Wed, 26 Oct 2011 08:44:53 +0200
7102191: G1: assert(_min_desired_young_length <= initial_region_num) failed: Initial young gen size too small
Summary: initial_region_num actually not needed.
Reviewed-by: tonyp, johnc
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 */
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"
39 #define PREDICTIONS_VERBOSE 0
41 // <NEW PREDICTION>
43 // Different defaults for different number of GC threads
44 // They were chosen by running GCOld and SPECjbb on debris with different
45 // numbers of GC threads and choosing them based on the results
47 // all the same
48 static double rs_length_diff_defaults[] = {
49 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
50 };
52 static double cost_per_card_ms_defaults[] = {
53 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015
54 };
56 // all the same
57 static double fully_young_cards_per_entry_ratio_defaults[] = {
58 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0
59 };
61 static double cost_per_entry_ms_defaults[] = {
62 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005
63 };
65 static double cost_per_byte_ms_defaults[] = {
66 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009
67 };
69 // these should be pretty consistent
70 static double constant_other_time_ms_defaults[] = {
71 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0
72 };
75 static double young_other_cost_per_region_ms_defaults[] = {
76 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1
77 };
79 static double non_young_other_cost_per_region_ms_defaults[] = {
80 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30
81 };
83 // </NEW PREDICTION>
85 // Help class for avoiding interleaved logging
86 class LineBuffer: public StackObj {
88 private:
89 static const int BUFFER_LEN = 1024;
90 static const int INDENT_CHARS = 3;
91 char _buffer[BUFFER_LEN];
92 int _indent_level;
93 int _cur;
95 void vappend(const char* format, va_list ap) {
96 int res = vsnprintf(&_buffer[_cur], BUFFER_LEN - _cur, format, ap);
97 if (res != -1) {
98 _cur += res;
99 } else {
100 DEBUG_ONLY(warning("buffer too small in LineBuffer");)
101 _buffer[BUFFER_LEN -1] = 0;
102 _cur = BUFFER_LEN; // vsnprintf above should not add to _buffer if we are called again
103 }
104 }
106 public:
107 explicit LineBuffer(int indent_level): _indent_level(indent_level), _cur(0) {
108 for (; (_cur < BUFFER_LEN && _cur < (_indent_level * INDENT_CHARS)); _cur++) {
109 _buffer[_cur] = ' ';
110 }
111 }
113 #ifndef PRODUCT
114 ~LineBuffer() {
115 assert(_cur == _indent_level * INDENT_CHARS, "pending data in buffer - append_and_print_cr() not called?");
116 }
117 #endif
119 void append(const char* format, ...) {
120 va_list ap;
121 va_start(ap, format);
122 vappend(format, ap);
123 va_end(ap);
124 }
126 void append_and_print_cr(const char* format, ...) {
127 va_list ap;
128 va_start(ap, format);
129 vappend(format, ap);
130 va_end(ap);
131 gclog_or_tty->print_cr("%s", _buffer);
132 _cur = _indent_level * INDENT_CHARS;
133 }
134 };
136 G1CollectorPolicy::G1CollectorPolicy() :
137 _parallel_gc_threads(G1CollectedHeap::use_parallel_gc_threads()
138 ? ParallelGCThreads : 1),
140 _n_pauses(0),
141 _recent_rs_scan_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
142 _recent_pause_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
143 _recent_rs_sizes(new TruncatedSeq(NumPrevPausesForHeuristics)),
144 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
145 _all_pause_times_ms(new NumberSeq()),
146 _stop_world_start(0.0),
147 _all_stop_world_times_ms(new NumberSeq()),
148 _all_yield_times_ms(new NumberSeq()),
149 _using_new_ratio_calculations(false),
151 _all_mod_union_times_ms(new NumberSeq()),
153 _summary(new Summary()),
155 _cur_clear_ct_time_ms(0.0),
157 _cur_ref_proc_time_ms(0.0),
158 _cur_ref_enq_time_ms(0.0),
160 #ifndef PRODUCT
161 _min_clear_cc_time_ms(-1.0),
162 _max_clear_cc_time_ms(-1.0),
163 _cur_clear_cc_time_ms(0.0),
164 _cum_clear_cc_time_ms(0.0),
165 _num_cc_clears(0L),
166 #endif
168 _region_num_young(0),
169 _region_num_tenured(0),
170 _prev_region_num_young(0),
171 _prev_region_num_tenured(0),
173 _aux_num(10),
174 _all_aux_times_ms(new NumberSeq[_aux_num]),
175 _cur_aux_start_times_ms(new double[_aux_num]),
176 _cur_aux_times_ms(new double[_aux_num]),
177 _cur_aux_times_set(new bool[_aux_num]),
179 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
180 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
182 // <NEW PREDICTION>
184 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
185 _prev_collection_pause_end_ms(0.0),
186 _pending_card_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
187 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
188 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
189 _fully_young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
190 _partially_young_cards_per_entry_ratio_seq(
191 new TruncatedSeq(TruncatedSeqLength)),
192 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
193 _partially_young_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
194 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
195 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)),
196 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
197 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
198 _non_young_other_cost_per_region_ms_seq(
199 new TruncatedSeq(TruncatedSeqLength)),
201 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)),
202 _scanned_cards_seq(new TruncatedSeq(TruncatedSeqLength)),
203 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)),
205 _pause_time_target_ms((double) MaxGCPauseMillis),
207 // </NEW PREDICTION>
209 _full_young_gcs(true),
210 _full_young_pause_num(0),
211 _partial_young_pause_num(0),
213 _during_marking(false),
214 _in_marking_window(false),
215 _in_marking_window_im(false),
217 _known_garbage_ratio(0.0),
218 _known_garbage_bytes(0),
220 _young_gc_eff_seq(new TruncatedSeq(TruncatedSeqLength)),
222 _recent_prev_end_times_for_all_gcs_sec(new TruncatedSeq(NumPrevPausesForHeuristics)),
224 _recent_CS_bytes_used_before(new TruncatedSeq(NumPrevPausesForHeuristics)),
225 _recent_CS_bytes_surviving(new TruncatedSeq(NumPrevPausesForHeuristics)),
227 _recent_avg_pause_time_ratio(0.0),
229 _all_full_gc_times_ms(new NumberSeq()),
231 // G1PausesBtwnConcMark defaults to -1
232 // so the hack is to do the cast QQQ FIXME
233 _pauses_btwn_concurrent_mark((size_t)G1PausesBtwnConcMark),
234 _initiate_conc_mark_if_possible(false),
235 _during_initial_mark_pause(false),
236 _should_revert_to_full_young_gcs(false),
237 _last_full_young_gc(false),
239 _eden_bytes_before_gc(0),
240 _survivor_bytes_before_gc(0),
241 _capacity_before_gc(0),
243 _prev_collection_pause_used_at_end_bytes(0),
245 _collection_set(NULL),
246 _collection_set_size(0),
247 _collection_set_bytes_used_before(0),
249 // Incremental CSet attributes
250 _inc_cset_build_state(Inactive),
251 _inc_cset_head(NULL),
252 _inc_cset_tail(NULL),
253 _inc_cset_size(0),
254 _inc_cset_young_index(0),
255 _inc_cset_bytes_used_before(0),
256 _inc_cset_max_finger(NULL),
257 _inc_cset_recorded_young_bytes(0),
258 _inc_cset_recorded_rs_lengths(0),
259 _inc_cset_predicted_elapsed_time_ms(0.0),
260 _inc_cset_predicted_bytes_to_copy(0),
262 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
263 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
264 #endif // _MSC_VER
266 _short_lived_surv_rate_group(new SurvRateGroup(this, "Short Lived",
267 G1YoungSurvRateNumRegionsSummary)),
268 _survivor_surv_rate_group(new SurvRateGroup(this, "Survivor",
269 G1YoungSurvRateNumRegionsSummary)),
270 // add here any more surv rate groups
271 _recorded_survivor_regions(0),
272 _recorded_survivor_head(NULL),
273 _recorded_survivor_tail(NULL),
274 _survivors_age_table(true),
276 _gc_overhead_perc(0.0) {
278 // Set up the region size and associated fields. Given that the
279 // policy is created before the heap, we have to set this up here,
280 // so it's done as soon as possible.
281 HeapRegion::setup_heap_region_size(Arguments::min_heap_size());
282 HeapRegionRemSet::setup_remset_size();
284 G1ErgoVerbose::initialize();
285 if (PrintAdaptiveSizePolicy) {
286 // Currently, we only use a single switch for all the heuristics.
287 G1ErgoVerbose::set_enabled(true);
288 // Given that we don't currently have a verboseness level
289 // parameter, we'll hardcode this to high. This can be easily
290 // changed in the future.
291 G1ErgoVerbose::set_level(ErgoHigh);
292 } else {
293 G1ErgoVerbose::set_enabled(false);
294 }
296 // Verify PLAB sizes
297 const size_t region_size = HeapRegion::GrainWords;
298 if (YoungPLABSize > region_size || OldPLABSize > region_size) {
299 char buffer[128];
300 jio_snprintf(buffer, sizeof(buffer), "%sPLABSize should be at most "SIZE_FORMAT,
301 OldPLABSize > region_size ? "Old" : "Young", region_size);
302 vm_exit_during_initialization(buffer);
303 }
305 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime());
306 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0;
308 _par_last_gc_worker_start_times_ms = new double[_parallel_gc_threads];
309 _par_last_ext_root_scan_times_ms = new double[_parallel_gc_threads];
310 _par_last_mark_stack_scan_times_ms = new double[_parallel_gc_threads];
312 _par_last_update_rs_times_ms = new double[_parallel_gc_threads];
313 _par_last_update_rs_processed_buffers = new double[_parallel_gc_threads];
315 _par_last_scan_rs_times_ms = new double[_parallel_gc_threads];
317 _par_last_obj_copy_times_ms = new double[_parallel_gc_threads];
319 _par_last_termination_times_ms = new double[_parallel_gc_threads];
320 _par_last_termination_attempts = new double[_parallel_gc_threads];
321 _par_last_gc_worker_end_times_ms = new double[_parallel_gc_threads];
322 _par_last_gc_worker_times_ms = new double[_parallel_gc_threads];
323 _par_last_gc_worker_other_times_ms = new double[_parallel_gc_threads];
325 // start conservatively
326 _expensive_region_limit_ms = 0.5 * (double) MaxGCPauseMillis;
328 // <NEW PREDICTION>
330 int index;
331 if (ParallelGCThreads == 0)
332 index = 0;
333 else if (ParallelGCThreads > 8)
334 index = 7;
335 else
336 index = ParallelGCThreads - 1;
338 _pending_card_diff_seq->add(0.0);
339 _rs_length_diff_seq->add(rs_length_diff_defaults[index]);
340 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]);
341 _fully_young_cards_per_entry_ratio_seq->add(
342 fully_young_cards_per_entry_ratio_defaults[index]);
343 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]);
344 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]);
345 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]);
346 _young_other_cost_per_region_ms_seq->add(
347 young_other_cost_per_region_ms_defaults[index]);
348 _non_young_other_cost_per_region_ms_seq->add(
349 non_young_other_cost_per_region_ms_defaults[index]);
351 // </NEW PREDICTION>
353 // Below, we might need to calculate the pause time target based on
354 // the pause interval. When we do so we are going to give G1 maximum
355 // flexibility and allow it to do pauses when it needs to. So, we'll
356 // arrange that the pause interval to be pause time target + 1 to
357 // ensure that a) the pause time target is maximized with respect to
358 // the pause interval and b) we maintain the invariant that pause
359 // time target < pause interval. If the user does not want this
360 // maximum flexibility, they will have to set the pause interval
361 // explicitly.
363 // First make sure that, if either parameter is set, its value is
364 // reasonable.
365 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
366 if (MaxGCPauseMillis < 1) {
367 vm_exit_during_initialization("MaxGCPauseMillis should be "
368 "greater than 0");
369 }
370 }
371 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
372 if (GCPauseIntervalMillis < 1) {
373 vm_exit_during_initialization("GCPauseIntervalMillis should be "
374 "greater than 0");
375 }
376 }
378 // Then, if the pause time target parameter was not set, set it to
379 // the default value.
380 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
381 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
382 // The default pause time target in G1 is 200ms
383 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200);
384 } else {
385 // We do not allow the pause interval to be set without the
386 // pause time target
387 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set "
388 "without setting MaxGCPauseMillis");
389 }
390 }
392 // Then, if the interval parameter was not set, set it according to
393 // the pause time target (this will also deal with the case when the
394 // pause time target is the default value).
395 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
396 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1);
397 }
399 // Finally, make sure that the two parameters are consistent.
400 if (MaxGCPauseMillis >= GCPauseIntervalMillis) {
401 char buffer[256];
402 jio_snprintf(buffer, 256,
403 "MaxGCPauseMillis (%u) should be less than "
404 "GCPauseIntervalMillis (%u)",
405 MaxGCPauseMillis, GCPauseIntervalMillis);
406 vm_exit_during_initialization(buffer);
407 }
409 double max_gc_time = (double) MaxGCPauseMillis / 1000.0;
410 double time_slice = (double) GCPauseIntervalMillis / 1000.0;
411 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time);
412 _sigma = (double) G1ConfidencePercent / 100.0;
414 // start conservatively (around 50ms is about right)
415 _concurrent_mark_remark_times_ms->add(0.05);
416 _concurrent_mark_cleanup_times_ms->add(0.20);
417 _tenuring_threshold = MaxTenuringThreshold;
418 // _max_survivor_regions will be calculated by
419 // update_young_list_target_length() during initialization.
420 _max_survivor_regions = 0;
422 assert(GCTimeRatio > 0,
423 "we should have set it to a default value set_g1_gc_flags() "
424 "if a user set it to 0");
425 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio));
427 uintx reserve_perc = G1ReservePercent;
428 // Put an artificial ceiling on this so that it's not set to a silly value.
429 if (reserve_perc > 50) {
430 reserve_perc = 50;
431 warning("G1ReservePercent is set to a value that is too large, "
432 "it's been updated to %u", reserve_perc);
433 }
434 _reserve_factor = (double) reserve_perc / 100.0;
435 // This will be set when the heap is expanded
436 // for the first time during initialization.
437 _reserve_regions = 0;
439 initialize_all();
440 _collectionSetChooser = new CollectionSetChooser();
441 }
443 // Increment "i", mod "len"
444 static void inc_mod(int& i, int len) {
445 i++; if (i == len) i = 0;
446 }
448 void G1CollectorPolicy::initialize_flags() {
449 set_min_alignment(HeapRegion::GrainBytes);
450 set_max_alignment(GenRemSet::max_alignment_constraint(rem_set_name()));
451 if (SurvivorRatio < 1) {
452 vm_exit_during_initialization("Invalid survivor ratio specified");
453 }
454 CollectorPolicy::initialize_flags();
455 }
457 // The easiest way to deal with the parsing of the NewSize /
458 // MaxNewSize / etc. parameteres is to re-use the code in the
459 // TwoGenerationCollectorPolicy class. This is similar to what
460 // ParallelScavenge does with its GenerationSizer class (see
461 // ParallelScavengeHeap::initialize()). We might change this in the
462 // future, but it's a good start.
463 class G1YoungGenSizer : public TwoGenerationCollectorPolicy {
464 private:
465 size_t size_to_region_num(size_t byte_size) {
466 return MAX2((size_t) 1, byte_size / HeapRegion::GrainBytes);
467 }
469 public:
470 G1YoungGenSizer() {
471 initialize_flags();
472 initialize_size_info();
473 }
474 size_t min_young_region_num() {
475 return size_to_region_num(_min_gen0_size);
476 }
477 size_t initial_young_region_num() {
478 return size_to_region_num(_initial_gen0_size);
479 }
480 size_t max_young_region_num() {
481 return size_to_region_num(_max_gen0_size);
482 }
483 };
485 void G1CollectorPolicy::update_young_list_size_using_newratio(size_t number_of_heap_regions) {
486 assert(number_of_heap_regions > 0, "Heap must be initialized");
487 size_t young_size = number_of_heap_regions / (NewRatio + 1);
488 _min_desired_young_length = young_size;
489 _max_desired_young_length = young_size;
490 }
492 void G1CollectorPolicy::init() {
493 // Set aside an initial future to_space.
494 _g1 = G1CollectedHeap::heap();
496 assert(Heap_lock->owned_by_self(), "Locking discipline.");
498 initialize_gc_policy_counters();
500 G1YoungGenSizer sizer;
501 _min_desired_young_length = sizer.min_young_region_num();
502 _max_desired_young_length = sizer.max_young_region_num();
504 if (FLAG_IS_CMDLINE(NewRatio)) {
505 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) {
506 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio");
507 } else {
508 // Treat NewRatio as a fixed size that is only recalculated when the heap size changes
509 update_young_list_size_using_newratio(_g1->n_regions());
510 _using_new_ratio_calculations = true;
511 }
512 }
514 assert(_min_desired_young_length <= _max_desired_young_length, "Invalid min/max young gen size values");
516 set_adaptive_young_list_length(_min_desired_young_length < _max_desired_young_length);
517 if (adaptive_young_list_length()) {
518 _young_list_fixed_length = 0;
519 } else {
520 assert(_min_desired_young_length == _max_desired_young_length, "Min and max young size differ");
521 _young_list_fixed_length = _min_desired_young_length;
522 }
523 _free_regions_at_end_of_collection = _g1->free_regions();
524 update_young_list_target_length();
525 _prev_eden_capacity = _young_list_target_length * HeapRegion::GrainBytes;
527 // We may immediately start allocating regions and placing them on the
528 // collection set list. Initialize the per-collection set info
529 start_incremental_cset_building();
530 }
532 // Create the jstat counters for the policy.
533 void G1CollectorPolicy::initialize_gc_policy_counters() {
534 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3);
535 }
537 bool G1CollectorPolicy::predict_will_fit(size_t young_length,
538 double base_time_ms,
539 size_t base_free_regions,
540 double target_pause_time_ms) {
541 if (young_length >= base_free_regions) {
542 // end condition 1: not enough space for the young regions
543 return false;
544 }
546 double accum_surv_rate = accum_yg_surv_rate_pred((int)(young_length - 1));
547 size_t bytes_to_copy =
548 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
549 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
550 double young_other_time_ms = predict_young_other_time_ms(young_length);
551 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms;
552 if (pause_time_ms > target_pause_time_ms) {
553 // end condition 2: prediction is over the target pause time
554 return false;
555 }
557 size_t free_bytes =
558 (base_free_regions - young_length) * HeapRegion::GrainBytes;
559 if ((2.0 * sigma()) * (double) bytes_to_copy > (double) free_bytes) {
560 // end condition 3: out-of-space (conservatively!)
561 return false;
562 }
564 // success!
565 return true;
566 }
568 void G1CollectorPolicy::record_new_heap_size(size_t new_number_of_regions) {
569 // re-calculate the necessary reserve
570 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor;
571 // We use ceiling so that if reserve_regions_d is > 0.0 (but
572 // smaller than 1.0) we'll get 1.
573 _reserve_regions = (size_t) ceil(reserve_regions_d);
575 if (_using_new_ratio_calculations) {
576 // -XX:NewRatio was specified so we need to update the
577 // young gen length when the heap size has changed.
578 update_young_list_size_using_newratio(new_number_of_regions);
579 }
580 }
582 size_t G1CollectorPolicy::calculate_young_list_desired_min_length(
583 size_t base_min_length) {
584 size_t desired_min_length = 0;
585 if (adaptive_young_list_length()) {
586 if (_alloc_rate_ms_seq->num() > 3) {
587 double now_sec = os::elapsedTime();
588 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
589 double alloc_rate_ms = predict_alloc_rate_ms();
590 desired_min_length = (size_t) ceil(alloc_rate_ms * when_ms);
591 } else {
592 // otherwise we don't have enough info to make the prediction
593 }
594 }
595 desired_min_length += base_min_length;
596 // make sure we don't go below any user-defined minimum bound
597 return MAX2(_min_desired_young_length, desired_min_length);
598 }
600 size_t G1CollectorPolicy::calculate_young_list_desired_max_length() {
601 // Here, we might want to also take into account any additional
602 // constraints (i.e., user-defined minimum bound). Currently, we
603 // effectively don't set this bound.
604 return _max_desired_young_length;
605 }
607 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) {
608 if (rs_lengths == (size_t) -1) {
609 // if it's set to the default value (-1), we should predict it;
610 // otherwise, use the given value.
611 rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
612 }
614 // Calculate the absolute and desired min bounds.
616 // This is how many young regions we already have (currently: the survivors).
617 size_t base_min_length = recorded_survivor_regions();
618 // This is the absolute minimum young length, which ensures that we
619 // can allocate one eden region in the worst-case.
620 size_t absolute_min_length = base_min_length + 1;
621 size_t desired_min_length =
622 calculate_young_list_desired_min_length(base_min_length);
623 if (desired_min_length < absolute_min_length) {
624 desired_min_length = absolute_min_length;
625 }
627 // Calculate the absolute and desired max bounds.
629 // We will try our best not to "eat" into the reserve.
630 size_t absolute_max_length = 0;
631 if (_free_regions_at_end_of_collection > _reserve_regions) {
632 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions;
633 }
634 size_t desired_max_length = calculate_young_list_desired_max_length();
635 if (desired_max_length > absolute_max_length) {
636 desired_max_length = absolute_max_length;
637 }
639 size_t young_list_target_length = 0;
640 if (adaptive_young_list_length()) {
641 if (full_young_gcs()) {
642 young_list_target_length =
643 calculate_young_list_target_length(rs_lengths,
644 base_min_length,
645 desired_min_length,
646 desired_max_length);
647 _rs_lengths_prediction = rs_lengths;
648 } else {
649 // Don't calculate anything and let the code below bound it to
650 // the desired_min_length, i.e., do the next GC as soon as
651 // possible to maximize how many old regions we can add to it.
652 }
653 } else {
654 if (full_young_gcs()) {
655 young_list_target_length = _young_list_fixed_length;
656 } else {
657 // A bit arbitrary: during partially-young GCs we allocate half
658 // the young regions to try to add old regions to the CSet.
659 young_list_target_length = _young_list_fixed_length / 2;
660 // We choose to accept that we might go under the desired min
661 // length given that we intentionally ask for a smaller young gen.
662 desired_min_length = absolute_min_length;
663 }
664 }
666 // Make sure we don't go over the desired max length, nor under the
667 // desired min length. In case they clash, desired_min_length wins
668 // which is why that test is second.
669 if (young_list_target_length > desired_max_length) {
670 young_list_target_length = desired_max_length;
671 }
672 if (young_list_target_length < desired_min_length) {
673 young_list_target_length = desired_min_length;
674 }
676 assert(young_list_target_length > recorded_survivor_regions(),
677 "we should be able to allocate at least one eden region");
678 assert(young_list_target_length >= absolute_min_length, "post-condition");
679 _young_list_target_length = young_list_target_length;
681 update_max_gc_locker_expansion();
682 }
684 size_t
685 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths,
686 size_t base_min_length,
687 size_t desired_min_length,
688 size_t desired_max_length) {
689 assert(adaptive_young_list_length(), "pre-condition");
690 assert(full_young_gcs(), "only call this for fully-young GCs");
692 // In case some edge-condition makes the desired max length too small...
693 if (desired_max_length <= desired_min_length) {
694 return desired_min_length;
695 }
697 // We'll adjust min_young_length and max_young_length not to include
698 // the already allocated young regions (i.e., so they reflect the
699 // min and max eden regions we'll allocate). The base_min_length
700 // will be reflected in the predictions by the
701 // survivor_regions_evac_time prediction.
702 assert(desired_min_length > base_min_length, "invariant");
703 size_t min_young_length = desired_min_length - base_min_length;
704 assert(desired_max_length > base_min_length, "invariant");
705 size_t max_young_length = desired_max_length - base_min_length;
707 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
708 double survivor_regions_evac_time = predict_survivor_regions_evac_time();
709 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
710 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
711 size_t scanned_cards = predict_young_card_num(adj_rs_lengths);
712 double base_time_ms =
713 predict_base_elapsed_time_ms(pending_cards, scanned_cards) +
714 survivor_regions_evac_time;
715 size_t available_free_regions = _free_regions_at_end_of_collection;
716 size_t base_free_regions = 0;
717 if (available_free_regions > _reserve_regions) {
718 base_free_regions = available_free_regions - _reserve_regions;
719 }
721 // Here, we will make sure that the shortest young length that
722 // makes sense fits within the target pause time.
724 if (predict_will_fit(min_young_length, base_time_ms,
725 base_free_regions, target_pause_time_ms)) {
726 // The shortest young length will fit into the target pause time;
727 // we'll now check whether the absolute maximum number of young
728 // regions will fit in the target pause time. If not, we'll do
729 // a binary search between min_young_length and max_young_length.
730 if (predict_will_fit(max_young_length, base_time_ms,
731 base_free_regions, target_pause_time_ms)) {
732 // The maximum young length will fit into the target pause time.
733 // We are done so set min young length to the maximum length (as
734 // the result is assumed to be returned in min_young_length).
735 min_young_length = max_young_length;
736 } else {
737 // The maximum possible number of young regions will not fit within
738 // the target pause time so we'll search for the optimal
739 // length. The loop invariants are:
740 //
741 // min_young_length < max_young_length
742 // min_young_length is known to fit into the target pause time
743 // max_young_length is known not to fit into the target pause time
744 //
745 // Going into the loop we know the above hold as we've just
746 // checked them. Every time around the loop we check whether
747 // the middle value between min_young_length and
748 // max_young_length fits into the target pause time. If it
749 // does, it becomes the new min. If it doesn't, it becomes
750 // the new max. This way we maintain the loop invariants.
752 assert(min_young_length < max_young_length, "invariant");
753 size_t diff = (max_young_length - min_young_length) / 2;
754 while (diff > 0) {
755 size_t young_length = min_young_length + diff;
756 if (predict_will_fit(young_length, base_time_ms,
757 base_free_regions, target_pause_time_ms)) {
758 min_young_length = young_length;
759 } else {
760 max_young_length = young_length;
761 }
762 assert(min_young_length < max_young_length, "invariant");
763 diff = (max_young_length - min_young_length) / 2;
764 }
765 // The results is min_young_length which, according to the
766 // loop invariants, should fit within the target pause time.
768 // These are the post-conditions of the binary search above:
769 assert(min_young_length < max_young_length,
770 "otherwise we should have discovered that max_young_length "
771 "fits into the pause target and not done the binary search");
772 assert(predict_will_fit(min_young_length, base_time_ms,
773 base_free_regions, target_pause_time_ms),
774 "min_young_length, the result of the binary search, should "
775 "fit into the pause target");
776 assert(!predict_will_fit(min_young_length + 1, base_time_ms,
777 base_free_regions, target_pause_time_ms),
778 "min_young_length, the result of the binary search, should be "
779 "optimal, so no larger length should fit into the pause target");
780 }
781 } else {
782 // Even the minimum length doesn't fit into the pause time
783 // target, return it as the result nevertheless.
784 }
785 return base_min_length + min_young_length;
786 }
788 double G1CollectorPolicy::predict_survivor_regions_evac_time() {
789 double survivor_regions_evac_time = 0.0;
790 for (HeapRegion * r = _recorded_survivor_head;
791 r != NULL && r != _recorded_survivor_tail->get_next_young_region();
792 r = r->get_next_young_region()) {
793 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, true);
794 }
795 return survivor_regions_evac_time;
796 }
798 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() {
799 guarantee( adaptive_young_list_length(), "should not call this otherwise" );
801 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths();
802 if (rs_lengths > _rs_lengths_prediction) {
803 // add 10% to avoid having to recalculate often
804 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000;
805 update_young_list_target_length(rs_lengths_prediction);
806 }
807 }
811 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size,
812 bool is_tlab,
813 bool* gc_overhead_limit_was_exceeded) {
814 guarantee(false, "Not using this policy feature yet.");
815 return NULL;
816 }
818 // This method controls how a collector handles one or more
819 // of its generations being fully allocated.
820 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size,
821 bool is_tlab) {
822 guarantee(false, "Not using this policy feature yet.");
823 return NULL;
824 }
827 #ifndef PRODUCT
828 bool G1CollectorPolicy::verify_young_ages() {
829 HeapRegion* head = _g1->young_list()->first_region();
830 return
831 verify_young_ages(head, _short_lived_surv_rate_group);
832 // also call verify_young_ages on any additional surv rate groups
833 }
835 bool
836 G1CollectorPolicy::verify_young_ages(HeapRegion* head,
837 SurvRateGroup *surv_rate_group) {
838 guarantee( surv_rate_group != NULL, "pre-condition" );
840 const char* name = surv_rate_group->name();
841 bool ret = true;
842 int prev_age = -1;
844 for (HeapRegion* curr = head;
845 curr != NULL;
846 curr = curr->get_next_young_region()) {
847 SurvRateGroup* group = curr->surv_rate_group();
848 if (group == NULL && !curr->is_survivor()) {
849 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name);
850 ret = false;
851 }
853 if (surv_rate_group == group) {
854 int age = curr->age_in_surv_rate_group();
856 if (age < 0) {
857 gclog_or_tty->print_cr("## %s: encountered negative age", name);
858 ret = false;
859 }
861 if (age <= prev_age) {
862 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing "
863 "(%d, %d)", name, age, prev_age);
864 ret = false;
865 }
866 prev_age = age;
867 }
868 }
870 return ret;
871 }
872 #endif // PRODUCT
874 void G1CollectorPolicy::record_full_collection_start() {
875 _cur_collection_start_sec = os::elapsedTime();
876 // Release the future to-space so that it is available for compaction into.
877 _g1->set_full_collection();
878 }
880 void G1CollectorPolicy::record_full_collection_end() {
881 // Consider this like a collection pause for the purposes of allocation
882 // since last pause.
883 double end_sec = os::elapsedTime();
884 double full_gc_time_sec = end_sec - _cur_collection_start_sec;
885 double full_gc_time_ms = full_gc_time_sec * 1000.0;
887 _all_full_gc_times_ms->add(full_gc_time_ms);
889 update_recent_gc_times(end_sec, full_gc_time_ms);
891 _g1->clear_full_collection();
893 // "Nuke" the heuristics that control the fully/partially young GC
894 // transitions and make sure we start with fully young GCs after the
895 // Full GC.
896 set_full_young_gcs(true);
897 _last_full_young_gc = false;
898 _should_revert_to_full_young_gcs = false;
899 clear_initiate_conc_mark_if_possible();
900 clear_during_initial_mark_pause();
901 _known_garbage_bytes = 0;
902 _known_garbage_ratio = 0.0;
903 _in_marking_window = false;
904 _in_marking_window_im = false;
906 _short_lived_surv_rate_group->start_adding_regions();
907 // also call this on any additional surv rate groups
909 record_survivor_regions(0, NULL, NULL);
911 _prev_region_num_young = _region_num_young;
912 _prev_region_num_tenured = _region_num_tenured;
914 _free_regions_at_end_of_collection = _g1->free_regions();
915 // Reset survivors SurvRateGroup.
916 _survivor_surv_rate_group->reset();
917 update_young_list_target_length();
918 _collectionSetChooser->updateAfterFullCollection();
919 }
921 void G1CollectorPolicy::record_stop_world_start() {
922 _stop_world_start = os::elapsedTime();
923 }
925 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec,
926 size_t start_used) {
927 if (PrintGCDetails) {
928 gclog_or_tty->stamp(PrintGCTimeStamps);
929 gclog_or_tty->print("[GC pause");
930 gclog_or_tty->print(" (%s)", full_young_gcs() ? "young" : "partial");
931 }
933 // We only need to do this here as the policy will only be applied
934 // to the GC we're about to start. so, no point is calculating this
935 // every time we calculate / recalculate the target young length.
936 update_survivors_policy();
938 assert(_g1->used() == _g1->recalculate_used(),
939 err_msg("sanity, used: "SIZE_FORMAT" recalculate_used: "SIZE_FORMAT,
940 _g1->used(), _g1->recalculate_used()));
942 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0;
943 _all_stop_world_times_ms->add(s_w_t_ms);
944 _stop_world_start = 0.0;
946 _cur_collection_start_sec = start_time_sec;
947 _cur_collection_pause_used_at_start_bytes = start_used;
948 _cur_collection_pause_used_regions_at_start = _g1->used_regions();
949 _pending_cards = _g1->pending_card_num();
950 _max_pending_cards = _g1->max_pending_card_num();
952 _bytes_in_collection_set_before_gc = 0;
953 _bytes_copied_during_gc = 0;
955 YoungList* young_list = _g1->young_list();
956 _eden_bytes_before_gc = young_list->eden_used_bytes();
957 _survivor_bytes_before_gc = young_list->survivor_used_bytes();
958 _capacity_before_gc = _g1->capacity();
960 #ifdef DEBUG
961 // initialise these to something well known so that we can spot
962 // if they are not set properly
964 for (int i = 0; i < _parallel_gc_threads; ++i) {
965 _par_last_gc_worker_start_times_ms[i] = -1234.0;
966 _par_last_ext_root_scan_times_ms[i] = -1234.0;
967 _par_last_mark_stack_scan_times_ms[i] = -1234.0;
968 _par_last_update_rs_times_ms[i] = -1234.0;
969 _par_last_update_rs_processed_buffers[i] = -1234.0;
970 _par_last_scan_rs_times_ms[i] = -1234.0;
971 _par_last_obj_copy_times_ms[i] = -1234.0;
972 _par_last_termination_times_ms[i] = -1234.0;
973 _par_last_termination_attempts[i] = -1234.0;
974 _par_last_gc_worker_end_times_ms[i] = -1234.0;
975 _par_last_gc_worker_times_ms[i] = -1234.0;
976 _par_last_gc_worker_other_times_ms[i] = -1234.0;
977 }
978 #endif
980 for (int i = 0; i < _aux_num; ++i) {
981 _cur_aux_times_ms[i] = 0.0;
982 _cur_aux_times_set[i] = false;
983 }
985 // These are initialized to zero here and they are set during
986 // the evacuation pause if marking is in progress.
987 _cur_satb_drain_time_ms = 0.0;
988 _last_satb_drain_processed_buffers = 0;
990 _last_young_gc_full = false;
992 // do that for any other surv rate groups
993 _short_lived_surv_rate_group->stop_adding_regions();
994 _survivors_age_table.clear();
996 assert( verify_young_ages(), "region age verification" );
997 }
999 void G1CollectorPolicy::record_mark_closure_time(double mark_closure_time_ms) {
1000 _mark_closure_time_ms = mark_closure_time_ms;
1001 }
1003 void G1CollectorPolicy::record_concurrent_mark_init_end(double
1004 mark_init_elapsed_time_ms) {
1005 _during_marking = true;
1006 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now");
1007 clear_during_initial_mark_pause();
1008 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms;
1009 }
1011 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
1012 _mark_remark_start_sec = os::elapsedTime();
1013 _during_marking = false;
1014 }
1016 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
1017 double end_time_sec = os::elapsedTime();
1018 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
1019 _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
1020 _cur_mark_stop_world_time_ms += elapsed_time_ms;
1021 _prev_collection_pause_end_ms += elapsed_time_ms;
1023 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true);
1024 }
1026 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
1027 _mark_cleanup_start_sec = os::elapsedTime();
1028 }
1030 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
1031 _should_revert_to_full_young_gcs = false;
1032 _last_full_young_gc = true;
1033 _in_marking_window = false;
1034 }
1036 void G1CollectorPolicy::record_concurrent_pause() {
1037 if (_stop_world_start > 0.0) {
1038 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0;
1039 _all_yield_times_ms->add(yield_ms);
1040 }
1041 }
1043 void G1CollectorPolicy::record_concurrent_pause_end() {
1044 }
1046 template<class T>
1047 T sum_of(T* sum_arr, int start, int n, int N) {
1048 T sum = (T)0;
1049 for (int i = 0; i < n; i++) {
1050 int j = (start + i) % N;
1051 sum += sum_arr[j];
1052 }
1053 return sum;
1054 }
1056 void G1CollectorPolicy::print_par_stats(int level,
1057 const char* str,
1058 double* data) {
1059 double min = data[0], max = data[0];
1060 double total = 0.0;
1061 LineBuffer buf(level);
1062 buf.append("[%s (ms):", str);
1063 for (uint i = 0; i < ParallelGCThreads; ++i) {
1064 double val = data[i];
1065 if (val < min)
1066 min = val;
1067 if (val > max)
1068 max = val;
1069 total += val;
1070 buf.append(" %3.1lf", val);
1071 }
1072 buf.append_and_print_cr("");
1073 double avg = total / (double) ParallelGCThreads;
1074 buf.append_and_print_cr(" Avg: %5.1lf, Min: %5.1lf, Max: %5.1lf, Diff: %5.1lf]",
1075 avg, min, max, max - min);
1076 }
1078 void G1CollectorPolicy::print_par_sizes(int level,
1079 const char* str,
1080 double* data) {
1081 double min = data[0], max = data[0];
1082 double total = 0.0;
1083 LineBuffer buf(level);
1084 buf.append("[%s :", str);
1085 for (uint i = 0; i < ParallelGCThreads; ++i) {
1086 double val = data[i];
1087 if (val < min)
1088 min = val;
1089 if (val > max)
1090 max = val;
1091 total += val;
1092 buf.append(" %d", (int) val);
1093 }
1094 buf.append_and_print_cr("");
1095 double avg = total / (double) ParallelGCThreads;
1096 buf.append_and_print_cr(" Sum: %d, Avg: %d, Min: %d, Max: %d, Diff: %d]",
1097 (int)total, (int)avg, (int)min, (int)max, (int)max - (int)min);
1098 }
1100 void G1CollectorPolicy::print_stats(int level,
1101 const char* str,
1102 double value) {
1103 LineBuffer(level).append_and_print_cr("[%s: %5.1lf ms]", str, value);
1104 }
1106 void G1CollectorPolicy::print_stats(int level,
1107 const char* str,
1108 int value) {
1109 LineBuffer(level).append_and_print_cr("[%s: %d]", str, value);
1110 }
1112 double G1CollectorPolicy::avg_value(double* data) {
1113 if (G1CollectedHeap::use_parallel_gc_threads()) {
1114 double ret = 0.0;
1115 for (uint i = 0; i < ParallelGCThreads; ++i) {
1116 ret += data[i];
1117 }
1118 return ret / (double) ParallelGCThreads;
1119 } else {
1120 return data[0];
1121 }
1122 }
1124 double G1CollectorPolicy::max_value(double* data) {
1125 if (G1CollectedHeap::use_parallel_gc_threads()) {
1126 double ret = data[0];
1127 for (uint i = 1; i < ParallelGCThreads; ++i) {
1128 if (data[i] > ret) {
1129 ret = data[i];
1130 }
1131 }
1132 return ret;
1133 } else {
1134 return data[0];
1135 }
1136 }
1138 double G1CollectorPolicy::sum_of_values(double* data) {
1139 if (G1CollectedHeap::use_parallel_gc_threads()) {
1140 double sum = 0.0;
1141 for (uint i = 0; i < ParallelGCThreads; i++) {
1142 sum += data[i];
1143 }
1144 return sum;
1145 } else {
1146 return data[0];
1147 }
1148 }
1150 double G1CollectorPolicy::max_sum(double* data1, double* data2) {
1151 double ret = data1[0] + data2[0];
1153 if (G1CollectedHeap::use_parallel_gc_threads()) {
1154 for (uint i = 1; i < ParallelGCThreads; ++i) {
1155 double data = data1[i] + data2[i];
1156 if (data > ret) {
1157 ret = data;
1158 }
1159 }
1160 }
1161 return ret;
1162 }
1164 // Anything below that is considered to be zero
1165 #define MIN_TIMER_GRANULARITY 0.0000001
1167 void G1CollectorPolicy::record_collection_pause_end() {
1168 double end_time_sec = os::elapsedTime();
1169 double elapsed_ms = _last_pause_time_ms;
1170 bool parallel = G1CollectedHeap::use_parallel_gc_threads();
1171 size_t rs_size =
1172 _cur_collection_pause_used_regions_at_start - collection_set_size();
1173 size_t cur_used_bytes = _g1->used();
1174 assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
1175 bool last_pause_included_initial_mark = false;
1176 bool update_stats = !_g1->evacuation_failed();
1178 #ifndef PRODUCT
1179 if (G1YoungSurvRateVerbose) {
1180 gclog_or_tty->print_cr("");
1181 _short_lived_surv_rate_group->print();
1182 // do that for any other surv rate groups too
1183 }
1184 #endif // PRODUCT
1186 last_pause_included_initial_mark = during_initial_mark_pause();
1187 if (last_pause_included_initial_mark)
1188 record_concurrent_mark_init_end(0.0);
1190 size_t marking_initiating_used_threshold =
1191 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent;
1193 if (!_g1->mark_in_progress() && !_last_full_young_gc) {
1194 assert(!last_pause_included_initial_mark, "invariant");
1195 if (cur_used_bytes > marking_initiating_used_threshold) {
1196 if (cur_used_bytes > _prev_collection_pause_used_at_end_bytes) {
1197 assert(!during_initial_mark_pause(), "we should not see this here");
1199 ergo_verbose3(ErgoConcCycles,
1200 "request concurrent cycle initiation",
1201 ergo_format_reason("occupancy higher than threshold")
1202 ergo_format_byte("occupancy")
1203 ergo_format_byte_perc("threshold"),
1204 cur_used_bytes,
1205 marking_initiating_used_threshold,
1206 (double) InitiatingHeapOccupancyPercent);
1208 // Note: this might have already been set, if during the last
1209 // pause we decided to start a cycle but at the beginning of
1210 // this pause we decided to postpone it. That's OK.
1211 set_initiate_conc_mark_if_possible();
1212 } else {
1213 ergo_verbose2(ErgoConcCycles,
1214 "do not request concurrent cycle initiation",
1215 ergo_format_reason("occupancy lower than previous occupancy")
1216 ergo_format_byte("occupancy")
1217 ergo_format_byte("previous occupancy"),
1218 cur_used_bytes,
1219 _prev_collection_pause_used_at_end_bytes);
1220 }
1221 }
1222 }
1224 _prev_collection_pause_used_at_end_bytes = cur_used_bytes;
1226 _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0,
1227 end_time_sec, false);
1229 guarantee(_cur_collection_pause_used_regions_at_start >=
1230 collection_set_size(),
1231 "Negative RS size?");
1233 // This assert is exempted when we're doing parallel collection pauses,
1234 // because the fragmentation caused by the parallel GC allocation buffers
1235 // can lead to more memory being used during collection than was used
1236 // before. Best leave this out until the fragmentation problem is fixed.
1237 // Pauses in which evacuation failed can also lead to negative
1238 // collections, since no space is reclaimed from a region containing an
1239 // object whose evacuation failed.
1240 // Further, we're now always doing parallel collection. But I'm still
1241 // leaving this here as a placeholder for a more precise assertion later.
1242 // (DLD, 10/05.)
1243 assert((true || parallel) // Always using GC LABs now.
1244 || _g1->evacuation_failed()
1245 || _cur_collection_pause_used_at_start_bytes >= cur_used_bytes,
1246 "Negative collection");
1248 size_t freed_bytes =
1249 _cur_collection_pause_used_at_start_bytes - cur_used_bytes;
1250 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes;
1252 double survival_fraction =
1253 (double)surviving_bytes/
1254 (double)_collection_set_bytes_used_before;
1256 _n_pauses++;
1258 // These values are used to update the summary information that is
1259 // displayed when TraceGen0Time is enabled, and are output as part
1260 // of the PrintGCDetails output, in the non-parallel case.
1262 double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms);
1263 double mark_stack_scan_time = avg_value(_par_last_mark_stack_scan_times_ms);
1264 double update_rs_time = avg_value(_par_last_update_rs_times_ms);
1265 double update_rs_processed_buffers =
1266 sum_of_values(_par_last_update_rs_processed_buffers);
1267 double scan_rs_time = avg_value(_par_last_scan_rs_times_ms);
1268 double obj_copy_time = avg_value(_par_last_obj_copy_times_ms);
1269 double termination_time = avg_value(_par_last_termination_times_ms);
1271 double known_time = ext_root_scan_time +
1272 mark_stack_scan_time +
1273 update_rs_time +
1274 scan_rs_time +
1275 obj_copy_time;
1277 double other_time_ms = elapsed_ms;
1279 // Subtract the SATB drain time. It's initialized to zero at the
1280 // start of the pause and is updated during the pause if marking
1281 // is in progress.
1282 other_time_ms -= _cur_satb_drain_time_ms;
1284 if (parallel) {
1285 other_time_ms -= _cur_collection_par_time_ms;
1286 } else {
1287 other_time_ms -= known_time;
1288 }
1290 // Subtract the time taken to clean the card table from the
1291 // current value of "other time"
1292 other_time_ms -= _cur_clear_ct_time_ms;
1294 // TraceGen0Time and TraceGen1Time summary info updating.
1295 _all_pause_times_ms->add(elapsed_ms);
1297 if (update_stats) {
1298 _recent_rs_scan_times_ms->add(scan_rs_time);
1299 _recent_pause_times_ms->add(elapsed_ms);
1300 _recent_rs_sizes->add(rs_size);
1302 _summary->record_total_time_ms(elapsed_ms);
1303 _summary->record_other_time_ms(other_time_ms);
1305 MainBodySummary* body_summary = _summary->main_body_summary();
1306 assert(body_summary != NULL, "should not be null!");
1308 // This will be non-zero iff marking is currently in progress (i.e.
1309 // _g1->mark_in_progress() == true) and the currrent pause was not
1310 // an initial mark pause. Since the body_summary items are NumberSeqs,
1311 // however, they have to be consistent and updated in lock-step with
1312 // each other. Therefore we unconditionally record the SATB drain
1313 // time - even if it's zero.
1314 body_summary->record_satb_drain_time_ms(_cur_satb_drain_time_ms);
1316 body_summary->record_ext_root_scan_time_ms(ext_root_scan_time);
1317 body_summary->record_mark_stack_scan_time_ms(mark_stack_scan_time);
1318 body_summary->record_update_rs_time_ms(update_rs_time);
1319 body_summary->record_scan_rs_time_ms(scan_rs_time);
1320 body_summary->record_obj_copy_time_ms(obj_copy_time);
1322 if (parallel) {
1323 body_summary->record_parallel_time_ms(_cur_collection_par_time_ms);
1324 body_summary->record_termination_time_ms(termination_time);
1326 double parallel_known_time = known_time + termination_time;
1327 double parallel_other_time = _cur_collection_par_time_ms - parallel_known_time;
1328 body_summary->record_parallel_other_time_ms(parallel_other_time);
1329 }
1331 body_summary->record_mark_closure_time_ms(_mark_closure_time_ms);
1332 body_summary->record_clear_ct_time_ms(_cur_clear_ct_time_ms);
1334 // We exempt parallel collection from this check because Alloc Buffer
1335 // fragmentation can produce negative collections. Same with evac
1336 // failure.
1337 // Further, we're now always doing parallel collection. But I'm still
1338 // leaving this here as a placeholder for a more precise assertion later.
1339 // (DLD, 10/05.
1340 assert((true || parallel)
1341 || _g1->evacuation_failed()
1342 || surviving_bytes <= _collection_set_bytes_used_before,
1343 "Or else negative collection!");
1345 _recent_CS_bytes_used_before->add(_collection_set_bytes_used_before);
1346 _recent_CS_bytes_surviving->add(surviving_bytes);
1348 // this is where we update the allocation rate of the application
1349 double app_time_ms =
1350 (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms);
1351 if (app_time_ms < MIN_TIMER_GRANULARITY) {
1352 // This usually happens due to the timer not having the required
1353 // granularity. Some Linuxes are the usual culprits.
1354 // We'll just set it to something (arbitrarily) small.
1355 app_time_ms = 1.0;
1356 }
1357 size_t regions_allocated =
1358 (_region_num_young - _prev_region_num_young) +
1359 (_region_num_tenured - _prev_region_num_tenured);
1360 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
1361 _alloc_rate_ms_seq->add(alloc_rate_ms);
1362 _prev_region_num_young = _region_num_young;
1363 _prev_region_num_tenured = _region_num_tenured;
1365 double interval_ms =
1366 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
1367 update_recent_gc_times(end_time_sec, elapsed_ms);
1368 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
1369 if (recent_avg_pause_time_ratio() < 0.0 ||
1370 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) {
1371 #ifndef PRODUCT
1372 // Dump info to allow post-facto debugging
1373 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds");
1374 gclog_or_tty->print_cr("-------------------------------------------");
1375 gclog_or_tty->print_cr("Recent GC Times (ms):");
1376 _recent_gc_times_ms->dump();
1377 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec);
1378 _recent_prev_end_times_for_all_gcs_sec->dump();
1379 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f",
1380 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio());
1381 // In debug mode, terminate the JVM if the user wants to debug at this point.
1382 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above");
1383 #endif // !PRODUCT
1384 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
1385 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
1386 if (_recent_avg_pause_time_ratio < 0.0) {
1387 _recent_avg_pause_time_ratio = 0.0;
1388 } else {
1389 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
1390 _recent_avg_pause_time_ratio = 1.0;
1391 }
1392 }
1393 }
1395 for (int i = 0; i < _aux_num; ++i) {
1396 if (_cur_aux_times_set[i]) {
1397 _all_aux_times_ms[i].add(_cur_aux_times_ms[i]);
1398 }
1399 }
1402 if (G1PolicyVerbose > 1) {
1403 gclog_or_tty->print_cr(" Recording collection pause(%d)", _n_pauses);
1404 }
1406 if (G1PolicyVerbose > 1) {
1407 gclog_or_tty->print_cr(" ET: %10.6f ms (avg: %10.6f ms)\n"
1408 " ET-RS: %10.6f ms (avg: %10.6f ms)\n"
1409 " |RS|: " SIZE_FORMAT,
1410 elapsed_ms, recent_avg_time_for_pauses_ms(),
1411 scan_rs_time, recent_avg_time_for_rs_scan_ms(),
1412 rs_size);
1414 gclog_or_tty->print_cr(" Used at start: " SIZE_FORMAT"K"
1415 " At end " SIZE_FORMAT "K\n"
1416 " garbage : " SIZE_FORMAT "K"
1417 " of " SIZE_FORMAT "K\n"
1418 " survival : %6.2f%% (%6.2f%% avg)",
1419 _cur_collection_pause_used_at_start_bytes/K,
1420 _g1->used()/K, freed_bytes/K,
1421 _collection_set_bytes_used_before/K,
1422 survival_fraction*100.0,
1423 recent_avg_survival_fraction()*100.0);
1424 gclog_or_tty->print_cr(" Recent %% gc pause time: %6.2f",
1425 recent_avg_pause_time_ratio() * 100.0);
1426 }
1428 // PrintGCDetails output
1429 if (PrintGCDetails) {
1430 bool print_marking_info =
1431 _g1->mark_in_progress() && !last_pause_included_initial_mark;
1433 gclog_or_tty->print_cr("%s, %1.8lf secs]",
1434 (last_pause_included_initial_mark) ? " (initial-mark)" : "",
1435 elapsed_ms / 1000.0);
1437 if (print_marking_info) {
1438 print_stats(1, "SATB Drain Time", _cur_satb_drain_time_ms);
1439 print_stats(2, "Processed Buffers", _last_satb_drain_processed_buffers);
1440 }
1442 if (parallel) {
1443 print_stats(1, "Parallel Time", _cur_collection_par_time_ms);
1444 print_par_stats(2, "GC Worker Start", _par_last_gc_worker_start_times_ms);
1445 print_par_stats(2, "Ext Root Scanning", _par_last_ext_root_scan_times_ms);
1446 if (print_marking_info) {
1447 print_par_stats(2, "Mark Stack Scanning", _par_last_mark_stack_scan_times_ms);
1448 }
1449 print_par_stats(2, "Update RS", _par_last_update_rs_times_ms);
1450 print_par_sizes(3, "Processed Buffers", _par_last_update_rs_processed_buffers);
1451 print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms);
1452 print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms);
1453 print_par_stats(2, "Termination", _par_last_termination_times_ms);
1454 print_par_sizes(3, "Termination Attempts", _par_last_termination_attempts);
1455 print_par_stats(2, "GC Worker End", _par_last_gc_worker_end_times_ms);
1457 for (int i = 0; i < _parallel_gc_threads; i++) {
1458 _par_last_gc_worker_times_ms[i] = _par_last_gc_worker_end_times_ms[i] - _par_last_gc_worker_start_times_ms[i];
1460 double worker_known_time = _par_last_ext_root_scan_times_ms[i] +
1461 _par_last_mark_stack_scan_times_ms[i] +
1462 _par_last_update_rs_times_ms[i] +
1463 _par_last_scan_rs_times_ms[i] +
1464 _par_last_obj_copy_times_ms[i] +
1465 _par_last_termination_times_ms[i];
1467 _par_last_gc_worker_other_times_ms[i] = _cur_collection_par_time_ms - worker_known_time;
1468 }
1469 print_par_stats(2, "GC Worker", _par_last_gc_worker_times_ms);
1470 print_par_stats(2, "GC Worker Other", _par_last_gc_worker_other_times_ms);
1471 } else {
1472 print_stats(1, "Ext Root Scanning", ext_root_scan_time);
1473 if (print_marking_info) {
1474 print_stats(1, "Mark Stack Scanning", mark_stack_scan_time);
1475 }
1476 print_stats(1, "Update RS", update_rs_time);
1477 print_stats(2, "Processed Buffers", (int)update_rs_processed_buffers);
1478 print_stats(1, "Scan RS", scan_rs_time);
1479 print_stats(1, "Object Copying", obj_copy_time);
1480 }
1481 print_stats(1, "Clear CT", _cur_clear_ct_time_ms);
1482 #ifndef PRODUCT
1483 print_stats(1, "Cur Clear CC", _cur_clear_cc_time_ms);
1484 print_stats(1, "Cum Clear CC", _cum_clear_cc_time_ms);
1485 print_stats(1, "Min Clear CC", _min_clear_cc_time_ms);
1486 print_stats(1, "Max Clear CC", _max_clear_cc_time_ms);
1487 if (_num_cc_clears > 0) {
1488 print_stats(1, "Avg Clear CC", _cum_clear_cc_time_ms / ((double)_num_cc_clears));
1489 }
1490 #endif
1491 print_stats(1, "Other", other_time_ms);
1492 print_stats(2, "Choose CSet", _recorded_young_cset_choice_time_ms);
1493 print_stats(2, "Ref Proc", _cur_ref_proc_time_ms);
1494 print_stats(2, "Ref Enq", _cur_ref_enq_time_ms);
1496 for (int i = 0; i < _aux_num; ++i) {
1497 if (_cur_aux_times_set[i]) {
1498 char buffer[96];
1499 sprintf(buffer, "Aux%d", i);
1500 print_stats(1, buffer, _cur_aux_times_ms[i]);
1501 }
1502 }
1503 }
1505 // Update the efficiency-since-mark vars.
1506 double proc_ms = elapsed_ms * (double) _parallel_gc_threads;
1507 if (elapsed_ms < MIN_TIMER_GRANULARITY) {
1508 // This usually happens due to the timer not having the required
1509 // granularity. Some Linuxes are the usual culprits.
1510 // We'll just set it to something (arbitrarily) small.
1511 proc_ms = 1.0;
1512 }
1513 double cur_efficiency = (double) freed_bytes / proc_ms;
1515 bool new_in_marking_window = _in_marking_window;
1516 bool new_in_marking_window_im = false;
1517 if (during_initial_mark_pause()) {
1518 new_in_marking_window = true;
1519 new_in_marking_window_im = true;
1520 }
1522 if (_last_full_young_gc) {
1523 if (!last_pause_included_initial_mark) {
1524 ergo_verbose2(ErgoPartiallyYoungGCs,
1525 "start partially-young GCs",
1526 ergo_format_byte_perc("known garbage"),
1527 _known_garbage_bytes, _known_garbage_ratio * 100.0);
1528 set_full_young_gcs(false);
1529 } else {
1530 ergo_verbose0(ErgoPartiallyYoungGCs,
1531 "do not start partially-young GCs",
1532 ergo_format_reason("concurrent cycle is about to start"));
1533 }
1534 _last_full_young_gc = false;
1535 }
1537 if ( !_last_young_gc_full ) {
1538 if (_should_revert_to_full_young_gcs) {
1539 ergo_verbose2(ErgoPartiallyYoungGCs,
1540 "end partially-young GCs",
1541 ergo_format_reason("partially-young GCs end requested")
1542 ergo_format_byte_perc("known garbage"),
1543 _known_garbage_bytes, _known_garbage_ratio * 100.0);
1544 set_full_young_gcs(true);
1545 } else if (_known_garbage_ratio < 0.05) {
1546 ergo_verbose3(ErgoPartiallyYoungGCs,
1547 "end partially-young GCs",
1548 ergo_format_reason("known garbage percent lower than threshold")
1549 ergo_format_byte_perc("known garbage")
1550 ergo_format_perc("threshold"),
1551 _known_garbage_bytes, _known_garbage_ratio * 100.0,
1552 0.05 * 100.0);
1553 set_full_young_gcs(true);
1554 } else if (adaptive_young_list_length() &&
1555 (get_gc_eff_factor() * cur_efficiency < predict_young_gc_eff())) {
1556 ergo_verbose5(ErgoPartiallyYoungGCs,
1557 "end partially-young GCs",
1558 ergo_format_reason("current GC efficiency lower than "
1559 "predicted fully-young GC efficiency")
1560 ergo_format_double("GC efficiency factor")
1561 ergo_format_double("current GC efficiency")
1562 ergo_format_double("predicted fully-young GC efficiency")
1563 ergo_format_byte_perc("known garbage"),
1564 get_gc_eff_factor(), cur_efficiency,
1565 predict_young_gc_eff(),
1566 _known_garbage_bytes, _known_garbage_ratio * 100.0);
1567 set_full_young_gcs(true);
1568 }
1569 }
1570 _should_revert_to_full_young_gcs = false;
1572 if (_last_young_gc_full && !_during_marking) {
1573 _young_gc_eff_seq->add(cur_efficiency);
1574 }
1576 _short_lived_surv_rate_group->start_adding_regions();
1577 // do that for any other surv rate groupsx
1579 // <NEW PREDICTION>
1581 if (update_stats) {
1582 double pause_time_ms = elapsed_ms;
1584 size_t diff = 0;
1585 if (_max_pending_cards >= _pending_cards)
1586 diff = _max_pending_cards - _pending_cards;
1587 _pending_card_diff_seq->add((double) diff);
1589 double cost_per_card_ms = 0.0;
1590 if (_pending_cards > 0) {
1591 cost_per_card_ms = update_rs_time / (double) _pending_cards;
1592 _cost_per_card_ms_seq->add(cost_per_card_ms);
1593 }
1595 size_t cards_scanned = _g1->cards_scanned();
1597 double cost_per_entry_ms = 0.0;
1598 if (cards_scanned > 10) {
1599 cost_per_entry_ms = scan_rs_time / (double) cards_scanned;
1600 if (_last_young_gc_full)
1601 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
1602 else
1603 _partially_young_cost_per_entry_ms_seq->add(cost_per_entry_ms);
1604 }
1606 if (_max_rs_lengths > 0) {
1607 double cards_per_entry_ratio =
1608 (double) cards_scanned / (double) _max_rs_lengths;
1609 if (_last_young_gc_full)
1610 _fully_young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1611 else
1612 _partially_young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1613 }
1615 size_t rs_length_diff = _max_rs_lengths - _recorded_rs_lengths;
1616 if (rs_length_diff >= 0)
1617 _rs_length_diff_seq->add((double) rs_length_diff);
1619 size_t copied_bytes = surviving_bytes;
1620 double cost_per_byte_ms = 0.0;
1621 if (copied_bytes > 0) {
1622 cost_per_byte_ms = obj_copy_time / (double) copied_bytes;
1623 if (_in_marking_window)
1624 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
1625 else
1626 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
1627 }
1629 double all_other_time_ms = pause_time_ms -
1630 (update_rs_time + scan_rs_time + obj_copy_time +
1631 _mark_closure_time_ms + termination_time);
1633 double young_other_time_ms = 0.0;
1634 if (_recorded_young_regions > 0) {
1635 young_other_time_ms =
1636 _recorded_young_cset_choice_time_ms +
1637 _recorded_young_free_cset_time_ms;
1638 _young_other_cost_per_region_ms_seq->add(young_other_time_ms /
1639 (double) _recorded_young_regions);
1640 }
1641 double non_young_other_time_ms = 0.0;
1642 if (_recorded_non_young_regions > 0) {
1643 non_young_other_time_ms =
1644 _recorded_non_young_cset_choice_time_ms +
1645 _recorded_non_young_free_cset_time_ms;
1647 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms /
1648 (double) _recorded_non_young_regions);
1649 }
1651 double constant_other_time_ms = all_other_time_ms -
1652 (young_other_time_ms + non_young_other_time_ms);
1653 _constant_other_time_ms_seq->add(constant_other_time_ms);
1655 double survival_ratio = 0.0;
1656 if (_bytes_in_collection_set_before_gc > 0) {
1657 survival_ratio = (double) _bytes_copied_during_gc /
1658 (double) _bytes_in_collection_set_before_gc;
1659 }
1661 _pending_cards_seq->add((double) _pending_cards);
1662 _scanned_cards_seq->add((double) cards_scanned);
1663 _rs_lengths_seq->add((double) _max_rs_lengths);
1665 double expensive_region_limit_ms =
1666 (double) MaxGCPauseMillis - predict_constant_other_time_ms();
1667 if (expensive_region_limit_ms < 0.0) {
1668 // this means that the other time was predicted to be longer than
1669 // than the max pause time
1670 expensive_region_limit_ms = (double) MaxGCPauseMillis;
1671 }
1672 _expensive_region_limit_ms = expensive_region_limit_ms;
1674 if (PREDICTIONS_VERBOSE) {
1675 gclog_or_tty->print_cr("");
1676 gclog_or_tty->print_cr("PREDICTIONS %1.4lf %d "
1677 "REGIONS %d %d %d "
1678 "PENDING_CARDS %d %d "
1679 "CARDS_SCANNED %d %d "
1680 "RS_LENGTHS %d %d "
1681 "RS_UPDATE %1.6lf %1.6lf RS_SCAN %1.6lf %1.6lf "
1682 "SURVIVAL_RATIO %1.6lf %1.6lf "
1683 "OBJECT_COPY %1.6lf %1.6lf OTHER_CONSTANT %1.6lf %1.6lf "
1684 "OTHER_YOUNG %1.6lf %1.6lf "
1685 "OTHER_NON_YOUNG %1.6lf %1.6lf "
1686 "VTIME_DIFF %1.6lf TERMINATION %1.6lf "
1687 "ELAPSED %1.6lf %1.6lf ",
1688 _cur_collection_start_sec,
1689 (!_last_young_gc_full) ? 2 :
1690 (last_pause_included_initial_mark) ? 1 : 0,
1691 _recorded_region_num,
1692 _recorded_young_regions,
1693 _recorded_non_young_regions,
1694 _predicted_pending_cards, _pending_cards,
1695 _predicted_cards_scanned, cards_scanned,
1696 _predicted_rs_lengths, _max_rs_lengths,
1697 _predicted_rs_update_time_ms, update_rs_time,
1698 _predicted_rs_scan_time_ms, scan_rs_time,
1699 _predicted_survival_ratio, survival_ratio,
1700 _predicted_object_copy_time_ms, obj_copy_time,
1701 _predicted_constant_other_time_ms, constant_other_time_ms,
1702 _predicted_young_other_time_ms, young_other_time_ms,
1703 _predicted_non_young_other_time_ms,
1704 non_young_other_time_ms,
1705 _vtime_diff_ms, termination_time,
1706 _predicted_pause_time_ms, elapsed_ms);
1707 }
1709 if (G1PolicyVerbose > 0) {
1710 gclog_or_tty->print_cr("Pause Time, predicted: %1.4lfms (predicted %s), actual: %1.4lfms",
1711 _predicted_pause_time_ms,
1712 (_within_target) ? "within" : "outside",
1713 elapsed_ms);
1714 }
1716 }
1718 _in_marking_window = new_in_marking_window;
1719 _in_marking_window_im = new_in_marking_window_im;
1720 _free_regions_at_end_of_collection = _g1->free_regions();
1721 update_young_list_target_length();
1723 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
1724 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
1725 adjust_concurrent_refinement(update_rs_time, update_rs_processed_buffers, update_rs_time_goal_ms);
1726 // </NEW PREDICTION>
1728 assert(assertMarkedBytesDataOK(), "Marked regions not OK at pause end.");
1729 }
1731 #define EXT_SIZE_FORMAT "%d%s"
1732 #define EXT_SIZE_PARAMS(bytes) \
1733 byte_size_in_proper_unit((bytes)), \
1734 proper_unit_for_byte_size((bytes))
1736 void G1CollectorPolicy::print_heap_transition() {
1737 if (PrintGCDetails) {
1738 YoungList* young_list = _g1->young_list();
1739 size_t eden_bytes = young_list->eden_used_bytes();
1740 size_t survivor_bytes = young_list->survivor_used_bytes();
1741 size_t used_before_gc = _cur_collection_pause_used_at_start_bytes;
1742 size_t used = _g1->used();
1743 size_t capacity = _g1->capacity();
1744 size_t eden_capacity =
1745 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_bytes;
1747 gclog_or_tty->print_cr(
1748 " [Eden: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT") "
1749 "Survivors: "EXT_SIZE_FORMAT"->"EXT_SIZE_FORMAT" "
1750 "Heap: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"
1751 EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")]",
1752 EXT_SIZE_PARAMS(_eden_bytes_before_gc),
1753 EXT_SIZE_PARAMS(_prev_eden_capacity),
1754 EXT_SIZE_PARAMS(eden_bytes),
1755 EXT_SIZE_PARAMS(eden_capacity),
1756 EXT_SIZE_PARAMS(_survivor_bytes_before_gc),
1757 EXT_SIZE_PARAMS(survivor_bytes),
1758 EXT_SIZE_PARAMS(used_before_gc),
1759 EXT_SIZE_PARAMS(_capacity_before_gc),
1760 EXT_SIZE_PARAMS(used),
1761 EXT_SIZE_PARAMS(capacity));
1763 _prev_eden_capacity = eden_capacity;
1764 } else if (PrintGC) {
1765 _g1->print_size_transition(gclog_or_tty,
1766 _cur_collection_pause_used_at_start_bytes,
1767 _g1->used(), _g1->capacity());
1768 }
1769 }
1771 // <NEW PREDICTION>
1773 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1774 double update_rs_processed_buffers,
1775 double goal_ms) {
1776 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1777 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1779 if (G1UseAdaptiveConcRefinement) {
1780 const int k_gy = 3, k_gr = 6;
1781 const double inc_k = 1.1, dec_k = 0.9;
1783 int g = cg1r->green_zone();
1784 if (update_rs_time > goal_ms) {
1785 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1786 } else {
1787 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1788 g = (int)MAX2(g * inc_k, g + 1.0);
1789 }
1790 }
1791 // Change the refinement threads params
1792 cg1r->set_green_zone(g);
1793 cg1r->set_yellow_zone(g * k_gy);
1794 cg1r->set_red_zone(g * k_gr);
1795 cg1r->reinitialize_threads();
1797 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1);
1798 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1799 cg1r->yellow_zone());
1800 // Change the barrier params
1801 dcqs.set_process_completed_threshold(processing_threshold);
1802 dcqs.set_max_completed_queue(cg1r->red_zone());
1803 }
1805 int curr_queue_size = dcqs.completed_buffers_num();
1806 if (curr_queue_size >= cg1r->yellow_zone()) {
1807 dcqs.set_completed_queue_padding(curr_queue_size);
1808 } else {
1809 dcqs.set_completed_queue_padding(0);
1810 }
1811 dcqs.notify_if_necessary();
1812 }
1814 double
1815 G1CollectorPolicy::
1816 predict_young_collection_elapsed_time_ms(size_t adjustment) {
1817 guarantee( adjustment == 0 || adjustment == 1, "invariant" );
1819 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1820 size_t young_num = g1h->young_list()->length();
1821 if (young_num == 0)
1822 return 0.0;
1824 young_num += adjustment;
1825 size_t pending_cards = predict_pending_cards();
1826 size_t rs_lengths = g1h->young_list()->sampled_rs_lengths() +
1827 predict_rs_length_diff();
1828 size_t card_num;
1829 if (full_young_gcs())
1830 card_num = predict_young_card_num(rs_lengths);
1831 else
1832 card_num = predict_non_young_card_num(rs_lengths);
1833 size_t young_byte_size = young_num * HeapRegion::GrainBytes;
1834 double accum_yg_surv_rate =
1835 _short_lived_surv_rate_group->accum_surv_rate(adjustment);
1837 size_t bytes_to_copy =
1838 (size_t) (accum_yg_surv_rate * (double) HeapRegion::GrainBytes);
1840 return
1841 predict_rs_update_time_ms(pending_cards) +
1842 predict_rs_scan_time_ms(card_num) +
1843 predict_object_copy_time_ms(bytes_to_copy) +
1844 predict_young_other_time_ms(young_num) +
1845 predict_constant_other_time_ms();
1846 }
1848 double
1849 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) {
1850 size_t rs_length = predict_rs_length_diff();
1851 size_t card_num;
1852 if (full_young_gcs())
1853 card_num = predict_young_card_num(rs_length);
1854 else
1855 card_num = predict_non_young_card_num(rs_length);
1856 return predict_base_elapsed_time_ms(pending_cards, card_num);
1857 }
1859 double
1860 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
1861 size_t scanned_cards) {
1862 return
1863 predict_rs_update_time_ms(pending_cards) +
1864 predict_rs_scan_time_ms(scanned_cards) +
1865 predict_constant_other_time_ms();
1866 }
1868 double
1869 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
1870 bool young) {
1871 size_t rs_length = hr->rem_set()->occupied();
1872 size_t card_num;
1873 if (full_young_gcs())
1874 card_num = predict_young_card_num(rs_length);
1875 else
1876 card_num = predict_non_young_card_num(rs_length);
1877 size_t bytes_to_copy = predict_bytes_to_copy(hr);
1879 double region_elapsed_time_ms =
1880 predict_rs_scan_time_ms(card_num) +
1881 predict_object_copy_time_ms(bytes_to_copy);
1883 if (young)
1884 region_elapsed_time_ms += predict_young_other_time_ms(1);
1885 else
1886 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
1888 return region_elapsed_time_ms;
1889 }
1891 size_t
1892 G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) {
1893 size_t bytes_to_copy;
1894 if (hr->is_marked())
1895 bytes_to_copy = hr->max_live_bytes();
1896 else {
1897 guarantee( hr->is_young() && hr->age_in_surv_rate_group() != -1,
1898 "invariant" );
1899 int age = hr->age_in_surv_rate_group();
1900 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
1901 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate);
1902 }
1904 return bytes_to_copy;
1905 }
1907 void
1908 G1CollectorPolicy::start_recording_regions() {
1909 _recorded_rs_lengths = 0;
1910 _recorded_young_regions = 0;
1911 _recorded_non_young_regions = 0;
1913 #if PREDICTIONS_VERBOSE
1914 _recorded_marked_bytes = 0;
1915 _recorded_young_bytes = 0;
1916 _predicted_bytes_to_copy = 0;
1917 _predicted_rs_lengths = 0;
1918 _predicted_cards_scanned = 0;
1919 #endif // PREDICTIONS_VERBOSE
1920 }
1922 void
1923 G1CollectorPolicy::record_cset_region_info(HeapRegion* hr, bool young) {
1924 #if PREDICTIONS_VERBOSE
1925 if (!young) {
1926 _recorded_marked_bytes += hr->max_live_bytes();
1927 }
1928 _predicted_bytes_to_copy += predict_bytes_to_copy(hr);
1929 #endif // PREDICTIONS_VERBOSE
1931 size_t rs_length = hr->rem_set()->occupied();
1932 _recorded_rs_lengths += rs_length;
1933 }
1935 void
1936 G1CollectorPolicy::record_non_young_cset_region(HeapRegion* hr) {
1937 assert(!hr->is_young(), "should not call this");
1938 ++_recorded_non_young_regions;
1939 record_cset_region_info(hr, false);
1940 }
1942 void
1943 G1CollectorPolicy::set_recorded_young_regions(size_t n_regions) {
1944 _recorded_young_regions = n_regions;
1945 }
1947 void G1CollectorPolicy::set_recorded_young_bytes(size_t bytes) {
1948 #if PREDICTIONS_VERBOSE
1949 _recorded_young_bytes = bytes;
1950 #endif // PREDICTIONS_VERBOSE
1951 }
1953 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) {
1954 _recorded_rs_lengths = rs_lengths;
1955 }
1957 void G1CollectorPolicy::set_predicted_bytes_to_copy(size_t bytes) {
1958 _predicted_bytes_to_copy = bytes;
1959 }
1961 void
1962 G1CollectorPolicy::end_recording_regions() {
1963 // The _predicted_pause_time_ms field is referenced in code
1964 // not under PREDICTIONS_VERBOSE. Let's initialize it.
1965 _predicted_pause_time_ms = -1.0;
1967 #if PREDICTIONS_VERBOSE
1968 _predicted_pending_cards = predict_pending_cards();
1969 _predicted_rs_lengths = _recorded_rs_lengths + predict_rs_length_diff();
1970 if (full_young_gcs())
1971 _predicted_cards_scanned += predict_young_card_num(_predicted_rs_lengths);
1972 else
1973 _predicted_cards_scanned +=
1974 predict_non_young_card_num(_predicted_rs_lengths);
1975 _recorded_region_num = _recorded_young_regions + _recorded_non_young_regions;
1977 _predicted_rs_update_time_ms =
1978 predict_rs_update_time_ms(_g1->pending_card_num());
1979 _predicted_rs_scan_time_ms =
1980 predict_rs_scan_time_ms(_predicted_cards_scanned);
1981 _predicted_object_copy_time_ms =
1982 predict_object_copy_time_ms(_predicted_bytes_to_copy);
1983 _predicted_constant_other_time_ms =
1984 predict_constant_other_time_ms();
1985 _predicted_young_other_time_ms =
1986 predict_young_other_time_ms(_recorded_young_regions);
1987 _predicted_non_young_other_time_ms =
1988 predict_non_young_other_time_ms(_recorded_non_young_regions);
1990 _predicted_pause_time_ms =
1991 _predicted_rs_update_time_ms +
1992 _predicted_rs_scan_time_ms +
1993 _predicted_object_copy_time_ms +
1994 _predicted_constant_other_time_ms +
1995 _predicted_young_other_time_ms +
1996 _predicted_non_young_other_time_ms;
1997 #endif // PREDICTIONS_VERBOSE
1998 }
2000 void G1CollectorPolicy::check_if_region_is_too_expensive(double
2001 predicted_time_ms) {
2002 // I don't think we need to do this when in young GC mode since
2003 // marking will be initiated next time we hit the soft limit anyway...
2004 if (predicted_time_ms > _expensive_region_limit_ms) {
2005 ergo_verbose2(ErgoPartiallyYoungGCs,
2006 "request partially-young GCs end",
2007 ergo_format_reason("predicted region time higher than threshold")
2008 ergo_format_ms("predicted region time")
2009 ergo_format_ms("threshold"),
2010 predicted_time_ms, _expensive_region_limit_ms);
2011 // no point in doing another partial one
2012 _should_revert_to_full_young_gcs = true;
2013 }
2014 }
2016 // </NEW PREDICTION>
2019 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
2020 double elapsed_ms) {
2021 _recent_gc_times_ms->add(elapsed_ms);
2022 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
2023 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
2024 }
2026 double G1CollectorPolicy::recent_avg_time_for_pauses_ms() {
2027 if (_recent_pause_times_ms->num() == 0) {
2028 return (double) MaxGCPauseMillis;
2029 }
2030 return _recent_pause_times_ms->avg();
2031 }
2033 double G1CollectorPolicy::recent_avg_time_for_rs_scan_ms() {
2034 if (_recent_rs_scan_times_ms->num() == 0) {
2035 return (double)MaxGCPauseMillis/3.0;
2036 }
2037 return _recent_rs_scan_times_ms->avg();
2038 }
2040 int G1CollectorPolicy::number_of_recent_gcs() {
2041 assert(_recent_rs_scan_times_ms->num() ==
2042 _recent_pause_times_ms->num(), "Sequence out of sync");
2043 assert(_recent_pause_times_ms->num() ==
2044 _recent_CS_bytes_used_before->num(), "Sequence out of sync");
2045 assert(_recent_CS_bytes_used_before->num() ==
2046 _recent_CS_bytes_surviving->num(), "Sequence out of sync");
2048 return _recent_pause_times_ms->num();
2049 }
2051 double G1CollectorPolicy::recent_avg_survival_fraction() {
2052 return recent_avg_survival_fraction_work(_recent_CS_bytes_surviving,
2053 _recent_CS_bytes_used_before);
2054 }
2056 double G1CollectorPolicy::last_survival_fraction() {
2057 return last_survival_fraction_work(_recent_CS_bytes_surviving,
2058 _recent_CS_bytes_used_before);
2059 }
2061 double
2062 G1CollectorPolicy::recent_avg_survival_fraction_work(TruncatedSeq* surviving,
2063 TruncatedSeq* before) {
2064 assert(surviving->num() == before->num(), "Sequence out of sync");
2065 if (before->sum() > 0.0) {
2066 double recent_survival_rate = surviving->sum() / before->sum();
2067 // We exempt parallel collection from this check because Alloc Buffer
2068 // fragmentation can produce negative collections.
2069 // Further, we're now always doing parallel collection. But I'm still
2070 // leaving this here as a placeholder for a more precise assertion later.
2071 // (DLD, 10/05.)
2072 assert((true || G1CollectedHeap::use_parallel_gc_threads()) ||
2073 _g1->evacuation_failed() ||
2074 recent_survival_rate <= 1.0, "Or bad frac");
2075 return recent_survival_rate;
2076 } else {
2077 return 1.0; // Be conservative.
2078 }
2079 }
2081 double
2082 G1CollectorPolicy::last_survival_fraction_work(TruncatedSeq* surviving,
2083 TruncatedSeq* before) {
2084 assert(surviving->num() == before->num(), "Sequence out of sync");
2085 if (surviving->num() > 0 && before->last() > 0.0) {
2086 double last_survival_rate = surviving->last() / before->last();
2087 // We exempt parallel collection from this check because Alloc Buffer
2088 // fragmentation can produce negative collections.
2089 // Further, we're now always doing parallel collection. But I'm still
2090 // leaving this here as a placeholder for a more precise assertion later.
2091 // (DLD, 10/05.)
2092 assert((true || G1CollectedHeap::use_parallel_gc_threads()) ||
2093 last_survival_rate <= 1.0, "Or bad frac");
2094 return last_survival_rate;
2095 } else {
2096 return 1.0;
2097 }
2098 }
2100 static const int survival_min_obs = 5;
2101 static double survival_min_obs_limits[] = { 0.9, 0.7, 0.5, 0.3, 0.1 };
2102 static const double min_survival_rate = 0.1;
2104 double
2105 G1CollectorPolicy::conservative_avg_survival_fraction_work(double avg,
2106 double latest) {
2107 double res = avg;
2108 if (number_of_recent_gcs() < survival_min_obs) {
2109 res = MAX2(res, survival_min_obs_limits[number_of_recent_gcs()]);
2110 }
2111 res = MAX2(res, latest);
2112 res = MAX2(res, min_survival_rate);
2113 // In the parallel case, LAB fragmentation can produce "negative
2114 // collections"; so can evac failure. Cap at 1.0
2115 res = MIN2(res, 1.0);
2116 return res;
2117 }
2119 size_t G1CollectorPolicy::expansion_amount() {
2120 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0;
2121 double threshold = _gc_overhead_perc;
2122 if (recent_gc_overhead > threshold) {
2123 // We will double the existing space, or take
2124 // G1ExpandByPercentOfAvailable % of the available expansion
2125 // space, whichever is smaller, bounded below by a minimum
2126 // expansion (unless that's all that's left.)
2127 const size_t min_expand_bytes = 1*M;
2128 size_t reserved_bytes = _g1->max_capacity();
2129 size_t committed_bytes = _g1->capacity();
2130 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
2131 size_t expand_bytes;
2132 size_t expand_bytes_via_pct =
2133 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
2134 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
2135 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
2136 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
2138 ergo_verbose5(ErgoHeapSizing,
2139 "attempt heap expansion",
2140 ergo_format_reason("recent GC overhead higher than "
2141 "threshold after GC")
2142 ergo_format_perc("recent GC overhead")
2143 ergo_format_perc("threshold")
2144 ergo_format_byte("uncommitted")
2145 ergo_format_byte_perc("calculated expansion amount"),
2146 recent_gc_overhead, threshold,
2147 uncommitted_bytes,
2148 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable);
2150 return expand_bytes;
2151 } else {
2152 return 0;
2153 }
2154 }
2156 class CountCSClosure: public HeapRegionClosure {
2157 G1CollectorPolicy* _g1_policy;
2158 public:
2159 CountCSClosure(G1CollectorPolicy* g1_policy) :
2160 _g1_policy(g1_policy) {}
2161 bool doHeapRegion(HeapRegion* r) {
2162 _g1_policy->_bytes_in_collection_set_before_gc += r->used();
2163 return false;
2164 }
2165 };
2167 void G1CollectorPolicy::count_CS_bytes_used() {
2168 CountCSClosure cs_closure(this);
2169 _g1->collection_set_iterate(&cs_closure);
2170 }
2172 void G1CollectorPolicy::print_summary(int level,
2173 const char* str,
2174 NumberSeq* seq) const {
2175 double sum = seq->sum();
2176 LineBuffer(level + 1).append_and_print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)",
2177 str, sum / 1000.0, seq->avg());
2178 }
2180 void G1CollectorPolicy::print_summary_sd(int level,
2181 const char* str,
2182 NumberSeq* seq) const {
2183 print_summary(level, str, seq);
2184 LineBuffer(level + 6).append_and_print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
2185 seq->num(), seq->sd(), seq->maximum());
2186 }
2188 void G1CollectorPolicy::check_other_times(int level,
2189 NumberSeq* other_times_ms,
2190 NumberSeq* calc_other_times_ms) const {
2191 bool should_print = false;
2192 LineBuffer buf(level + 2);
2194 double max_sum = MAX2(fabs(other_times_ms->sum()),
2195 fabs(calc_other_times_ms->sum()));
2196 double min_sum = MIN2(fabs(other_times_ms->sum()),
2197 fabs(calc_other_times_ms->sum()));
2198 double sum_ratio = max_sum / min_sum;
2199 if (sum_ratio > 1.1) {
2200 should_print = true;
2201 buf.append_and_print_cr("## CALCULATED OTHER SUM DOESN'T MATCH RECORDED ###");
2202 }
2204 double max_avg = MAX2(fabs(other_times_ms->avg()),
2205 fabs(calc_other_times_ms->avg()));
2206 double min_avg = MIN2(fabs(other_times_ms->avg()),
2207 fabs(calc_other_times_ms->avg()));
2208 double avg_ratio = max_avg / min_avg;
2209 if (avg_ratio > 1.1) {
2210 should_print = true;
2211 buf.append_and_print_cr("## CALCULATED OTHER AVG DOESN'T MATCH RECORDED ###");
2212 }
2214 if (other_times_ms->sum() < -0.01) {
2215 buf.append_and_print_cr("## RECORDED OTHER SUM IS NEGATIVE ###");
2216 }
2218 if (other_times_ms->avg() < -0.01) {
2219 buf.append_and_print_cr("## RECORDED OTHER AVG IS NEGATIVE ###");
2220 }
2222 if (calc_other_times_ms->sum() < -0.01) {
2223 should_print = true;
2224 buf.append_and_print_cr("## CALCULATED OTHER SUM IS NEGATIVE ###");
2225 }
2227 if (calc_other_times_ms->avg() < -0.01) {
2228 should_print = true;
2229 buf.append_and_print_cr("## CALCULATED OTHER AVG IS NEGATIVE ###");
2230 }
2232 if (should_print)
2233 print_summary(level, "Other(Calc)", calc_other_times_ms);
2234 }
2236 void G1CollectorPolicy::print_summary(PauseSummary* summary) const {
2237 bool parallel = G1CollectedHeap::use_parallel_gc_threads();
2238 MainBodySummary* body_summary = summary->main_body_summary();
2239 if (summary->get_total_seq()->num() > 0) {
2240 print_summary_sd(0, "Evacuation Pauses", summary->get_total_seq());
2241 if (body_summary != NULL) {
2242 print_summary(1, "SATB Drain", body_summary->get_satb_drain_seq());
2243 if (parallel) {
2244 print_summary(1, "Parallel Time", body_summary->get_parallel_seq());
2245 print_summary(2, "Ext Root Scanning", body_summary->get_ext_root_scan_seq());
2246 print_summary(2, "Mark Stack Scanning", body_summary->get_mark_stack_scan_seq());
2247 print_summary(2, "Update RS", body_summary->get_update_rs_seq());
2248 print_summary(2, "Scan RS", body_summary->get_scan_rs_seq());
2249 print_summary(2, "Object Copy", body_summary->get_obj_copy_seq());
2250 print_summary(2, "Termination", body_summary->get_termination_seq());
2251 print_summary(2, "Parallel Other", body_summary->get_parallel_other_seq());
2252 {
2253 NumberSeq* other_parts[] = {
2254 body_summary->get_ext_root_scan_seq(),
2255 body_summary->get_mark_stack_scan_seq(),
2256 body_summary->get_update_rs_seq(),
2257 body_summary->get_scan_rs_seq(),
2258 body_summary->get_obj_copy_seq(),
2259 body_summary->get_termination_seq()
2260 };
2261 NumberSeq calc_other_times_ms(body_summary->get_parallel_seq(),
2262 6, other_parts);
2263 check_other_times(2, body_summary->get_parallel_other_seq(),
2264 &calc_other_times_ms);
2265 }
2266 } else {
2267 print_summary(1, "Ext Root Scanning", body_summary->get_ext_root_scan_seq());
2268 print_summary(1, "Mark Stack Scanning", body_summary->get_mark_stack_scan_seq());
2269 print_summary(1, "Update RS", body_summary->get_update_rs_seq());
2270 print_summary(1, "Scan RS", body_summary->get_scan_rs_seq());
2271 print_summary(1, "Object Copy", body_summary->get_obj_copy_seq());
2272 }
2273 }
2274 print_summary(1, "Mark Closure", body_summary->get_mark_closure_seq());
2275 print_summary(1, "Clear CT", body_summary->get_clear_ct_seq());
2276 print_summary(1, "Other", summary->get_other_seq());
2277 {
2278 if (body_summary != NULL) {
2279 NumberSeq calc_other_times_ms;
2280 if (parallel) {
2281 // parallel
2282 NumberSeq* other_parts[] = {
2283 body_summary->get_satb_drain_seq(),
2284 body_summary->get_parallel_seq(),
2285 body_summary->get_clear_ct_seq()
2286 };
2287 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
2288 3, other_parts);
2289 } else {
2290 // serial
2291 NumberSeq* other_parts[] = {
2292 body_summary->get_satb_drain_seq(),
2293 body_summary->get_update_rs_seq(),
2294 body_summary->get_ext_root_scan_seq(),
2295 body_summary->get_mark_stack_scan_seq(),
2296 body_summary->get_scan_rs_seq(),
2297 body_summary->get_obj_copy_seq()
2298 };
2299 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
2300 6, other_parts);
2301 }
2302 check_other_times(1, summary->get_other_seq(), &calc_other_times_ms);
2303 }
2304 }
2305 } else {
2306 LineBuffer(1).append_and_print_cr("none");
2307 }
2308 LineBuffer(0).append_and_print_cr("");
2309 }
2311 void G1CollectorPolicy::print_tracing_info() const {
2312 if (TraceGen0Time) {
2313 gclog_or_tty->print_cr("ALL PAUSES");
2314 print_summary_sd(0, "Total", _all_pause_times_ms);
2315 gclog_or_tty->print_cr("");
2316 gclog_or_tty->print_cr("");
2317 gclog_or_tty->print_cr(" Full Young GC Pauses: %8d", _full_young_pause_num);
2318 gclog_or_tty->print_cr(" Partial Young GC Pauses: %8d", _partial_young_pause_num);
2319 gclog_or_tty->print_cr("");
2321 gclog_or_tty->print_cr("EVACUATION PAUSES");
2322 print_summary(_summary);
2324 gclog_or_tty->print_cr("MISC");
2325 print_summary_sd(0, "Stop World", _all_stop_world_times_ms);
2326 print_summary_sd(0, "Yields", _all_yield_times_ms);
2327 for (int i = 0; i < _aux_num; ++i) {
2328 if (_all_aux_times_ms[i].num() > 0) {
2329 char buffer[96];
2330 sprintf(buffer, "Aux%d", i);
2331 print_summary_sd(0, buffer, &_all_aux_times_ms[i]);
2332 }
2333 }
2335 size_t all_region_num = _region_num_young + _region_num_tenured;
2336 gclog_or_tty->print_cr(" New Regions %8d, Young %8d (%6.2lf%%), "
2337 "Tenured %8d (%6.2lf%%)",
2338 all_region_num,
2339 _region_num_young,
2340 (double) _region_num_young / (double) all_region_num * 100.0,
2341 _region_num_tenured,
2342 (double) _region_num_tenured / (double) all_region_num * 100.0);
2343 }
2344 if (TraceGen1Time) {
2345 if (_all_full_gc_times_ms->num() > 0) {
2346 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s",
2347 _all_full_gc_times_ms->num(),
2348 _all_full_gc_times_ms->sum() / 1000.0);
2349 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times_ms->avg());
2350 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]",
2351 _all_full_gc_times_ms->sd(),
2352 _all_full_gc_times_ms->maximum());
2353 }
2354 }
2355 }
2357 void G1CollectorPolicy::print_yg_surv_rate_info() const {
2358 #ifndef PRODUCT
2359 _short_lived_surv_rate_group->print_surv_rate_summary();
2360 // add this call for any other surv rate groups
2361 #endif // PRODUCT
2362 }
2364 void G1CollectorPolicy::update_region_num(bool young) {
2365 if (young) {
2366 ++_region_num_young;
2367 } else {
2368 ++_region_num_tenured;
2369 }
2370 }
2372 #ifndef PRODUCT
2373 // for debugging, bit of a hack...
2374 static char*
2375 region_num_to_mbs(int length) {
2376 static char buffer[64];
2377 double bytes = (double) (length * HeapRegion::GrainBytes);
2378 double mbs = bytes / (double) (1024 * 1024);
2379 sprintf(buffer, "%7.2lfMB", mbs);
2380 return buffer;
2381 }
2382 #endif // PRODUCT
2384 size_t G1CollectorPolicy::max_regions(int purpose) {
2385 switch (purpose) {
2386 case GCAllocForSurvived:
2387 return _max_survivor_regions;
2388 case GCAllocForTenured:
2389 return REGIONS_UNLIMITED;
2390 default:
2391 ShouldNotReachHere();
2392 return REGIONS_UNLIMITED;
2393 };
2394 }
2396 void G1CollectorPolicy::update_max_gc_locker_expansion() {
2397 size_t expansion_region_num = 0;
2398 if (GCLockerEdenExpansionPercent > 0) {
2399 double perc = (double) GCLockerEdenExpansionPercent / 100.0;
2400 double expansion_region_num_d = perc * (double) _young_list_target_length;
2401 // We use ceiling so that if expansion_region_num_d is > 0.0 (but
2402 // less than 1.0) we'll get 1.
2403 expansion_region_num = (size_t) ceil(expansion_region_num_d);
2404 } else {
2405 assert(expansion_region_num == 0, "sanity");
2406 }
2407 _young_list_max_length = _young_list_target_length + expansion_region_num;
2408 assert(_young_list_target_length <= _young_list_max_length, "post-condition");
2409 }
2411 // Calculates survivor space parameters.
2412 void G1CollectorPolicy::update_survivors_policy() {
2413 double max_survivor_regions_d =
2414 (double) _young_list_target_length / (double) SurvivorRatio;
2415 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but
2416 // smaller than 1.0) we'll get 1.
2417 _max_survivor_regions = (size_t) ceil(max_survivor_regions_d);
2419 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
2420 HeapRegion::GrainWords * _max_survivor_regions);
2421 }
2423 #ifndef PRODUCT
2424 class HRSortIndexIsOKClosure: public HeapRegionClosure {
2425 CollectionSetChooser* _chooser;
2426 public:
2427 HRSortIndexIsOKClosure(CollectionSetChooser* chooser) :
2428 _chooser(chooser) {}
2430 bool doHeapRegion(HeapRegion* r) {
2431 if (!r->continuesHumongous()) {
2432 assert(_chooser->regionProperlyOrdered(r), "Ought to be.");
2433 }
2434 return false;
2435 }
2436 };
2438 bool G1CollectorPolicy::assertMarkedBytesDataOK() {
2439 HRSortIndexIsOKClosure cl(_collectionSetChooser);
2440 _g1->heap_region_iterate(&cl);
2441 return true;
2442 }
2443 #endif
2445 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(
2446 GCCause::Cause gc_cause) {
2447 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2448 if (!during_cycle) {
2449 ergo_verbose1(ErgoConcCycles,
2450 "request concurrent cycle initiation",
2451 ergo_format_reason("requested by GC cause")
2452 ergo_format_str("GC cause"),
2453 GCCause::to_string(gc_cause));
2454 set_initiate_conc_mark_if_possible();
2455 return true;
2456 } else {
2457 ergo_verbose1(ErgoConcCycles,
2458 "do not request concurrent cycle initiation",
2459 ergo_format_reason("concurrent cycle already in progress")
2460 ergo_format_str("GC cause"),
2461 GCCause::to_string(gc_cause));
2462 return false;
2463 }
2464 }
2466 void
2467 G1CollectorPolicy::decide_on_conc_mark_initiation() {
2468 // We are about to decide on whether this pause will be an
2469 // initial-mark pause.
2471 // First, during_initial_mark_pause() should not be already set. We
2472 // will set it here if we have to. However, it should be cleared by
2473 // the end of the pause (it's only set for the duration of an
2474 // initial-mark pause).
2475 assert(!during_initial_mark_pause(), "pre-condition");
2477 if (initiate_conc_mark_if_possible()) {
2478 // We had noticed on a previous pause that the heap occupancy has
2479 // gone over the initiating threshold and we should start a
2480 // concurrent marking cycle. So we might initiate one.
2482 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2483 if (!during_cycle) {
2484 // The concurrent marking thread is not "during a cycle", i.e.,
2485 // it has completed the last one. So we can go ahead and
2486 // initiate a new cycle.
2488 set_during_initial_mark_pause();
2489 // We do not allow non-full young GCs during marking.
2490 if (!full_young_gcs()) {
2491 set_full_young_gcs(true);
2492 ergo_verbose0(ErgoPartiallyYoungGCs,
2493 "end partially-young GCs",
2494 ergo_format_reason("concurrent cycle is about to start"));
2495 }
2497 // And we can now clear initiate_conc_mark_if_possible() as
2498 // we've already acted on it.
2499 clear_initiate_conc_mark_if_possible();
2501 ergo_verbose0(ErgoConcCycles,
2502 "initiate concurrent cycle",
2503 ergo_format_reason("concurrent cycle initiation requested"));
2504 } else {
2505 // The concurrent marking thread is still finishing up the
2506 // previous cycle. If we start one right now the two cycles
2507 // overlap. In particular, the concurrent marking thread might
2508 // be in the process of clearing the next marking bitmap (which
2509 // we will use for the next cycle if we start one). Starting a
2510 // cycle now will be bad given that parts of the marking
2511 // information might get cleared by the marking thread. And we
2512 // cannot wait for the marking thread to finish the cycle as it
2513 // periodically yields while clearing the next marking bitmap
2514 // and, if it's in a yield point, it's waiting for us to
2515 // finish. So, at this point we will not start a cycle and we'll
2516 // let the concurrent marking thread complete the last one.
2517 ergo_verbose0(ErgoConcCycles,
2518 "do not initiate concurrent cycle",
2519 ergo_format_reason("concurrent cycle already in progress"));
2520 }
2521 }
2522 }
2524 class KnownGarbageClosure: public HeapRegionClosure {
2525 CollectionSetChooser* _hrSorted;
2527 public:
2528 KnownGarbageClosure(CollectionSetChooser* hrSorted) :
2529 _hrSorted(hrSorted)
2530 {}
2532 bool doHeapRegion(HeapRegion* r) {
2533 // We only include humongous regions in collection
2534 // sets when concurrent mark shows that their contained object is
2535 // unreachable.
2537 // Do we have any marking information for this region?
2538 if (r->is_marked()) {
2539 // We don't include humongous regions in collection
2540 // sets because we collect them immediately at the end of a marking
2541 // cycle. We also don't include young regions because we *must*
2542 // include them in the next collection pause.
2543 if (!r->isHumongous() && !r->is_young()) {
2544 _hrSorted->addMarkedHeapRegion(r);
2545 }
2546 }
2547 return false;
2548 }
2549 };
2551 class ParKnownGarbageHRClosure: public HeapRegionClosure {
2552 CollectionSetChooser* _hrSorted;
2553 jint _marked_regions_added;
2554 jint _chunk_size;
2555 jint _cur_chunk_idx;
2556 jint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end)
2557 int _worker;
2558 int _invokes;
2560 void get_new_chunk() {
2561 _cur_chunk_idx = _hrSorted->getParMarkedHeapRegionChunk(_chunk_size);
2562 _cur_chunk_end = _cur_chunk_idx + _chunk_size;
2563 }
2564 void add_region(HeapRegion* r) {
2565 if (_cur_chunk_idx == _cur_chunk_end) {
2566 get_new_chunk();
2567 }
2568 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition");
2569 _hrSorted->setMarkedHeapRegion(_cur_chunk_idx, r);
2570 _marked_regions_added++;
2571 _cur_chunk_idx++;
2572 }
2574 public:
2575 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
2576 jint chunk_size,
2577 int worker) :
2578 _hrSorted(hrSorted), _chunk_size(chunk_size), _worker(worker),
2579 _marked_regions_added(0), _cur_chunk_idx(0), _cur_chunk_end(0),
2580 _invokes(0)
2581 {}
2583 bool doHeapRegion(HeapRegion* r) {
2584 // We only include humongous regions in collection
2585 // sets when concurrent mark shows that their contained object is
2586 // unreachable.
2587 _invokes++;
2589 // Do we have any marking information for this region?
2590 if (r->is_marked()) {
2591 // We don't include humongous regions in collection
2592 // sets because we collect them immediately at the end of a marking
2593 // cycle.
2594 // We also do not include young regions in collection sets
2595 if (!r->isHumongous() && !r->is_young()) {
2596 add_region(r);
2597 }
2598 }
2599 return false;
2600 }
2601 jint marked_regions_added() { return _marked_regions_added; }
2602 int invokes() { return _invokes; }
2603 };
2605 class ParKnownGarbageTask: public AbstractGangTask {
2606 CollectionSetChooser* _hrSorted;
2607 jint _chunk_size;
2608 G1CollectedHeap* _g1;
2609 public:
2610 ParKnownGarbageTask(CollectionSetChooser* hrSorted, jint chunk_size) :
2611 AbstractGangTask("ParKnownGarbageTask"),
2612 _hrSorted(hrSorted), _chunk_size(chunk_size),
2613 _g1(G1CollectedHeap::heap())
2614 {}
2616 void work(int i) {
2617 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size, i);
2618 // Back to zero for the claim value.
2619 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, i,
2620 HeapRegion::InitialClaimValue);
2621 jint regions_added = parKnownGarbageCl.marked_regions_added();
2622 _hrSorted->incNumMarkedHeapRegions(regions_added);
2623 if (G1PrintParCleanupStats) {
2624 gclog_or_tty->print_cr(" Thread %d called %d times, added %d regions to list.",
2625 i, parKnownGarbageCl.invokes(), regions_added);
2626 }
2627 }
2628 };
2630 void
2631 G1CollectorPolicy::record_concurrent_mark_cleanup_end() {
2632 double start_sec;
2633 if (G1PrintParCleanupStats) {
2634 start_sec = os::elapsedTime();
2635 }
2637 _collectionSetChooser->clearMarkedHeapRegions();
2638 double clear_marked_end_sec;
2639 if (G1PrintParCleanupStats) {
2640 clear_marked_end_sec = os::elapsedTime();
2641 gclog_or_tty->print_cr(" clear marked regions: %8.3f ms.",
2642 (clear_marked_end_sec - start_sec) * 1000.0);
2643 }
2645 if (G1CollectedHeap::use_parallel_gc_threads()) {
2646 const size_t OverpartitionFactor = 4;
2647 const size_t MinWorkUnit = 8;
2648 const size_t WorkUnit =
2649 MAX2(_g1->n_regions() / (ParallelGCThreads * OverpartitionFactor),
2650 MinWorkUnit);
2651 _collectionSetChooser->prepareForAddMarkedHeapRegionsPar(_g1->n_regions(),
2652 WorkUnit);
2653 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser,
2654 (int) WorkUnit);
2655 _g1->workers()->run_task(&parKnownGarbageTask);
2657 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2658 "sanity check");
2659 } else {
2660 KnownGarbageClosure knownGarbagecl(_collectionSetChooser);
2661 _g1->heap_region_iterate(&knownGarbagecl);
2662 }
2663 double known_garbage_end_sec;
2664 if (G1PrintParCleanupStats) {
2665 known_garbage_end_sec = os::elapsedTime();
2666 gclog_or_tty->print_cr(" compute known garbage: %8.3f ms.",
2667 (known_garbage_end_sec - clear_marked_end_sec) * 1000.0);
2668 }
2670 _collectionSetChooser->sortMarkedHeapRegions();
2671 double end_sec = os::elapsedTime();
2672 if (G1PrintParCleanupStats) {
2673 gclog_or_tty->print_cr(" sorting: %8.3f ms.",
2674 (end_sec - known_garbage_end_sec) * 1000.0);
2675 }
2677 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0;
2678 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
2679 _cur_mark_stop_world_time_ms += elapsed_time_ms;
2680 _prev_collection_pause_end_ms += elapsed_time_ms;
2681 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec, true);
2682 }
2684 // Add the heap region at the head of the non-incremental collection set
2685 void G1CollectorPolicy::
2686 add_to_collection_set(HeapRegion* hr) {
2687 assert(_inc_cset_build_state == Active, "Precondition");
2688 assert(!hr->is_young(), "non-incremental add of young region");
2690 if (_g1->mark_in_progress())
2691 _g1->concurrent_mark()->registerCSetRegion(hr);
2693 assert(!hr->in_collection_set(), "should not already be in the CSet");
2694 hr->set_in_collection_set(true);
2695 hr->set_next_in_collection_set(_collection_set);
2696 _collection_set = hr;
2697 _collection_set_size++;
2698 _collection_set_bytes_used_before += hr->used();
2699 _g1->register_region_with_in_cset_fast_test(hr);
2700 }
2702 // Initialize the per-collection-set information
2703 void G1CollectorPolicy::start_incremental_cset_building() {
2704 assert(_inc_cset_build_state == Inactive, "Precondition");
2706 _inc_cset_head = NULL;
2707 _inc_cset_tail = NULL;
2708 _inc_cset_size = 0;
2709 _inc_cset_bytes_used_before = 0;
2711 _inc_cset_young_index = 0;
2713 _inc_cset_max_finger = 0;
2714 _inc_cset_recorded_young_bytes = 0;
2715 _inc_cset_recorded_rs_lengths = 0;
2716 _inc_cset_predicted_elapsed_time_ms = 0;
2717 _inc_cset_predicted_bytes_to_copy = 0;
2718 _inc_cset_build_state = Active;
2719 }
2721 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) {
2722 // This routine is used when:
2723 // * adding survivor regions to the incremental cset at the end of an
2724 // evacuation pause,
2725 // * adding the current allocation region to the incremental cset
2726 // when it is retired, and
2727 // * updating existing policy information for a region in the
2728 // incremental cset via young list RSet sampling.
2729 // Therefore this routine may be called at a safepoint by the
2730 // VM thread, or in-between safepoints by mutator threads (when
2731 // retiring the current allocation region) or a concurrent
2732 // refine thread (RSet sampling).
2734 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true);
2735 size_t used_bytes = hr->used();
2737 _inc_cset_recorded_rs_lengths += rs_length;
2738 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms;
2740 _inc_cset_bytes_used_before += used_bytes;
2742 // Cache the values we have added to the aggregated informtion
2743 // in the heap region in case we have to remove this region from
2744 // the incremental collection set, or it is updated by the
2745 // rset sampling code
2746 hr->set_recorded_rs_length(rs_length);
2747 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms);
2749 #if PREDICTIONS_VERBOSE
2750 size_t bytes_to_copy = predict_bytes_to_copy(hr);
2751 _inc_cset_predicted_bytes_to_copy += bytes_to_copy;
2753 // Record the number of bytes used in this region
2754 _inc_cset_recorded_young_bytes += used_bytes;
2756 // Cache the values we have added to the aggregated informtion
2757 // in the heap region in case we have to remove this region from
2758 // the incremental collection set, or it is updated by the
2759 // rset sampling code
2760 hr->set_predicted_bytes_to_copy(bytes_to_copy);
2761 #endif // PREDICTIONS_VERBOSE
2762 }
2764 void G1CollectorPolicy::remove_from_incremental_cset_info(HeapRegion* hr) {
2765 // This routine is currently only called as part of the updating of
2766 // existing policy information for regions in the incremental cset that
2767 // is performed by the concurrent refine thread(s) as part of young list
2768 // RSet sampling. Therefore we should not be at a safepoint.
2770 assert(!SafepointSynchronize::is_at_safepoint(), "should not be at safepoint");
2771 assert(hr->is_young(), "it should be");
2773 size_t used_bytes = hr->used();
2774 size_t old_rs_length = hr->recorded_rs_length();
2775 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms();
2777 // Subtract the old recorded/predicted policy information for
2778 // the given heap region from the collection set info.
2779 _inc_cset_recorded_rs_lengths -= old_rs_length;
2780 _inc_cset_predicted_elapsed_time_ms -= old_elapsed_time_ms;
2782 _inc_cset_bytes_used_before -= used_bytes;
2784 // Clear the values cached in the heap region
2785 hr->set_recorded_rs_length(0);
2786 hr->set_predicted_elapsed_time_ms(0);
2788 #if PREDICTIONS_VERBOSE
2789 size_t old_predicted_bytes_to_copy = hr->predicted_bytes_to_copy();
2790 _inc_cset_predicted_bytes_to_copy -= old_predicted_bytes_to_copy;
2792 // Subtract the number of bytes used in this region
2793 _inc_cset_recorded_young_bytes -= used_bytes;
2795 // Clear the values cached in the heap region
2796 hr->set_predicted_bytes_to_copy(0);
2797 #endif // PREDICTIONS_VERBOSE
2798 }
2800 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, size_t new_rs_length) {
2801 // Update the collection set information that is dependent on the new RS length
2802 assert(hr->is_young(), "Precondition");
2804 remove_from_incremental_cset_info(hr);
2805 add_to_incremental_cset_info(hr, new_rs_length);
2806 }
2808 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) {
2809 assert( hr->is_young(), "invariant");
2810 assert( hr->young_index_in_cset() == -1, "invariant" );
2811 assert(_inc_cset_build_state == Active, "Precondition");
2813 // We need to clear and set the cached recorded/cached collection set
2814 // information in the heap region here (before the region gets added
2815 // to the collection set). An individual heap region's cached values
2816 // are calculated, aggregated with the policy collection set info,
2817 // and cached in the heap region here (initially) and (subsequently)
2818 // by the Young List sampling code.
2820 size_t rs_length = hr->rem_set()->occupied();
2821 add_to_incremental_cset_info(hr, rs_length);
2823 HeapWord* hr_end = hr->end();
2824 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end);
2826 assert(!hr->in_collection_set(), "invariant");
2827 hr->set_in_collection_set(true);
2828 assert( hr->next_in_collection_set() == NULL, "invariant");
2830 _inc_cset_size++;
2831 _g1->register_region_with_in_cset_fast_test(hr);
2833 hr->set_young_index_in_cset((int) _inc_cset_young_index);
2834 ++_inc_cset_young_index;
2835 }
2837 // Add the region at the RHS of the incremental cset
2838 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) {
2839 // We should only ever be appending survivors at the end of a pause
2840 assert( hr->is_survivor(), "Logic");
2842 // Do the 'common' stuff
2843 add_region_to_incremental_cset_common(hr);
2845 // Now add the region at the right hand side
2846 if (_inc_cset_tail == NULL) {
2847 assert(_inc_cset_head == NULL, "invariant");
2848 _inc_cset_head = hr;
2849 } else {
2850 _inc_cset_tail->set_next_in_collection_set(hr);
2851 }
2852 _inc_cset_tail = hr;
2853 }
2855 // Add the region to the LHS of the incremental cset
2856 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) {
2857 // Survivors should be added to the RHS at the end of a pause
2858 assert(!hr->is_survivor(), "Logic");
2860 // Do the 'common' stuff
2861 add_region_to_incremental_cset_common(hr);
2863 // Add the region at the left hand side
2864 hr->set_next_in_collection_set(_inc_cset_head);
2865 if (_inc_cset_head == NULL) {
2866 assert(_inc_cset_tail == NULL, "Invariant");
2867 _inc_cset_tail = hr;
2868 }
2869 _inc_cset_head = hr;
2870 }
2872 #ifndef PRODUCT
2873 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) {
2874 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be");
2876 st->print_cr("\nCollection_set:");
2877 HeapRegion* csr = list_head;
2878 while (csr != NULL) {
2879 HeapRegion* next = csr->next_in_collection_set();
2880 assert(csr->in_collection_set(), "bad CS");
2881 st->print_cr(" [%08x-%08x], t: %08x, P: %08x, N: %08x, C: %08x, "
2882 "age: %4d, y: %d, surv: %d",
2883 csr->bottom(), csr->end(),
2884 csr->top(),
2885 csr->prev_top_at_mark_start(),
2886 csr->next_top_at_mark_start(),
2887 csr->top_at_conc_mark_count(),
2888 csr->age_in_surv_rate_group_cond(),
2889 csr->is_young(),
2890 csr->is_survivor());
2891 csr = next;
2892 }
2893 }
2894 #endif // !PRODUCT
2896 void G1CollectorPolicy::choose_collection_set(double target_pause_time_ms) {
2897 // Set this here - in case we're not doing young collections.
2898 double non_young_start_time_sec = os::elapsedTime();
2900 YoungList* young_list = _g1->young_list();
2902 start_recording_regions();
2904 guarantee(target_pause_time_ms > 0.0,
2905 err_msg("target_pause_time_ms = %1.6lf should be positive",
2906 target_pause_time_ms));
2907 guarantee(_collection_set == NULL, "Precondition");
2909 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
2910 double predicted_pause_time_ms = base_time_ms;
2912 double time_remaining_ms = target_pause_time_ms - base_time_ms;
2914 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2915 "start choosing CSet",
2916 ergo_format_ms("predicted base time")
2917 ergo_format_ms("remaining time")
2918 ergo_format_ms("target pause time"),
2919 base_time_ms, time_remaining_ms, target_pause_time_ms);
2921 // the 10% and 50% values are arbitrary...
2922 double threshold = 0.10 * target_pause_time_ms;
2923 if (time_remaining_ms < threshold) {
2924 double prev_time_remaining_ms = time_remaining_ms;
2925 time_remaining_ms = 0.50 * target_pause_time_ms;
2926 _within_target = false;
2927 ergo_verbose3(ErgoCSetConstruction,
2928 "adjust remaining time",
2929 ergo_format_reason("remaining time lower than threshold")
2930 ergo_format_ms("remaining time")
2931 ergo_format_ms("threshold")
2932 ergo_format_ms("adjusted remaining time"),
2933 prev_time_remaining_ms, threshold, time_remaining_ms);
2934 } else {
2935 _within_target = true;
2936 }
2938 size_t expansion_bytes = _g1->expansion_regions() * HeapRegion::GrainBytes;
2940 HeapRegion* hr;
2941 double young_start_time_sec = os::elapsedTime();
2943 _collection_set_bytes_used_before = 0;
2944 _collection_set_size = 0;
2945 _young_cset_length = 0;
2946 _last_young_gc_full = full_young_gcs() ? true : false;
2948 if (_last_young_gc_full) {
2949 ++_full_young_pause_num;
2950 } else {
2951 ++_partial_young_pause_num;
2952 }
2954 // The young list is laid with the survivor regions from the previous
2955 // pause are appended to the RHS of the young list, i.e.
2956 // [Newly Young Regions ++ Survivors from last pause].
2958 size_t survivor_region_num = young_list->survivor_length();
2959 size_t eden_region_num = young_list->length() - survivor_region_num;
2960 size_t old_region_num = 0;
2961 hr = young_list->first_survivor_region();
2962 while (hr != NULL) {
2963 assert(hr->is_survivor(), "badly formed young list");
2964 hr->set_young();
2965 hr = hr->get_next_young_region();
2966 }
2968 // Clear the fields that point to the survivor list - they are all young now.
2969 young_list->clear_survivors();
2971 if (_g1->mark_in_progress())
2972 _g1->concurrent_mark()->register_collection_set_finger(_inc_cset_max_finger);
2974 _young_cset_length = _inc_cset_young_index;
2975 _collection_set = _inc_cset_head;
2976 _collection_set_size = _inc_cset_size;
2977 _collection_set_bytes_used_before = _inc_cset_bytes_used_before;
2978 time_remaining_ms -= _inc_cset_predicted_elapsed_time_ms;
2979 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms;
2981 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2982 "add young regions to CSet",
2983 ergo_format_region("eden")
2984 ergo_format_region("survivors")
2985 ergo_format_ms("predicted young region time"),
2986 eden_region_num, survivor_region_num,
2987 _inc_cset_predicted_elapsed_time_ms);
2989 // The number of recorded young regions is the incremental
2990 // collection set's current size
2991 set_recorded_young_regions(_inc_cset_size);
2992 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths);
2993 set_recorded_young_bytes(_inc_cset_recorded_young_bytes);
2994 #if PREDICTIONS_VERBOSE
2995 set_predicted_bytes_to_copy(_inc_cset_predicted_bytes_to_copy);
2996 #endif // PREDICTIONS_VERBOSE
2998 assert(_inc_cset_size == young_list->length(), "Invariant");
3000 double young_end_time_sec = os::elapsedTime();
3001 _recorded_young_cset_choice_time_ms =
3002 (young_end_time_sec - young_start_time_sec) * 1000.0;
3004 // We are doing young collections so reset this.
3005 non_young_start_time_sec = young_end_time_sec;
3007 if (!full_young_gcs()) {
3008 bool should_continue = true;
3009 NumberSeq seq;
3010 double avg_prediction = 100000000000000000.0; // something very large
3012 size_t prev_collection_set_size = _collection_set_size;
3013 double prev_predicted_pause_time_ms = predicted_pause_time_ms;
3014 do {
3015 hr = _collectionSetChooser->getNextMarkedRegion(time_remaining_ms,
3016 avg_prediction);
3017 if (hr != NULL) {
3018 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false);
3019 time_remaining_ms -= predicted_time_ms;
3020 predicted_pause_time_ms += predicted_time_ms;
3021 add_to_collection_set(hr);
3022 record_non_young_cset_region(hr);
3023 seq.add(predicted_time_ms);
3024 avg_prediction = seq.avg() + seq.sd();
3025 }
3027 should_continue = true;
3028 if (hr == NULL) {
3029 // No need for an ergo verbose message here,
3030 // getNextMarkRegion() does this when it returns NULL.
3031 should_continue = false;
3032 } else {
3033 if (adaptive_young_list_length()) {
3034 if (time_remaining_ms < 0.0) {
3035 ergo_verbose1(ErgoCSetConstruction,
3036 "stop adding old regions to CSet",
3037 ergo_format_reason("remaining time is lower than 0")
3038 ergo_format_ms("remaining time"),
3039 time_remaining_ms);
3040 should_continue = false;
3041 }
3042 } else {
3043 if (_collection_set_size >= _young_list_fixed_length) {
3044 ergo_verbose2(ErgoCSetConstruction,
3045 "stop adding old regions to CSet",
3046 ergo_format_reason("CSet length reached target")
3047 ergo_format_region("CSet")
3048 ergo_format_region("young target"),
3049 _collection_set_size, _young_list_fixed_length);
3050 should_continue = false;
3051 }
3052 }
3053 }
3054 } while (should_continue);
3056 if (!adaptive_young_list_length() &&
3057 _collection_set_size < _young_list_fixed_length) {
3058 ergo_verbose2(ErgoCSetConstruction,
3059 "request partially-young GCs end",
3060 ergo_format_reason("CSet length lower than target")
3061 ergo_format_region("CSet")
3062 ergo_format_region("young target"),
3063 _collection_set_size, _young_list_fixed_length);
3064 _should_revert_to_full_young_gcs = true;
3065 }
3067 old_region_num = _collection_set_size - prev_collection_set_size;
3069 ergo_verbose2(ErgoCSetConstruction | ErgoHigh,
3070 "add old regions to CSet",
3071 ergo_format_region("old")
3072 ergo_format_ms("predicted old region time"),
3073 old_region_num,
3074 predicted_pause_time_ms - prev_predicted_pause_time_ms);
3075 }
3077 stop_incremental_cset_building();
3079 count_CS_bytes_used();
3081 end_recording_regions();
3083 ergo_verbose5(ErgoCSetConstruction,
3084 "finish choosing CSet",
3085 ergo_format_region("eden")
3086 ergo_format_region("survivors")
3087 ergo_format_region("old")
3088 ergo_format_ms("predicted pause time")
3089 ergo_format_ms("target pause time"),
3090 eden_region_num, survivor_region_num, old_region_num,
3091 predicted_pause_time_ms, target_pause_time_ms);
3093 double non_young_end_time_sec = os::elapsedTime();
3094 _recorded_non_young_cset_choice_time_ms =
3095 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0;
3096 }