain_time_ms);
1281
1282 body_summary->record_ext_root_scan_time_ms(ext_root_scan_time);
1283 body_summary->record_mark_stack_scan_time_ms(mark_stack_scan_time);
1284 body_summary->record_update_rs_time_ms(update_rs_time);
1285 body_summary->record_scan_rs_time_ms(scan_rs_time);
1286 body_summary->record_obj_copy_time_ms(obj_copy_time);
1287
1288 if (parallel) {
1289 body_summary->record_parallel_time_ms(_cur_collection_par_time_ms);
1290 body_summary->record_termination_time_ms(termination_time);
1291
1292 double parallel_known_time = known_time + termination_time;
1293 double parallel_other_time = _cur_collection_par_time_ms - parallel_known_time;
1294 body_summary->record_parallel_other_time_ms(parallel_other_time);
1295 }
1296
1297 body_summary->record_mark_closure_time_ms(_mark_closure_time_ms);
1298 body_summary->record_clear_ct_time_ms(_cur_clear_ct_time_ms);
1299
1300 // We exempt parallel collection from this check because Alloc Buffer
1301 // fragmentation can produce negative collections. Same with evac
1302 // failure.
1303 // Further, we're now always doing parallel collection. But I'm still
1304 // leaving this here as a placeholder for a more precise assertion later.
1305 // (DLD, 10/05.
1306 assert((true || parallel)
1307 || _g1->evacuation_failed()
1308 || surviving_bytes <= _collection_set_bytes_used_before,
1309 "Or else negative collection!");
1310
1311 // this is where we update the allocation rate of the application
1312 double app_time_ms =
1313 (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms);
1314 if (app_time_ms < MIN_TIMER_GRANULARITY) {
1315 // This usually happens due to the timer not having the required
1316 // granularity. Some Linuxes are the usual culprits.
1317 // We'll just set it to something (arbitrarily) small.
1318 app_time_ms = 1.0;
1319 }
1320 // We maintain the invariant that all objects allocated by mutator
1321 // threads will be allocated out of eden regions. So, we can use
1322 // the eden region number allocated since the previous GC to
1323 // calculate the application's allocate rate. The only exception
1324 // to that is humongous objects that are allocated separately. But
1325 // given that humongous object allocations do not really affect
1326 // either the pause's duration nor when the next pause will take
1327 // place we can safely ignore them here.
1328 size_t regions_allocated = eden_cset_region_length();
1329 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
1330 _alloc_rate_ms_seq->add(alloc_rate_ms);
1331
1332 double interval_ms =
1333 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
1334 update_recent_gc_times(end_time_sec, elapsed_ms);
1335 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
1336 if (recent_avg_pause_time_ratio() < 0.0 ||
1337 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) {
1338 #ifndef PRODUCT
1339 // Dump info to allow post-facto debugging
1340 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds");
1341 gclog_or_tty->print_cr("-------------------------------------------");
1342 gclog_or_tty->print_cr("Recent GC Times (ms):");
1343 _recent_gc_times_ms->dump();
1344 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec);
1345 _recent_prev_end_times_for_all_gcs_sec->dump();
1346 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f",
1347 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio());
1348 // In debug mode, terminate the JVM if the user wants to debug at this point.
1349 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above");
1350 #endif // !PRODUCT
1351 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
1352 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
1353 if (_recent_avg_pause_time_ratio < 0.0) {
1354 _recent_avg_pause_time_ratio = 0.0;
1355 } else {
1356 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
1357 _recent_avg_pause_time_ratio = 1.0;
1358 }
1359 }
1360 }
1361
1362 for (int i = 0; i < _aux_num; ++i) {
1363 if (_cur_aux_times_set[i]) {
1364 _all_aux_times_ms[i].add(_cur_aux_times_ms[i]);
1365 }
1366 }
1367
1368 // PrintGCDetails output
1369 if (PrintGCDetails) {
1370 bool print_marking_info =
1371 _g1->mark_in_progress() && !last_pause_included_initial_mark;
1372
1373 gclog_or_tty->print_cr("%s, %1.8lf secs]",
1374 (last_pause_included_initial_mark) ? " (initial-mark)" : "",
1375 elapsed_ms / 1000.0);
1376
1377 if (print_marking_info) {
1378 print_stats(1, "SATB Drain Time", _cur_satb_drain_time_ms);
1379 }
1380
1381 if (parallel) {
1382 print_stats(1, "Parallel Time", _cur_collection_par_time_ms);
1383 print_par_stats(2, "GC Worker Start", _par_last_gc_worker_start_times_ms);
1384 print_par_stats(2, "Ext Root Scanning", _par_last_ext_root_scan_times_ms);
1385 if (print_marking_info) {
1386 print_par_stats(2, "Mark Stack Scanning", _par_last_mark_stack_scan_times_ms);
1387 }
1388 print_par_stats(2, "Update RS", _par_last_update_rs_times_ms);
1389 print_par_sizes(3, "Processed Buffers", _par_last_update_rs_processed_buffers);
1390 print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms);
1391 print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms);
1392 print_par_stats(2, "Termination", _par_last_termination_times_ms);
1393 print_par_sizes(3, "Termination Attempts", _par_last_termination_attempts);
1394 print_par_stats(2, "GC Worker End", _par_last_gc_worker_end_times_ms);
1395
1396 for (int i = 0; i < _parallel_gc_threads; i++) {
1397 _par_last_gc_worker_times_ms[i] = _par_last_gc_worker_end_times_ms[i] - _par_last_gc_worker_start_times_ms[i];
1398
1399 double worker_known_time = _par_last_ext_root_scan_times_ms[i] +
1400 _par_last_mark_stack_scan_times_ms[i] +
1401 _par_last_update_rs_times_ms[i] +
1402 _par_last_scan_rs_times_ms[i] +
1403 _par_last_obj_copy_times_ms[i] +
1404 _par_last_termination_times_ms[i];
1405
1406 _par_last_gc_worker_other_times_ms[i] = _cur_collection_par_time_ms - worker_known_time;
1407 }
1408 print_par_stats(2, "GC Worker", _par_last_gc_worker_times_ms);
1409 print_par_stats(2, "GC Worker Other", _par_last_gc_worker_other_times_ms);
1410 } else {
1411 print_stats(1, "Ext Root Scanning", ext_root_scan_time);
1412 if (print_marking_info) {
1413 print_stats(1, "Mark Stack Scanning", mark_stack_scan_time);
1414 }
1415 print_stats(1, "Update RS", update_rs_time);
1416 print_stats(2, "Processed Buffers", (int)update_rs_processed_buffers);
1417 print_stats(1, "Scan RS", scan_rs_time);
1418 print_stats(1, "Object Copying", obj_copy_time);
1419 }
1420 if (print_marking_info) {
1421 print_stats(1, "Complete CSet Marking", _mark_closure_time_ms);
1422 }
1423 print_stats(1, "Clear CT", _cur_clear_ct_time_ms);
1424 #ifndef PRODUCT
1425 print_stats(1, "Cur Clear CC", _cur_clear_cc_time_ms);
1426 print_stats(1, "Cum Clear CC", _cum_clear_cc_time_ms);
1427 print_stats(1, "Min Clear CC", _min_clear_cc_time_ms);
1428 print_stats(1, "Max Clear CC", _max_clear_cc_time_ms);
1429 if (_num_cc_clears > 0) {
1430 print_stats(1, "Avg Clear CC", _cum_clear_cc_time_ms / ((double)_num_cc_clears));
1431 }
1432 #endif
1433 print_stats(1, "Other", other_time_ms);
1434 print_stats(2, "Choose CSet",
1435 (_recorded_young_cset_choice_time_ms +
1436 _recorded_non_young_cset_choice_time_ms));
1437 print_stats(2, "Ref Proc", _cur_ref_proc_time_ms);
1438 print_stats(2, "Ref Enq", _cur_ref_enq_time_ms);
1439 print_stats(2, "Free CSet",
1440 (_recorded_young_free_cset_time_ms +
1441 _recorded_non_young_free_cset_time_ms));
1442
1443 for (int i = 0; i < _aux_num; ++i) {
1444 if (_cur_aux_times_set[i]) {
1445 char buffer[96];
1446 sprintf(buffer, "Aux%d", i);
1447 print_stats(1, buffer, _cur_aux_times_ms[i]);
1448 }
1449 }
1450 }
1451
1452 // Update the efficiency-since-mark vars.
1453 double proc_ms = elapsed_ms * (double) _parallel_gc_threads;
1454 if (elapsed_ms < MIN_TIMER_GRANULARITY) {
1455 // This usually happens due to the timer not having the required
1456 // granularity. Some Linuxes are the usual culprits.
1457 // We'll just set it to something (arbitrarily) small.
1458 proc_ms = 1.0;
1459 }
1460 double cur_efficiency = (double) freed_bytes / proc_ms;
1461
1462 bool new_in_marking_window = _in_marking_window;
1463 bool new_in_marking_window_im = false;
1464 if (during_initial_mark_pause()) {
1465 new_in_marking_window = true;
1466 new_in_marking_window_im = true;
1467 }
1468
1469 if (_last_young_gc) {
1470 if (!last_pause_included_initial_mark) {
1471 ergo_verbose2(ErgoMixedGCs,
1472 "start mixed GCs",
1473 ergo_format_byte_perc("known garbage"),
1474 _known_garbage_bytes, _known_garbage_ratio * 100.0);
1475 set_gcs_are_young(false);
1476 } else {
1477 ergo_verbose0(ErgoMixedGCs,
1478 "do not start mixed GCs",
1479 ergo_format_reason("concurrent cycle is about to start"));
1480 }
1481 _last_young_gc = false;
1482 }
1483
1484 if (!_last_gc_was_young) {
1485 if (_should_revert_to_young_gcs) {
1486 ergo_verbose2(ErgoMixedGCs,
1487 "end mixed GCs",
1488 ergo_format_reason("mixed GCs end requested")
1489 ergo_format_byte_perc("known garbage"),
1490 _known_garbage_bytes, _known_garbage_ratio * 100.0);
1491 set_gcs_are_young(true);
1492 } else if (_known_garbage_ratio < 0.05) {
1493 ergo_verbose3(ErgoMixedGCs,
1494 "end mixed GCs",
1495 ergo_format_reason("known garbage percent lower than threshold")
1496 ergo_format_byte_perc("known garbage")
1497 ergo_format_perc("threshold"),
1498 _known_garbage_bytes, _known_garbage_ratio * 100.0,
1499 0.05 * 100.0);
1500 set_gcs_are_young(true);
1501 } else if (adaptive_young_list_length() &&
1502 (get_gc_eff_factor() * cur_efficiency < predict_young_gc_eff())) {
1503 ergo_verbose5(ErgoMixedGCs,
1504 "end mixed GCs",
1505 ergo_format_reason("current GC efficiency lower than "
1506 "predicted young GC efficiency")
1507 ergo_format_double("GC efficiency factor")
1508 ergo_format_double("current GC efficiency")
1509 ergo_format_double("predicted young GC efficiency")
1510 ergo_format_byte_perc("known garbage"),
1511 get_gc_eff_factor(), cur_efficiency,
1512 predict_young_gc_eff(),
1513 _known_garbage_bytes, _known_garbage_ratio * 100.0);
1514 set_gcs_are_young(true);
1515 }
1516 }
1517 _should_revert_to_young_gcs = false;
1518
1519 if (_last_gc_was_young && !_during_marking) {
1520 _young_gc_eff_seq->add(cur_efficiency);
1521 }
1522
1523 _short_lived_surv_rate_group->start_adding_regions();
1524 // do that for any other surv rate groupsx
1525
1526 if (update_stats) {
1527 double pause_time_ms = elapsed_ms;
1528
1529 size_t diff = 0;
1530 if (_max_pending_cards >= _pending_cards)
1531 diff = _max_pending_cards - _pending_cards;
1532 _pending_card_diff_seq->add((double) diff);
1533
1534 double cost_per_card_ms = 0.0;
1535 if (_pending_cards > 0) {
1536 cost_per_card_ms = update_rs_time / (double) _pending_cards;
1537 _cost_per_card_ms_seq->add(cost_per_card_ms);
1538 }
1539
1540 size_t cards_scanned = _g1->cards_scanned();
1541
1542 double cost_per_entry_ms = 0.0;
1543 if (cards_scanned > 10) {
1544 cost_per_entry_ms = scan_rs_time / (double) cards_scanned;
1545 if (_last_gc_was_young) {
1546 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
1547 } else {
1548 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms);
1549 }
1550 }
1551
1552 if (_max_rs_lengths > 0) {
1553 double cards_per_entry_ratio =
1554 (double) cards_scanned / (double) _max_rs_lengths;
1555 if (_last_gc_was_young) {
1556 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1557 } else {
1558 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1559 }
1560 }
1561
1562 // It turns out that, sometimes, _max_rs_lengths can get smaller
1563 // than _recorded_rs_lengths which causes rs_length_diff to get
1564 // very large and mess up the RSet length predictions. We'll be
1565 // defensive until we work out why this happens.
1566 size_t rs_length_diff = 0;
1567 if (_max_rs_lengths > _recorded_rs_lengths) {
1568 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths;
1569 }
1570 _rs_length_diff_seq->add((double) rs_length_diff);
1571
1572 size_t copied_bytes = surviving_bytes;
1573 double cost_per_byte_ms = 0.0;
1574 if (copied_bytes > 0) {
1575 cost_per_byte_ms = obj_copy_time / (double) copied_bytes;
1576 if (_in_marking_window) {
1577 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
1578 } else {
1579 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
1580 }
1581 }
1582
1583 double all_other_time_ms = pause_time_ms -
1584 (update_rs_time + scan_rs_time + obj_copy_time +
1585 _mark_closure_time_ms + termination_time);
1586
1587 double young_other_time_ms = 0.0;
1588 if (young_cset_region_length() > 0) {
1589 young_other_time_ms =
1590 _recorded_young_cset_choice_time_ms +
1591 _recorded_young_free_cset_time_ms;
1592 _young_other_cost_per_region_ms_seq->add(young_other_time_ms /
1593 (double) young_cset_region_length());
1594 }
1595 double non_young_other_time_ms = 0.0;
1596 if (old_cset_region_length() > 0) {
1597 non_young_other_time_ms =
1598 _recorded_non_young_cset_choice_time_ms +
1599 _recorded_non_young_free_cset_time_ms;
1600
1601 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms /
1602 (double) old_cset_region_length());
1603 }
1604
1605 double constant_other_time_ms = all_other_time_ms -
1606 (young_other_time_ms + non_young_other_time_ms);
1607 _constant_other_time_ms_seq->add(constant_other_time_ms);
1608
1609 double survival_ratio = 0.0;
1610 if (_bytes_in_collection_set_before_gc > 0) {
1611 survival_ratio = (double) _bytes_copied_during_gc /
1612 (double) _bytes_in_collection_set_before_gc;
1613 }
1614
1615 _pending_cards_seq->add((double) _pending_cards);
1616 _rs_lengths_seq->add((double) _max_rs_lengths);
1617
1618 double expensive_region_limit_ms =
1619 (double) MaxGCPauseMillis - predict_constant_other_time_ms();
1620 if (expensive_region_limit_ms < 0.0) {
1621 // this means that the other time was predicted to be longer than
1622 // than the max pause time
1623 expensive_region_limit_ms = (double) MaxGCPauseMillis;
1624 }
1625 _expensive_region_limit_ms = expensive_region_limit_ms;
1626 }
1627
1628 _in_marking_window = new_in_marking_window;
1629 _in_marking_window_im = new_in_marking_window_im;
1630 _free_regions_at_end_of_collection = _g1->free_regions();
1631 update_young_list_target_length();
1632
1633 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
1634 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
1635 adjust_concurrent_refinement(update_rs_time, update_rs_processed_buffers, update_rs_time_goal_ms);
1636
1637 assert(assertMarkedBytesDataOK(), "Marked regions not OK at pause end.");
1638 }
1639
1640 #define EXT_SIZE_FORMAT "%d%s"
1641 #define EXT_SIZE_PARAMS(bytes) \
1642 byte_size_in_proper_unit((bytes)), \
1643 proper_unit_for_byte_size((bytes))
1644
1645 void G1CollectorPolicy::print_heap_transition() {
1646 if (PrintGCDetails) {
1647 YoungList* young_list = _g1->young_list();
1648 size_t eden_bytes = young_list->eden_used_bytes();
1649 size_t survivor_bytes = young_list->survivor_used_bytes();
1650 size_t used_before_gc = _cur_collection_pause_used_at_start_bytes;
1651 size_t used = _g1->used();
1652 size_t capacity = _g1->capacity();
1653 size_t eden_capacity =
1654 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_bytes;
1655
1656 gclog_or_tty->print_cr(
1657 " [Eden: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT") "
1658 "Survivors: "EXT_SIZE_FORMAT"->"EXT_SIZE_FORMAT" "
1659 "Heap: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"
1660 EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")]",
1661 EXT_SIZE_PARAMS(_eden_bytes_before_gc),
1662 EXT_SIZE_PARAMS(_prev_eden_capacity),
1663 EXT_SIZE_PARAMS(eden_bytes),
1664 EXT_SIZE_PARAMS(eden_capacity),
1665 EXT_SIZE_PARAMS(_survivor_bytes_before_gc),
1666 EXT_SIZE_PARAMS(survivor_bytes),
1667 EXT_SIZE_PARAMS(used_before_gc),
1668 EXT_SIZE_PARAMS(_capacity_before_gc),
1669 EXT_SIZE_PARAMS(used),
1670 EXT_SIZE_PARAMS(capacity));
1671
1672 _prev_eden_capacity = eden_capacity;
1673 } else if (PrintGC) {
1674 _g1->print_size_transition(gclog_or_tty,
1675 _cur_collection_pause_used_at_start_bytes,
1676 _g1->used(), _g1->capacity());
1677 }
1678 }
1679
1680 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1681 double update_rs_processed_buffers,
1682 double goal_ms) {
1683 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1684 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1685
1686 if (G1UseAdaptiveConcRefinement) {
1687 const int k_gy = 3, k_gr = 6;
1688 const double inc_k = 1.1, dec_k = 0.9;
1689
1690 int g = cg1r->green_zone();
1691 if (update_rs_time > goal_ms) {
1692 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1693 } else {
1694 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1695 g = (int)MAX2(g * inc_k, g + 1.0);
1696 }
1697 }
1698 // Change the refinement threads params
1699 cg1r->set_green_zone(g);
1700 cg1r->set_yellow_zone(g * k_gy);
1701 cg1r->set_red_zone(g * k_gr);
1702 cg1r->reinitialize_threads();
1703
1704 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1);
1705 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1706 cg1r->yellow_zone());
1707 // Change the barrier params
1708 dcqs.set_process_completed_threshold(processing_threshold);
1709 dcqs.set_max_completed_queue(cg1r->red_zone());
1710 }
1711
1712 int curr_queue_size = dcqs.completed_buffers_num();
1713 if (curr_queue_size >= cg1r->yellow_zone()) {
1714 dcqs.set_completed_queue_padding(curr_queue_size);
1715 } else {
1716 dcqs.set_completed_queue_padding(0);
1717 }
1718 dcqs.notify_if_necessary();
1719 }
1720
1721 double
1722 G1CollectorPolicy::
1723 predict_young_collection_elapsed_time_ms(size_t adjustment) {
1724 guarantee( adjustment == 0 || adjustment == 1, "invariant" );
1725
1726 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1727 size_t young_num = g1h->young_list()->length();
1728 if (young_num == 0)
1729 return 0.0;
1730
1731 young_num += adjustment;
1732 size_t pending_cards = predict_pending_cards();
1733 size_t rs_lengths = g1h->young_list()->sampled_rs_lengths() +
1734 predict_rs_length_diff();
1735 size_t card_num;
1736 if (gcs_are_young()) {
1737 card_num = predict_young_card_num(rs_lengths);
1738 } else {
1739 card_num = predict_non_young_card_num(rs_lengths);
1740 }
1741 size_t young_byte_size = young_num * HeapRegion::GrainBytes;
1742 double accum_yg_surv_rate =
1743 _short_lived_surv_rate_group->accum_surv_rate(adjustment);
1744
1745 size_t bytes_to_copy =
1746 (size_t) (accum_yg_surv_rate * (double) HeapRegion::GrainBytes);
1747
1748 return
1749 predict_rs_update_time_ms(pending_cards) +
1750 predict_rs_scan_time_ms(card_num) +
1751 predict_object_copy_time_ms(bytes_to_copy) +
1752 predict_young_other_time_ms(young_num) +
1753 predict_constant_other_time_ms();
1754 }
1755
1756 double
1757 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) {
1758 size_t rs_length = predict_rs_length_diff();
1759 size_t card_num;
1760 if (gcs_are_young()) {
1761 card_num = predict_young_card_num(rs_length);
1762 } else {
1763 card_num = predict_non_young_card_num(rs_length);
1764 }
1765 return predict_base_elapsed_time_ms(pending_cards, card_num);
1766 }
1767
1768 double
1769 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
1770 size_t scanned_cards) {
1771 return
1772 predict_rs_update_time_ms(pending_cards) +
1773 predict_rs_scan_time_ms(scanned_cards) +
1774 predict_constant_other_time_ms();
1775 }
1776
1777 double
1778 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
1779 bool young) {
1780 size_t rs_length = hr->rem_set()->occupied();
1781 size_t card_num;
1782 if (gcs_are_young()) {
1783 card_num = predict_young_card_num(rs_length);
1784 } else {
1785 card_num = predict_non_young_card_num(rs_length);
1786 }
1787 size_t bytes_to_copy = predict_bytes_to_copy(hr);
1788
1789 double region_elapsed_time_ms =
1790 predict_rs_scan_time_ms(card_num) +
1791 predict_object_copy_time_ms(bytes_to_copy);
1792
1793 if (young)
1794 region_elapsed_time_ms += predict_young_other_time_ms(1);
1795 else
1796 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
1797
1798 return region_elapsed_time_ms;
1799 }
1800
1801 size_t
1802 G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) {
1803 size_t bytes_to_copy;
1804 if (hr->is_marked())
1805 bytes_to_copy = hr->max_live_bytes();
1806 else {
1807 guarantee( hr->is_young() && hr->age_in_surv_rate_group() != -1,
1808 "invariant" );
1809 int age = hr->age_in_surv_rate_group();
1810 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
1811 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate);
1812 }
1813
1814 return bytes_to_copy;
1815 }
1816
1817 void
1818 G1CollectorPolicy::init_cset_region_lengths(size_t eden_cset_region_length,
1819 size_t survivor_cset_region_length) {
1820 _eden_cset_region_length = eden_cset_region_length;
1821 _survivor_cset_region_length = survivor_cset_region_length;
1822 _old_cset_region_length = 0;
1823 }
1824
1825 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) {
1826 _recorded_rs_lengths = rs_lengths;
1827 }
1828
1829 void G1CollectorPolicy::check_if_region_is_too_expensive(double
1830 predicted_time_ms) {
1831 // I don't think we need to do this when in young GC mode since
1832 // marking will be initiated next time we hit the soft limit anyway...
1833 if (predicted_time_ms > _expensive_region_limit_ms) {
1834 ergo_verbose2(ErgoMixedGCs,
1835 "request mixed GCs end",
1836 ergo_format_reason("predicted region time higher than threshold")
1837 ergo_format_ms("predicted region time")
1838 ergo_format_ms("threshold"),
1839 predicted_time_ms, _expensive_region_limit_ms);
1840 // no point in doing another mixed GC
1841 _should_revert_to_young_gcs = true;
1842 }
1843 }
1844
1845 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
1846 double elapsed_ms) {
1847 _recent_gc_times_ms->add(elapsed_ms);
1848 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
1849 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
1850 }
1851
1852 size_t G1CollectorPolicy::expansion_amount() {
1853 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0;
1854 double threshold = _gc_overhead_perc;
1855 if (recent_gc_overhead > threshold) {
1856 // We will double the existing space, or take
1857 // G1ExpandByPercentOfAvailable % of the available expansion
1858 // space, whichever is smaller, bounded below by a minimum
1859 // expansion (unless that's all that's left.)
1860 const size_t min_expand_bytes = 1*M;
1861 size_t reserved_bytes = _g1->max_capacity();
1862 size_t committed_bytes = _g1->capacity();
1863 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
1864 size_t expand_bytes;
1865 size_t expand_bytes_via_pct =
1866 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
1867 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
1868 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
1869 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
1870
1871 ergo_verbose5(ErgoHeapSizing,
1872 "attempt heap expansion",
1873 ergo_format_reason("recent GC overhead higher than "
1874 "threshold after GC")
1875 ergo_format_perc("recent GC overhead")
1876 ergo_format_perc("threshold")
1877 ergo_format_byte("uncommitted")
1878 ergo_format_byte_perc("calculated expansion amount"),
1879 recent_gc_overhead, threshold,
1880 uncommitted_bytes,
1881 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable);
1882
1883 return expand_bytes;
1884 } else {
1885 return 0;
1886 }
1887 }
1888
1889 class CountCSClosure: public HeapRegionClosure {
1890 G1CollectorPolicy* _g1_policy;
1891 public:
1892 CountCSClosure(G1CollectorPolicy* g1_policy) :
1893 _g1_policy(g1_policy) {}
1894 bool doHeapRegion(HeapRegion* r) {
1895 _g1_policy->_bytes_in_collection_set_before_gc += r->used();
1896 return false;
1897 }
1898 };
1899
1900 void G1CollectorPolicy::count_CS_bytes_used() {
1901 CountCSClosure cs_closure(this);
1902 _g1->collection_set_iterate(&cs_closure);
1903 }
1904
1905 void G1CollectorPolicy::print_summary(int level,
1906 const char* str,
1907 NumberSeq* seq) const {
1908 double sum = seq->sum();
1909 LineBuffer(level + 1).append_and_print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)",
1910 str, sum / 1000.0, seq->avg());
1911 }
1912
1913 void G1CollectorPolicy::print_summary_sd(int level,
1914 const char* str,
1915 NumberSeq* seq) const {
1916 print_summary(level, str, seq);
1917 LineBuffer(level + 6).append_and_print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
1918 seq->num(), seq->sd(), seq->maximum());
1919 }
1920
1921 void G1CollectorPolicy::check_other_times(int level,
1922 NumberSeq* other_times_ms,
1923 NumberSeq* calc_other_times_ms) const {
1924 bool should_print = false;
1925 LineBuffer buf(level + 2);
1926
1927 double max_sum = MAX2(fabs(other_times_ms->sum()),
1928 fabs(calc_other_times_ms->sum()));
1929 double min_sum = MIN2(fabs(other_times_ms->sum()),
1930 fabs(calc_other_times_ms->sum()));
1931 double sum_ratio = max_sum / min_sum;
1932 if (sum_ratio > 1.1) {
1933 should_print = true;
1934 buf.append_and_print_cr("## CALCULATED OTHER SUM DOESN'T MATCH RECORDED ###");
1935 }
1936
1937 double max_avg = MAX2(fabs(other_times_ms->avg()),
1938 fabs(calc_other_times_ms->avg()));
1939 double min_avg = MIN2(fabs(other_times_ms->avg()),
1940 fabs(calc_other_times_ms->avg()));
1941 double avg_ratio = max_avg / min_avg;
1942 if (avg_ratio > 1.1) {
1943 should_print = true;
1944 buf.append_and_print_cr("## CALCULATED OTHER AVG DOESN'T MATCH RECORDED ###");
1945 }
1946
1947 if (other_times_ms->sum() < -0.01) {
1948 buf.append_and_print_cr("## RECORDED OTHER SUM IS NEGATIVE ###");
1949 }
1950
1951 if (other_times_ms->avg() < -0.01) {
1952 buf.append_and_print_cr("## RECORDED OTHER AVG IS NEGATIVE ###");
1953 }
1954
1955 if (calc_other_times_ms->sum() < -0.01) {
1956 should_print = true;
1957 buf.append_and_print_cr("## CALCULATED OTHER SUM IS NEGATIVE ###");
1958 }
1959
1960 if (calc_other_times_ms->avg() < -0.01) {
1961 should_print = true;
1962 buf.append_and_print_cr("## CALCULATED OTHER AVG IS NEGATIVE ###");
1963 }
1964
1965 if (should_print)
1966 print_summary(level, "Other(Calc)", calc_other_times_ms);
1967 }
1968
1969 void G1CollectorPolicy::print_summary(PauseSummary* summary) const {
1970 bool parallel = G1CollectedHeap::use_parallel_gc_threads();
1971 MainBodySummary* body_summary = summary->main_body_summary();
1972 if (summary->get_total_seq()->num() > 0) {
1973 print_summary_sd(0, "Evacuation Pauses", summary->get_total_seq());
1974 if (body_summary != NULL) {
1975 print_summary(1, "SATB Drain", body_summary->get_satb_drain_seq());
1976 if (parallel) {
1977 print_summary(1, "Parallel Time", body_summary->get_parallel_seq());
1978 print_summary(2, "Ext Root Scanning", body_summary->get_ext_root_scan_seq());
1979 print_summary(2, "Mark Stack Scanning", body_summary->get_mark_stack_scan_seq());
1980 print_summary(2, "Update RS", body_summary->get_update_rs_seq());
1981 print_summary(2, "Scan RS", body_summary->get_scan_rs_seq());
1982 print_summary(2, "Object Copy", body_summary->get_obj_copy_seq());
1983 print_summary(2, "Termination", body_summary->get_termination_seq());
1984 print_summary(2, "Parallel Other", body_summary->get_parallel_other_seq());
1985 {
1986 NumberSeq* other_parts[] = {
1987 body_summary->get_ext_root_scan_seq(),
1988 body_summary->get_mark_stack_scan_seq(),
1989 body_summary->get_update_rs_seq(),
1990 body_summary->get_scan_rs_seq(),
1991 body_summary->get_obj_copy_seq(),
1992 body_summary->get_termination_seq()
1993 };
1994 NumberSeq calc_other_times_ms(body_summary->get_parallel_seq(),
1995 6, other_parts);
1996 check_other_times(2, body_summary->get_parallel_other_seq(),
1997 &calc_other_times_ms);
1998 }
1999 } else {
2000 print_summary(1, "Ext Root Scanning", body_summary->get_ext_root_scan_seq());
2001 print_summary(1, "Mark Stack Scanning", body_summary->get_mark_stack_scan_seq());
2002 print_summary(1, "Update RS", body_summary->get_update_rs_seq());
2003 print_summary(1, "Scan RS", body_summary->get_scan_rs_seq());
2004 print_summary(1, "Object Copy", body_summary->get_obj_copy_seq());
2005 }
2006 }
2007 print_summary(1, "Mark Closure", body_summary->get_mark_closure_seq());
2008 print_summary(1, "Clear CT", body_summary->get_clear_ct_seq());
2009 print_summary(1, "Other", summary->get_other_seq());
2010 {
2011 if (body_summary != NULL) {
2012 NumberSeq calc_other_times_ms;
2013 if (parallel) {
2014 // parallel
2015 NumberSeq* other_parts[] = {
2016 body_summary->get_satb_drain_seq(),
2017 body_summary->get_parallel_seq(),
2018 body_summary->get_clear_ct_seq()
2019 };
2020 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
2021 3, other_parts);
2022 } else {
2023 // serial
2024 NumberSeq* other_parts[] = {
2025 body_summary->get_satb_drain_seq(),
2026 body_summary->get_update_rs_seq(),
2027 body_summary->get_ext_root_scan_seq(),
2028 body_summary->get_mark_stack_scan_seq(),
2029 body_summary->get_scan_rs_seq(),
2030 body_summary->get_obj_copy_seq()
2031 };
2032 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
2033 6, other_parts);
2034 }
2035 check_other_times(1, summary->get_other_seq(), &calc_other_times_ms);
2036 }
2037 }
2038 } else {
2039 LineBuffer(1).append_and_print_cr("none");
2040 }
2041 LineBuffer(0).append_and_print_cr("");
2042 }
2043
2044 void G1CollectorPolicy::print_tracing_info() const {
2045 if (TraceGen0Time) {
2046 gclog_or_tty->print_cr("ALL PAUSES");
2047 print_summary_sd(0, "Total", _all_pause_times_ms);
2048 gclog_or_tty->print_cr("");
2049 gclog_or_tty->print_cr("");
2050 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num);
2051 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num);
2052 gclog_or_tty->print_cr("");
2053
2054 gclog_or_tty->print_cr("EVACUATION PAUSES");
2055 print_summary(_summary);
2056
2057 gclog_or_tty->print_cr("MISC");
2058 print_summary_sd(0, "Stop World", _all_stop_world_times_ms);
2059 print_summary_sd(0, "Yields", _all_yield_times_ms);
2060 for (int i = 0; i < _aux_num; ++i) {
2061 if (_all_aux_times_ms[i].num() > 0) {
2062 char buffer[96];
2063 sprintf(buffer, "Aux%d", i);
2064 print_summary_sd(0, buffer, &_all_aux_times_ms[i]);
2065 }
2066 }
2067 }
2068 if (TraceGen1Time) {
2069 if (_all_full_gc_times_ms->num() > 0) {
2070 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s",
2071 _all_full_gc_times_ms->num(),
2072 _all_full_gc_times_ms->sum() / 1000.0);
2073 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times_ms->avg());
2074 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]",
2075 _all_full_gc_times_ms->sd(),
2076 _all_full_gc_times_ms->maximum());
2077 }
2078 }
2079 }
2080
2081 void G1CollectorPolicy::print_yg_surv_rate_info() const {
2082 #ifndef PRODUCT
2083 _short_lived_surv_rate_group->print_surv_rate_summary();
2084 // add this call for any other surv rate groups
2085 #endif // PRODUCT
2086 }
2087
2088 #ifndef PRODUCT
2089 // for debugging, bit of a hack...
2090 static char*
2091 region_num_to_mbs(int length) {
2092 static char buffer[64];
2093 double bytes = (double) (length * HeapRegion::GrainBytes);
2094 double mbs = bytes / (double) (1024 * 1024);
2095 sprintf(buffer, "%7.2lfMB", mbs);
2096 return buffer;
2097 }
2098 #endif // PRODUCT
2099
2100 size_t G1CollectorPolicy::max_regions(int purpose) {
2101 switch (purpose) {
2102 case GCAllocForSurvived:
2103 return _max_survivor_regions;
2104 case GCAllocForTenured:
2105 return REGIONS_UNLIMITED;
2106 default:
2107 ShouldNotReachHere();
2108 return REGIONS_UNLIMITED;
2109 };
2110 }
2111
2112 void G1CollectorPolicy::update_max_gc_locker_expansion() {
2113 size_t expansion_region_num = 0;
2114 if (GCLockerEdenExpansionPercent > 0) {
2115 double perc = (double) GCLockerEdenExpansionPercent / 100.0;
2116 double expansion_region_num_d = perc * (double) _young_list_target_length;
2117 // We use ceiling so that if expansion_region_num_d is > 0.0 (but
2118 // less than 1.0) we'll get 1.
2119 expansion_region_num = (size_t) ceil(expansion_region_num_d);
2120 } else {
2121 assert(expansion_region_num == 0, "sanity");
2122 }
2123 _young_list_max_length = _young_list_target_length + expansion_region_num;
2124 assert(_young_list_target_length <= _young_list_max_length, "post-condition");
2125 }
2126
2127 // Calculates survivor space parameters.
2128 void G1CollectorPolicy::update_survivors_policy() {
2129 double max_survivor_regions_d =
2130 (double) _young_list_target_length / (double) SurvivorRatio;
2131 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but
2132 // smaller than 1.0) we'll get 1.
2133 _max_survivor_regions = (size_t) ceil(max_survivor_regions_d);
2134
2135 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
2136 HeapRegion::GrainWords * _max_survivor_regions);
2137 }
2138
2139 #ifndef PRODUCT
2140 class HRSortIndexIsOKClosure: public HeapRegionClosure {
2141 CollectionSetChooser* _chooser;
2142 public:
2143 HRSortIndexIsOKClosure(CollectionSetChooser* chooser) :
2144 _chooser(chooser) {}
2145
2146 bool doHeapRegion(HeapRegion* r) {
2147 if (!r->continuesHumongous()) {
2148 assert(_chooser->regionProperlyOrdered(r), "Ought to be.");
2149 }
2150 return false;
2151 }
2152 };
2153
2154 bool G1CollectorPolicy::assertMarkedBytesDataOK() {
2155 HRSortIndexIsOKClosure cl(_collectionSetChooser);
2156 _g1->heap_region_iterate(&cl);
2157 return true;
2158 }
2159 #endif
2160
2161 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(
2162 GCCause::Cause gc_cause) {
2163 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2164 if (!during_cycle) {
2165 ergo_verbose1(ErgoConcCycles,
2166 "request concurrent cycle initiation",
2167 ergo_format_reason("requested by GC cause")
2168 ergo_format_str("GC cause"),
2169 GCCause::to_string(gc_cause));
2170 set_initiate_conc_mark_if_possible();
2171 return true;
2172 } else {
2173 ergo_verbose1(ErgoConcCycles,
2174 "do not request concurrent cycle initiation",
2175 ergo_format_reason("concurrent cycle already in progress")
2176 ergo_format_str("GC cause"),
2177 GCCause::to_string(gc_cause));
2178 return false;
2179 }
2180 }
2181
2182 void
2183 G1CollectorPolicy::decide_on_conc_mark_initiation() {
2184 // We are about to decide on whether this pause will be an
2185 // initial-mark pause.
2186
2187 // First, during_initial_mark_pause() should not be already set. We
2188 // will set it here if we have to. However, it should be cleared by
2189 // the end of the pause (it's only set for the duration of an
2190 // initial-mark pause).
2191 assert(!during_initial_mark_pause(), "pre-condition");
2192
2193 if (initiate_conc_mark_if_possible()) {
2194 // We had noticed on a previous pause that the heap occupancy has
2195 // gone over the initiating threshold and we should start a
2196 // concurrent marking cycle. So we might initiate one.
2197
2198 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2199 if (!during_cycle) {
2200 // The concurrent marking thread is not "during a cycle", i.e.,
2201 // it has completed the last one. So we can go ahead and
2202 // initiate a new cycle.
2203
2204 set_during_initial_mark_pause();
2205 // We do not allow mixed GCs during marking.
2206 if (!gcs_are_young()) {
2207 set_gcs_are_young(true);
2208 ergo_verbose0(ErgoMixedGCs,
2209 "end mixed GCs",
2210 ergo_format_reason("concurrent cycle is about to start"));
2211 }
2212
2213 // And we can now clear initiate_conc_mark_if_possible() as
2214 // we've already acted on it.
2215 clear_initiate_conc_mark_if_possible();
2216
2217 ergo_verbose0(ErgoConcCycles,
2218 "initiate concurrent cycle",
2219 ergo_format_reason("concurrent cycle initiation requested"));
2220 } else {
2221 // The concurrent marking thread is still finishing up the
2222 // previous cycle. If we start one right now the two cycles
2223 // overlap. In particular, the concurrent marking thread might
2224 // be in the process of clearing the next marking bitmap (which
2225 // we will use for the next cycle if we start one). Starting a
2226 // cycle now will be bad given that parts of the marking
2227 // information might get cleared by the marking thread. And we
2228 // cannot wait for the marking thread to finish the cycle as it
2229 // periodically yields while clearing the next marking bitmap
2230 // and, if it's in a yield point, it's waiting for us to
2231 // finish. So, at this point we will not start a cycle and we'll
2232 // let the concurrent marking thread complete the last one.
2233 ergo_verbose0(ErgoConcCycles,
2234 "do not initiate concurrent cycle",
2235 ergo_format_reason("concurrent cycle already in progress"));
2236 }
2237 }
2238 }
2239
2240 class KnownGarbageClosure: public HeapRegionClosure {
2241 CollectionSetChooser* _hrSorted;
2242
2243 public:
2244 KnownGarbageClosure(CollectionSetChooser* hrSorted) :
2245 _hrSorted(hrSorted)
2246 {}
2247
2248 bool doHeapRegion(HeapRegion* r) {
2249 // We only include humongous regions in collection
2250 // sets when concurrent mark shows that their contained object is
2251 // unreachable.
2252
2253 // Do we have any marking information for this region?
2254 if (r->is_marked()) {
2255 // We don't include humongous regions in collection
2256 // sets because we collect them immediately at the end of a marking
2257 // cycle. We also don't include young regions because we *must*
2258 // include them in the next collection pause.
2259 if (!r->isHumongous() && !r->is_young()) {
2260 _hrSorted->addMarkedHeapRegion(r);
2261 }
2262 }
2263 return false;
2264 }
2265 };
2266
2267 class ParKnownGarbageHRClosure: public HeapRegionClosure {
2268 CollectionSetChooser* _hrSorted;
2269 jint _marked_regions_added;
2270 jint _chunk_size;
2271 jint _cur_chunk_idx;
2272 jint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end)
2273 int _worker;
2274 int _invokes;
2275
2276 void get_new_chunk() {
2277 _cur_chunk_idx = _hrSorted->getParMarkedHeapRegionChunk(_chunk_size);
2278 _cur_chunk_end = _cur_chunk_idx + _chunk_size;
2279 }
2280 void add_region(HeapRegion* r) {
2281 if (_cur_chunk_idx == _cur_chunk_end) {
2282 get_new_chunk();
2283 }
2284 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition");
2285 _hrSorted->setMarkedHeapRegion(_cur_chunk_idx, r);
2286 _marked_regions_added++;
2287 _cur_chunk_idx++;
2288 }
2289
2290 public:
2291 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
2292 jint chunk_size,
2293 int worker) :
2294 _hrSorted(hrSorted), _chunk_size(chunk_size), _worker(worker),
2295 _marked_regions_added(0), _cur_chunk_idx(0), _cur_chunk_end(0),
2296 _invokes(0)
2297 {}
2298
2299 bool doHeapRegion(HeapRegion* r) {
2300 // We only include humongous regions in collection
2301 // sets when concurrent mark shows that their contained object is
2302 // unreachable.
2303 _invokes++;
2304
2305 // Do we have any marking information for this region?
2306 if (r->is_marked()) {
2307 // We don't include humongous regions in collection
2308 // sets because we collect them immediately at the end of a marking
2309 // cycle.
2310 // We also do not include young regions in collection sets
2311 if (!r->isHumongous() && !r->is_young()) {
2312 add_region(r);
2313 }
2314 }
2315 return false;
2316 }
2317 jint marked_regions_added() { return _marked_regions_added; }
2318 int invokes() { return _invokes; }
2319 };
2320
2321 class ParKnownGarbageTask: public AbstractGangTask {
2322 CollectionSetChooser* _hrSorted;
2323 jint _chunk_size;
2324 G1CollectedHeap* _g1;
2325 public:
2326 ParKnownGarbageTask(CollectionSetChooser* hrSorted, jint chunk_size) :
2327 AbstractGangTask("ParKnownGarbageTask"),
2328 _hrSorted(hrSorted), _chunk_size(chunk_size),
2329 _g1(G1CollectedHeap::heap())
2330 {}
2331
2332 void work(int i) {
2333 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size, i);
2334 // Back to zero for the claim value.
2335 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, i,
2336 _g1->workers()->active_workers(),
2337 HeapRegion::InitialClaimValue);
2338 jint regions_added = parKnownGarbageCl.marked_regions_added();
2339 _hrSorted->incNumMarkedHeapRegions(regions_added);
2340 if (G1PrintParCleanupStats) {
2341 gclog_or_tty->print_cr(" Thread %d called %d times, added %d regions to list.",
2342 i, parKnownGarbageCl.invokes(), regions_added);
2343 }
2344 }
2345 };
2346
2347 void
2348 G1CollectorPolicy::record_concurrent_mark_cleanup_end(int no_of_gc_threads) {
2349 double start_sec;
2350 if (G1PrintParCleanupStats) {
2351 start_sec = os::elapsedTime();
2352 }
2353
2354 _collectionSetChooser->clearMarkedHeapRegions();
2355 double clear_marked_end_sec;
2356 if (G1PrintParCleanupStats) {
2357 clear_marked_end_sec = os::elapsedTime();
2358 gclog_or_tty->print_cr(" clear marked regions: %8.3f ms.",
2359 (clear_marked_end_sec - start_sec) * 1000.0);
2360 }
2361
2362 if (G1CollectedHeap::use_parallel_gc_threads()) {
2363 const size_t OverpartitionFactor = 4;
2364 size_t WorkUnit;
2365 // The use of MinChunkSize = 8 in the original code
2366 // causes some assertion failures when the total number of
2367 // region is less than 8. The code here tries to fix that.
2368 // Should the original code also be fixed?
2369 if (no_of_gc_threads > 0) {
2370 const size_t MinWorkUnit =
2371 MAX2(_g1->n_regions() / no_of_gc_threads, (size_t) 1U);
2372 WorkUnit =
2373 MAX2(_g1->n_regions() / (no_of_gc_threads * OverpartitionFactor),
2374 MinWorkUnit);
2375 } else {
2376 assert(no_of_gc_threads > 0,
2377 "The active gc workers should be greater than 0");
2378 // In a product build do something reasonable to avoid a crash.
2379 const size_t MinWorkUnit =
2380 MAX2(_g1->n_regions() / ParallelGCThreads, (size_t) 1U);
2381 WorkUnit =
2382 MAX2(_g1->n_regions() / (ParallelGCThreads * OverpartitionFactor),
2383 MinWorkUnit);
2384 }
2385 _collectionSetChooser->prepareForAddMarkedHeapRegionsPar(_g1->n_regions(),
2386 WorkUnit);
2387 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser,
2388 (int) WorkUnit);
2389 _g1->workers()->run_task(&parKnownGarbageTask);
2390
2391 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2392 "sanity check");
2393 } else {
2394 KnownGarbageClosure knownGarbagecl(_collectionSetChooser);
2395 _g1->heap_region_iterate(&knownGarbagecl);
2396 }
2397 double known_garbage_end_sec;
2398 if (G1PrintParCleanupStats) {
2399 known_garbage_end_sec = os::elapsedTime();
2400 gclog_or_tty->print_cr(" compute known garbage: %8.3f ms.",
2401 (known_garbage_end_sec - clear_marked_end_sec) * 1000.0);
2402 }
2403
2404 _collectionSetChooser->sortMarkedHeapRegions();
2405 double end_sec = os::elapsedTime();
2406 if (G1PrintParCleanupStats) {
2407 gclog_or_tty->print_cr(" sorting: %8.3f ms.",
2408 (end_sec - known_garbage_end_sec) * 1000.0);
2409 }
2410
2411 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0;
2412 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
2413 _cur_mark_stop_world_time_ms += elapsed_time_ms;
2414 _prev_collection_pause_end_ms += elapsed_time_ms;
2415 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec, true);
2416 }
2417
2418 // Add the heap region at the head of the non-incremental collection set
2419 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) {
2420 assert(_inc_cset_build_state == Active, "Precondition");
2421 assert(!hr->is_young(), "non-incremental add of young region");
2422
2423 if (_g1->mark_in_progress())
2424 _g1->concurrent_mark()->registerCSetRegion(hr);
2425
2426 assert(!hr->in_collection_set(), "should not already be in the CSet");
2427 hr->set_in_collection_set(true);
2428 hr->set_next_in_collection_set(_collection_set);
2429 _collection_set = hr;
2430 _collection_set_bytes_used_before += hr->used();
2431 _g1->register_region_with_in_cset_fast_test(hr);
2432 size_t rs_length = hr->rem_set()->occupied();
2433 _recorded_rs_lengths += rs_length;
2434 _old_cset_region_length += 1;
2435 }
2436
2437 // Initialize the per-collection-set information
2438 void G1CollectorPolicy::start_incremental_cset_building() {
2439 assert(_inc_cset_build_state == Inactive, "Precondition");
2440
2441 _inc_cset_head = NULL;
2442 _inc_cset_tail = NULL;
2443 _inc_cset_bytes_used_before = 0;
2444
2445 _inc_cset_max_finger = 0;
2446 _inc_cset_recorded_rs_lengths = 0;
2447 _inc_cset_predicted_elapsed_time_ms = 0;
2448 _inc_cset_build_state = Active;
2449 }
2450
2451 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) {
2452 // This routine is used when:
2453 // * adding survivor regions to the incremental cset at the end of an
2454 // evacuation pause,
2455 // * adding the current allocation region to the incremental cset
2456 // when it is retired, and
2457 // * updating existing policy information for a region in the
2458 // incremental cset via young list RSet sampling.
2459 // Therefore this routine may be called at a safepoint by the
2460 // VM thread, or in-between safepoints by mutator threads (when
2461 // retiring the current allocation region) or a concurrent
2462 // refine thread (RSet sampling).
2463
2464 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true);
2465 size_t used_bytes = hr->used();
2466
2467 _inc_cset_recorded_rs_lengths += rs_length;
2468 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms;
2469
2470 _inc_cset_bytes_used_before += used_bytes;
2471
2472 // Cache the values we have added to the aggregated informtion
2473 // in the heap region in case we have to remove this region from
2474 // the incremental collection set, or it is updated by the
2475 // rset sampling code
2476 hr->set_recorded_rs_length(rs_length);
2477 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms);
2478 }
2479
2480 void G1CollectorPolicy::remove_from_incremental_cset_info(HeapRegion* hr) {
2481 // This routine is currently only called as part of the updating of
2482 // existing policy information for regions in the incremental cset that
2483 // is performed by the concurrent refine thread(s) as part of young list
2484 // RSet sampling. Therefore we should not be at a safepoint.
2485
2486 assert(!SafepointSynchronize::is_at_safepoint(), "should not be at safepoint");
2487 assert(hr->is_young(), "it should be");
2488
2489 size_t used_bytes = hr->used();
2490 size_t old_rs_length = hr->recorded_rs_length();
2491 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms();
2492
2493 // Subtract the old recorded/predicted policy information for
2494 // the given heap region from the collection set info.
2495 _inc_cset_recorded_rs_lengths -= old_rs_length;
2496 _inc_cset_predicted_elapsed_time_ms -= old_elapsed_time_ms;
2497
2498 _inc_cset_bytes_used_before -= used_bytes;
2499
2500 // Clear the values cached in the heap region
2501 hr->set_recorded_rs_length(0);
2502 hr->set_predicted_elapsed_time_ms(0);
2503 }
2504
2505 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, size_t new_rs_length) {
2506 // Update the collection set information that is dependent on the new RS length
2507 assert(hr->is_young(), "Precondition");
2508
2509 remove_from_incremental_cset_info(hr);
2510 add_to_incremental_cset_info(hr, new_rs_length);
2511 }
2512
2513 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) {
2514 assert(hr->is_young(), "invariant");
2515 assert(hr->young_index_in_cset() > -1, "should have already been set");
2516 assert(_inc_cset_build_state == Active, "Precondition");
2517
2518 // We need to clear and set the cached recorded/cached collection set
2519 // information in the heap region here (before the region gets added
2520 // to the collection set). An individual heap region's cached values
2521 // are calculated, aggregated with the policy collection set info,
2522 // and cached in the heap region here (initially) and (subsequently)
2523 // by the Young List sampling code.
2524
2525 size_t rs_length = hr->rem_set()->occupied();
2526 add_to_incremental_cset_info(hr, rs_length);
2527
2528 HeapWord* hr_end = hr->end();
2529 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end);
2530
2531 assert(!hr->in_collection_set(), "invariant");
2532 hr->set_in_collection_set(true);
2533 assert( hr->next_in_collection_set() == NULL, "invariant");
2534
2535 _g1->register_region_with_in_cset_fast_test(hr);
2536 }
2537
2538 // Add the region at the RHS of the incremental cset
2539 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) {
2540 // We should only ever be appending survivors at the end of a pause
2541 assert( hr->is_survivor(), "Logic");
2542
2543 // Do the 'common' stuff
2544 add_region_to_incremental_cset_common(hr);
2545
2546 // Now add the region at the right hand side
2547 if (_inc_cset_tail == NULL) {
2548 assert(_inc_cset_head == NULL, "invariant");
2549 _inc_cset_head = hr;
2550 } else {
2551 _inc_cset_tail->set_next_in_collection_set(hr);
2552 }
2553 _inc_cset_tail = hr;
2554 }
2555
2556 // Add the region to the LHS of the incremental cset
2557 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) {
2558 // Survivors should be added to the RHS at the end of a pause
2559 assert(!hr->is_survivor(), "Logic");
2560
2561 // Do the 'common' stuff
2562 add_region_to_incremental_cset_common(hr);
2563
2564 // Add the region at the left hand side
2565 hr->set_next_in_collection_set(_inc_cset_head);
2566 if (_inc_cset_head == NULL) {
2567 assert(_inc_cset_tail == NULL, "Invariant");
2568 _inc_cset_tail = hr;
2569 }
2570 _inc_cset_head = hr;
2571 }
2572
2573 #ifndef PRODUCT
2574 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) {
2575 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be");
2576
2577 st->print_cr("\nCollection_set:");
2578 HeapRegion* csr = list_head;
2579 while (csr != NULL) {
2580 HeapRegion* next = csr->next_in_collection_set();
2581 assert(csr->in_collection_set(), "bad CS");
2582 st->print_cr(" [%08x-%08x], t: %08x, P: %08x, N: %08x, C: %08x, "
2583 "age: %4d, y: %d, surv: %d",
2584 csr->bottom(), csr->end(),
2585 csr->top(),
2586 csr->prev_top_at_mark_start(),
2587 csr->next_top_at_mark_start(),
2588 csr->top_at_conc_mark_count(),
2589 csr->age_in_surv_rate_group_cond(),
2590 csr->is_young(),
2591 csr->is_survivor());
2592 csr = next;
2593 }
2594 }
2595 #endif // !PRODUCT
2596
2597 void G1CollectorPolicy::choose_collection_set(double target_pause_time_ms) {
2598 // Set this here - in case we're not doing young collections.
2599 double non_young_start_time_sec = os::elapsedTime();
2600
2601 YoungList* young_list = _g1->young_list();
2602
2603 guarantee(target_pause_time_ms > 0.0,
2604 err_msg("target_pause_time_ms = %1.6lf should be positive",
2605 target_pause_time_ms));
2606 guarantee(_collection_set == NULL, "Precondition");
2607
2608 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
2609 double predicted_pause_time_ms = base_time_ms;
2610
2611 double time_remaining_ms = target_pause_time_ms - base_time_ms;
2612
2613 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2614 "start choosing CSet",
2615 ergo_format_ms("predicted base time")
2616 ergo_format_ms("remaining time")
2617 ergo_format_ms("target pause time"),
2618 base_time_ms, time_remaining_ms, target_pause_time_ms);
2619
2620 // the 10% and 50% values are arbitrary...
2621 double threshold = 0.10 * target_pause_time_ms;
2622 if (time_remaining_ms < threshold) {
2623 double prev_time_remaining_ms = time_remaining_ms;
2624 time_remaining_ms = 0.50 * target_pause_time_ms;
2625 ergo_verbose3(ErgoCSetConstruction,
2626 "adjust remaining time",
2627 ergo_format_reason("remaining time lower than threshold")
2628 ergo_format_ms("remaining time")
2629 ergo_format_ms("threshold")
2630 ergo_format_ms("adjusted remaining time"),
2631 prev_time_remaining_ms, threshold, time_remaining_ms);
2632 }
2633
2634 size_t expansion_bytes = _g1->expansion_regions() * HeapRegion::GrainBytes;
2635
2636 HeapRegion* hr;
2637 double young_start_time_sec = os::elapsedTime();
2638
2639 _collection_set_bytes_used_before = 0;
2640 _last_gc_was_young = gcs_are_young() ? true : false;
2641
2642 if (_last_gc_was_young) {
2643 ++_young_pause_num;
2644 } else {
2645 ++_mixed_pause_num;
2646 }
2647
2648 // The young list is laid with the survivor regions from the previous
2649 // pause are appended to the RHS of the young list, i.e.
2650 // [Newly Young Regions ++ Survivors from last pause].
2651
2652 size_t survivor_region_length = young_list->survivor_length();
2653 size_t eden_region_length = young_list->length() - survivor_region_length;
2654 init_cset_region_lengths(eden_region_length, survivor_region_length);
2655 hr = young_list->first_survivor_region();
2656 while (hr != NULL) {
2657 assert(hr->is_survivor(), "badly formed young list");
2658 hr->set_young();
2659 hr = hr->get_next_young_region();
2660 }
2661
2662 // Clear the fields that point to the survivor list - they are all young now.
2663 young_list->clear_survivors();
2664
2665 if (_g1->mark_in_progress())
2666 _g1->concurrent_mark()->register_collection_set_finger(_inc_cset_max_finger);
2667
2668 _collection_set = _inc_cset_head;
2669 _collection_set_bytes_used_before = _inc_cset_bytes_used_before;
2670 time_remaining_ms -= _inc_cset_predicted_elapsed_time_ms;
2671 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms;
2672
2673 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2674 "add young regions to CSet",
2675 ergo_format_region("eden")
2676 ergo_format_region("survivors")
2677 ergo_format_ms("predicted young region time"),
2678 eden_region_length, survivor_region_length,
2679 _inc_cset_predicted_elapsed_time_ms);
2680
2681 // The number of recorded young regions is the incremental
2682 // collection set's current size
2683 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths);
2684
2685 double young_end_time_sec = os::elapsedTime();
2686 _recorded_young_cset_choice_time_ms =
2687 (young_end_time_sec - young_start_time_sec) * 1000.0;
2688
2689 // We are doing young collections so reset this.
2690 non_young_start_time_sec = young_end_time_sec;
2691
2692 if (!gcs_are_young()) {
2693 bool should_continue = true;
2694 NumberSeq seq;
2695 double avg_prediction = 100000000000000000.0; // something very large
2696
2697 double prev_predicted_pause_time_ms = predicted_pause_time_ms;
2698 do {
2699 // Note that add_old_region_to_cset() increments the
2700 // _old_cset_region_length field and cset_region_length() returns the
2701 // sum of _eden_cset_region_length, _survivor_cset_region_length, and
2702 // _old_cset_region_length. So, as old regions are added to the
2703 // CSet, _old_cset_region_length will be incremented and
2704 // cset_region_length(), which is used below, will always reflect
2705 // the the total number of regions added up to this point to the CSet.
2706
2707 hr = _collectionSetChooser->getNextMarkedRegion(time_remaining_ms,
2708 avg_prediction);
2709 if (hr != NULL) {
2710 _g1->old_set_remove(hr);
2711 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false);
2712 time_remaining_ms -= predicted_time_ms;
2713 predicted_pause_time_ms += predicted_time_ms;
2714 add_old_region_to_cset(hr);
2715 seq.add(predicted_time_ms);
2716 avg_prediction = seq.avg() + seq.sd();
2717 }
2718
2719 should_continue = true;
2720 if (hr == NULL) {
2721 // No need for an ergo verbose message here,
2722 // getNextMarkRegion() does this when it returns NULL.
2723 should_continue = false;
2724 } else {
2725 if (adaptive_young_list_length()) {
2726 if (time_remaining_ms < 0.0) {
2727 ergo_verbose1(ErgoCSetConstruction,
2728 "stop adding old regions to CSet",
2729 ergo_format_reason("remaining time is lower than 0")
2730 ergo_format_ms("remaining time"),
2731 time_remaining_ms);
2732 should_continue = false;
2733 }
2734 } else {
2735 if (cset_region_length() >= _young_list_fixed_length) {
2736 ergo_verbose2(ErgoCSetConstruction,
2737 "stop adding old regions to CSet",
2738 ergo_format_reason("CSet length reached target")
2739 ergo_format_region("CSet")
2740 ergo_format_region("young target"),
2741 cset_region_length(), _young_list_fixed_length);
2742 should_continue = false;
2743 }
2744 }
2745 }
2746 } while (should_continue);
2747
2748 if (!adaptive_young_list_length() &&
2749 cset_region_length() < _young_list_fixed_length) {
2750 ergo_verbose2(ErgoCSetConstruction,
2751 "request mixed GCs end",
2752 ergo_format_reason("CSet length lower than target")
2753 ergo_format_region("CSet")
2754 ergo_format_region("young target"),
2755 cset_region_length(), _young_list_fixed_length);
2756 _should_revert_to_young_gcs = true;
2757 }
2758
2759 ergo_verbose2(ErgoCSetConstruction | ErgoHigh,
2760 "add old regions to CSet",
2761 ergo_format_region("old")
2762 ergo_format_ms("predicted old region time"),
2763 old_cset_region_length(),
2764 predicted_pause_time_ms - prev_predicted_pause_time_ms);
2765 }
2766
2767 stop_incremental_cset_building();
2768
2769 count_CS_bytes_used();
2770
2771 ergo_verbose5(ErgoCSetConstruction,
2772 "finish choosing CSet",
2773 ergo_format_region("eden")
2774 ergo_format_region("survivors")
2775 ergo_format_region("old")
2776 ergo_format_ms("predicted pause time")
2777 ergo_format_ms("target pause time"),
2778 eden_region_length, survivor_region_length,
2779 old_cset_region_length(),
2780 predicted_pause_time_ms, target_pause_time_ms);
2781
2782 double non_young_end_time_sec = os::elapsedTime();
2783 _recorded_non_young_cset_choice_time_ms =
2784 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0;
2785 }